Interpretable Deep Learning Reveals Biologically Relevant Spatial Gene Expression Patterns in Lung Tumors and their Microenvironment

Interpretable Deep Learning Reveals Biologically Relevant Spatial Gene Expression Patterns in Lung Tumors and their Microenvironment | bioRxiv /* */ /* */ <!-- <!-- /*! * yepnope1.5.4 * (c) WTFPL, GPLv2 */ (function(a,b,c){function d(a){return"[object Function]"==o.call(a)}function e(a){return"string"==typeof a}function f(){}function g(a){return!a||"loaded"==a||"complete"==a||"uninitialized"==a}function h(){var a=p.shift();q=1,a?a.t?m(function(){("c"==a.t?B.injectCss:B.injectJs)(a.s,0,a.a,a.x,a.e,1)},0):(a(),h()):q=0}function i(a,c,d,e,f,i,j){function k(b){if(!o&&g(l.readyState)&&(u.r=o=1,!q&&h(),l.onload=l.onreadystatechange=null,b)){"img"!=a&&m(function(){t.removeChild(l)},50);for(var d in y[c])y[c].hasOwnProperty(d)&&y[c][d].onload()}}var j=j||B.errorTimeout,l=b.createElement(a),o=0,r=0,u={t:d,s:c,e:f,a:i,x:j};1===y[c]&&(r=1,y[c]=[]),"object"==a?l.data=c:(l.src=c,l.type=a),l.width=l.height="0",l.onerror=l.onload=l.onreadystatechange=function(){k.call(this,r)},p.splice(e,0,u),"img"!=a&&(r||2===y[c]?(t.insertBefore(l,s?null:n),m(k,j)):y[c].push(l))}function j(a,b,c,d,f){return q=0,b=b||"j",e(a)?i("c"==b?v:u,a,b,this.i++,c,d,f):(p.splice(this.i++,0,a),1==p.length&&h()),this}function k(){var a=B;return a.loader={load:j,i:0},a}var l=b.documentElement,m=a.setTimeout,n=b.getElementsByTagName("script")[0],o={}.toString,p=[],q=0,r="MozAppearance"in l.style,s=r&&!!b.createRange().compareNode,t=s?l:n.parentNode,l=a.opera&&"[object Opera]"==o.call(a.opera),l=!!b.attachEvent&&!l,u=r?"object":l?"script":"img",v=l?"script":u,w=Array.isArray||function(a){return"[object Array]"==o.call(a)},x=[],y={},z={timeout:function(a,b){return b.length&&(a.timeout=b[0]),a}},A,B;B=function(a){function b(a){var a=a.split("!"),b=x.length,c=a.pop(),d=a.length,c={url:c,origUrl:c,prefixes:a},e,f,g;for(f=0;f<d;f++)g=a[f].split("="),(e=z[g.shift()])&&(c=e(c,g));for(f=0;f<b;f++)c=x[f](c);return c}function g(a,e,f,g,h){var i=b(a),j=i.autoCallback;i.url.split(".").pop().split("?").shift(),i.bypass||(e&&(e=d(e)?e:e[a]||e[g]||e[a.split("/").pop().split("?")[0]]),i.instead?i.instead(a,e,f,g,h):(y[i.url]?i.noexec=!0:y[i.url]=1,f.load(i.url,i.forceCSS||!i.forceJS&&"css"==i.url.split(".").pop().split("?").shift()?"c":c,i.noexec,i.attrs,i.timeout),(d(e)||d(j))&&f.load(function(){k(),e&&e(i.origUrl,h,g),j&&j(i.origUrl,h,g),y[i.url]=2})))}function h(a,b){function c(a,c){if(a){if(e(a))c||(j=function(){var a=[].slice.call(arguments);k.apply(this,a),l()}),g(a,j,b,0,h);else if(Object(a)===a)for(n in m=function(){var b=0,c;for(c in a)a.hasOwnProperty(c)&&b++;return b}(),a)a.hasOwnProperty(n)&&(!c&&!--m&&(d(j)?j=function(){var a=[].slice.call(arguments);k.apply(this,a),l()}:j[n]=function(a){return function(){var b=[].slice.call(arguments);a&&a.apply(this,b),l()}}(k[n])),g(a[n],j,b,n,h))}else!c&&l()}var h=!!a.test,i=a.load||a.both,j=a.callback||f,k=j,l=a.complete||f,m,n;c(h?a.yep:a.nope,!!i),i&&c(i)}var i,j,l=this.yepnope.loader;if(e(a))g(a,0,l,0);else if(w(a))for(i=0;i (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];var j=d.createElement(s);var dl=l!='dataLayer'?'&l='+l:'';j.src='//www.googletagmanager.com/gtm.js?id='+i+dl;j.type='text/javascript';j.async=true;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-M677548'); Skip to main content Home About Submit ALERTS / RSS Search for this keyword Advanced Search New Results Interpretable Deep Learning Reveals Biologically Relevant Spatial Gene Expression Patterns in Lung Tumors and their Microenvironment Vibha R. Rao , Adrienne A. Workman , Liang Lu , Xiaoying Liu , Shrey S. Sukhadia doi: https://doi.org/10.1101/2025.09.17.674367 Vibha R. Rao 1 Dartmouth Hitchcock Medical Center , Lebanon, NH Find this author on Google Scholar Find this author on PubMed Search for this author on this site Adrienne A. Workman 1 Dartmouth Hitchcock Medical Center , Lebanon, NH Find this author on Google Scholar Find this author on PubMed Search for this author on this site Liang Lu 1 Dartmouth Hitchcock Medical Center , Lebanon, NH Find this author on Google Scholar Find this author on PubMed Search for this author on this site Xiaoying Liu 1 Dartmouth Hitchcock Medical Center , Lebanon, NH Find this author on Google Scholar Find this author on PubMed Search for this author on this site For correspondence: Shrey.Sukhadia{at}dartmouth.edu Xiaoying.Liu{at}dartmouth.edu Shrey S. Sukhadia 1 Dartmouth Hitchcock Medical Center , Lebanon, NH Find this author on Google Scholar Find this author on PubMed Search for this author on this site For correspondence: Shrey.Sukhadia{at}dartmouth.edu Xiaoying.Liu{at}dartmouth.edu Abstract Full Text Info/History Metrics Supplementary material Preview PDF Abstract Lung adenocarcinoma (LUAD), the most common subtype of non–small cell lung cancer (NSCLC) exhibits profound histological and molecular heterogeneity, hindering accurate prognosis and effective treatment. Current approaches to assess this heterogeneity, such as histopathology, molecular profiling, and spatial transcriptomics are constrained by high costs, long turnaround times, and limited tissue availability, making them challenging for widespread prognostic use. To address this gap, we developed XpressO-Lung, an explanatory deep learning model that predicts gene expression heterogeneity, spatially, in tumor and its microenvironment, on hematoxylin and eosin (H&E) based diagnostic (Dx) whole-slide images (WSIs) by learning associations between tissue morphology and the corresponding bulk-transcriptomic data. Utilizing 200 LUAD cases from The Cancer Genome Atlas (TCGA), XpressO-Lung predicted spatial expression patterns of NAPSA, SLC47A1, TP53I3, KLRB1 , FAM189A1, TICAM1, CD8A, CXCL13, TTF, CDH3 , KRT7 and CDKN2A genes (biomarkers) on the respective Dx-WSIs with AUCs ranging up to 0.92. More importantly, the predicted spatial gene expression patterns aligned with the known morphologic interactions of the tumor and its microenvironment, capturing biological events directly on Dx-WSIs. By coupling predictive performance with spatial interpretability of gene expression on Dx-WSIs, XpressO-Lung bridges histopathology and bulk-transcriptomics, enabling explainable morpho-genomic analyses to advance biomarker discovery and offer prognostic insights to inform precision oncology in LUAD, especially in low-resource settings. Introduction Lung adenocarcinoma (LUAD) is the most prevalent histological subtype of non-small cell lung cancer (NSCLC), accounting for approximately 40% of all lung cancer diagnoses 1 , 2 . Originating from glandular epithelial cells in the distal airways, LUAD is characterized by considerable histological and molecular heterogeneity, often presenting with distinct architectural growth patterns such as lepidic, acinar, papillary, micropapillary, complex glandular, and solid subtypes. 3 . This heterogeneity extends beyond the malignant epithelial cells to include the surrounding tumor microenvironment (TME), which comprises stromal, immune, and vascular elements that significantly influence tumor progression, treatment response and prognosis 3 . Genomic diversity is also prominent, with frequent oncogenic driver mutations in genes such as EGFR, KRAS, ALK, STK11, and TP53 4 – 6 . Current guidelines, including the WHO and CAP lung cancer biomarker reporting protocols, recommend reflex testing for ALK and ROS1 rearrangements, while the NCCN and IASLC-CAP-AMP guidelines further advocate broad molecular profiling to include EGFR, KRAS, BRAF, MET, RET, NTRK, and PD-L1 as part of routine diagnostic workup 7 .These alterations inform therapeutic strategies, particularly the use of tyrosine kinase inhibitors (TKIs) and immune checkpoint inhibitors, both of which have substantially improved outcomes for subsets of LUAD patients 8 . Nevertheless, a substantial subset of LUAD cases show unpredictable responses to therapy and widely varying prognoses, underscoring the need for more comprehensive molecular characterization 9 . A major obstacle to improving prognostic accuracy in LUAD is the high degree of inter- and intra-tumoral heterogeneity, which includes variation in both cancer cell states and TME composition. Even among patients with identical histological subtypes or driver mutations, differences in gene expression programs can significantly influence tumor progression, immune interactions, and treatment response 10 . Addressing this complexity requires an expanded genetic landscape and scalable tools capable of capturing transcriptomic diversity across LUAD cohorts. Spatial transcriptomics technologies have emerged as powerful tools for mapping gene expression across tumor sections, enabling in situ profiling of both cancer and non-cancer compartments with high resolution 11 . However, their technical demands, high costs, and limited clinical availability pose significant barriers to widespread implementation. Moreover, these methods often require specialized protocols and non-standard tissue preservation techniques, limiting their applicability to large-scale or retrospective prognostic studies 12 . As a result, there is growing interest in computational approaches that can infer molecular features from more accessible sources, such as routine histopathology slides. Hematoxylin and eosin (H&E)-stained diagnostic (Dx) whole slide images (WSIs) offer a rich but underutilized source of phenotypic data. Recent developments in computational pathology have leveraged deep learning (DL) models to extract biologically relevant features from WSIs, demonstrating the ability to classify LUAD subtypes 13 , predict driver mutations 14 , infer tumor microenvironment composition 15 , and even approximate transcriptomic signatures 16 . These models, often based on convolutional neural networks (CNNs) or attention-based architectures, have shown strong performance across aforementioned predictive tasks. However, a persistent challenge in these approaches is their lack of interpretability 17 . Most DL models in computational pathology function as “black boxes,” providing slide-level predictions without clear explanations of how these decisions are made or which tissue regions contribute most to the outcome 18 , 19 . This lack of interpretability remains a major barrier to its clinical application, particularly for prognostic tasks, as it hinders trust in models’ predictions and complicates biological validation 18 . In the context of prediction of spatial gene expression that drives tumor progression and patient outcomes, the inability to interpret histomorphologic regions along with the predicted spatial gene expression in both tumor and TME, downgrades the utility of these models for prognostic assessment and clinical decision-making for cancer patients. The importance of interpretability in clinical applications has also been recognized at the policy level, with the White House’s National AI Action Plan designating interpretability as a national research priority and emphasizing its critical role in ensuring safe, trustworthy, and clinically meaningful AI deployment in healthcare 20 . To address these challenges, we introduce XpressO-Lung, an explanatory deep learning model that predicts spatial gene expression heterogeneity in tumors and TME from H&E-stained Dx WSIs by learning associations between tissue morphology and bulk transcriptomic profiles using our XpressO pipeline 21 . XpressO-Lung builds upon our prior work in breast 21 and melanoma 22 cancers, extending the same modular, transparent framework to LUAD. The XpressO pipeline employs a weakly supervised attention-based multiple instance learning (MIL) approach, generating gene-specific spatial heatmaps that highlight high- and low-attention regions of gene expression across both tumor and its microenvironment 21 . We then analyse these heatmaps by comparing them with the pathologists’ annotations and interpret specific tissue regions driving the prediction of each gene (biomarker), thereby uncovering biologically meaningful spatial biomarker expression patterns within the tumor and TME, that highlight prognostic differences in LUAD. Therefore, XpressO-Lung not only offers the predictive power of state-of-the-art DL models, but also defines a clear methodology for leveraging and interpreting such models to reveal spatial morpho-genomic relationships in cancer, enabling the level of interpretability necessary for trustworthy prognostic application in the clinic. Materials and Methods Data Collection and Preprocessing We used an explainable deep learning pipeline ‘XpressO’ 21 to build the XpressO-Lung model to predict tissue-based gene expression heterogeneity from the corresponding Dx-WSIs in a weakly supervised and interpretable manner. XpressO spatially connects tissue morphology with gene expression patterns, enabling us to discern morphologic patterns, such as co- or alternate expression of genes in both tumor and TME. It incorporates four core modules: Segmentation, Feature Extraction, Classification, and Heatmap that enable seamless deployment across diverse cancer types with minimal reconfiguration, while effectively capturing spatially informative features from both tumor and TME regions. Dx-WSIs of LUAD from 200 patients were obtained from The Cancer Genome Atlas (TCGA-LUAD) via the Genomic Data Commons (GDC) portal 23 . These WSIs were processed using the segmentation module in XpressO 21 which leverages the CLAM framework 24 to identify tumor-enriched regions by assigning attention scores to image patches and selecting the most informative regions for analysis. The WSIs were tiled into non-overlapping 256×256 pixel patches at 20× magnification 25 . The resulting patches were passed through the feature extraction module in XpressO 21 , which uses a pretrained Vision Transformer (ViT-L/16) 26 called Unified Network for Instance-level Representation Learning (UNI) 27 , trained with DINOv2 28 to extract high-dimensional feature embeddings optimized for histopathological image analysis. These features were then retained for downstream classification. Collection and Processing of RNA-seq Data Corresponding bulk RNA sequencing (RNA-seq) data for the same 200 LUAD patients were downloaded from the TCGA-LUAD portal in the form of fragments per kilobase of transcript per million mapped reads (FPKM). To prepare the dataset for gene expression classification, FPKM values were binarized into “high” and “low” categories using XpressO’s custom binarization script 21 . For each gene, the median expression value across all patients was used as the binarization threshold. Samples with expression values greater than or equal to the median were labelled as high-expression (“1”), while those below the median were labelled as low-expression (“0”). Following binarization, the dataset was split into training, validation, and testing subsets using a predefined XpressO script. A range of k values (k = 4 to 14) was used to create k-folds for each gene. The value of k varied by gene (e.g., k = 10 or 13), resulting in different training, validation, and testing ratios (e.g., 80:10:10 [training:validation:testing] for k = 10; 74:13:13 for k = 13). Binary expression labels were assigned at the slide level, allowing each WSI to be categorized as high (“1”) or low (“0”) expression for a given gene. These labelled subsets were then used to train and evaluate the deep learning model for gene expression prediction using WSIs Model Training Gene expression prediction was performed using the classification module in XpressO, which is based on the CLAM-SB architecture 24 . For each gene, a weakly supervised model was trained to classify binarized expression status using the top-ranked patch embeddings from each tissue-slide in the training set. Training was conducted using binary cross-entropy loss, with the Adam optimizer (learning rate = 2×10 −4 ) over 200 epochs. The model was trained to weigh instance-level features using attention-based pooling and output a slide-level prediction for each gene’s expression class. Model Evaluation and Metrics The trained model was evaluated on independent test sets for each gene of interest based on the data-split corresponding to the best-performing k-fold during training. For each gene, the model that achieved the best validation performance during cross-validation was selected, and its performance was then evaluated on the corresponding held-out (or independent) testing set (i.e. 10% of the total set, in case of k=10-as an example). Evaluation metrics included area under the receiver operating characteristic curve (AUC-ROC), accuracy, precision, recall and F1-score. These metrics were computed separately for each gene based on the performance of the model on the independent test set. Visualization of the Distribution of High and Low Gene Expression on WSIs Interpretation of spatial heatmaps of gene expression on WSIs were conducted using XpressO’s HeatMap module wherein the attention scores for each patch of the test WSI got mapped to their corresponding spatial coordinates. These heatmaps aided visualization of spatial patterns associated with high versus low expression and co-versus alternating expression of gene pairs on WSI, allowing for the identification of histologic features contributing to the patterns of expression across multiple genes using gene-specific spatial heatmaps from the same test WSI clubbed together for a paired interpretation. Predicted heatmaps were compared with pathologists’ annotations to evaluate alignment between model-derived predictions and established histomorphologic patterns. Results Genes of interest Analysis of bulk RNA-seq profiles from 200 LUAD tumors in the TCGA cohort revealed a broad range of gene expression patterns across patients. Twelve genes were selected based on prior literature implicating their varying expression in tumor and TME in LUAD: CD8A , CDH3 , CDKN2A , CXCL13 , FAM189A1 , KRT7 , NAPSA , SLC47A1 , TICAM1 , TP53I3 , TTF1 , and KLRB1 ( Table 1a ). Their fpkm values are depicted in Table 1b . These genes encompass diverse facets of LUAD pathogenesis, including tumor lineage specification, cell cycle regulation, immune infiltration, epithelial plasticity, and therapeutic response. TTF1 (NKX2-1) 29 – 32 , NAPSA (Napsin A) 33 – 37 , and KRT7 38 – 42 are canonical immunohistochemical (IHC) markers routinely used to distinguish LUAD from squamous and metastatic tumors. CD8A 43 – 48 and KLRB1 (CD161) 49 – 51 reflect cytotoxic lymphocyte infiltration and immune competence, while CXCL13 52 – 57 organizes tertiary lymphoid structures and predicts favorable response to immunotherapy. View this table: View inline View popup Table 1a: Genes found to be under- or over-expressed in tumor and/or TME regions in LUAD in the literature. View this table: View inline View popup Download powerpoint Table 1b: Genes found to be under under- or over-expressed in tumor and TME regions in the TCGA-LUAD cohort. CDKN2A 58 – 64 , frequently lost in LUAD, drives cell cycle deregulation and poor outcomes, whereas TP53I3 (PIG3) 65 – 69 , a p53-inducible DNA damage response gene, promotes proliferation, invasion, and therapy resistance. TICAM1 70 – 74 links Toll-like receptor signaling to type I interferon pathways and has emerged as a prognostic marker, while CDH3 (P-cadherin) 75 – 77 enhances epithelial–mesenchymal transition and metastatic potential. SLC47A1 (MATE1) 78 – 80 influences chemotherapeutic sensitivity as a solute transporter, and FAM189A1 , a multi-pass transmembrane protein, has been reported to be upregulated in LUAD and implicated in treatment adaptation 81 – 83 . Model Evaluation of Gene Expression The performance metrics for the predicted expression of 12 LUAD biomarkers from the testing set are shown in Table 2 . The model demonstrated strong predictive ability for NAPSA , TP53I3 , and SLC47A1 , that achieved AUC values of 0.92, 0.84, and 0.84, respectively. All three genes showed high accuracy (0.85), with TP53I3 and SLC47A1 achieving consistently strong precision and recall (0.85), indicating confident classification of expression status across samples. KLRB1 and FAM189A1 also performed well, with AUCs of 0.83 and 0.8, respectively; KLRB1 showed balanced precision and recall (0.75-0.76), while FAM189A1 achieved its highest performance in test accuracy (0.75). View this table: View inline View popup Table 2: Performance metrics for the predicted expression of twelve genes in the test set as a function of the best performing kth fold for each gene. Moderate predictive performance was observed for CD8A and TICAM1 , with AUCs of 0.77 and 0.78, respectively. Both genes maintained test accuracies above 0.70, and TICAM1 showed good precision (∼0.76). CXCL13 and TTF1 were predicted with slightly lower AUCs (0.76 and 0.72, respectively), but TTF1 retained high precision (0.84). Finally, KRT7 , CDH3 , and CDKN2A had lower AUC values, ranging from 0.64 to 0.66, though all three still achieved test accuracies of 0.65 and demonstrated moderate classification performance across other metrics. Visualization and Interpretation of Biomarker Expression To further investigate the model’s spatial attention and interpretability of ROIs, we examined the attention heatmaps produced by XpressO’s HeatMaps module for test WSIs across biomarkers. Depending on the model’s association of the biomarkers’ expression with the tissue-morphological patterns, a biomarker could be predicted as either high or low-expressed in several WSI-patches, where red-patches indicate regions that influence the model’s prediction the most, followed by yellow and green-patches that have moderate and low influence on that prediction ( Figures 1 - 5 and Supplementary Figures S1-S11). The p_0 and p_1 indicate probabilities for the predicted class (high or low-expression) using top-10 patches and non-predicted class (low or high expression) using bottom-10 patches ( Table 3 ). The heatmaps revealed spatial associations between predicted biomarker expression and histomorphologic features in tumor and TME regions ( Figures 1 - 5 and Supplementary Figures S1-S11). The following subsections illustrate representative examples for individual biomarkers and biomarker pairs, linking predicted expression with corresponding histomorphologic patterns. Download figure Open in new tab Figure 1. Attention heatmaps and corresponding H&E views for NAPSA and TTF1 expression predictions in LUAD whole-slide images. (a) Case (TCGA-55-A493) showing low NAPSA expression (a-i) and high TTF1 expression (a-ii). (b) Case (TCGA-75-7027 ) showing low expression for both NAPSA (b-i) and TTF1 (b-ii). High or low expression status is indicated by red arrows, with the predicted probability of high or low expression ( p_0 ) shown in boxes. Black circles mark regions of interest selected for zoom-in visualization. View this table: View inline View popup Table 3. Model-predicted biomarker expression status across TCGA lung cancer samples. For each TCGA case, the model predicts whether the biomarker is in high or low expression state; where pred_0 = Predicted expression class (high/low expression) using the top-10 attended patches; p_0 = Probability of the predicted class using top-10 attention; pred_1 = Predicted expression class (high/low) using the bottom-10 attended patches; p_1 = Probability of the predicted class using bottom-10 attention. Associated figure panels (e.g., 1-a-i, S3-b-ii) point to corresponding heatmaps and H&E visualization. a) NAPSA and TTF1 NAPSA (Napsin A) and TTF1 , two well-established markers of LUAD, were frequently predicted to have expression heterogeneity and exhibited high attention weights within similar tumor regions ( Table 3 , Fig 1-a ,b and Supplementary Fig S1). In WSIs TCGA-55-A493 and TCGA-55-1592, TTF1 was highly expressed while NAPSA showed low expression. In WSI TCGA-75-7027, both markers exhibited low expression. Heatmaps highlighted a focused spatial pattern for TTF1 , consistent with its stable nuclear localization in the bulk adenocarcinoma clone ( Fig 1-a,b , Supplementary Fig S1). In contrast, NAPSA ’s scattered expression appeared more diffuse within tumor regions, reflecting its cytoplasmic granular staining and possible dilution by adjacent normal pneumocytes ( Fig 1-a,b , Supplementary Fig S1) 84 , 85 . Morphologically, high- TTF1 regions showed cohesive gland-forming tumor nests with uniform nuclear features, while low- NAPSA areas often corresponded to poorly differentiated foci lacking clear cytoplasmic granularity ( Fig 1-a,b , Supplementary Fig S1). These observations are consistent with known histopathologic features and demonstrate the model’s capacity to distinguish subcellular expression patterns directly from WSIs 86 . While primarily used as diagnostic markers, their heterogeneous expression may also reflect underlying clonal diversity, which has been associated with variable clinical outcomes 87 . b) CD8A and KRT7 The predicted expression of CD8A was consistently high across the three WSIs evaluated (TCGA-44-5643, TCGA-49-AAR3, and TCGA-69-7973) in the test set ( Table 3 , Fig 2-a,b and Supplementary Fig S2). KRT7 expression was also predicted to be high in two of these slides (TCGA-44-5643 and TCGA-69-7973) ( Table 3 , Fig 2a and Supplementary Fig S2), with only one case (TCGA-49-AAR3) ( Table 3 and Fig 2b ) showing low predicted KRT7 . Notably, in TCGA-44-5643 and TCGA-69-7973, where both CD8A and KRT7 expression were high, the model-generated heatmaps showed spatial co-localization of CD8A -positive regions with KRT7 -positive epithelial nests. Morphologically, this translated to lymphocytic cuffs and infiltrates directly abutting or penetrating keratin-rich tumor islands, a configuration characteristic of an immune-inflamed TME. This observation is consistent with established histopathologic patterns in LUAD, where a CD8A rim penetrating KRT7 -positive nests is indicative of an immune-inflamed TME 88 . This immune-inflamed architecture is associated with improved response to immunotherapy and favorable prognosis in LUAD 89 . Download figure Open in new tab Figure 2. Attention heatmaps and corresponding H&E views for CD8A and KRT7 expression predictions in LUAD whole-slide images. (a) Case (TCGA-49-AAR3) showing high CD8A expression (a-i) and low KRT7 expression (a-ii). (b) Case (TCGA-69-7973) showing high expression for both CD8A (b-i) and KRT7 (b-ii). High or low expression status is indicated by red arrows, with the predicted probability of high or low expression ( p_0 ) shown in boxes. Black circles mark regions of interest selected for zoom-in visualization. Areas labeled “T” indicate tumor nests, while areas labeled “TME” denote the surrounding tumor microenvironment. c) KRT7 and CDKN2A In both test cases analyzed (TCGA-44-8119, TCGA-97-A4M5), KRT7 expression was predicted to be high, while CDKN2A expression was predicted to be low ( Table 3 and Figure 3-a,b ). This aligns with previously described histologic patterns in lung adenocarcinoma, where regions that are KRT7 -bright but CDKN2A -silent often reflect proliferative LUAD clones with 9p21 loss, a genotype associated with poorer prognosis and potential responsiveness to CDK4/6 or PRMT5-targeted therapies over PD-(L)1 monotherapy 90 , 91 . Morphologically, such regions displayed densely cellular epithelial nests with preserved keratin cytoskeleton but minimal intervening stroma, supporting a high proliferative index. Download figure Open in new tab Figure 3. Attention heatmaps and corresponding H&E views for KRT7 and CDKN2A expression predictions in LUAD whole-slide images. (a) Case (TCGA-44-8119) showing high KRT7 expression (a-i) and low CDKN2A expression (a-ii). (b) Case (TCGA-97-A4M5) showing high KRT7 expression (b-i) and low CDKN2A expression (b-ii). High or low expression status is indicated by red arrows, with the predicted probability of high or low expression ( p_0 ) shown in boxes. Black circles mark regions of interest selected for zoom-in visualization. d) NAPSA and CDKN2A Across four test tissues (TCGA-55-7907, TCGA-55-8506, TCGA-97-A4M1, TCGA-97-A4M5), NAPSA was predicted as high in three and low in one, while CDKN2A was consistently low in each tissue, resulting in the most probable pattern of high NAPSA coinciding with low CDKN2A expression ( Table 3 , Fig 4-a,b and Supplementary Fig S3-a,b). This pattern, NAPSA –bright yet CDKN2A –silent characterizes an alveolar-differentiated but cell-cycle-unleashed LUAD clone, frequently associated with 9p21/MTAP deletions. Morphologically, these regions had abundant, finely vacuolated cytoplasm and centrally placed nuclei, mimicking type II pneumocytes, but arranged in crowded acinar or papillary formations without significant maturation. Tumors with this genotype are associated with poorer prognosis and reduced benefit from PD-(L)1 monotherapy, yet they present actionable vulnerabilities: they are susceptible to CDK4/6 blockade and PRMT5 / MAT2A-directed therapies and are candidates for combination immunotherapy strategies 90 , 92 , 93 . Download figure Open in new tab Figure 4. Attention heatmaps and corresponding H&E views for NAPSA and CDKN2A expression predictions in LUAD whole-slide images. (a) Case (TCGA-97-A4M1) showing high NAPSA expression (a-i) and low CDKN2A expression (a-ii). (b) Case (TCGA-55-7907) showing high NAPSA expression (b-i) and low CDKN2A expression (b-ii). High or low expression status is indicated by red arrows, with the predicted probability of high or low expression ( p_0 ) shown in boxes. Black circles mark regions of interest selected for zoom-in visualization. Download figure Open in new tab Figure 5. Attention heatmaps and corresponding H&E views for TTF1 and KRT7 expression predictions in LUAD whole-slide images. (a) Case (TCGA-69-7973) showing low TTF1 expression (a-i) and high KRT7 expression (a-ii). (b) Case (TCGA-L9-A50W) showing low TTF1 expression (b-i) and high KRT7 expression (b-ii). High or low expression status is indicated by red arrows, with the predicted probability of high or low expression ( p_0 ) shown in boxes. Black circles mark regions of interest selected for zoom-in visualization. e) KRT7 and TTF1 In tissues TCGA-69-7973, TCGA-69-8254, and TCGA-L9-A50W, KRT7 expression was consistently high, while TTF1 (NKX2-1) showed variable expression, with low or scattered levels ( Table 3 , Fig 5-a,b , Supplementary Fig S4). This pattern of sharp KRT7 alongside reduced or dispersed TTF1 suggests a histologic landscape containing both normal type II pneumocytes scattered within the tumor and intra-tumoral clones that have selectively downregulated NKX2-1 while retaining keratin cytoskeletal integrity 94 . Morphologically, this was seen as irregular nests and cords of keratin-rich cells interspersed with alveolar spaces lined by residual TTF1 -positive pneumocytes. f) TP53I3 TP53I3 , a transcriptional target of the tumor suppressor TP53, plays a role in oxidative stress response and DNA damage-induced apoptosis 95 . Predicted expression of TP53I3 was spatially enriched in areas with architectural disruption or nuclear atypia ( Table 3 and Supplementary Fig S5), suggesting that the model may detect regions undergoing genotoxic stress or p53 pathway activation, both of which have implications for tumor aggressiveness and treatment response in LUAD 96 . Regions with high TP53I3 expression localized to tumor nests with marked nuclear pleomorphism and structural disruption, consistent with stress-induced p53 signalling (Supplementary Fig S5-a) 97 , while areas with low TP53I3 expression showed preserved epithelial organization and uniform nuclei, appearing morphologically uniform and possibly indicative of lower DNA damage response activity (Supplementary Fig S5-b) 98 . g) TICAM1 TICAM1 , a signalling adaptor in the TLR3 and cGAS–STING pathways, is increasingly recognized as a marker of innate immune activation in NSCLC 99 , 100 . In our analysis, TICAM1 expression localized to immune-infiltrated tumor borders and stromal pockets ( Table 3 and Supplementary Fig S6), consistent with its inclusion in cGAS–STING pathway-associated gene (CSPAG) signatures that stratify NSCLC patients into high- and low-risk groups 101 , 102 . Regions with high TICAM1 expression showed dense lymphocytic cuffs and immune infiltrates at tumor– stromal interfaces, reflecting an immune-“hot” phenotype associated with robust cGAS–STING signaling and favorable outcomes (low-risk group) (Supplementary Fig S6-a) 102 . In contrast, low TICAM1 expression was observed in morphologically cold tumor nests with limited stromal immune infiltration, consistent with diminished innate immune activation and poorer prognosis (high-risk group) (Supplementary Fig S6-b) 102 . These findings reinforce TICAM1 ’s role as a classifier of tumor immunogenicity and support its emerging relevance as both a prognostic marker and a therapeutic target in LUAD 70 . h) SLC47A1 SLC47A1 , a solute carrier transporter and actionable therapeutic target, has been shown to sensitize cancer cells to platinum–acridine agents, a novel class of chemotherapeutics with enhanced DNA-damaging potential 103 . In our model, predicted low expression of SLC47A1 was consistent across three LUAD test cases ( Table 3 and Supplementary Fig. S7). Attention heatmaps localized these predictions to well-differentiated glandular regions with preserved acinar structures and uniform nuclei, indicating that reduced SLC47A1 expression can occur within histologically stable epithelium. This highlights an additional layer of expression-level heterogeneity in LUAD that may be relevant for therapeutic response 104 . Low SLC47A1 expression has also been associated with resistance to conventional chemotherapy and poorer survival outcomes 105 . i) FAM189A1 FAM189A1 , a CD20-like multi-pass transmembrane protein, has been implicated in both lung and hematologic malignancies through mechanisms such as gene-disrupting translocations and promoter methylation, suggesting a role in cell signaling and tumor adaptation 82 , 83 . In LUAD, it has also been reported as upregulated in radioresistant NSCLC sublines 83 . In our analysis, regions with high predicted FAM189A1 expression localized to cohesive, gland-forming acinar nests with preserved epithelial organization, consistent with a role in maintaining signaling integrity and epithelial polarity 82 ( Table 3 and Supplementary Fig S8-a). By contrast, regions with low predicted expression were enriched in papillary structures characterized by elongated nuclei, nuclear crowding, and irregular alignment along fibrovascular cores, suggesting a shift toward less cohesive growth and altered epithelial architecture 82 ( Table 3 and Supplementary Fig S8-b). Together, these findings highlight FAM189A1 as a potentially important biomarker of tumor plasticity and therapeutic resistance in LUAD 83 . j) CXCL13 CXCL13 , a B cell–attracting chemokine, plays a critical role in tertiary lymphoid structure (TLS) formation and has been associated with improved immunotherapy response in NSCLC 52 . In our analysis, predicted expression of CXCL13 was found localized to stromal patches adjacent to tumor epithelial fronts, frequently overlapping with lymphoid aggregates or immune-infiltrated regions ( Table 3 and Supplementary Fig S9). Morphologically, high-expression regions showed densely clustered lymphoid aggregates within stroma, consistent with TLS-like organization (Supplementary Fig S9-b), whereas low-expression cases displayed more dispersed stromal lymphocytes without clear clustering (Supplementary Fig S9-a) 52 . These findings align with reports that CXCL13 -expressing T and B cells are enriched in TLS-positive tumors and highlight the model’s sensitivity to spatially organized immune niches 106 . Additionally, high CXCL13 expression and TLS presence have been associated with enhanced immunotherapy response and improved prognosis in LUAD 107 . k) KLRB1 KLRB1 , which encodes CD161, is a marker of natural killer (NK) cells and subsets of memory-like CD8 and CD4 T cells, particularly tissue-resident T cells implicated in antitumor immunity 108 . In LUAD, KLRB1 expression has been associated with favorable prognosis and enhanced immune infiltration. In our cohort, three test cases were predicted to have low KLRB1 expression, with attention maps predominantly highlighting immune-poor stroma and tumor epithelial regions lacking dense lymphocytic infiltrates ( Table 3 and Supplementary Fig S10-a). These areas were characterized by sparse or absent small, round lymphocyte nuclei within the stromal matrix. In contrast, the single high-expression case showed model attention concentrated in lymphocyte-rich stromal zones and peritumoral lymphoid aggregates, consistent with organized immune niches ( Table 3 and Supplementary Fig S10-a) . These morphologic features align with prior reports linking reduced KLRB1 activity to diminished immune cell presence, while higher expression in the literature is associated with enriched lymphocyte-rich microenvironments and improved response to checkpoint blockade in NSCLC 49 . l) CDH3 CDH3 (P-cadherin), a calcium-dependent cell–cell adhesion protein typically expressed in basal or progenitor epithelial cells, is frequently upregulated in lung adenocarcinoma (LUAD) and has been associated with increased tumor cell proliferation, migration, and poor prognosis 109 . Mechanistically, CDH3 overexpression can promote tumor aggressiveness by modulating adherens junction stability and activating downstream Wnt/β-catenin signaling pathways 110 . In our attention heatmap analysis, regions predicted to have high CDH3 expression corresponded to cohesive nests of malignant epithelial cells with preserved gland-forming architecture, a pattern consistent with its role in maintaining cell–cell adhesion while supporting proliferative tumor growth ( Table 3 and Supplementary Fig S11-a) 111 . Low CDH3 expression localized to disorganized tumor regions with looser epithelial cohesion, irregular cell borders, and nuclear variability, reflecting weakened junctional integrity ( Table 3 and Supplementary Fig S11-b). Together, these patterns align with CDH3 ’s known role in regulating epithelial adhesion and tumor morphology in LUAD 109 . The biomarker pairs identified through our model also align with disease associations and pathways highlighted in STRING (version 12.5) 112 , reinforcing their biological relevance. For instance, CD8A , CDKN2A , and KRT7 showed strong associations with non-small cell lung carcinoma, vulvar carcinoma, and melanoma, with a very low false discovery rate in the Disease-gene Associations (DISEASES) 113 (Supplementary Fig. S12a). In this context, the FDR represents the probability that a reported association is a false positive after adjusting for multiple hypothesis testing, and values near zero indicate highly confident enrichment. Literature linked to these markers underscores their translational importance (Supplementary Fig. S12b): case reports highlight how immune contexture and PD-L1–driven dynamics in lung cancer can shape outcomes, implicating CD8A , CDKN2A , and KRT7 in pathways that govern both therapeutic resistance and progression across subtypes 114 . Other studies emphasize their roles in intratumoral heterogeneity that drives drug resistance 115 and spatial heterogeneity influencing survival in lung adenocarcinoma patients 116 . Furthermore, guidelines for diagnostic immunohistochemistry in lung cancer recommend markers like KRT7 and CDKN2A as lineage classifiers, directly supporting their inclusion in predictive modelling 117 . Another STRING-derived network connecting CXCL13 , KLRB1 , and CD8A captures immune regulatory processes with direct relevance to the TME (Supplementary Fig. S13). These markers appear in prognostic gene signatures such as necroptosis-associated immune profiles in hepatocellular carcinoma 118 , and in studies distinguishing intratumoral vs. circulating lymphoid populations as predictive biomarkers in lung cancer patients undergoing immune checkpoint blockade 119 . Additional work in single-cell settings reveals their role in defining exhausted CD8+ T cells recognizing tumor neoantigens, consistent with their function as immune infiltration markers in NSCLC 120 . The inclusion of KLRB1 , a regulator of NK and T cell function, situates this network at the interface of adaptive and innate immunity, providing a mechanistic complement to our model’s predictions. Finally, STRING associations among TTF1 , KRT7 , CDKN2A , and NAPSA connect these canonical lung adenocarcinoma markers to both disease associations (NSCLC, intestinal cancer, vulvar carcinoma) and specific pathways such as surfactant metabolism, in which TTF1 and NAPSA play key regulatory roles (Supplementary Fig. S14a). Here again, FDR-adjusted enrichment values were very low, indicating robust statistical confidence. Literature linked to this cluster highlights their clinical importance in molecularly complex or therapy-resistant settings (Supplementary Fig. S14b): ALK-rearranged lung tumors lacking typical lineage markers 121 , ALK fusion–driven resistance to osimertinib 122 , and histomorphological transformation from NSCLC to small cell carcinoma after immunotherapy 123 all emphasize the diagnostic and prognostic value of these markers. Meta-analyses distinguishing adenocarcinoma from squamous carcinoma further reinforce the utility of KRT7 and TTF1 as diagnostic anchors 124 . Notably, STRING highlights a local cluster of NAPSA and KRT7 with signet-ring cell adenocarcinoma, a rarer subtype, while in our LUAD cohort, these markers instead aligned with classical adenocarcinoma morphologies, reinforcing their broader lineage-defining role. (Supplementary Fig. S14a). These STRING-derived connections illustrate how the biomarker pairs identified by our model intersect with established cancer pathways, disease associations, and prior studies. Discussion In this study, we present XpressO-Lung, an explainable deep learning model that predicts gene expression heterogeneity spatially in LUAD tumors and their microenvironment directly from H&E-stained Dx-WSIs using bulk-transcriptomics data. The model captured histomorphologic patterns predictive of both immune and epithelial gene expression signatures within LUAD tissues. Our results demonstrate that the model not only recapitulates known LUAD biomarkers, i.e., NAPSA , TTF1 , KRT7 , CDKN2A , CD8A , SLC47A1 , TP53I3 , KLRB1 , FAM189A1 , TICAM1 , CXCL13 , CDH3 but also reveals their biologically meaningful spatial expression patterns either in a paired or individual format, mirroring underlying tumor and TME interactions that can inform patient prognosis. Across these biomarkers, the model’s predictive performance ranged from AUC 0.64-0.92, providing a strong foundation for interpreting the spatial patterns it uncovered across distinct epithelial lineages, proliferative states, and immune–stromal niches in LUAD. XpressO-Lung uncovered diverse epithelial lineages and proliferative states within LUAD. NAPSA and TTF1 frequently co-localized in cohesive glandular nests, anchoring well-differentiated adenocarcinoma clones 84 – 86 , while KRT7 often appeared in sharp keratin-rich nests alongside scattered TTF1, suggesting lineage drift 93 . In contrast, KRT7-bright/CDKN2A-silent regions marked 9p21-loss proliferative clones 89 – 90 , and NAPSA-bright/CDKN2A-silent zones reflected alveolar-differentiated yet cell-cycle-unrestrained phenotypes, both linked to poor prognosis yet potentially targetable by CDK4/6 or PRMT5 inhibitors, highlighting their relevance as prognostic and therapeutic biomarkers 91 – 92 . CDH3 was enriched in cohesive nests with intact architecture, whereas its loss accompanied disorganized epithelial growth and has been linked to poorer clinical outcomes 108 . These spatial phenotypes align with STRING associations connecting NAPSA, TTF1, KRT7, CDKN2A, and CDH3 to surfactant metabolism, cell-cycle regulation, and NSCLC pathways 118 – 121 . Beyond epithelial compartments, the model also mapped immune and stromal microenvironments with clear spatial resolution, offering insights into tumor-immune interactions that modulate patient prognosis. CD8A frequently rimmed KRT7-positive tumor islands, forming lymphocytic cuffs characteristic of an immune-inflamed TME, often linked to better survival 88 . CXCL13 clustered at peritumoral lymphoid aggregates, consistent with TLS-like niches and improved prognosis 106 – 107 , while KLRB1 marked lymphocyte-rich stromal pockets and was absent from immune-poor zones 81 . TICAM1 concentrated at tumor-stromal borders with dense lymphocytic infiltrates, suggesting cGAS–STING activation, whereas its loss delineated immune-cold nests often associated with poor clinical outcomes 101 – 102 . These findings mirror STRING-derived networks linking CD8A, CXCL13, and KLRB1 to immune regulatory pathways and checkpoint blockade response, underscoring their value as spatial markers of immune activity 115 – 117 . Finally, XpressO-Lung identified stress-responsive and therapy-relevant niches within LUAD tissue. TP53I3 marked disordered regions with nuclear pleomorphism, hinting at p53 pathway activation and oxidative stress, while its low expression was observed in morphologically uniform zones 97 . SLC47A1 was uniformly low in well-differentiated acini, echoing its reported role in sensitizing tumors to platinum–acridine therapeutics 102 . FAM189A1 showed a striking spatial dichotomy, high in cohesive acinar nests and low in papillary structures with disordered polarity consistent with its reported links to therapy adaptation and tumor plasticity 53 . Together, these patterns illustrate how XpressO-Lung characterizes the spatial architecture of the LUAD tumor–TME interface, revealing how morphology encodes both lineage state and clinically-relevant phenotypes. While XpressO-Lung demonstrates strong predictive performance and spatial interpretability, certain limitations remain. Although the model is trained to associate global expression with local histologic features, validation at the spatial level remains a challenge. The TCGA-LUAD dataset does not include spatial transcriptomics or immunohistochemistry (IHC) data for most biomarkers analyzed, limiting our ability to directly validate patch-level expression predictions. In future work, we plan to close this gap by performing targeted spatial transcriptomics and multiplexed IHC staining on in-house LUAD tissue cohorts. This will enable finer-grained validation of XpressO-Lung’s attention maps and expression predictions, further strengthening its clinical relevance and increasing its utility for prognostic applications in translational settings. XpressO-Lung advances the field by offering an accessible, scalable, and interpretable tool for morpho-genomic profiling in LUAD. It bridges histopathology and bulk-transcriptomics using standard clinical slides, circumventing the need for expensive molecular assays, and enabling prognostic biomarker discovery, stratification of immune versus epithelial tumor states, and improved patient-specific therapy selection. In deep learning for computational pathology, interpreting model outputs is as critical as achieving high predictive performance, and XpressO-Lung delivers on both fronts. The model’s framework reliably infers spatial gene expression patterns from histologic images and highlights biologically meaningful tumor-TME interactions that are predictive of patient prognosis. The model was trained, validated, and tested using a carefully structured design that ensured sufficient degrees of freedom for each gene, and incorporated a custom script for gene expression analysis to systematically reveal morphology-expression relationships. By combining rigorous methodology with interpretable outputs, XpressO-Lung offers a powerful complement to molecular profiling, with the potential to support prognostic assessment, inform clinical decision-making, and advance precision medicine in LUAD, especially in low-resource settings. Data Availability Statement All data analyzed during this study are either included in this article or at the repository: https://github.com/skr1/XpressO . Conflict of Interest No conflict of interest is reported by any of the authors. Funding N/A SUPPLEMENTARY FIGURE LEGENDS Supplementary Figure S1. Attention heatmaps and corresponding H&E views for NAPSA and TTF1 expression predictions in LUAD whole-slide images. (a) Case (TCGA-55-1592) showing low NAPSA expression (a-i) and high TTF1 expression (a-ii). High or low expression status is indicated by red arrows, with the predicted probability of high or low expression ( p_0 ) shown in boxes. Black circles mark regions of interest selected for zoom-in visualization. Supplementary Figure S2. Attention heatmaps and corresponding H&E views for CD8A and KRT7 expression predictions in LUAD whole-slide images. (a) Case (TCGA-44-5643) showing high CD8A expression (a-i) and high KRT7 expression (a-ii). High or low expression status is indicated by red arrows, with the predicted probability of high or low expression ( p_0 ) shown in boxes. Black circles mark regions of interest selected for zoom-in visualization. Areas labeled “T” indicate tumor nests, while areas labeled “TME” denote the surrounding tumor microenvironment. Supplementary Figure S3. Attention heatmaps and corresponding H&E views for NAPSA and CDKN2A expression predictions in LUAD whole-slide images. (a) Case (TCGA-55-8506) showing low NAPSA expression (a-i) and low CDKN2A expression (a-ii). (b) Case (TCGA-97-A4M5) showing high NAPSA expression (b-i) and low CDKN2A expression (b-ii). High or low expression status is indicated by red arrows, with the predicted probability of high or low expression ( p_0 ) shown in boxes. Black circles mark regions of interest selected for zoom-in visualization. Supplementary Figure S4. Attention heatmaps and corresponding H&E views for TTF1 and KRT7 expression predictions in LUAD whole-slide images. (a) Case (TCGA-69-8254) showing high TTF1 expression (a-i) and high KRT7 expression (a-ii). High or low expression status is indicated by red arrows, with the predicted probability of high or low expression ( p_0 ) shown in boxes. Black circles mark regions of interest selected for zoom-in visualization. Supplementary Figure S5. Attention heatmaps and corresponding H&E views for TP53I3 expression predictions in LUAD whole-slide images. (a) Case (TCGA-05-4397) showing high TP53I3 expression. (b) Case (TCGA-78-7153) showing low TP53I3 expression High or low expression status is indicated by red arrows, with the predicted probability of high or low expression ( p_0 ) shown in boxes. Black circles mark regions of interest selected for zoom-in visualization. Supplementary Figure S6. Attention heatmaps and corresponding H&E views for TICAM1 expression predictions in LUAD whole-slide images. (a) Case (TCGA-97-A4M5) showing high TICAM1 expression. (b) Case (TCGA-69-8254) showing low TICAM1 expression. High or low expression status is indicated by red arrows, with the predicted probability of high or low expression ( p_0 ) shown in boxes. Black circles mark regions of interest selected for zoom-in visualization. Areas labeled “T” indicate tumor nests, while areas labeled “TME” denote the surrounding tumor microenvironment Supplementary Figure S7. Attention heatmaps and corresponding H&E views for SLC47A1 expression predictions in LUAD whole-slide images. Case (TCGA-97-8172) showing low SLC47A1 expression. High or low expression status is indicated by red arrows, with the predicted probability of high or low expression ( p_0 ) shown in boxes. Black circles mark regions of interest selected for zoom-in visualization. Supplementary Figure S8. Attention heatmaps and corresponding H&E views for FAM189A1 expression predictions in LUAD whole-slide images. (a) Case (TCGA-05-4397) showing high FAM189A1 expression. (b) Case (TCGA-97-8179) showing low FAM189A1 expression High or low expression status is indicated by red arrows, with the predicted probability of high or low expression ( p_0 ) shown in boxes. Black circles mark regions of interest selected for zoom-in visualization. Areas labeled “T” indicate tumor nests, while areas labeled “TME” denote the surrounding tumor microenvironment. Supplementary Figure S9. Attention heatmaps and corresponding H&E views for CXCL13 expression predictions in LUAD whole-slide images. (a) Case (TCGA-75-7027) showing low CXCL13 expression. (b) Case (TCGA-78-7153) showing high CXCL13 expression. High or low expression status is indicated by red arrows, with the predicted probability of high or low expression ( p_0 ) shown in boxes. Black circles mark regions of interest selected for zoom-in visualization. Areas labeled “T” indicate tumor nests, while areas labeled “TME” denote the surrounding tumor microenvironment. Supplementary Figure S10. Attention heatmaps and corresponding H&E views for KLRB1 expression predictions in LUAD whole-slide images. (a) Case (TCGA-97-A4M5) showing low KLRB1 expression. (b) Case (TCGA-L4-A4E6) showing high KLRB1 expression. High or low expression status is indicated by red arrows, with the predicted probability of high or low expression ( p_0 ) shown in boxes. Black circles mark regions of interest selected for zoom-in visualization. Areas labeled “T” indicate tumor nests, while areas labeled “TME” denote the surrounding tumor microenvironment. Supplementary Figure S11. Attention heatmaps and corresponding H&E views for CDH3 expression predictions in LUAD whole-slide images. (a) Case (TCGA-55-A493) showing high CDH3 expression. (b) Case (TCGA-L9-A443) showing low CDH3 expression. High or low expression status is indicated by red arrows, with the predicted probability of high or low expression ( p_0 ) shown in boxes. Black circles mark regions of interest selected for zoom-in visualization. Supplementary Figure S12. (a) STRING disease–gene associations for CD8A , CDKN2A , and KRT7 from DISEASES. (b) Reference publications linked to these markers. Node colors reflect supporting evidence categories; edges (green line) represent text-mining associations. Supplementary Figure S13. STRING network of CXCL13 , KLRB1 , and CD8A with reference publications. Node colors reflect supporting evidence categories; edges (green line) represent text-mining associations. Supplementary Figure S14. (a) STRING Reactome pathway, network cluster, and disease gene associations for TTF1 , KRT7 , CDKN2A , and NAPSA . (b) Reference publications linked to this cluster. Node colors reflect supporting evidence categories; edges (green line) represent text-mining associations. Acknowledgements Author Contributions Statement: Conceptualization SS, XL; Data curation VR, SS, AW, XL; Formal analysis VR, SS, AW, XL; Funding acquisition SS; Investigation SS, VR, XL, LL, AW; Methodology VR, SS, AW, XL; Project administration SS; Resources SS; Software SS, VR; Supervision SS, XL; Validation; Visualization SS, VR, XL, AW; Writing – original draft VR, SS; Writing – review & editing VR, SS, XL, LL, AW Footnotes Emails: vibha.r.rao{at}gmail.com , Adrienne.A.Workman{at}hitchcock.org , Liang.Lu{at}hitchcock.org , Xiaoying.Liu{at}dartmouth.edu , Shrey.Sukhadia{at}dartmouth.edu References 1. ↵ Nicholson AG , Tsao MS , Beasley MB , Borczuk AC , Brambilla E , Cooper WA , Dacic S , Jain D , Kerr KM , Lantuejoul S , Noguchi M , Papotti M , Rekhtman N , Scagliotti G , Schil P van, Sholl L , Yatabe Y , Yoshida A , Travis WD : The 2021 WHO Classification of Lung Tumors: Impact of Advances Since 2015 . [cited 2025 Aug 24], . Available from: https://www.jto.org/article/S1556-0864(21)03316-5/fulltext 2. ↵ Cancer statistics, 2025 - Siegel - 2025 - CA: A Cancer Journal for Clinicians - Wiley Online Library [Internet] . [cited 2025 Aug 24], . Available from: https://acsjournals.onlinelibrary.wiley.com/doi/10.3322/caac.21871 3. ↵ Solis LM , Behrens C , Raso MG , Lin HY , Kadara H , Yuan P , Galindo H , Tang X , Lee JJ , Kalhor N , Wistuba II , Moran CA: Histologic patterns and molecular characteristics of lung adenocarcinoma associated with clinical outcome . Cancer 2012 , 118 : 2889 – 2899 . OpenUrl CrossRef PubMed Web of Science 4. ↵ Skoulidis F , Heymach JV: Co-occurring genomic alterations in non-small cell lung cancer biology and therapy . Nat Rev Cancer 2019 , 19 : 495 – 509 . OpenUrl CrossRef PubMed 5. Chen J , Yang H , Teo ASM , Amer LB , Sherbaf FG , Tan CQ , Alvarez JJS , Lu B , Lim JQ , Takano A , Nahar R , Lee YY , Phua CZJ , Chua KP , Suteja L , Chen PJ , Chang MM , Koh TPT , Ong B-H , Anantham D , Hsu AAL , Gogna A , Too CW , Aung ZW , Lee YF , Wang L , Lim TKH , Wilm A , Choi PS , Ng PY , Toh CK , Lim W-T , Ma S , Lim B , Liu J , Tam WL , Skanderup AJ , Yeong JPS , Tan E-H , Creasy CL , Tan DSW , Hillmer AM , Zhai W : Genomic landscape of lung adenocarcinoma in East Asians . Nature Genetics , Nature Publishing Group , 2020 , 52 : 177 – 186 . OpenUrl CrossRef PubMed 6. ↵ Comprehensive molecular profiling of lung adenocarcinoma . Nature 2014 , 511 : 543 – 550 . OpenUrl CrossRef PubMed Web of Science 7. ↵ Lindeman NI , Cagle PT , Aisner DL , Arcila ME , Beasley MB , Bernicker EH , Colasacco C , Dacic S , Hirsch FR , Kerr K , Kwiatkowski DJ , Ladanyi M , Nowak JA , Sholl L , Temple-Smolkin R , Solomon B , Souter LH , Thunnissen E , Tsao MS , Ventura CB , Wynes MW , Yatabe Y : Updated Molecular Testing Guideline for the Selection of Lung Cancer Patients for Treatment With Targeted Tyrosine Kinase Inhibitors: Guideline From the College of American Pathologists, the International Association for the Study of Lung Cancer, and the Association for Molecular Pathology . Arch Pathol Lab Med 2018 , 142 : 321 – 346 . OpenUrl CrossRef PubMed 8. ↵ Kaur P , Singh SK , Mishra MK , Singh S , Singh R: Promising Combinatorial Therapeutic Strategies against Non-Small Cell Lung Cancer. Cancers , Multidisciplinary Digital Publishing Institute , 2024 , 16 : 2205 . OpenUrl 9. ↵ Chen Q , Wan W , Zhao Q , Li J , Xiong Y , Yuan Y , Tang L , Wu X , Xing W , Guo W , Lu D , Ao L , Xu X , Ao X: Decoding the prognostic landscape of LUAD: the interplay between N6-methyladenosine modification and immune microenvironment . Front Immunol 2024 , 15 : 1514497 . 10. ↵ Lim Z-F , Ma PC: Emerging insights of tumor heterogeneity and drug resistance mechanisms in lung cancer targeted therapy . Journal of Hematology & Oncology 2019 , 12 : 134 . 11. ↵ Rao A , Barkley D , França GS , Yanai I : Exploring tissue architecture using spatial transcriptomics . Nature, Nature Publishing Group , 2021 , 596 : 211 – 220 . 12. ↵ Spatially Resolved Transcriptomes—Next Generation Tools for Tissue Exploration - Asp - 2020 - BioEssays - Wiley Online Library [Internet] . [cited 2025 Aug 24], . Available from: https://onlinelibrary.wiley.com/doi/10.1002/bies.201900221 13. ↵ Coudray N , Ocampo PS , Sakellaropoulos T , Narula N , Snuderl M , Fenyö D , Moreira AL , Razavian N , Tsirigos A: Classification and Mutation Prediction from Non-Small Cell Lung Cancer Histopathology Images using Deep Learning . Nat Med 2018 , 24 : 1559 – 1567 . OpenUrl CrossRef PubMed 14. ↵ Lu MY , Chen TY , Williamson DFK , Zhao M , Shady M , Lipkova J , Mahmood F : AI-based pathology predicts origins for cancers of unknown primary . Nature, Nature Publishing Group , 2021 , 594 : 106 – 110 . 15. ↵ Fu Y , Jung AW , Torne RV , Gonzalez S , Vöhringer H , Shmatko A , Yates LR , Jimenez- Linan M , Moore L , Gerstung M : Pan-cancer computational histopathology reveals mutations, tumor composition and prognosis . Nature Cancer, Nature Publishing Group , 2020 , 1: 800 – 810 . 16. ↵ Schmauch B , Romagnoni A , Pronier E , Saillard C , Maillé P , Calderaro J , Kamoun A , Sefta M , Toldo S , Zaslavskiy M , Clozel T , Moarii M , Courtiol P , Wainrib G: A deep learning model to predict RNA-Seq expression of tumours from whole slide images. Nature Communications , Nature Publishing Group , 2020 , 11 : 3877 . OpenUrl 17. ↵ Holzinger A , Langs G , Denk H , Zatloukal K , Müller H: Causability and explainability of artificial intelligence in medicine . WIREs Data Mining and Knowledge Discovery 2019 , 9 : e1312 . OpenUrl 18. ↵ Tonekaboni S , Joshi S , McCradden MD , Goldenberg A : What Clinicians Want: Contextualizing Explainable Machine Learning for Clinical End Use [Internet] . arXiv, 2019 [cited 2025 Aug 24], . Available from: http://arxiv.org/abs/1905.05134 19. ↵ Ghassemi M , Oakden-Rayner L , Beam AL : The false hope of current approaches to explainable artificial intelligence in health care . [cited 2025 June 27], . Available from: https://www.thelancet.com/journals/landig/article/PIIS2589-7500(21)00208-9/fulltext 20. ↵ AI Action Plan [Internet] . [cited 2025 Aug 3], . Available from: https://www.ai.gov/action-plan 21. ↵ Sukhadia SS , Yadav D , Rao V , Kalra M , Muller K , Head J : XpressO: An explanatory deep learning pipeline for the prediction and visualization of gene expression heterogeneity in breast tumors [Internet]. Research Square , 2025 [cited 2025 June 26], . Available from: https://www.researchsquare.com/article/rs-6167948/v1 22. ↵ Rao VR , Phung TL , Sukhadia SS : XpressO-Melanoma: An Explainable Deep Learning Model for the Prediction of BRAF V600 Mutation Status in Cutaneous Melanomas . bioRxiv 2025 , :2025.07.03.662694. 23. ↵ The Cancer Genome Atlas Program (TCGA) - NCI [Internet] . [cited 2025 June 25], . Available from: https://www.cancer.gov/ccg/research/genome-sequencing/tcga 24. ↵ Lu MY , Williamson DFK , Chen TY , Chen RJ , Barbieri M , Mahmood F : Data-efficient and weakly supervised computational pathology on whole-slide images. Nature Biomedical Engineering , Nature Publishing Group , 2021 , 5 : 555 – 570 . 25. ↵ Welcome to Python.org [Internet] . Python.org. 2025 [cited 2025 June 25], . Available from: https://www.python.org/ 26. ↵ Chen RJ , Ding T , Lu MY , Williamson DFK , Jaume G , Song AH , Chen B , Zhang A , Shao D , Shaban M , Williams M , Oldenburg L , Weishaupt LL , Wang JJ , Vaidya A , Le LP , Gerber G , Sahai S , Williams W , Mahmood F : Towards a general-purpose foundation model for computational pathology . Nature Medicine , Nature Publishing Group , 2024 , 30 : 850 – 862 . OpenUrl CrossRef PubMed 27. ↵ Chu J , Jin L , Teng Y , Li J , Wei Y , Wang Z , Xing J , Yan S , Zhao J : UniParser: Multi- Human Parsing With Unified Correlation Representation Learning . IEEE Transactions on Image Processing 2024 , 33 : 5159 – 5171 . OpenUrl PubMed 28. ↵ Oquab M , Darcet T , Moutakanni T , Vo H , Szafraniec M , Khalidov V , Fernandez P , Haziza D , Massa F , El-Nouby A , Assran M , Ballas N , Galuba W , Howes R , Huang P-Y , Li S-W , Misra I , Rabbat M , Sharma V , Synnaeve G , Xu H , Jegou H , Mairal J , Labatut P , Joulin A , Bojanowski P : DINOv2: Learning Robust Visual Features without Supervision [Internet] . arXiv, 2024 [cited 2025 June 26], . Available from: http://arxiv.org/abs/2304.07193 29. ↵ Schallenberg S , Dernbach G , Dragomir MP , Schlachtenberger G , Boschung K , Friedrich C , Standvoss K , Ruff L , Anders P , Grohé C , Randerath W , Merkelbach-Bruse S , Quaas A , Heldwein M , Keilholz U , Hekmat JK , Rückert C , Büttner R , Horst D , Klauschen F , Frost N : TTF-1 status in early-stage lung adenocarcinoma is an independent predictor of relapse and survival superior to tumor grading . [cited 2025 Aug 24], . Available from: https://www.ejcancer.com/article/S0959-8049(23)00776-1/fulltext 30. Iso H , Hisakane K , Mikami E , Suzuki T , Matsuki S , Atsumi K , Nagata K , Seike M , Hirose T: Thyroid transcription factor-1 (TTF-1) expression and the efficacy of combination therapy with immune checkpoint inhibitors and cytotoxic chemotherapy in non-squamous non-small cell lung cancer . Transl Lung Cancer Res 2023 , 12 : 1850 – 1861 . OpenUrl PubMed 31. Park JY , Jang SH , Kim HI , Kim J-H , Park S , Hwang YI , Jung K-S , Seo J , Lee CY , Ko Y , Park Y-B: Thyroid transcription factor-1 as a prognostic indicator for stage IV lung adenocarcinoma with and without EGFR-sensitizing mutations . BMC Cancer 2019 , 19 : 574 . 32. ↵ Galland L , Le Page AL , Lecuelle J , Bibeau F , Oulkhouir Y , Derangère V , Truntzer C , Ghiringhelli F: Prognostic value of Thyroid Transcription Factor-1 expression in lung adenocarcinoma in patients treated with anti PD-1/PD-L1 . Oncoimmunology 10 : 1957603 . 33. ↵ Identification of immunohistochemical markers for distinguishing lung adenocarcinoma from squamous cell carcinoma - Zhan - Journal of Thoracic Disease [Internet]. [cited 2025 Aug 24], . Available from: https://jtd.amegroups.org/article/view/4756/5207 34. Turner BM , Cagle PT , Sainz IM , Fukuoka J , Shen SS , Jagirdar J : Napsin A , a New Marker for Lung Adenocarcinoma, Is Complementary and More Sensitive and Specific Than Thyroid Transcription Factor 1 in the Differential Diagnosis of Primary Pulmonary Carcinoma: Evaluation of 1674 Cases by Tissue Microarray . Arch Pathol Lab Med 2012 , 136 : 163 – 171 . OpenUrl CrossRef PubMed 35. Lee JG , Kim S , Shim HS: Napsin A is an independent prognostic factor in surgically resected adenocarcinoma of the lung . Lung Cancer 2012 , 77 : 156 – 161 . OpenUrl CrossRef PubMed Web of Science 36. Ueno T , Linder S , Elmberger G: Aspartic proteinase napsin is a useful marker for diagnosis of primary lung adenocarcinoma . Br J Cancer 2003 , 88 : 1229 – 1233 . OpenUrl CrossRef PubMed 37. ↵ Weidemann S , Lukas Böhle J , Luebke AM , Kluth M , Büscheck F , Hube-Magg C , Höflmayer D , Möller K , Fraune C , Bernreuther C , Rink M , Simon R , Menz A , Hinsch A , Lebok P , Clauditz TS , Sauter G , Uhlig R , Wilczak W , Steurer S , Burandt E , Krech R , Dum D , Krech T , Marx A , Minner S : Pathology & Oncology Research | Napsin A Expression in Human Tumors and Normal Tissues . [cited 2025 Sept 11], . Available from: https://www.por-journal.com/journals/pathology-and-oncology-research/articles/10.3389/pore.2021.613099/full?utm_source=chatgpt.com 38. ↵ Li G , Guo J , Mou Y , Luo Q , Wang X , Xue W , Hou T , Zeng T , Yang Y: Keratin gene signature expression drives epithelial-mesenchymal transition through enhanced TGF-β signaling pathway activation and correlates with adverse prognosis in lung adenocarcinoma . Heliyon 2024 , 10 : e24549 . OpenUrl 39. Hosseinalizadeh H , Hussain QM , Poshtchaman Z , Ahsan M , Amin AH , Naghavi S , Mahabady MK: Emerging insights into keratin 7 roles in tumor progression and metastasis of cancers . Front Oncol 2024 , 13 : 1243871 . 40. Mei J , Cai Y , Mussafi O , Zheng M , Xu Y , Chen R , Jiang G , Mao W , Xia W , Wan Y: Identification of adenoid subtype characterized with immune-escaped phenotype in lung squamous carcinoma based on transcriptomics . Experimental Hematology & Oncology 2022 , 11 : 70 . 41. Luo H-T , Liang C-X , Luo R-C , Gu W-G : Identification of relevant prognostic values of cytokeratin 20 and cytokeratin 7 expressions in lung cancer . Biosci Rep 2017 , 37 :BSR20171086. 42. ↵ Camilo R , Capelozzi VL , Coelho Siqueira SA , Del Carlo Bernardi F: Expression of p63, keratin 5/6, keratin 7, and surfactant-A in non–small cell lung carcinomas . Human Pathology 2006 , 37 : 542 – 546 . OpenUrl CrossRef PubMed Web of Science 43. ↵ Ma K , Qiao Y , Wang H , Wang S: Comparative expression analysis of PD-1, PD-L1, and CD8A in lung adenocarcinoma . Annals of Translational Medicine 2020 , 8 : 1478 . OpenUrl PubMed 44. Kist de Ruijter L , van de Donk PP , Hooiveld-Noeken JS , Giesen D , Elias SG , Lub-de Hooge MN , Oosting SF , Jalving M , Timens W , Brouwers AH , Kwee TC , Gietema JA , Fehrmann RSN , Fine BM , Sanabria Bohórquez SM , Yadav M , Koeppen H , Jing J , Guelman S , Lin MT , Mamounas MJ , Eastham JR , Kimes PK , Williams SP , Ungewickell A , de Groot DJA , de Vries EGE : Whole-body CD8+ T cell visualization before and during cancer immunotherapy: a phase 1/2 trial . Nature Medicine , Nature Publishing Group , 2022 , 28 : 2601 – 2610 . OpenUrl CrossRef PubMed 45. Chen Y , Xin Z , Huang L , Zhao L , Wang S , Cheng J , Wu P , Chai Y : Frontiers | CD8+ T Cells Form the Predominant Subset of NKG2A+ Cells in Human Lung Cancer . [cited 2025 Sept 11], . Available from: https://www.frontiersin.org/journals/immunology/articles/10.3389/fimmu.2019.03002/full 46. Komuro H , Shinohara S , Fukushima Y , Demachi-Okamura A , Muraoka D , Masago K , Matsui T , Sugita Y , Takahashi Y , Nishida R , Takashima C , Ohki T , Shigematsu Y , Watanabe F , Adachi K , Fukuyama T , Hamana H , Kishi H , Miura D , Tanaka Y , Onoue K , Onoguchi K , Yamashita Y , Stratford R , Clancy T , Yamaguchi R , Kuroda H , Doi K , Iwata H , Matsushita H: Single-cell sequencing on CD8+ TILs revealed the nature of exhausted T cells recognizing neoantigen and cancer/testis antigen in non-small cell lung cancer. Journal for ImmunoTherapy of Cancer [Internet] , BMJ Publishing Group Ltd , 2023 [cited 2025 Sept 11], . Available from: https://jitc.bmj.com/content/11/8/e007180 47. Feng J , Xu L , Zhang S , Geng L , Zhang T , Yu Y , Yuan R , He Y , Nan Z , Lin M , Guo H : Frontiers | A robust CD8+ T cell-related classifier for predicting the prognosis and efficacy of immunotherapy in stage III lung adenocarcinoma . [cited 2025 Sept 11], . Available from: https://www.frontiersin.org/journals/immunology/articles/10.3389/fimmu.2022.993187/full 48. ↵ Liu C , Hu Q , Hu K , Su H , Shi F , Kong L , Zhu H , Yu J: Increased CD8+CD28+ T cells independently predict better early response to stereotactic ablative radiotherapy in patients with lung metastases from non-small cell lung cancer . J Transl Med 2019 , 17 : 120 . 49. ↵ Xu S , Xu Y , Chai W , Liu X , Li J , Sun L , Pan H , Yan M : KLRB1 expression is associated with lung adenocarcinoma prognosis and immune infiltration and regulates lung adenocarcinoma cell proliferation and metastasis through the MAPK/ERK pathway . [cited 2025 Aug 24], . Available from: https://jtd.amegroups.org/article/view/87711/html 50. Chen J-L , Wu C-Y , Luo X-Y , Wang X-Y , Wang F-M , Huang X , Yuan W , Guo Q: Down- regulation of KLRB1 is associated with increased cell growth, metastasis, poor prognosis, as well as a dysfunctional immune microenvironment in LUAD. Scientific Reports , Nature Publishing Group , 2024 , 14 : 11782 . 51. ↵ Cheng X , Cao Y , Wang X , Cheng L , Liu Y , Lei J , Peng W , Shi D: Systematic Pan-Cancer Analysis of KLRB1 with Prognostic Value and Immunological Activity across Human Tumors . Journal of Immunology Research 2022 , 2022: 5254911 . 52. ↵ Park S , Cha H , Kim HS , Lee B , Kim S , Kim TM , Jung HA , Sun J , Ahn JS , Ahn M , Park K , Park W , Lee S: Transcriptional upregulation of CXCL13 is correlated with a favorable response to immune checkpoint inhibitors in lung adenocarcinoma . Cancer Med 2022 , 12 : 7639 – 7650 . OpenUrl PubMed 53. ↵ Wang G-Z , Cheng X , Zhou B , Wen Z-S , Huang Y-C , Chen H-B , Li G-F , Huang Z-L , Zhou Y-C , Feng L , Wei M-M , Qu L-W , Cao Y , Zhou G-B: The chemokine CXCL13 in lung cancers associated with environmental polycyclic aromatic hydrocarbons pollution . eLife 4 : e09419 . 54. Gao S-H , Liu S-Z , Wang G-Z , Zhou G-B : CXCL13 in Cancer and Other Diseases: Biological Functions , Clinical Significance, and Therapeutic Opportunities. Life (Basel ) 2021 , 11 : 1282 . OpenUrl PubMed 55. Park S , Cha H , Kim HS , Lee B , Kim S , Kim TM , Jung HA , Sun J-M , Ahn JS , Ahn M-J , Park K , Park W-Y , Lee S-H: Transcriptional upregulation of CXCL13 is correlated with a favorable response to immune checkpoint inhibitors in lung adenocarcinoma . Cancer Medicine 2023 , 12 : 7639 – 7650 . OpenUrl PubMed 56. Vahidian F , Lamaze FC , Bouffard C , Coulombe F , Gagné A , Blais F , Tonneau M , Orain M , Routy B , Manem VSK , Joubert P: CXCL13 Positive Cells Localization Predict Response to Anti-PD-1/PD-L1 in Pulmonary Non-Small Cell Carcinoma . Cancers (Basel ) 2024 , 16 : 708 . 57. ↵ Sakai SA , Oyoshi H , Nakamura M , Taki T , Nomura K , Hojo H , Hirata H , Motegi A , Nakamura Y , Zenkoh J , Aokage K , Hamada A , Kojima M , Kuwata T , Tsuchihara K , Akimoto T , Soh J , Mitsudomi T , Tsuboi M , Ishii G , Suzuki Y , Suzuki A , Yamashita R , Kageyama S-I: Single-cell spatial analysis with Xenium reveals anti-tumour responses of CXCL13 + T and CXCL9+ cells after radiotherapy combined with anti-PD-L1 therapy. British Journal of Cancer , Nature Publishing Group , 2025 , : 1 – 14 . 58. ↵ Song T , Li S , Zhao K , Zou D , Zhang M , Wang H: Comprehensive analysis of prognosis and tumor immune microenvironment of cuproptosis-related gene CDKN2A in lung adenocarcinoma . BMC Pulmonary Medicine 2025 , 25 : 179 . 59. Kim N , Song M , Kim S , Seo Y , Kim Y , Yoon S : Differential regulation and synthetic lethality of exclusive RB1 and CDKN2A mutations in lung cancer . [cited 2025 Sept 11], . Available from: https://www.spandidos-publications.com/10.3892/ijo.2015.3262 60. Tam KW , Zhang W , Soh J , Stastny V , Chen M , Sun H , Thu K , Rios JJ , Yang C , Marconett CN , Selamat SA , Laird-Offringa IA , Taguchi A , Hanash S , Shames D , Ma X , Zhang MQ , Lam WL , Gazdar A: CDKN2A/p16 inactivation mechanisms and their relationship to smoke exposure and molecular features in non-small cell lung cancer . J Thorac Oncol 2013 , 8 : 1378 – 1388 . OpenUrl CrossRef PubMed 61. Schuster K , Venkateswaran N , Rabellino A , Girard L , Peña-Llopis S , Scaglioni PP: Nullifying the CDKN2AB Locus Promotes Mutant K-ras Lung Tumorigenesis . Mol Cancer Res 2014 , 12 : 912 – 923 . OpenUrl Abstract / FREE Full Text 62. Gutiontov SI , Turchan WT , Spurr LF , Rouhani SJ , Chervin CS , Steinhardt G , Lager AM , Wanjari P , Malik R , Connell PP , Chmura SJ , Juloori A , Hoffman PC , Ferguson MK , Donington JS , Patel JD , Vokes EE , Weichselbaum RR , Bestvina CM , Segal JP , Pitroda SP: CDKN2A loss-of-function predicts immunotherapy resistance in non-small cell lung cancer . Sci Rep 2021 , 11 : 20059 . 63. Song T , Li S , Zhao K , Zou D , Zhang M , Wang H: Comprehensive analysis of prognosis and tumor immune microenvironment of cuproptosis-related gene CDKN2A in lung adenocarcinoma . BMC Pulmonary Medicine 2025 , 25 : 179 . 64. ↵ Li D-M , Li G-S , Li J-D , Chen F , Huang H , Huang W-Y , Huang Z-G , Dang Y-W , Tang Y- L , Tang Z-Q , Tang W-J , Chen G , Lu H-P: Clinical significance and prospective mechanism of increased CDKN2A expression in small cell lung cancer . Clin Transl Oncol 2024 , 26 : 1519 – 1531 . OpenUrl PubMed 65. ↵ Chen T , Ashwood LM , Kondrashova O , Strasser A , Kelly G , Sutherland KD : Breathing new insights into the role of mutant p53 in lung cancer. Oncogene , Nature Publishing Group , 2025 , 44 : 115 – 129 . 66. Li M , Li S , Liu B , Gu M-M , Zou S , Xiao B-B , Yu L , Ding W-Q , Zhou P-K , Zhou J , Shang Z-F: PIG3 promotes NSCLC cell mitotic progression and is associated with poor prognosis of NSCLC patients . J Exp Clin Cancer Res 2017 , 36 : 39 . 67. Gu M , Gao D , Yao P , Yu L , Yang X , Xing C , Zhou J , Shang Z , Li M: p53 inducible gene 3 promotes cell migration and invasion by activating the FAK/Src pathway in lung adenocarcinoma . Cancer Sci 2018 , 109 : 3783 – 3793 . OpenUrl CrossRef PubMed 68. Chen X , Zhang W , Xu X: Cyanidin-3-glucoside suppresses the progression of lung adenocarcinoma by downregulating TP53I3 and inhibiting PI3K/AKT/mTOR pathway . World J Surg Oncol 2021 , 19 : 232 . 69. ↵ Zhou H , Tang L , Shang Z-F , Zhou P-K , Li M: PIG3 downregulation enhances the radiosensitivity of NSCLC cells by promoting G2/M cell cycle arrest and apoptosis . Radiation Medicine and Protection 2023 , 4 : 19 – 25 . OpenUrl 70. ↵ Lu Z , Sun F: Downregulated TICAM1 is a prognostic biomarker and associated with immune tolerance of Wilms tumor patients . BMC Medical Genomics 2022 , 15 : 174 . 71. Miao Z , Yu W: Significance of novel PANoptosis genes to predict prognosis and therapy effect in the lung adenocarcinoma. Scientific Reports , Nature Publishing Group , 2024 , 14 : 20934 . 72. Lu Z , Sun F: Downregulated TICAM1 is a prognostic biomarker and associated with immune tolerance of Wilms tumor patients . BMC Medical Genomics 2022 , 15 : 174 . 73. Ono J , Shime H , Takaki H , Takashima K , Funami K , Yoshida S , Takeda Y , Matsumoto M , Kasahara M , Seya T: The TLR3/TICAM-1 signal constitutively controls spontaneous polyposis through suppression of c-Myc in ApcMin/+mice . Journal of Biomedical Science 2017 , 24 : 79 . 74. ↵ Miyashita Y , Kouwaki T , Tsukamoto H , Okamoto M , Nakamura K , Oshiumi H: TICAM- 1/TRIF associates with Act1 and suppresses IL-17 receptor–mediated inflammatory responses . Life Sci Alliance 2021 , 5 : e202101181 . OpenUrl Abstract / FREE Full Text 75. ↵ Ma W , Hu J: Downregulated CDH3 is correlated with a better prognosis for LUAD and decreases proliferation and migration of lung cancer cells. Genes & Genomics , Springer , 2023 , 46 : 713 – 731 . OpenUrl 76. Gong W , Gao F , He H , Zhang Z , Shao F , Gao Y , He J: CDH3 promotes the progression of lung adenocarcinoma through driving epithelial-mesenchymal transition progress . J Transl Med 2025 , 23 : 887 . 77. ↵ Hsiao T-F , Wang C-L , Wu Y-C , Feng H-P , Chiu Y-C , Lin H-Y , Liu K-J , Chang G-C , Chien K-Y , Yu J-S , Yu C-J: Integrative Omics Analysis Reveals Soluble Cadherin-3 as a Survival Predictor and an Early Monitoring Marker of EGFR Tyrosine Kinase Inhibitor Therapy in Lung Cancer . Clin Cancer Res 2020 , 26 : 3220 – 3229 . OpenUrl Abstract / FREE Full Text 78. ↵ Huang N , Liu Z , Li R: Pan-cancer analysis reveals the relationship between SLC47A1 immune infiltration and clinical prognosis of human tumors . Journal of Holistic Integrative Pharmacy 2023 , 4 : 103 – 118 . OpenUrl 79. Yao X , Watkins NH , Brown-Harding H , Bierbach U: A membrane transporter determines the spectrum of activity of a potent platinum–acridine hybrid anticancer agent. Scientific Reports , Nature Publishing Group , 2020 , 10 : 15201 . 80. ↵ Zhang S , Wu H , Day CS , Bierbach U: Platinum–Acridine Agents with High Activity in Cancers Expressing the Solute Carrier MATE1 (SLC47A1) . ACS Med Chem Lett 2023 , 14 : 1122 – 1128 . OpenUrl PubMed 81. ↵ Song H , Wu J , Liu W , Cai K , Xie Z , Liu Y , Huang J , Gan S , Xiong Y , Sun Y: Key genes involved with prognosis were identified in lung adenocarcinoma by integrated bioinformatics analysis . Heliyon 2023 , 9 : e16789 . OpenUrl PubMed 82. ↵ Chen C , Bartenhagen C , Gombert M , Okpanyi V , Binder V , Röttgers S , Bradtke J , Teigler- Schlegel A , Harbott J , Ginzel S , Thiele R , Fischer U , Dugas M , Hu J , Borkhardt A: Next-generation-sequencing-based risk stratification and identification of new genes involved in structural and sequence variations in near haploid lymphoblastic leukemia. Genes , Chromosomes and Cancer 2013 , 52 : 564 – 579 . OpenUrl 83. ↵ Pustovalova M , Malakhov P , Guryanova A , Sorokin M , Suntsova M , Buzdin A , Osipov AN , Leonov S: Transcriptome-Based Traits of Radioresistant Sublines of Non-Small Cell Lung Cancer Cells. International Journal of Molecular Sciences , Multidisciplinary Digital Publishing Institute , 2023 , 24 : 3042 . OpenUrl 84. ↵ Medical Science Monitor | The Utility of Thyroid Transcription Factor 1 (TTF-1), Napsin A, Excision Repair Cross-Complementing 1 (ERCC1), Anaplastic Lymphoma Kinase (ALK) and the Epidermal Growth Factor Receptor (EGFR) Expression in Small Biopsy in Prognosis of Patients with Lung Adenocarcinoma – A Retrograde Single-Center Study from Croatia - Article abstract #899378 [Internet]. [cited 2025 Aug 24], . Available from: https://medscimonit.com/abstract/index/idArt/899378 85. ↵ Ye J , Findeis-Hosey JJ , Yang Q , McMahon LA , Yao JL , Li F , Xu H : Combination of Napsin A and TTF-1 Immunohistochemistry… : Applied Immunohistochemistry & Molecular Morphology . [cited 2025 Aug 24], . Available from: https://journals.lww.com/appliedimmunohist/abstract/2011/07000/combination_of_napsin_a_and_ttf_1.6.aspx 86. ↵ Ren X , Wen X , Ren Y-J , Liu X , Wang J , Hao M , Ma Q , Ren J , Jin B , Qiao X , Li B , Wu J , Li X , Liu Z: Significance of thyroid transcription factor 1 and Napsin A for prompting the status of EGFR mutations in lung adenocarcinoma patients . J Thorac Dis 2022 , 14 : 4395 – 4404 . OpenUrl PubMed 87. ↵ Katayama Y , Yamada T , Morimoto K , Fujii H , Morita S , Tanimura K , Takeda T , Okada A , Shiotsu S , Chihara Y , Hiranuma O , Yamada T , Ota T , Harada T , Hasegawa I , Yoshimura A , Iwasaku M , Tokuda S , Kim YH , Takayama K : TTF-1 Expression and Clinical Outcomes of Combined Chemoimmunotherapy in Patients With Advanced Lung Adenocarcinoma: A Prospective Observational Study . JTO Clin Res Rep 2023 , 4 : 100494 . OpenUrl PubMed 88. ↵ Hosseinalizadeh H , Hussain QM , Poshtchaman Z , Ahsan M , Amin AH , Naghavi S , Mahabady MK : Frontiers | Emerging insights into keratin 7 roles in tumor progression and metastasis of cancers . [cited 2025 Aug 24], . Available from: https://www.frontiersin.org/journals/oncology/articles/10.3389/fonc.2023.1243871/full 89. ↵ Guo D , Wang M , Shen Z , Zhu J: A new immune signature for survival prediction and immune checkpoint molecules in lung adenocarcinoma . Journal of Translational Medicine 2020 , 18 : 123 . 90. ↵ Gutiontov SI , Turchan WT , Spurr LF , Rouhani SJ , Chervin CS , Steinhardt G , Lager AM , Wanjari P , Malik R , Connell PP , Chmura SJ , Juloori A , Hoffman PC , Ferguson MK , Donington JS , Patel JD , Vokes EE , Weichselbaum RR , Bestvina CM , Segal JP , Pitroda SP: CDKN2A loss-of-function predicts immunotherapy resistance in non-small cell lung cancer. Scientific Reports , Nature Publishing Group , 2021 , 11 : 20059 . 91. ↵ Lin C-C , Chang T-C , Wang Y , Guo L , Gao Y , Bikorimana E , Lemoff A , Fang YV , Zhang H , Zhang Y , Ye D , Soria-Bretones I , Servetto A , Lee K , Luo X , Otto JJ , Akamatsu H , Napolitano F , Mani R , Cescon DW , Xu L , Xie Y , Mendell JT , Hanker AB , Arteaga CL: PRMT5 is an actionable therapeutic target in CDK4/6 inhibitor-resistant ER+/RB- deficient breast cancer. Nature Communications , Nature Publishing Group , 2024 , 15 : 2287 . OpenUrl 92. ↵ Han G , Yang G , Hao D , Lu Y , Thein K , Simpson BS , Chen J , Sun R , Alhalabi O , Wang R , Dang M , Dai E , Zhang S , Nie F , Zhao S , Guo C , Hamza A , Czerniak B , Cheng C , Siefker- Radtke A , Bhat K , Futreal A , Peng G , Wargo J , Peng W , Kadara H , Ajani J , Swanton C , Litchfield K , Ahnert JR , Gao J , Wang L: 9p21 loss confers a cold tumor immune microenvironment and primary resistance to immune checkpoint therapy. Nature Communications , Nature Publishing Group , 2021 , 12 : 5606 . OpenUrl 93. ↵ Frontiers | PRMT5 Inhibition Promotes PD-L1 Expression and Immuno-Resistance in Lung Cancer [Internet] . [cited 2025 Aug 24], . Available from: https://www.frontiersin.org/journals/immunology/articles/10.3389/fimmu.2021.722188/ful l 94. ↵ Ingram K , Samson SC , Zewdu R , Zitnay RG , Snyder EL , Mendoza MC : NKX2-1 controls lung cancer progression by inducing DUSP6 to dampen ERK activity. Oncogene , Nature Publishing Group , 2022 , 41 : 293 – 300 . 95. ↵ Kotsinas A , Aggarwal V , Tan E-J , Levy B , Gorgoulis VG : PIG3: A novel link between oxidative stress and DNA damage response in cancer . Cancer Letters 2012 , 327 : 97 – 102 . OpenUrl CrossRef PubMed Web of Science 96. ↵ Li H , Yang L , Wang Y , Wang L , Chen G , Zhang L , Wang D: Integrative analysis of TP53 mutations in lung adenocarcinoma for immunotherapies and prognosis . BMC Bioinformatics 2023 , 24 : 155 . 97. ↵ Gu M , Gao D , Yao P , Yu L , Yang X , Xing C , Zhou J , Shang Z , Li M: p53 inducible gene 3 promotes cell migration and invasion by activating the FAK/Src pathway in lung adenocarcinoma . Cancer Sci 2018 , 109 : 3783 – 3793 . OpenUrl CrossRef PubMed 98. ↵ Kim T , Kim S , Lee S , Hwang S , Han J , Jeong H , Choi Y-L : Linking p53 immunostaining to TP53 mutation status in patients with non-small cell lung cancer . Pathology [Internet] 2025 [cited 2025 Sept 8], . Available from: https://www.sciencedirect.com/science/article/pii/S0031302525002235 99. ↵ Shen M , Jiang X , Peng Q , Oyang L , Ren Z , Wang J , Peng M , Zhou Y , Deng X , Liao Q: The cGAS STING pathway in cancer immunity: mechanisms, challenges, and therapeutic implications . Journal of Hematology & Oncology 2025 , 18 : 40 . 100. ↵ Takashima K , Oshiumi H , Matsumoto M , Seya T: TICAM-1 is dispensable in STING- mediated innate immune responses in myeloid immune cells . Biochemical and Biophysical Research Communications 2018 , 499 : 985 – 991 . OpenUrl CrossRef PubMed 101. ↵ Wu J , Chen Y , Xie M , Yu X , Su C: cGAS-STING signaling pathway in lung cancer: Regulation on antitumor immunity and application in immunotherapy . Chinese Medical Journal Pulmonary and Critical Care Medicine 2024 , 2 : 257 – 264 . OpenUrl 102. ↵ Gammelgaard KR , Sandfeld-Paulsen B , Godsk SH , Demuth C , Meldgaard P , Sorensen BS , Jakobsen MR : cGAS-STING pathway expression as a prognostic tool in NSCLC . [cited 2025 Aug 24], . Available from: https://tlcr.amegroups.org/article/view/46144/html 103. ↵ Zhang S , Wu H , Day CS , Bierbach U: Platinum–Acridine Agents with High Activity in Cancers Expressing the Solute Carrier MATE1 (SLC47A1) . ACS Med Chem Lett 2023 , 14 : 1122 – 1128 . OpenUrl PubMed 104. ↵ Batara DC , Park SW , Kim H-J , Choi S-Y , Ohn T , Choi M-C , Park S-I , Kim S-H: Targeting the multidrug and toxin extrusion 1 gene (SLC47A1) sensitizes glioma stem cells to temozolomide . Am J Cancer Res 2023 , 13 : 4021 – 4038 . OpenUrl PubMed 105. ↵ Huang N , Liu Z , Li R: Pan-cancer analysis reveals the relationship between SLC47A1 immune infiltration and clinical prognosis of human tumors . Journal of Holistic Integrative Pharmacy 2023 , 4 : 103 – 118 . OpenUrl 106. ↵ Chen Y , Wu Y , Zhao Z , Wen L , Wu M , Song D , Zeng Q , Liu Y , Yan G , Zhang G: Retrospective study on the correlation between CXCL13, immune infiltration, and tertiary lymphoid structures in cutaneous squamous cell carcinoma . [ cited 2025 Aug 24], . Available from: https://peerj.com/articles/19398 107. ↵ Park S , Cha H , Kim HS , Lee B , Kim S , Kim TM , Jung HA , Sun J , Ahn JS , Ahn M , Park K , Park W , Lee S: Transcriptional upregulation of CXCL13 is correlated with a favorable response to immune checkpoint inhibitors in lung adenocarcinoma . Cancer Med 2022 , 12 : 7639 – 7650 . OpenUrl PubMed 108. ↵ Mathewson ND , Ashenberg O , Tirosh I , Gritsch S , Perez EM , Marx S , Jerby-Arnon L , Chanoch-Myers R , Hara T , Richman AR , Ito Y , Pyrdol J , Friedrich M , Schumann K , Poitras MJ , Gokhale PC , Castro LNG , Shore ME , Hebert CM , Shaw B , Cahill HL , Drummond M , Zhang W , Olawoyin O , Wakimoto H , Rozenblatt-Rosen O , Brastianos PK , Liu XS , Jones PS , Cahill DP , Frosch MP , Louis DN , Freeman GJ , Ligon KL , Marson A , Chiocca EA , Reardon DA , Regev A , Suvà ML , Wucherpfennig KW : Inhibitory CD161 receptor identified in glioma-infiltrating T cells by single-cell analysis . [cited 2025 Aug 24], . Available from: https://www.cell.com/cell/abstract/S0092-8674(21)00065-9 109. ↵ Wang Y , Qin X , Dong W , Lei C , Zheng S , Salem MM , Bekhit MM , Almuraikhi N: Cadherin family genes in non-small cell lung cancer: implications for diagnosis, prognosis, and targeted therapy . Am J Transl Res 2025 , 17 : 3718 – 3737 . OpenUrl PubMed 110. ↵ Vieira AF , Paredes J: P-cadherin and the journey to cancer metastasis . Molecular Cancer 2015 , 14 : 178 . 111. ↵ Xu Y , Zhao J , Dai X , Xie Y , Dong M: High expression of CDH3 predicts a good prognosis for colon adenocarcinoma patients . Exp Ther Med 2019 , 18 : 841 – 847 . OpenUrl PubMed 112. ↵ Szklarczyk D , Nastou K , Koutrouli M , Kirsch R , Mehryary F , Hachilif R , Hu D , Peluso ME , Huang Q , Fang T , Doncheva NT , Pyysalo S , Bork P , Jensen LJ , von Mering C : The STRING database in 2025: protein networks with directionality of regulation . Nucleic Acids Res 2025 , 53 : D730 – D737 . OpenUrl CrossRef PubMed 113. ↵ Grissa D , Junge A , Oprea TI , Jensen LJ: Diseases 2.0: a weekly updated database of disease–gene associations from text mining and data integration . Database (Oxford ) 2022 , 2022:baac019. 114. ↵ Xia L , Yu Y , Lan F , Yan J , Li J , Li W , Xia Y : Case Report: Tumor Microenvironment Characteristics in a Patient With HER2 Mutant Lung Squamous Cell Carcinoma Harboring High PD-L1 Expression Who Presented Hyperprogressive Disease . Front Oncol 2021 , 11 : 760703 . 115. ↵ Gregorc V , Lazzari C , Mandalá M , Ippati S , Bulotta A , Cangi MG , Khater A , Viganò MG , Mirabile A , Pecciarini L , Ogliari FR , Arrigoni G , Grassini G , Veronesi G , Doglioni C : Intratumoral Cellular Heterogeneity: Implications for Drug Resistance in Patients with Non-Small Cell Lung Cancer . Cancers (Basel ) 2021 , 13 : 2023 . OpenUrl PubMed 116. ↵ Wu H-J , Temko D , Maliga Z , Moreira AL , Sei E , Minussi DC , Dean J , Lee C , Xu Q , Hochart G , Jacobson CA , Yapp C , Schapiro D , Sorger PK , Seeley EH , Navin N , Downey RJ , Michor F: Spatial intra-tumor heterogeneity is associated with survival of lung adenocarcinoma patients . Cell Genom 2022 , 2 : 100165 . 117. ↵ Yatabe Y , Dacic S , Borczuk AC , Warth A , Russell PA , Lantuejoul S , Beasley MB , Thunnissen E , Pelosi G , Rekhtman N , Bubendorf L , Mino-Kenudson M , Yoshida A , Geisinger KR , Noguchi M , Chirieac LR , Bolting J , Chung J-H , Chou T-Y , Chen G , Poleri C , Lopez-Rios F , Papotti M , Sholl LM , Roden AC , Travis WD , Hirsch FR , Kerr KM , Tsao M-S , Nicholson AG , Wistuba I , Moreira AL: Best Practices Recommendations for Diagnostic Immunohistochemistry in Lung Cancer . J Thorac Oncol 2019 , 14 : 377 – 407 . OpenUrl PubMed 118. ↵ Zhang K , Li J , Yuan E: A necroptosis-related gene signature to predict prognosis and immune features in hepatocellular carcinoma . BMC Cancer 2023 , 23 : 660 . 119. ↵ Gascón M , Isla D , Cruellas M , Gálvez EM , Lastra R , Ocáriz M , Paño JR , Ramírez A , Sesma A , Torres-Ramón I , Yubero A , Pardo J , Martínez-Lostao L : Intratumoral versus Circulating Lymphoid Cells as Predictive Biomarkers in Lung Cancer Patients Treated with Immune Checkpoint Inhibitors: Is the Easiest Path the Best One? Cells 2020 , 9 : 1525 . OpenUrl 120. ↵ Komuro H , Shinohara S , Fukushima Y , Demachi-Okamura A , Muraoka D , Masago K , Matsui T , Sugita Y , Takahashi Y , Nishida R , Takashima C , Ohki T , Shigematsu Y , Watanabe F , Adachi K , Fukuyama T , Hamana H , Kishi H , Miura D , Tanaka Y , Onoue K , Onoguchi K , Yamashita Y , Stratford R , Clancy T , Yamaguchi R , Kuroda H , Doi K , Iwata H , Matsushita H: Single-cell sequencing on CD8+ TILs revealed the nature of exhausted T cells recognizing neoantigen and cancer/testis antigen in non-small cell lung cancer . J Immunother Cancer 2023 , 11 : e007180 . OpenUrl Abstract / FREE Full Text 121. ↵ Chung JH , Ali SM , Davis J , Robstad K , McNally R , Gay LM , Erlich RL , Palma NA , Stephens PJ , Miller VA , Cutugno A , Ross JS: A Poorly Differentiated Malignant Neoplasm Lacking Lung Markers Harbors an EML4-ALK Rearrangement and Responds to Crizotinib . Case Rep Oncol 2014 , 7 : 628 – 632 . OpenUrl PubMed 122. ↵ Jiao Y , Liu M , Luo N , Guo H , Li J : Novel MRPL13-ALK and PPP1CB-ALK Double Fusion As a Potential Mechanism of Acquired Resistance to First-Line Osimertinib in EGFR-Mutant High-Grade Neuroendocrine Tumor of the Lung . JTO Clin Res Rep 2020 , 1 : 100079 . OpenUrl PubMed 123. ↵ Liu H , Chen L-H , Zhang Z-H , Wang N , Zhuang S-H , Chen H , Du J , Pang L-J , Qi Y : Histomorphological transformation from non-small cell lung carcinoma to small cell lung carcinoma after targeted therapy or immunotherapy: A report of two cases . Front Oncol 2022 , 12 : 1022705 . 124. ↵ Huang T , Li J , Zhang C , Hong Q , Jiang D , Ye M , Duan S: Distinguishing Lung Adenocarcinoma from Lung Squamous Cell Carcinoma by Two Hypomethylated and Three Hypermethylated Genes: A Meta-Analysis . PLoS One 2016 , 11 : e0149088 . OpenUrl PubMed View the discussion thread. Back to top Previous Next Posted September 20, 2025. Download PDF Supplementary Material Email Thank you for your interest in spreading the word about bioRxiv. NOTE: Your email address is requested solely to identify you as the sender of this article. Your Email * Your Name * Send To * Enter multiple addresses on separate lines or separate them with commas. You are going to email the following Interpretable Deep Learning Reveals Biologically Relevant Spatial Gene Expression Patterns in Lung Tumors and their Microenvironment Message Subject (Your Name) has forwarded a page to you from bioRxiv Message Body (Your Name) thought you would like to see this page from the bioRxiv website. Your Personal Message CAPTCHA This question is for testing whether or not you are a human visitor and to prevent automated spam submissions. Share Interpretable Deep Learning Reveals Biologically Relevant Spatial Gene Expression Patterns in Lung Tumors and their Microenvironment Vibha R. Rao , Adrienne A. Workman , Liang Lu , Xiaoying Liu , Shrey S. Sukhadia bioRxiv 2025.09.17.674367; doi: https://doi.org/10.1101/2025.09.17.674367 Share This Article: Copy Citation Tools Interpretable Deep Learning Reveals Biologically Relevant Spatial Gene Expression Patterns in Lung Tumors and their Microenvironment Vibha R. Rao , Adrienne A. Workman , Liang Lu , Xiaoying Liu , Shrey S. Sukhadia bioRxiv 2025.09.17.674367; doi: https://doi.org/10.1101/2025.09.17.674367 Citation Manager Formats BibTeX Bookends EasyBib EndNote (tagged) EndNote 8 (xml) Medlars Mendeley Papers RefWorks Tagged Ref Manager RIS Zotero Tweet Widget Facebook Like Google Plus One Subject Area Pathology Subject Areas All Articles Animal Behavior and Cognition (7635) Biochemistry (17697) Bioengineering (13895) Bioinformatics (41951) Biophysics (21456) Cancer Biology (18594) Cell Biology (25520) Clinical Trials (138) Developmental Biology (13381) Ecology (19903) Epidemiology (2067) Evolutionary Biology (24323) Genetics (15612) Genomics (22510) Immunology (17738) Microbiology (40401) Molecular Biology (17184) Neuroscience (88622) Paleontology (667) Pathology (2833) Pharmacology and Toxicology (4825) Physiology (7644) Plant Biology (15158) Scientific Communication and Education (2046) Synthetic Biology (4296) Systems Biology (9825) Zoology (2271)