Spatial Multiomic Profiling Identifies Metabolic and Inflammatory Signatures Driving Recurrence in Adjuvant-Treated NSCLC

Spatial Multiomic Profiling Identifies Metabolic and Inflammatory Signatures Driving Recurrence in Adjuvant-Treated NSCLC | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article Spatial Multiomic Profiling Identifies Metabolic and Inflammatory Signatures Driving Recurrence in Adjuvant-Treated NSCLC Arutha Kulasinghe, Naomi Berrell, Chin Wee Tan, Aaron Kilgallon, and 13 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-7692105/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted You are reading this latest preprint version Abstract While there have been therapeutic advancements in treatments for non-small cell lung cancer (NSCLC), we have a limited understanding of how the tumour composition and spatial heterogeneity influences clinical outcomes. Here, we performed a multiomic analysis of 61 NSCLC patients treated with adjuvant chemotherapy and curative resection, integrating spatial transcriptomics, spatial proteomics, and deep learning to profile the tumour microenvironment (TME). We identified spatial-metabolic alterations associated with disease recurrence in and between cells in the TME, involving glutamine import, glycolytic, and lipid oxidation pathways. Based on protein expression, we found spatially distinct regions linked to tumour recurrence, implicating specific roles for tumour associated macrophages, metabolic pathways, and major histocompatibility complex (MHC) proteins. Moreover, transcriptomic analysis revealed histology-specific gene expression changes and metabolic and inflammation pathway associations with tumour recurrence. This comprehensive multiomic analysis highlights the complex immunological and metabolic dynamics characteristic of disease recurrence in adjuvant chemotherapy treated NSCLC. Biological sciences/Cancer/Lung cancer/Non-small-cell lung cancer Health sciences/Oncology/Cancer/Cancer microenvironment Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 Introduction Non-small cell lung cancer (NSCLC) is the predominant subtype of lung cancer and is the leading cause of cancer-related death worldwide, accounting for 1.8 million deaths annually 1 . The two most common subtypes of NSCLC include lung adenocarcinoma (LUAD) and lung squamous cell carcinoma (LUSC), contributing to 40% and 25% of lung cancer diagnoses, respectively 1 . Specific driver mutations that currently allow targeted therapies include kirsten rat sarcoma viral oncogene homologue ( KRAS ), epithelial growth factor receptor (EGFR) , and anaplastic lymphoma kinase (ALK) genes, however, a greater understanding of the tumour microenvironment (TME) that spans histology, immunological composition and cancer stage is needed to aid prognoses for other therapeutic strategies, such as chemotherapy and immunotherapy. Spatial multiomics can provide new insights into how the TME contributes towards clinical outcomes measured for endpoints such as recurrence free survival (RFS) or overall survival (OS) 2 . While bulk and single-cell approaches have yielded important insights into NSCLC biology, spatial characterisation of functionally and metabolically distinct cellular phenotypes and how they contribute to inter-cellular networks are likely to provide a deeper understanding of tumour–immune interactions relevant to therapy and survival. Here, we applied spatial multiomic approaches by combining high-plex spatial proteomics, whole-transcriptome spatial transcriptomics, and deep learning in an adjuvant chemotherapy–treated NSCLC tissue cohort to profile 61 patients and over 1.1 million cells. High-plex cyclic immunofluorescence was combined with tumour/stromal-compartment whole transcriptome profiling to identify tumour-immune, functional, and metabolic characteristics of the NSCLC TME associated with tumour recurrence. A geometric deep learning model was developed to gain further insights into the recurrence-risk association of protein-differentiable regions in the tissue structure. Compartment-wise assessment of differentially expressed genes from whole-transcriptome profiling found recurrence-associated enriched pathways, such as metabolism and inflammation. This multifaceted overview of recurrence-risk in resected NSCLC tissues provides pointed insights into the distinct immunological and metabolic profiles of tumours and their associated recurrence risk after adjuvant chemotherapy. Results Patient cohort characteristics The tissue microarrays (TMAs) used in this retrospective study were constructed using locally collected samples by medical oncologists from the Princess Alexandra Hospital (PAH). After filtering patients in the hospital database who satisfied all inclusion criteria (histology, age, treatment, tissue quality), two TMAs comprised of 176 samples from 88 patients were created using 2mm cores from pathology-reviewed, representative tumour regions. These TMAs were profiled using cyclic immunofluorescence (48-plex), H&E staining, and spatial transcriptomics (regional, whole transcriptome profiling). Pathological assessment of H&E images, evaluation of immunofluorescent staining patterns, and bioinformatic inspection of protein and RNA expression was performed to assess both tissue and assay quality. After quality control (QC) of the proteomics dataset, 61 patients contributed at least one sample to the analysis, with 48 of these patients contributing 2 cores. Transcriptome profiling of tumour and stromal compartments was performed on one core per patient. Following QC, transcriptomic data from 52 patients were used for downstream analysis (Fig. 1a, Table 2 ). To identify statistically robust signatures, we distributed patient duplicates between “training” and “validation” groups (Fig. 1b). These groups were jointly used throughout the analysis and in the predictive modelling to statistically cross-validate findings from each group. In the instances where only one tumour core passed QC, these were randomly allocated across the training and validation groups. Spatial features were considered robust only if significant ( p < 0.05) in both training and validation groups. Table 1 Non-small cell lung cancer (NSCLC) patient cohort characteristics Deceased Recurrence Overall Y N Y N Not known All 61 16 45 28 31 2 Stage IIIA 16 26.2% 9 7 11 4 1 IIB 28 45.9% 4 24 13 15 0 IIA 6 9.8% 1 5 2 4 0 IB 6 9.8% 1 5 0 6 0 IIIB 3 4.9% 1 2 2 1 0 IA3 2 3.3% 0 2 0 1 1 Histology * Adenocarcinoma 45 73.7% 13 32 21 22 2 Squamous 12 19.7% 1 11 6 6 0 Large cell carcinoma 3 4.9% 2 1 1 2 0 Adenosquamous 1 1.6% 0 1 0 1 0 Smoking history Y 54 88.5% 14 40 23 29 2 N 6 9.8% 1 5 4 2 0 N/A 1 1.6% 1 0 1 0 0 Mutation None 48 78.7% 14 34 20 26 2 KRAS 5 8.2% 0 5 0 5 0 EGFR 8 13.1% 2 6 8 0 0 * Adenocarcinoma or Squamous histologies were used for transcriptomic analysis. Multivariate analysis was performed on all histologies. Table 2 Custom Phenocycler Fusion Panel Immune characterisation Functional markers Structural markers Metabolic enzymes Lymphocyte CD4 CD8 CD45 CD3e NaK-ATPase CD45RO PD-1 Granzyme B PD-L1 FOXP3 PAX5 HLA-A ICOS(CD278) LAG3 CD44 Ki67 HLA-DR Collagen IV SMA E-cadherin CD31 CD34 Vimentin CPT1A G6PD pNRF2 GLUT1 ATPA5 SDHA IDH2 ASCT2 Citrate synthase LDHA Hexokinase1 B cell CD20 CD21 CD19 Monocyte CD68 CD14 CD163 DC CD11c, CD11B NK CD57 Tumour Pan-Cytokeratin Defined proteomic characteristics for cell type annotation Cell typing was performed by iterative unsupervised clustering based on mean protein expression. Cells were first labelled as either tumour/non-tumour cells and then each class further sub-clustered to resolve distinct immune and structural cell types 3 . The mean protein expression of markers in each cluster used to annotate cells is shown in Fig. 2a , with subsequent normalised cell type proportions per patient shown in Fig. 2b . Tumour cells were further striated based on HLA-DR, HLA-A, Ki67 and CD44 expression. Immune subsets encompassed B cells, CD4 T follicular helper (TFH) cells, CD4 + / CD8 + T cells (including cytotoxic and PD-1 + subsets), and macrophages (M1, M2, PD-L1 + , PD-L1 + M2). Structural categories included fibroblasts, vessels, and miscellaneous stromal cells. These annotations were manually curated by mapping cells back to raw immunofluorescence and H&E images to confirm annotation accuracy. Neighbourhood clustering was performed using the tumour/non-tumour cell classes to create tumour, non-tumour and tumour-interface regions (Fig. 2c). These annotations enabled spatial dissection of tissue regions for downstream spatial analysis with disease recurrence. Metabolic annotation of cell phenotypes Pathway definitions based on positive expression of key markers (Fig. 3a) were utilised to annotate tumour cells with activity for seven discrete metabolic pathways: glycolytic metabolism (GLYC high ), oxidative phosphorylation (OXPH high ), mitochondrial metabolism (MITC high ), oxidative stress (OXSTR high ), lipid oxidation (LPDOX high ), glutamine import (GLN high ), and pentose phosphate pathway (PPP high ). Each cell was annotated by their joint metabolic pathways and confirmed via visual inspection of expression heatmaps and raw data (Fig. 3b-c). Figure 3b shows the average protein expression in tumour cells for each metabolic phenotype. These metabolic phenotypes were assessed both independently as well as with their associated cell type for downstream analysis in both tumour and non-tumour cell populations. Spatial features reveal T cell-driven immunological interactions associated with non-recurrence To assess the cellular spatial characteristics of NSCLC tissues for their association with recurrence, we first calculated cellular densities at a whole core level and within tumour, stroma and tumour-stroma interface cellular neighbourhoods (CNs). Spatial features (proximity density, G-Cross AUC, JSD scores, and entropy/diversity measurements – see Methods section) were calculated on both broad and granular cell-types. After filtering out sparse features, 16,384 spatial features were retained for analysis. Univariate Mann–Whitney U-tests (MW) and Kaplan–Meier (KM) tests were used to measure associations with recurrence and RFS 4 respectively. KM analysis identified 42 recurrence-associated features, while MW tests identified 23 significant features consistent between training (p value = p t ,) and validation (p value = p v) groups (Supplementary Tables 1–2). When the cohort was split into distinct histology types, 66 and 119 features were identified for LUSC samples using KM and MW tests, respectively. Conversely, 10 and 3 features were significant in the LUAD cohort ( Supplementary Tables 1 and 2 ). Univariate analyses highlighted specific spatial features within defined tissue compartments that were associated with tumour recurrence. The proximity of cytotoxic CD8 T cells to tumour cells was indicative of tumour recurrence in both the training and validation sets ( p t = 0.026, p v = 0.0113) (Fig. 2d ) . Broadly across immune types, PD-L1, PD-1, and Granzyme-B expression differed by recurrence status. In the stroma, PD-L1 + Granzyme-B + immune cells were enriched in non-recurrence patients ( p t = 0.041, p v = 0.041) (Fig. 2e). Conversely, higher proportions of PD-1 + Granzyme-B + immune cells in the tumour-stroma interface neighbourhood were associated with shorter time-to-recurrence ( p t = 0.031, p v = 0.034) (Fig. 2f). Across all compartments, specifically stroma, tumour, and tumour–stroma interface CNs, patients with recurrence had higher proportions of PD-1 + immune cells (StromaCN: p t = 0.0196, p v = 0.0198, TumourCN: p t = 0.0061, p v = 0.037, Tumour/Stroma InterfaceCN: p t = 0.0105, p v = 0.0094) (Fig. 2g) . Taken together, these results highlight the conflicting roles that PD-1 and PD-L1 expression on immune cells have on disease recurrence in adjuvant treated NSCLC. Spatial features revealed fibroblast-driven and immune cell-driven metabolic niches associated with tumour recurrence We observed that significant recurrence-associated cell-cell proximities defined by the G-cross function most frequently included combinations of a cell phenotype and a metabolic pathway, indicating the prominent role that cell metabolism plays in the TME (Fig. 3e). In particular, higher proximity of PPP high /GLN high fibroblasts to GLN high fibroblasts were associated with more favourable outcomes ( p t = 0.0099, p v = 0.0059), and a similar association was seen between GLYC high /PPP high /GLN high fibroblasts to GLN high fibroblasts ( p t = 0.0037, p v = 0.0092) (Fig. 3f). Conversely, higher self-proximity of PPP high /GLN high CD4 T cells was unfavourable ( p t = 0.029, p v = 0.011) (Fig. 3g). Other significant associations were identified largely between immune cells with distinct metabolic types ( Supplementary Table 1 ); to illustrate this, a relationship connectivity web of the main findings with respect to the cell-pair proximities are shown in Fig. 3d (favourable associations are shown with green arrows and unfavourable associations with red arrows). In the LUSC samples, significant changes in metabolic-type proportions at a whole core level for CD4 T cells, fibroblasts, and macrophages were associated with tumour recurrence. Proportions of GLN high /GLYC high CD4 T cells ( p t = 0.013, p v = 0.038), GLN high /GLYC high macrophages ( p t = 0.0078, p v = 0.040), GLN high macrophages ( p t = 0.026, p v = 0.0095), and GLN high fibroblasts ( p t = 0.015, p v = 0.0095) ( Supplementary Fig. 2c ) were higher in patients where recurrence did not occur. Conversely, LPDOX high /PPP high CD4 T cells ( p t = 0.0078, p v = 0.013) and LPDOX high /PPP high fibroblasts ( p t = 0.034, p v = 0.042) were in greater proportion in patients who did have recurrence ( Supplementary Table 2, Supplementary Fig. 2b ). Overall, we found that PPP high , GLN high , and LPDOX high in fibroblasts, macrophages and CD8 + T cells, as well as PPP high and GLN high in CD4 T cells and macrophages to be associated with RFS. This highlights a potential feature of immune suppressive regulatory CD4 T cells and M2 macrophages and of their metabolic relationships with the TME. Characterisation of tissue metabolic neighbourhoods identifies disease stage-dependence of the metabolic diversity and non-tumour cell co-occurrence We utilised an orthogonal approach to assigning metabolic phenotypes by generating metabolic marker expression-derived neighbourhoods (Fig. 4a, Supplementary Fig. 3 ) 5 . Eight unique neighbourhoods were established and annotated based on the combination of marker expressions (Fig. 4a ). For entropy/diversity analyses, tumour cells were assigned to these metabolic neighbourhoods, while non-tumour cells were annotated by functional state. Figure 4b showcases the differences in spatial neighbourhood distributions between an early-stage and an advanced sample. Based on Shannon diversity scores calculated at different scales, there were significant differences between early- and late-stage disease at scale 48 ( p t = 0.030, p v = 0.004), and partial significances at scales 24 ( p t = 0.053, p v = 0.029) and 32 ( p t = 0.051, p v = 0.023) (Fig. 4c). The multiscale diversity index (MDI) (Fig. 4b–c) , which evaluates changes in Shannon diversity across spatial scales, demonstrated that advanced staged samples possessed higher MDI scores ( p t = 0.011, p v = 0.0079) (Fig. 4c) , indicating that these samples have noticeable changes in diversity across spatial scales. Global diversity index (GDI) was then derived from the MDI results, to evaluate the spatial adjacency of patches with high or low diversity. The results showed a tendency of lower GDI scores in advanced stage samples ( p t = 0.051, p v = 0.023), indicating that the tissue architecture of later stage of disease is typically less structured and has more blended cellular distributions (Fig. 4d). Moran’s I was used to identify statistically significant areas of similar diversity, based on local diversity index (LDI) calculations, where regions of high entropy were considered hot spots and areas of low entropy were cold spots. Example visualisations of the cell phenotypes considered in the analysis, local diversity index, Moran’s I scores, and the identified hot/cold spots are shown in Fig. 4e . The results found that stage-dependent co-occurrence patterns were evident across the cohort. At a global level, PD-L1 + macrophages co-occurred more frequently with CD8 + T cells ( p t = 0.0335, p v = 0.0364), fibroblasts ( p t = 0.0229, p v = 0.0217), and stromal cells ( p t = 0.0453, p v = 0.0260) in advanced-stage samples (Fig. 4f). Within hot-spots, early-stage samples showed a higher co-occurrence of M2 macrophages with PD-L1 + macrophages ( p t = 0.0376, p v = 0.0187), and PD-L1 + M2 macrophages with stromal cells ( p t = 0.0154, p v = 0.0399) (Fig. 4g). Taken together, these findings suggest that the TME of later-stage disease is more heterogeneous and less structured, showing a broad energy dysregulation that may contribute to recurrence risk. Multivariable modelling found signatures of coordinated immune and metabolic interactions contributing to tumour recurrence We built an integrated model of clinical response using spatial metrics that capture cell–cell interactions and tissue organisation in the TME. Feature selection was applied to spatial features computed across spatial scales in the core (Fig. 5a). These features, G-Cross AUC, JSD Scores, Entropy/Diversity, and proportions, capture the tissue composition, cell configuration, and local cell-cell interactions at a variety of scales and were computed for each CN and/or over the whole core. Features were selected to model two clinical endpoints, recurrence status and RFS, and were then validated by CoxPH and Boosted Cox Survival model fits to data that were unseen by the feature selection algorithm. A distinct 10% difference in time-dependent AUCs was observed between the model fits to features seen by the feature selection algorithm and the held-out validation data (Fig. 5c /5d ); this demonstrates the intractable nature of clinical modelling that are complicated by the tumour heterogeneity, even for the relatively large spatial coverage of these 2mm cores. Features selected using this method are profiled in Fig. 5e/5f and in Supplementary Figs. 5–8 , and exhibit a diverse set of proximity features, density overlap signatures, and proportions for a variety of cell types. The contribution of these features to the overall Boosted Cox Survival model fit to all cores in the cohort is highlighted in Supplementary Fig. 5 , showing SHAP values that exhibit the diverse response of the model to each feature in the core. Further Kaplan–Meier analyses (Fig. 5g; Supplementary Fig. 8 ) confirmed associations of these selected features with RFS and OS. This profile of the TME shows the coordinated architecture of the interactions of stromal cells, immune cells, and tumour cells that collectively model the TME, highlighting the dynamics of M2 macrophage proximity to cytotoxic CD8 T cells, the role of exhausted immune cells, and of the PPP pathway in driving tumour recurrence. Geometric deep learning found recurrence-associated spatial protein regions Traditional measurement and modelling of tissue features is complicated by tumour heterogeneity. Given the broad range of immunological, metabolic and tumour regions in our 2mm cores, we implemented a graph neural network (GNN) method to annotate regions of the core that may distinguish recurrence status. We utilized contrastive learning to separate a single output embedding from subregions of the core, where subregions were represented as graphs containing representations of cell-types, functional type positivity (PD-L1 + , PD-1 + , Ki67 + , etc.), metabolic marker positivity, and metabolic pathway positivity. Figure 6a shows a method and training flowchart, and Fig. 6b gives an overview of the architecture of the model. The model was trained on the training dataset, which corresponded to approximately 600k cells. Post-training, the subgraph output embeddings were pooled for each core and plotted by the actual recurrence status in Fig. 6c , indicating that the output embeddings tend to predominantly differentiate recurrence-like regions from non-recurrence-like regions. Figure 6d shows the averaged embedding score for each core on the left, and the accuracy and precision metrics for the validation data. Examples of these embedding annotations are shown overlaid on cores in Fig. 6e , indicating that the model distinctly stratifies regions of the core that exhibit a diversity of cell-types, tissue structures, and immunological niches. Figure 6f shows the relationship between proportions of cell-types, marker positivity, and metabolic pathways with the embedding annotations, showing the characteristics of these predicted regions. Overall, the highest cell-type and marker positivity values exhibited within non-recurrence regions tended to be antigen expressing tumour phenotypes such as CD44 + , MHC class I (HLA-A), and PD-L1 + . Also highly expressed within these non-recurrence regions were cytotoxic CD8 + T cells and PD-L1 + M2 macrophages. The differentiable and primarily recurrence-associated regions were composed of B cells, MHC class II (HLA-DR) expressing tumour cells, and proliferating (Ki67 + ) tumour cells, as well as tumours without antigen expression. Upregulation of PD-L1, HLA-A, GLUT1, ASCT2, and PD-1 expression were noted in non-recurrence regions, as opposed to an upregulation of the PPP and Oxidative Phosphorylation pathways in the recurrence regions. Glutamine import, GLUT1 + , and glycolytic pathways were also enriched in these non-recurrence regions broadly across tumours, immune cells, and structural cells ( Supplementary Fig. 9 ). Cell proportions in radial neighbourhoods associated with each annotated subgraph showed similar cell-type distributions as in the compartments. Spatial transcriptomics revealed gene and pathway expression changes associated with tumour recurrence and histological classifications To validate the proteomic findings, we performed spatial transcriptomics on adjacent TMA sections using the Nanostring GeoMx Digital Spatial Profiler. Each Area of Interest (AOI), containing one Region of Interest (ROI) separated into tumour or stromal compartment regions, was digitally profiled as previously described 6 . The data were subjected to processing using R package standR ’s quality control (QC) pipeline 7 . After QC, 104 ROIs samples were assessed for sample and gene quality, with 6 ROIs removed based on assessment of the nuclei counts and library size. Principal component analysis (PCA) did not identify any confounding factors and suggested the main variance is attributable to differences in segmentation (i.e. location) and to histological subtype ( Supplementary Fig. 10a ). The data was then Trimmed Mean of M-values (TMM) normalised, correcting for any systemic bias ( Supplementary Fig. 10b ). When comparing the RNA expression of the tumour regions between LUAD and LUSC samples, we identified 369 upregulated transcripts and 230 downregulated transcripts ( Supplementary Table 3) . Of these significantly differentially expressed genes, keratin 5 ( KRT5 ) was upregulated in LUSC samples whereas KRT7 was higher in LUAD (Fig. 7a). Amongst the DEGs, fructosamine-3-kinase ( FN3K ) and macrophage-stimulating 1 receptor ( MST1R ) were upregulated in LUAD while cytochrome P450, family 4, subfamily F, polypeptide 11 ( CYP4F11 ) were upregulated in LUSC. These gene expressions were also significant in the tumour segments, where they were associated with a worse OS (Fig. 7b). Cell differentiation, goblet cell, epithelial, and gland development pathways were all identified to be upregulated within the LUAD cohort (Fig. 7c). Differential expression analysis was also performed between patients with tumour recurrence and those without, on all samples and on each histological segment. No DEGs were identified in any of the non-tumour comparisons with limited DEGs identified in the tumour segment. Melanoma associated antigen family A9B ( MAGEA9B ) was upregulated in LUAD and is associated with poor survival 8 . Cyclin D2 ( CCND2 ) was upregulated in LUSC and is associated with survival and immune infiltration (Fig. 7d) 9 . CCND2 was identified in our feature association analysis where higher expression was associated with better RFS (Fig. 7e). Pathway analysis revealed downregulation of viral KEGG pathways including measles, HPV, EBV, and viral carcinogenesis in recurrence patients (Fig. 7f ) . Leveraging on the capabilities of the GNN approach, we utilised the GNN based modelling to estimate recurrence-like compartments and subsequently predicted recurrence status per sample. Similarly, DE analysis was conducted comparing patients by GNN-predicted recurrence either in all samples or in either segment. Limited DEGs were identified for all the contrasts with most DEGs found in the tumour segments ( Supplementary Table 3 ). Specifically, matrix metallopeptidase 10 ( MMP10 ) has been known to be required for lung cancer maintenance, metastasis, and recurrence 10 , 11 (Fig. 7g). The gene encoding solute carrier family 38 member 5 ( SLC38A5 ) is a regulator of glutaminolysis. Feature association analysis identified statherin ( STATH ) gene expression associated with worse OS in the tumour while the expression of the gene encoding RAB6B was associated with better OS in the tumour (Fig. 7h). Pathway analysis suggested an enrichment of metabolic processes (including lipid metabolism) in the adenocarcinoma recurrence samples (Fig. 7i). For the non-tumour segments, DEGs were again limited to 4 DEGs upregulated in the patients with GNN-predicted recurrence ( Supplementary Table 3 ). Of these, peptidyl arginine deiminase 4 ( PADI4 ) is overexpressed in lung cancer and linked to EMT and metastasis 12 , 13 , and tumour necrosis factor ligand superfamily member 18 ( TNFSF18 ) is a potential biomarker for predicting the efficacy of anti-PD-1 treatments 14 (Fig. 7j). Feature association analysis linked the gene encoding guanylate binding protein 7 ( GBP7 ) (a negative regulator of innate immune responses, specifically IFN signalling) 15 expression in the non-tumour segment with a worse RFS (Fig. 7k). Pathway analysis suggested an overall enrichment of T cell receptor signalling processes in non-recurrence patients. In the non-tumour segments, NOD-like receptor signalling and neutrophil extracellular trap formation were enriched in the LUSC patients who did not have tumour recurrence while the complement and coagulation pathways were enriched in LUAD patients (Fig. 7l). Taken together, the findings from the RNA analysis align with the major functional outcomes from the protein analysis and the GNN-modelling, specifically the observation of the enrichment of lipid metabolism and other metabolic processes in recurrence patients when stratified using GNN-predicted classifications of patients. Discussion Spatial multiomic profiling provides deeper insight of the NSCLC TME beyond bulk or single-cell approaches. We identified cellular phenotypes, functional markers, and metabolic alterations associated with tumour recurrence and disease stage in the adjuvant chemotherapy setting. Altered cell dependency and metabolism of glutamine, glucose, and lipids are well characterised and tend to support mechanisms enabling the survival and adaptation of cancer cells 16 – 20 . Our data suggests that metabolic diversity shifts with disease stage, and this may be reflective of cellular adaptation to changes in extracellular nutrient availability. This allows cancer cells to acquire the metabolic characteristics required for metastasis spread, driven by an energy independence to energy shortfalls experienced during circulation 21 . Here, we show that the compositional diversity of NSCLC metabolic phenotypes was found to be associated with disease stage, with global and “metabolic hot spot” co-occurrence of critical immune cells, such as PD-L1 + macrophages, decreasing across stages 22 . Global metabolic diversity across length scales (from local neighbourhoods to the whole tissue core) was found to decrease in late-stage disease, reflecting an adaptation to changing energy regimes across the tissue. In both the transcriptomic and proteomic data, lipid metabolism emerged as a consistent determinant of recurrence risk. In particular, pathway analysis identified lipid, sterol, and cholesterol transport pathways enriched in tumour regions of recurrence patients. In the protein analysis, carnitine palmitoyltransferase I (CPT1a), a key enzyme in fatty acid oxidation, was used to represent the lipid oxidation pathway. In tandem with CPT1a, our analysis highlighted a role for G6PD, which functions in the pentose phosphate pathway (PPP) pathway. G6PD is a rate-limiting enzyme involved in the PPP and is well described in lung cancer oncogenesis, chemotherapy resistance, and immune checkpoint efficacy 17 , 23 , 24 . G6PD contributes to nucleotide synthesis and lipid maintenance through generation of nicotinamide adenine dinucleotide phosphate (NADPH), required for reductions in de novo lipogenesis 25 . G6PD and CPT1a are commonly overexpressed in tumour cells, and their dual inhibition has caused significant reduction in tumour growth and improvement to OS in breast cancer and glioblastoma 19 , 26 . Notably, our study identified that the dual expression of CPT1a and G6PD and their association with tumour recurrence was dependent upon cellular organisation and pathways. For example, favourable outcomes were identified when these metabolic populations were in closer proximity: LPDOX high /PPP high /GLN high cells to GLN high cells, and LPDOX high /PPP high /GLN high cells to GLYC high /GLN high cells. Higher proportions of LPDOX high /PPP high /GLN high were identified to be unfavourable unless these populations were in proximity to the GLN high (negative for other metabolic markers) population. While these findings highlight the involvement of CPT1a and G6PD in the metabolic reprogramming that supports recurrence, it also points to potential involvement and competition for extracellular glutamine. This could support previous studies that showed the inhibition of ASCT2 or the depletion of glutamine availability can contribute to decreases in lipid and fatty acid biosynthesis 27 . In the context of this study, while the dual expression of ASCT2 and CPT1a is a less established phenomenon, there is recent evidence to suggest that ASCT2 may be stabilising CPT1a activity 27 . In non-tumour cells, our study identified the combination of glutamine import and PPP had unfavourable outcomes when enriched in fibroblasts, CD4 T cells, CD8 T cells, and macrophages. Our multivariate feature analysis also identified that the proximity of CD8 T cells with PPP high /GLN high tumour cells was associated with a shorter RFS. Conversely, GLN high /GLYC high or GLN high /GLYC high /PPP high populations in proximity to CD8 T cells and fibroblasts was found to be favourable. Past research has shown, in mouse models, that tumour and immune cells compete for glutamine, and that selectively blocking glutamine entry to the TCA cycle in tumour cells increases T-cell activation 18 . Our data also shows that in cells enriched in oxidative, glycolytic, PPP, and lipid oxidising pathways, there are distinct ASCT2 + and ASCT2- subpopulations. While it is difficult to infer the metabolite availability in these environments, we can speculate that ASCT2 expression in these populations could be due, at least partly, to their proximity to other cell types which may be competing for extracellular glutamine in the TME 18 . In NSCLC, macrophages make up the majority of infiltrating immune cells, where they can exhibit both pro-tumour and anti-tumour effects 28 . Both proteomic and transcriptomic analysis identified macrophage-related features associated with tumour recurrence. In the transcriptomic analysis, these included MSTR1 29 within the tumour and TNFAIP6 in the stroma, where higher expression was linked to shorter RFS. TNFAIP6 is known to be involved with M2 polarisation, extracellular matrix modelling, and cell migration, which could contribute to disease aggression 30 , 31 . Mechanistic studies of PD-L1 + macrophages typically have found roles in suppression of anti-tumour immune cells, but some results show immunostimulatory and beneficial effects in a range of cancers, including NSCLC 32 , 33 . Our multivariable modelling showed that PD-L1 + and M2 macrophages play a significant role, with the separation of PD-L1 + macrophages from PD-L1 − macrophages being highly favourable and the separation of PD-L1 + M2 macrophages from stem-like antigen expressing tumour cells also contributing to longer RFS. The interplay between tumour-associated macrophages (TAMs) and tumour metabolism has been shown to be an important aspect in TME remodelling 34 – 36 , and is also dependent on the macrophage metabolism – M1 macrophages are typically associated with glycolytic metabolism, whereas the anti-inflammatory M2 macrophages typically rely on oxidative metabolism 37 . TMEs associated with high lactate content and extended lactic acidosis have enhanced monocyte to macrophage differentiation with pro-tumour and inflammatory characteristics 38 . Of particular interest, results from our univariate analysis identified that the co-localisation of GLN high macrophages with PPP high /GLN high CD4 T cells was unfavourable and that an increased proportion of GLN high macrophages and GLN high CD4 T cells contributed to a longer RFS. The multivariate analysis also revealed that increased proximity of LPDOX high macrophages with GLYC high /PPP high /GLN high CD4 T cells was favourable, potentially indicating a role for a metabolically atypical M1 TAM state associated with CD4 T cells in tumour non-recurrence. Using geometric deep learning for the stratification of tissue core regions into protein niches differentiable by tumour recurrence status, we found a diverse set of functional and metabolic states, highlighting the effects of several distinct immunological interactions. Major histocompatibility (MHC) proteins are essential for an adaptive immune system but have differing reported effects in the literature. Increased expression of MHC class I protein has shown to be associated with increased survival in NSCLC 39 ; particularly, increased MHC class II protein expression has been shown to improve OS in resection surgery-treated LUAD NSCLC patients 40 . However, our analyses results suggests that this MHC I associated effect may be dependent on the composition of the tumour cells when expressing CD44, PD-L1, Ki67, or HLA-A. The model predictions also point to distinct, well established immunological interactions, such as the presence of PD-L1 expression in the tumour in conjunction with proliferating or stem-like gene expression – all in the presence of cytotoxic CD8 T cells 41 . Clinically unfavourable associations were found for proliferating tumour cells both with or without antigen expression, the presence of B cells (in our study, this study could not differentiate between plasma, memory, or effector B cells), and MHC class II-expressing tumours without additional functional properties. This is validated in the transcriptomics data, where relapse-associated DEGs in both tumour and stroma regions include CEACAM1 , and CD3D , which are indirectly associated with MHC processes via MHC-class-I inhibition and antigen detection 42 , 43 . Our study has built a comprehensive profile of the TME, revealing additional cellular interactions that profile the TME and contribute to tumour recurrence. This profile highlighted the complex metabolic and immunological characteristics of the TME across disease stage and across the clinical response. We further validated this profile using spatial transcriptomics in a standalone manner and in combination with a deep learning technique to detangle signatures of tumour recurrence in gene expression. In addition, we validated previously identified immunological associations in the NSCLC TME and the mechanistic and metabolic relationships in the TME contributing to tumour recurrence. Figure Legends Methods Study design This study has Metro South Human Research Ethics approval (LNR/2019/QMS/51117) and ratification by the Queensland University of Technology. This study has University of Queensland Human Research Ethics Approval (2021/HE001936). Samples for this retrospective study were collected by medical oncologists at the Princess Alexandra Hospital (PAH). Using the lung cancer database, patients were filtered with the intent to collate tumour-microarrays from resected lung cancer tissues. Of these, 2,678 patients with NSCLC were identified to have received platinum-based chemotherapy. The cohort was established based on curative treatment intent (fully resected, early-stage disease, who received adjuvant platinum-based chemotherapy). The inclusion criteria were patients over 18 years of age, NSCLC cases treated only with adjuvant cisplatin/carboplatin in combination with vinorelbine, and tumour specimens of adequate tissue availability for IHC staining. Whole tissue blocks were sourced, pathology reviewed, and dual cores punched from representative areas at least ~ 0.5cm apart. A normal non-malignant lung sample was included on each TMA to act as a control and assist with orientation. The samples were formalin-fixed paraffin embedded, and cores were of a 2mm diameter. The final TMAs were comprised of 176 samples from 88 patients, where all samples were treatment naïve and from fully resected specimens. Spatial single-cell proteomic profiling Patient cohort 84 patients contributed 168 cores to cyclic-IF analysis. Quality control was performed to assess the staining and tissue quality across the two TMAs, with cores excluded if they were deemed to be majority of poor quality, or a subsection of the core was used if part of the core was disturbed by tissue lifting or staining artifacts such as staining edge effect. After quality control, 61 patients were included in analysis (48 paired samples and 13 unpaired samples). Multiplex IF tissue profiling Tissue processing was performed by the AKOYA step team. Tumour microarrays were subject to the standard PhenoCycler Fusion imaging protocol 44 . The cyclic imaging protocol requires up to three fluorophores to be imaged simultaneously, before being washed away and three new reporters applied to the tissue. The order of reporter application was designed so that antibodies known to co-localise were not put in the same cycle, and antibodies known to have high expression were placed at the end of protocol. Once all images were captured, post processing combined the layers into one qptiff file, using a blank cycle to adjust for background staining. Proteomic bioinformatic analysis Cell segmentation and data QC Qptiff files were imported into Qupath for initial image processing, including marker and tissue quality control and cell segmentation. Segmentation was performed using the Cellpose (v2.0) plugin in Qupath using the nucleus model with 5µm nuclear expansion, constrained at 1.8 nuclear size. The median fluorescent intensities of each marker and cellular morphology features for each cell were exported into a .csv file containing a unique cellular ID, appropriate core ID, and x-y co-ordinates. Based on edge effects of imaging on the x-axis, boundaries were introduced to remove cells and cores that were affected by edge effects. For some cores, this resulted in the whole core being excluded from analysis while some cores were split roughly in half. Cell typing Cell typing was performed based on the median protein expression of each cell. Cells were first clustered into tumour and non-tumour phenotypes, using a subset of markers and morphology features; CD4, E-cadherin, CD20, CD68, CD8, CD14, Vimentin, CD45, CD34, CD3e, CD31, SMA, Ki67, Pan-Cytokeratin, nuclear size, nuclear solidity, and nuclear diameter. Cellular expression was normalised based on an arcsinh transform with a co-factor of 150 and a percentile transform of 97 along axis 0. A Z-score transform was then applied to the normalised expression. Batch correction was applied by PCA transform and integration using harmony integrate (rapids single cell package) between images 45 . Phenograph unsupervised clustering was performed based on neighbour analysis (n = 20), a resolution of 0.4 and minimum size of 10. 21 clusters were established and then annotated based on protein expression on a heatmap and after mapping clusters manually onto the images 46 . Using the raw data, the data was subset into non-tumour cells for further subtyping. The markers used for broad annotations were CD4, CD20, CD68, CD8, CD14, Vimentin, CD45, CD34, CD3e, CD31, SMA, nuclear solidity, and nuclear diameter. The same normalisation and integration method as earlier was applied to the non-tumour cells. Phenograph clusters were calculated over a range of resolutions and k-nearest neighbours. The heatmaps were evaluated using ARI, NMI, and AMI metrics to select the most reliable cluster parameters. The top 6 silhouette scores were visually inspected to identify the most reasonable cluster parameters. Based on heatmap expression, the parameters K = 40 and resolution = 0.2 were used for broad non-tumour phenotypes, these clusters were then annotated based on mean expression of markers. These non-tumour annotations were then merged into the original raw data file. For CD4, CD8, macrophage, and tumour populations, these were then further annotated, individually, by sub-setting the raw data into only the CD4, CD8, macrophage, or tumour populations and the same normalisation and integration methods, as used earlier, were applied to each subset. Following this, a binary gaussian mixture model (GMM) was applied to relevant functional markers for each cell type, then based on the combinations of positive markers the cells were functionally annotated. For the CD4 population, PD-1 and ICOS were assessed, and two populations were established: CD4 cells (~ 106 k) and CD4 T Follicular Helper (TFH) cells (ICOS + , PD-1 + ~ 6k cells). For CD8 T cells, Granzyme-B and PD-1 were assessed to establish CD8 populations of Cytotoxic CD8 and PD-1 + CD8 T cells. GMM positivity for CD163 and PD-L1 were assessed for the macrophage population, creating four subtypes: macrophages, M2 macrophages, PD-L1 + macrophages, and PD-L1 + M2 macrophages. For the tumour phenotyping, positive staining of CD44, HLA-DR, HLA-A, KI67, and PD-L1 were assessed to assign phenotypes of all combinations of CD44 + , MHC class I or class II positive, proliferating, or PD-L1 + types. Similarly, using the metabolic markers in the panel (Table 2 ) based on the combination of positive or negative markers, metabolic pathways were annotated for all cells. These definitions were created based on the metabolic pathways present in our panel through consultation with an expert in cancer metabolomics (Fig. 3a). Due to the number of cell types this produced, two cell phenotypes were created: 1. Broad cell types (tumour and non-tumour) + metabolic pathways and 2. Broad cell phenotypes of non-tumour cells + tumour cells annotated with metabolic pathways. Neighbourhoods Cellular neighbourhoods (CNs) were defined using a k-means clustering of KNN neighbourhood counts of base tumour/non-tumour types 47 . CNs were classified into “Tumour” or “Stroma” (k = 2) on the tumour/non-tumour cell proportions, and k = 3 CNs were labelled as “Tumour,” “Tumour/Stroma Interface,” and “Stroma” regions using tumour/non-tumour cell proportions. Metabolic neighbourhoods were generated based on the protein expression of G6PD, pNRF2, ATPA5, GLUT1, Hexokinase1, ASCT2, Citrate synthase, SDHA, CPT1A, PanCK and Ki67 using the MESA package 5 [ks = 20, K-means method, kd tree]. We identified 8 distinct metabolic neighbourhoods using this method, including a ‘low’ neighbourhood that had little expression of metabolic markers in our panel. Most immune cells were captured in ‘low’ neighbourhoods. Mostly, this method captured metabolic diversity within the tumour cells, with only a small proportion of tumour cells captured in the ‘low’ metabolic neighbourhood. Entropy analysis We employed spatial diversity algorithms in MESA to calculate the local and global diversity of the samples at several scales 5 . Cells were annotated by their metabolic neighbourhood for the tumour cells and the functional annotation of non-tumour cells. This approach was used to capture the metabolic diversity of tumours at a niche level and identify differences in the immune cell populations in these neighbourhoods. MDI (multiscale diversity index) score was calculated based on the slope of Shannon diversity scores calculated over a range of scales (2 (~ 1mm patches), 4 (~ 500 µm), 8 (~ 250 µm), 16 (~ 130 µm), 24 (~ 80 µm), 32 (~ 60 µm), 48 (~ 40 µm), 64 (~ 30 µm), 72 (~ 27 µm)). To establish scales, samples were dissected into grids, where the number of rows and columns was that of the selected scale – i.e. a scale of 72 would divide the samples by applying a grid of 72x72 onto the tissue. Based on MDI results, the optimal scale was 32, and this resolution was used in subsequent analyses. Shannon-diversity and Moran’s I was used as the entropy metric for sub-regions; these values were then compared between local (LDI) and global scales (GDI). Hot and cold spots were identified by identifying regions that had similar Moran I scores. Patches that have high entropy scores were labelled as hot spots and patches that have low entropy scores were labelled cold spots. Cell frequency and cellular co-occurrence was calculated across each tissue at a global scale and then within every hot and cold spot. Proximity features Proximity metrics were computed for the cohort to capture cell-cell and cell-functional interactions at a range of scales. Global scales were captured by various entropy metrics and cell proportions within each core and within each CN. Medium-scale interactions were captured by the JSD Score, G-Cross AUC metrics, and diversity metrics. JSD measures probability overlap within each CN or core of cell-type or cell-functional type pairs and is bounded by 0 (complete density overlap) and 1 (complete density separation) 48 . Cell-cell interactions were described quantitatively using the G-Gross function, which measures the cumulative distribution of the nearest-neighbour distance of each target cell type from each reference type. The edge correction proposed by Hanisch was used to account for potential biases induced by a cell’s proximity to the core edge, which restricts the calculation to reference cells which are closer to the nearest target cell than to the edge of the core 49 . As G-Cross is a function that incorporate information from the range of computed radii, the area under the curve (AUC) was computed for the G-Cross function between 0 and ~ 250 µm. The smallest-scale interactions were captured by the proximity density metric. Proximity density scores were calculated between tumour metabolic neighbourhoods, with immune cells and other tumour phenotypes to identify patterns of cell-pair co-localisation. Proximity density was based on KNN within a 50-pixel radius and calculated the ratio of the number of two specific cell types within that radius compared to the total number of those two cell types within the core 50 . Multivariable Feature Selection Feature selection was performed using the Stabl package, a feature selection method designed to select model-informative features from high-dimensional data in a false discovery-minimal manner 51 . Metrics for each feature family (G-Cross AUC, JSD Scores, proportions, and entropy) were computed and were also combined into a single feature set by a quantile transform to enforce a similar cumulative distribution for all features. Sparse features were filtered by removal of features that had less than 60 unique values among the 106 training and validation cores, reducing this down to approximately 6k features, and these remaining features were standard scaled before feature selection was performed. Four complementary model types and feature selection methods were tested in parallel: adaptive Lasso fits to recurrence status for each feature family, adaptive lasso fits to the quantile transformed (QT) features, Cox Proportional Hazards model (CoxPH) fits to the feature family-selected features for RFS, and CoxPH fits to the QT transformed features for RFS. We found that feature-family fitting to model RFS produced marginally better time-dependent AUCs than the QT features for a BoostedCoxSurvival model to RFS and that QT-feature selection worked marginally better for fitting recurrence status. BoostedCoxSurvival models were fit with 100 estimators and were trained on the training set and tested on the validation set. SHAP values were estimated for the BoostedCoxSurvival model fits to quantify the core-to-core variation in the overall estimate of the feature importance and to measure the contribution to each feature’s hazard ratio. Models to quantify the feature selection in the context of patient survival were implemented in scikit-survival , with linear model fits, standard scaling transforms, and quantile transforms performed in scikit-learn . Feature importances were estimated by the SHAP python library using the kernel explainer function. Deep Learning Training Graphs were formed using cells as nodes, with edges formed using KNNs with radial connections from the cell centroid x and y positions, up to 80 pixels away. Node features were formed using one-hot encoded cell-types from the list of 14 cell types formed during the cell-typing process. Additional functional markers such as the MHC class, ICOS, etc., positivity scores were generated using a Gaussian Mixture Model (GMM) fit to the median pixel values within each cell mask and appended to the feature list, followed by the metabolic pathway positivity scores formed from the pathway assessment logic outlined in Fig. 3a . PD-1 and PD-L1 markers exhibited high background expression due to ineffective reporter removal during the imaging cycles, so rolling ball subtraction of these two channels was applied before expression measurements were taken. The distance between each cell pair per 100 pixels was appended as an edge feature to the graphs. Graphs were implemented in Deep Graph Library, with subgraphs sampled using k = 2 hops. The model was constructed with two Convolutional Transformer layers using the torch-geometric library, with the layers containing 32 embedding size, single headed attention, mean pooling before output into a single dimension embedding, and final normalization using a sigmoid layer. The contrastive loss function was implemented from the PyTorch Metrics Learning library with default parameters 52 . Training was performed on the training cores to contrast batched subgraphs by the patient recurrence status and prediction scores from the agglomerated cell predictions were tested on the validation cores. Graph augmentation was applied during training by random masking of feature values, random node and edge dropping, and gaussian smearing of the edge distance feature. The training was stopped after 12 epochs as preliminary tests showed that additional training merely pushed marginal regions in the “differentiable”- “non-differentiable” interface apart in the embedding space with little additional benefit to interpretability or prediction power. Histograms of the embedding scores stratified by the patient’s disease recurrence status showed that the predominant and undifferentiable regions of the cores were non-recurrence-like and that differentiable regions tended to be recurrence-associated. Spatial transcriptomic profiling (compartment-specific) Whole transcriptome spatial profiling The TMAs were subject to the GeoMx DSP Whole Transcriptome Atlas (WTA) assay. Tissue morphology markers included Pan-cytokeratin, CD45 and cyto-13 as markers for tumour, lymphocytes, and nuclei, respectively. 660µm round regions were drawn and masked by either PanCK + masks (tumour) or PanCK- masks (non-tumour) from a single sample per patient. Based on these selections, a laser directed UV-light onto the selected area, cleaving the oligonucleotide barcodes. These cleaved barcodes were removed from the slide surface and put into a well in a 96-well plate by the instrument. One well corresponds to one AOI. barcodes were prepared for sequencing according to manufacturer protocol and sequencing was performed by Australian Genome Research Facility (AGRF). Raw data was processed through the Nanostring GeoMx Next Generation Sequencing pipeline and digital count files reinstated on machine for mapping of counts. A QC level csv file was exported from the instrument for bioinformatic analysis. Transcriptomic bioinformatics analysis Quality control and normalisation The transcriptomics data from the GeoMx DSP comprising 120 regions of interest (ROIs)/AOIs (each representing either tumour or non-tumour segments) from 60 patients were processed using the standR package workflow (Liu, et al., 2024). Firstly, only adenocarcinoma and squamous carcinoma samples were kept (110 ROIs, 55 patients). Secondly, 3 patients lacking clinical information were removed, leaving 104 ROIs and 52 patients. Sample filtering was performed by excluding ROIs with low detection counts (< 500,000) and low nuclei counts ( 90% of the ROIs), with no genes removed. The logCPM counts were evaluated using relative log expression (RLE) and principal components analysis (PCA) plots to identify confounding factors and to identify batch effects. Normalisation was then performed using the trimmed mean of M-values (TMM) method across all genes in the panel to adjust for library size variations and unwanted compositional biases. Differential expression analysis Differential expression (DE) analysis was carried out using the R packages edgeR (v3.34.0) (Chen, et al., 2025) and limma (3.48.0) (Ritchie, et al., 2015), using linear modelling incorporating experimental factors as predictors. The variation in gene expression was modelled by considering a common dispersion affecting all genes and a gene-specific dispersion. To estimate these dispersions, an empirical Bayes approach was employed, borrowing information from all genes and treating patient variation as a random effect using duplicatecorrelation in edgeR::voomLmFit function using sample weights. The linear model was then fitted to the experimental design containing the relevant factors, and DE analysis was conducted for specific contrasts of interest where in some cases, the T test relative to a threshold (TREAT) criterion was used 53 . Contrast includes comparing between histology (adenocarcinoma or squamous), recurrence status (Yes or No) or predicted recurrence status (based on modelling) within segments (i.e. Tumour or Stroma). The resulting statistic is an empirical Bayes moderated t-statistic with multiple testing adjustment performed using the Benjamini–Hochberg procedure where an adjusted p-value of < 0.05 was used to determine significantly DE genes. Predicted recurrence status for each sample were determined using a median cutoff of GNN-predicted recurrence score (from GNN-modelling) strategy, into either yes or no recurrence. Geneset enrichment analysis (GSEA) and visualisations GSEA was performed using the statistics provided by the DE analysis. The limma::fry function was utilised for testing gene sets from the Molecular Signatures Database’s (MsigDB) categories including Hallmarks , C2 (curated gene sets), C5 (Ontology gene sets) and C7 (immunological signature gene sets, IMMUNESIGDB). Limma::kegga function was used to test terms from the Kyoto Encyclopedia of Genes and Genomes (KEGG) database while limma::goana function was used for testing gene ontology (GO) enrichment. The respective contrasts in the DE analyses were interrogated to identify functional enrichment of specific gene sets in the tumour and stromal segments associated with different survival/recurrence groups. The results from GSEA were inputted into the vissE (1.8.0) (Bhuva, et al., 2024) R package where clusters of enriched gene sets with common biological themes were identified and visualized. This visualization assisted in identification of biologically relevant pathways of interest in the study. Feature association with survival outcomes Analysis of the feature association with survival or recurrence outcomes were implemented as per described previously 6 . Briefly, univariate cox proportional hazards (CoxPH) regression model and Kaplan Meier (KM) survival analyses were performed using the R packages survival and survminer . Features tested were collated from the significant DE genes from either 1) Histology in either segment or 2) GNN-predicted recurrence status in either segment. For the CoxPH modelling, RFS or OS durations were treated as the time variable with the event defined by occurrence of RFS events (DFS < OS) or deceased status respectively. Features having a Wald value of < 0.05 were followed with a KM survival analysis for further validation. In the KM analysis, the samples were grouped based on a median cutoff of the feature’s gene expression as either high or low. Features with a p-value < 0.05 in the KM were considered significant. Multivariate CoxPH and KM analysis were also conducted with either histology or clinical stage as a covariate. Statistics Univariate CoxPH modelling, and Kaplan Meier (KM) survival analysis was performed on the data for OS and RFS using R packages survival and survminer . Median cell proportions with logit transformation were utilized for cohort stratification when assessing the association between cell proportion of ROIs and OS/recurrence status. Cox proportional hazards models were applied for non-discrete cell proportions. For the proteomics data, as there were instances of multiple samples per patient “training” and “validation” groups were created, with samples from the same patient distributed randomly between the two groups, and the remaining samples randomly distributed between the two groups. The cores from the cohort were grouped with 55 patient samples in the training group and 54 in the validation group. Unpaired samples were randomly split into two roughly even groups, with 7 samples assigned to the training and 6 to the validation cohort. In the univariate feature analysis, p -values were initially adjusted for false discovery rate using the Benjamini-Hochberg correction, however this yielded no significant results due to the large number of features, so the raw p -value was subsequently used. Features were filtered down to those with over 60 unique value to reduce feature sparsity. Features were compared using the Mann-Whitney U-test for binary tests (Histological subtype, recurrence (yes/no), status (alive/dead), stage (low/high)) and Kaplan Meier fits were performed using feature median splits to differentiate high and low feature values in relation to the RFS. To avoid diluting the impact of intra-tumoral heterogeneity by averaging the features per patient, features were considered statistically robust if the feature was significant in both the training and validation sets. Transcriptomics Spatial RNA-seq data only contained one sample per patient and thus was not subject to training and validation splits. Statistical significance was calculated using binary comparisons between histological subtypes, recurrence, and survival status. For differential gene expression the resulting statistic is an empirical Bayes moderated t-statistic with multiple testing adjustment performed using the Benjamini–Hochberg procedure where an adjusted p -value of < 0.05 was used to determine significantly DE genes. Declarations Data Availability Data from this study, including the spatial proteomics data, spatial transcriptomics data and metadata are available at https://doi.org/10.48610/3244c1b Code for this study is available at https://github.com/clinicalomx/Adjuvant_Spatial_Multiomic_Profiling Conflicts of Interest JH and GM is an employee of Surge Care. AK is on the Scientific Advisory Board for Omapix Solutions, European Spatial Biology Centre, Predxbio, Molecular Instruments, and Visiopharm. Author Contributions Concept: CO, KOB, AK, MA Acknowledgements The authors would like to thank the Translational Research Institute Histology Facilities for assisting with building the tissue microarray. The authors acknowledge the contribution of Akila Wijeranthna Yapa in the manuscript writing and review. The study investigators AK, CWT, CB, NL, KOB, MA are supported by the MRFF METASPATIAL Study (2031100). AK is supported by Cure Cancer and the Princess Alexandra Research Foundation. NB, MLD, JM, KE, AK, AK1, JF are supported by the Queensland Spatial Biology Centre (QSBC) at the Wesley Research Institute. AK1 is supported by the Harding Family Fellowship (Wesley Research Institute). AK = Arutha Kulasinghe, AK1 = Aaron Kilgallon. References Garg P et al (2024) Advances in Non-Small Cell Lung Cancer: Current Insights and Future Directions. J Clin Med 13. https://doi.org/10.3390/jcm13144189 Rad HS et al (2021) The Pandora's box of novel technologies that may revolutionize lung cancer. Lung Cancer 159:34–41. https://doi.org/10.1016/j.lungcan.2021.06.022 Jhaveri N et al (2023) Mapping the Spatial Proteome of Head and Neck Tumors: Key Immune Mediators and Metabolic Determinants in the Tumor Microenvironment. GEN Biotechnol 2:418–434. https://doi.org/10.1089/genbio.2023.0029 Monkman J et al (2025) Metabolic characterisation of tumor-immune interactions by multiplexed immunofluorescence reveals spatial mechanisms of immunotherapy response PREPRINT Research Square https://doi.org/ https://doi.org/10.21203/rs.3.rs-6695435/v1 Ding DY et al (2025) Quantitative characterization of tissue states using multiomics and ecological spatial analysis. Nat Genet 57:910–921. https://doi.org/10.1038/s41588-025-02119-z Tan CW et al (2025) The development of a high-plex spatial proteomic methodology for the characterisation of the head and neck tumour microenvironment. npj Precision Oncol 9:191. https://doi.org/10.1038/s41698-025-00963-0 Liu N et al (2024) standR: spatial transcriptomic analysis for GeoMx DSP data. Nucleic Acids Res 52:e2. https://doi.org/10.1093/nar/gkad1026 Zhang S et al (2015) High expression of MAGE-A9 in tumor and stromal cells of non-small cell lung cancer was correlated with patient poor survival. Int J Clin Exp Pathol 8:541–550 Hung CS et al (2018) Hypermethylation of CCND2 in Lung and Breast Cancer Is a Potential Biomarker and Drug Target. Int J Mol Sci 19. https://doi.org/10.3390/ijms19103096 Cho NH et al (2004) MMP expression profiling in recurred stage IB lung cancer. Oncogene 23:845–851. https://doi.org/10.1038/sj.onc.1207140 Gill JH et al (2004) MMP-10 Is Overexpressed, Proteolytically Active, and a Potential Target for Therapeutic Intervention in Human Lung Carcinomas. Neoplasia 6:777–785. https://doi.org/https://doi.org/10.1593/neo.04283 Lesbon JCC et al (2022) A Screening of Epigenetic Therapeutic Targets for Non-Small Cell Lung Cancer Reveals PADI4 and KDM6B as Promising Candidates. Int J Mol Sci 23:11911 Liu M et al (2019) PADI4–mediated epithelial–mesenchymal transition in lung cancer cells Retraction in /10.3892/mmr.2025.13610. Mol Med Rep 19:3087–3094. https://doi.org/10.3892/mmr.2019.9968 Chen B et al (2021) Immune-related genes and gene sets for predicting the response to anti-programmed death 1 therapy in patients with primary or metastatic non-small cell lung cancer. Oncol Lett 22:540. https://doi.org/10.3892/ol.2021.12801 Ahmetoglu D, Zheng H, Swart A, Zhu H, Li M (2025) Multifaceted Roles of Guanylate-Binding Proteins in Cancer. Int J Mol Sci 26:5477 Hanahan D, Weinberg RA (2011) Hallmarks of Cancer: The Next Generation. Cell 144:646–674. https://doi.org/10.1016/j.cell.2011.02.013 Lan T et al (2024) Glucose-6-phosphate dehydrogenase maintains redox homeostasis and biosynthesis in LKB1-deficient KRAS-driven lung cancer. Nat Commun 15:5857. https://doi.org/10.1038/s41467-024-50157-8 Edwards DN et al (2021) Selective glutamine metabolism inhibition in tumor cells improves antitumor T lymphocyte activity in triple-negative breast cancer. J Clin Investig 131. https://doi.org/10.1172/JCI140100 Liu Z et al (2023) CPT1A-mediated fatty acid oxidation confers cancer cell resistance to immune-mediated cytolytic killing. Proceedings of the National Academy of Sciences 120, e2302878120 https://doi.org/doi:10.1073/pnas.2302878120 Tang M et al (2022) CPT1A-mediated fatty acid oxidation promotes cell proliferation via nucleoside metabolism in nasopharyngeal carcinoma. Cell Death Dis 13:331. https://doi.org/10.1038/s41419-022-04730-y Weber GF (2016) (2016) Time and Circumstances: Cancer Cell Metabolism at Various Stages of Disease Progression. Frontiers in Oncology Volume 6– https://doi.org/10.3389/fonc.2016.00257 Liu Y et al (2020) Immune Cell PD-L1 Colocalizes with Macrophages and Is Associated with Outcome in PD-1 Pathway Blockade Therapy. Clin Cancer Res 26:970–977. https://doi.org/10.1158/1078-0432.Ccr-19-1040 Hong W et al (2018) Inhibition of Glucose-6-Phosphate Dehydrogenase Reverses Cisplatin Resistance in Lung Cancer Cells via the Redox System. Front Pharmacol 9–2018. https://doi.org/10.3389/fphar.2018.00043 Nakamura M et al (2024) Blockade of glucose-6-phosphate dehydrogenase induces immunogenic cell death and accelerates immunotherapy. J Immunother Cancer 12:e008441. https://doi.org/10.1136/jitc-2023-008441 Koundouros N, Poulogiannis G (2020) Reprogramming of fatty acid metabolism in cancer. Br J Cancer 122:4–22. https://doi.org/10.1038/s41416-019-0650-z Kim SJ et al (2022) Dual inhibition of CPT1A and G6PD suppresses glioblastoma tumorspheres. J Neurooncol 160:677–689. https://doi.org/10.1007/s11060-022-04189-z Wang J et al (2024) ASCT2 Regulates Fatty Acid Metabolism to Trigger Glutamine Addiction in Basal-like Breast Cancer. Cancers (Basel) 16. https://doi.org/10.3390/cancers16173028 Zhou Y et al (2024) The role of tumor-associated macrophages in lung cancer: From mechanism to small molecule therapy. Biomed Pharmacother 170:116014. https://doi.org/https://doi.org/10.1016/j.biopha.2023.116014 Cazes A, Childers BG, Esparza E, Lowy AM (2022) The MST1R/RON Tyrosine Kinase in Cancer: Oncogenic Functions and Therapeutic Strategies. Cancers (Basel) 14. https://doi.org/10.3390/cancers14082037 Li L et al (2022) TNFAIP6 defines the MSC subpopulation with enhanced immune suppression activities. Stem Cell Res Ther 13:479. https://doi.org/10.1186/s13287-022-03176-5 Liu S, Chen H, Yang X, Wen Y, Chen L (2024) Identification and validation of up-regulated TNFAIP6 in osteoarthritis with type 2 diabetes mellitus. Sci Rep 14:31450. https://doi.org/10.1038/s41598-024-82985-5 (!!! INVALID CITATION !!! 32) Wang L et al (2024) PD-L1-expressing tumor-associated macrophages are immunostimulatory and associate with good clinical outcome in human breast cancer. Cell Rep Med 5. https://doi.org/10.1016/j.xcrm.2024.101420 Schuurbiers OCJ et al (2014) Glucose Metabolism in NSCLC Is Histology-Specific and Diverges the Prognostic Potential of 18FDG-PET for Adenocarcinoma and Squamous Cell Carcinoma. J Thorac Oncol 9:1485–1493. https://doi.org/https://doi.org/10.1097/JTO.0000000000000286 He A et al (2025) Decoding the role of lipid metabolism in NSCLC: From macrophage subtype identification to prognostic model development. The FASEB Journal 39, e70467 https://doi.org/https://doi.org/10.1096/fj.202500124 Netea-Maier RT, Smit JWA, Netea MG (2018) Metabolic changes in tumor cells and tumor-associated macrophages: A mutual relationship. Cancer Lett 413:102–109. https://doi.org/https://doi.org/10.1016/j.canlet.2017.10.037 Li M et al (2023) Metabolism, metabolites, and macrophages in cancer. J Hematol Oncol 16:80. https://doi.org/10.1186/s13045-023-01478-6 Colegio OR et al (2014) Functional polarization of tumour-associated macrophages by tumour-derived lactic acid. Nature 513:559–563. https://doi.org/10.1038/nature13490 Datar IJ et al (2021) Spatial Analysis and Clinical Significance of HLA Class-I and Class-II Subunit Expression in Non–Small Cell Lung Cancer. Clin Cancer Res 27:2837–2847. https://doi.org/10.1158/1078-0432.Ccr-20-3655 Johnson AM et al (2021) Cancer Cell-Specific Major Histocompatibility Complex II Expression as a Determinant of the Immune Infiltrate Organization and Function in the NSCLC Tumor Microenvironment. J Thorac Oncol 16:1694–1704. https://doi.org/https://doi.org/10.1016/j.jtho.2021.05.004 Wu S-P et al (2018) Stromal PD-L1–Positive Regulatory T cells and PD-1–Positive CD8-Positive T cells Define the Response of Different Subsets of Non–Small Cell Lung Cancer to PD-1/PD-L1 Blockade Immunotherapy. J Thorac Oncol 13:521–532. https://doi.org/https://doi.org/10.1016/j.jtho.2017.11.132 Gray-Owen SD, Blumberg RS (2006) CEACAM1: contact-dependent control of immunity. Nat Rev Immunol 6:433–446. https://doi.org/10.1038/nri1864 Wei Z et al (2022) CD3D: a prognostic biomarker associated with immune infiltration and immunotherapeutic response in head and neck squamous cell carcinoma. Bioengineered 13:13784–13800. https://doi.org/10.1080/21655979.2022.2084254 Donovan ML et al (2024) Protocol for high-plex, whole-slide imaging of human formalin-fixed paraffin-embedded tissue using PhenoCycler-Fusion. STAR Protoc 5:103226. https://doi.org/10.1016/j.xpro.2024.103226 scverse/rapids_singlecell: v0.10.6 v. 0.10.6 (2024) Levine JH et al (2015) Data-Driven Phenotypic Dissection of AML Reveals Progenitor-like Cells that Correlate with Prognosis. Cell 162:184–197. https://doi.org/10.1016/j.cell.2015.05.047 Schürch CM et al (2020) Coordinated Cellular Neighborhoods Orchestrate Antitumoral Immunity at the Colorectal Cancer Invasive Front. Cell 182:1341–1359e1319. https://doi.org/10.1016/j.cell.2020.07.005 Masotti M, Osher N, Eliason J, Rao A, Baladandayuthapani VDIMPLE (2023) An R package to quantify, visualize, and model spatial cellular interactions from multiplex imaging with distance matrices. Patterns 4:100879. https://doi.org/https://doi.org/10.1016/j.patter.2023.100879 Hanisch KH (1984) Some remarks on estimators of the distribution function of nearest neighbour distance in stationary spatial point processes. Ser Stat 15:409–412. https://doi.org/10.1080/02331888408801788 Nirmal AJ, Sorger PK (2024) SCIMAP: A Python Toolkit for Integrated Spatial Analysis of Multiplexed Imaging Data. J Open Source Softw 9. https://doi.org/10.21105/joss.06604 Hédou J et al (2024) Discovery of sparse, reliable omic biomarkers with Stabl. Nat Biotechnol 42:1581–1593. https://doi.org/10.1038/s41587-023-02033-x Musgrave K, Belongie S, Lim S-N Pytorch metric learning. arXiv preprint arXiv :(2008). 09164 (2020) McCarthy DJ, Smyth GK (2009) Testing significance relative to a fold-change threshold is a TREAT. Bioinformatics 25:765–771. https://doi.org/10.1093/bioinformatics/btp053 Additional Declarations Yes there is potential Competing Interest. JH and GM are employees of Surge Care. AK is on the Scientific Advisory Board for Omapix Solutions, European Spatial Biology Centre, Predxbio, Molecular Instruments, and Visiopharm. Supplementary Files SupplementaryFigures.docx Supplementary Figures with legends Supplementalfigure1.png Supplementary Figure 1 Supplementalfigure2.png Supplementary Figure 2 Supplementalfigure3.png Supplementary Figure 3 SupplementalFigure4.png Supplementary Figure 4 SupplementalFigure5.png Supplementary Figure 5 SupplementalFigure6.png Supplementary Figure 6 SupplementalFigure7.png Supplementary Figure 7 SupplementalFigure8.png Supplementary Figure 8 SupplementalFigure9.png Supplementary Figure 9 SupplementalFigure10.png Supplementary Figure 10 sp1.pdf Code and Software Submission Checklist rs1.pdf Reporting Summary Supplementarytables.xlsx Supplementary tables Cite Share Download PDF Status: Under Review Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-7692105","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":532932342,"identity":"5c87dd10-5506-4277-b416-fcdb45b5b341","order_by":0,"name":"Arutha Kulasinghe","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAABJElEQVRIie3QsUrDQBjA8a8EkuVK1sugvsIdgVhoHyaHYJcglYI4pCUQuG7tWshLRFzcLByky+FccZKCU4eUguBSeokGFM6Cm8P9SY4k3O8uCYDJ9E+z1EnAUWNYHU7SPPy1VloTqyFo8RdSKxweJ24mz9eDGHwirHX5+jC6cr0NXcFtlyWOJDqCnyKazgsIiLADzORy6GWR3wHZZwmKtAQkoimyoUcEBMB4wfKXKMAtLlgCenJWk31FnF1ZkcdnqcheEXejJaQibV69GCKY8ZjlGCmSKIL1u1BpX2ftKfY9gW4UWbC5vBx2wqLvc/w20JFTaeU79N6j0+XkfvvBx2w2EXerMu6ezNyLXPv5Xz+uuRCft6Ea7CPzvzf+uYLJZDKZVAdOlF/fUaaWSQAAAABJRU5ErkJggg==","orcid":"https://orcid.org/0000-0003-3224-7350","institution":"The University of Queensland","correspondingAuthor":true,"prefix":"","firstName":"Arutha","middleName":"","lastName":"Kulasinghe","suffix":""},{"id":532932343,"identity":"b3747be6-98c1-406a-9ec9-4d87325e1502","order_by":1,"name":"Naomi Berrell","email":"","orcid":"","institution":"Wesley Research Institute","correspondingAuthor":false,"prefix":"","firstName":"Naomi","middleName":"","lastName":"Berrell","suffix":""},{"id":532932344,"identity":"c154fe12-a3a2-4cb9-a363-638422040871","order_by":2,"name":"Chin Wee Tan","email":"","orcid":"","institution":"Walter and Eliza Hall Institute","correspondingAuthor":false,"prefix":"","firstName":"Chin","middleName":"Wee","lastName":"Tan","suffix":""},{"id":532932345,"identity":"73174e9a-3279-4b4e-b7e9-e18205acd99d","order_by":3,"name":"Aaron Kilgallon","email":"","orcid":"https://orcid.org/0000-0003-1450-0009","institution":"Wesley Research Institute","correspondingAuthor":false,"prefix":"","firstName":"Aaron","middleName":"","lastName":"Kilgallon","suffix":""},{"id":532932346,"identity":"21cb41eb-d23d-44eb-bd17-c6f9b26c9d5b","order_by":4,"name":"James Monkman","email":"","orcid":"https://orcid.org/0000-0002-7219-8402","institution":"University of Queensland","correspondingAuthor":false,"prefix":"","firstName":"James","middleName":"","lastName":"Monkman","suffix":""},{"id":532932347,"identity":"321f727a-3697-410c-8d33-41d164abfff6","order_by":5,"name":"Charles Bidgood","email":"","orcid":"","institution":"Queensland University of Technology","correspondingAuthor":false,"prefix":"","firstName":"Charles","middleName":"","lastName":"Bidgood","suffix":""},{"id":532932348,"identity":"3c261d33-339d-494c-9a61-b7734bf9cf58","order_by":6,"name":"Cooper Janke","email":"","orcid":"","institution":"University of Queensland","correspondingAuthor":false,"prefix":"","firstName":"Cooper","middleName":"","lastName":"Janke","suffix":""},{"id":532932349,"identity":"b9d73c1d-0dc8-4c12-a233-8bdd3d36495c","order_by":7,"name":"Meg Donovan","email":"","orcid":"https://orcid.org/0000-0002-9875-8914","institution":"Wesley Research Institute","correspondingAuthor":false,"prefix":"","firstName":"Meg","middleName":"","lastName":"Donovan","suffix":""},{"id":532932350,"identity":"4d1dee3b-2497-4c7e-aa16-5799a37c5a3d","order_by":8,"name":"Ning Liu","email":"","orcid":"","institution":"University of Adelaide","correspondingAuthor":false,"prefix":"","firstName":"Ning","middleName":"","lastName":"Liu","suffix":""},{"id":532932351,"identity":"38808c7b-7ddd-4cdb-9a2c-aa4e4b276229","order_by":9,"name":"Kidane Embaye","email":"","orcid":"","institution":"University of Queensland","correspondingAuthor":false,"prefix":"","firstName":"Kidane","middleName":"","lastName":"Embaye","suffix":""},{"id":532932352,"identity":"ecb4afb7-1298-45f5-bac7-8a56fe3c6a3a","order_by":10,"name":"Joseph Yunis","email":"","orcid":"https://orcid.org/0000-0002-0272-0605","institution":"uni","correspondingAuthor":false,"prefix":"","firstName":"Joseph","middleName":"","lastName":"Yunis","suffix":""},{"id":532932353,"identity":"90cba63c-90fe-42ca-96be-4b56dcb138d0","order_by":11,"name":"Gwendoline Mendes","email":"","orcid":"","institution":"Surge care","correspondingAuthor":false,"prefix":"","firstName":"Gwendoline","middleName":"","lastName":"Mendes","suffix":""},{"id":532932354,"identity":"e336e225-373f-41df-88d5-81b50646635e","order_by":12,"name":"Julien Hedou","email":"","orcid":"","institution":"Surge care","correspondingAuthor":false,"prefix":"","firstName":"Julien","middleName":"","lastName":"Hedou","suffix":""},{"id":532932355,"identity":"a949fbcb-603d-4e73-9e2b-7e960bbcc1d4","order_by":13,"name":"John Fraser","email":"","orcid":"","institution":"Univer","correspondingAuthor":false,"prefix":"","firstName":"John","middleName":"","lastName":"Fraser","suffix":""},{"id":532932356,"identity":"0d9fe8f7-bcf5-499a-bb29-49a167b44d5e","order_by":14,"name":"Connor O'Leary","email":"","orcid":"","institution":"Mater Hospital","correspondingAuthor":false,"prefix":"","firstName":"Connor","middleName":"","lastName":"O'Leary","suffix":""},{"id":532932357,"identity":"98f94f8f-b7fb-4297-8635-2d7a1dc9a5b1","order_by":15,"name":"Ken O'Byrne","email":"","orcid":"","institution":"Princess Alexandra Hospital","correspondingAuthor":false,"prefix":"","firstName":"Ken","middleName":"","lastName":"O'Byrne","suffix":""},{"id":532932358,"identity":"ae92314a-00df-4c1f-bc80-34048c6527f7","order_by":16,"name":"Mark Adams","email":"","orcid":"https://orcid.org/0000-0003-1906-5018","institution":"Queensland University of Technology","correspondingAuthor":false,"prefix":"","firstName":"Mark","middleName":"","lastName":"Adams","suffix":""}],"badges":[],"createdAt":"2025-09-23 09:01:54","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-7692105/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-7692105/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":94198009,"identity":"ceddb9fb-6514-4fc9-8c39-6520f8bb2a37","added_by":"auto","created_at":"2025-10-23 13:26:03","extension":"png","order_by":0,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1775513,"visible":true,"origin":"","legend":"","description":"","filename":"Figure1.png","url":"https://assets-eu.researchsquare.com/files/rs-7692105/v1/9b63f75aff0d187ae4001239.png"},{"id":94198010,"identity":"fb27c08f-1b7f-4358-a6cd-f8d01d08f9dc","added_by":"auto","created_at":"2025-10-23 13:26:03","extension":"docx","order_by":1,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":306049,"visible":true,"origin":"","legend":"","description":"","filename":"Berrelletal2025.docx","url":"https://assets-eu.researchsquare.com/files/rs-7692105/v1/687d61007f5cfff598a03928.docx"},{"id":94198016,"identity":"a75f5857-f1d5-4364-99a5-3a7bbd2e21a3","added_by":"auto","created_at":"2025-10-23 13:26:03","extension":"png","order_by":2,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":8106427,"visible":true,"origin":"","legend":"","description":"","filename":"Figure2.png","url":"https://assets-eu.researchsquare.com/files/rs-7692105/v1/b15c3170fd186e58d277578b.png"},{"id":94198020,"identity":"bae28a84-13c7-43f1-9028-959bbdade8a4","added_by":"auto","created_at":"2025-10-23 13:26:03","extension":"png","order_by":3,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":7541071,"visible":true,"origin":"","legend":"","description":"","filename":"Figure3.png","url":"https://assets-eu.researchsquare.com/files/rs-7692105/v1/e6db7f51842ae196b48c74f3.png"},{"id":94198013,"identity":"2ab57320-c329-4884-ace5-91bad6ae9514","added_by":"auto","created_at":"2025-10-23 13:26:03","extension":"png","order_by":4,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":7429430,"visible":true,"origin":"","legend":"","description":"","filename":"Figure4.png","url":"https://assets-eu.researchsquare.com/files/rs-7692105/v1/38bcc847caf7c47226d1b0da.png"},{"id":94198031,"identity":"d8df009c-52cc-46b7-b870-9752e3ea52f6","added_by":"auto","created_at":"2025-10-23 13:26:04","extension":"png","order_by":5,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":3330765,"visible":true,"origin":"","legend":"","description":"","filename":"Figure5.png","url":"https://assets-eu.researchsquare.com/files/rs-7692105/v1/1f58e6fa669aa81bc2238122.png"},{"id":94198040,"identity":"6c95b3d0-0c7c-48bf-a121-e0628bc18035","added_by":"auto","created_at":"2025-10-23 13:26:04","extension":"png","order_by":6,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":7194315,"visible":true,"origin":"","legend":"","description":"","filename":"Figure6.png","url":"https://assets-eu.researchsquare.com/files/rs-7692105/v1/de2c461436a7f94fadca70c8.png"},{"id":94198038,"identity":"10db776a-9de1-4ffa-98f8-ca28fe57acd8","added_by":"auto","created_at":"2025-10-23 13:26:04","extension":"jpg","order_by":7,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":5218404,"visible":true,"origin":"","legend":"","description":"","filename":"Figure7.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7692105/v1/53f8e41e247737e009660591.jpg"},{"id":94198508,"identity":"d3fbdc0f-60d3-4197-88c5-9b691fb533bf","added_by":"auto","created_at":"2025-10-23 13:34:04","extension":"json","order_by":8,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":15483,"visible":true,"origin":"","legend":"","description":"","filename":"NCOMMS2580635T.json","url":"https://assets-eu.researchsquare.com/files/rs-7692105/v1/2deeb55c3b335dd6f5286124.json"},{"id":94198509,"identity":"e72799a2-33b9-4c68-be8b-6568499bb639","added_by":"auto","created_at":"2025-10-23 13:34:04","extension":"png","order_by":9,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":553988,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementalFigure10.png","url":"https://assets-eu.researchsquare.com/files/rs-7692105/v1/a54d106f3127d77bfa765d3f.png"},{"id":94198511,"identity":"bee988c6-53ca-4e19-9e65-de0b4e032019","added_by":"auto","created_at":"2025-10-23 13:34:04","extension":"png","order_by":10,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1899760,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementalFigure4.png","url":"https://assets-eu.researchsquare.com/files/rs-7692105/v1/a955a3108daaf52fb35e6e6c.png"},{"id":94198042,"identity":"550d2bcb-6fef-4424-bfa3-55a683b2edfe","added_by":"auto","created_at":"2025-10-23 13:26:04","extension":"png","order_by":11,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":2899851,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementalFigure5.png","url":"https://assets-eu.researchsquare.com/files/rs-7692105/v1/63228f6457de81b1abab7b1a.png"},{"id":94198047,"identity":"faaebf3b-8a71-45ad-98d7-e1c762edaea6","added_by":"auto","created_at":"2025-10-23 13:26:04","extension":"png","order_by":12,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":590447,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementalFigure6.png","url":"https://assets-eu.researchsquare.com/files/rs-7692105/v1/d34e08277dd8081f3a55778e.png"},{"id":94198033,"identity":"1c3074aa-aaf1-45df-82f6-dff1930c0368","added_by":"auto","created_at":"2025-10-23 13:26:04","extension":"png","order_by":13,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":2219638,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementalFigure7.png","url":"https://assets-eu.researchsquare.com/files/rs-7692105/v1/0b60ba4a63a2c0c6932d474e.png"},{"id":94199811,"identity":"19d2d79a-836b-420e-a01f-3ccd8eb94fe3","added_by":"auto","created_at":"2025-10-23 13:42:05","extension":"png","order_by":14,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1430154,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementalFigure8.png","url":"https://assets-eu.researchsquare.com/files/rs-7692105/v1/791fe86ea89521815d3751e0.png"},{"id":94198513,"identity":"9e1b6129-2d2f-4861-b7f0-fe463b22bc94","added_by":"auto","created_at":"2025-10-23 13:34:04","extension":"png","order_by":15,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":799930,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementalFigure9.png","url":"https://assets-eu.researchsquare.com/files/rs-7692105/v1/c9c8ef8edf006a8c6fbc284f.png"},{"id":94198059,"identity":"e976e60c-cf92-4b1e-bf91-49e6dd15d0ec","added_by":"auto","created_at":"2025-10-23 13:26:05","extension":"png","order_by":16,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":969221,"visible":true,"origin":"","legend":"","description":"","filename":"Supplementalfigure1.png","url":"https://assets-eu.researchsquare.com/files/rs-7692105/v1/863acfb97c461f761ffcb274.png"},{"id":94198052,"identity":"1572790d-9779-4ae5-b3c5-0c6d55249d5d","added_by":"auto","created_at":"2025-10-23 13:26:05","extension":"png","order_by":17,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":453039,"visible":true,"origin":"","legend":"","description":"","filename":"Supplementalfigure2.png","url":"https://assets-eu.researchsquare.com/files/rs-7692105/v1/63765d6d0ded5e43586b97ca.png"},{"id":94199808,"identity":"e661c32a-7597-4ead-bf08-97e0da1af2f1","added_by":"auto","created_at":"2025-10-23 13:42:04","extension":"png","order_by":18,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":427500,"visible":true,"origin":"","legend":"","description":"","filename":"Supplementalfigure3.png","url":"https://assets-eu.researchsquare.com/files/rs-7692105/v1/43abe7975898fa8a4d38ba57.png"},{"id":94198058,"identity":"d471adaf-ad81-46c2-b16c-c5c3e4ace9d0","added_by":"auto","created_at":"2025-10-23 13:26:05","extension":"docx","order_by":19,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":3248430,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementaryFigures.docx","url":"https://assets-eu.researchsquare.com/files/rs-7692105/v1/ef295363b143687e2dc27026.docx"},{"id":94198048,"identity":"fb0c50b7-b796-4645-a5e6-80b4ef38ce26","added_by":"auto","created_at":"2025-10-23 13:26:05","extension":"xlsx","order_by":20,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":47494,"visible":true,"origin":"","legend":"","description":"","filename":"Supplementarytables.xlsx","url":"https://assets-eu.researchsquare.com/files/rs-7692105/v1/b8e441338cc812e83f286b04.xlsx"},{"id":94198067,"identity":"ba54df4a-768c-4863-a9c0-d3f68886d3ed","added_by":"auto","created_at":"2025-10-23 13:26:05","extension":"xml","order_by":21,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":196224,"visible":true,"origin":"","legend":"","description":"","filename":"NCOMMS2580635T0enriched.xml","url":"https://assets-eu.researchsquare.com/files/rs-7692105/v1/4d88b9e0c447171bebfaf324.xml"},{"id":94198512,"identity":"f0a55008-11e3-4a61-a54e-fcbecf9f81f1","added_by":"auto","created_at":"2025-10-23 13:34:04","extension":"png","order_by":22,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1775513,"visible":true,"origin":"","legend":"","description":"","filename":"Figure1.png","url":"https://assets-eu.researchsquare.com/files/rs-7692105/v1/1d1655db76a034cac410580b.png"},{"id":94199807,"identity":"e96b011c-36cc-476b-a52a-46a901b0fe0a","added_by":"auto","created_at":"2025-10-23 13:42:04","extension":"png","order_by":23,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":8106427,"visible":true,"origin":"","legend":"","description":"","filename":"Figure2.png","url":"https://assets-eu.researchsquare.com/files/rs-7692105/v1/017b6a5293083544c00b6a0f.png"},{"id":94198056,"identity":"59507138-34fc-4546-9627-081aad7e8f4c","added_by":"auto","created_at":"2025-10-23 13:26:05","extension":"png","order_by":24,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":7541071,"visible":true,"origin":"","legend":"","description":"","filename":"Figure3.png","url":"https://assets-eu.researchsquare.com/files/rs-7692105/v1/cb8f14b243eb252a911c2fbf.png"},{"id":94198053,"identity":"7f2918b0-5b27-4aa6-a846-e14b9a78a6bd","added_by":"auto","created_at":"2025-10-23 13:26:05","extension":"png","order_by":25,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":7429430,"visible":true,"origin":"","legend":"","description":"","filename":"Figure4.png","url":"https://assets-eu.researchsquare.com/files/rs-7692105/v1/becf4b86ae0e114e848f4848.png"},{"id":94198515,"identity":"be1c3124-3903-4375-9dcd-2d250303f0df","added_by":"auto","created_at":"2025-10-23 13:34:04","extension":"png","order_by":26,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":3330765,"visible":true,"origin":"","legend":"","description":"","filename":"Figure5.png","url":"https://assets-eu.researchsquare.com/files/rs-7692105/v1/3aed04fd79887571cf15dd11.png"},{"id":94199818,"identity":"b32201b1-d073-4e5c-9f4a-3d9966c46d01","added_by":"auto","created_at":"2025-10-23 13:42:05","extension":"png","order_by":27,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":7194315,"visible":true,"origin":"","legend":"","description":"","filename":"Figure6.png","url":"https://assets-eu.researchsquare.com/files/rs-7692105/v1/9d2ced8a05a4147afdd8f96b.png"},{"id":94198526,"identity":"4cbd15b1-51f0-4e13-9bff-6b9014abf10b","added_by":"auto","created_at":"2025-10-23 13:34:06","extension":"jpg","order_by":28,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":5218404,"visible":true,"origin":"","legend":"","description":"","filename":"Figure7.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7692105/v1/388a12f833432d1b126eccce.jpg"},{"id":94200308,"identity":"b6333d40-0a8c-4520-b516-3f7637a46e02","added_by":"auto","created_at":"2025-10-23 13:50:05","extension":"png","order_by":29,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":484051,"visible":true,"origin":"","legend":"","description":"","filename":"OnlineFigure1.png","url":"https://assets-eu.researchsquare.com/files/rs-7692105/v1/1f40a6c4d95eea019339f696.png"},{"id":94199810,"identity":"18f8e75a-1a78-4fec-b168-26613e699256","added_by":"auto","created_at":"2025-10-23 13:42:05","extension":"png","order_by":30,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1644661,"visible":true,"origin":"","legend":"","description":"","filename":"OnlineFigure2.png","url":"https://assets-eu.researchsquare.com/files/rs-7692105/v1/10375c778abe972414b13079.png"},{"id":94199813,"identity":"f68b34d0-feb5-4924-91c6-fc7d58319ddf","added_by":"auto","created_at":"2025-10-23 13:42:05","extension":"png","order_by":31,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1606456,"visible":true,"origin":"","legend":"","description":"","filename":"OnlineFigure3.png","url":"https://assets-eu.researchsquare.com/files/rs-7692105/v1/cda540b1e96e4ee44c8193be.png"},{"id":94198051,"identity":"51035249-474c-45ec-9e66-101630c86de8","added_by":"auto","created_at":"2025-10-23 13:26:05","extension":"png","order_by":32,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1522163,"visible":true,"origin":"","legend":"","description":"","filename":"OnlineFigure4.png","url":"https://assets-eu.researchsquare.com/files/rs-7692105/v1/350247fb73971c8c07272dfc.png"},{"id":94198065,"identity":"ac30827a-9ae3-4eed-ab8b-d78df999a027","added_by":"auto","created_at":"2025-10-23 13:26:05","extension":"png","order_by":33,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":663513,"visible":true,"origin":"","legend":"","description":"","filename":"OnlineFigure5.png","url":"https://assets-eu.researchsquare.com/files/rs-7692105/v1/c562e54a8eed87c6670caaf6.png"},{"id":94198517,"identity":"22333c7a-cfd5-4395-bb72-2e5dea1a4234","added_by":"auto","created_at":"2025-10-23 13:34:05","extension":"png","order_by":34,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1262760,"visible":true,"origin":"","legend":"","description":"","filename":"OnlineFigure6.png","url":"https://assets-eu.researchsquare.com/files/rs-7692105/v1/e3ceec62570a98d268ce1fd1.png"},{"id":94200309,"identity":"9a3202e9-4524-4e9f-9ff1-a6bbe687ec4b","added_by":"auto","created_at":"2025-10-23 13:50:05","extension":"png","order_by":35,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":351492,"visible":true,"origin":"","legend":"","description":"","filename":"OnlineFigure7.png","url":"https://assets-eu.researchsquare.com/files/rs-7692105/v1/8f9e4cedc038e19c755eabea.png"},{"id":94200310,"identity":"281f1c7e-8aed-4503-8989-17db93dc4297","added_by":"auto","created_at":"2025-10-23 13:50:05","extension":"xml","order_by":36,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":192625,"visible":true,"origin":"","legend":"","description":"","filename":"NCOMMS2580635T0structuring.xml","url":"https://assets-eu.researchsquare.com/files/rs-7692105/v1/7f49f4950fa541d109f83f2f.xml"},{"id":94198069,"identity":"42a9aa12-56f2-4185-bea1-6875ec31f32a","added_by":"auto","created_at":"2025-10-23 13:26:06","extension":"html","order_by":37,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":215771,"visible":true,"origin":"","legend":"","description":"","filename":"earlyproof.html","url":"https://assets-eu.researchsquare.com/files/rs-7692105/v1/05c2de32be4be63a302f04e4.html"},{"id":94198500,"identity":"76917078-f54b-4549-b686-7cf2df1bbe70","added_by":"auto","created_at":"2025-10-23 13:34:03","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":1775513,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eNon-small cell lung cancer (NSCLC) patient characteristics and study overview. a)\u003c/strong\u003e The NSCLC patient clinicopathological findings, showing the clinical characteristics of the study cohort. All patients included in this plot contributed at least 1 tissue core to the proteomic analysis. Patient ages (top row) are represented by both the dot positions and sizes on the y-axis. \u003cstrong\u003eb)\u003c/strong\u003e Overview of study design (right) and a diagram of how patient tissue samples were distributed between training and validation groups (left).\u003c/p\u003e","description":"","filename":"Figure1.png","url":"https://assets-eu.researchsquare.com/files/rs-7692105/v1/d16ed9d44d6d9bc167beebf5.png"},{"id":94198019,"identity":"0f588587-9f6f-4129-b96e-43b3e27b1b7c","added_by":"auto","created_at":"2025-10-23 13:26:03","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":8106427,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eOverview of the data types, analysis approach, and cell-typing results in this study.\u0026nbsp; a)\u003c/strong\u003e Heatmap of the mean protein expression used for base cell typing. 14 cell types were established using unsupervised clustering approaches.\u0026nbsp; \u003cstrong\u003eb)\u003c/strong\u003e Cell-type proportions of patients, normalised to all cells (top) and non-tumour cells (bottom), stratified across tumour recurrence status. \u003cstrong\u003ec) \u003c/strong\u003eHematoxylin and Eosin (H\u0026amp;E), immunofluorescence, cell type, and cellular neighbourhood plots of a representative core.\u0026nbsp; KNN neighbourhoods were generated using tumour and non-tumour cell types to establish tumour, stroma and tumour/stroma interface neighbourhoods.\u0026nbsp; \u003cstrong\u003ed)\u003c/strong\u003e Box plots of G-Cross AUCs of cytotoxic CD8 T cells to tumour cells for patients with recurrent versus non-recurrent disease. \u003cstrong\u003ee)\u003c/strong\u003e Box plots showing the proportion of PD-L1\u003csup\u003e+\u003c/sup\u003e Granzyme-B\u003csup\u003e+\u003c/sup\u003e immune cells within the stroma neighbourhood show higher proportions in patients who did not have tumour recurrence. \u003cstrong\u003ef)\u003c/strong\u003e Kaplan Meier analysis calculated using time to recurrence indicates higher proportion of PD-1\u003csup\u003e+\u003c/sup\u003e Granzyme-B\u003csup\u003e+\u003c/sup\u003e immune cells within the tumour/stroma interface was associated with tumour recurrence. \u003cstrong\u003eg)\u003c/strong\u003e Box plots showing that a higher proportion of PD-1\u003csup\u003e+\u003c/sup\u003e immune cells within the stroma (i), tumour (ii) and tumour/stroma interface (iii) cellular neighbourhoods was found in patients who had recurrence.\u003c/p\u003e","description":"","filename":"Figure2.png","url":"https://assets-eu.researchsquare.com/files/rs-7692105/v1/cd362da6cb6bf7851fb873da.png"},{"id":94199800,"identity":"7b7187ca-7cd6-4229-a426-f919dc72ef60","added_by":"auto","created_at":"2025-10-23 13:42:03","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":7541071,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eAn overview of the metabolic characteristics associated with tumour recurrence\u003c/strong\u003e. \u003cstrong\u003ea)\u003c/strong\u003e\u0026nbsp; A table showing the metabolic pathway definitions used to phenotype cells by positivity/negativity of metabolic proteins. \u003cstrong\u003eb)\u003c/strong\u003e A heatmap of the metabolic protein expressions for each tumour metabolic phenotype. \u003cstrong\u003ec)\u003c/strong\u003e Spatial maps of representative tissue cores based on the metabolic pathway phenotypes. \u003cstrong\u003ed)\u003c/strong\u003e A schematic diagram of significant G-Cross AUCs based on Kaplan Meier analyses. Interactions associated with favourable outcomes are shown with green arrows, and less favourable outcomes are shown with red arrows. \u003cstrong\u003ee-g)\u003c/strong\u003e Kaplan Meier analyses of G-Cross AUC between two cells using the base cell type \u003csup\u003e+\u003c/sup\u003e metabolic pathways, where high and low groups are relative to the feature median values. Kaplan Meier analyses were computed using recurrence-free-survival time, using log-rank tests to determine significance.\u003c/p\u003e","description":"","filename":"Figure3.png","url":"https://assets-eu.researchsquare.com/files/rs-7692105/v1/4f886a7f2acccc0d0d1a22b4.png"},{"id":94200306,"identity":"170a437b-0379-4c11-9299-7debbb7822f4","added_by":"auto","created_at":"2025-10-23 13:50:03","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":7429430,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eMetabolic neighbourhoods and expression diversity analysis. a\u003c/strong\u003e) Seven neighbourhoods were generated using the protein expression values of metabolic markers. \u003cstrong\u003eb)\u003c/strong\u003eVisual representations of the metabolic neighbourhoods for two tumour cores – one an early-stage sample and the other an advanced sample. Both cores are from patients who experienced disease recurrence. \u003cstrong\u003ec)\u003c/strong\u003e Multiscale diversity index plot showing the overall Shannon diversity values over a range of scales, and the slope of change. Statistical significance, where both the training and validation groups have \u003cem\u003ep\u003c/em\u003e-value \u0026lt; 0.05, is indicated by * and was calculated between early and advanced samples. Where either the training or validation groups have a \u003cem\u003ep-\u003c/em\u003evalue \u0026lt;0\u003cem\u003e.\u003c/em\u003e05 and the other group has a \u003cem\u003ep-\u003c/em\u003evalue \u0026lt;0.06, this is represented by • on the graph. \u003cstrong\u003ed)\u003c/strong\u003e Violin plot of global diversity index scores between early and late-stage samples indicate that lower GDI scores are associated with later stages of disease. \u003cstrong\u003ee)\u003c/strong\u003e Visualisations of diversity scores on two representative tumour cores. Spatial maps showing the location of immune cells and tumour cells, where tumour cells are characterised by the metabolic neighbourhoods (first column). Maps of local diversity scores calculated at a scale of 32, where darker red represents a higher diversity index and darker blue represents lower diversity index (second column). Visual representation of Moran’s I diversity calculation where darker green represents a lower p-value (third column). Moran’s I is used to characterise areas that share similar diversity scores. Based on Moran’s I calculations to identify statistically significant regions of similar diversity levels, hot (areas of high diversity, shown by red patches) and cold patches (areas of low diversity, shown by blue patches) were assigned (fourth column). \u003cstrong\u003ef, g)\u003c/strong\u003e Bar plots showing significant features between early and late-stage disease of cellular co-occurrence scores at a global (f) and within hot-spots (g). Global (f) and hot-spot (g) co-occurrence was calculated between all combinations of tumour cells characterised by metabolic neighbourhoods and immune cells, including tumour/tumour and non-tumour/non-tumour combinations.\u003c/p\u003e","description":"","filename":"Figure4.png","url":"https://assets-eu.researchsquare.com/files/rs-7692105/v1/c62fa93a65cc401e016b0b91.png"},{"id":94198018,"identity":"efd9b2dc-fcf5-4621-a6c3-14ef9550acbf","added_by":"auto","created_at":"2025-10-23 13:26:03","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":3330765,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eOverview of the multi-scale feature generation and clinical modelling results from this study. a) \u003c/strong\u003eFeatures on the local to global scale were generated to capture tissue cell-type and metabolic composition and cell-cell / cell-functional interactions, from G-Cross Area Under the Curve (AUC) (left), JSD Score cell density overlaps (middle), to proportions and entropy measures at the scale of the core. A demonstration of the censoring mechanism that is integral to the Hanisch edge correction in the G-Cross calculation is shown in the left subpanel. \u003cstrong\u003eb)\u003c/strong\u003e An overview of the linear modelling and feature selection method to select clinically informative features that model binary recurrence using linear model fits (left) and time-to-recurrence using Cox Proportional Hazards (CoxPH) models (right). \u003cstrong\u003ec)\u003c/strong\u003e Time-dependent AUCs of features selected to model binary recurrence for a model fit to data seen in the feature selection pipeline (blue) and unseen data (magenta). \u003cstrong\u003ed)\u003c/strong\u003e Time-dependent AUCs using a Boosted Cox Survival model for selected features from CoxPH fits to model RFS – data seen in the feature selection is blue and data that is unseen is in magenta. \u003cstrong\u003ee)\u003c/strong\u003e Chord diagrams of selected G-Cross AUC features (top left) and JSD features (bottom left) for features that model binary recurrence, and \u003cstrong\u003ef)\u003c/strong\u003e JSD features (top left) and G-Cross features (bottom right) that model time-to-recurrence. To simplify the overall interpretation, the few metabolic cell type-free features are specified under the “tumour” classification. \u003cstrong\u003eg) \u003c/strong\u003eKaplan Meier analysis of significant features pulled from the multivariable selection, where blue represents lower values and grey higher values. Solid lines indicate the results from the validation group and dashed lines represent results from the training group.\u003c/p\u003e","description":"","filename":"Figure5.png","url":"https://assets-eu.researchsquare.com/files/rs-7692105/v1/e51c688d34bbaccf97146f20.png"},{"id":94198024,"identity":"96c86b51-84a1-4013-a88c-f51435ce4fc3","added_by":"auto","created_at":"2025-10-23 13:26:04","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":7194315,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eOverview of the deep learning method to differentiate clinically relevant regions of tissues in this study. a) \u003c/strong\u003eSubgraphs are formed from a KNN graph of cells, with subgraphs formed by k-hop sampling. Subgraphs are masked and augmented during training, with contrastive learning applied to differentiate the output embedding score by the patient recurrence. \u003cstrong\u003eb)\u003c/strong\u003eAn overview of the general architecture of the Graph Neural Network (GNN) that utilizes the cell-type labels and metabolic/functional positivity scores to profile the core subregions. \u003cstrong\u003ec)\u003c/strong\u003e An overview of the contrastive loss function optimized during the training epochs. \u003cstrong\u003ed)\u003c/strong\u003e Output predictions scores, sorted by patient recurrence values, show that subregions profiled with higher embedding scores are more associated with recurrence-likely cores. \u003cstrong\u003ee)\u003c/strong\u003e Averaged embedding scores across both training and validation cores show general alignment with the patient recurrence (left), with precision-recall (PR) AUC and accuracy values peaking around 0.7 for the validation cores (right) when varying the majority-vote classification threshold. \u003cstrong\u003ef)\u003c/strong\u003e Representative cores showing diverse regions that the model is learning to differentiate by cell-type, functional, and metabolic composition. \u003cstrong\u003eg)\u003c/strong\u003e Top: mean cell proportions and non-recurrence/recurrence proportion ratios within prediction compartments defined by recurrence-like and non-recurrence-like subgraph classifications. Bars represent mean measurements, and error bars represent 5% to 95% percentile ranges. Bottom:\u003cstrong\u003e \u003c/strong\u003emean protein and pathway positivity proportions and non-recurrence/recurrence positivity ratios in core compartments defined by the mean subgraph classifications for all cells.\u003c/p\u003e","description":"","filename":"Figure6.png","url":"https://assets-eu.researchsquare.com/files/rs-7692105/v1/6554cf566130c196a4806b8e.png"},{"id":94198505,"identity":"f7fce8b5-a753-4b32-957d-c92e75b68063","added_by":"auto","created_at":"2025-10-23 13:34:04","extension":"jpg","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":5218404,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eDifferential expression, gene set enrichment and feature association (with survival and recurrence) analyses results of the RNA data. a)\u003c/strong\u003eMA plot for LUAD vs LUSC in the tumour segments. \u003cstrong\u003eb)\u003c/strong\u003e Univariate Kaplan–Meier (KM) plots for \u003cem\u003eFN3K\u003c/em\u003e, \u003cem\u003eMST1R\u003c/em\u003e and \u003cem\u003eCYP4F11\u003c/em\u003e for OS in tumour segments. \u003cstrong\u003ec)\u003c/strong\u003e Word clouds for the top 6 functional enriched gene set clusters for LUAD vs LUSC in tumour segments (upregulated) and the corresponding gene sets names, DEG counts and significance for cluster 5. \u003cstrong\u003ed)\u003c/strong\u003e MA plots for recurrence vs non-recurrence in either tumour segments (left) of all samples or (right) LUSC only samples. \u003cstrong\u003ee)\u003c/strong\u003eUnivariate KM plot of \u003cem\u003eCCND2\u003c/em\u003e for Recurrence status in tumour segments. \u003cstrong\u003ef)\u003c/strong\u003e GSEA KEGG analysis (selected) for recurrence vs non-recurrence in all samples (Up/Down) or tumour LUSC segments (Down) showing the gene sets names, DEG counts and significance. \u003cstrong\u003eg) \u003c/strong\u003eGNN-predictedrecurrence vs non-recurrence in either tumour segments (left) of all samples or for LUSC only samples. \u003cstrong\u003eh)\u003c/strong\u003e Univariate KM plot of \u003cem\u003eSTATH\u003c/em\u003e and \u003cem\u003eRAB6B\u003c/em\u003efor OS status in tumour segments. \u003cstrong\u003ei)\u003c/strong\u003eWord clouds for the top 4 functional enriched gene set clusters for GNN-predicted recurrence vs Non-recurrence in LUAD segments (upregulated) and the corresponding gene sets names, DEG counts and significance for clusters 3 and 4. \u0026nbsp;\u003cstrong\u003ej) \u003c/strong\u003eGNN-predicted recurrence vs non-recurrence in non-tumour segments (left) of LUSC samples. \u003cstrong\u003ek)\u003c/strong\u003e Univariate KM plot of \u003cem\u003eGBP7\u003c/em\u003e for Recurrence in non-tumour segments.\u003cstrong\u003e l)\u003c/strong\u003e GSEA KEGG analysis (selected) for recurrence vs non-recurrence in all samples (Down) or non-tumour LUAD (Up/Down) and LUSC (Up) segments, showing the gene sets names, DEG counts and significance. DE analysis results presented as M (log ratio)-A (mean average) plots. \u003cem\u003eLimma-voom eBayes\u003c/em\u003e (with TREAT) pipeline used for DE analysis with a significance cutoff based on the adjusted \u003cem\u003ep\u003c/em\u003e value of ≤0.05 using the Benjamini–Hochberg procedure while the GNN-predicted recurrence groups were constructed based on a median cutoff. \u0026nbsp;GSEA significances are shown as -log10 of the p-values.\u003c/p\u003e","description":"","filename":"Figure7.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7692105/v1/c663741cfbda7a69182b606d.jpg"},{"id":94290496,"identity":"c3da48b5-7804-4a4b-9843-dd3c8e3398b2","added_by":"auto","created_at":"2025-10-27 11:19:40","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":38058101,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7692105/v1/0a8718ae-8bc4-48dc-882a-6a53a4eafe78.pdf"},{"id":94199803,"identity":"d0be7c90-2377-4495-bb60-eb57ef8ff8b8","added_by":"auto","created_at":"2025-10-23 13:42:04","extension":"docx","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":3248430,"visible":true,"origin":"","legend":"\u003cp\u003eSupplementary Figures with legends\u003c/p\u003e","description":"","filename":"SupplementaryFigures.docx","url":"https://assets-eu.researchsquare.com/files/rs-7692105/v1/e1b367f2384c723d913fd105.docx"},{"id":94198012,"identity":"a465bcb8-9a37-4002-b9e7-7201141432f0","added_by":"auto","created_at":"2025-10-23 13:26:03","extension":"png","order_by":2,"title":"","display":"","copyAsset":false,"role":"supplement","size":969221,"visible":true,"origin":"","legend":"Supplementary Figure 1","description":"","filename":"Supplementalfigure1.png","url":"https://assets-eu.researchsquare.com/files/rs-7692105/v1/70e2f49f43f7d9b132d7ba6e.png"},{"id":94201474,"identity":"3cf00924-625a-4e8a-8f93-ca0ca950969c","added_by":"auto","created_at":"2025-10-23 13:58:03","extension":"png","order_by":3,"title":"","display":"","copyAsset":false,"role":"supplement","size":453039,"visible":true,"origin":"","legend":"Supplementary Figure 2","description":"","filename":"Supplementalfigure2.png","url":"https://assets-eu.researchsquare.com/files/rs-7692105/v1/461bd416ebf9916d0f71b002.png"},{"id":94199802,"identity":"dac924db-d905-47d1-beb8-4d339b62ca2f","added_by":"auto","created_at":"2025-10-23 13:42:03","extension":"png","order_by":4,"title":"","display":"","copyAsset":false,"role":"supplement","size":427500,"visible":true,"origin":"","legend":"Supplementary Figure 3","description":"","filename":"Supplementalfigure3.png","url":"https://assets-eu.researchsquare.com/files/rs-7692105/v1/08ca8b4c43060e94db593293.png"},{"id":94198502,"identity":"ec400422-eb94-4a26-88fe-f7f9f43bf465","added_by":"auto","created_at":"2025-10-23 13:34:03","extension":"png","order_by":5,"title":"","display":"","copyAsset":false,"role":"supplement","size":1899760,"visible":true,"origin":"","legend":"Supplementary Figure 4","description":"","filename":"SupplementalFigure4.png","url":"https://assets-eu.researchsquare.com/files/rs-7692105/v1/19d6f76375a6153f22410d89.png"},{"id":94198029,"identity":"885f5aad-0260-475a-b467-4e360ecef31d","added_by":"auto","created_at":"2025-10-23 13:26:04","extension":"png","order_by":6,"title":"","display":"","copyAsset":false,"role":"supplement","size":2899851,"visible":true,"origin":"","legend":"Supplementary Figure 5","description":"","filename":"SupplementalFigure5.png","url":"https://assets-eu.researchsquare.com/files/rs-7692105/v1/89b6dfc2575ff7b40756e71c.png"},{"id":94198050,"identity":"905e5ee8-9c0d-4a3b-b775-3235b79df9e4","added_by":"auto","created_at":"2025-10-23 13:26:05","extension":"png","order_by":7,"title":"","display":"","copyAsset":false,"role":"supplement","size":590447,"visible":true,"origin":"","legend":"Supplementary Figure 6","description":"","filename":"SupplementalFigure6.png","url":"https://assets-eu.researchsquare.com/files/rs-7692105/v1/8bf131dd55e68fb9850ed41c.png"},{"id":94199805,"identity":"96c1c8ba-b74b-4323-a76c-7642c45bbda9","added_by":"auto","created_at":"2025-10-23 13:42:04","extension":"png","order_by":8,"title":"","display":"","copyAsset":false,"role":"supplement","size":2219638,"visible":true,"origin":"","legend":"\u003cp\u003eSupplementary Figure 7\u003c/p\u003e","description":"","filename":"SupplementalFigure7.png","url":"https://assets-eu.researchsquare.com/files/rs-7692105/v1/a1ed7cb5cc2145cb3c30615a.png"},{"id":94200307,"identity":"a52266cd-5388-4d0c-99ba-3f523458bb50","added_by":"auto","created_at":"2025-10-23 13:50:04","extension":"png","order_by":9,"title":"","display":"","copyAsset":false,"role":"supplement","size":1430154,"visible":true,"origin":"","legend":"\u003cp\u003eSupplementary Figure 8\u003c/p\u003e","description":"","filename":"SupplementalFigure8.png","url":"https://assets-eu.researchsquare.com/files/rs-7692105/v1/e745b10c272e0bb30ebcd2db.png"},{"id":94198027,"identity":"519b126c-f308-49b2-8089-d79b9c285adb","added_by":"auto","created_at":"2025-10-23 13:26:04","extension":"png","order_by":10,"title":"","display":"","copyAsset":false,"role":"supplement","size":799930,"visible":true,"origin":"","legend":"\u003cp\u003eSupplementary Figure 9\u003c/p\u003e","description":"","filename":"SupplementalFigure9.png","url":"https://assets-eu.researchsquare.com/files/rs-7692105/v1/53da046ba2ddf74ce6024f53.png"},{"id":94198054,"identity":"80c3e127-f2db-4869-b1c2-7b8d813a805f","added_by":"auto","created_at":"2025-10-23 13:26:05","extension":"png","order_by":11,"title":"","display":"","copyAsset":false,"role":"supplement","size":553988,"visible":true,"origin":"","legend":"Supplementary Figure 10","description":"","filename":"SupplementalFigure10.png","url":"https://assets-eu.researchsquare.com/files/rs-7692105/v1/0e4d4f07de6f541ed3f8c54f.png"},{"id":94198046,"identity":"a00be76b-2fd2-459c-8f1f-0dd08595f0cf","added_by":"auto","created_at":"2025-10-23 13:26:04","extension":"pdf","order_by":12,"title":"","display":"","copyAsset":false,"role":"supplement","size":44378,"visible":true,"origin":"","legend":"Code and Software Submission Checklist","description":"","filename":"sp1.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7692105/v1/408b476e74ab987401b7b17a.pdf"},{"id":94198036,"identity":"0f876a15-17dc-427b-9733-22e63d14f0bf","added_by":"auto","created_at":"2025-10-23 13:26:04","extension":"pdf","order_by":13,"title":"","display":"","copyAsset":false,"role":"supplement","size":86149,"visible":true,"origin":"","legend":"Reporting Summary","description":"","filename":"rs1.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7692105/v1/ab8a453f9804b6a90b7b339b.pdf"},{"id":94199816,"identity":"0721086b-8045-4f0e-9ae8-02f019f477ed","added_by":"auto","created_at":"2025-10-23 13:42:05","extension":"xlsx","order_by":14,"title":"","display":"","copyAsset":false,"role":"supplement","size":47494,"visible":true,"origin":"","legend":"Supplementary tables","description":"","filename":"Supplementarytables.xlsx","url":"https://assets-eu.researchsquare.com/files/rs-7692105/v1/ec7b97359d379f41ae2783a9.xlsx"}],"financialInterests":"\u003cb\u003eYes\u003c/b\u003e there is potential Competing Interest.\nJH and GM are employees of Surge Care. AK is on the Scientific Advisory Board for Omapix Solutions, European Spatial Biology Centre, Predxbio, Molecular Instruments, and Visiopharm.","formattedTitle":"Spatial Multiomic Profiling Identifies Metabolic and Inflammatory Signatures Driving Recurrence in Adjuvant-Treated NSCLC","fulltext":[{"header":"Introduction","content":"\u003cp\u003eNon-small cell lung cancer (NSCLC) is the predominant subtype of lung cancer and is the leading cause of cancer-related death worldwide, accounting for 1.8\u0026nbsp;million deaths annually\u003csup\u003e\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e\u003c/sup\u003e. The two most common subtypes of NSCLC include lung adenocarcinoma (LUAD) and lung squamous cell carcinoma (LUSC), contributing to 40% and 25% of lung cancer diagnoses, respectively\u003csup\u003e\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e\u003c/sup\u003e. Specific driver mutations that currently allow targeted therapies include kirsten rat sarcoma viral oncogene homologue (\u003cem\u003eKRAS\u003c/em\u003e), epithelial growth factor receptor \u003cem\u003e(EGFR)\u003c/em\u003e, and anaplastic lymphoma kinase \u003cem\u003e(ALK)\u003c/em\u003e genes, however, a greater understanding of the tumour microenvironment (TME) that spans histology, immunological composition and cancer stage is needed to aid prognoses for other therapeutic strategies, such as chemotherapy and immunotherapy.\u003c/p\u003e\u003cp\u003eSpatial multiomics can provide new insights into how the TME contributes towards clinical outcomes measured for endpoints such as recurrence free survival (RFS) or overall survival (OS)\u003csup\u003e\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e\u003c/sup\u003e. While bulk and single-cell approaches have yielded important insights into NSCLC biology, spatial characterisation of functionally and metabolically distinct cellular phenotypes and how they contribute to inter-cellular networks are likely to provide a deeper understanding of tumour\u0026ndash;immune interactions relevant to therapy and survival.\u003c/p\u003e\u003cp\u003eHere, we applied spatial multiomic approaches by combining high-plex spatial proteomics, whole-transcriptome spatial transcriptomics, and deep learning in an adjuvant chemotherapy\u0026ndash;treated NSCLC tissue cohort to profile 61 patients and over 1.1\u0026nbsp;million cells. High-plex cyclic immunofluorescence was combined with tumour/stromal-compartment whole transcriptome profiling to identify tumour-immune, functional, and metabolic characteristics of the NSCLC TME associated with tumour recurrence. A geometric deep learning model was developed to gain further insights into the recurrence-risk association of protein-differentiable regions in the tissue structure. Compartment-wise assessment of differentially expressed genes from whole-transcriptome profiling found recurrence-associated enriched pathways, such as metabolism and inflammation. This multifaceted overview of recurrence-risk in resected NSCLC tissues provides pointed insights into the distinct immunological and metabolic profiles of tumours and their associated recurrence risk after adjuvant chemotherapy.\u003c/p\u003e"},{"header":"Results","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e\u003ch2\u003ePatient cohort characteristics\u003c/h2\u003e\u003cp\u003eThe tissue microarrays (TMAs) used in this retrospective study were constructed using locally collected samples by medical oncologists from the Princess Alexandra Hospital (PAH). After filtering patients in the hospital database who satisfied all inclusion criteria (histology, age, treatment, tissue quality), two TMAs comprised of 176 samples from 88 patients were created using 2mm cores from pathology-reviewed, representative tumour regions.\u003c/p\u003e\u003cp\u003eThese TMAs were profiled using cyclic immunofluorescence (48-plex), H\u0026amp;E staining, and spatial transcriptomics (regional, whole transcriptome profiling). Pathological assessment of H\u0026amp;E images, evaluation of immunofluorescent staining patterns, and bioinformatic inspection of protein and RNA expression was performed to assess both tissue and assay quality. After quality control (QC) of the proteomics dataset, 61 patients contributed at least one sample to the analysis, with 48 of these patients contributing 2 cores. Transcriptome profiling of tumour and stromal compartments was performed on one core per patient. Following QC, transcriptomic data from 52 patients were used for downstream analysis (Fig.\u0026nbsp;1a, Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e). To identify statistically robust signatures, we distributed patient duplicates between \u0026ldquo;training\u0026rdquo; and \u0026ldquo;validation\u0026rdquo; groups (Fig.\u0026nbsp;1b). These groups were jointly used throughout the analysis and in the predictive modelling to statistically cross-validate findings from each group. In the instances where only one tumour core passed QC, these were randomly allocated across the training and validation groups. Spatial features were considered robust only if significant (\u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.05) in both training and validation groups.\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003eNon-small cell lung cancer (NSCLC) patient cohort characteristics\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"8\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c8\" colnum=\"8\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003eDeceased\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/th\u003e\u003cth align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/th\u003e\u003cth align=\"left\" colname=\"c7\"\u003e\u003cp\u003eRecurrence\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e\u003cb\u003eOverall\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003eY\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003eN\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003eY\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u003cp\u003eN\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c8\"\u003e\u003cp\u003eNot known\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eAll\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e\u003cb\u003e61\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e16\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e45\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003e28\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u003cp\u003e31\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c8\"\u003e\u003cp\u003e2\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eStage\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c4\" namest=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colspan=\"2\" nameend=\"c6\" namest=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eIIIA\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e16\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e26.2%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e9\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e7\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003e11\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u003cp\u003e4\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c8\"\u003e\u003cp\u003e1\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eIIB\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e28\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e45.9%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e4\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e24\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003e13\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u003cp\u003e15\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c8\"\u003e\u003cp\u003e0\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eIIA\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e6\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e9.8%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e1\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e5\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003e2\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u003cp\u003e4\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c8\"\u003e\u003cp\u003e0\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eIB\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e6\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e9.8%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e1\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e5\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003e0\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u003cp\u003e6\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c8\"\u003e\u003cp\u003e0\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eIIIB\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e3\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e4.9%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e1\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e2\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003e2\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u003cp\u003e1\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c8\"\u003e\u003cp\u003e0\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eIA3\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e2\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e3.3%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e0\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e2\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003e0\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u003cp\u003e1\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c8\"\u003e\u003cp\u003e1\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eHistology *\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eAdenocarcinoma\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e45\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e73.7%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e13\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e32\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003e21\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u003cp\u003e22\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c8\"\u003e\u003cp\u003e2\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eSquamous\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e12\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e19.7%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e1\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e11\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003e6\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u003cp\u003e6\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c8\"\u003e\u003cp\u003e0\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eLarge cell carcinoma\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e3\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e4.9%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e2\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e1\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003e1\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u003cp\u003e2\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c8\"\u003e\u003cp\u003e0\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eAdenosquamous\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e1\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e1.6%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e0\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e1\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003e0\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u003cp\u003e1\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c8\"\u003e\u003cp\u003e0\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eSmoking history\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eY\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e54\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e88.5%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e14\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e40\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003e23\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u003cp\u003e29\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c8\"\u003e\u003cp\u003e2\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eN\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e6\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e9.8%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e1\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e5\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003e4\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u003cp\u003e2\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c8\"\u003e\u003cp\u003e0\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eN/A\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e1\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e1.6%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e1\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e0\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003e1\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u003cp\u003e0\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c8\"\u003e\u003cp\u003e0\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eMutation\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c8\"\u003e\u0026nbsp;\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eNone\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e48\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e78.7%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e14\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e34\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003e20\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u003cp\u003e26\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c8\"\u003e\u003cp\u003e2\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eKRAS\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e5\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e8.2%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e0\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e5\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003e0\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u003cp\u003e5\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c8\"\u003e\u003cp\u003e0\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eEGFR\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e8\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e13.1%\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e2\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003e6\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c6\"\u003e\u003cp\u003e8\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c7\"\u003e\u003cp\u003e0\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c8\"\u003e\u003cp\u003e0\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003ctfoot\u003e\u003ctr\u003e\u003ctd colspan=\"8\"\u003e* \u003cem\u003eAdenocarcinoma or Squamous histologies were used for transcriptomic analysis. Multivariate analysis was performed on all histologies.\u003c/em\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tfoot\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003eCustom Phenocycler Fusion Panel\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"5\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eImmune characterisation\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eFunctional markers\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003eStructural markers\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c5\"\u003e\u003cp\u003eMetabolic enzymes\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eLymphocyte\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eCD4\u003c/p\u003e\u003cp\u003eCD8\u003c/p\u003e\u003cp\u003eCD45\u003c/p\u003e\u003cp\u003eCD3e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\" morerows=\"5\" rowspan=\"6\"\u003e\u003cp\u003eNaK-ATPase\u003c/p\u003e\u003cp\u003eCD45RO\u003c/p\u003e\u003cp\u003ePD-1\u003c/p\u003e\u003cp\u003eGranzyme B\u003c/p\u003e\u003cp\u003ePD-L1\u003c/p\u003e\u003cp\u003eFOXP3\u003c/p\u003e\u003cp\u003ePAX5\u003c/p\u003e\u003cp\u003eHLA-A\u003c/p\u003e\u003cp\u003eICOS(CD278)\u003c/p\u003e\u003cp\u003eLAG3\u003c/p\u003e\u003cp\u003eCD44\u003c/p\u003e\u003cp\u003eKi67\u003c/p\u003e\u003cp\u003eHLA-DR\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\" morerows=\"5\" rowspan=\"6\"\u003e\u003cp\u003eCollagen IV SMA\u003c/p\u003e\u003cp\u003eE-cadherin\u003c/p\u003e\u003cp\u003eCD31\u003c/p\u003e\u003cp\u003eCD34\u003c/p\u003e\u003cp\u003eVimentin\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\" morerows=\"5\" rowspan=\"6\"\u003e\u003cp\u003eCPT1A\u003c/p\u003e\u003cp\u003eG6PD\u003c/p\u003e\u003cp\u003epNRF2\u003c/p\u003e\u003cp\u003eGLUT1\u003c/p\u003e\u003cp\u003eATPA5\u003c/p\u003e\u003cp\u003eSDHA\u003c/p\u003e\u003cp\u003eIDH2\u003c/p\u003e\u003cp\u003eASCT2\u003c/p\u003e\u003cp\u003eCitrate synthase\u003c/p\u003e\u003cp\u003eLDHA\u003c/p\u003e\u003cp\u003eHexokinase1\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eB cell\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eCD20\u003c/p\u003e\u003cp\u003eCD21\u003c/p\u003e\u003cp\u003eCD19\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eMonocyte\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eCD68\u003c/p\u003e\u003cp\u003eCD14\u003c/p\u003e\u003cp\u003eCD163\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eDC\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eCD11c, CD11B\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eNK\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eCD57\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eTumour\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003ePan-Cytokeratin\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003c/div\u003e\n\u003ch3\u003eDefined proteomic characteristics for cell type annotation\u003c/h3\u003e\n\u003cp\u003eCell typing was performed by iterative unsupervised clustering based on mean protein expression. Cells were first labelled as either tumour/non-tumour cells and then each class further sub-clustered to resolve distinct immune and structural cell types\u003csup\u003e\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e\u003c/sup\u003e. The mean protein expression of markers in each cluster used to annotate cells is shown in \u003cb\u003eFig.\u0026nbsp;2a\u003c/b\u003e, with subsequent normalised cell type proportions per patient shown in \u003cb\u003eFig.\u0026nbsp;2b\u003c/b\u003e. Tumour cells were further striated based on HLA-DR, HLA-A, Ki67 and CD44 expression. Immune subsets encompassed B cells, CD4 T follicular helper (TFH) cells, CD4\u003csup\u003e+\u003c/sup\u003e / CD8\u003csup\u003e+\u003c/sup\u003e T cells (including cytotoxic and PD-1\u003csup\u003e+\u003c/sup\u003e subsets), and macrophages (M1, M2, PD-L1\u003csup\u003e+\u003c/sup\u003e, PD-L1\u003csup\u003e+\u003c/sup\u003e M2). Structural categories included fibroblasts, vessels, and miscellaneous stromal cells. These annotations were manually curated by mapping cells back to raw immunofluorescence and H\u0026amp;E images to confirm annotation accuracy. Neighbourhood clustering was performed using the tumour/non-tumour cell classes to create tumour, non-tumour and tumour-interface regions (Fig.\u0026nbsp;2c). These annotations enabled spatial dissection of tissue regions for downstream spatial analysis with disease recurrence.\u003c/p\u003e\n\u003ch3\u003eMetabolic annotation of cell phenotypes\u003c/h3\u003e\n\u003cp\u003ePathway definitions based on positive expression of key markers (Fig.\u0026nbsp;3a) were utilised to annotate tumour cells with activity for seven discrete metabolic pathways: glycolytic metabolism (GLYC\u003csup\u003ehigh\u003c/sup\u003e), oxidative phosphorylation (OXPH\u003csup\u003ehigh\u003c/sup\u003e), mitochondrial metabolism (MITC\u003csup\u003ehigh\u003c/sup\u003e), oxidative stress (OXSTR\u003csup\u003ehigh\u003c/sup\u003e), lipid oxidation (LPDOX\u003csup\u003ehigh\u003c/sup\u003e), glutamine import (GLN\u003csup\u003ehigh\u003c/sup\u003e), and pentose phosphate pathway (PPP\u003csup\u003ehigh\u003c/sup\u003e). Each cell was annotated by their joint metabolic pathways and confirmed via visual inspection of expression heatmaps and raw data (Fig.\u0026nbsp;3b-c). Figure\u0026nbsp;3b shows the average protein expression in tumour cells for each metabolic phenotype. These metabolic phenotypes were assessed both independently as well as with their associated cell type for downstream analysis in both tumour and non-tumour cell populations.\u003c/p\u003e\n\u003ch3\u003eSpatial features reveal T cell-driven immunological interactions associated with non-recurrence\u003c/h3\u003e\n\u003cp\u003eTo assess the cellular spatial characteristics of NSCLC tissues for their association with recurrence, we first calculated cellular densities at a whole core level and within tumour, stroma and tumour-stroma interface cellular neighbourhoods (CNs). Spatial features (proximity density, G-Cross AUC, JSD scores, and entropy/diversity measurements \u0026ndash; see Methods section) were calculated on both broad and granular cell-types. After filtering out sparse features, 16,384 spatial features were retained for analysis. Univariate Mann\u0026ndash;Whitney U-tests (MW) and Kaplan\u0026ndash;Meier (KM) tests were used to measure associations with recurrence and RFS\u003csup\u003e\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e\u003c/sup\u003e respectively. KM analysis identified 42 recurrence-associated features, while MW tests identified 23 significant features consistent between training \u003cem\u003e(p\u003c/em\u003e value\u0026thinsp;\u003cem\u003e=\u0026thinsp;p\u003c/em\u003e\u003csub\u003e\u003cem\u003et\u003c/em\u003e\u003c/sub\u003e,) and validation (p value\u0026thinsp;=\u0026thinsp;\u003cem\u003ep\u003c/em\u003e\u003csub\u003e\u003cem\u003ev)\u003c/em\u003e\u003c/sub\u003e groups (Supplementary Tables\u0026nbsp;1\u0026ndash;2). When the cohort was split into distinct histology types, 66 and 119 features were identified for LUSC samples using KM and MW tests, respectively. Conversely, 10 and 3 features were significant in the LUAD cohort (\u003cb\u003eSupplementary Tables\u0026nbsp;1 and 2\u003c/b\u003e).\u003c/p\u003e\u003cp\u003eUnivariate analyses highlighted specific spatial features within defined tissue compartments that were associated with tumour recurrence. The proximity of cytotoxic CD8 T cells to tumour cells was indicative of tumour recurrence in both the training and validation sets (\u003cem\u003ep\u003c/em\u003e\u003csub\u003e\u003cem\u003et\u003c/em\u003e\u003c/sub\u003e = 0.026, \u003cem\u003ep\u003c/em\u003e\u003csub\u003e\u003cem\u003ev\u003c/em\u003e\u003c/sub\u003e = 0.0113) (Fig.\u0026nbsp;2d\u003cb\u003e)\u003c/b\u003e. Broadly across immune types, PD-L1, PD-1, and Granzyme-B expression differed by recurrence status. In the stroma, PD-L1\u003csup\u003e+\u003c/sup\u003e Granzyme-B\u003csup\u003e+\u003c/sup\u003e immune cells were enriched in non-recurrence patients (\u003cem\u003ep\u003c/em\u003e\u003csub\u003e\u003cem\u003et\u003c/em\u003e\u003c/sub\u003e = 0.041, \u003cem\u003ep\u003c/em\u003e\u003csub\u003e\u003cem\u003ev\u003c/em\u003e\u003c/sub\u003e = 0.041) (Fig.\u0026nbsp;2e). Conversely, higher proportions of PD-1\u003csup\u003e+\u003c/sup\u003e Granzyme-B\u003csup\u003e+\u003c/sup\u003e immune cells in the tumour-stroma interface neighbourhood were associated with shorter time-to-recurrence (\u003cem\u003ep\u003c/em\u003e\u003csub\u003e\u003cem\u003et\u003c/em\u003e\u003c/sub\u003e = 0.031, \u003cem\u003ep\u003c/em\u003e\u003csub\u003e\u003cem\u003ev\u003c/em\u003e\u003c/sub\u003e = 0.034) (Fig.\u0026nbsp;2f). Across all compartments, specifically stroma, tumour, and tumour\u0026ndash;stroma interface CNs, patients with recurrence had higher proportions of PD-1\u003csup\u003e+\u003c/sup\u003e immune cells (StromaCN: \u003cem\u003ep\u003c/em\u003e\u003csub\u003e\u003cem\u003et\u003c/em\u003e\u003c/sub\u003e = 0.0196, \u003cem\u003ep\u003c/em\u003e\u003csub\u003e\u003cem\u003ev\u003c/em\u003e\u003c/sub\u003e = 0.0198, TumourCN: \u003cem\u003ep\u003c/em\u003e\u003csub\u003e\u003cem\u003et\u003c/em\u003e\u003c/sub\u003e = 0.0061, \u003cem\u003ep\u003c/em\u003e\u003csub\u003e\u003cem\u003ev\u003c/em\u003e\u003c/sub\u003e = 0.037, Tumour/Stroma InterfaceCN: \u003cem\u003ep\u003c/em\u003e\u003csub\u003e\u003cem\u003et\u003c/em\u003e\u003c/sub\u003e = 0.0105, \u003cem\u003ep\u003c/em\u003e\u003csub\u003e\u003cem\u003ev\u003c/em\u003e\u003c/sub\u003e = 0.0094) \u003cb\u003e(Fig.\u0026nbsp;2g)\u003c/b\u003e. Taken together, these results highlight the conflicting roles that PD-1 and PD-L1 expression on immune cells have on disease recurrence in adjuvant treated NSCLC.\u003c/p\u003e\n\u003ch3\u003eSpatial features revealed fibroblast-driven and immune cell-driven metabolic niches associated with tumour recurrence\u003c/h3\u003e\n\u003cp\u003eWe observed that significant recurrence-associated cell-cell proximities defined by the G-cross function most frequently included combinations of a cell phenotype and a metabolic pathway, indicating the prominent role that cell metabolism plays in the TME (Fig.\u0026nbsp;3e). In particular, higher proximity of PPP\u003csup\u003ehigh\u003c/sup\u003e/GLN\u003csup\u003ehigh\u003c/sup\u003e fibroblasts to GLN\u003csup\u003ehigh\u003c/sup\u003e fibroblasts were associated with more favourable outcomes (\u003cem\u003ep\u003c/em\u003e\u003csub\u003e\u003cem\u003et\u003c/em\u003e\u003c/sub\u003e = 0.0099, \u003cem\u003ep\u003c/em\u003e\u003csub\u003e\u003cem\u003ev\u003c/em\u003e\u003c/sub\u003e = 0.0059), and a similar association was seen between GLYC\u003csup\u003ehigh\u003c/sup\u003e/PPP\u003csup\u003ehigh\u003c/sup\u003e/GLN\u003csup\u003ehigh\u003c/sup\u003e fibroblasts to GLN\u003csup\u003ehigh\u003c/sup\u003e fibroblasts (\u003cem\u003ep\u003c/em\u003e\u003csub\u003e\u003cem\u003et\u003c/em\u003e\u003c/sub\u003e = 0.0037, \u003cem\u003ep\u003c/em\u003e\u003csub\u003e\u003cem\u003ev\u003c/em\u003e\u003c/sub\u003e = 0.0092) (Fig.\u0026nbsp;3f). Conversely, higher self-proximity of PPP\u003csup\u003ehigh\u003c/sup\u003e/GLN\u003csup\u003ehigh\u003c/sup\u003e CD4 T cells was unfavourable (\u003cem\u003ep\u003c/em\u003e\u003csub\u003e\u003cem\u003et\u003c/em\u003e\u003c/sub\u003e = 0.029, \u003cem\u003ep\u003c/em\u003e\u003csub\u003e\u003cem\u003ev\u003c/em\u003e\u003c/sub\u003e = 0.011) (Fig.\u0026nbsp;3g). Other significant associations were identified largely between immune cells with distinct metabolic types (\u003cb\u003eSupplementary Table\u0026nbsp;1\u003c/b\u003e); to illustrate this, a relationship connectivity web of the main findings with respect to the cell-pair proximities are shown in \u003cb\u003eFig.\u0026nbsp;3d\u003c/b\u003e (favourable associations are shown with green arrows and unfavourable associations with red arrows). In the LUSC samples, significant changes in metabolic-type proportions at a whole core level for CD4 T cells, fibroblasts, and macrophages were associated with tumour recurrence. Proportions of GLN\u003csup\u003ehigh\u003c/sup\u003e/GLYC\u003csup\u003ehigh\u003c/sup\u003e CD4 T cells (\u003cem\u003ep\u003c/em\u003e\u003csub\u003e\u003cem\u003et\u003c/em\u003e\u003c/sub\u003e = 0.013, \u003cem\u003ep\u003c/em\u003e\u003csub\u003e\u003cem\u003ev\u003c/em\u003e\u003c/sub\u003e = 0.038), GLN\u003csup\u003ehigh\u003c/sup\u003e/GLYC\u003csup\u003ehigh\u003c/sup\u003e macrophages (\u003cem\u003ep\u003c/em\u003e\u003csub\u003e\u003cem\u003et\u003c/em\u003e\u003c/sub\u003e = 0.0078, \u003cem\u003ep\u003c/em\u003e\u003csub\u003e\u003cem\u003ev\u003c/em\u003e\u003c/sub\u003e = 0.040), GLN\u003csup\u003ehigh\u003c/sup\u003e macrophages (\u003cem\u003ep\u003c/em\u003e\u003csub\u003e\u003cem\u003et\u003c/em\u003e\u003c/sub\u003e = 0.026, \u003cem\u003ep\u003c/em\u003e\u003csub\u003e\u003cem\u003ev\u003c/em\u003e\u003c/sub\u003e = 0.0095), and GLN\u003csup\u003ehigh\u003c/sup\u003e fibroblasts (\u003cem\u003ep\u003c/em\u003e\u003csub\u003e\u003cem\u003et\u003c/em\u003e\u003c/sub\u003e = 0.015, \u003cem\u003ep\u003c/em\u003e\u003csub\u003e\u003cem\u003ev\u003c/em\u003e\u003c/sub\u003e = 0.0095) (\u003cb\u003eSupplementary Fig.\u0026nbsp;2c\u003c/b\u003e) were higher in patients where recurrence did not occur. Conversely, LPDOX\u003csup\u003ehigh\u003c/sup\u003e/PPP\u003csup\u003ehigh\u003c/sup\u003e CD4 T cells (\u003cem\u003ep\u003c/em\u003e\u003csub\u003e\u003cem\u003et\u003c/em\u003e\u003c/sub\u003e = 0.0078, \u003cem\u003ep\u003c/em\u003e\u003csub\u003e\u003cem\u003ev\u003c/em\u003e\u003c/sub\u003e = 0.013) and LPDOX\u003csup\u003ehigh\u003c/sup\u003e/PPP\u003csup\u003ehigh\u003c/sup\u003e fibroblasts (\u003cem\u003ep\u003c/em\u003e\u003csub\u003e\u003cem\u003et\u003c/em\u003e\u003c/sub\u003e = 0.034, \u003cem\u003ep\u003c/em\u003e\u003csub\u003e\u003cem\u003ev\u003c/em\u003e\u003c/sub\u003e = 0.042) were in greater proportion in patients who did have recurrence (\u003cb\u003eSupplementary Table\u0026nbsp;2, Supplementary Fig.\u0026nbsp;2b\u003c/b\u003e). Overall, we found that PPP\u003csup\u003ehigh\u003c/sup\u003e, GLN\u003csup\u003ehigh\u003c/sup\u003e, and LPDOX\u003csup\u003ehigh\u003c/sup\u003e in fibroblasts, macrophages and CD8\u003csup\u003e+\u003c/sup\u003e T cells, as well as PPP\u003csup\u003ehigh\u003c/sup\u003e and GLN\u003csup\u003ehigh\u003c/sup\u003e in CD4 T cells and macrophages to be associated with RFS. This highlights a potential feature of immune suppressive regulatory CD4 T cells and M2 macrophages and of their metabolic relationships with the TME.\u003c/p\u003e\u003cp\u003e\u003cb\u003eCharacterisation of tissue metabolic neighbourhoods identifies disease stage-dependence of the metabolic diversity and non-tumour cell co-occurrence\u003c/b\u003e\u003c/p\u003e\u003cp\u003eWe utilised an orthogonal approach to assigning metabolic phenotypes by generating metabolic marker expression-derived neighbourhoods (Fig.\u0026nbsp;4a, \u003cb\u003eSupplementary Fig.\u0026nbsp;3\u003c/b\u003e)\u003csup\u003e\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e\u003c/sup\u003e. Eight unique neighbourhoods were established and annotated based on the combination of marker expressions (Fig.\u0026nbsp;4a\u003cb\u003e).\u003c/b\u003e\u003c/p\u003e\u003cp\u003eFor entropy/diversity analyses, tumour cells were assigned to these metabolic neighbourhoods, while non-tumour cells were annotated by functional state. Figure\u0026nbsp;4b showcases the differences in spatial neighbourhood distributions between an early-stage and an advanced sample. Based on Shannon diversity scores calculated at different scales, there were significant differences between early- and late-stage disease at scale 48 (\u003cem\u003ep\u003c/em\u003e\u003csub\u003e\u003cem\u003et\u003c/em\u003e\u003c/sub\u003e = 0.030, \u003cem\u003ep\u003c/em\u003e\u003csub\u003e\u003cem\u003ev\u003c/em\u003e\u003c/sub\u003e = 0.004), and partial significances at scales 24 (\u003cem\u003ep\u003c/em\u003e\u003csub\u003e\u003cem\u003et\u003c/em\u003e\u003c/sub\u003e = 0.053, \u003cem\u003ep\u003c/em\u003e\u003csub\u003e\u003cem\u003ev\u003c/em\u003e\u003c/sub\u003e = 0.029) and 32 (\u003cem\u003ep\u003c/em\u003e\u003csub\u003e\u003cem\u003et\u003c/em\u003e\u003c/sub\u003e = 0.051, \u003cem\u003ep\u003c/em\u003e\u003csub\u003e\u003cem\u003ev\u003c/em\u003e\u003c/sub\u003e = 0.023) \u003cb\u003e(Fig.\u0026nbsp;4c).\u003c/b\u003e The multiscale diversity index (MDI) \u003cb\u003e(Fig.\u0026nbsp;4b\u0026ndash;c)\u003c/b\u003e, which evaluates changes in Shannon diversity across spatial scales, demonstrated that advanced staged samples possessed higher MDI scores (\u003cem\u003ep\u003c/em\u003e\u003csub\u003e\u003cem\u003et\u003c/em\u003e\u003c/sub\u003e = 0.011, \u003cem\u003ep\u003c/em\u003e\u003csub\u003e\u003cem\u003ev\u003c/em\u003e\u003c/sub\u003e = 0.0079) \u003cb\u003e(Fig.\u0026nbsp;4c)\u003c/b\u003e, indicating that these samples have noticeable changes in diversity across spatial scales. Global diversity index (GDI) was then derived from the MDI results, to evaluate the spatial adjacency of patches with high or low diversity. The results showed a tendency of lower GDI scores in advanced stage samples (\u003cem\u003ep\u003c/em\u003e\u003csub\u003e\u003cem\u003et\u003c/em\u003e\u003c/sub\u003e = 0.051, \u003cem\u003ep\u003c/em\u003e\u003csub\u003e\u003cem\u003ev\u003c/em\u003e\u003c/sub\u003e = 0.023), indicating that the tissue architecture of later stage of disease is typically less structured and has more blended cellular distributions (Fig.\u0026nbsp;4d). Moran\u0026rsquo;s I was used to identify statistically significant areas of similar diversity, based on local diversity index (LDI) calculations, where regions of high entropy were considered hot spots and areas of low entropy were cold spots. Example visualisations of the cell phenotypes considered in the analysis, local diversity index, Moran\u0026rsquo;s I scores, and the identified hot/cold spots are shown in \u003cb\u003eFig.\u0026nbsp;4e\u003c/b\u003e. The results found that stage-dependent co-occurrence patterns were evident across the cohort. At a global level, PD-L1\u003csup\u003e+\u003c/sup\u003e macrophages co-occurred more frequently with CD8\u003csup\u003e+\u003c/sup\u003e T cells (\u003cem\u003ep\u003c/em\u003e\u003csub\u003e\u003cem\u003et\u003c/em\u003e\u003c/sub\u003e = 0.0335, \u003cem\u003ep\u003c/em\u003e\u003csub\u003e\u003cem\u003ev\u003c/em\u003e\u003c/sub\u003e = 0.0364), fibroblasts (\u003cem\u003ep\u003c/em\u003e\u003csub\u003e\u003cem\u003et\u003c/em\u003e\u003c/sub\u003e = 0.0229, \u003cem\u003ep\u003c/em\u003e\u003csub\u003e\u003cem\u003ev\u003c/em\u003e\u003c/sub\u003e = 0.0217), and stromal cells (\u003cem\u003ep\u003c/em\u003e\u003csub\u003e\u003cem\u003et\u003c/em\u003e\u003c/sub\u003e = 0.0453, \u003cem\u003ep\u003c/em\u003e\u003csub\u003e\u003cem\u003ev\u003c/em\u003e\u003c/sub\u003e = 0.0260) in advanced-stage samples (Fig.\u0026nbsp;4f). Within hot-spots, early-stage samples showed a higher co-occurrence of M2 macrophages with PD-L1\u003csup\u003e+\u003c/sup\u003e macrophages (\u003cem\u003ep\u003c/em\u003e\u003csub\u003e\u003cem\u003et\u003c/em\u003e\u003c/sub\u003e = 0.0376, \u003cem\u003ep\u003c/em\u003e\u003csub\u003e\u003cem\u003ev\u003c/em\u003e\u003c/sub\u003e = 0.0187), and PD-L1\u003csup\u003e+\u003c/sup\u003e M2 macrophages with stromal cells (\u003cem\u003ep\u003c/em\u003e\u003csub\u003e\u003cem\u003et\u003c/em\u003e\u003c/sub\u003e = 0.0154, \u003cem\u003ep\u003c/em\u003e\u003csub\u003e\u003cem\u003ev\u003c/em\u003e\u003c/sub\u003e = 0.0399) (Fig.\u0026nbsp;4g). Taken together, these findings suggest that the TME of later-stage disease is more heterogeneous and less structured, showing a broad energy dysregulation that may contribute to recurrence risk.\u003c/p\u003e\u003cdiv id=\"Sec8\" class=\"Section2\"\u003e\u003ch2\u003eMultivariable modelling found signatures of coordinated immune and metabolic interactions contributing to tumour recurrence\u003c/h2\u003e\u003cp\u003eWe built an integrated model of clinical response using spatial metrics that capture cell\u0026ndash;cell interactions and tissue organisation in the TME. Feature selection was applied to spatial features computed across spatial scales in the core (Fig.\u0026nbsp;5a). These features, G-Cross AUC, JSD Scores, Entropy/Diversity, and proportions, capture the tissue composition, cell configuration, and local cell-cell interactions at a variety of scales and were computed for each CN and/or over the whole core. Features were selected to model two clinical endpoints, recurrence status and RFS, and were then validated by CoxPH and Boosted Cox Survival model fits to data that were unseen by the feature selection algorithm. A distinct 10% difference in time-dependent AUCs was observed between the model fits to features seen by the feature selection algorithm and the held-out validation data (Fig.\u0026nbsp;5c\u003cb\u003e/5d\u003c/b\u003e); this demonstrates the intractable nature of clinical modelling that are complicated by the tumour heterogeneity, even for the relatively large spatial coverage of these 2mm cores.\u003c/p\u003e\u003cp\u003eFeatures selected using this method are profiled in \u003cb\u003eFig.\u0026nbsp;5e/5f\u003c/b\u003e and in \u003cb\u003eSupplementary Figs.\u0026nbsp;5\u0026ndash;8\u003c/b\u003e, and exhibit a diverse set of proximity features, density overlap signatures, and proportions for a variety of cell types. The contribution of these features to the overall Boosted Cox Survival model fit to all cores in the cohort is highlighted in \u003cb\u003eSupplementary Fig.\u0026nbsp;5\u003c/b\u003e, showing SHAP values that exhibit the diverse response of the model to each feature in the core. Further Kaplan\u0026ndash;Meier analyses (Fig.\u0026nbsp;5g; \u003cb\u003eSupplementary Fig.\u0026nbsp;8\u003c/b\u003e) confirmed associations of these selected features with RFS and OS. This profile of the TME shows the coordinated architecture of the interactions of stromal cells, immune cells, and tumour cells that collectively model the TME, highlighting the dynamics of M2 macrophage proximity to cytotoxic CD8 T cells, the role of exhausted immune cells, and of the PPP pathway in driving tumour recurrence.\u003c/p\u003e\u003c/div\u003e\n\u003ch3\u003eGeometric deep learning found recurrence-associated spatial protein regions\u003c/h3\u003e\n\u003cp\u003eTraditional measurement and modelling of tissue features is complicated by tumour heterogeneity. Given the broad range of immunological, metabolic and tumour regions in our 2mm cores, we implemented a graph neural network (GNN) method to annotate regions of the core that may distinguish recurrence status. We utilized contrastive learning to separate a single output embedding from subregions of the core, where subregions were represented as graphs containing representations of cell-types, functional type positivity (PD-L1\u003csup\u003e+\u003c/sup\u003e, PD-1\u003csup\u003e+\u003c/sup\u003e, Ki67\u003csup\u003e+\u003c/sup\u003e, etc.), metabolic marker positivity, and metabolic pathway positivity. Figure\u0026nbsp;6a shows a method and training flowchart, and \u003cb\u003eFig.\u0026nbsp;6b\u003c/b\u003e gives an overview of the architecture of the model. The model was trained on the training dataset, which corresponded to approximately 600k cells. Post-training, the subgraph output embeddings were pooled for each core and plotted by the actual recurrence status in \u003cb\u003eFig.\u0026nbsp;6c\u003c/b\u003e, indicating that the output embeddings tend to predominantly differentiate recurrence-like regions from non-recurrence-like regions. Figure\u0026nbsp;6d shows the averaged embedding score for each core on the left, and the accuracy and precision metrics for the validation data. Examples of these embedding annotations are shown overlaid on cores in \u003cb\u003eFig.\u0026nbsp;6e\u003c/b\u003e, indicating that the model distinctly stratifies regions of the core that exhibit a diversity of cell-types, tissue structures, and immunological niches.\u003c/p\u003e\u003cp\u003e\u003cb\u003eFigure 6f\u003c/b\u003e shows the relationship between proportions of cell-types, marker positivity, and metabolic pathways with the embedding annotations, showing the characteristics of these predicted regions. Overall, the highest cell-type and marker positivity values exhibited within non-recurrence regions tended to be antigen expressing tumour phenotypes such as CD44\u003csup\u003e+\u003c/sup\u003e, MHC class I (HLA-A), and PD-L1\u003csup\u003e+\u003c/sup\u003e. Also highly expressed within these non-recurrence regions were cytotoxic CD8\u003csup\u003e+\u003c/sup\u003e T cells and PD-L1\u003csup\u003e+\u003c/sup\u003e M2 macrophages. The differentiable and primarily recurrence-associated regions were composed of B cells, MHC class II (HLA-DR) expressing tumour cells, and proliferating (Ki67\u003csup\u003e+\u003c/sup\u003e) tumour cells, as well as tumours without antigen expression. Upregulation of PD-L1, HLA-A, GLUT1, ASCT2, and PD-1 expression were noted in non-recurrence regions, as opposed to an upregulation of the PPP and Oxidative Phosphorylation pathways in the recurrence regions. Glutamine import, GLUT1\u003csup\u003e+\u003c/sup\u003e, and glycolytic pathways were also enriched in these non-recurrence regions broadly across tumours, immune cells, and structural cells (\u003cb\u003eSupplementary Fig.\u0026nbsp;9\u003c/b\u003e). Cell proportions in radial neighbourhoods associated with each annotated subgraph showed similar cell-type distributions as in the compartments.\u003c/p\u003e\n\u003ch3\u003eSpatial transcriptomics revealed gene and pathway expression changes associated with tumour recurrence and histological classifications\u003c/h3\u003e\n\u003cp\u003eTo validate the proteomic findings, we performed spatial transcriptomics on adjacent TMA sections using the Nanostring GeoMx Digital Spatial Profiler. Each Area of Interest (AOI), containing one Region of Interest (ROI) separated into tumour or stromal compartment regions, was digitally profiled as previously described\u003csup\u003e\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e\u003c/sup\u003e. The data were subjected to processing using R package \u003cem\u003estandR\u003c/em\u003e\u0026rsquo;s quality control (QC) pipeline\u003csup\u003e\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e\u003c/sup\u003e. After QC, 104 ROIs samples were assessed for sample and gene quality, with 6 ROIs removed based on assessment of the nuclei counts and library size. Principal component analysis (PCA) did not identify any confounding factors and suggested the main variance is attributable to differences in segmentation (i.e. location) and to histological subtype (\u003cb\u003eSupplementary Fig.\u0026nbsp;10a\u003c/b\u003e). The data was then Trimmed Mean of M-values (TMM) normalised, correcting for any systemic bias (\u003cb\u003eSupplementary Fig.\u0026nbsp;10b\u003c/b\u003e).\u003c/p\u003e\u003cp\u003eWhen comparing the RNA expression of the tumour regions between LUAD and LUSC samples, we identified 369 upregulated transcripts and 230 downregulated transcripts (\u003cb\u003eSupplementary Table\u0026nbsp;3)\u003c/b\u003e. Of these significantly differentially expressed genes, keratin 5 (\u003cem\u003eKRT5\u003c/em\u003e) was upregulated in LUSC samples whereas \u003cem\u003eKRT7\u003c/em\u003e was higher in LUAD (Fig.\u0026nbsp;7a). Amongst the DEGs, fructosamine-3-kinase (\u003cem\u003eFN3K\u003c/em\u003e) and macrophage-stimulating 1 receptor (\u003cem\u003eMST1R\u003c/em\u003e) were upregulated in LUAD while cytochrome P450, family 4, subfamily F, polypeptide 11 (\u003cem\u003eCYP4F11\u003c/em\u003e) were upregulated in LUSC. These gene expressions were also significant in the tumour segments, where they were associated with a worse OS (Fig.\u0026nbsp;7b). Cell differentiation, goblet cell, epithelial, and gland development pathways were all identified to be upregulated within the LUAD cohort (Fig.\u0026nbsp;7c).\u003c/p\u003e\u003cp\u003eDifferential expression analysis was also performed between patients with tumour recurrence and those without, on all samples and on each histological segment. No DEGs were identified in any of the non-tumour comparisons with limited DEGs identified in the tumour segment. Melanoma associated antigen family A9B (\u003cem\u003eMAGEA9B\u003c/em\u003e) was upregulated in LUAD and is associated with poor survival\u003csup\u003e\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e\u003c/sup\u003e. Cyclin D2 (\u003cem\u003eCCND2\u003c/em\u003e) was upregulated in LUSC and is associated with survival and immune infiltration (Fig.\u0026nbsp;7d)\u003csup\u003e\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e\u003c/sup\u003e. \u003cem\u003eCCND2\u003c/em\u003e was identified in our feature association analysis where higher expression was associated with better RFS (Fig.\u0026nbsp;7e). Pathway analysis revealed downregulation of viral KEGG pathways including measles, HPV, EBV, and viral carcinogenesis in recurrence patients (Fig.\u0026nbsp;7f\u003cb\u003e)\u003c/b\u003e.\u003c/p\u003e\u003cp\u003eLeveraging on the capabilities of the GNN approach, we utilised the GNN based modelling to estimate recurrence-like compartments and subsequently predicted recurrence status per sample. Similarly, DE analysis was conducted comparing patients by GNN-predicted recurrence either in all samples or in either segment. Limited DEGs were identified for all the contrasts with most DEGs found in the tumour segments (\u003cb\u003eSupplementary Table\u0026nbsp;3\u003c/b\u003e). Specifically, matrix metallopeptidase 10 (\u003cem\u003eMMP10\u003c/em\u003e) has been known to be required for lung cancer maintenance, metastasis, and recurrence\u003csup\u003e\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e,\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e\u003c/sup\u003e (Fig.\u0026nbsp;7g). The gene encoding solute carrier family 38 member 5 (\u003cem\u003eSLC38A5\u003c/em\u003e) is a regulator of glutaminolysis. Feature association analysis identified statherin (\u003cem\u003eSTATH\u003c/em\u003e) gene expression associated with worse OS in the tumour while the expression of the gene encoding \u003cem\u003eRAB6B\u003c/em\u003e was associated with better OS in the tumour (Fig.\u0026nbsp;7h). Pathway analysis suggested an enrichment of metabolic processes (including lipid metabolism) in the adenocarcinoma recurrence samples (Fig.\u0026nbsp;7i).\u003c/p\u003e\u003cp\u003eFor the non-tumour segments, DEGs were again limited to 4 DEGs upregulated in the patients with GNN-predicted recurrence (\u003cb\u003eSupplementary Table\u0026nbsp;3\u003c/b\u003e). Of these, peptidyl arginine deiminase 4 (\u003cem\u003ePADI4\u003c/em\u003e) is overexpressed in lung cancer and linked to EMT and metastasis\u003csup\u003e\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e,\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e\u003c/sup\u003e, and tumour necrosis factor ligand superfamily member 18 (\u003cem\u003eTNFSF18\u003c/em\u003e) is a potential biomarker for predicting the efficacy of anti-PD-1 treatments\u003csup\u003e\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e\u003c/sup\u003e (Fig.\u0026nbsp;7j). Feature association analysis linked the gene encoding guanylate binding protein 7 (\u003cem\u003eGBP7\u003c/em\u003e) (a negative regulator of innate immune responses, specifically IFN signalling)\u003csup\u003e\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e\u003c/sup\u003e expression in the non-tumour segment with a worse RFS (Fig.\u0026nbsp;7k). Pathway analysis suggested an overall enrichment of T cell receptor signalling processes in non-recurrence patients. In the non-tumour segments, NOD-like receptor signalling and neutrophil extracellular trap formation were enriched in the LUSC patients who did not have tumour recurrence while the complement and coagulation pathways were enriched in LUAD patients (Fig.\u0026nbsp;7l).\u003c/p\u003e\u003cp\u003eTaken together, the findings from the RNA analysis align with the major functional outcomes from the protein analysis and the GNN-modelling, specifically the observation of the enrichment of lipid metabolism and other metabolic processes in recurrence patients when stratified using GNN-predicted classifications of patients.\u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003eSpatial multiomic profiling provides deeper insight of the NSCLC TME beyond bulk or single-cell approaches. We identified cellular phenotypes, functional markers, and metabolic alterations associated with tumour recurrence and disease stage in the adjuvant chemotherapy setting. Altered cell dependency and metabolism of glutamine, glucose, and lipids are well characterised and tend to support mechanisms enabling the survival and adaptation of cancer cells\u003csup\u003e\u003cspan additionalcitationids=\"CR17 CR18 CR19\" citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e–\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e\u003c/sup\u003e. Our data suggests that metabolic diversity shifts with disease stage, and this may be reflective of cellular adaptation to changes in extracellular nutrient availability. This allows cancer cells to acquire the metabolic characteristics required for metastasis spread, driven by an energy independence to energy shortfalls experienced during circulation\u003csup\u003e\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e\u003c/sup\u003e. Here, we show that the compositional diversity of NSCLC metabolic phenotypes was found to be associated with disease stage, with global and “metabolic hot spot” co-occurrence of critical immune cells, such as PD-L1\u003csup\u003e+\u003c/sup\u003e macrophages, decreasing across stages\u003csup\u003e\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e\u003c/sup\u003e. Global metabolic diversity across length scales (from local neighbourhoods to the whole tissue core) was found to decrease in late-stage disease, reflecting an adaptation to changing energy regimes across the tissue.\u003c/p\u003e\u003cp\u003eIn both the transcriptomic and proteomic data, lipid metabolism emerged as a consistent determinant of recurrence risk. In particular, pathway analysis identified lipid, sterol, and cholesterol transport pathways enriched in tumour regions of recurrence patients. In the protein analysis, carnitine palmitoyltransferase I (CPT1a), a key enzyme in fatty acid oxidation, was used to represent the lipid oxidation pathway. In tandem with CPT1a, our analysis highlighted a role for G6PD, which functions in the pentose phosphate pathway (PPP) pathway. G6PD is a rate-limiting enzyme involved in the PPP and is well described in lung cancer oncogenesis, chemotherapy resistance, and immune checkpoint efficacy\u003csup\u003e\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e,\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e,\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e\u003c/sup\u003e. G6PD contributes to nucleotide synthesis and lipid maintenance through generation of nicotinamide adenine dinucleotide phosphate (NADPH), required for reductions in \u003cem\u003ede novo\u003c/em\u003e lipogenesis\u003csup\u003e\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e\u003c/sup\u003e. G6PD and CPT1a are commonly overexpressed in tumour cells, and their dual inhibition has caused significant reduction in tumour growth and improvement to OS in breast cancer and glioblastoma\u003csup\u003e\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e,\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e\u003c/sup\u003e. Notably, our study identified that the dual expression of CPT1a and G6PD and their association with tumour recurrence was dependent upon cellular organisation and pathways. For example, favourable outcomes were identified when these metabolic populations were in closer proximity: LPDOX\u003csup\u003ehigh\u003c/sup\u003e/PPP\u003csup\u003ehigh\u003c/sup\u003e/GLN\u003csup\u003ehigh\u003c/sup\u003e cells to GLN\u003csup\u003ehigh\u003c/sup\u003e cells, and LPDOX\u003csup\u003ehigh\u003c/sup\u003e/PPP\u003csup\u003ehigh\u003c/sup\u003e/GLN\u003csup\u003ehigh\u003c/sup\u003e cells to GLYC\u003csup\u003ehigh\u003c/sup\u003e/GLN\u003csup\u003ehigh\u003c/sup\u003e cells. Higher proportions of LPDOX\u003csup\u003ehigh\u003c/sup\u003e/PPP\u003csup\u003ehigh\u003c/sup\u003e/GLN\u003csup\u003ehigh\u003c/sup\u003e were identified to be unfavourable unless these populations were in proximity to the GLN\u003csup\u003ehigh\u003c/sup\u003e (negative for other metabolic markers) population. While these findings highlight the involvement of CPT1a and G6PD in the metabolic reprogramming that supports recurrence, it also points to potential involvement and competition for extracellular glutamine. This could support previous studies that showed the inhibition of ASCT2 or the depletion of glutamine availability can contribute to decreases in lipid and fatty acid biosynthesis\u003csup\u003e\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e\u003c/sup\u003e. In the context of this study, while the dual expression of ASCT2 and CPT1a is a less established phenomenon, there is recent evidence to suggest that ASCT2 may be stabilising CPT1a activity \u003csup\u003e\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e\u003cp\u003eIn non-tumour cells, our study identified the combination of glutamine import and PPP had unfavourable outcomes when enriched in fibroblasts, CD4 T cells, CD8 T cells, and macrophages. Our multivariate feature analysis also identified that the proximity of CD8 T cells with PPP\u003csup\u003ehigh\u003c/sup\u003e/GLN\u003csup\u003ehigh\u003c/sup\u003e tumour cells was associated with a shorter RFS. Conversely, GLN\u003csup\u003ehigh\u003c/sup\u003e/GLYC\u003csup\u003ehigh\u003c/sup\u003e or GLN\u003csup\u003ehigh\u003c/sup\u003e/GLYC\u003csup\u003ehigh\u003c/sup\u003e/PPP\u003csup\u003ehigh\u003c/sup\u003e populations in proximity to CD8 T cells and fibroblasts was found to be favourable. Past research has shown, in mouse models, that tumour and immune cells compete for glutamine, and that selectively blocking glutamine entry to the TCA cycle in tumour cells increases T-cell activation\u003csup\u003e\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e\u003c/sup\u003e. Our data also shows that in cells enriched in oxidative, glycolytic, PPP, and lipid oxidising pathways, there are distinct ASCT2\u003csup\u003e+\u003c/sup\u003e and ASCT2- subpopulations. While it is difficult to infer the metabolite availability in these environments, we can speculate that ASCT2 expression in these populations could be due, at least partly, to their proximity to other cell types which may be competing for extracellular glutamine in the TME\u003csup\u003e\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e\u003cp\u003eIn NSCLC, macrophages make up the majority of infiltrating immune cells, where they can exhibit both pro-tumour and anti-tumour effects\u003csup\u003e\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e\u003c/sup\u003e. Both proteomic and transcriptomic analysis identified macrophage-related features associated with tumour recurrence. In the transcriptomic analysis, these included \u003cem\u003eMSTR1\u003c/em\u003e\u003csup\u003e\u003cem\u003e29\u003c/em\u003e\u003c/sup\u003e within the tumour and \u003cem\u003eTNFAIP6\u003c/em\u003e in the stroma, where higher expression was linked to shorter RFS. \u003cem\u003eTNFAIP6\u003c/em\u003e is known to be involved with M2 polarisation, extracellular matrix modelling, and cell migration, which could contribute to disease aggression\u003csup\u003e\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e,\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e\u003c/sup\u003e. Mechanistic studies of PD-L1\u003csup\u003e+\u003c/sup\u003e macrophages typically have found roles in suppression of anti-tumour immune cells, but some results show immunostimulatory and beneficial effects in a range of cancers, including NSCLC\u003csup\u003e\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e,\u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e\u003c/sup\u003e. Our multivariable modelling showed that PD-L1\u003csup\u003e+\u003c/sup\u003e and M2 macrophages play a significant role, with the separation of PD-L1\u003csup\u003e+\u003c/sup\u003e macrophages from PD-L1\u003csup\u003e−\u003c/sup\u003e macrophages being highly favourable and the separation of PD-L1\u003csup\u003e+\u003c/sup\u003e M2 macrophages from stem-like antigen expressing tumour cells also contributing to longer RFS. The interplay between tumour-associated macrophages (TAMs) and tumour metabolism has been shown to be an important aspect in TME remodelling\u003csup\u003e\u003cspan additionalcitationids=\"CR35\" citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e–\u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e36\u003c/span\u003e\u003c/sup\u003e, and is also dependent on the macrophage metabolism – M1 macrophages are typically associated with glycolytic metabolism, whereas the anti-inflammatory M2 macrophages typically rely on oxidative metabolism\u003csup\u003e\u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e37\u003c/span\u003e\u003c/sup\u003e. TMEs associated with high lactate content and extended lactic acidosis have enhanced monocyte to macrophage differentiation with pro-tumour and inflammatory characteristics\u003csup\u003e\u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e38\u003c/span\u003e\u003c/sup\u003e. Of particular interest, results from our univariate analysis identified that the co-localisation of GLN\u003csup\u003ehigh\u003c/sup\u003e macrophages with PPP\u003csup\u003ehigh\u003c/sup\u003e/GLN\u003csup\u003ehigh\u003c/sup\u003e CD4 T cells was unfavourable and that an increased proportion of GLN\u003csup\u003ehigh\u003c/sup\u003e macrophages and GLN\u003csup\u003ehigh\u003c/sup\u003e CD4 T cells contributed to a longer RFS. The multivariate analysis also revealed that increased proximity of LPDOX\u003csup\u003ehigh\u003c/sup\u003e macrophages with GLYC\u003csup\u003ehigh\u003c/sup\u003e/PPP\u003csup\u003ehigh\u003c/sup\u003e/GLN\u003csup\u003ehigh\u003c/sup\u003e CD4 T cells was favourable, potentially indicating a role for a metabolically atypical M1 TAM state associated with CD4 T cells in tumour non-recurrence.\u003c/p\u003e\u003cp\u003eUsing geometric deep learning for the stratification of tissue core regions into protein niches differentiable by tumour recurrence status, we found a diverse set of functional and metabolic states, highlighting the effects of several distinct immunological interactions. Major histocompatibility (MHC) proteins are essential for an adaptive immune system but have differing reported effects in the literature. Increased expression of MHC class I protein has shown to be associated with increased survival in NSCLC\u003csup\u003e\u003cspan citationid=\"CR39\" class=\"CitationRef\"\u003e39\u003c/span\u003e\u003c/sup\u003e; particularly, increased MHC class II protein expression has been shown to improve OS in resection surgery-treated LUAD NSCLC patients\u003csup\u003e\u003cspan citationid=\"CR40\" class=\"CitationRef\"\u003e40\u003c/span\u003e\u003c/sup\u003e. However, our analyses results suggests that this MHC I associated effect may be dependent on the composition of the tumour cells when expressing CD44, PD-L1, Ki67, or HLA-A. The model predictions also point to distinct, well established immunological interactions, such as the presence of PD-L1 expression in the tumour in conjunction with proliferating or stem-like gene expression – all in the presence of cytotoxic CD8 T cells\u003csup\u003e\u003cspan citationid=\"CR41\" class=\"CitationRef\"\u003e41\u003c/span\u003e\u003c/sup\u003e. Clinically unfavourable associations were found for proliferating tumour cells both with or without antigen expression, the presence of B cells (in our study, this study could not differentiate between plasma, memory, or effector B cells), and MHC class II-expressing tumours without additional functional properties. This is validated in the transcriptomics data, where relapse-associated DEGs in both tumour and stroma regions include \u003cem\u003eCEACAM1\u003c/em\u003e, and \u003cem\u003eCD3D\u003c/em\u003e, which are indirectly associated with MHC processes via MHC-class-I inhibition and antigen detection\u003csup\u003e\u003cspan citationid=\"CR42\" class=\"CitationRef\"\u003e42\u003c/span\u003e,\u003cspan citationid=\"CR43\" class=\"CitationRef\"\u003e43\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e\u003cp\u003eOur study has built a comprehensive profile of the TME, revealing additional cellular interactions that profile the TME and contribute to tumour recurrence. This profile highlighted the complex metabolic and immunological characteristics of the TME across disease stage and across the clinical response. We further validated this profile using spatial transcriptomics in a standalone manner and in combination with a deep learning technique to detangle signatures of tumour recurrence in gene expression. In addition, we validated previously identified immunological associations in the NSCLC TME and the mechanistic and metabolic relationships in the TME contributing to tumour recurrence.\u003c/p\u003e\u003cdiv id=\"Sec12\" class=\"Section2\"\u003e\u003cdiv id=\"Sec13\" class=\"Section3\"\u003e\u003c/div\u003e\u003c/div\u003e\u003cdiv id=\"Sec14\" class=\"Section2\"\u003e\u003cdiv id=\"Sec15\" class=\"Section3\"\u003e\u003c/div\u003e\u003c/div\u003e\u003cdiv id=\"Sec17\" class=\"Section2\"\u003e\u003cdiv id=\"Sec18\" class=\"Section3\"\u003e\u003c/div\u003e\u003c/div\u003e\u003cdiv id=\"Sec22\" class=\"Section2\"\u003e\u003cdiv id=\"Sec23\" class=\"Section3\"\u003e\u003c/div\u003e\u003c/div\u003e\u003cdiv id=\"Sec24\" class=\"Section2\"\u003e\u003cdiv id=\"Sec25\" class=\"Section3\"\u003e\u003cdiv id=\"Sec26\" class=\"Section4\"\u003e\u003c/div\u003e\u003c/div\u003e\u003cdiv id=\"Sec27\" class=\"Section3\"\u003e\u003cdiv id=\"Sec28\" class=\"Section4\"\u003e\u003c/div\u003e\u003c/div\u003e\u003c/div\u003e\n\n\u003cdiv id=\"Sec32\" class=\"Section2\"\u003e\u003cdiv id=\"Sec33\" class=\"Section3\"\u003e\u003c/div\u003e\u003cdiv id=\"Sec34\" class=\"Section3\"\u003e\u003cp\u003e\u003cb\u003eFigure Legends\u003c/b\u003e\u003c/p\u003e\u003c/div\u003e\u003c/div\u003e"},{"header":"Methods","content":"\u003ch2\u003eStudy design\u003c/h2\u003e\u003cp\u003e This study has Metro South Human Research Ethics approval (LNR/2019/QMS/51117) and ratification by the Queensland University of Technology. This study has University of Queensland Human Research Ethics Approval (2021/HE001936). Samples for this retrospective study were collected by medical oncologists at the Princess Alexandra Hospital (PAH). Using the lung cancer database, patients were filtered with the intent to collate tumour-microarrays from resected lung cancer tissues. Of these, 2,678 patients with NSCLC were identified to have received platinum-based chemotherapy. The cohort was established based on curative treatment intent (fully resected, early-stage disease, who received adjuvant platinum-based chemotherapy). The inclusion criteria were patients over 18 years of age, NSCLC cases treated only with adjuvant cisplatin/carboplatin in combination with vinorelbine, and tumour specimens of adequate tissue availability for IHC staining. Whole tissue blocks were sourced, pathology reviewed, and dual cores punched from representative areas at least ~ 0.5cm apart. A normal non-malignant lung sample was included on each TMA to act as a control and assist with orientation. The samples were formalin-fixed paraffin embedded, and cores were of a 2mm diameter. The final TMAs were comprised of 176 samples from 88 patients, where all samples were treatment naïve and from fully resected specimens.\u003c/p\u003e\u003ch2\u003eSpatial single-cell proteomic profiling\u003c/h2\u003e\u003ch2\u003ePatient cohort\u003c/h2\u003e\u003cp\u003e84 patients contributed 168 cores to cyclic-IF analysis. Quality control was performed to assess the staining and tissue quality across the two TMAs, with cores excluded if they were deemed to be majority of poor quality, or a subsection of the core was used if part of the core was disturbed by tissue lifting or staining artifacts such as staining edge effect. After quality control, 61 patients were included in analysis (48 paired samples and 13 unpaired samples).\u003c/p\u003e\u003ch2\u003eMultiplex IF tissue profiling\u003c/h2\u003e\u003cp\u003eTissue processing was performed by the AKOYA step team. Tumour microarrays were subject to the standard PhenoCycler Fusion imaging protocol\u003csup\u003e\u003cspan citationid=\"CR44\" class=\"CitationRef\"\u003e44\u003c/span\u003e\u003c/sup\u003e. The cyclic imaging protocol requires up to three fluorophores to be imaged simultaneously, before being washed away and three new reporters applied to the tissue. The order of reporter application was designed so that antibodies known to co-localise were not put in the same cycle, and antibodies known to have high expression were placed at the end of protocol. Once all images were captured, post processing combined the layers into one qptiff file, using a blank cycle to adjust for background staining.\u003c/p\u003e\u003ch2\u003eProteomic bioinformatic analysis\u003c/h2\u003e\u003ch2\u003eCell segmentation and data QC\u003c/h2\u003e\u003cp\u003eQptiff files were imported into \u003cem\u003eQupath\u003c/em\u003e for initial image processing, including marker and tissue quality control and cell segmentation. Segmentation was performed using the \u003cem\u003eCellpose\u003c/em\u003e (v2.0) plugin in \u003cem\u003eQupath\u003c/em\u003e using the nucleus model with 5µm nuclear expansion, constrained at 1.8 nuclear size. The median fluorescent intensities of each marker and cellular morphology features for each cell were exported into a .csv file containing a unique cellular ID, appropriate core ID, and x-y co-ordinates. Based on edge effects of imaging on the x-axis, boundaries were introduced to remove cells and cores that were affected by edge effects. For some cores, this resulted in the whole core being excluded from analysis while some cores were split roughly in half.\u003c/p\u003e\u003ch2\u003eCell typing\u003c/h2\u003e\u003cp\u003eCell typing was performed based on the median protein expression of each cell. Cells were first clustered into tumour and non-tumour phenotypes, using a subset of markers and morphology features; CD4, E-cadherin, CD20, CD68, CD8, CD14, Vimentin, CD45, CD34, CD3e, CD31, SMA, Ki67, Pan-Cytokeratin, nuclear size, nuclear solidity, and nuclear diameter. Cellular expression was normalised based on an \u003cem\u003earcsinh\u003c/em\u003e transform with a co-factor of 150 and a percentile transform of 97 along axis 0. A Z-score transform was then applied to the normalised expression. Batch correction was applied by PCA transform and integration using harmony integrate (rapids single cell package) between images\u003csup\u003e\u003cspan citationid=\"CR45\" class=\"CitationRef\"\u003e45\u003c/span\u003e\u003c/sup\u003e. \u003cem\u003ePhenograph\u003c/em\u003e unsupervised clustering was performed based on neighbour analysis (n = 20), a resolution of 0.4 and minimum size of 10. 21 clusters were established and then annotated based on protein expression on a heatmap and after mapping clusters manually onto the images \u003csup\u003e\u003cspan citationid=\"CR46\" class=\"CitationRef\"\u003e46\u003c/span\u003e\u003c/sup\u003e. Using the raw data, the data was subset into non-tumour cells for further subtyping. The markers used for broad annotations were CD4, CD20, CD68, CD8, CD14, Vimentin, CD45, CD34, CD3e, CD31, SMA, nuclear solidity, and nuclear diameter. The same normalisation and integration method as earlier was applied to the non-tumour cells. \u003cem\u003ePhenograph\u003c/em\u003e clusters were calculated over a range of resolutions and k-nearest neighbours. The heatmaps were evaluated using ARI, NMI, and AMI metrics to select the most reliable cluster parameters. The top 6 silhouette scores were visually inspected to identify the most reasonable cluster parameters. Based on heatmap expression, the parameters K = 40 and resolution = 0.2 were used for broad non-tumour phenotypes, these clusters were then annotated based on mean expression of markers.\u003c/p\u003e\u003cp\u003eThese non-tumour annotations were then merged into the original raw data file. For CD4, CD8, macrophage, and tumour populations, these were then further annotated, individually, by sub-setting the raw data into only the CD4, CD8, macrophage, or tumour populations and the same normalisation and integration methods, as used earlier, were applied to each subset. Following this, a binary gaussian mixture model (GMM) was applied to relevant functional markers for each cell type, then based on the combinations of positive markers the cells were functionally annotated. For the CD4 population, PD-1 and ICOS were assessed, and two populations were established: CD4 cells (~ 106 k) and CD4 T Follicular Helper (TFH) cells (ICOS\u003csup\u003e+\u003c/sup\u003e, PD-1\u003csup\u003e+\u003c/sup\u003e~ 6k cells). For CD8 T cells, Granzyme-B and PD-1 were assessed to establish CD8 populations of Cytotoxic CD8 and PD-1\u003csup\u003e+\u003c/sup\u003e CD8 T cells. GMM positivity for CD163 and PD-L1 were assessed for the macrophage population, creating four subtypes: macrophages, M2 macrophages, PD-L1\u003csup\u003e+\u003c/sup\u003e macrophages, and PD-L1\u003csup\u003e+\u003c/sup\u003e M2 macrophages. For the tumour phenotyping, positive staining of CD44, HLA-DR, HLA-A, KI67, and PD-L1 were assessed to assign phenotypes of all combinations of CD44\u003csup\u003e+\u003c/sup\u003e, MHC class I or class II positive, proliferating, or PD-L1\u003csup\u003e+\u003c/sup\u003e types.\u003c/p\u003e\u003cp\u003eSimilarly, using the metabolic markers in the panel (Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e) based on the combination of positive or negative markers, metabolic pathways were annotated for all cells. These definitions were created based on the metabolic pathways present in our panel through consultation with an expert in cancer metabolomics (Fig.\u0026nbsp;3a). Due to the number of cell types this produced, two cell phenotypes were created: 1. Broad cell types (tumour and non-tumour) \u003csup\u003e+\u003c/sup\u003e metabolic pathways and 2. Broad cell phenotypes of non-tumour cells \u003csup\u003e+\u003c/sup\u003e tumour cells annotated with metabolic pathways.\u003c/p\u003e\u003ch2\u003eNeighbourhoods\u003c/h2\u003e\u003cp\u003eCellular neighbourhoods (CNs) were defined using a k-means clustering of KNN neighbourhood counts of base tumour/non-tumour types \u003csup\u003e\u003cspan citationid=\"CR47\" class=\"CitationRef\"\u003e47\u003c/span\u003e\u003c/sup\u003e. CNs were classified into “Tumour” or “Stroma” (k = 2) on the tumour/non-tumour cell proportions, and k = 3 CNs were labelled as “Tumour,” “Tumour/Stroma Interface,” and “Stroma” regions using tumour/non-tumour cell proportions.\u003c/p\u003e\u003cp\u003eMetabolic neighbourhoods were generated based on the protein expression of G6PD, pNRF2, ATPA5, GLUT1, Hexokinase1, ASCT2, Citrate synthase, SDHA, CPT1A, PanCK and Ki67 using the \u003cem\u003eMESA\u003c/em\u003e package\u003csup\u003e5\u003c/sup\u003e[ks = 20, K-means method, kd tree]. We identified 8 distinct metabolic neighbourhoods using this method, including a ‘low’ neighbourhood that had little expression of metabolic markers in our panel. Most immune cells were captured in ‘low’ neighbourhoods. Mostly, this method captured metabolic diversity within the tumour cells, with only a small proportion of tumour cells captured in the ‘low’ metabolic neighbourhood.\u003c/p\u003e\u003ch2\u003eEntropy analysis\u003c/h2\u003e\u003cp\u003eWe employed spatial diversity algorithms in \u003cem\u003eMESA\u003c/em\u003e to calculate the local and global diversity of the samples at several scales \u003csup\u003e\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e\u003c/sup\u003e. Cells were annotated by their metabolic neighbourhood for the tumour cells and the functional annotation of non-tumour cells. This approach was used to capture the metabolic diversity of tumours at a niche level and identify differences in the immune cell populations in these neighbourhoods. MDI (multiscale diversity index) score was calculated based on the slope of Shannon diversity scores calculated over a range of scales (2 (~ 1mm patches), 4 (~ 500 µm), 8 (~ 250 µm), 16 (~ 130 µm), 24 (~ 80 µm), 32 (~ 60 µm), 48 (~ 40 µm), 64 (~ 30 µm), 72 (~ 27 µm)). To establish scales, samples were dissected into grids, where the number of rows and columns was that of the selected scale – i.e. a scale of 72 would divide the samples by applying a grid of 72x72 onto the tissue. Based on MDI results, the optimal scale was 32, and this resolution was used in subsequent analyses. Shannon-diversity and Moran’s I was used as the entropy metric for sub-regions; these values were then compared between local (LDI) and global scales (GDI). Hot and cold spots were identified by identifying regions that had similar Moran I scores. Patches that have high entropy scores were labelled as hot spots and patches that have low entropy scores were labelled cold spots. Cell frequency and cellular co-occurrence was calculated across each tissue at a global scale and then within every hot and cold spot.\u003c/p\u003e\u003ch2\u003eProximity features\u003c/h2\u003e\u003cp\u003eProximity metrics were computed for the cohort to capture cell-cell and cell-functional interactions at a range of scales. Global scales were captured by various entropy metrics and cell proportions within each core and within each CN. Medium-scale interactions were captured by the JSD Score, G-Cross AUC metrics, and diversity metrics. JSD measures probability overlap within each CN or core of cell-type or cell-functional type pairs and is bounded by 0 (complete density overlap) and 1 (complete density separation) \u003csup\u003e\u003cspan citationid=\"CR48\" class=\"CitationRef\"\u003e48\u003c/span\u003e\u003c/sup\u003e. Cell-cell interactions were described quantitatively using the G-Gross function, which measures the cumulative distribution of the nearest-neighbour distance of each target cell type from each reference type. The edge correction proposed by Hanisch was used to account for potential biases induced by a cell’s proximity to the core edge, which restricts the calculation to reference cells which are closer to the nearest target cell than to the edge of the core\u003csup\u003e\u003cspan citationid=\"CR49\" class=\"CitationRef\"\u003e49\u003c/span\u003e\u003c/sup\u003e. As G-Cross is a function that incorporate information from the range of computed radii, the area under the curve (AUC) was computed for the G-Cross function between 0 and ~ 250 µm.\u003c/p\u003e\u003cp\u003eThe smallest-scale interactions were captured by the proximity density metric. Proximity density scores were calculated between tumour metabolic neighbourhoods, with immune cells and other tumour phenotypes to identify patterns of cell-pair co-localisation. Proximity density was based on KNN within a 50-pixel radius and calculated the ratio of the number of two specific cell types within that radius compared to the total number of those two cell types within the core\u003csup\u003e\u003cspan citationid=\"CR50\" class=\"CitationRef\"\u003e50\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e\u003ch2\u003eMultivariable Feature Selection\u003c/h2\u003e\u003cp\u003eFeature selection was performed using the \u003cem\u003eStabl\u003c/em\u003e package, a feature selection method designed to select model-informative features from high-dimensional data in a false discovery-minimal manner\u003csup\u003e\u003cspan citationid=\"CR51\" class=\"CitationRef\"\u003e51\u003c/span\u003e\u003c/sup\u003e. Metrics for each feature family (G-Cross AUC, JSD Scores, proportions, and entropy) were computed and were also combined into a single feature set by a quantile transform to enforce a similar cumulative distribution for all features. Sparse features were filtered by removal of features that had less than 60 unique values among the 106 training and validation cores, reducing this down to approximately 6k features, and these remaining features were standard scaled before feature selection was performed. Four complementary model types and feature selection methods were tested in parallel: adaptive Lasso fits to recurrence status for each feature family, adaptive lasso fits to the quantile transformed (QT) features, Cox Proportional Hazards model (CoxPH) fits to the feature family-selected features for RFS, and CoxPH fits to the QT transformed features for RFS. We found that feature-family fitting to model RFS produced marginally better time-dependent AUCs than the QT features for a \u003cem\u003eBoostedCoxSurvival\u003c/em\u003e model to RFS and that QT-feature selection worked marginally better for fitting recurrence status. \u003cem\u003eBoostedCoxSurvival\u003c/em\u003e models were fit with 100 estimators and were trained on the training set and tested on the validation set. SHAP values were estimated for the \u003cem\u003eBoostedCoxSurvival\u003c/em\u003e model fits to quantify the core-to-core variation in the overall estimate of the feature importance and to measure the contribution to each feature’s hazard ratio.\u003c/p\u003e\u003cp\u003eModels to quantify the feature selection in the context of patient survival were implemented in \u003cem\u003escikit-survival\u003c/em\u003e, with linear model fits, standard scaling transforms, and quantile transforms performed in \u003cem\u003escikit-learn\u003c/em\u003e. Feature importances were estimated by the \u003cem\u003eSHAP\u003c/em\u003e python library using the \u003cem\u003ekernel explainer\u003c/em\u003e function.\u003c/p\u003e\u003ch2\u003eDeep Learning Training\u003c/h2\u003e\u003cp\u003eGraphs were formed using cells as nodes, with edges formed using KNNs with radial connections from the cell centroid \u003cem\u003ex\u003c/em\u003e and \u003cem\u003ey\u003c/em\u003e positions, up to 80 pixels away. Node features were formed using one-hot encoded cell-types from the list of 14 cell types formed during the cell-typing process. Additional functional markers such as the MHC class, ICOS, etc., positivity scores were generated using a Gaussian Mixture Model (GMM) fit to the median pixel values within each cell mask and appended to the feature list, followed by the metabolic pathway positivity scores formed from the pathway assessment logic outlined in \u003cb\u003eFig.\u0026nbsp;3a\u003c/b\u003e. PD-1 and PD-L1 markers exhibited high background expression due to ineffective reporter removal during the imaging cycles, so rolling ball subtraction of these two channels was applied before expression measurements were taken. The distance between each cell pair per 100 pixels was appended as an edge feature to the graphs.\u003c/p\u003e\u003cp\u003eGraphs were implemented in Deep Graph Library, with subgraphs sampled using k = 2 hops. The model was constructed with two Convolutional Transformer layers using the \u003cem\u003etorch-geometric\u003c/em\u003e library, with the layers containing 32 embedding size, single headed attention, mean pooling before output into a single dimension embedding, and final normalization using a sigmoid layer. The contrastive loss function was implemented from the \u003cem\u003ePyTorch Metrics Learning\u003c/em\u003e library with default parameters\u003csup\u003e\u003cspan citationid=\"CR52\" class=\"CitationRef\"\u003e52\u003c/span\u003e\u003c/sup\u003e. Training was performed on the training cores to contrast batched subgraphs by the patient recurrence status and prediction scores from the agglomerated cell predictions were tested on the validation cores. Graph augmentation was applied during training by random masking of feature values, random node and edge dropping, and gaussian smearing of the edge distance feature. The training was stopped after 12 epochs as preliminary tests showed that additional training merely pushed marginal regions in the “differentiable”- “non-differentiable” interface apart in the embedding space with little additional benefit to interpretability or prediction power. Histograms of the embedding scores stratified by the patient’s disease recurrence status showed that the predominant and undifferentiable regions of the cores were non-recurrence-like and that differentiable regions tended to be recurrence-associated.\u003c/p\u003e\u003ch2\u003eSpatial transcriptomic profiling (compartment-specific)\u003c/h2\u003e\u003ch2\u003eWhole transcriptome spatial profiling\u003c/h2\u003e\u003cp\u003eThe TMAs were subject to the GeoMx DSP Whole Transcriptome Atlas (WTA) assay. Tissue morphology markers included Pan-cytokeratin, CD45 and cyto-13 as markers for tumour, lymphocytes, and nuclei, respectively. 660µm round regions were drawn and masked by either PanCK\u003csup\u003e+\u003c/sup\u003e masks (tumour) or PanCK- masks (non-tumour) from a single sample per patient. Based on these selections, a laser directed UV-light onto the selected area, cleaving the oligonucleotide barcodes. These cleaved barcodes were removed from the slide surface and put into a well in a 96-well plate by the instrument. One well corresponds to one AOI. barcodes were prepared for sequencing according to manufacturer protocol and sequencing was performed by Australian Genome Research Facility (AGRF). Raw data was processed through the Nanostring GeoMx Next Generation Sequencing pipeline and digital count files reinstated on machine for mapping of counts. A QC level csv file was exported from the instrument for bioinformatic analysis.\u003c/p\u003e\u003ch2\u003eTranscriptomic bioinformatics analysis\u003c/h2\u003e\u003ch2\u003eQuality control and normalisation\u003c/h2\u003e\u003cp\u003eThe transcriptomics data from the GeoMx DSP comprising 120 regions of interest (ROIs)/AOIs (each representing either tumour or non-tumour segments) from 60 patients were processed using the \u003cem\u003estandR\u003c/em\u003e package workflow (Liu, et al., 2024). Firstly, only adenocarcinoma and squamous carcinoma samples were kept (110 ROIs, 55 patients). Secondly, 3 patients lacking clinical information were removed, leaving 104 ROIs and 52 patients. Sample filtering was performed by excluding ROIs with low detection counts (\u0026lt; 500,000) and low nuclei counts (\u0026lt; 200), resulting in the removal of 6 ROIs from the analysis. Gene filtering was conducted to remove low-expression genes (lowly expressed in \u0026gt; 90% of the ROIs), with no genes removed. The logCPM counts were evaluated using relative log expression (RLE) and principal components analysis (PCA) plots to identify confounding factors and to identify batch effects. Normalisation was then performed using the trimmed mean of M-values (TMM) method across all genes in the panel to adjust for library size variations and unwanted compositional biases.\u003c/p\u003e\u003ch2\u003eDifferential expression analysis\u003c/h2\u003e\u003cp\u003eDifferential expression (DE) analysis was carried out using the R packages \u003cem\u003eedgeR\u003c/em\u003e (v3.34.0) (Chen, et al., 2025) and \u003cem\u003elimma\u003c/em\u003e (3.48.0) (Ritchie, et al., 2015), using linear modelling incorporating experimental factors as predictors. The variation in gene expression was modelled by considering a common dispersion affecting all genes and a gene-specific dispersion. To estimate these dispersions, an empirical Bayes approach was employed, borrowing information from all genes and treating patient variation as a random effect using \u003cem\u003eduplicatecorrelation\u003c/em\u003e in \u003cem\u003eedgeR::voomLmFit\u003c/em\u003e function using sample weights. The linear model was then fitted to the experimental design containing the relevant factors, and DE analysis was conducted for specific contrasts of interest where in some cases, the T test relative to a threshold (TREAT) criterion was used\u003csup\u003e\u003cspan citationid=\"CR53\" class=\"CitationRef\"\u003e53\u003c/span\u003e\u003c/sup\u003e. Contrast includes comparing between histology (adenocarcinoma or squamous), recurrence status (Yes or No) or predicted recurrence status (based on modelling) within segments (i.e. Tumour or Stroma). The resulting statistic is an empirical Bayes moderated t-statistic with multiple testing adjustment performed using the Benjamini–Hochberg procedure where an adjusted p-value of \u0026lt; 0.05 was used to determine significantly DE genes. Predicted recurrence status for each sample were determined using a median cutoff of GNN-predicted recurrence score (from GNN-modelling) strategy, into either yes or no recurrence.\u003c/p\u003e\u003ch3\u003eGeneset enrichment analysis (GSEA) and visualisations\u003c/h3\u003e\u003cp\u003eGSEA was performed using the statistics provided by the DE analysis. The \u003cem\u003elimma::fry\u003c/em\u003e function was utilised for testing gene sets from the Molecular Signatures Database’s (MsigDB) categories including \u003cem\u003eHallmarks\u003c/em\u003e, \u003cem\u003eC2\u003c/em\u003e (curated gene sets), \u003cem\u003eC5\u003c/em\u003e (Ontology gene sets) and \u003cem\u003eC7\u003c/em\u003e (immunological signature gene sets, IMMUNESIGDB). \u003cem\u003eLimma::kegga\u003c/em\u003e function was used to test terms from the Kyoto Encyclopedia of Genes and Genomes (KEGG) database while \u003cem\u003elimma::goana\u003c/em\u003e function was used for testing gene ontology (GO) enrichment. The respective contrasts in the DE analyses were interrogated to identify functional enrichment of specific gene sets in the tumour and stromal segments associated with different survival/recurrence groups. The results from GSEA were inputted into the \u003cem\u003evissE\u003c/em\u003e (1.8.0) (Bhuva, et al., 2024) R package where clusters of enriched gene sets with common biological themes were identified and visualized. This visualization assisted in identification of biologically relevant pathways of interest in the study.\u003c/p\u003e\u003ch2\u003eFeature association with survival outcomes\u003c/h2\u003e\u003cp\u003eAnalysis of the feature association with survival or recurrence outcomes were implemented as per described previously\u003csup\u003e\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e\u003c/sup\u003e. Briefly, univariate cox proportional hazards (CoxPH) regression model and Kaplan Meier (KM) survival analyses were performed using the R packages \u003cem\u003esurvival\u003c/em\u003e and \u003cem\u003esurvminer\u003c/em\u003e. Features tested were collated from the significant DE genes from either 1) Histology in either segment or 2) GNN-predicted recurrence status in either segment. For the CoxPH modelling, RFS or OS durations were treated as the time variable with the event defined by occurrence of RFS events (DFS \u0026lt; OS) or deceased status respectively. Features having a Wald value of \u0026lt; 0.05 were followed with a KM survival analysis for further validation. In the KM analysis, the samples were grouped based on a median cutoff of the feature’s gene expression as either high or low. Features with a p-value \u0026lt; 0.05 in the KM were considered significant. Multivariate CoxPH and KM analysis were also conducted with either histology or clinical stage as a covariate.\u003c/p\u003e\u003ch2\u003eStatistics\u003c/h2\u003e\u003cp\u003eUnivariate CoxPH modelling, and Kaplan Meier (KM) survival analysis was performed on the data for OS and RFS using R packages \u003cem\u003esurvival\u003c/em\u003e and \u003cem\u003esurvminer\u003c/em\u003e. Median cell proportions with logit transformation were utilized for cohort stratification when assessing the association between cell proportion of ROIs and OS/recurrence status. Cox proportional hazards models were applied for non-discrete cell proportions.\u003c/p\u003e\u003cp\u003eFor the proteomics data, as there were instances of multiple samples per patient “training” and “validation” groups were created, with samples from the same patient distributed randomly between the two groups, and the remaining samples randomly distributed between the two groups. The cores from the cohort were grouped with 55 patient samples in the training group and 54 in the validation group. Unpaired samples were randomly split into two roughly even groups, with 7 samples assigned to the training and 6 to the validation cohort.\u003c/p\u003e\u003cp\u003eIn the univariate feature analysis, \u003cem\u003ep\u003c/em\u003e-values were initially adjusted for false discovery rate using the Benjamini-Hochberg correction, however this yielded no significant results due to the large number of features, so the raw \u003cem\u003ep\u003c/em\u003e-value was subsequently used. Features were filtered down to those with over 60 unique value to reduce feature sparsity.\u003c/p\u003e\u003cp\u003eFeatures were compared using the Mann-Whitney U-test for binary tests (Histological subtype, recurrence (yes/no), status (alive/dead), stage (low/high)) and Kaplan Meier fits were performed using feature median splits to differentiate high and low feature values in relation to the RFS. To avoid diluting the impact of intra-tumoral heterogeneity by averaging the features per patient, features were considered statistically robust if the feature was significant in both the training and validation sets.\u003c/p\u003e\u003ch2\u003eTranscriptomics\u003c/h2\u003e\u003cp\u003eSpatial RNA-seq data only contained one sample per patient and thus was not subject to training and validation splits. Statistical significance was calculated using binary comparisons between histological subtypes, recurrence, and survival status. For differential gene expression the resulting statistic is an empirical Bayes moderated t-statistic with multiple testing adjustment performed using the Benjamini–Hochberg procedure where an adjusted \u003cem\u003ep\u003c/em\u003e-value of \u0026lt; 0.05 was used to determine significantly DE genes.\u003c/p\u003e"},{"header":"Declarations","content":"\u003ch2\u003eData Availability\u003c/h2\u003e\n\u003cp\u003eData from this study, including the spatial proteomics data, spatial transcriptomics data and metadata are available at \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.48610/3244c1b\u003c/span\u003e\u003c/span\u003e\u003c/p\u003e\n\u003cp\u003eCode for this study is available at \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://github.com/clinicalomx/Adjuvant_Spatial_Multiomic_Profiling\u003c/span\u003e\u003c/span\u003e\u003c/p\u003e\u003cp\u003e\u003ch2\u003eConflicts of Interest\u003c/h2\u003e\u003cp\u003eJH and GM is an employee of Surge Care. AK is on the Scientific Advisory Board for Omapix Solutions, European Spatial Biology Centre, Predxbio, Molecular Instruments, and Visiopharm.\u003c/p\u003e\u003c/p\u003e\u003ch2\u003eAuthor Contributions\u003c/h2\u003e\u003cp\u003eConcept: CO, KOB, AK, MA\u003c/p\u003e\u003ch2\u003eAcknowledgements\u003c/h2\u003e\u003cp\u003eThe authors would like to thank the Translational Research Institute Histology Facilities for assisting with building the tissue microarray. The authors acknowledge the contribution of Akila Wijeranthna Yapa in the manuscript writing and review. The study investigators AK, CWT, CB, NL, KOB, MA are supported by the MRFF METASPATIAL Study (2031100). AK is supported by Cure Cancer and the Princess Alexandra Research Foundation. NB, MLD, JM, KE, AK, AK1, JF are supported by the Queensland Spatial Biology Centre (QSBC) at the Wesley Research Institute. AK1 is supported by the Harding Family Fellowship (Wesley Research Institute). AK\u0026thinsp;=\u0026thinsp;Arutha Kulasinghe, AK1\u0026thinsp;=\u0026thinsp;Aaron Kilgallon.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eGarg P et al (2024) Advances in Non-Small Cell Lung Cancer: Current Insights and Future Directions. J Clin Med 13. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.3390/jcm13144189\u003c/span\u003e\u003cspan address=\"10.3390/jcm13144189\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eRad HS et al (2021) The Pandora's box of novel technologies that may revolutionize lung cancer. Lung Cancer 159:34\u0026ndash;41. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.lungcan.2021.06.022\u003c/span\u003e\u003cspan address=\"10.1016/j.lungcan.2021.06.022\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eJhaveri N et al (2023) Mapping the Spatial Proteome of Head and Neck Tumors: Key Immune Mediators and Metabolic Determinants in the Tumor Microenvironment. GEN Biotechnol 2:418\u0026ndash;434. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1089/genbio.2023.0029\u003c/span\u003e\u003cspan address=\"10.1089/genbio.2023.0029\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eMonkman J et al (2025) Metabolic characterisation of tumor-immune interactions by multiplexed immunofluorescence reveals spatial mechanisms of immunotherapy response PREPRINT \u003cem\u003eResearch Square\u003c/em\u003e \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/\u003c/span\u003e\u003cspan address=\"https://doi.org/\" targettype=\"URL\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.21203/rs.3.rs-6695435/v1\u003c/span\u003e\u003cspan address=\"10.21203/rs.3.rs-6695435/v1\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eDing DY et al (2025) Quantitative characterization of tissue states using multiomics and ecological spatial analysis. Nat Genet 57:910\u0026ndash;921. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1038/s41588-025-02119-z\u003c/span\u003e\u003cspan address=\"10.1038/s41588-025-02119-z\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eTan CW et al (2025) The development of a high-plex spatial proteomic methodology for the characterisation of the head and neck tumour microenvironment. npj Precision Oncol 9:191. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1038/s41698-025-00963-0\u003c/span\u003e\u003cspan address=\"10.1038/s41698-025-00963-0\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLiu N et al (2024) standR: spatial transcriptomic analysis for GeoMx DSP data. Nucleic Acids Res 52:e2. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1093/nar/gkad1026\u003c/span\u003e\u003cspan address=\"10.1093/nar/gkad1026\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eZhang S et al (2015) High expression of MAGE-A9 in tumor and stromal cells of non-small cell lung cancer was correlated with patient poor survival. Int J Clin Exp Pathol 8:541\u0026ndash;550\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eHung CS et al (2018) Hypermethylation of CCND2 in Lung and Breast Cancer Is a Potential Biomarker and Drug Target. Int J Mol Sci 19. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.3390/ijms19103096\u003c/span\u003e\u003cspan address=\"10.3390/ijms19103096\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eCho NH et al (2004) MMP expression profiling in recurred stage IB lung cancer. Oncogene 23:845\u0026ndash;851. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1038/sj.onc.1207140\u003c/span\u003e\u003cspan address=\"10.1038/sj.onc.1207140\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eGill JH et al (2004) MMP-10 Is Overexpressed, Proteolytically Active, and a Potential Target for Therapeutic Intervention in Human Lung Carcinomas. Neoplasia 6:777\u0026ndash;785. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/https://doi.org/10.1593/neo.04283\u003c/span\u003e\u003cspan address=\"10.1593/neo.04283\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLesbon JCC et al (2022) A Screening of Epigenetic Therapeutic Targets for Non-Small Cell Lung Cancer Reveals PADI4 and KDM6B as Promising Candidates. Int J Mol Sci 23:11911\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLiu M et al (2019) PADI4\u0026ndash;mediated epithelial\u0026ndash;mesenchymal transition in lung cancer cells Retraction in /10.3892/mmr.2025.13610. Mol Med Rep 19:3087\u0026ndash;3094. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.3892/mmr.2019.9968\u003c/span\u003e\u003cspan address=\"10.3892/mmr.2019.9968\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eChen B et al (2021) Immune-related genes and gene sets for predicting the response to anti-programmed death 1 therapy in patients with primary or metastatic non-small cell lung cancer. Oncol Lett 22:540. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.3892/ol.2021.12801\u003c/span\u003e\u003cspan address=\"10.3892/ol.2021.12801\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eAhmetoglu D, Zheng H, Swart A, Zhu H, Li M (2025) Multifaceted Roles of Guanylate-Binding Proteins in Cancer. Int J Mol Sci 26:5477\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eHanahan D, Weinberg RA (2011) Hallmarks of Cancer: The Next Generation. Cell 144:646\u0026ndash;674. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.cell.2011.02.013\u003c/span\u003e\u003cspan address=\"10.1016/j.cell.2011.02.013\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLan T et al (2024) Glucose-6-phosphate dehydrogenase maintains redox homeostasis and biosynthesis in LKB1-deficient KRAS-driven lung cancer. Nat Commun 15:5857. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1038/s41467-024-50157-8\u003c/span\u003e\u003cspan address=\"10.1038/s41467-024-50157-8\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eEdwards DN et al (2021) Selective glutamine metabolism inhibition in tumor cells improves antitumor T lymphocyte activity in triple-negative breast cancer. J Clin Investig 131. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1172/JCI140100\u003c/span\u003e\u003cspan address=\"10.1172/JCI140100\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLiu Z et al (2023) CPT1A-mediated fatty acid oxidation confers cancer cell resistance to immune-mediated cytolytic killing. \u003cem\u003eProceedings of the National Academy of Sciences\u003c/em\u003e 120, e2302878120 \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/doi:10.1073/pnas.2302878120\u003c/span\u003e\u003cspan address=\"doi:10.1073/pnas.2302878120\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eTang M et al (2022) CPT1A-mediated fatty acid oxidation promotes cell proliferation via nucleoside metabolism in nasopharyngeal carcinoma. Cell Death Dis 13:331. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1038/s41419-022-04730-y\u003c/span\u003e\u003cspan address=\"10.1038/s41419-022-04730-y\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eWeber GF (2016) (2016) Time and Circumstances: Cancer Cell Metabolism at Various Stages of Disease Progression. \u003cem\u003eFrontiers in Oncology\u003c/em\u003e Volume 6\u0026ndash; \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.3389/fonc.2016.00257\u003c/span\u003e\u003cspan address=\"10.3389/fonc.2016.00257\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLiu Y et al (2020) Immune Cell PD-L1 Colocalizes with Macrophages and Is Associated with Outcome in PD-1 Pathway Blockade Therapy. Clin Cancer Res 26:970\u0026ndash;977. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1158/1078-0432.Ccr-19-1040\u003c/span\u003e\u003cspan address=\"10.1158/1078-0432.Ccr-19-1040\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eHong W et al (2018) Inhibition of Glucose-6-Phosphate Dehydrogenase Reverses Cisplatin Resistance in Lung Cancer Cells via the Redox System. Front Pharmacol 9\u0026ndash;2018. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.3389/fphar.2018.00043\u003c/span\u003e\u003cspan address=\"10.3389/fphar.2018.00043\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eNakamura M et al (2024) Blockade of glucose-6-phosphate dehydrogenase induces immunogenic cell death and accelerates immunotherapy. J Immunother Cancer 12:e008441. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1136/jitc-2023-008441\u003c/span\u003e\u003cspan address=\"10.1136/jitc-2023-008441\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eKoundouros N, Poulogiannis G (2020) Reprogramming of fatty acid metabolism in cancer. Br J Cancer 122:4\u0026ndash;22. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1038/s41416-019-0650-z\u003c/span\u003e\u003cspan address=\"10.1038/s41416-019-0650-z\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eKim SJ et al (2022) Dual inhibition of CPT1A and G6PD suppresses glioblastoma tumorspheres. J Neurooncol 160:677\u0026ndash;689. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1007/s11060-022-04189-z\u003c/span\u003e\u003cspan address=\"10.1007/s11060-022-04189-z\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eWang J et al (2024) ASCT2 Regulates Fatty Acid Metabolism to Trigger Glutamine Addiction in Basal-like Breast Cancer. Cancers (Basel) 16. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.3390/cancers16173028\u003c/span\u003e\u003cspan address=\"10.3390/cancers16173028\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eZhou Y et al (2024) The role of tumor-associated macrophages in lung cancer: From mechanism to small molecule therapy. Biomed Pharmacother 170:116014. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/https://doi.org/10.1016/j.biopha.2023.116014\u003c/span\u003e\u003cspan address=\"10.1016/j.biopha.2023.116014\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eCazes A, Childers BG, Esparza E, Lowy AM (2022) The MST1R/RON Tyrosine Kinase in Cancer: Oncogenic Functions and Therapeutic Strategies. Cancers (Basel) 14. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.3390/cancers14082037\u003c/span\u003e\u003cspan address=\"10.3390/cancers14082037\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLi L et al (2022) TNFAIP6 defines the MSC subpopulation with enhanced immune suppression activities. Stem Cell Res Ther 13:479. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1186/s13287-022-03176-5\u003c/span\u003e\u003cspan address=\"10.1186/s13287-022-03176-5\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLiu S, Chen H, Yang X, Wen Y, Chen L (2024) Identification and validation of up-regulated TNFAIP6 in osteoarthritis with type 2 diabetes mellitus. Sci Rep 14:31450. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1038/s41598-024-82985-5\u003c/span\u003e\u003cspan address=\"10.1038/s41598-024-82985-5\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003e(!!! INVALID CITATION !!! 32)\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eWang L et al (2024) PD-L1-expressing tumor-associated macrophages are immunostimulatory and associate with good clinical outcome in human breast cancer. Cell Rep Med 5. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.xcrm.2024.101420\u003c/span\u003e\u003cspan address=\"10.1016/j.xcrm.2024.101420\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eSchuurbiers OCJ et al (2014) Glucose Metabolism in NSCLC Is Histology-Specific and Diverges the Prognostic Potential of 18FDG-PET for Adenocarcinoma and Squamous Cell Carcinoma. J Thorac Oncol 9:1485\u0026ndash;1493. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/https://doi.org/10.1097/JTO.0000000000000286\u003c/span\u003e\u003cspan address=\"10.1097/JTO.0000000000000286\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eHe A et al (2025) Decoding the role of lipid metabolism in NSCLC: From macrophage subtype identification to prognostic model development. \u003cem\u003eThe FASEB Journal\u003c/em\u003e 39, e70467 \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/https://doi.org/10.1096/fj.202500124\u003c/span\u003e\u003cspan address=\"10.1096/fj.202500124\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eNetea-Maier RT, Smit JWA, Netea MG (2018) Metabolic changes in tumor cells and tumor-associated macrophages: A mutual relationship. Cancer Lett 413:102\u0026ndash;109. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/https://doi.org/10.1016/j.canlet.2017.10.037\u003c/span\u003e\u003cspan address=\"10.1016/j.canlet.2017.10.037\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLi M et al (2023) Metabolism, metabolites, and macrophages in cancer. J Hematol Oncol 16:80. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1186/s13045-023-01478-6\u003c/span\u003e\u003cspan address=\"10.1186/s13045-023-01478-6\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eColegio OR et al (2014) Functional polarization of tumour-associated macrophages by tumour-derived lactic acid. Nature 513:559\u0026ndash;563. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1038/nature13490\u003c/span\u003e\u003cspan address=\"10.1038/nature13490\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eDatar IJ et al (2021) Spatial Analysis and Clinical Significance of HLA Class-I and Class-II Subunit Expression in Non\u0026ndash;Small Cell Lung Cancer. Clin Cancer Res 27:2837\u0026ndash;2847. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1158/1078-0432.Ccr-20-3655\u003c/span\u003e\u003cspan address=\"10.1158/1078-0432.Ccr-20-3655\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eJohnson AM et al (2021) Cancer Cell-Specific Major Histocompatibility Complex II Expression as a Determinant of the Immune Infiltrate Organization and Function in the NSCLC Tumor Microenvironment. J Thorac Oncol 16:1694\u0026ndash;1704. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/https://doi.org/10.1016/j.jtho.2021.05.004\u003c/span\u003e\u003cspan address=\"10.1016/j.jtho.2021.05.004\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eWu S-P et al (2018) Stromal PD-L1\u0026ndash;Positive Regulatory T cells and PD-1\u0026ndash;Positive CD8-Positive T cells Define the Response of Different Subsets of Non\u0026ndash;Small Cell Lung Cancer to PD-1/PD-L1 Blockade Immunotherapy. J Thorac Oncol 13:521\u0026ndash;532. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/https://doi.org/10.1016/j.jtho.2017.11.132\u003c/span\u003e\u003cspan address=\"10.1016/j.jtho.2017.11.132\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eGray-Owen SD, Blumberg RS (2006) CEACAM1: contact-dependent control of immunity. Nat Rev Immunol 6:433\u0026ndash;446. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1038/nri1864\u003c/span\u003e\u003cspan address=\"10.1038/nri1864\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eWei Z et al (2022) CD3D: a prognostic biomarker associated with immune infiltration and immunotherapeutic response in head and neck squamous cell carcinoma. Bioengineered 13:13784\u0026ndash;13800. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1080/21655979.2022.2084254\u003c/span\u003e\u003cspan address=\"10.1080/21655979.2022.2084254\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eDonovan ML et al (2024) Protocol for high-plex, whole-slide imaging of human formalin-fixed paraffin-embedded tissue using PhenoCycler-Fusion. STAR Protoc 5:103226. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.xpro.2024.103226\u003c/span\u003e\u003cspan address=\"10.1016/j.xpro.2024.103226\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003escverse/rapids_singlecell: v0.10.6 v. 0.10.6 (2024)\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eLevine JH et al (2015) Data-Driven Phenotypic Dissection of AML Reveals Progenitor-like Cells that Correlate with Prognosis. Cell 162:184\u0026ndash;197. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.cell.2015.05.047\u003c/span\u003e\u003cspan address=\"10.1016/j.cell.2015.05.047\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eSch\u0026uuml;rch CM et al (2020) Coordinated Cellular Neighborhoods Orchestrate Antitumoral Immunity at the Colorectal Cancer Invasive Front. Cell 182:1341\u0026ndash;1359e1319. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1016/j.cell.2020.07.005\u003c/span\u003e\u003cspan address=\"10.1016/j.cell.2020.07.005\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eMasotti M, Osher N, Eliason J, Rao A, Baladandayuthapani VDIMPLE (2023) An R package to quantify, visualize, and model spatial cellular interactions from multiplex imaging with distance matrices. Patterns 4:100879. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/https://doi.org/10.1016/j.patter.2023.100879\u003c/span\u003e\u003cspan address=\"10.1016/j.patter.2023.100879\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eHanisch KH (1984) Some remarks on estimators of the distribution function of nearest neighbour distance in stationary spatial point processes. Ser Stat 15:409\u0026ndash;412. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1080/02331888408801788\u003c/span\u003e\u003cspan address=\"10.1080/02331888408801788\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eNirmal AJ, Sorger PK (2024) SCIMAP: A Python Toolkit for Integrated Spatial Analysis of Multiplexed Imaging Data. J Open Source Softw 9. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.21105/joss.06604\u003c/span\u003e\u003cspan address=\"10.21105/joss.06604\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eH\u0026eacute;dou J et al (2024) Discovery of sparse, reliable omic biomarkers with Stabl. Nat Biotechnol 42:1581\u0026ndash;1593. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1038/s41587-023-02033-x\u003c/span\u003e\u003cspan address=\"10.1038/s41587-023-02033-x\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eMusgrave K, Belongie S, Lim S-N Pytorch metric learning. \u003cem\u003earXiv preprint arXiv\u003c/em\u003e:(2008). \u003cem\u003e09164\u003c/em\u003e (2020)\u003c/span\u003e\u003c/li\u003e\u003cli\u003e\u003cspan\u003eMcCarthy DJ, Smyth GK (2009) Testing significance relative to a fold-change threshold is a TREAT. Bioinformatics 25:765\u0026ndash;771. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.1093/bioinformatics/btp053\u003c/span\u003e\u003cspan address=\"10.1093/bioinformatics/btp053\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"nature-portfolio","isNatureJournal":true,"hasQc":false,"allowDirectSubmit":false,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"","title":"Nature Portfolio","twitterHandle":"","acdcEnabled":false,"dfaEnabled":false,"editorialSystem":"ejp","reportingPortfolio":"","inReviewEnabled":true,"inReviewRevisionsEnabled":false},"keywords":"","lastPublishedDoi":"10.21203/rs.3.rs-7692105/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-7692105/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eWhile there have been therapeutic advancements in treatments for non-small cell lung cancer (NSCLC), we have a limited understanding of how the tumour composition and spatial heterogeneity influences clinical outcomes. Here, we performed a multiomic analysis of 61 NSCLC patients treated with adjuvant chemotherapy and curative resection, integrating spatial transcriptomics, spatial proteomics, and deep learning to profile the tumour microenvironment (TME). We identified spatial-metabolic alterations associated with disease recurrence in and between cells in the TME, involving glutamine import, glycolytic, and lipid oxidation pathways. Based on protein expression, we found spatially distinct regions linked to tumour recurrence, implicating specific roles for tumour associated macrophages, metabolic pathways, and major histocompatibility complex (MHC) proteins. Moreover, transcriptomic analysis revealed histology-specific gene expression changes and metabolic and inflammation pathway associations with tumour recurrence. This comprehensive multiomic analysis highlights the complex immunological and metabolic dynamics characteristic of disease recurrence in adjuvant chemotherapy treated NSCLC.\u003c/p\u003e","manuscriptTitle":"Spatial Multiomic Profiling Identifies Metabolic and Inflammatory Signatures Driving Recurrence in Adjuvant-Treated NSCLC","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-10-23 13:25:58","doi":"10.21203/rs.3.rs-7692105/v1","editorialEvents":[],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"nature-communications","isNatureJournal":true,"hasQc":false,"allowDirectSubmit":false,"externalIdentity":"NCOMMS","sideBox":"Learn more about [Nature Communications](http://www.nature.com/ncomms/)","snPcode":"","submissionUrl":"https://mts-ncomms.nature.com/","title":"Nature Communications","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"ejp","reportingPortfolio":"Nature Communications","inReviewEnabled":true,"inReviewRevisionsEnabled":false}}],"origin":"","ownerIdentity":"270fdb4d-c6b3-4b8a-b71d-5762e24f6372","owner":[],"postedDate":"October 23rd, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[{"id":56655010,"name":"Biological sciences/Cancer/Lung cancer/Non-small-cell lung cancer"},{"id":56655011,"name":"Health sciences/Oncology/Cancer/Cancer microenvironment"}],"tags":[],"updatedAt":"2026-02-06T10:01:46+00:00","versionOfRecord":[],"versionCreatedAt":"2025-10-23 13:25:58","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-7692105","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-7692105","identity":"rs-7692105","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}