Modern genomic sequencing can identify thousands of mutations in a single tumor, but most are passenger mutations -- bystander alterations that accumulate by chance and have no effect on tumor behavior. The biologically critical alterations are driver mutations: the small subset that actually promotes cancer growth, treatment resistance, or metastasis. Identifying true drivers from the background noise is one of the central unsolved problems in cancer genomics.
In castration-resistant prostate cancer (CRPC), this challenge is particularly acute. The tumor has already survived androgen deprivation therapy and often multiple other treatments, selecting for complex combinations of genomic alterations. Known resistance-associated changes include RB1/TP53 loss (in 25-60% of metastatic CRPC), PTEN deletion (hyperactivating PI3K-AKT signaling in up to 60%), AR amplification (approximately 60% of metastatic tumors), DNA repair defects (BRCA1/2, ATM), and transcriptional regulator changes (FOXA1, TMPRSS2-ERG).
Current genomic models treat mutations as independent variables -- asking whether gene A is mutated, and separately whether gene B is mutated. This ignores epistatic interactions: the phenomenon where two mutations together produce an effect that neither alone would create. The combination of PTEN loss and AR amplification, for example, may drive resistance far more powerfully than either event alone, but linear models miss this relationship entirely.
Beyond statistical limitations, many clinically important mutations are low-frequency or subclonal -- present in only a subset of tumor cells and detectable only when sequencing is performed with enough depth and the right analytical tools. Bulk DNA sequencing often misses these events entirely, especially if they are only present in a small fraction of cells or are expressed at low levels.
OncoVar is a computational pipeline developed to address the core limitation of standard genomic variant calling: it requires evidence from both DNA and RNA sequencing before accepting a mutation as a candidate driver. If a mutation is present in the DNA but not being actively transcribed into RNA, it is unlikely to be functionally relevant. By requiring transcriptional support, OncoVar filters out a large fraction of irrelevant background mutations.
The pipeline uses well-validated variant calling tools (Mutect2 and HaplotypeCaller for DNA; HaplotypeCaller via the nf-core/rnavar pipeline for RNA) and then applies a stringent consensus requirement: a mutation is retained only if it was detected in at least 80% of RNA sequencing replicates for a given resistant cell line. This ensures that only robustly and consistently expressed alterations -- not technical noise -- survive the filtering.
A key innovation is the integration of AlphaMissense -- an AI-based pathogenicity scoring tool built on DeepMind's AlphaFold protein structure prediction platform. AlphaMissense evaluates each candidate mutation in the context of the protein's three-dimensional structure and evolutionary conservation across species. Mutations at conserved structural positions are much more likely to alter protein function than mutations in flexible, non-conserved regions. This structural-evolutionary scoring layer adds a layer of functional plausibility on top of the genomic and transcriptomic evidence.
To identify mutations specific to each stage of resistance, the researchers applied OncoVar to an isogenic cell line progression model: a series of cell lines all derived from the same parent (LNCaP, representing castration-sensitive prostate cancer), with each derivative representing a distinct resistance stage -- ADT-resistant (LNCaP-abl), AR-variant-driven (LNCaP95, expressing the AR-V7 splice variant), and bone-metastatic CRPC (C4-2B). Subtracting the parent cell line's mutations from each derivative isolated the mutations that appeared specifically during resistance development.
Applying OncoVar across the three resistance cell lines produced three distinct stage-specific driver mutation signatures: one enriched in the ADT-resistant CRPC model, one in the AR-variant-driven CRPC model, and one in the metastatic CRPC model. The number of acquired driver mutations increased from CSPC through CRPC-ARv7, then dropped sharply in the bone-metastatic model -- a pattern consistent with a metastatic bottleneck that selects a refined, highly potent subset of alterations.
To determine whether these cell-line-derived signatures reflect real human disease, the researchers mapped all three signatures onto the genomes of 404 metastatic CRPC patients from the SU2C clinical cohort. All three signatures were recurrently mutated across patient tumors, confirming that the isogenic cell model captured genuine biological patterns of clinical resistance rather than laboratory artifacts.
Prognostic analysis using Kaplan-Meier survival curves showed that patients carrying a high burden of mCRPC-signature driver mutations had significantly reduced overall survival (p less than 0.05). The CRPC and CRPC-ARv7 signatures showed trends in the same direction but did not reach statistical significance individually. This finding suggests the metastatic signature captures the most clinically lethal alterations.
The functional annotation of prioritized mutations revealed that missense mutations formed the largest class, followed by stop-gain (premature termination) and splice-site alterations. These are the mutation types most likely to fundamentally alter protein function, providing biological plausibility for their role as drivers. The mutations were distributed broadly across the genome rather than clustered at known cancer hotspots, highlighting the value of an unbiased discovery approach.
ATHENA (Attention-based Therapeutic Network Analyzer) is a deep learning model specifically designed to learn how combinations of driver mutations interact to influence patient outcomes. It takes the three OncoVar-derived driver signatures as inputs and uses a multi-branch architecture with one branch per signature. Each branch contains a gated attention module that learns to weight the most important mutations within that signature, dynamically emphasizing the ones most relevant to prognosis in each patient context.
Training proceeded in two phases: pretraining on the large MSK IMPACT cohort (which covers thousands of cancer patients profiled with a targeted gene panel) to learn broadly applicable mutation-outcome patterns, followed by fine-tuning on the smaller SU2C metastatic CRPC cohort to refine performance specifically for advanced prostate cancer. This two-phase strategy allows the model to benefit from the statistical power of large general datasets while retaining specificity for the biology of interest.
A critical technical challenge was that different genomic datasets use different sequencing panels, leaving some genes profiled in one cohort but not another. ATHENA explicitly encodes a missingness flag for each gene in each patient's data -- allowing the model to distinguish genes that were tested and found wild-type from genes that were simply not tested. Without this distinction, models incorrectly treat unsequenced genes as normal, introducing systematic biases that distort both training and predictions.
The model's prognostic performance across the full patient cohort demonstrated a C-index greater than 0.68 -- a measure of how well the model correctly ranks patients by survival risk (0.5 = random, 1.0 = perfect). Dividing patients into high- and low-risk groups by their ATHENA score yielded highly significant separation in overall survival (p less than 0.001), demonstrating that the model's learned mutation interaction patterns translate directly to clinically meaningful risk stratification.
To move from prediction to biological understanding, the researchers applied SHAP (Shapley Additive Explanations) -- an explainable AI method that quantifies each mutation's contribution to the model's risk score. Critically, SHAP can also evaluate pairwise contributions, revealing how two mutations together influence the prediction beyond what either mutation contributes alone. This makes it possible to identify synergistic mutation pairs: combinations that cooperate to drive poor outcomes.
SHAP analysis generated mutational synergy heatmaps for each of ATHENA's three signature branches. Rather than diffuse or random patterns typical of high-dimensional black-box models, the maps revealed well-defined interaction clusters among genes with established biological roles in CRPC. Prominent synergy hotspots included genes in the endoplasmic reticulum-associated protein degradation (ERAD) pathway (notably SYVN1, an E3 ubiquitin ligase), lipid and sterol biosynthesis (HMGCS1), and cell cycle regulation (CDK11B).
The strongest synergy signal involved the pair SYVN1 and STC2. SYVN1 helps cancer cells manage protein-folding stress during AR blockade, while STC2 is involved in metabolic stress responses. Their cooperative action suggests that the tumor exploits both proteostasis and metabolic adaptation simultaneously to survive androgen deprivation -- a combination that may represent a targetable vulnerability if both pathways are blocked at once.
The UMAP visualization of ATHENA's latent space showed that high-risk patients (yellow) and low-risk patients (purple) formed distinct, well-separated clusters with a clear gradient between them. This structured embedding indicates that ATHENA has learned a biologically coherent representation of the prostate cancer mutational landscape -- not just a statistical pattern -- reflecting genuine differences in tumor biology between patient risk groups.
The most immediate clinical application of ATHENA is patient risk stratification -- identifying which metastatic CRPC patients are at highest risk of rapid disease progression and death. This information can guide decisions about treatment intensity: patients identified as high-risk by ATHENA's score may benefit from more aggressive combination regimens or earlier escalation to experimental therapies, while lower-risk patients may safely avoid the toxicity of over-treatment.
Beyond prognosis, ATHENA's synergistic mutation pairs provide a mechanistic rationale for combination therapy design. If SYVN1 and STC2 cooperate to maintain tumor survival during AR blockade, co-targeting the ERAD pathway and metabolic stress response alongside AR inhibition may be more effective than adding any single agent. This represents a departure from empirical combination testing toward mechanistically informed drug selection.
The OncoVar-ATHENA framework also has direct utility for experimental validation. The refined list of high-confidence driver mutations produced by OncoVar is precisely calibrated for base editing experiments, CRISPR knock-in models, and patient-derived organoid systems -- allowing researchers to test the functional impact of specific mutations in controlled laboratory settings. This closes the loop between computational prediction and biological proof-of-concept.
The authors also propose that the framework's design philosophy is generalizable. The three-step paradigm -- domain-specific knowledge to define biologically meaningful modules, architectures that reflect those modules, and interpretable AI to generate testable hypotheses -- could be applied to other cancers where resistance evolves through complex genomic trajectories. ATHENA is therefore not just a prostate cancer tool but a model for biology-driven AI in precision oncology.
The OncoVar-ATHENA framework represents a meaningful advance over existing genomic analysis approaches for prostate cancer. By requiring both DNA presence and RNA expression before accepting a mutation as a candidate driver, and by incorporating protein-structural pathogenicity scoring, OncoVar delivers a substantially cleaner signal than conventional variant calling pipelines -- a critical foundation for the downstream deep learning analysis.
ATHENA's multi-branch attention architecture, trained to recognize how combinations of stage-specific driver mutations jointly determine prognosis, achieves a C-index above 0.68 in large patient cohorts and identifies biologically coherent synergistic mutation pairs through SHAP analysis. These pairs -- such as SYVN1-STC2 -- represent hypotheses about cooperative molecular mechanisms underlying lethal CRPC that can now be directly tested in laboratory models.
The significance of this work extends beyond the specific mutations identified. Current precision oncology for prostate cancer relies heavily on single-gene biomarkers (BRCA2 mutation for PARP inhibitor eligibility, PSMA expression for radioligand therapy). ATHENA demonstrates that multi-gene interaction scores can outperform single-marker approaches, pointing toward a more comprehensive genomic classification of treatment-resistant disease.
Future validation steps include prospective clinical studies testing whether ATHENA risk scores predict treatment outcomes in patients receiving specific therapies, and laboratory experiments confirming the functional significance of the top-ranked synergistic mutation pairs. If these validations succeed, the OncoVar-ATHENA pipeline could become a practical tool for identifying actionable vulnerabilities and designing combination regimens personalized to the specific genomic profile of each patient's tumor.