PRESSnet: a novel framework for patient stratification and biomarker discovery using clinical knowledge graphs

PRESSnet: a novel framework for patient stratification and biomarker discovery using clinical knowledge graphs | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article PRESSnet: a novel framework for patient stratification and biomarker discovery using clinical knowledge graphs Jake Cohen-Setton, Shruti Shikhare, Ioannis Kagiampakis, Domingo Salazar, and 7 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8289747/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted You are reading this latest preprint version Abstract Harnessing multiomic data for patient stratification and biomarker discovery is vital for effective personalized medicine. However, as the volume and heterogeneity of data increases, so do the challenges of data integration and explainable hypotheses. To address this, we present PRESSnet (Patient REcommendation via Stratification and Selection using networks), an end-to-end framework leveraging multimodal patient knowledge graphs (KGs) for stratification and biomarker discovery. PRESSnet incorporates graph artificial intelligence and network algorithms in scalable, flexible analysis pipelines that can integrate underlying multiomic patient features with prior knowledge such as curated gene pathway data. Applied to patients from two different cancer types, PRESSnet generates explainable stratification hypotheses and captured known survival biomarkers as well as novel composite signatures that comparatively increased statistically significant survival separation compared to univariate markers. These biomarkers were validated for their translatability to unseen patients within cohorts in IO-treated NSCLC patients (MSK 2022) and across independent datasets in AML patients (TCGA and Beat AML). PRESSnet compared favourably to benchmark models such as MOFA + for stratification and Random Forests for biomarker generation and survival risk classification. We also demonstrate that PRESSnet’s ability to model prior knowledge can improve patient survival prediction, including in small datasets, and offer context-relevant insights into signalling pathways and regulatory networks involved in therapy resistance. PRESSnet is provided as a lightweight, adaptable framework for the scientific community to inform research into patient selection, asset positioning and trial design. Biological sciences/Computational biology and bioinformatics/Machine learning Biological sciences/Drug discovery/Biomarkers Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Introduction Against a background of diverse technological and scientific advances in recent years, the rate of success for clinical trials has remained low. 1 Currently, approximately 9 of 10 new molecular entities entering a clinical study are likely to fail. 2 A key contributor to success, or lack thereof, is patient selection. Personalised medicine has become a ubiquitous goal in pharmaceutical R&D: identifying the optimum therapy for a clearly definable population of patients saves lives and improves the quality of life for trial participants while making medicines more affordable. Major pharmaceutical companies have adopted personalised medicine as a core part of their R&D strategy (for example, AstraZeneca’s 5Rs framework 3 ), impacting all stages of drug discovery from target discovery to identification of disease subgroups. These advances have led to demonstrable improvements in numerous areas, including phase 3 trial completion. 3 As part of these developments, the use of biomarkers to define populations and stratify patients during trial design has become the norm across diseases. Multiple biomarker-driven oncology therapies have been approved in the last decade 4 : these include osimertinib, which targets patients with the T790m EGFR mutation in non–squamous-cell lung cancer (NSCLC) 5 ; anti-HER2 therapies such as trastuzumab deruxtecan, 6 which target HER2-positive breast and other cancers; and immune checkpoint blockade therapies such as pembrolizumab, 7 for which high programmed cell death ligand-1 (PD-L1) status is a key patient selection criterion. Other major biomarker-driven therapies include those for patients with BRCA -mutant breast and ovarian cancers, who are typically treated with poly(ADP-ribose) polymerase (PARP) inhibitors, 8 and patients with high microsatellite instability (MSI) disease across multiple cancer types, who are more likely to benefit from immunotherapy (IO). 9 The fields of data science and artificial intelligence (AI) are now increasingly focusing on biomarker identification and patient stratification, and multiple computational pipelines for these, some of which are AI driven, have recently been developed 10 . Statistical models such as the Cox proportional hazards model have also long been applied to survival analysis 11 and are still commonly used today as interpretable tools to investigate the relationship between variables and survival time. However, all current computational approaches face multiple challenges. In many cases, the number of features available in a dataset far exceeds the number of patients, which for early-phase trial arms often falls below 100. These challenges raise potential issues including algorithm overfitting, feature redundancy, and spurious relationship generation. Patient data are typically heterogeneous and multimodal, incorporating clinical features, omics, and other patient-level readouts from imaging or other sources. Some feature modalities, such as those based on mutations, necessitate the accommodation of high levels of feature sparsity. In addition, univariate biomarkers extracted from these data can lack the granularity to distinguish patient survival trajectories across multiple datasets for the same indication, line of treatment, and other fixed conditions, especially when measurements are not standardised. 12 An approach that can easily integrate multimodal data while also extending beyond univariate recommendations would therefore be highly advantageous. For these reasons, knowledge graphs (KGs) carry promise for this area of research. A network graph is a logical map of relationships (‘edges’) between entities (‘nodes’). Computational approaches based on homogeneous networks, in which all nodes and edges are of the same type (e.g., a gene-gene interaction network), have already been implemented for integration of multiomic data and hypothesis generation for patient stratification. For example, similarity network fusion is used to create homogeneous networks of patient-patient similarity edges by merging these edges from different omics, 13 whereas network-based statistics generate patient-derived features by propagating data from a patient’s mutations across a gene interaction network and using the resulting patient-feature matrix to generate subtypes. 14 More recently, homogeneous network approaches have also been applied to biomarker discovery, for example, by propagating signals from drug targets or disease signatures through protein-protein interaction networks, to identify biomarkers of drug response. 15–17 Unlike homogeneous networks, KGs can be composed of multiple types of nodes and edges. In biomedicine, examples of such edges between two node types are ‘Gene → is expressed in → Patient’ and ‘Patient → responds to → Drug’. KGs can be extended beyond homogeneous networks by capturing heterogeneous data and relationships between entities in a highly flexible and scalable data structure, and as such their applications to the drug discovery pipeline have been steadily expanding. KGs have been used to rank targets for resistance to EGFRm inhibitors in NSCLC 18 and to capture pharmacological and disease data in frameworks that can be leveraged for drug repurposing. 19,20 In the clinical setting, a recent study used KGs to predict cohort-level response rates in trial arms, 21 using the accompanying graph node embeddings as inputs for asset prioritisation models. The use of KGs to model data at the individual patient level is also still relatively new. Notably, patient data have been encoded in one edge type of a mutation-based KG to capture patient-specific mutation co-occurrences alongside other omic data for the identification of downstream targets. 22 They have also been represented as patient nodes in a KG incorporating patients, genomics, and biomedical prior knowledge that can be used to derive patient features for survival prediction. 23 There is a clear opportunity to combine the ability of KGs to capture individual patient data with their ability to generate multimodal insights and to leverage this function for personalised medicine via a systematic, reusable, generalisable framework. Here we present PRESSnet ( P atient RE commendation via S tratification and S election using net works), a framework for generating hypotheses for patient stratification and biomarker discovery by analysing patient KGs from clinical trial data. The user can choose which trial modality files to include from their dataset (e.g., clinical and genomic features) and use their choice of graph tool to analyse the data. PRESSnet then automatedly creates a KG of the input data in which nodes represent patients, their associated features, and, optionally, their clinical outcomes. In addition, biomedical prior knowledge—for example, gene-pathway or gene-gene relationship data—can also be integrated with the graphs. PRESSnet generates insights from the KG via multiple algorithms. As graph algorithms capture interrelationships between nodes, and the graphs contain nodes representing features, PRESSnet can offer biomarker signatures that are ‘composite’, i.e., that are made up of more than one biomarker and can integrate multiple modalities within one signature (e.g., ‘Gene A mutated AND Tumor Mutational Burden low’). We implemented PRESSnet on patient data from two independent patient populations: one in which we assessed the power of this framework in a treatment-specific setting, and the other across a range of treatments and multiple datasets. We demonstrate PRESSnet’s ability to generate explainable patient stratification hypotheses and to suggest a mixture of known and novel potential biomarkers with statistical significance estimation across datasets. Results The PRESSnet framework PRESSnet works end-to-end to convert raw data into graph-derived insights in a customizable and scalable approach (Fig. 1 ). The framework consists of multiple stages: User-defined input parameters: The user can define choices around input data and algorithm parameters before executing the methodology. Automated KG creation: PRESSnet creates a KG containing patient, feature, and (optionally) clinical outcome data directly from the user’s input files. Graph algorithms for patient stratification and biomarker discovery: The user’s choice of graph algorithm (e.g., a graph-embedding algorithm) is applied to the graph for downstream analysis. Robustness assessment: Optionally, robustness assessment (bootstrapping and permutation testing) is applied to the results produced by the algorithm. Generation of patient and feature insights: PRESSnet outputs a set of files containing information such as patient clusters, biomarker information, and clinical outcome statistics. Each patient is represented by a unique node, and a patient node is connected to nodes representing the features that the patient possesses, creating a network that links patients through shared features. For continuous data, a node is created corresponding to categories (e.g., tumor mutational burden [TMB] score > 10, TMB score ≤ 10; see Methods: KG creation). PRESSnet offers two types of KG analysis: supervised and unsupervised. Both approaches can be used for patient stratification. Supervised analysis incorporates patients’ clinical outcome data in the graph, and this can also be leveraged for biomarker discovery by analysing the connectivity between outcome nodes and feature nodes via shared patient nodes. Unsupervised analysis excludes this outcome data and therefore provides more flexibility for integrating datasets that include treatment or outcome data with those that do not for applications such as genomics-driven subtyping. Including both supervised and unsupervised approaches in one framework provides the user with the advantages of both. As referenced above, since features are themselves nodes, we can generate hypotheses for not just how patients stratify or cluster, but also how specific sets of features are associated with each other, either with respect to survival or not. PRESSnet also serves as an importable Python library for custom analyses. Users can create graphs directly from data frames and can benefit from ‘bring your own graph’ (BYOG) approaches, whereby they can apply its stratification or biomarker discovery functions to pre-made graphs containing patient and feature nodes. This is useful when the user has already assembled a graph or has a custom graph that cannot be assembled in one step from raw data files. A key benefit of KGs for preclinical or clinical data is the ability to connect gene data to prior knowledge through relationships such as gene → pathway edges. PRESSnet offers the ability to integrate these prior knowledge nodes with the gene nodes in the trial KG (Supplementary Fig. 1; see Methods: KG creation). The algorithms in PRESSnet can be used for patient stratification and biomarker discovery applications. For patient stratification, community detection algorithms separate patients immediately into communities, and PRESSnet provides the user with the option to generate the most statistically significantly enriched features of each community; this offers explainability by revealing the features driving the stratification hypothesis. Graph embeddings (e.g., from random walk–based algorithms) can be used to cluster patients in downstream analysis, using tools like UMAP or TSNE with DBSCAN, K-means, etc. (see Methods: Graph algorithms for patient stratification and biomarker discovery). For biomarker discovery, both community detection and graph embedding–led approaches can generate biomarkers of response or survival; they do so via the supervised approach, whereby the connectivity of the nodes for these endpoints is leveraged to generate insights (e.g., biomarkers of response ending up in communities with the ‘responder’ node). For biomarker discovery, PRESSnet also offers an adapted version of personalised PageRank, a link analysis algorithm which in this case measures the importance of feature nodes in the graph relative to endpoint nodes (see Methods: Graph algorithms for patient stratification and biomarker discovery). For any analysis framework working on relatively small datasets, maximising robustness in the face of potential overfitting and selection bias is highly desirable. As such, the PRESSnet workflow incorporates methods to address the robustness of its recommendations through the PageRank and community detection workflows. The user can choose whether to apply bootstrapping and/or permutation testing, in which the underlying graph is respectively resampled or shuffled (see Methods: In-built meta-analysis and robustness assessment). Benchmarks for statistically significant biomarker discovery and patient stratification For benchmarking, we tested PRESSnet on a patient dataset containing widely established (‘known’) biomarkers of overall survival (OS). The 2022 Memorial Sloan Kettering (MSK) IO TMB dataset 24 is a public dataset of patients with stage IV NSCLC treated with immune checkpoint inhibitors. It contains survival, response, and treatment data for each patient, as well as genomic and clinical features that include multiple widely published biomarkers for IO-treated NSCLC. We conducted an analysis of the overall dataset (240 patients with mutations and clinical data) and also carried out a subpopulation-specific analysis by splitting the data for treatment type (e.g., anti–programmed cell death-1[PD-1] vs. anti–CTLA-4 therapy) and cancer subtype (lung adenocarcinoma [LUAD] vs. lung squamous-cell carcinoma [LUSC]); from the latter, the only subpopulations large enough for meaningful analysis were anti-PD-1–treated LUSC ( n = 27) and anti-PD-1–treated LUAD ( n = 168) (Supplementary Table 1). Within the latter subpopulation, we examined whether PRESSnet could meaningfully stratify patients by clinical outcome and uncover biomarkers of OS. We implemented PRESSnet in a supervised setting for patient stratification using both direct, one-pass community detection on the whole patient graph and a metaclustering of community detection runs across bootstrapped graphs (Fig. 2 ). For both approaches, PRESSnet generated communities of patients with clearly distinguished survival trajectories (Fig. 3 ) and associated biomarkers. We also implemented the same type of analysis in an unsupervised setting with survival data hidden. In both the direct and metaclustering-based unsupervised approaches, PRESSnet was able to generate two distinct communities of patients with statistically significant survival differentiation. A visualisation and statistics for subtypes produced by PRESSnet versus benchmark methods for stratification including MOFA + is shown in Supplementary Fig. 2. Compared to these methods, PRESSnet displays superior cluster separation by silhouette and C-H score (e.g. unsupervised C-H score of 20.043 for PRESSnet’s metaclustering approach v.s. 6.156 for MOFA + with K-Means) and evidences a better balance between patient survival differentiation and practical, plausible clustering in a supervised setting (e.g. 2 clusters with a C-H score of 9.404 and concordance index of 0.659 for PRESSnet’s metaclustering approach, v.s. 9 clusters for MOFA + with a concordance index of 0.915 but a C-H score of 2.367). We next generated a final list of biomarker hypotheses for the LUAD population, using the process visualised in Fig. 2 . The biomarkers associated with the stratifications from the community detection analysis described above were combined with biomarkers from a separate PageRank-based analysis that also used bootstrapping. These biomarkers included known markers of poor prognosis in immune checkpoint inhibitor–treated NSCLC, such as STK11 and KEAP1 driver mutation, TMB, and PD-L1 status, 25,26 that were consistent across bootstrapping and passed permutation testing. Examples of composite biomarkers in the LUAD population from the bootstrapping approaches are shown in Fig. 3 . A comparison of the overlap in the LUAD population between PRESSnet-derived univariate markers, a benchmark machine learning (ML)–based biomarker discovery technique (see Methods: Benchmarks for patient stratification and biomarker discovery on MSK 2022 LUAD data), and the statistically top biomarkers from log-rank testing or univariate Cox modelling are shown in Fig. 4 a. PRESSnet outperformed the ML benchmark method for retrieving statistically significant univariate markers according to log-rank testing or Cox modelling. We assessed the PRESSnet-derived signature biomarkers against univariate features and observed that composite biomarkers were superior to univariate biomarkers for statistically significant patient separation according to multiple metrics (Fig. 4 b and 4 c). We also used PRESSnet to generate biomarker hypotheses in the LUSC population, using bootstrapping and permutation testing. Despite the small patient sample size of 27, PRESSnet was able to generate 58 univariate and composite biomarkers of OS that passed the graph-based permutation tests and carried statistical significance according to log-rank testing. An example of a significant composite biomarker in the LUSC population is shown in Supplementary Fig. 3. These results demonstrate the ability of the framework to generate insights in a small dataset. We also assessed the translatability of PRESSnet’s generated biomarkers and model weights by performing a train/test split analysis within the LUAD population. We performed 10 train/test splits of patients into 50% training and 50% test patients ( n = 84 for each category) and assessed whether (1) biomarkers generated from the training set would maintain statistical significance in the test set and (2) PRESSnet could be used to generate meaningful ‘risk scores’ for unseen test set patients. For (1), PRESSnet generated on average 26 univariate and composite biomarkers of OS in each test split that maintained significance; biomarkers that were found to be significant across multiple of the 10 splits included ‘true positives’ such as STK11 mutation, PD-L1–high scoring, and composite signatures involving known markers such as low albumin 27,28 and high neutrophil-to-lymphocyte ratio. 29 The biomarkers found to be significant in the most test sets can be found in Supplementary Table 2; along with Fig. 4 b and 4 c., this demonstrates the utility of composite signatures for producing strong and persistent signal compared to univariate markers. For (2), we fed PRESSnet-derived weights for test patient nodes into a Cox regression model and found that they outperformed equivalent scores and categorisations from a benchmark deep neural network model according to multiple metrics (see Supplementary Table 3). Supplementary Fig. 4 shows an example of Kaplan-Meier (KM) curves for test set patients stratified into high-risk and low-risk groups by PRESSnet, demonstrating a clear and statistically significant inferior survival trajectory for the high-risk group. This highlights the potential of the framework to generate meaningful graph-derived features for assessing survival in unseen patients. Including biomedical prior knowledge for additional IO insights As the treatments in this dataset were IO, we hypothesised that the inclusion of tumor microenvironment (TME)/IO-relevant prior knowledge might increase the power of graph-derived features and generate an extra layer of potential insights. We constructed a TME subgraph using public data (see Methods: Dataset preparation). By linking genes to pathways and the biological processes in which they are involved, we investigated how adding prior knowledge would (1) alter the performance of PRESSnet, e.g., for downstream predictive tasks, and (2) reveal extra context around drivers of OS in the original feature set. For (1), we used PRESSnet to generate embeddings of patients given anti-PD1 therapy, keeping algorithm parameters and train/test splits identical but changing the input data to exclude or include prior knowledge. The resultant patient embeddings underwent downstream classification using XGBoost 30 to predict binary response status across five splits. We observed that in both supervised and unsupervised settings (i.e. with outcome data nodes included or excluded from the graph), prior knowledge could scale to improve the downstream predictive performance of patient embeddings in a small dataset (the LUSC population, n = 27) and a substantially larger one (all patients, n = 240). In the latter, we observed that adding prior knowledge pathway data improved accuracy, F1 score and area under the curve (AUC) of the classifier compared to using the original dataset alone. In the smaller LUSC population, we observed the same pattern in the supervised setting, and while we did not see improvements in performance in the unsupervised setting with pathway inclusion alone, the addition of both pathway and gene-gene relationship produced by far the best performance compared to the other data options in this setting (see Supplementary Table 5). A visualisation of the embeddings can be found in Supplementary Fig. 5. For (2), prior knowledge KG nodes were separated or encoded like the original feature nodes by the algorithms in PRESSnet. Thus, in addition to looking at how biomarkers and patients separated when prior knowledge was included, we also examined what the outputs of PRESSnet revealed about the prior knowledge nodes themselves, by extracting their PageRank weights (Supplementary Fig. 6a). The results indicated the presence of a ‘tail’ at both ends of the curve, where a small selection of nodes were either notably higher or lower weighted (Supplementary Fig. 6b, 6c), suggesting potential association with clinical outcome. In the above-described subset of prior knowledge nodes, we found multiple examples of features with established links to the efficacy of IO (and specifically anti–PD-1 therapy) in the literature. Some nodes did not pass permutation testing with P < 0.05 due to signal attenuation, …. ,; however, these still contained highly relevant nodes such as regulation of leukocyte activation ( P = 0.1356; the contributory genes for this pathway include PD-1 , CTLA-4 , RICTOR , and other proposed or existing NSCLC IO targets), 24,31 activation of B cells ( P = 0.565; levels of B cell activation have been shown to correlate with poor response to anti–PD-1 therapy 32 ), positive regulation of macrophage apoptosis ( P = 0.132; this is a process that can increase the efficacy of IO therapy by reducing immunosuppression 33 ), and osteoblast development ( P = 0.0854; many constituent pathway member genes for this process have been linked to T-cell exhaustion across cancer types 34–36 ). Nodes that passed permutation testing included the node for the vascular endothelial growth factor receptor signaling pathway ( P = 0.00740; inhibition of this pathway alongside anti–PD-1 therapy has been shown to improve anti-tumor response over anti–PD-1 therapy alone 37 ), and positive regulation of reactive oxygen species biosynthetic process ( P = 0.00589; this has been shown to increase the efficacy of anti–PD-1 therapy 38 ). These findings provide evidence that PRESSnet can identify highly context-relevant nodes from prior knowledge data for explaining therapy-specific survival outcome, highlighting its utility for mechanistic biological interpretation and raising the possibility of future applications to diseases where therapy resistance and prognosis is less well understood. Generalisable AML biomarker discovery across TCGA and Beat AML with PRESSnet For biomarker discovery frameworks, the ability to uncover biomarkers that are applicable across datasets (rather than ‘overfit’ on a specific patient cohort) is generally highly desirable. We therefore assessed PRESSnet for generalizability by applying it to two public acute myeloid leukemia (AML) patient datasets, Beat AML 2022 39 and The Cancer Genome Atlas (TCGA). 40 Both datasets contain treatment history, survival data, genomics, transcriptomics, and clinical data for AML patients (Supplementary Fig. 7); as these datasets contain a range of treatment modalities administered to patients across different time periods, we chose to perform a treatment-agnostic analysis, focusing on prognostic biomarker discovery. To explore generalisability, we implemented PRESSnet on both datasets separately and then identified biomarkers that were significant in, and overlapped across, both sets of results. PRESSnet generated 430 univariate and composite biomarkers from analysis of the TCGA dataset alone that turned out to be statistically significant (log-rank test P < 0.05) in both the TCGA dataset and the Beat AML dataset, including 217 with P < 0.001 in at least 1 dataset and 45 with P < 0.001 in both datasets. From analysis of the Beat AML dataset alone, we found that PRESSnet generated 413 biomarkers that were statistically significant in both datasets, including 274 with P < 0.001 in at least 1 dataset, and 26 with P < 0.001 in both datasets. The resulting markers contained a mixture of known AML biomarkers and novel biomarkers (Fig. 5 ). To leverage the scalability of KGs, we also investigated whether combining datasets would increase our ability to find statistically powerful biomarkers that are present across a large patient population. This resulted in a single KG that contained patients from both TCGA and Beat AML, linked by shared features and clinical outcomes. We implemented PRESSnet on the combined graph, using experimental conditions that were identical to those of our individual dataset analysis. This generated 2425 biomarker hypotheses, including 1504 with P < 0.001 and 389 with P < 0.00001 according to log rank testing. Two examples of resulting composite biomarker hypotheses are shown in Fig. 5 b. We also identified signatures for which statistical significance only occurred in the combined dataset (an example of which can be found in Supplementary Fig. 8). Together, these results serve as examples of how KGs and PRESSnet have the potential to facilitate cross–dataset analysis for generating novel patient insights. Discussion In the era of personalised medicine, the ability to accelerate data-to-decision timelines, the integration of multimodal data for modelling, and the generation of targeted novel hypotheses are highly desirable. We have developed a generalisable framework for building patient knowledge graphs to identify explainable and translatable stratification and biomarker hypotheses. As a lightweight, customisable, end-to-end framework, PRESSnet empowers users to easily build a graph representation of their patient data and run iterative analyses across different analytical algorithms, data splits, modalities, and thresholds. PRESSnet can identify both univariate and composite biomarkers of outcomes in patient populations and can perform unsupervised or supervised analysis to drive patient separation by multiomic features or clinical outcomes. It can also integrate context from prior knowledge such as pathways and causal gene networks that can improve model performance and provide additional biologically grounded and explainable insights beyond the original dataset that can be directly interrogated experimentally, further eliminating the problem with interpreting context in the predictions from black-box ML models. We applied PRESSnet to multiple independent datasets, demonstrating its ability to find known, novel, and generalisable biomarkers across multiple datasets, data modalities, and indications. These included biomarkers of poor prognosis for anti–PD-1 treatment in NSCLC patients and generalisable prognostic markers of improved outcome in AML patients. We also showed the power of composite biomarkers that PRESSnet offers for outperforming univariate biomarkers for separating patients by survival and translating to unseen patients. PRESSnet is published as an adaptable framework for the scientific community to build and analyse KGs of patient data for their own use cases. There is scope to evolve PRESSnet further to meet the challenges of modelling multiomic and clinical data. Representing patient feature data in graphs is not a solved problem—for example, deliberately pursuing sparsity (e.g., by including only one biomarker per underlying variable, or removing ‘noisy’ edges for which the density of connections exceeds a certain threshold) might offer performance benefits for certain datasets—but we regard a systematic exploration of this issue as requiring a separate, future analysis. For our purposes, the common practice for biomarker discovery is to define a biomarker categorically (e.g., ‘TMB low’, ‘PD-L1 negative’), and as such, we developed PRESSnet to capture biomarkers in this way. However, the distribution of a given variable is often not conducive to straightforward categorisation (prompting many discussions about so-called ‘dichotomania’ 41 ). This potentially results in a loss of expressiveness once thresholds are defined. Additionally, the notion of ‘high’ or ‘low’ might vary greatly across contexts. PRESSnet provides the flexibility to apply thresholds in various ways during the pipeline, and users can also apply thresholds to their data before passing it through the framework to allow for column-specific data decisions. Another angle from which to approach this issue is to recommend variables rather than thresholded biomarkers and to encode continuous data in our biomarker discovery models. Graph neural networks (GNNs) can deal with weighted graphs that capture continuous values in nodes or edges. This raises the challenge of reconciling heterogeneous weighted edge types, but advances in heterogeneous GNNs, 42 particularly in the past several years, have facilitated the analysis of such data and enabled applications for multiomic data. 43 Future versions of PRESSnet explore the incorporation of GNNs for biomarker discovery via explainability mechanisms such as graph attention or retrofitted GNN explainers. PRESSnet itself can also be extended to a wider range of applications than OS or treatment response. The framework can be easily adapted to uncover signatures of readouts such as adverse events for applications to patient safety, for example by representing the events as nodes in the KG and looking at features that co-occur with them in communities. For trial design, PRESSnet can also find markers of features such as HER2 or PD-L1 status. This would be useful in settings where such readouts are not available for some patients: generating a novel expression signature associated with HER2 status from immunohistochemistry imaging, for example, would potentially enable inference of HER2 status in patients who have not had this imaging carried out. PRESSnet can also be applied to early discovery, for example preclinical target identification. Nodes representing cell lines, organoids or xenograft models can easily replace patient nodes in the workflow, replicating the same KG structure (e.g., a cell line node linked to nodes for mutations present in the cell line). This could be useful for deriving preclinical signatures of response or efficacy that can then be translated to a clinical setting. Conversely, uncovering druggable resistance markers for a given therapy can provide novel combination hypotheses or can be back-translated to inform target recommendations preclinical experimental validation. PRESSnet’s flexibility and lack of computational intensity make it particularly suited for incorporation into a ‘lab-in-the-loop’ setting, where feedback from experiments can be used to adjust feature processing, model architecture, causal prior knowledge, and robustness constraints for subsequent rounds of hypothesis generation. A very important next step for PRESSnet will be leveraging opportunities offered by the rise of ‘Agentic AI’ and Foundation Models. We foresee PRESSnet benefitting greatly from a multi-agent framework that supports multiple parts of its analysis framework: a ‘setup agent’ can handle parameter configuration and feature selection, for example, and an ‘evidence agent’ can provide context from literature or other supporting databases for PRESSnet’s recommendations. Large language models serve as natural tools for intuitively querying KGs, enriching KG data, and providing semantic understanding of KGs, and as such, they could be integrated with PRESSnet to make generated KGs more accessible to non–data scientists and offer additional insights into the data. Knowledge derived from Foundation Models, for example, around gene function from GenePT, 44 can be integrated with PRESSnet in the future as part of prior knowledge or feature selection. As PRESSnet is a tool that can assign weights to individual patients and features or generate AI-derived numerical representations of them (e.g., embeddings), its outputs can also be fed into a multi-model framework governed by a central agent for downstream tasks such as survival prediction or target identification. In summary, our work indicates that PRESSnet demonstrates the capability of KG-based frameworks to rapidly analyse complex multiomic clinical datasets towards delivering effective targeted therapies to patients. It both serves as a touchstone for further research in this area and offers opportunities for extension to new domains in the future. Methods Running PRESSnet PRESSnet can work as either an importable library or a command line tool in Python, in which the user specifies a script corresponding to the type of analysis they wish to perform, followed by a file path to the .yaml file they are using to configure the run. For example, if the user wished to run the embeddings generation pipeline using parameters from a .yaml file called test_run.yaml , they would execute the following command: python pressnet_embeddings_generation.py -y test_run.yaml The script runs end-to-end without any further input needed from the user. Print commands during the running of the script update the user on the progress of the analysis and provide relevant statistics such as survival rates of communities discovered by community detection. PRESSnet is computationally inexpensive to run and practical in terms of time consumption. For example, generating communities from the graph of MSK LUAD patients (total edges = 70,488) took 1.3 seconds when run on a 2020 MacBook Pro with a 2-GHz Quad-Core Intel Core i5 processor, and applying adapted Personalised PageRank to the graph of merged Beat AML and TCGA patients (totalling 646,609 edges) required 4 minutes and 14 seconds to execute on 50 bootstrapped samples and 3 minutes and 45 seconds to execute on 50 shuffled graphs for permutation testing on the same device. Extracting and aggregating log-rank test statistics across n samples of the data (see Methods: In-built meta-analysis and robustness assessment) is the most time-intensive step (e.g., 43 minutes and 50 seconds for 2450 biomarkers evaluated across 50 samples of 615 patients from the combined AML dataset). This is a parameter that can be customised by the user (using the variable n _resamples_composite_assessment in the .yaml) or switched off entirely, depending on the user’s task requirements. All outputs described in Results were generated by running PRESSnet on CPUs, and no pretraining is required for any of the algorithms. Creation of KGs PRESSnet uses a straightforward, customizable workflow to create a KG directly from a .yaml file (or a Python dictionary corresponding to such a file). The file sections for graph creation can be summarized as follows: File definitions: The user specifies the paths to raw files they want to include in the graph, as well as the location to save output files. Column definitions: The user specifies the name of the patient identifier column, the endpoints column for survival data, if applicable, and the columns to use to split the dataset; they also specify suffixes to distinguish shared underlying feature names (e.g., KRAS mutation vs. KRAS expression) and define continuous columns for thresholding and the thresholding technique to use. Features with a shared underlying entity (e.g., KRAS for KRAS mutation and KRAS expression) are not linked by default, as interrelationships, such as between expression level and mutation status, are not automatically inferable from underlying data. Inclusion of outcome data: The user specifies whether to include nodes for binary clinical outcome and categorised survival duration data in the graph, for supervised or unsupervised analysis. Prior knowledge: The user toggles whether to include prior knowledge and specifies the location of this data if applicable. The first set of inputs allow the user to (1) specify the files they wish to use to create the graph and (2) provide feature annotations for these. The latter are on a per-file basis: for example, if a dataset has copy number, mutations, and expression data for the gene CD274 , in order to distinguish each modality in the final graph (rather than having one duplicated ‘CD274’ node), the user might specify suffixes such that the features ‘CD274_exp’, ‘CD274_CNA’, and ‘CD274_mut’ are created, and a graph node is subsequently created for each of these. Each file is read into a pandas data frame in Python, and the eventual group of data frames are merged on a specified patient identifier column. A graph is then constructed out of combined data. Once column naming is defined, the user can also indicate columns to be used to split the dataset. For example, the user may want to conduct per-indication or per-treatment analyses within a trial, allowing for the discovery of predictive biomarkers. A separate KG will therefore be constructed for the patients within each split. Finally, the user can also specify how to apply thresholds to continuous data for biomarker creation. For continuous data that have been categorized, a node is created corresponding to each of the categories (e.g., TMB score > 10, TMB score ≤ 10). There are multiple options for how to apply thresholds: Median cutoff categorises each biomarker as greater than or less than equal to the median of its values. Mean cutoff categorises each biomarker as greater than or less than equal to the mean of its values. Quantile cutoff categorises each biomarker into ‘top quantile’, ‘bottom quantile’, and ‘between quantiles’ based on its values. Youden’s J-statistic-based cutoff uses Youden’s J statistic based on the receiving operator characteristic curve of the feature versus the endpoint event to determine the optimal cutoff from which to categorise the feature. Log-rank test-based cutoff thresholds are applied across the range of values of the feature and performs iterative log-rank testing of the ‘feature-high’ versus ‘feature-low’ subpopulations at each threshold to determine the optimal cutoff for the feature to generate the greatest survival separation. The user has the option to use a ‘fast’ sliding window approach, which speeds up implementation of this approach for a slight loss of granularity. An initial set of uniformly spaced values (default n = 10) are explored, and from these, a neighbouring pair is extracted for which the log-rank separation was greatest. The final cutoff is determined by iterating across the values that fall within the value range of the pair. PRESSnet offers two types of KG analysis: supervised and unsupervised. Supervised analysis is used for biomarker discovery pipelines (see Methods: Graph algorithms for patient stratification and biomarker discovery). In a supervised analysis, nodes corresponding to outcomes (e.g., ‘responder’, ‘survival status: alive’) are included in the graph, just like feature nodes. As well as binary clinical outcome, the user may be interested in looking for biomarkers that reflect progression of disease related to time, i.e., biomarkers of ‘fast progressors’ or ‘long survivors’. Duration nodes can be included in the KG, both in terms of quantile (‘1st quantile duration’, ‘2nd quantile duration’, etc.) and in terms of the concept of progression (‘fast progressor’, ‘intermediate group’, ‘long survivor’). The user can define the time thresholds for categorising the latter. For example, if the fast progressor threshold is set at 60 days, any patient with a survival duration less than or equal to that value will be connected to a ‘fast progressor’ node in the graph. The shared connections between these feature nodes and these outcome nodes are leveraged to find biomarkers. Supervised analysis can also be used for stratification: the dense network connections between outcome nodes and patient nodes should cause the separation of patients to be influenced by these outcomes, leading to better stratification than the unsupervised approach. Unsupervised analysis prioritizes the features themselves driving separation, as the outcome data are not converted into graph edges and are excluded from any analysis. This type of analysis also offers more scope for integrating patient cohorts from different datasets in which treatments or endpoint types might differ. Represented as ‘gene → entity’ graph edges (e.g., gene → associates → disease), biomedical prior knowledge can be incorporated with a trial KG. This feature may be highly relevant to the dataset at hand. For example, if a gene G is linked to an immune-related biological process B from Gene Ontology annotations, 45,46 patients that express G might be less receptive to IO. If the user has a file of gene → biological process graph edges, they can specify to add this file to their KG, such that a patient P who is linked to a gene G through a gene expression relationship will now be connected at one extra ‘hop’ away to the immune process B. This will potentially provide better stratification; moreover, the node for B may itself represent a form of biomarker and co-occur in communities with nodes corresponding to poor prognosis. Graphs are constructed directly from data frames using the networkx library in Python and can easily be converted into a Pytorch Geometric data object for use with certain embedding algorithms (see below). Graph algorithms for patient stratification and biomarker discovery Once a KG is created, a variety of algorithms are available to users to leverage the graph according to the intended task. These can be categorised as community detection algorithms, embedding generation algorithms, and personalised PageRank. There is a unique script for each type of algorithm: pressnet_embeddings_generation.py , pressnet_community_detection.py , and pressnet_pagerank.py . The user can choose the specific algorithm within these in the .yaml file and can edit certain parameters for each algorithm, such as learning rate or resolution (a parameter specific to Louvain community detection that affects the size and number of communities found). Community detection is the process by which graph nodes are separated into groups using the topology of the graph itself, typically with the aim of these groups being tightly connected. At present, PRESSnet offers two community detection techniques that work in different ways: Louvain 47 tries to optimise a modularity score that represents the density of edges within a community as opposed to outside it. This algorithm was implemented using the networkx library in Python. EdMOT 48 samples sequences of neighbouring nodes using random walks and minimises a neighbourhood observation loss in the construction of the embeddings. This algorithm was implemented using the karateclub library in Python. After running a community detection process, the user receives a list of node identifiers (including feature nodes and, optionally, endpoint nodes) with their corresponding community number. PRESSnet also provides information for the user about each community, including statistics and KM curves for survival if specified and, optionally, biomarkers associated with each community. The latter is measured by looking at the difference between the proportion of patients with the signature in the community versus the proportion of those in other communities. There is a parameter that allows users to set the minimum number of patient nodes for a given cluster; if a very small cluster is produced with patient n below this, it will be merged with the next-biggest cluster that is most similar by graph connectivity. For biomarker discovery, PRESSnet leverages the fact that communities can contain any type of node. As such, a node for any biomarker B with a strong association with a clinical outcome O (e.g., ‘survival status: deceased’) will end up in a community containing the node for O (Fig. 6 a). This community will also contain nodes for patients P1 , P2 .. . Pn who exhibit O and B , as they are connected to these nodes. As graph nodes with similar connectivity should end up in the same community, biomarkers that co-occur with B in patients should also be found in B ’s community. PRESSnet therefore searches the communities containing positive and negative endpoint nodes to find biomarkers of those endpoints and performs pairwise searches to identify composite biomarkers from these. A graph node embedding is a numerical vector representation of a node that captures information about its connectivity in the graph and, optionally, its encoded numerical data. Embeddings are useful for a variety of functions: they can be clustered like raw features to find patterns in data, used in search tools to find related entities to a given input, or fed into downstream predictive algorithms for classification or regression analysis. Over the past several years, two methods have become predominant for generating KG embeddings: random walks and matrix factorisation. In walk-based algorithms, random walks through a graph create sequences of nodes, and these sequences are used to create embeddings that capture neighbourhood information and the overall graph structure. PRESSnet currently offers the user multiple random walk–based algorithms, details of which can be found below. DeepWalk, Node2Vec, and Walklets are implemented using the karateclub library and are initialised with the default parameters of n_walks = 40 , walk_length = 40 , but these can be altered by the user. DeepWalk 49 leverages the principles behind the Skip-Gram algorithm (originally popularised for learning word embeddings in natural language processing) for the graph domain: treating random walks as sequences of nodes, it tries to maximise the log probability of neighbouring nodes appearing next to a given node using embeddings. Node2Vec 50 is an extension of DeepWalk that uses negative sampling and leverages search algorithms to extract random walks. Walklets 51 subsamples short walks using ‘skips’ within longer walks to capture higher-order relationships between nodes. MetaPath2Vec 52 defines walks by node type, in a so-called ‘metapath’; the metapath in PRESSnet consists of [Outcome → Patient → Feature → Patient → Outcome], so an example walk might consist of ‘survival status: deceased’ → ‘Patient P1’ → ‘ KRAS mutation’ → ‘Patient P2’ → ‘survival status: alive’.” The default parameters for MetaPath2Vec in PRESSnet are n _walks = 40, walk_length = 40, n _epochs = 100, learning rate = 0.1, but these can be altered by the user. Matrix factorisation techniques achieve embeddings that capture graph and neighbourhood information by decomposing matrices from the KG (typically, the Laplacian or adjacency matrix) to derive the embeddings. They typically use negative sampling as well as true relations to learn correct embeddings. PRESSnet offers matrix factorisation algorithms implemented in the Pytorch Geometric library, with default parameters of n _epochs = 50 and learning rate = 0.005, which can be altered by the user. DistMult 53 is a bilinear model in which a relationship between two entities is modelled as a diagonal matrix and the vectors for those entities are multiplied with this matrix element-wise to produce a representation of the overall entity-entity interaction. This is then aggregated to output a scalar score representing the likelihood of that relationship existing or not. ComplEx 54 is also a bilinear model that uses dot product multiplication, but unlike DistMult, which leverages real-valued embeddings, ComplEx uses complex-valued embeddings and is geared more suitably for asymmetric relationships (e.g., gene-gene regulatory data). TransE 55 is a so-called translational distance model that encodes entities and relations in vectors in a shared target semantic space and uses a distance-based scoring function within this space to generate embeddings. The underlying principle is that addition of the vector of the source entity to that of the relation itself should be able to approximate the vector of the target entity. For patient stratification, patient embeddings from PRESSnet can be clustered to generate potential subtypes. As embeddings are generated for every node in the graph, distance-based analysis of these embeddings is used to identify biomarkers. Nodes with similar connectivity should have numerically similar embeddings. Returning to the example of biomarker B associated with outcome O , this means that the node for B should be closer in embedding space to the node for O than another feature, F , which has no association with survival. Similarly, the node for B should theoretically be further from the node that represents the converse of O (i.e., ‘survival status: alive’) than F . We can thus frame biomarker discovery as a multi-objective optimisation problem in embedding space: for biomarkers of poor prognosis in terms of binary overall survival, for example, we are simultaneously minimising the distance to the embedding for ‘deceased’ and maximising the distance to the embedding for ‘alive’ (Fig. 6 b). PRESSnet also offers an adaptation of personalised PageRank to generate biomarkers (Fig. 6 c). Node values for outcome and duration nodes are given weights corresponding to their directionality; nodes corresponding to longer survival and duration are weighted positively, the converse for negative outcome, and feature nodes and patient nodes are comparatively unweighted. After the adapted personalised PageRank is computed, the feature nodes with the largest positive weights are considered as potential biomarkers of good prognosis and vice versa for features with the most negative weights. Outcome predictions for unseen patients As PRESSnet’s PageRank weights capture associations between features and endpoints such as survival, we can use these weights to generate risk scores for ‘unseen’ patients. Training set patients (for whom survival information is known) are connected to test set patients via shared features; survival information for the training set is propagated through the graph to generate PageRank weights associated with survival outcome in the test set patients (Supplementary Fig. 9). We fed these weights into a Cox regression model to generate survival association metrics. We also repeated the process separately for binarized versions of these weights by categorising the PageRank weights into ‘high risk’ and ‘low risk’ by using K-means clustering with n = 2 clusters. Benchmarks for patient stratification and biomarker discovery on MSK 2022 LUAD data We assessed PRESSnet against benchmark methods for subtyping and biomarker discovery on the MSK 2022 dataset. For patient stratification, we implemented K-means clustering on (1) raw tabular data and (2) latent factors that had been discovered by MOFA+. This allowed for a direct comparison with the graph methods in supervised settings (i.e., where outcome data were included in the tabular data) and unsupervised settings. In unsupervised approaches, optimal K-number was derived by identifying the cluster number corresponding to the highest Calinski-Harabasz (CH) score, whereas in supervised approaches, it was derived by identifying the number corresponding to the highest concordance index. For biomarker discovery benchmarking, we trained a random-forest classifier to predict OS on both the raw and PRESSnet-thresholded data and extracted the features that were most important for the prediction to serve as biomarker hypotheses. For risk prediction in unseen patients, we benchmarked PRESSnet against a deep neural network (an MLPClassifier from scikit-learn, with random_state = 1 and 500 max iterations) trained on raw data to predict binary survival outcome. Across 10 train/test splits with 50% training size, we generated c-index and hazard ratio P values from Cox proportional hazards models, which were implemented using the Lifelines library in Python 56 with no parameter tuning, fitted on the following: Raw PageRank weights for test set patients only Binarised PageRank weights (‘high risk’ vs. ‘low risk’ categorisation using k-means clustering with n = 2) for test set patients only Raw prediction scores for binary OS from the deep neural network Binarised prediction (output > 0.5) scores for OS binary OS from the deep neural network We took the average of these c-indices and P values across the 10 iterations to generate our metrics in the Results. In-built meta-analysis and robustness assessment In PRESSnet, the user can choose whether to output results that use both, either, or neither bootstrapping and permutation testing when generating results from PageRank and community detection. We experimented with applying these to the embedding generation modules as well, but found that the time considerations involved were impractical given the training time of the complex embedding algorithms themselves. Addressing this issue can be considered a future extension of the framework alongside those mentioned in the Discussion. A summary of how robustness is implemented for PageRank is presented in Supplementary Fig. 10. The user can choose to implement bootstrapping before running PageRank, at which point the PageRank algorithm is run on a user-specified number of bootstrapped graphs. The weights for each node across the different PageRanks are then averaged to produce an aggregated final weight, while statistics for the variance of these weights are used as a robustness flag to highlight those that fluctuate significantly across runs. The other robustness assessment is via permutation testing. Here, PageRank is run on a user-specified number of permuted graphs; for each node, the weights produced by the bootstrapping analysis are assessed against all the weights from the permuted runs to determine whether we can reject the null hypothesis that the original weights fall within the distribution of the permuted (i.e., randomly generated) weights. Features that fail permutation testing and/or fall outside the acceptable variance of weights are then filtered out before biomarker selection. When biomarkers are combined for composite signatures, a final set of log-rank testing is performed on multiple random samples of the data to assess the stability of significant signatures; the user can set the sampling proportion and number of samples. PRESSnet offers the users multiple methods for performing bootstrapping and permutation testing. The user can resample from (1) all nodes in the graph, (2) all edges in the graph, or (3) patient nodes only in the graph to produce a bootstrapped graph. For permutation testing, the user can randomly shuffle the identity of the target nodes from either (1) all the edges in the graph or (2) only the patient endpoint edges. Examples of how bootstrapped and permuted graphs work with PageRank are shown in Supplementary Fig. 10. For community detection, variability in results can come from both the composition of the graph and the model parameters, particularly the initialisation (i.e., seed or random state) of the chosen algorithm. As such, if the user chooses to enable robustness assessment, community detection is run for both (1) the original graph n times with the seed initialisation ranging from 1 to n , and (2) the n bootstrapped graphs, where the algorithm is run once per graph with a fixed seed. For stratification, the resulting 2 n sets of communities are concatenated and one-hot encoded to generate inputs for a final metaclustering, which is carried out using hierarchical clustering implemented in scikit-learn. The optimal number of communities is assessed using a weighted Calinski-Harabasz (CH) score approach, 0.7 * CH1 + 0.3 * CH2 , where CH1 is calculated by comparing cluster labels to the underlying patient data, and CH2 by comparing them to the 2n sets of community labels. This attempts to balance biological meaningfulness with algorithmic stability. For biomarker discovery, the 2 n clusterings capture (1) the features that most frequently co-occur in communities with endpoint nodes across the iterations and (2) the features that most frequently co-occur with the features from (1) across the iterations. As such, the features from (1) are evaluated as potential individual biomarkers of the endpoint, and for each feature, a user-defined number of nearest neighbours (default = 10) are considered alongside the feature as potential paired composite signatures. Dataset preparation The MSK 2022 data were downloaded from the publicly accessible cBioPortal database. 57 The TME prior knowledge graph that we built to supplement the original data was also constructed using public data and consists of the following edges: (1) TME-specific ‘Biological Process’ nodes linked to participating genes, curated with HetionetDB, 58 CellPhoneDB, 59 and Gene Ontology; (2) TME-specific cell type nodes linked to genes involved in the cell signature, curated using CellMarker2.0; 60 and (3) gene-to-gene edges modelling gene interactions between wild-type gene nodes using HetionetDB and CellPhoneDB. We also downloaded the TCGA and Beat AML datasets from cBioPortal. Because the gene set for expression data is large (e.g., n = 16383 for TCGA), for preliminary feature selection we identified a set of genes with non-zero confidence scores for links to AML according to Open Targets. 61 We then processed the data to identify patients with complete survival and treatment data, retrieved overlapping features between the two datasets, and used these as inputs for our framework. Expression data were taken as the z-score of the log 2 transcripts per million for each gene to combat batch effects. This resulted in 834 features for expression, 463 for mutations, and 2 for clinical (TMB and age), in 167 patients from TCGA and 603 from Beat AML. For generating results on AML data, we found that age as a feature dominated results in terms of composite biomarker creation (itself a true positive), so we implemented the framework with and without its inclusion, focusing in our Results section on outputs from the latter. Declarations Competing interests J.C.-S., S.S., I.K., D.S., M.G., S.J., D.B., B.S., E.J., and K.B. are or were employees of AstraZeneca when this work was performed and may have stock ownership, options, and/or interests in the company. E.A.C. is an employee of The Bridge and receives a salary from The Bridge, but her work is contracted out to AstraZeneca. Contributions J.C.-S., conceptualization, methodology, software, validation, data curation, visualization, analysis, investigation, writing (original draft, reviewing and editing); S.S., software, validation, data curation, visualisation, analysis, investigation, writing (original draft, reviewing and editing); I.K., conceptualization, methodology, software, investigation; D.S., methodology, investigation; M.G. and S.J., software, investigation; E.A.C., validation, writing (reviewing); D.B., investigation; B.S., writing (reviewing); E.J., and K.B., conceptualization, methodology, supervision, investigation, writing (reviewing and editing). Deborah Shuman (Scientific Publications, Oncology R&D, AstraZeneca, Gaithersburg, MD, USA) provided editorial and graphics support. Corresponding Authors Correspondence to Jake Cohen-Setton. Code Availability The code used in this study is available at 10.5281/zenodo.17100620. Data Availability The specific versions of the datasets used in this study are available publicly at https://www.cbioportal.org/datasets. These are as follows: MSK IO NSCLC: https://www.cbioportal.org/study/summary?id=lung_msk_mind_2020 Beat AML: https://www.cbioportal.org/study/summary?id=aml_ohsu_2022 TCGA AML: https://www.cbioportal.org/study/summary?id=laml_tcga_pan_can_atlas_2018 The cBioPortal website also provides URL links to the corresponding publications, citations for which can be found in References. References Markey, N. et al. Clinical trials are becoming more complex: a machine learning analysis of data from over 16,000 trials. Sci Rep 14 , 3514 (2024). King, E. A., Davis, J. W. & Degner, J. F. Are drug targets with genetic support twice as likely to be approved? Revised estimates of the impact of genetic support for drug mechanisms on the probability of drug approval. PLoS Genet 15 , e1008489 (2019). Morgan, P. et al. Impact of a five-dimensional framework on R&D productivity at AstraZeneca. Nat Rev Drug Discov 17 , 167-181 (2018). Hu, C. & Dignam, J. J. Biomarker-driven oncology clinical trials: key design elements, types, features, and practical considerations. JCO Precis Oncol 3 (2019). Greig, S. L. Osimertinib: first global approval. Drugs 76 , 263-273 (2016). Keam, S. J. Trastuzumab deruxtecan: first approval. Drugs 80 , 501-508 (2020). Garon, E. B. et al. Pembrolizumab for the treatment of non-small-cell lung cancer. N Engl J Med 372 , 2018-2028 (2015). Cortesi, L., Rugo, H. S. & Jackisch, C. An overview of PARP inhibitors for the treatment of breast cancer. Target Oncol 16 , 255-282 (2021). Chang, L., Chang, M., Chang, H. M. & Chang, F. Microsatellite instability: a predictive biomarker for cancer immunotherapy. Appl Immunohistochem Mol Morphol 26 , e15-e21 (2018). Sun, T. H., Wang, C. C., Wu, Y. L., Hsu, K. C. & Lee, T. H. Machine learning approaches for biomarker discovery to predict large-artery atherosclerosis. Sci Rep 13 , 15139 (2023). McLernon, D. J. et al. Assessing performance and clinical usefulness in prediction models with survival outcomes: practical guidance for Cox proportional hazards models. Ann Intern Med 176 , 105-114 (2023). McNamara, M. G. et al. 89PD: prognostic and predictive impact of high tumor mutation burden (TMB) in solid tumors: a systematic review and meta-analysis. Ann Oncol 30 , v25–v54 (2019). Wang, B. et al. Similarity network fusion for aggregating data types on a genomic scale. Nat Methods 11 , 333-337 (2014). Hofree, M., Shen, J. P., Carter, H., Gross, A. & Ideker, T. Network-based stratification of tumor mutations. Nat Methods 10 , 1108-1115 (2013). Kong, J. et al. Network-based machine learning approach to predict immunotherapy response in cancer patients. Nat Commun 13 , 3703 (2022). Shanthamallu, U. S. et al. A network-based framework to discover treatment-response-predicting biomarkers for complex diseases. J Mol Diagn 26 , 917-930 (2024). Lin, C., Sun, G. J., Bulusu, K. C., Dry, J. R. & Hernandez, M. Graph neural networks including sparse interpretability. arXiv (2020). DOI: 10.48550/arXiv.2007.00119. https://arxiv.org/abs/2007.00119. Accessed August 15, 2025. Gogleva, A. et al. Knowledge graph-based recommendation framework identifies drivers of resistance in EGFR mutant non-small cell lung cancer. Nat Commun 13 , 1667 (2022). Chandak, P., Huang, K. & Zitnik, M. Building a knowledge graph to enable precision medicine. Sci Data 10 , 67 (2023). Huang, K. et al. A foundation model for clinician-centered drug repurposing. Nat Med 30 , 3601-3613 (2024). Brbić, M., Yasunaga, M., Agarwal, P. & Leskovec, J. Predicting drug outcome of population via clinical knowledge graph. medRxiv (2024). DOI: 10.1101/2024.03.06.24303800. https://www.medrxiv.org/content/10.1101/2024.03.06.24303800v2. Accessed February 17, 2025. Gorostiola González, M., Ijzerman, A. P. & van Westen, G. J. P. A patient-centric knowledge graph approach to prioritize mutants for selective anti-cancer targeting. bioRxiv (2024). DOI: 10.1101/2024.09.29.615658. https://www.biorxiv.org/content/10.1101/2024.09.29.615658v1. Accessed February 17, 2025. Fang, C. et al. Integrating knowledge graphs into machine learning models for survival prediction and biomarker discovery in patients with non-small-cell lung cancer. J Transl Med 22 , 726 (2024). Vanguri, R. S. et al. Multimodal integration of radiology, pathology and genomics for prediction of response to PD-(L)1 blockade in patients with non-small cell lung cancer. Nat Cancer 3 , 1151-1164 (2022). Khwaja, R. M. & Chu, Q. S. Present and emerging biomarkers in immunotherapy for metastatic non-small cell lung cancer: a review. Curr Oncol 29 , 479-489 (2022). Xiao, Y. et al. Recent progress in targeted therapy for non-small cell lung cancer. Front Pharmacol 14 , 1125547 (2023). Stares, M. et al. Hypoalbuminaemia as a prognostic biomarker of first-line treatment resistance in metastatic non-small cell lung cancer. Front Nutr 8 , 734735 (2021). Yildirim, M., Yildiz, M., Duman, E., Goktas, S. & Kaya, V. Prognostic importance of the nutritional status and systemic inflammatory response in non-small cell lung cancer. J BUON 18 , 728-732 (2013). Yu, Y., Qian, L. & Cui, J. Value of neutrophil-to-lymphocyte ratio for predicting lung cancer prognosis: a meta-analysis of 7,219 patients. Mol Clin Oncol 7 , 498-506 (2017). Chen, T. & Guestrin, C. Xgboost: a scalable tree boosting system. Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, San Franciso, CA, August 13–17, 2016. https://dl.acm.org/doi/10.1145/2939672.2939785. Accessed August 15, 2025. Zhao, D., Jiang, M., Zhang, X. & Hou, H. The role of RICTOR amplification in targeted therapy and drug resistance. Mol Med 26 , 20 (2020). Yuan, S., Liu, Y., Till, B., Song, Y. & Wang, Z. Pretreatment peripheral B cells are associated with tumor response to anti-PD-1-based immunotherapy. Front Immunol 11 , 563653 (2020). Liu, R. et al. ABL1 is a prognostic marker and associated with immune infiltration in hepatocellular carcinoma. J Oncol 2021 , 1379706 (2021). Liu, L. et al. BCL-2 expression promotes immunosuppression in chronic lymphocytic leukemia by enhancing regulatory T cell differentiation and cytotoxic T cell exhaustion. Mol Cancer 21 , 59 (2022). Liu, R. et al. Infiltration of apoptotic M2 macrophage subpopulation is negatively correlated with the immunotherapy response in colorectal cancer. Int J Mol Sci 23 (2022). Zhou, F. et al. CTHRC1 is a prognostic biomarker and correlated with immune infiltrates in kidney renal papillary cell carcinoma and kidney renal clear cell carcinoma. Front Oncol 10 , 570819 (2020). Shigeta, K. et al. Dual programmed death receptor-1 and vascular endothelial growth factor receptor-2 blockade promotes vascular normalization and enhances antitumor immune responses in hepatocellular carcinoma. Hepatology 71 , 1247-1261 (2020). Wang, J., Liu, N., Jiang, H., Li, Q. & Xing, D. Reactive oxygen species in anticancer immunity: a double-edged sword. Front Bioeng Biotechnol 9 , 784612 (2021). Bottomly, D. et al. Integrative analysis of drug response and clinical outcome in acute myeloid leukemia. Cancer Cell 40 , 850-864 e859 (2022). National Cancer Institute & Center for Cancer Genomics. The Cancer Genome Atlas Program (TCGA). https://www.cancer.gov/ccg/research/genome-sequencing/tcga. Accessed February 17, 2025. Senn, S. Dichotomania: an obsessive compulsive disorder that is badly affecting the quality of analysis of pharmaceutical trials. (2005). https://www.semanticscholar.org/paper/Dichotomania%3A-An-Obsessive-Compulsive-Disorder-that-Senn/8c390cb90465ad1f08b86cee52474859e4dffd77. Accessed February 18, 2025. Zhang, C., Song, D., Huang, C., Swami, A. & Chawla, N. V. Heterogeneous graph neural network. 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining, Anchorage, AK, August 4-8, 2019. https://dl.acm.org/doi/10.1145/3292500.3330961. Accessed February 18, 2025. Valous, N. A., Popp, F., Zornig, I., Jager, D. & Charoentong, P. Graph machine learning for integrated multi-omics analysis. Br J Cancer 131 , 205-211 (2024). Chen, Y. & Zou, J. GenePT: a simple but effective foundation model for genes and cells built from ChatGPT. bioRxiv (2024). Ashburner, M. et al. Gene ontology: tool for the unification of biology. The Gene Ontology Consortium. Nat Genet 25 , 25-29 (2000). Gene Ontology, C. et al. The Gene Ontology knowledgebase in 2023. Genetics 224 (2023). De Meo, P., Ferrara, E., Fiumara, G. & Provetti, A. Generalized Louvain method for community detection in large networks. 11th International Conference on Intelligent Systems Design and Applications, November 22-24, 2011. https://ieeexplore.ieee.org/document/6121636/authors#authors. Accessed February 17, 2025. Li, P.-Z., Huang, L., Wang, C.-D. & Lai, J.-H. EdMot: an edge enhancement approach for motif-aware community detection. (2019). DOI: https://arxiv.org/abs/1906.04560. https://arxiv.org/abs/1906.04560. Accessed February 18, 2025. Perozzi, B., Al-Rfou, R. & Skiena, S. DeepWalk: online learning of social representations. 20th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, New York, NY, August 24-27, 2014. https://dl.acm.org/doi/10.1145/2623330.2623732. Accessed February 18, 2025. Grover, A. & Leskovec, J. node2vec: scalable feature learning for networks. 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, San Francisco, CA, August 13-17, 2016. https://dl.acm.org/doi/10.1145/2939672.2939754. Accessed February 18, 2025. Perozzi, B., Kulkarni, V., Chen, H. & Skiena, S. Don't walk, skip! Online learning of multi-scale network embeddings. ASONAM '17: Advances in Social Networks Analysis and Mining, Sydney, Australia, July 31-August 3, 2017. https://dl.acm.org/doi/10.1145/3110025.3110086. Accessed February 18, 2025. Dong, Y., Chawla, N. V. & Swami, A. metapath2vec: Scalable representation learning for heterogeneous networks. 23rd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, Halifax, NS, Canada, August 13-17, 2017. https://dl.acm.org/doi/10.1145/3097983.3098036. Accessed February 18, 2025. Yang, B., Yih, W., He, X., Gao, J. & Deng, L. Embedding entities and relations for learning and inference in knowledge bases. arXiv (2015). DOI: 10.48550/arXiv.1412.6575. https://arxiv.org/abs/1412.6575. Accessed Febuary 18, 2025. Trouillon, T., Welbl, J., Riedel, S., Gaussier, E. & Bouchard, G. Complex embeddings for simple link prediction. Proceedings of the 33rd International Conference on Machine Learning, New York, NY. https://proceedings.mlr.press/v48/trouillon16.html. Accessed February 18, 2025. Bordes, A., Usunier, N., Garcia-Duran, A., Weston, J. & Yakhnenko, O. Translating embeddings for modeling multi-relational data. In Advances in Neural Information Processing Systems 26. (eds C. J. Burges et al. ). Davidson-Pilon, D. Lifelines: survival analysis in Python. J Open Source Softw 4 , 1317 (2019). Cerami, E. et al. The cBio cancer genomics portal: an open platform for exploring multidimensional cancer genomics data. Cancer Discov 2 , 401-404 (2012). Himmelstein, D. S. et al. Systematic integration of biomedical knowledge prioritizes drugs for repurposing. Elife 6 , e26726 (2017). Efremova, M., Vento-Tormo, M., Teichmann, S. A. & Vento-Tormo, R. CellPhoneDB: inferring cell-cell communication from combined expression of multi-subunit ligand-receptor complexes. Nat Protoc 15 , 1484-1506 (2020). Hu, C. et al. CellMarker 2.0: an updated database of manually curated cell markers in human/mouse and web tools based on scRNA-seq data. Nucleic Acids Res 51 , D870-D876 (2023). Ochoa, D. et al. The next-generation Open Targets Platform: reimagined, redesigned, rebuilt. Nucleic Acids Res 51 , D1353-D1359 (2023). Zhang, J. et al. High expression levels of SMAD3 and SMAD7 at diagnosis predict poor prognosis in acute myeloid leukemia patients undergoing chemotherapy. Cancer Gene Ther 26 , 119-127 (2019). Chen, J. et al. Homoharringtonine targets Smad3 and TGF-beta pathway to inhibit the proliferation of acute myeloid leukemia cells. Oncotarget 8 , 40318-40326 (2017). Zhang, Y. et al. Galectin-9 and PSMB8 overexpression predict unfavorable prognosis in patients with AML. J Cancer 12 , 4257-4263 (2021). Park, S. et al. Calreticulin mRNA expression and clinicopathological characteristics in acute myeloid leukemia. Cancer Genet 208 , 630-635 (2015). Aly, M. et al. Distinct clinical and biological implications of CUX1 in myeloid neoplasms. Blood Adv 3 , 2164-2178 (2019). Additional Declarations There is NO Competing Interest. Supplementary Files CohenSettonetal.Supplementary.docx Supplementary Figures and Tables Cite Share Download PDF Status: Under Review Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-8289747","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":559489389,"identity":"cf06c089-0752-488a-a660-3b2fd3d522d6","order_by":0,"name":"Jake Cohen-Setton","email":"data:image/png;base64,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","orcid":"https://orcid.org/0009-0001-9585-1320","institution":"AstraZeneca","correspondingAuthor":true,"prefix":"","firstName":"Jake","middleName":"","lastName":"Cohen-Setton","suffix":""},{"id":559489390,"identity":"76258660-93ba-413b-b193-d2b8e4e641cb","order_by":1,"name":"Shruti Shikhare","email":"","orcid":"","institution":"AstraZeneca","correspondingAuthor":false,"prefix":"","firstName":"Shruti","middleName":"","lastName":"Shikhare","suffix":""},{"id":559489392,"identity":"7c19915e-224c-44ed-8f2b-19c7db9ec4c4","order_by":2,"name":"Ioannis Kagiampakis","email":"","orcid":"","institution":"AstraZeneca","correspondingAuthor":false,"prefix":"","firstName":"Ioannis","middleName":"","lastName":"Kagiampakis","suffix":""},{"id":559489394,"identity":"e7808887-aea3-4de3-afa5-f9c517ca6175","order_by":3,"name":"Domingo Salazar","email":"","orcid":"","institution":"AstraZeneca","correspondingAuthor":false,"prefix":"","firstName":"Domingo","middleName":"","lastName":"Salazar","suffix":""},{"id":559489396,"identity":"6f3b6b51-6509-4b81-91ca-461b2276fef8","order_by":4,"name":"Miguel Goncalves","email":"","orcid":"","institution":"AstraZeneca","correspondingAuthor":false,"prefix":"","firstName":"Miguel","middleName":"","lastName":"Goncalves","suffix":""},{"id":559489397,"identity":"cc2fbf44-441b-465f-aefa-77172406e2e3","order_by":5,"name":"Elizabeth Coker","email":"","orcid":"https://orcid.org/0000-0003-4934-8887","institution":"AstraZeneca","correspondingAuthor":false,"prefix":"","firstName":"Elizabeth","middleName":"","lastName":"Coker","suffix":""},{"id":559489398,"identity":"0e996d81-7f80-4d31-8b3d-7f7a8930064a","order_by":6,"name":"Sanddhya Jayabalan","email":"","orcid":"","institution":"Technical University, Dresden","correspondingAuthor":false,"prefix":"","firstName":"Sanddhya","middleName":"","lastName":"Jayabalan","suffix":""},{"id":559489399,"identity":"500f18f6-bd20-4b8a-b005-7a565bcc23b7","order_by":7,"name":"Damian Bikiel","email":"","orcid":"","institution":"AstraZeneca","correspondingAuthor":false,"prefix":"","firstName":"Damian","middleName":"","lastName":"Bikiel","suffix":""},{"id":559489400,"identity":"185eb84e-99de-4146-bb11-01f4b4c9c5cd","order_by":8,"name":"Ben Sidders","email":"","orcid":"","institution":"Biorelate","correspondingAuthor":false,"prefix":"","firstName":"Ben","middleName":"","lastName":"Sidders","suffix":""},{"id":559489401,"identity":"1fe87c9c-cfb1-45c3-ac92-eb36adbe8141","order_by":9,"name":"Etai Jacob","email":"","orcid":"https://orcid.org/0009-0007-4599-0457","institution":"AstraZeneca","correspondingAuthor":false,"prefix":"","firstName":"Etai","middleName":"","lastName":"Jacob","suffix":""},{"id":559489402,"identity":"8eabca6d-8900-4afb-907f-bad7449cf6b0","order_by":10,"name":"Krishna Bulusu","email":"","orcid":"","institution":"Turbine AI","correspondingAuthor":false,"prefix":"","firstName":"Krishna","middleName":"","lastName":"Bulusu","suffix":""}],"badges":[],"createdAt":"2025-12-05 16:57:17","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-8289747/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-8289747/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":98431685,"identity":"c3f7ebcc-fc37-4c4b-910f-7c1005be0bda","added_by":"auto","created_at":"2025-12-17 16:48:08","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":2398514,"visible":true,"origin":"","legend":"","description":"","filename":"CohenSettonetal.Main.docx","url":"https://assets-eu.researchsquare.com/files/rs-8289747/v1/80bd1e2c7a4aeaad3cb5123d.docx"},{"id":98432579,"identity":"522e7be0-7a16-4703-b465-c8aad0407dbd","added_by":"auto","created_at":"2025-12-17 16:49:43","extension":"json","order_by":1,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":11828,"visible":true,"origin":"","legend":"","description":"","filename":"NCOMMS25098768.json","url":"https://assets-eu.researchsquare.com/files/rs-8289747/v1/bd2f0f2c971734b7c0d61ea8.json"},{"id":98200883,"identity":"26d8f75c-36f4-4f3d-bf8d-91281155e477","added_by":"auto","created_at":"2025-12-15 07:50:15","extension":"docx","order_by":2,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":11646865,"visible":true,"origin":"","legend":"","description":"","filename":"CohenSettonetal.Supplementary.docx","url":"https://assets-eu.researchsquare.com/files/rs-8289747/v1/57716b75f9ff1d4c7bffbf33.docx"},{"id":98431910,"identity":"398006fa-5e3a-4f59-8048-b72d3ddfbbe9","added_by":"auto","created_at":"2025-12-17 16:48:36","extension":"xml","order_by":3,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":116974,"visible":true,"origin":"","legend":"","description":"","filename":"NCOMMS250987680enriched.xml","url":"https://assets-eu.researchsquare.com/files/rs-8289747/v1/012c4abd773b9128ffaf614c.xml"},{"id":98432939,"identity":"d45c60d2-a1ba-4dee-b1a5-d237b02b8138","added_by":"auto","created_at":"2025-12-17 16:50:07","extension":"jpeg","order_by":4,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":435024,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage1.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-8289747/v1/3fbc8b287b8b7aeb6a8d9ab6.jpeg"},{"id":98432311,"identity":"2277a1ad-e45c-4fce-8040-49eac99bcb18","added_by":"auto","created_at":"2025-12-17 16:49:22","extension":"jpeg","order_by":5,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":932711,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage2.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-8289747/v1/ba87405295ba43ab364c159d.jpeg"},{"id":98431942,"identity":"498afa08-e4b0-42d5-a316-cbc94d5abefe","added_by":"auto","created_at":"2025-12-17 16:48:40","extension":"jpeg","order_by":6,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":613433,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage3.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-8289747/v1/78680a211ba982824381baf6.jpeg"},{"id":98431903,"identity":"742526d8-c45d-4e5c-ac3e-c126dfcfc406","added_by":"auto","created_at":"2025-12-17 16:48:36","extension":"jpeg","order_by":7,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":379320,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage4.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-8289747/v1/baf44bfd5cedacbecd12fdce.jpeg"},{"id":98432541,"identity":"3986fdac-ac1c-4621-b0f5-a8f3b89c9272","added_by":"auto","created_at":"2025-12-17 16:49:40","extension":"jpeg","order_by":8,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":911382,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage5.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-8289747/v1/329ef249e683c597f19ab031.jpeg"},{"id":98431547,"identity":"c8eaeb41-92a0-4d12-bdc4-97bfbce09f84","added_by":"auto","created_at":"2025-12-17 16:47:52","extension":"jpeg","order_by":9,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":287612,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage6.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-8289747/v1/414d9f53ea3a82d06947360d.jpeg"},{"id":98200887,"identity":"c751d542-4e24-4e6d-838d-848699df033f","added_by":"auto","created_at":"2025-12-15 07:50:15","extension":"png","order_by":10,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":130201,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-8289747/v1/5f5752b2020066da32121600.png"},{"id":98200885,"identity":"28dcbc78-42f1-4718-97ec-4b78aef4fdc0","added_by":"auto","created_at":"2025-12-15 07:50:15","extension":"png","order_by":11,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":221458,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-8289747/v1/b670de7769b7394592e1f432.png"},{"id":98431512,"identity":"dcf7a3cc-bbc4-4058-80ac-790c9283dde9","added_by":"auto","created_at":"2025-12-17 16:47:50","extension":"png","order_by":12,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":98058,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-8289747/v1/08f54908d572486050825fac.png"},{"id":98200882,"identity":"44a0314c-eb35-4761-b654-8d35a553b3b5","added_by":"auto","created_at":"2025-12-15 07:50:15","extension":"png","order_by":13,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":84487,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-8289747/v1/3496e0dc65f1e8cda5b26637.png"},{"id":98200888,"identity":"36578fb4-f129-44c8-8475-c1bd196a33e7","added_by":"auto","created_at":"2025-12-15 07:50:15","extension":"png","order_by":14,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":168201,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-8289747/v1/fd8207b956554c9e379ccdc6.png"},{"id":98200892,"identity":"d2aed545-1e75-4fdc-9e06-903125cd13a3","added_by":"auto","created_at":"2025-12-15 07:50:15","extension":"png","order_by":15,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":98763,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage6.png","url":"https://assets-eu.researchsquare.com/files/rs-8289747/v1/f6935abc511fa73a9666a2ca.png"},{"id":98433137,"identity":"e267ac0a-1dc5-4825-85bd-c80b1c8d0a16","added_by":"auto","created_at":"2025-12-17 16:50:20","extension":"xml","order_by":16,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":114182,"visible":true,"origin":"","legend":"","description":"","filename":"NCOMMS250987680structuring.xml","url":"https://assets-eu.researchsquare.com/files/rs-8289747/v1/fd1df9989c2c60258bb4273a.xml"},{"id":98200894,"identity":"3224122d-b5a8-42c4-8f35-102a945e8c62","added_by":"auto","created_at":"2025-12-15 07:50:15","extension":"html","order_by":17,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":130001,"visible":true,"origin":"","legend":"","description":"","filename":"earlyproof.html","url":"https://assets-eu.researchsquare.com/files/rs-8289747/v1/f4acac087637f90de9bd70ee.html"},{"id":98200869,"identity":"a412ba50-13d9-4d22-935e-4d94c4a04f7e","added_by":"auto","created_at":"2025-12-15 07:50:13","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":719158,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003ePRESSnet is an end-to-end framework for patient stratification and biomarker discovery using knowledge graphs.\u003c/strong\u003e \u003cstrong\u003ea\u003c/strong\u003e, Overview of how PRESSnet turns multimodal clinical trial data into knowledge graphs of patient data and uses these to generate biomarker discovery and patient stratification hypotheses. \u003cstrong\u003eb\u003c/strong\u003e, Visualisation of how PRESSnet turns raw multimodal data, including treatment outcome data, into a knowledge graph that connects patient nodes to feature and endpoint nodes. \u003cstrong\u003ec\u003c/strong\u003e, PRESSnet offers a suite of graph algorithms for separating patient nodes by their connectivity to biomarker and clinical endpoint nodes; from these algorithms, PRESSnet generates novel patient stratification and biomarker hypotheses, including composite biomarker signatures that can be made up of multiple biomarkers from different modalities. \u003cstrong\u003ed\u003c/strong\u003e, PRESSnet has in-built and post-hoc methods to assess the robustness of generated hypotheses and/or perform meta-analyses to generate results.\u003c/p\u003e","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-8289747/v1/149279d814d083e79a3b4480.png"},{"id":98431380,"identity":"bcd4aba4-e88a-4ed1-ba53-f14595bb3668","added_by":"auto","created_at":"2025-12-17 16:47:36","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":1077718,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eVisualisation workflow for the generation of stratification and biomarker hypotheses using supervised PRESSnet analysis with bootstrapping.\u003c/strong\u003e \u003cstrong\u003ea\u003c/strong\u003e, A graph is created of the data linking patients to their features and OS endpoints. \u003cstrong\u003eb\u003c/strong\u003e, Community detection algorithms are applied to the data across multiple iterations and \u003cstrong\u003ec\u003c/strong\u003e, the extracted communities are aggregated to produce a final patient stratification hypothesis. \u003cstrong\u003ed\u003c/strong\u003e, PageRank weights corresponding to survival association are generated across multiple iterations with permutation testing. The extracted communities are aggregated to produce a final patient stratification hypothesis. \u003cstrong\u003ee\u003c/strong\u003e, Biomarkers from both approaches (based respectively on PageRank weights and co-occurrences in communities of feature nodes with endpoint nodes) are aggregated to produce a final set of univariate and composite biomarker hypotheses.\u003c/p\u003e","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-8289747/v1/63dd44b4a488e00d289c5577.png"},{"id":98431698,"identity":"dea0282e-aecb-4cc1-b953-bd826db127df","added_by":"auto","created_at":"2025-12-17 16:48:11","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":779090,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eStratification and biomarker hypotheses produced by PRESSnet for LUAD patients given anti-PD1 therapy.\u003c/strong\u003e Examining outcome node and feature node co-occurrences in communities, as well as feature enrichment within patient groups, allows us to link stratification and biomarker hypotheses; the biomarkers on the right were associated with the communities on the left that displayed the worst survival trajectory. a, Hypotheses generated by the direct community detection approach on the original patient graph. Log-rank\u003cem\u003e P \u003c/em\u003e(sig. +ve v.s. overall population) = 2.17e-05, HR (95% CI, sig. +ve) = 3.22 (1.98 – 5.23) \u003cstrong\u003eb\u003c/strong\u003e, Hypotheses generated by the meta-clustering approach for community detection on bootstrapped versions of the graph. Log-rank\u003cem\u003e P \u003c/em\u003e(sig. +ve v.s. overall population) = 4.59e-04, HR (95% CI, sig. +ve) = 2.46 (1.63 – 3.71).\u003c/p\u003e","description":"","filename":"floatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-8289747/v1/43c0b60e3f637ab5dbeea463.png"},{"id":98431063,"identity":"df4bfdb6-3942-4418-b3e6-a0b3e4de6710","added_by":"auto","created_at":"2025-12-17 16:46:54","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":690870,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eBenchmarking biomarker hypotheses generated by PRESSnet for LUAD patients given anti-PD1 therapy. a\u003c/strong\u003e, Visualisation of normalised bootstrapped patient weights from PRESSnet’s Adapted Personalised PageRank model versus normalised feature weights from a random forest classifier (RFC), fitted on the multiomic tabular MSK IO data and trained to predict the OS event. True positives and statistically significant weights were ranked higher by PageRank than by RFC. \u003cstrong\u003eb\u003c/strong\u003e, Assessing the significance of composite biomarker metrics against those of univariate features according to their log rank and Cox Proportional Hazard survival statistics. More composite biomarkers have statistically significant metrics than univariate, noted by the frequency in the bottom left box. Gray lines indicate the cutoff for \u003cem\u003eP\u003c/em\u003e \u0026lt; 0.05 for both assessment type. \u003cstrong\u003ec. \u003c/strong\u003e\u0026nbsp;Box plots comparing survival metrics for univariate v.s. composite biomarkers, plus their associated statistical significances. Composite biomarkers displayed elevated hazard ratios compared to their univariate counterparts, with a wider distribution of values, and the statistical significance of these ratios was also higher (as reflected in b.).\u003c/p\u003e","description":"","filename":"floatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-8289747/v1/e8ba82fd52084c6ce8addcb5.png"},{"id":98200889,"identity":"362e5040-4b4d-4adb-8e4e-f594d18141c1","added_by":"auto","created_at":"2025-12-15 07:50:15","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":1095336,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eBiomarker hypotheses produced by PRESSnet for AML patients in TCGA, Beat AML and a merged dataset.\u003c/strong\u003e \u003cstrong\u003ea\u003c/strong\u003e, Examples of biomarkers derived from PRESSnet analysis of the TCGA dataset that were found to be significant in the Beat AML dataset. SMAD3 is a biomarker of poor prognosis in AML patients receiving chemotherapy\u003csup\u003e62\u003c/sup\u003e and both it and PSMB8 have been proposed as therapeutic targets for AML.\u003csup\u003e63,64\u003c/sup\u003e \u003cstrong\u003eb\u003c/strong\u003e, Examples of biomarkers found by analysing the merged dataset. CALR has been found to be more highly expressed in AML than in other haematological cancers,\u003csup\u003e65\u003c/sup\u003e and CUX1 low expression has been linked to myelodysplastic syndromes,\u003csup\u003e66\u003c/sup\u003e but their prognostic relevance as suggested here has not been documented before. Statistics for \u003cstrong\u003ea\u003c/strong\u003e. SMAD3: TCGA\u003cem\u003e; \u003c/em\u003eLog-rank\u003cem\u003e P \u003c/em\u003e(sig. +ve v.s. overall population) = 0.00891, HR (95% CI, sig. +ve) = 0.42 (0.25 – 0.70). Beat AML; Log-rank\u003cem\u003e P \u003c/em\u003e(sig. +ve v.s. overall population) = 0.0359, HR (95% CI, sig. +ve) = 0.71 (0.55 – 0.91). PSMB8 \u0026amp; TCF19: TCGA; Log-rank \u003cem\u003eP\u003c/em\u003e (sig. +ve v.s. overall population) = 0.0022, HR (95% CI, sig. +ve) = 2.36 (1.45 – 3.85). Beat AML; Log-rank\u003cem\u003e P\u003c/em\u003e (sig. +ve v.s. overall population) = 3.4e-04, HR (95% CI, sig. +ve) = 1.71 (1.33 – 2.2). \u0026nbsp;\u003cstrong\u003eb\u003c/strong\u003e. CALR \u0026amp; CUX1, merged dataset; Log-rank \u003cem\u003eP\u003c/em\u003e (sig. +ve v.s. overall population) = 2.67e-05, HR (95% CI, sig. +ve) = 0.53 (0.42 – 0.66). SRC \u0026amp; TP53, merged dataset; Log-rank \u003cem\u003eP\u003c/em\u003e (sig. +ve v.s. overall population) = 6.99e-16, HR (95% CI, sig. +ve) = 3.28 (2.48 – 4.32).\u003c/p\u003e","description":"","filename":"floatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-8289747/v1/4c22757aaa698a7a38d603ec.png"},{"id":98432928,"identity":"4f54c7a5-5719-430d-95b5-09f7b458cd57","added_by":"auto","created_at":"2025-12-17 16:50:07","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":516675,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003ePRESSnet offers multiple algorithm types for biomarker hypothesis generation. a\u003c/strong\u003e, Community detection algorithms create communities that can be made up of any node type. As communities comprise closely connected nodes, potential biomarkers should end up in communities with their corresponding outcome nodes, and co-occurring biomarkers should end up in the same community. Here, we would look for signatures of response from patients who have combinations of Gene A mutation, TMB high and Gene B wildtype. \u003cstrong\u003eb\u003c/strong\u003e, Nodes with similar connectivity should have similar embeddings, so we can find biomarkers close to their corresponding outcome node in embedding space. PRESSnet treats this as a multi-objective problem to identify biomarkers, simultaneously minimising distances to desired nodes and maximising distances to the opposite. A hypothetical ideal response biomarker is represented as a green node. \u003cstrong\u003ec\u003c/strong\u003e, PRESSnet’s Adapted Personalised PageRank approach propagates signal from endpoint nodes to biomarkers, using the connectivity of the graph to learn weights representing the strength of relationship between them.\u003c/p\u003e","description":"","filename":"floatimage6.png","url":"https://assets-eu.researchsquare.com/files/rs-8289747/v1/aae5ed9a5a1adb9177ccd70c.png"},{"id":98445032,"identity":"a3f4d929-3525-429d-9190-6dc5fda9f6a4","added_by":"auto","created_at":"2025-12-17 17:18:44","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":6249098,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8289747/v1/60ac2600-0bdf-4368-b2b0-b64a159d910d.pdf"},{"id":98200891,"identity":"8be43830-91ba-45bd-a3cc-1e93a833d75f","added_by":"auto","created_at":"2025-12-15 07:50:15","extension":"docx","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":11646865,"visible":true,"origin":"","legend":"Supplementary Figures and Tables","description":"","filename":"CohenSettonetal.Supplementary.docx","url":"https://assets-eu.researchsquare.com/files/rs-8289747/v1/e5d8a41b52dc92260d6cbb20.docx"}],"financialInterests":"There is \u003cb\u003eNO\u003c/b\u003e Competing Interest.","formattedTitle":"PRESSnet: a novel framework for patient stratification and biomarker discovery using clinical knowledge graphs","fulltext":[{"header":"Introduction","content":"\u003cp\u003eAgainst a background of diverse technological and scientific advances in recent years, the rate of success for clinical trials has remained low.\u003csup\u003e1\u003c/sup\u003e Currently, approximately 9 of 10 new molecular entities entering a clinical study are likely to fail.\u003csup\u003e2\u003c/sup\u003e A key contributor to success, or lack thereof, is patient selection. Personalised medicine has become a ubiquitous goal in pharmaceutical R\u0026amp;D: identifying the optimum therapy for a clearly definable population of patients saves lives and improves the quality of life for trial participants while making medicines more affordable. Major pharmaceutical companies have adopted personalised medicine as a core part of their R\u0026amp;D strategy (for example, AstraZeneca\u0026rsquo;s 5Rs framework\u003csup\u003e3\u003c/sup\u003e), impacting all stages of drug discovery from target discovery to identification of disease subgroups. These advances have led to demonstrable improvements in numerous areas, including phase 3 trial completion.\u003csup\u003e3\u003c/sup\u003e As part of these developments, the use of biomarkers to define populations and stratify patients during trial design has become the norm across diseases. Multiple biomarker-driven oncology therapies have been approved in the last decade\u003csup\u003e4\u003c/sup\u003e: these include osimertinib, which targets patients with the T790m \u003cem\u003eEGFR\u003c/em\u003e mutation in non\u0026ndash;squamous-cell lung cancer (NSCLC)\u003csup\u003e5\u003c/sup\u003e; anti-HER2 therapies such as trastuzumab deruxtecan,\u003csup\u003e6\u003c/sup\u003e which target HER2-positive breast and other cancers; and immune checkpoint blockade therapies such as pembrolizumab,\u003csup\u003e7\u003c/sup\u003e for which high programmed cell death ligand-1 (PD-L1) status is a key patient selection criterion. Other major biomarker-driven therapies include those for patients with \u003cem\u003eBRCA\u003c/em\u003e-mutant breast and ovarian cancers, who are typically treated with poly(ADP-ribose) polymerase (PARP) inhibitors,\u003csup\u003e8\u003c/sup\u003e and patients with high microsatellite instability (MSI) disease across multiple cancer types, who are more likely to benefit from immunotherapy (IO).\u003csup\u003e9\u003c/sup\u003e\u003c/p\u003e \u003cp\u003eThe fields of data science and artificial intelligence (AI) are now increasingly focusing on biomarker identification and patient stratification, and multiple computational pipelines for these, some of which are AI driven, have recently been developed\u003csup\u003e10\u003c/sup\u003e. Statistical models such as the Cox proportional hazards model have also long been applied to survival analysis\u003csup\u003e11\u003c/sup\u003e and are still commonly used today as interpretable tools to investigate the relationship between variables and survival time. However, all current computational approaches face multiple challenges. In many cases, the number of features available in a dataset far exceeds the number of patients, which for early-phase trial arms often falls below 100. These challenges raise potential issues including algorithm overfitting, feature redundancy, and spurious relationship generation. Patient data are typically heterogeneous and multimodal, incorporating clinical features, omics, and other patient-level readouts from imaging or other sources. Some feature modalities, such as those based on mutations, necessitate the accommodation of high levels of feature sparsity. In addition, univariate biomarkers extracted from these data can lack the granularity to distinguish patient survival trajectories across multiple datasets for the same indication, line of treatment, and other fixed conditions, especially when measurements are not standardised.\u003csup\u003e12\u003c/sup\u003e An approach that can easily integrate multimodal data while also extending beyond univariate recommendations would therefore be highly advantageous. For these reasons, knowledge graphs (KGs) carry promise for this area of research.\u003c/p\u003e \u003cp\u003eA network graph is a logical map of relationships (\u0026lsquo;edges\u0026rsquo;) between entities (\u0026lsquo;nodes\u0026rsquo;). Computational approaches based on homogeneous networks, in which all nodes and edges are of the same type (e.g., a gene-gene interaction network), have already been implemented for integration of multiomic data and hypothesis generation for patient stratification. For example, similarity network fusion is used to create homogeneous networks of patient-patient similarity edges by merging these edges from different omics,\u003csup\u003e13\u003c/sup\u003e whereas network-based statistics generate patient-derived features by propagating data from a patient\u0026rsquo;s mutations across a gene interaction network and using the resulting patient-feature matrix to generate subtypes.\u003csup\u003e14\u003c/sup\u003e More recently, homogeneous network approaches have also been applied to biomarker discovery, for example, by propagating signals from drug targets or disease signatures through protein-protein interaction networks, to identify biomarkers of drug response.\u003csup\u003e15\u0026ndash;17\u003c/sup\u003e\u003c/p\u003e \u003cp\u003eUnlike homogeneous networks, KGs can be composed of multiple types of nodes and edges. In biomedicine, examples of such edges between two node types are \u0026lsquo;Gene \u0026rarr; is expressed in \u0026rarr; Patient\u0026rsquo; and \u0026lsquo;Patient \u0026rarr; responds to \u0026rarr; Drug\u0026rsquo;. KGs can be extended beyond homogeneous networks by capturing heterogeneous data and relationships between entities in a highly flexible and scalable data structure, and as such their applications to the drug discovery pipeline have been steadily expanding. KGs have been used to rank targets for resistance to EGFRm inhibitors in NSCLC\u003csup\u003e18\u003c/sup\u003e and to capture pharmacological and disease data in frameworks that can be leveraged for drug repurposing.\u003csup\u003e19,20\u003c/sup\u003e In the clinical setting, a recent study used KGs to predict cohort-level response rates in trial arms,\u003csup\u003e21\u003c/sup\u003e using the accompanying graph node embeddings as inputs for asset prioritisation models. The use of KGs to model data at the individual patient level is also still relatively new. Notably, patient data have been encoded in one edge type of a mutation-based KG to capture patient-specific mutation co-occurrences alongside other omic data for the identification of downstream targets.\u003csup\u003e22\u003c/sup\u003e They have also been represented as patient nodes in a KG incorporating patients, genomics, and biomedical prior knowledge that can be used to derive patient features for survival prediction.\u003csup\u003e23\u003c/sup\u003e There is a clear opportunity to combine the ability of KGs to capture individual patient data with their ability to generate multimodal insights and to leverage this function for personalised medicine via a systematic, reusable, generalisable framework.\u003c/p\u003e \u003cp\u003eHere we present PRESSnet (\u003cb\u003eP\u003c/b\u003eatient \u003cb\u003eRE\u003c/b\u003ecommendation via \u003cb\u003eS\u003c/b\u003etratification and \u003cb\u003eS\u003c/b\u003eelection using \u003cb\u003enet\u003c/b\u003eworks), a framework for generating hypotheses for patient stratification and biomarker discovery by analysing patient KGs from clinical trial data. The user can choose which trial modality files to include from their dataset (e.g., clinical and genomic features) and use their choice of graph tool to analyse the data. PRESSnet then automatedly creates a KG of the input data in which nodes represent patients, their associated features, and, optionally, their clinical outcomes. In addition, biomedical prior knowledge\u0026mdash;for example, gene-pathway or gene-gene relationship data\u0026mdash;can also be integrated with the graphs. PRESSnet generates insights from the KG via multiple algorithms. As graph algorithms capture interrelationships between nodes, and the graphs contain nodes representing features, PRESSnet can offer biomarker signatures that are \u0026lsquo;composite\u0026rsquo;, i.e., that are made up of more than one biomarker and can integrate multiple modalities within one signature (e.g., \u0026lsquo;Gene A mutated AND Tumor Mutational Burden low\u0026rsquo;). We implemented PRESSnet on patient data from two independent patient populations: one in which we assessed the power of this framework in a treatment-specific setting, and the other across a range of treatments and multiple datasets. We demonstrate PRESSnet\u0026rsquo;s ability to generate explainable patient stratification hypotheses and to suggest a mixture of known and novel potential biomarkers with statistical significance estimation across datasets.\u003c/p\u003e"},{"header":"Results","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003eThe PRESSnet framework\u003c/h2\u003e \u003cp\u003ePRESSnet works end-to-end to convert raw data into graph-derived insights in a customizable and scalable approach (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e1\u003c/span\u003e). The framework consists of multiple stages:\u003c/p\u003e \u003cp\u003e \u003col\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eUser-defined input parameters: The user can define choices around input data and algorithm parameters before executing the methodology.\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eAutomated KG creation: PRESSnet creates a KG containing patient, feature, and (optionally) clinical outcome data directly from the user\u0026rsquo;s input files.\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eGraph algorithms for patient stratification and biomarker discovery: The user\u0026rsquo;s choice of graph algorithm (e.g., a graph-embedding algorithm) is applied to the graph for downstream analysis.\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eRobustness assessment: Optionally, robustness assessment (bootstrapping and permutation testing) is applied to the results produced by the algorithm.\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eGeneration of patient and feature insights: PRESSnet outputs a set of files containing information such as patient clusters, biomarker information, and clinical outcome statistics.\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003c/ol\u003e \u003c/p\u003e \u003cp\u003eEach patient is represented by a unique node, and a patient node is connected to nodes representing the features that the patient possesses, creating a network that links patients through shared features. For continuous data, a node is created corresponding to categories (e.g., tumor mutational burden [TMB] score\u0026thinsp;\u0026gt;\u0026thinsp;10, TMB score\u0026thinsp;\u0026le;\u0026thinsp;10; see Methods: KG creation).\u003c/p\u003e \u003cp\u003ePRESSnet offers two types of KG analysis: supervised and unsupervised. Both approaches can be used for patient stratification. Supervised analysis incorporates patients\u0026rsquo; clinical outcome data in the graph, and this can also be leveraged for biomarker discovery by analysing the connectivity between outcome nodes and feature nodes via shared patient nodes. Unsupervised analysis excludes this outcome data and therefore provides more flexibility for integrating datasets that include treatment or outcome data with those that do not for applications such as genomics-driven subtyping. Including both supervised and unsupervised approaches in one framework provides the user with the advantages of both. As referenced above, since features are themselves nodes, we can generate hypotheses for not just how patients stratify or cluster, but also how specific sets of features are associated with each other, either with respect to survival or not.\u003c/p\u003e \u003cp\u003ePRESSnet also serves as an importable Python library for custom analyses. Users can create graphs directly from data frames and can benefit from \u0026lsquo;bring your own graph\u0026rsquo; (BYOG) approaches, whereby they can apply its stratification or biomarker discovery functions to pre-made graphs containing patient and feature nodes. This is useful when the user has already assembled a graph or has a custom graph that cannot be assembled in one step from raw data files. A key benefit of KGs for preclinical or clinical data is the ability to connect gene data to prior knowledge through relationships such as gene \u0026rarr; pathway edges. PRESSnet offers the ability to integrate these prior knowledge nodes with the gene nodes in the trial KG (Supplementary Fig.\u0026nbsp;1; see Methods: KG creation).\u003c/p\u003e \u003cp\u003eThe algorithms in PRESSnet can be used for patient stratification and biomarker discovery applications. For patient stratification, community detection algorithms separate patients immediately into communities, and PRESSnet provides the user with the option to generate the most statistically significantly enriched features of each community; this offers explainability by revealing the features driving the stratification hypothesis. Graph embeddings (e.g., from random walk\u0026ndash;based algorithms) can be used to cluster patients in downstream analysis, using tools like UMAP or TSNE with DBSCAN, K-means, etc. (see Methods: Graph algorithms for patient stratification and biomarker discovery). For biomarker discovery, both community detection and graph embedding\u0026ndash;led approaches can generate biomarkers of response or survival; they do so via the supervised approach, whereby the connectivity of the nodes for these endpoints is leveraged to generate insights (e.g., biomarkers of response ending up in communities with the \u0026lsquo;responder\u0026rsquo; node). For biomarker discovery, PRESSnet also offers an adapted version of personalised PageRank, a link analysis algorithm which in this case measures the importance of feature nodes in the graph relative to endpoint nodes (see Methods: Graph algorithms for patient stratification and biomarker discovery).\u003c/p\u003e \u003cp\u003eFor any analysis framework working on relatively small datasets, maximising robustness in the face of potential overfitting and selection bias is highly desirable. As such, the PRESSnet workflow incorporates methods to address the robustness of its recommendations through the PageRank and community detection workflows. The user can choose whether to apply bootstrapping and/or permutation testing, in which the underlying graph is respectively resampled or shuffled (see Methods: In-built meta-analysis and robustness assessment).\u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eBenchmarks for statistically significant biomarker discovery and patient stratification\u003c/h3\u003e\n\u003cp\u003eFor benchmarking, we tested PRESSnet on a patient dataset containing widely established (\u0026lsquo;known\u0026rsquo;) biomarkers of overall survival (OS). The 2022 Memorial Sloan Kettering (MSK) IO TMB dataset\u003csup\u003e24\u003c/sup\u003e is a public dataset of patients with stage IV NSCLC treated with immune checkpoint inhibitors. It contains survival, response, and treatment data for each patient, as well as genomic and clinical features that include multiple widely published biomarkers for IO-treated NSCLC. We conducted an analysis of the overall dataset (240 patients with mutations and clinical data) and also carried out a subpopulation-specific analysis by splitting the data for treatment type (e.g., anti\u0026ndash;programmed cell death-1[PD-1] vs. anti\u0026ndash;CTLA-4 therapy) and cancer subtype (lung adenocarcinoma [LUAD] vs. lung squamous-cell carcinoma [LUSC]); from the latter, the only subpopulations large enough for meaningful analysis were anti-PD-1\u0026ndash;treated LUSC (\u003cem\u003en\u003c/em\u003e\u0026thinsp;=\u0026thinsp;27) and anti-PD-1\u0026ndash;treated LUAD (\u003cem\u003en\u003c/em\u003e\u0026thinsp;=\u0026thinsp;168) (Supplementary Table\u0026nbsp;1). Within the latter subpopulation, we examined whether PRESSnet could meaningfully stratify patients by clinical outcome and uncover biomarkers of OS.\u003c/p\u003e \u003cp\u003eWe implemented PRESSnet in a supervised setting for patient stratification using both direct, one-pass community detection on the whole patient graph and a metaclustering of community detection runs across bootstrapped graphs (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e2\u003c/span\u003e). For both approaches, PRESSnet generated communities of patients with clearly distinguished survival trajectories (Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e3\u003c/span\u003e) and associated biomarkers. We also implemented the same type of analysis in an unsupervised setting with survival data hidden. In both the direct and metaclustering-based unsupervised approaches, PRESSnet was able to generate two distinct communities of patients with statistically significant survival differentiation. A visualisation and statistics for subtypes produced by PRESSnet versus benchmark methods for stratification including MOFA\u0026thinsp;+\u0026thinsp;is shown in Supplementary Fig.\u0026nbsp;2. Compared to these methods, PRESSnet displays superior cluster separation by silhouette and C-H score (e.g. unsupervised C-H score of 20.043 for PRESSnet\u0026rsquo;s metaclustering approach v.s. 6.156 for MOFA\u0026thinsp;+\u0026thinsp;with K-Means) and evidences a better balance between patient survival differentiation and practical, plausible clustering in a supervised setting (e.g. 2 clusters with a C-H score of 9.404 and concordance index of 0.659 for PRESSnet\u0026rsquo;s metaclustering approach, v.s. 9 clusters for MOFA\u0026thinsp;+\u0026thinsp;with a concordance index of 0.915 but a C-H score of 2.367).\u003c/p\u003e \u003cp\u003eWe next generated a final list of biomarker hypotheses for the LUAD population, using the process visualised in Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e2\u003c/span\u003e. The biomarkers associated with the stratifications from the community detection analysis described above were combined with biomarkers from a separate PageRank-based analysis that also used bootstrapping. These biomarkers included known markers of poor prognosis in immune checkpoint inhibitor\u0026ndash;treated NSCLC, such as STK11 and KEAP1 driver mutation, TMB, and PD-L1 status,\u003csup\u003e25,26\u003c/sup\u003e that were consistent across bootstrapping and passed permutation testing. Examples of composite biomarkers in the LUAD population from the bootstrapping approaches are shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e3\u003c/span\u003e. A comparison of the overlap in the LUAD population between PRESSnet-derived univariate markers, a benchmark machine learning (ML)\u0026ndash;based biomarker discovery technique (see Methods: Benchmarks for patient stratification and biomarker discovery on MSK 2022 LUAD data), and the statistically top biomarkers from log-rank testing or univariate Cox modelling are shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e4\u003c/span\u003ea. PRESSnet outperformed the ML benchmark method for retrieving statistically significant univariate markers according to log-rank testing or Cox modelling. We assessed the PRESSnet-derived signature biomarkers against univariate features and observed that composite biomarkers were superior to univariate biomarkers for statistically significant patient separation according to multiple metrics (Fig.\u0026nbsp;\u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e4\u003c/span\u003eb and \u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e4\u003c/span\u003ec).\u003c/p\u003e \u003cp\u003eWe also used PRESSnet to generate biomarker hypotheses in the LUSC population, using bootstrapping and permutation testing. Despite the small patient sample size of 27, PRESSnet was able to generate 58 univariate and composite biomarkers of OS that passed the graph-based permutation tests and carried statistical significance according to log-rank testing. An example of a significant composite biomarker in the LUSC population is shown in Supplementary Fig.\u0026nbsp;3. These results demonstrate the ability of the framework to generate insights in a small dataset.\u003c/p\u003e \u003cp\u003eWe also assessed the translatability of PRESSnet\u0026rsquo;s generated biomarkers and model weights by performing a train/test split analysis within the LUAD population. We performed 10 train/test splits of patients into 50% training and 50% test patients (\u003cem\u003en\u003c/em\u003e\u0026thinsp;=\u0026thinsp;84 for each category) and assessed whether (1) biomarkers generated from the training set would maintain statistical significance in the test set and (2) PRESSnet could be used to generate meaningful \u0026lsquo;risk scores\u0026rsquo; for unseen test set patients. For (1), PRESSnet generated on average 26 univariate and composite biomarkers of OS in each test split that maintained significance; biomarkers that were found to be significant across multiple of the 10 splits included \u0026lsquo;true positives\u0026rsquo; such as \u003cem\u003eSTK11\u003c/em\u003e mutation, PD-L1\u0026ndash;high scoring, and composite signatures involving known markers such as low albumin\u003csup\u003e27,28\u003c/sup\u003e and high neutrophil-to-lymphocyte ratio.\u003csup\u003e29\u003c/sup\u003e The biomarkers found to be significant in the most test sets can be found in Supplementary Table\u0026nbsp;2; along with Fig.\u0026nbsp;\u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e4\u003c/span\u003eb and \u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e4\u003c/span\u003ec., this demonstrates the utility of composite signatures for producing strong and persistent signal compared to univariate markers. For (2), we fed PRESSnet-derived weights for test patient nodes into a Cox regression model and found that they outperformed equivalent scores and categorisations from a benchmark deep neural network model according to multiple metrics (see Supplementary Table\u0026nbsp;3). Supplementary Fig.\u0026nbsp;4 shows an example of Kaplan-Meier (KM) curves for test set patients stratified into high-risk and low-risk groups by PRESSnet, demonstrating a clear and statistically significant inferior survival trajectory for the high-risk group. This highlights the potential of the framework to generate meaningful graph-derived features for assessing survival in unseen patients.\u003c/p\u003e\n\u003ch3\u003eIncluding biomedical prior knowledge for additional IO insights\u003c/h3\u003e\n\u003cp\u003eAs the treatments in this dataset were IO, we hypothesised that the inclusion of tumor microenvironment (TME)/IO-relevant prior knowledge might increase the power of graph-derived features and generate an extra layer of potential insights. We constructed a TME subgraph using public data (see Methods: Dataset preparation). By linking genes to pathways and the biological processes in which they are involved, we investigated how adding prior knowledge would (1) alter the performance of PRESSnet, e.g., for downstream predictive tasks, and (2) reveal extra context around drivers of OS in the original feature set.\u003c/p\u003e \u003cp\u003eFor (1), we used PRESSnet to generate embeddings of patients given anti-PD1 therapy, keeping algorithm parameters and train/test splits identical but changing the input data to exclude or include prior knowledge. The resultant patient embeddings underwent downstream classification using XGBoost\u003csup\u003e30\u003c/sup\u003e to predict binary response status across five splits. We observed that in both supervised and unsupervised settings (i.e. with outcome data nodes included or excluded from the graph), prior knowledge could scale to improve the downstream predictive performance of patient embeddings in a small dataset (the LUSC population, n\u0026thinsp;=\u0026thinsp;27) and a substantially larger one (all patients, n\u0026thinsp;=\u0026thinsp;240). In the latter, we observed that adding prior knowledge pathway data improved accuracy, F1 score and area under the curve (AUC) of the classifier compared to using the original dataset alone. In the smaller LUSC population, we observed the same pattern in the supervised setting, and while we did not see improvements in performance in the unsupervised setting with pathway inclusion alone, the addition of both pathway and gene-gene relationship produced by far the best performance compared to the other data options in this setting (see Supplementary Table\u0026nbsp;5). A visualisation of the embeddings can be found in Supplementary Fig.\u0026nbsp;5.\u003c/p\u003e \u003cp\u003eFor (2), prior knowledge KG nodes were separated or encoded like the original feature nodes by the algorithms in PRESSnet. Thus, in addition to looking at how biomarkers and patients separated when prior knowledge was included, we also examined what the outputs of PRESSnet revealed about the prior knowledge nodes themselves, by extracting their PageRank weights (Supplementary Fig.\u0026nbsp;6a). The results indicated the presence of a \u0026lsquo;tail\u0026rsquo; at both ends of the curve, where a small selection of nodes were either notably higher or lower weighted (Supplementary Fig.\u0026nbsp;6b, 6c), suggesting potential association with clinical outcome.\u003c/p\u003e \u003cp\u003eIn the above-described subset of prior knowledge nodes, we found multiple examples of features with established links to the efficacy of IO (and specifically anti\u0026ndash;PD-1 therapy) in the literature. Some nodes did not pass permutation testing with \u003cem\u003eP\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.05 due to signal attenuation, \u0026hellip;. ,; however, these still contained highly relevant nodes such as regulation of leukocyte activation (\u003cem\u003eP\u0026thinsp;=\u003c/em\u003e\u0026thinsp;0.1356; the contributory genes for this pathway include \u003cem\u003ePD-1\u003c/em\u003e, \u003cem\u003eCTLA-4\u003c/em\u003e, \u003cem\u003eRICTOR\u003c/em\u003e, and other proposed or existing NSCLC IO targets),\u003csup\u003e24,31\u003c/sup\u003e activation of B cells (\u003cem\u003eP\u0026thinsp;=\u003c/em\u003e\u0026thinsp;0.565; levels of B cell activation have been shown to correlate with poor response to anti\u0026ndash;PD-1 therapy\u003csup\u003e32\u003c/sup\u003e), positive regulation of macrophage apoptosis (\u003cem\u003eP\u0026thinsp;=\u003c/em\u003e\u0026thinsp;0.132; this is a process that can increase the efficacy of IO therapy by reducing immunosuppression\u003csup\u003e33\u003c/sup\u003e), and osteoblast development (\u003cem\u003eP\u0026thinsp;=\u003c/em\u003e\u0026thinsp;0.0854; many constituent pathway member genes for this process have been linked to T-cell exhaustion across cancer types\u003csup\u003e34\u0026ndash;36\u003c/sup\u003e). Nodes that passed permutation testing included the node for the vascular endothelial growth factor receptor signaling pathway (\u003cem\u003eP\u0026thinsp;=\u003c/em\u003e\u0026thinsp;0.00740; inhibition of this pathway alongside anti\u0026ndash;PD-1 therapy has been shown to improve anti-tumor response over anti\u0026ndash;PD-1 therapy alone\u003csup\u003e37\u003c/sup\u003e), and positive regulation of reactive oxygen species biosynthetic process (\u003cem\u003eP\u0026thinsp;=\u003c/em\u003e\u0026thinsp;0.00589; this has been shown to increase the efficacy of anti\u0026ndash;PD-1 therapy\u003csup\u003e38\u003c/sup\u003e). These findings provide evidence that PRESSnet can identify highly context-relevant nodes from prior knowledge data for explaining therapy-specific survival outcome, highlighting its utility for mechanistic biological interpretation and raising the possibility of future applications to diseases where therapy resistance and prognosis is less well understood.\u003c/p\u003e\n\u003ch3\u003eGeneralisable AML biomarker discovery across TCGA and Beat AML with PRESSnet\u003c/h3\u003e\n\u003cp\u003eFor biomarker discovery frameworks, the ability to uncover biomarkers that are applicable across datasets (rather than \u0026lsquo;overfit\u0026rsquo; on a specific patient cohort) is generally highly desirable. We therefore assessed PRESSnet for generalizability by applying it to two public acute myeloid leukemia (AML) patient datasets, Beat AML 2022\u003csup\u003e39\u003c/sup\u003e and The Cancer Genome Atlas (TCGA).\u003csup\u003e40\u003c/sup\u003e Both datasets contain treatment history, survival data, genomics, transcriptomics, and clinical data for AML patients (Supplementary Fig.\u0026nbsp;7); as these datasets contain a range of treatment modalities administered to patients across different time periods, we chose to perform a treatment-agnostic analysis, focusing on prognostic biomarker discovery.\u003c/p\u003e \u003cp\u003eTo explore generalisability, we implemented PRESSnet on both datasets separately and then identified biomarkers that were significant in, and overlapped across, both sets of results. PRESSnet generated 430 univariate and composite biomarkers from analysis of the TCGA dataset alone that turned out to be statistically significant (log-rank test \u003cem\u003eP\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.05) in both the TCGA dataset and the Beat AML dataset, including 217 with \u003cem\u003eP\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001 in at least 1 dataset and 45 with \u003cem\u003eP\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001 in both datasets. From analysis of the Beat AML dataset alone, we found that PRESSnet generated 413 biomarkers that were statistically significant in both datasets, including 274 with \u003cem\u003eP\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001 in at least 1 dataset, and 26 with \u003cem\u003eP\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001 in both datasets. The resulting markers contained a mixture of known AML biomarkers and novel biomarkers (Fig.\u0026nbsp;\u003cspan refid=\"Fig9\" class=\"InternalRef\"\u003e5\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eTo leverage the scalability of KGs, we also investigated whether combining datasets would increase our ability to find statistically powerful biomarkers that are present across a large patient population. This resulted in a single KG that contained patients from both TCGA and Beat AML, linked by shared features and clinical outcomes. We implemented PRESSnet on the combined graph, using experimental conditions that were identical to those of our individual dataset analysis. This generated 2425 biomarker hypotheses, including 1504 with \u003cem\u003eP\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001 and 389 with \u003cem\u003eP\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.00001 according to log rank testing. Two examples of resulting composite biomarker hypotheses are shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig9\" class=\"InternalRef\"\u003e5\u003c/span\u003eb. We also identified signatures for which statistical significance only occurred in the combined dataset (an example of which can be found in Supplementary Fig.\u0026nbsp;8). Together, these results serve as examples of how KGs and PRESSnet have the potential to facilitate cross\u0026ndash;dataset analysis for generating novel patient insights.\u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003eIn the era of personalised medicine, the ability to accelerate data-to-decision timelines, the integration of multimodal data for modelling, and the generation of targeted novel hypotheses are highly desirable. We have developed a generalisable framework for building patient knowledge graphs to identify explainable and translatable stratification and biomarker hypotheses. As a lightweight, customisable, end-to-end framework, PRESSnet empowers users to easily build a graph representation of their patient data and run iterative analyses across different analytical algorithms, data splits, modalities, and thresholds. PRESSnet can identify both univariate and composite biomarkers of outcomes in patient populations and can perform unsupervised or supervised analysis to drive patient separation by multiomic features or clinical outcomes. It can also integrate context from prior knowledge such as pathways and causal gene networks that can improve model performance and provide additional biologically grounded and explainable insights beyond the original dataset that can be directly interrogated experimentally, further eliminating the problem with interpreting context in the predictions from black-box ML models. We applied PRESSnet to multiple independent datasets, demonstrating its ability to find known, novel, and generalisable biomarkers across multiple datasets, data modalities, and indications. These included biomarkers of poor prognosis for anti–PD-1 treatment in NSCLC patients and generalisable prognostic markers of improved outcome in AML patients. We also showed the power of composite biomarkers that PRESSnet offers for outperforming univariate biomarkers for separating patients by survival and translating to unseen patients. PRESSnet is published as an adaptable framework for the scientific community to build and analyse KGs of patient data for their own use cases.\u003c/p\u003e \u003cp\u003eThere is scope to evolve PRESSnet further to meet the challenges of modelling multiomic and clinical data. Representing patient feature data in graphs is not a solved problem—for example, deliberately pursuing sparsity (e.g., by including only one biomarker per underlying variable, or removing ‘noisy’ edges for which the density of connections exceeds a certain threshold) might offer performance benefits for certain datasets—but we regard a systematic exploration of this issue as requiring a separate, future analysis. For our purposes, the common practice for biomarker discovery is to define a biomarker categorically (e.g., ‘TMB low’, ‘PD-L1 negative’), and as such, we developed PRESSnet to capture biomarkers in this way. However, the distribution of a given variable is often not conducive to straightforward categorisation (prompting many discussions about so-called ‘dichotomania’\u003csup\u003e41\u003c/sup\u003e). This potentially results in a loss of expressiveness once thresholds are defined. Additionally, the notion of ‘high’ or ‘low’ might vary greatly across contexts. PRESSnet provides the flexibility to apply thresholds in various ways during the pipeline, and users can also apply thresholds to their data before passing it through the framework to allow for column-specific data decisions. Another angle from which to approach this issue is to recommend variables rather than thresholded biomarkers and to encode continuous data in our biomarker discovery models. Graph neural networks (GNNs) can deal with weighted graphs that capture continuous values in nodes or edges. This raises the challenge of reconciling heterogeneous weighted edge types, but advances in heterogeneous GNNs,\u003csup\u003e42\u003c/sup\u003e particularly in the past several years, have facilitated the analysis of such data and enabled applications for multiomic data.\u003csup\u003e43\u003c/sup\u003e Future versions of PRESSnet explore the incorporation of GNNs for biomarker discovery via explainability mechanisms such as graph attention or retrofitted GNN explainers.\u003c/p\u003e \u003cp\u003ePRESSnet itself can also be extended to a wider range of applications than OS or treatment response. The framework can be easily adapted to uncover signatures of readouts such as adverse events for applications to patient safety, for example by representing the events as nodes in the KG and looking at features that co-occur with them in communities. For trial design, PRESSnet can also find markers of features such as HER2 or PD-L1 status. This would be useful in settings where such readouts are not available for some patients: generating a novel expression signature associated with HER2 status from immunohistochemistry imaging, for example, would potentially enable inference of HER2 status in patients who have not had this imaging carried out.\u003c/p\u003e \u003cp\u003ePRESSnet can also be applied to early discovery, for example preclinical target identification. Nodes representing cell lines, organoids or xenograft models can easily replace patient nodes in the workflow, replicating the same KG structure (e.g., a cell line node linked to nodes for mutations present in the cell line). This could be useful for deriving preclinical signatures of response or efficacy that can then be translated to a clinical setting. Conversely, uncovering druggable resistance markers for a given therapy can provide novel combination hypotheses or can be back-translated to inform target recommendations preclinical experimental validation. PRESSnet’s flexibility and lack of computational intensity make it particularly suited for incorporation into a ‘lab-in-the-loop’ setting, where feedback from experiments can be used to adjust feature processing, model architecture, causal prior knowledge, and robustness constraints for subsequent rounds of hypothesis generation.\u003c/p\u003e \u003cp\u003eA very important next step for PRESSnet will be leveraging opportunities offered by the rise of ‘Agentic AI’ and Foundation Models. We foresee PRESSnet benefitting greatly from a multi-agent framework that supports multiple parts of its analysis framework: a ‘setup agent’ can handle parameter configuration and feature selection, for example, and an ‘evidence agent’ can provide context from literature or other supporting databases for PRESSnet’s recommendations. Large language models serve as natural tools for intuitively querying KGs, enriching KG data, and providing semantic understanding of KGs, and as such, they could be integrated with PRESSnet to make generated KGs more accessible to non–data scientists and offer additional insights into the data. Knowledge derived from Foundation Models, for example, around gene function from GenePT,\u003csup\u003e44\u003c/sup\u003e can be integrated with PRESSnet in the future as part of prior knowledge or feature selection. As PRESSnet is a tool that can assign weights to individual patients and features or generate AI-derived numerical representations of them (e.g., embeddings), its outputs can also be fed into a multi-model framework governed by a central agent for downstream tasks such as survival prediction or target identification.\u003c/p\u003e \u003cp\u003eIn summary, our work indicates that PRESSnet demonstrates the capability of KG-based frameworks to rapidly analyse complex multiomic clinical datasets towards delivering effective targeted therapies to patients. It both serves as a touchstone for further research in this area and offers opportunities for extension to new domains in the future.\u003c/p\u003e \u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003cdiv id=\"Sec9\" class=\"Section3\"\u003e \u003c/div\u003e \u003c/div\u003e"},{"header":"Methods","content":"\u003ch2\u003eRunning PRESSnet\u003c/h2\u003e\u003cp\u003ePRESSnet can work as either an importable library or a command line tool in Python, in which the user specifies a script corresponding to the type of analysis they wish to perform, followed by a file path to the .yaml file they are using to configure the run. For example, if the user wished to run the embeddings generation pipeline using parameters from a .yaml file called \u003cem\u003etest_run.yaml\u003c/em\u003e, they would execute the following command:\u003c/p\u003e\n\u003ch3\u003epython pressnet_embeddings_generation.py -y test_run.yaml\u003c/h3\u003e\n\u003cp\u003eThe script runs end-to-end without any further input needed from the user. Print commands during the running of the script update the user on the progress of the analysis and provide relevant statistics such as survival rates of communities discovered by community detection.\u003c/p\u003e \u003cp\u003ePRESSnet is computationally inexpensive to run and practical in terms of time consumption. For example, generating communities from the graph of MSK LUAD patients (total edges\u0026thinsp;=\u0026thinsp;70,488) took 1.3 seconds when run on a 2020 MacBook Pro with a 2-GHz Quad-Core Intel Core i5 processor, and applying adapted Personalised PageRank to the graph of merged Beat AML and TCGA patients (totalling 646,609 edges) required 4 minutes and 14 seconds to execute on 50 bootstrapped samples and 3 minutes and 45 seconds to execute on 50 shuffled graphs for permutation testing on the same device. Extracting and aggregating log-rank test statistics across \u003cem\u003en\u003c/em\u003e samples of the data (see Methods: In-built meta-analysis and robustness assessment) is the most time-intensive step (e.g., 43 minutes and 50 seconds for 2450 biomarkers evaluated across 50 samples of 615 patients from the combined AML dataset). This is a parameter that can be customised by the user (using the variable \u003cem\u003en\u003c/em\u003e_resamples_composite_assessment in the .yaml) or switched off entirely, depending on the user\u0026rsquo;s task requirements. All outputs described in Results were generated by running PRESSnet on CPUs, and no pretraining is required for any of the algorithms.\u003c/p\u003e \u003cdiv id=\"Sec11\" class=\"Section2\"\u003e \u003ch2\u003eCreation of KGs\u003c/h2\u003e \u003cp\u003ePRESSnet uses a straightforward, customizable workflow to create a KG directly from a .yaml file (or a Python dictionary corresponding to such a file). The file sections for graph creation can be summarized as follows:\u003c/p\u003e \u003cp\u003e \u003cul\u003e \u003cli\u003e \u003cp\u003eFile definitions: The user specifies the paths to raw files they want to include in the graph, as well as the location to save output files.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eColumn definitions: The user specifies the name of the patient identifier column, the endpoints column for survival data, if applicable, and the columns to use to split the dataset; they also specify suffixes to distinguish shared underlying feature names (e.g., \u003cem\u003eKRAS\u003c/em\u003e mutation vs. KRAS expression) and define continuous columns for thresholding and the thresholding technique to use. Features with a shared underlying entity (e.g., KRAS for \u003cem\u003eKRAS\u003c/em\u003e mutation and KRAS expression) are not linked by default, as interrelationships, such as between expression level and mutation status, are not automatically inferable from underlying data.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003eInclusion of outcome data: The user specifies whether to include nodes for binary clinical outcome and categorised survival duration data in the graph, for supervised or unsupervised analysis.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003ePrior knowledge: The user toggles whether to include prior knowledge and specifies the location of this data if applicable.\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003c/p\u003e \u003cp\u003eThe first set of inputs allow the user to (1) specify the files they wish to use to create the graph and (2) provide feature annotations for these. The latter are on a per-file basis: for example, if a dataset has copy number, mutations, and expression data for the gene \u003cem\u003eCD274\u003c/em\u003e, in order to distinguish each modality in the final graph (rather than having one duplicated \u0026lsquo;CD274\u0026rsquo; node), the user might specify suffixes such that the features \u0026lsquo;CD274_exp\u0026rsquo;, \u0026lsquo;CD274_CNA\u0026rsquo;, and \u0026lsquo;CD274_mut\u0026rsquo; are created, and a graph node is subsequently created for each of these. Each file is read into a pandas data frame in Python, and the eventual group of data frames are merged on a specified patient identifier column. A graph is then constructed out of combined data.\u003c/p\u003e \u003cp\u003eOnce column naming is defined, the user can also indicate columns to be used to split the dataset. For example, the user may want to conduct per-indication or per-treatment analyses within a trial, allowing for the discovery of predictive biomarkers. A separate KG will therefore be constructed for the patients within each split.\u003c/p\u003e \u003cp\u003eFinally, the user can also specify how to apply thresholds to continuous data for biomarker creation. For continuous data that have been categorized, a node is created corresponding to each of the categories (e.g., TMB score\u0026thinsp;\u0026gt;\u0026thinsp;10, TMB score\u0026thinsp;\u0026le;\u0026thinsp;10).\u003c/p\u003e \u003cp\u003eThere are multiple options for how to apply thresholds:\u003c/p\u003e \u003cp\u003e \u003cul\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eMedian cutoff\u003c/b\u003e categorises each biomarker as greater than or less than equal to the median of its values.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eMean cutoff\u003c/b\u003e categorises each biomarker as greater than or less than equal to the mean of its values.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eQuantile cutoff\u003c/b\u003e categorises each biomarker into \u0026lsquo;top quantile\u0026rsquo;, \u0026lsquo;bottom quantile\u0026rsquo;, and \u0026lsquo;between quantiles\u0026rsquo; based on its values.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eYouden\u0026rsquo;s J-statistic-based cutoff\u003c/b\u003e uses Youden\u0026rsquo;s J statistic based on the receiving operator characteristic curve of the feature versus the endpoint event to determine the optimal cutoff from which to categorise the feature.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eLog-rank test-based cutoff thresholds\u003c/b\u003e are applied across the range of values of the feature and performs iterative log-rank testing of the \u0026lsquo;feature-high\u0026rsquo; versus \u0026lsquo;feature-low\u0026rsquo; subpopulations at each threshold to determine the optimal cutoff for the feature to generate the greatest survival separation. The user has the option to use a \u0026lsquo;fast\u0026rsquo; sliding window approach, which speeds up implementation of this approach for a slight loss of granularity. An initial set of uniformly spaced values (default \u003cem\u003en\u003c/em\u003e\u0026thinsp;=\u0026thinsp;10) are explored, and from these, a neighbouring pair is extracted for which the log-rank separation was greatest. The final cutoff is determined by iterating across the values that fall within the value range of the pair.\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003c/p\u003e \u003cp\u003ePRESSnet offers two types of KG analysis: supervised and unsupervised. Supervised analysis is used for biomarker discovery pipelines (see Methods: Graph algorithms for patient stratification and biomarker discovery). In a supervised analysis, nodes corresponding to outcomes (e.g., \u0026lsquo;responder\u0026rsquo;, \u0026lsquo;survival status: alive\u0026rsquo;) are included in the graph, just like feature nodes. As well as binary clinical outcome, the user may be interested in looking for biomarkers that reflect progression of disease related to time, i.e., biomarkers of \u0026lsquo;fast progressors\u0026rsquo; or \u0026lsquo;long survivors\u0026rsquo;. Duration nodes can be included in the KG, both in terms of quantile (\u0026lsquo;1st quantile duration\u0026rsquo;, \u0026lsquo;2nd quantile duration\u0026rsquo;, etc.) and in terms of the concept of progression (\u0026lsquo;fast progressor\u0026rsquo;, \u0026lsquo;intermediate group\u0026rsquo;, \u0026lsquo;long survivor\u0026rsquo;). The user can define the time thresholds for categorising the latter. For example, if the fast progressor threshold is set at 60 days, any patient with a survival duration less than or equal to that value will be connected to a \u0026lsquo;fast progressor\u0026rsquo; node in the graph. The shared connections between these feature nodes and these outcome nodes are leveraged to find biomarkers. Supervised analysis can also be used for stratification: the dense network connections between outcome nodes and patient nodes should cause the separation of patients to be influenced by these outcomes, leading to better stratification than the unsupervised approach. Unsupervised analysis prioritizes the features themselves driving separation, as the outcome data are not converted into graph edges and are excluded from any analysis. This type of analysis also offers more scope for integrating patient cohorts from different datasets in which treatments or endpoint types might differ.\u003c/p\u003e \u003cp\u003eRepresented as \u0026lsquo;gene \u0026rarr; entity\u0026rsquo; graph edges (e.g., gene \u0026rarr; associates \u0026rarr; disease), biomedical prior knowledge can be incorporated with a trial KG. This feature may be highly relevant to the dataset at hand. For example, if a gene \u003cem\u003eG\u003c/em\u003e is linked to an immune-related biological process \u003cem\u003eB\u003c/em\u003e from Gene Ontology annotations,\u003csup\u003e45,46\u003c/sup\u003e patients that express \u003cem\u003eG\u003c/em\u003e might be less receptive to IO. If the user has a file of gene \u0026rarr; biological process graph edges, they can specify to add this file to their KG, such that a patient \u003cem\u003eP\u003c/em\u003e who is linked to a gene \u003cem\u003eG\u003c/em\u003e through a gene expression relationship will now be connected at one extra \u0026lsquo;hop\u0026rsquo; away to the immune process \u003cem\u003eB.\u003c/em\u003e This will potentially provide better stratification; moreover, the node for \u003cem\u003eB\u003c/em\u003e may itself represent a form of biomarker and co-occur in communities with nodes corresponding to poor prognosis.\u003c/p\u003e \u003cp\u003eGraphs are constructed directly from data frames using the networkx library in Python and can easily be converted into a Pytorch Geometric data object for use with certain embedding algorithms (see below).\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec12\" class=\"Section2\"\u003e \u003ch2\u003eGraph algorithms for patient stratification and biomarker discovery\u003c/h2\u003e \u003cp\u003eOnce a KG is created, a variety of algorithms are available to users to leverage the graph according to the intended task. These can be categorised as community detection algorithms, embedding generation algorithms, and personalised PageRank. There is a unique script for each type of algorithm: \u003cem\u003epressnet_embeddings_generation.py\u003c/em\u003e, \u003cem\u003epressnet_community_detection.py\u003c/em\u003e, and \u003cem\u003epressnet_pagerank.py\u003c/em\u003e. The user can choose the specific algorithm within these in the .yaml file and can edit certain parameters for each algorithm, such as learning rate or resolution (a parameter specific to Louvain community detection that affects the size and number of communities found).\u003c/p\u003e \u003cp\u003eCommunity detection is the process by which graph nodes are separated into groups using the topology of the graph itself, typically with the aim of these groups being tightly connected. At present, PRESSnet offers two community detection techniques that work in different ways:\u003c/p\u003e \u003cp\u003e \u003cul\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eLouvain\u003c/b\u003e \u003csup\u003e47\u003c/sup\u003e tries to optimise a modularity score that represents the density of edges within a community as opposed to outside it. This algorithm was implemented using the networkx library in Python.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eEdMOT\u003c/b\u003e \u003csup\u003e48\u003c/sup\u003e samples sequences of neighbouring nodes using random walks and minimises a neighbourhood observation loss in the construction of the embeddings. This algorithm was implemented using the karateclub library in Python.\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003c/p\u003e \u003cp\u003eAfter running a community detection process, the user receives a list of node identifiers (including feature nodes and, optionally, endpoint nodes) with their corresponding community number. PRESSnet also provides information for the user about each community, including statistics and KM curves for survival if specified and, optionally, biomarkers associated with each community. The latter is measured by looking at the difference between the proportion of patients with the signature in the community versus the proportion of those in other communities. There is a parameter that allows users to set the minimum number of patient nodes for a given cluster; if a very small cluster is produced with patient \u003cem\u003en\u003c/em\u003e below this, it will be merged with the next-biggest cluster that is most similar by graph connectivity.\u003c/p\u003e \u003cp\u003eFor biomarker discovery, PRESSnet leverages the fact that communities can contain any type of node. As such, a node for any biomarker \u003cem\u003eB\u003c/em\u003e with a strong association with a clinical outcome \u003cem\u003eO\u003c/em\u003e (e.g., \u0026lsquo;survival status: deceased\u0026rsquo;) will end up in a community containing the node for \u003cem\u003eO\u003c/em\u003e (Fig.\u0026nbsp;\u003cspan refid=\"Fig11\" class=\"InternalRef\"\u003e6\u003c/span\u003ea). This community will also contain nodes for patients \u003cem\u003eP1\u003c/em\u003e, \u003cem\u003eP2\u003c/em\u003e.. . \u003cem\u003ePn\u003c/em\u003e who exhibit \u003cem\u003eO\u003c/em\u003e and \u003cem\u003eB\u003c/em\u003e, as they are connected to these nodes. As graph nodes with similar connectivity should end up in the same community, biomarkers that co-occur with \u003cem\u003eB\u003c/em\u003e in patients should also be found in \u003cem\u003eB\u003c/em\u003e\u0026rsquo;s community. PRESSnet therefore searches the communities containing positive and negative endpoint nodes to find biomarkers of those endpoints and performs pairwise searches to identify composite biomarkers from these.\u003c/p\u003e \u003cp\u003eA graph node embedding is a numerical vector representation of a node that captures information about its connectivity in the graph and, optionally, its encoded numerical data. Embeddings are useful for a variety of functions: they can be clustered like raw features to find patterns in data, used in search tools to find related entities to a given input, or fed into downstream predictive algorithms for classification or regression analysis. Over the past several years, two methods have become predominant for generating KG embeddings: random walks and matrix factorisation.\u003c/p\u003e \u003cp\u003eIn walk-based algorithms, random walks through a graph create sequences of nodes, and these sequences are used to create embeddings that capture neighbourhood information and the overall graph structure. PRESSnet currently offers the user multiple random walk\u0026ndash;based algorithms, details of which can be found below. DeepWalk, Node2Vec, and Walklets are implemented using the karateclub library and are initialised with the default parameters of \u003cem\u003en_walks\u0026thinsp;=\u0026thinsp;40\u003c/em\u003e, \u003cem\u003ewalk_length\u0026thinsp;=\u0026thinsp;40\u003c/em\u003e, but these can be altered by the user.\u003c/p\u003e \u003cp\u003e \u003cul\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eDeepWalk\u003c/b\u003e \u003csup\u003e49\u003c/sup\u003e leverages the principles behind the Skip-Gram algorithm (originally popularised for learning word embeddings in natural language processing) for the graph domain: treating random walks as sequences of nodes, it tries to maximise the log probability of neighbouring nodes appearing next to a given node using embeddings.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eNode2Vec\u003c/b\u003e \u003csup\u003e50\u003c/sup\u003e is an extension of DeepWalk that uses negative sampling and leverages search algorithms to extract random walks.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eWalklets\u003c/b\u003e \u003csup\u003e51\u003c/sup\u003e subsamples short walks using \u0026lsquo;skips\u0026rsquo; within longer walks to capture higher-order relationships between nodes.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eMetaPath2Vec\u003c/b\u003e \u003csup\u003e52\u003c/sup\u003e defines walks by node type, in a so-called \u0026lsquo;metapath\u0026rsquo;; the metapath in PRESSnet consists of [Outcome \u0026rarr; Patient \u0026rarr; Feature \u0026rarr; Patient \u0026rarr; Outcome], so an example walk might consist of \u0026lsquo;survival status: deceased\u0026rsquo; \u0026rarr; \u0026lsquo;Patient P1\u0026rsquo; \u0026rarr; \u0026lsquo;\u003cem\u003eKRAS\u003c/em\u003e mutation\u0026rsquo; \u0026rarr; \u0026lsquo;Patient P2\u0026rsquo; \u0026rarr; \u0026lsquo;survival status: alive\u0026rsquo;.\u0026rdquo; The default parameters for MetaPath2Vec in PRESSnet are \u003cem\u003en\u003c/em\u003e_walks\u0026thinsp;=\u0026thinsp;40, walk_length\u0026thinsp;=\u0026thinsp;40, \u003cem\u003en\u003c/em\u003e_epochs\u0026thinsp;=\u0026thinsp;100, learning rate\u0026thinsp;=\u0026thinsp;0.1, but these can be altered by the user.\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003c/p\u003e \u003cp\u003eMatrix factorisation techniques achieve embeddings that capture graph and neighbourhood information by decomposing matrices from the KG (typically, the Laplacian or adjacency matrix) to derive the embeddings. They typically use negative sampling as well as true relations to learn correct embeddings. PRESSnet offers matrix factorisation algorithms implemented in the Pytorch Geometric library, with default parameters of \u003cem\u003en\u003c/em\u003e_epochs\u0026thinsp;\u003cem\u003e=\u003c/em\u003e\u0026thinsp;50 and learning rate\u0026thinsp;=\u0026thinsp;0.005, which can be altered by the user.\u003c/p\u003e \u003cp\u003e \u003cul\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eDistMult\u003c/b\u003e \u003csup\u003e53\u003c/sup\u003e is a bilinear model in which a relationship between two entities is modelled as a diagonal matrix and the vectors for those entities are multiplied with this matrix element-wise to produce a representation of the overall entity-entity interaction. This is then aggregated to output a scalar score representing the likelihood of that relationship existing or not.\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eComplEx\u003c/b\u003e \u003csup\u003e54\u003c/sup\u003e is also a bilinear model that uses dot product multiplication, but unlike DistMult, which leverages real-valued embeddings, ComplEx uses complex-valued embeddings and is geared more suitably for asymmetric relationships (e.g., gene-gene regulatory data).\u003c/p\u003e \u003c/li\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eTransE\u003c/b\u003e \u003csup\u003e55\u003c/sup\u003e is a so-called translational distance model that encodes entities and relations in vectors in a shared target semantic space and uses a distance-based scoring function within this space to generate embeddings. The underlying principle is that addition of the vector of the source entity to that of the relation itself should be able to approximate the vector of the target entity.\u003c/p\u003e \u003c/li\u003e \u003c/ul\u003e \u003c/p\u003e \u003cp\u003eFor patient stratification, patient embeddings from PRESSnet can be clustered to generate potential subtypes.\u003c/p\u003e \u003cp\u003eAs embeddings are generated for every node in the graph, distance-based analysis of these embeddings is used to identify biomarkers. Nodes with similar connectivity should have numerically similar embeddings. Returning to the example of biomarker \u003cem\u003eB\u003c/em\u003e associated with outcome \u003cem\u003eO\u003c/em\u003e, this means that the node for \u003cem\u003eB\u003c/em\u003e should be closer in embedding space to the node for \u003cem\u003eO\u003c/em\u003e than another feature, \u003cem\u003eF\u003c/em\u003e, which has no association with survival. Similarly, the node for \u003cem\u003eB\u003c/em\u003e should theoretically be further from the node that represents the converse of \u003cem\u003eO\u003c/em\u003e (i.e., \u0026lsquo;survival status: alive\u0026rsquo;) than \u003cem\u003eF\u003c/em\u003e. We can thus frame biomarker discovery as a multi-objective optimisation problem in embedding space: for biomarkers of poor prognosis in terms of binary overall survival, for example, we are simultaneously minimising the distance to the embedding for \u0026lsquo;deceased\u0026rsquo; and maximising the distance to the embedding for \u0026lsquo;alive\u0026rsquo; (Fig.\u0026nbsp;\u003cspan refid=\"Fig11\" class=\"InternalRef\"\u003e6\u003c/span\u003eb).\u003c/p\u003e \u003cp\u003ePRESSnet also offers an adaptation of personalised PageRank to generate biomarkers (Fig.\u0026nbsp;\u003cspan refid=\"Fig11\" class=\"InternalRef\"\u003e6\u003c/span\u003ec). Node values for outcome and duration nodes are given weights corresponding to their directionality; nodes corresponding to longer survival and duration are weighted positively, the converse for negative outcome, and feature nodes and patient nodes are comparatively unweighted. After the adapted personalised PageRank is computed, the feature nodes with the largest positive weights are considered as potential biomarkers of good prognosis and vice versa for features with the most negative weights.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec13\" class=\"Section2\"\u003e \u003ch2\u003eOutcome predictions for unseen patients\u003c/h2\u003e \u003cp\u003eAs PRESSnet\u0026rsquo;s PageRank weights capture associations between features and endpoints such as survival, we can use these weights to generate risk scores for \u0026lsquo;unseen\u0026rsquo; patients. Training set patients (for whom survival information is known) are connected to test set patients via shared features; survival information for the training set is propagated through the graph to generate PageRank weights associated with survival outcome in the test set patients (Supplementary Fig.\u0026nbsp;9). We fed these weights into a Cox regression model to generate survival association metrics. We also repeated the process separately for binarized versions of these weights by categorising the PageRank weights into \u0026lsquo;high risk\u0026rsquo; and \u0026lsquo;low risk\u0026rsquo; by using K-means clustering with \u003cem\u003en\u003c/em\u003e\u0026thinsp;=\u0026thinsp;2 clusters.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec14\" class=\"Section2\"\u003e \u003ch2\u003eBenchmarks for patient stratification and biomarker discovery on MSK 2022 LUAD data\u003c/h2\u003e \u003cp\u003eWe assessed PRESSnet against benchmark methods for subtyping and biomarker discovery on the MSK 2022 dataset. For patient stratification, we implemented K-means clustering on (1) raw tabular data and (2) latent factors that had been discovered by MOFA+. This allowed for a direct comparison with the graph methods in supervised settings (i.e., where outcome data were included in the tabular data) and unsupervised settings. In unsupervised approaches, optimal K-number was derived by identifying the cluster number corresponding to the highest Calinski-Harabasz (CH) score, whereas in supervised approaches, it was derived by identifying the number corresponding to the highest concordance index.\u003c/p\u003e \u003cp\u003eFor biomarker discovery benchmarking, we trained a random-forest classifier to predict OS on both the raw and PRESSnet-thresholded data and extracted the features that were most important for the prediction to serve as biomarker hypotheses. For risk prediction in unseen patients, we benchmarked PRESSnet against a deep neural network (an MLPClassifier from scikit-learn, with random_state\u0026thinsp;=\u0026thinsp;1 and 500 max iterations) trained on raw data to predict binary survival outcome. Across 10 train/test splits with 50% training size, we generated c-index and hazard ratio \u003cem\u003eP\u003c/em\u003e values from Cox proportional hazards models, which were implemented using the Lifelines library in Python\u003csup\u003e56\u003c/sup\u003e with no parameter tuning, fitted on the following:\u003c/p\u003e \u003cp\u003e \u003col\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eRaw PageRank weights for test set patients only\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eBinarised PageRank weights (\u0026lsquo;high risk\u0026rsquo; vs. \u0026lsquo;low risk\u0026rsquo; categorisation using k-means clustering with \u003cem\u003en\u003c/em\u003e\u0026thinsp;=\u0026thinsp;2) for test set patients only\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eRaw prediction scores for binary OS from the deep neural network\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eBinarised prediction (output\u0026thinsp;\u0026gt;\u0026thinsp;0.5) scores for OS binary OS from the deep neural network\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003c/ol\u003e \u003c/p\u003e \u003cp\u003eWe took the average of these c-indices and \u003cem\u003eP\u003c/em\u003e values across the 10 iterations to generate our metrics in the Results.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec15\" class=\"Section2\"\u003e \u003ch2\u003eIn-built meta-analysis and robustness assessment\u003c/h2\u003e \u003cp\u003eIn PRESSnet, the user can choose whether to output results that use both, either, or neither bootstrapping and permutation testing when generating results from PageRank and community detection. We experimented with applying these to the embedding generation modules as well, but found that the time considerations involved were impractical given the training time of the complex embedding algorithms themselves. Addressing this issue can be considered a future extension of the framework alongside those mentioned in the Discussion.\u003c/p\u003e \u003cp\u003eA summary of how robustness is implemented for PageRank is presented in Supplementary Fig.\u0026nbsp;10. The user can choose to implement bootstrapping before running PageRank, at which point the PageRank algorithm is run on a user-specified number of bootstrapped graphs. The weights for each node across the different PageRanks are then averaged to produce an aggregated final weight, while statistics for the variance of these weights are used as a robustness flag to highlight those that fluctuate significantly across runs. The other robustness assessment is via permutation testing. Here, PageRank is run on a user-specified number of permuted graphs; for each node, the weights produced by the bootstrapping analysis are assessed against all the weights from the permuted runs to determine whether we can reject the null hypothesis that the original weights fall within the distribution of the permuted (i.e., randomly generated) weights. Features that fail permutation testing and/or fall outside the acceptable variance of weights are then filtered out before biomarker selection. When biomarkers are combined for composite signatures, a final set of log-rank testing is performed on multiple random samples of the data to assess the stability of significant signatures; the user can set the sampling proportion and number of samples.\u003c/p\u003e \u003cp\u003ePRESSnet offers the users multiple methods for performing bootstrapping and permutation testing. The user can resample from (1) all nodes in the graph, (2) all edges in the graph, or (3) patient nodes only in the graph to produce a bootstrapped graph. For permutation testing, the user can randomly shuffle the identity of the target nodes from either (1) all the edges in the graph or (2) only the patient endpoint edges. Examples of how bootstrapped and permuted graphs work with PageRank are shown in Supplementary Fig.\u0026nbsp;10.\u003c/p\u003e \u003cp\u003eFor community detection, variability in results can come from both the composition of the graph and the model parameters, particularly the initialisation (i.e., seed or random state) of the chosen algorithm. As such, if the user chooses to enable robustness assessment, community detection is run for both (1) the original graph \u003cem\u003en\u003c/em\u003e times with the seed initialisation ranging from 1 to \u003cem\u003en\u003c/em\u003e, and (2) the \u003cem\u003en\u003c/em\u003e bootstrapped graphs, where the algorithm is run once per graph with a fixed seed. For stratification, the resulting 2\u003cem\u003en\u003c/em\u003e sets of communities are concatenated and one-hot encoded to generate inputs for a final metaclustering, which is carried out using hierarchical clustering implemented in scikit-learn. The optimal number of communities is assessed using a weighted Calinski-Harabasz (CH) score approach, \u003cem\u003e0.7 * CH1\u0026thinsp;+\u0026thinsp;0.3 * CH2\u003c/em\u003e, where \u003cem\u003eCH1\u003c/em\u003e is calculated by comparing cluster labels to the underlying patient data, and \u003cem\u003eCH2\u003c/em\u003e by comparing them to the \u003cem\u003e2n\u003c/em\u003e sets of community labels. This attempts to balance biological meaningfulness with algorithmic stability.\u003c/p\u003e \u003cp\u003eFor biomarker discovery, the 2\u003cem\u003en\u003c/em\u003e clusterings capture (1) the features that most frequently co-occur in communities with endpoint nodes across the iterations and (2) the features that most frequently co-occur with the features from (1) across the iterations. As such, the features from (1) are evaluated as potential individual biomarkers of the endpoint, and for each feature, a user-defined number of nearest neighbours (default\u0026thinsp;=\u0026thinsp;10) are considered alongside the feature as potential paired composite signatures.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec16\" class=\"Section2\"\u003e \u003ch2\u003eDataset preparation\u003c/h2\u003e \u003cp\u003eThe MSK 2022 data were downloaded from the publicly accessible cBioPortal database.\u003csup\u003e57\u003c/sup\u003e The TME prior knowledge graph that we built to supplement the original data was also constructed using public data and consists of the following edges: (1) TME-specific \u0026lsquo;Biological Process\u0026rsquo; nodes linked to participating genes, curated with HetionetDB,\u003csup\u003e58\u003c/sup\u003e CellPhoneDB,\u003csup\u003e59\u003c/sup\u003e and Gene Ontology; (2) TME-specific cell type nodes linked to genes involved in the cell signature, curated using CellMarker2.0;\u003csup\u003e60\u003c/sup\u003e and (3) gene-to-gene edges modelling gene interactions between wild-type gene nodes using HetionetDB and CellPhoneDB.\u003c/p\u003e \u003cp\u003eWe also downloaded the TCGA and Beat AML datasets from cBioPortal. Because the gene set for expression data is large (e.g., \u003cem\u003en\u003c/em\u003e\u0026thinsp;=\u0026thinsp;16383 for TCGA), for preliminary feature selection we identified a set of genes with non-zero confidence scores for links to AML according to Open Targets.\u003csup\u003e61\u003c/sup\u003e We then processed the data to identify patients with complete survival and treatment data, retrieved overlapping features between the two datasets, and used these as inputs for our framework. Expression data were taken as the z-score of the log\u003csub\u003e2\u003c/sub\u003e transcripts per million for each gene to combat batch effects. This resulted in 834 features for expression, 463 for mutations, and 2 for clinical (TMB and age), in 167 patients from TCGA and 603 from Beat AML. For generating results on AML data, we found that age as a feature dominated results in terms of composite biomarker creation (itself a true positive), so we implemented the framework with and without its inclusion, focusing in our Results section on outputs from the latter.\u003c/p\u003e \u003c/div\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eCompeting interests\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eJ.C.-S., S.S., I.K., D.S., M.G., S.J., D.B., B.S., E.J., and K.B. are or were employees of AstraZeneca when this work was performed and may have stock ownership, options, and/or interests in the company.\u003c/p\u003e\n\u003cp\u003eE.A.C. is an employee of The Bridge and receives a salary from The Bridge, but her work is contracted out to AstraZeneca.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eContributions\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eJ.C.-S., conceptualization, methodology, software, validation, data curation, visualization, analysis, investigation, writing (original draft, reviewing and editing); S.S., software, validation, data curation, visualisation, analysis, investigation, writing (original draft, reviewing and editing); I.K., conceptualization, methodology, software, investigation; D.S., methodology, investigation; M.G. and S.J., software, investigation; E.A.C., validation, writing (reviewing); D.B., investigation; B.S., writing (reviewing); E.J., and K.B., conceptualization, methodology, supervision, investigation, writing (reviewing and editing).\u003c/p\u003e\n\u003cp\u003eDeborah Shuman (Scientific Publications, Oncology R\u0026amp;D, AstraZeneca, Gaithersburg, MD, USA) provided editorial and graphics support.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCorresponding Authors\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eCorrespondence to Jake Cohen-Setton.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCode Availability\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe code used in this study is available at 10.5281/zenodo.17100620.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eData Availability\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe specific versions of the datasets used in this study are available publicly at https://www.cbioportal.org/datasets. These are as follows:\u003c/p\u003e\n\u003cp\u003eMSK IO NSCLC: https://www.cbioportal.org/study/summary?id=lung_msk_mind_2020\u003c/p\u003e\n\u003cp\u003eBeat AML: https://www.cbioportal.org/study/summary?id=aml_ohsu_2022\u003c/p\u003e\n\u003cp\u003eTCGA AML: https://www.cbioportal.org/study/summary?id=laml_tcga_pan_can_atlas_2018\u003c/p\u003e\n\u003cp\u003eThe cBioPortal website also provides URL links to the corresponding publications, citations for which can be found in References.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eMarkey, N.\u003cem\u003e et al.\u003c/em\u003e Clinical trials are becoming more complex: a machine learning analysis of data from over 16,000 trials. \u003cem\u003eSci Rep\u003c/em\u003e \u003cstrong\u003e14\u003c/strong\u003e, 3514 (2024).\u003c/li\u003e\n\u003cli\u003eKing, E. A., Davis, J. W. \u0026amp; Degner, J. F. Are drug targets with genetic support twice as likely to be approved? Revised estimates of the impact of genetic support for drug mechanisms on the probability of drug approval. \u003cem\u003ePLoS Genet\u003c/em\u003e \u003cstrong\u003e15\u003c/strong\u003e, e1008489 (2019).\u003c/li\u003e\n\u003cli\u003eMorgan, P.\u003cem\u003e et al.\u003c/em\u003e Impact of a five-dimensional framework on R\u0026amp;D productivity at AstraZeneca. \u003cem\u003eNat Rev Drug Discov\u003c/em\u003e \u003cstrong\u003e17\u003c/strong\u003e, 167-181 (2018).\u003c/li\u003e\n\u003cli\u003eHu, C. \u0026amp; Dignam, J. J. Biomarker-driven oncology clinical trials: key design elements, types, features, and practical considerations. \u003cem\u003eJCO Precis Oncol\u003c/em\u003e \u003cstrong\u003e3\u003c/strong\u003e (2019).\u003c/li\u003e\n\u003cli\u003eGreig, S. L. Osimertinib: first global approval. \u003cem\u003eDrugs\u003c/em\u003e \u003cstrong\u003e76\u003c/strong\u003e, 263-273 (2016).\u003c/li\u003e\n\u003cli\u003eKeam, S. J. Trastuzumab deruxtecan: first approval. \u003cem\u003eDrugs\u003c/em\u003e \u003cstrong\u003e80\u003c/strong\u003e, 501-508 (2020).\u003c/li\u003e\n\u003cli\u003eGaron, E. B.\u003cem\u003e et al.\u003c/em\u003e Pembrolizumab for the treatment of non-small-cell lung cancer. \u003cem\u003eN Engl J Med\u003c/em\u003e \u003cstrong\u003e372\u003c/strong\u003e, 2018-2028 (2015).\u003c/li\u003e\n\u003cli\u003eCortesi, L., Rugo, H. S. \u0026amp; Jackisch, C. An overview of PARP inhibitors for the treatment of breast cancer. \u003cem\u003eTarget Oncol\u003c/em\u003e \u003cstrong\u003e16\u003c/strong\u003e, 255-282 (2021).\u003c/li\u003e\n\u003cli\u003eChang, L., Chang, M., Chang, H. M. \u0026amp; Chang, F. Microsatellite instability: a predictive biomarker for cancer immunotherapy. \u003cem\u003eAppl Immunohistochem Mol Morphol\u003c/em\u003e \u003cstrong\u003e26\u003c/strong\u003e, e15-e21 (2018).\u003c/li\u003e\n\u003cli\u003eSun, T. H., Wang, C. C., Wu, Y. L., Hsu, K. C. \u0026amp; Lee, T. H. Machine learning approaches for biomarker discovery to predict large-artery atherosclerosis. \u003cem\u003eSci Rep\u003c/em\u003e \u003cstrong\u003e13\u003c/strong\u003e, 15139 (2023).\u003c/li\u003e\n\u003cli\u003eMcLernon, D. J.\u003cem\u003e et al.\u003c/em\u003e Assessing performance and clinical usefulness in prediction models with survival outcomes: practical guidance for Cox proportional hazards models. \u003cem\u003eAnn Intern Med\u003c/em\u003e \u003cstrong\u003e176\u003c/strong\u003e, 105-114 (2023).\u003c/li\u003e\n\u003cli\u003eMcNamara, M. G.\u003cem\u003e et al.\u003c/em\u003e 89PD: prognostic and predictive impact of high tumor mutation burden (TMB) in solid tumors: a systematic review and meta-analysis. \u003cem\u003eAnn Oncol\u003c/em\u003e \u003cstrong\u003e30\u003c/strong\u003e, v25\u0026ndash;v54 (2019).\u003c/li\u003e\n\u003cli\u003eWang, B.\u003cem\u003e et al.\u003c/em\u003e Similarity network fusion for aggregating data types on a genomic scale. \u003cem\u003eNat Methods\u003c/em\u003e \u003cstrong\u003e11\u003c/strong\u003e, 333-337 (2014).\u003c/li\u003e\n\u003cli\u003eHofree, M., Shen, J. P., Carter, H., Gross, A. \u0026amp; Ideker, T. Network-based stratification of tumor mutations. \u003cem\u003eNat Methods\u003c/em\u003e \u003cstrong\u003e10\u003c/strong\u003e, 1108-1115 (2013).\u003c/li\u003e\n\u003cli\u003eKong, J.\u003cem\u003e et al.\u003c/em\u003e Network-based machine learning approach to predict immunotherapy response in cancer patients. \u003cem\u003eNat Commun\u003c/em\u003e \u003cstrong\u003e13\u003c/strong\u003e, 3703 (2022).\u003c/li\u003e\n\u003cli\u003eShanthamallu, U. S.\u003cem\u003e et al.\u003c/em\u003e A network-based framework to discover treatment-response-predicting biomarkers for complex diseases. \u003cem\u003eJ Mol Diagn\u003c/em\u003e \u003cstrong\u003e26\u003c/strong\u003e, 917-930 (2024).\u003c/li\u003e\n\u003cli\u003eLin, C., Sun, G. J., Bulusu, K. C., Dry, J. R. \u0026amp; Hernandez, M. Graph neural networks including sparse interpretability. arXiv (2020). DOI: 10.48550/arXiv.2007.00119. https://arxiv.org/abs/2007.00119. Accessed August 15, 2025.\u003c/li\u003e\n\u003cli\u003eGogleva, A.\u003cem\u003e et al.\u003c/em\u003e Knowledge graph-based recommendation framework identifies drivers of resistance in EGFR mutant non-small cell lung cancer. \u003cem\u003eNat Commun\u003c/em\u003e \u003cstrong\u003e13\u003c/strong\u003e, 1667 (2022).\u003c/li\u003e\n\u003cli\u003eChandak, P., Huang, K. \u0026amp; Zitnik, M. Building a knowledge graph to enable precision medicine. \u003cem\u003eSci Data\u003c/em\u003e \u003cstrong\u003e10\u003c/strong\u003e, 67 (2023).\u003c/li\u003e\n\u003cli\u003eHuang, K.\u003cem\u003e et al.\u003c/em\u003e A foundation model for clinician-centered drug repurposing. \u003cem\u003eNat Med\u003c/em\u003e \u003cstrong\u003e30\u003c/strong\u003e, 3601-3613 (2024).\u003c/li\u003e\n\u003cli\u003eBrbić, M., Yasunaga, M., Agarwal, P. \u0026amp; Leskovec, J. Predicting drug outcome of population via clinical knowledge graph. medRxiv (2024). DOI: 10.1101/2024.03.06.24303800. https://www.medrxiv.org/content/10.1101/2024.03.06.24303800v2. Accessed February 17, 2025.\u003c/li\u003e\n\u003cli\u003eGorostiola Gonz\u0026aacute;lez, M., Ijzerman, A. P. \u0026amp; van Westen, G. J. P. A patient-centric knowledge graph approach to prioritize mutants for selective anti-cancer targeting. bioRxiv (2024). DOI: 10.1101/2024.09.29.615658. https://www.biorxiv.org/content/10.1101/2024.09.29.615658v1. Accessed February 17, 2025.\u003c/li\u003e\n\u003cli\u003eFang, C.\u003cem\u003e et al.\u003c/em\u003e Integrating knowledge graphs into machine learning models for survival prediction and biomarker discovery in patients with non-small-cell lung cancer. \u003cem\u003eJ Transl Med\u003c/em\u003e \u003cstrong\u003e22\u003c/strong\u003e, 726 (2024).\u003c/li\u003e\n\u003cli\u003eVanguri, R. S.\u003cem\u003e et al.\u003c/em\u003e Multimodal integration of radiology, pathology and genomics for prediction of response to PD-(L)1 blockade in patients with non-small cell lung cancer. \u003cem\u003eNat Cancer\u003c/em\u003e \u003cstrong\u003e3\u003c/strong\u003e, 1151-1164 (2022).\u003c/li\u003e\n\u003cli\u003eKhwaja, R. M. \u0026amp; Chu, Q. S. Present and emerging biomarkers in immunotherapy for metastatic non-small cell lung cancer: a review. \u003cem\u003eCurr Oncol\u003c/em\u003e \u003cstrong\u003e29\u003c/strong\u003e, 479-489 (2022).\u003c/li\u003e\n\u003cli\u003eXiao, Y.\u003cem\u003e et al.\u003c/em\u003e Recent progress in targeted therapy for non-small cell lung cancer. \u003cem\u003eFront Pharmacol\u003c/em\u003e \u003cstrong\u003e14\u003c/strong\u003e, 1125547 (2023).\u003c/li\u003e\n\u003cli\u003eStares, M.\u003cem\u003e et al.\u003c/em\u003e Hypoalbuminaemia as a prognostic biomarker of first-line treatment resistance in metastatic non-small cell lung cancer. \u003cem\u003eFront Nutr\u003c/em\u003e \u003cstrong\u003e8\u003c/strong\u003e, 734735 (2021).\u003c/li\u003e\n\u003cli\u003eYildirim, M., Yildiz, M., Duman, E., Goktas, S. \u0026amp; Kaya, V. Prognostic importance of the nutritional status and systemic inflammatory response in non-small cell lung cancer. \u003cem\u003eJ BUON\u003c/em\u003e \u003cstrong\u003e18\u003c/strong\u003e, 728-732 (2013).\u003c/li\u003e\n\u003cli\u003eYu, Y., Qian, L. \u0026amp; Cui, J. Value of neutrophil-to-lymphocyte ratio for predicting lung cancer prognosis: a meta-analysis of 7,219 patients. \u003cem\u003eMol Clin Oncol\u003c/em\u003e \u003cstrong\u003e7\u003c/strong\u003e, 498-506 (2017).\u003c/li\u003e\n\u003cli\u003eChen, T. \u0026amp; Guestrin, C. Xgboost: a scalable tree boosting system. Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, San Franciso, CA, August 13\u0026ndash;17, 2016. https://dl.acm.org/doi/10.1145/2939672.2939785. Accessed August 15, 2025.\u003c/li\u003e\n\u003cli\u003eZhao, D., Jiang, M., Zhang, X. \u0026amp; Hou, H. The role of RICTOR amplification in targeted therapy and drug resistance. \u003cem\u003eMol Med\u003c/em\u003e \u003cstrong\u003e26\u003c/strong\u003e, 20 (2020).\u003c/li\u003e\n\u003cli\u003eYuan, S., Liu, Y., Till, B., Song, Y. \u0026amp; Wang, Z. Pretreatment peripheral B cells are associated with tumor response to anti-PD-1-based immunotherapy. \u003cem\u003eFront Immunol\u003c/em\u003e \u003cstrong\u003e11\u003c/strong\u003e, 563653 (2020).\u003c/li\u003e\n\u003cli\u003eLiu, R.\u003cem\u003e et al.\u003c/em\u003e ABL1 is a prognostic marker and associated with immune infiltration in hepatocellular carcinoma. \u003cem\u003eJ Oncol\u003c/em\u003e \u003cstrong\u003e2021\u003c/strong\u003e, 1379706 (2021).\u003c/li\u003e\n\u003cli\u003eLiu, L.\u003cem\u003e et al.\u003c/em\u003e BCL-2 expression promotes immunosuppression in chronic lymphocytic leukemia by enhancing regulatory T cell differentiation and cytotoxic T cell exhaustion. \u003cem\u003eMol Cancer\u003c/em\u003e \u003cstrong\u003e21\u003c/strong\u003e, 59 (2022).\u003c/li\u003e\n\u003cli\u003eLiu, R.\u003cem\u003e et al.\u003c/em\u003e Infiltration of apoptotic M2 macrophage subpopulation is negatively correlated with the immunotherapy response in colorectal cancer. \u003cem\u003eInt J Mol Sci\u003c/em\u003e \u003cstrong\u003e23\u003c/strong\u003e (2022).\u003c/li\u003e\n\u003cli\u003eZhou, F.\u003cem\u003e et al.\u003c/em\u003e CTHRC1 is a prognostic biomarker and correlated with immune infiltrates in kidney renal papillary cell carcinoma and kidney renal clear cell carcinoma. \u003cem\u003eFront Oncol\u003c/em\u003e \u003cstrong\u003e10\u003c/strong\u003e, 570819 (2020).\u003c/li\u003e\n\u003cli\u003eShigeta, K.\u003cem\u003e et al.\u003c/em\u003e Dual programmed death receptor-1 and vascular endothelial growth factor receptor-2 blockade promotes vascular normalization and enhances antitumor immune responses in hepatocellular carcinoma. \u003cem\u003eHepatology\u003c/em\u003e \u003cstrong\u003e71\u003c/strong\u003e, 1247-1261 (2020).\u003c/li\u003e\n\u003cli\u003eWang, J., Liu, N., Jiang, H., Li, Q. \u0026amp; Xing, D. Reactive oxygen species in anticancer immunity: a double-edged sword. \u003cem\u003eFront Bioeng Biotechnol\u003c/em\u003e \u003cstrong\u003e9\u003c/strong\u003e, 784612 (2021).\u003c/li\u003e\n\u003cli\u003eBottomly, D.\u003cem\u003e et al.\u003c/em\u003e Integrative analysis of drug response and clinical outcome in acute myeloid leukemia. \u003cem\u003eCancer Cell\u003c/em\u003e \u003cstrong\u003e40\u003c/strong\u003e, 850-864 e859 (2022).\u003c/li\u003e\n\u003cli\u003eNational Cancer Institute \u0026amp; Center for Cancer Genomics. The Cancer Genome Atlas Program (TCGA). https://www.cancer.gov/ccg/research/genome-sequencing/tcga. Accessed February 17, 2025.\u003c/li\u003e\n\u003cli\u003eSenn, S. Dichotomania: an obsessive compulsive disorder that is badly affecting the quality of analysis of pharmaceutical trials. (2005). https://www.semanticscholar.org/paper/Dichotomania%3A-An-Obsessive-Compulsive-Disorder-that-Senn/8c390cb90465ad1f08b86cee52474859e4dffd77. Accessed February 18, 2025.\u003c/li\u003e\n\u003cli\u003eZhang, C., Song, D., Huang, C., Swami, A. \u0026amp; Chawla, N. V. Heterogeneous graph neural network. 25th ACM SIGKDD International Conference on Knowledge Discovery \u0026amp; Data Mining, Anchorage, AK, August 4-8, 2019. https://dl.acm.org/doi/10.1145/3292500.3330961. Accessed February 18, 2025.\u003c/li\u003e\n\u003cli\u003eValous, N. A., Popp, F., Zornig, I., Jager, D. \u0026amp; Charoentong, P. Graph machine learning for integrated multi-omics analysis. \u003cem\u003eBr J Cancer\u003c/em\u003e \u003cstrong\u003e131\u003c/strong\u003e, 205-211 (2024).\u003c/li\u003e\n\u003cli\u003eChen, Y. \u0026amp; Zou, J. GenePT: a simple but effective foundation model for genes and cells built from ChatGPT. \u003cem\u003ebioRxiv\u003c/em\u003e (2024).\u003c/li\u003e\n\u003cli\u003eAshburner, M.\u003cem\u003e et al.\u003c/em\u003e Gene ontology: tool for the unification of biology. The Gene Ontology Consortium. \u003cem\u003eNat Genet\u003c/em\u003e \u003cstrong\u003e25\u003c/strong\u003e, 25-29 (2000).\u003c/li\u003e\n\u003cli\u003eGene Ontology, C.\u003cem\u003e et al.\u003c/em\u003e The Gene Ontology knowledgebase in 2023. \u003cem\u003eGenetics\u003c/em\u003e \u003cstrong\u003e224\u003c/strong\u003e (2023).\u003c/li\u003e\n\u003cli\u003eDe Meo, P., Ferrara, E., Fiumara, G. \u0026amp; Provetti, A. Generalized Louvain method for community detection in large networks. 11th International Conference on Intelligent Systems Design and Applications, November 22-24, 2011. https://ieeexplore.ieee.org/document/6121636/authors#authors. Accessed February 17, 2025.\u003c/li\u003e\n\u003cli\u003eLi, P.-Z., Huang, L., Wang, C.-D. \u0026amp; Lai, J.-H. EdMot: an edge enhancement approach for motif-aware community detection. (2019). DOI: https://arxiv.org/abs/1906.04560. https://arxiv.org/abs/1906.04560. Accessed February 18, 2025.\u003c/li\u003e\n\u003cli\u003ePerozzi, B., Al-Rfou, R. \u0026amp; Skiena, S. DeepWalk: online learning of social representations. 20th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, New York, NY, August 24-27, 2014. https://dl.acm.org/doi/10.1145/2623330.2623732. Accessed February 18, 2025.\u003c/li\u003e\n\u003cli\u003eGrover, A. \u0026amp; Leskovec, J. node2vec: scalable feature learning for networks. 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, San Francisco, CA, August 13-17, 2016. https://dl.acm.org/doi/10.1145/2939672.2939754. Accessed February 18, 2025.\u003c/li\u003e\n\u003cli\u003ePerozzi, B., Kulkarni, V., Chen, H. \u0026amp; Skiena, S. Don\u0026apos;t walk, skip! Online learning of multi-scale network embeddings. ASONAM \u0026apos;17: Advances in Social Networks Analysis and Mining, Sydney, Australia, July 31-August 3, 2017. https://dl.acm.org/doi/10.1145/3110025.3110086. Accessed February 18, 2025.\u003c/li\u003e\n\u003cli\u003eDong, Y., Chawla, N. V. \u0026amp; Swami, A. metapath2vec: Scalable representation learning for heterogeneous networks. 23rd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, Halifax, NS, Canada, August 13-17, 2017. https://dl.acm.org/doi/10.1145/3097983.3098036. Accessed February 18, 2025.\u003c/li\u003e\n\u003cli\u003eYang, B., Yih, W., He, X., Gao, J. \u0026amp; Deng, L. Embedding entities and relations for learning and inference in knowledge bases. arXiv (2015). DOI: 10.48550/arXiv.1412.6575. https://arxiv.org/abs/1412.6575. Accessed Febuary 18, 2025.\u003c/li\u003e\n\u003cli\u003eTrouillon, T., Welbl, J., Riedel, S., Gaussier, E. \u0026amp; Bouchard, G. Complex embeddings for simple link prediction. Proceedings of the 33rd International Conference on Machine Learning, New York, NY. https://proceedings.mlr.press/v48/trouillon16.html. Accessed February 18, 2025.\u003c/li\u003e\n\u003cli\u003eBordes, A., Usunier, N., Garcia-Duran, A., Weston, J. \u0026amp; Yakhnenko, O. Translating embeddings for modeling multi-relational data. In Advances in Neural Information Processing Systems 26. (eds C. J. Burges\u003cem\u003e et al.\u003c/em\u003e).\u003c/li\u003e\n\u003cli\u003eDavidson-Pilon, D. Lifelines: survival analysis in Python. \u003cem\u003eJ Open Source Softw\u003c/em\u003e \u003cstrong\u003e4\u003c/strong\u003e, 1317 (2019).\u003c/li\u003e\n\u003cli\u003eCerami, E.\u003cem\u003e et al.\u003c/em\u003e The cBio cancer genomics portal: an open platform for exploring multidimensional cancer genomics data. \u003cem\u003eCancer Discov\u003c/em\u003e \u003cstrong\u003e2\u003c/strong\u003e, 401-404 (2012).\u003c/li\u003e\n\u003cli\u003eHimmelstein, D. S.\u003cem\u003e et al.\u003c/em\u003e Systematic integration of biomedical knowledge prioritizes drugs for repurposing. \u003cem\u003eElife\u003c/em\u003e \u003cstrong\u003e6\u003c/strong\u003e, e26726 (2017).\u003c/li\u003e\n\u003cli\u003eEfremova, M., Vento-Tormo, M., Teichmann, S. A. \u0026amp; Vento-Tormo, R. CellPhoneDB: inferring cell-cell communication from combined expression of multi-subunit ligand-receptor complexes. \u003cem\u003eNat Protoc\u003c/em\u003e \u003cstrong\u003e15\u003c/strong\u003e, 1484-1506 (2020).\u003c/li\u003e\n\u003cli\u003eHu, C.\u003cem\u003e et al.\u003c/em\u003e CellMarker 2.0: an updated database of manually curated cell markers in human/mouse and web tools based on scRNA-seq data. \u003cem\u003eNucleic Acids Res\u003c/em\u003e \u003cstrong\u003e51\u003c/strong\u003e, D870-D876 (2023).\u003c/li\u003e\n\u003cli\u003eOchoa, D.\u003cem\u003e et al.\u003c/em\u003e The next-generation Open Targets Platform: reimagined, redesigned, rebuilt. \u003cem\u003eNucleic Acids Res\u003c/em\u003e \u003cstrong\u003e51\u003c/strong\u003e, D1353-D1359 (2023).\u003c/li\u003e\n\u003cli\u003eZhang, J.\u003cem\u003e et al.\u003c/em\u003e High expression levels of SMAD3 and SMAD7 at diagnosis predict poor prognosis in acute myeloid leukemia patients undergoing chemotherapy. \u003cem\u003eCancer Gene Ther\u003c/em\u003e \u003cstrong\u003e26\u003c/strong\u003e, 119-127 (2019).\u003c/li\u003e\n\u003cli\u003eChen, J.\u003cem\u003e et al.\u003c/em\u003e Homoharringtonine targets Smad3 and TGF-beta pathway to inhibit the proliferation of acute myeloid leukemia cells. \u003cem\u003eOncotarget\u003c/em\u003e \u003cstrong\u003e8\u003c/strong\u003e, 40318-40326 (2017).\u003c/li\u003e\n\u003cli\u003eZhang, Y.\u003cem\u003e et al.\u003c/em\u003e Galectin-9 and PSMB8 overexpression predict unfavorable prognosis in patients with AML. \u003cem\u003eJ Cancer\u003c/em\u003e \u003cstrong\u003e12\u003c/strong\u003e, 4257-4263 (2021).\u003c/li\u003e\n\u003cli\u003ePark, S.\u003cem\u003e et al.\u003c/em\u003e Calreticulin mRNA expression and clinicopathological characteristics in acute myeloid leukemia. \u003cem\u003eCancer Genet\u003c/em\u003e \u003cstrong\u003e208\u003c/strong\u003e, 630-635 (2015).\u003c/li\u003e\n\u003cli\u003eAly, M.\u003cem\u003e et al.\u003c/em\u003e Distinct clinical and biological implications of CUX1 in myeloid neoplasms. \u003cem\u003eBlood Adv\u003c/em\u003e \u003cstrong\u003e3\u003c/strong\u003e, 2164-2178 (2019).\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"nature-portfolio","isNatureJournal":true,"hasQc":false,"allowDirectSubmit":false,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"","title":"Nature Portfolio","twitterHandle":"","acdcEnabled":false,"dfaEnabled":false,"editorialSystem":"ejp","reportingPortfolio":"","inReviewEnabled":true,"inReviewRevisionsEnabled":false},"keywords":"","lastPublishedDoi":"10.21203/rs.3.rs-8289747/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-8289747/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eHarnessing multiomic data for patient stratification and biomarker discovery is vital for effective personalized medicine. However, as the volume and heterogeneity of data increases, so do the challenges of data integration and explainable hypotheses. To address this, we present PRESSnet (Patient REcommendation via Stratification and Selection using networks), an end-to-end framework leveraging multimodal patient knowledge graphs (KGs) for stratification and biomarker discovery. PRESSnet incorporates graph artificial intelligence and network algorithms in scalable, flexible analysis pipelines that can integrate underlying multiomic patient features with prior knowledge such as curated gene pathway data. Applied to patients from two different cancer types, PRESSnet generates explainable stratification hypotheses and captured known survival biomarkers as well as novel composite signatures that comparatively increased statistically significant survival separation compared to univariate markers. These biomarkers were validated for their translatability to unseen patients within cohorts in IO-treated NSCLC patients (MSK 2022) and across independent datasets in AML patients (TCGA and Beat AML). PRESSnet compared favourably to benchmark models such as MOFA\u0026thinsp;+\u0026thinsp;for stratification and Random Forests for biomarker generation and survival risk classification. We also demonstrate that PRESSnet\u0026rsquo;s ability to model prior knowledge can improve patient survival prediction, including in small datasets, and offer context-relevant insights into signalling pathways and regulatory networks involved in therapy resistance. PRESSnet is provided as a lightweight, adaptable framework for the scientific community to inform research into patient selection, asset positioning and trial design.\u003c/p\u003e","manuscriptTitle":"PRESSnet: a novel framework for patient stratification and biomarker discovery using clinical knowledge graphs","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-12-15 07:50:08","doi":"10.21203/rs.3.rs-8289747/v1","editorialEvents":[],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"nature-communications","isNatureJournal":true,"hasQc":false,"allowDirectSubmit":false,"externalIdentity":"NCOMMS","sideBox":"Learn more about [Nature Communications](http://www.nature.com/ncomms/)","snPcode":"","submissionUrl":"https://mts-ncomms.nature.com/","title":"Nature Communications","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"ejp","reportingPortfolio":"Nature Communications","inReviewEnabled":true,"inReviewRevisionsEnabled":false}}],"origin":"","ownerIdentity":"844a5473-d751-4d5f-921b-2a2bffc8632e","owner":[],"postedDate":"December 15th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[{"id":59565261,"name":"Biological sciences/Computational biology and bioinformatics/Machine learning"},{"id":59565262,"name":"Biological sciences/Drug discovery/Biomarkers"}],"tags":[],"updatedAt":"2026-01-19T15:43:08+00:00","versionOfRecord":[],"versionCreatedAt":"2025-12-15 07:50:08","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-8289747","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-8289747","identity":"rs-8289747","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}