Prostate-VarBench: A Benchmark with Interpretable TabNet Framework for Prostate Cancer Variant Classification

TL;DR

Prostate-VarBench tackles the pervasive issue of Variants of Uncertain Significance in prostate cancer genomics by building a prostate-specific, leakage-controlled benchmark that harmonizes data from COSMIC, ClinVar, and TCGA-PRAD into a 193,278-variant corpus. It combines a VEP-annotation correction, an eight-tier clinical feature ontology, and the AlphaMissense scores to feed a TabNet classifier that delivers native per-case explanations via step-wise attention masks. The approach achieves an on-par 89.9% test accuracy with balanced metrics, while the VEP correction reduces VUS by 6.5 percentage points, enhancing actionable interpretation in a clinically realistic setting. Overall, the work demonstrates that interpretable AI can meet clinical performance standards and provide transparent rationales suitable for tumor boards and regulatory review, offering a reproducible benchmark for future prostate cancer genomics research.

Abstract

Variants of Uncertain Significance (VUS) limit the clinical utility of prostate cancer genomics by delaying diagnosis and therapy when evidence for pathogenicity or benignity is incomplete. Progress is further limited by inconsistent annotations across sources and the absence of a prostate-specific benchmark for fair comparison. We introduce Prostate-VarBench, a curated pipeline for creating prostate-specific benchmarks that integrates COSMIC (somatic cancer mutations), ClinVar (expert-curated clinical variants), and TCGA-PRAD (prostate tumor genomics from The Cancer Genome Atlas) into a harmonized dataset of 193,278 variants supporting patient- or gene-aware splits to prevent data leakage. To ensure data integrity, we corrected a Variant Effect Predictor (VEP) issue that merged multiple transcript records, introducing ambiguity in clinical significance fields. We then standardized 56 interpretable features across eight clinically relevant tiers, including population frequency, variant type, and clinical context. AlphaMissense pathogenicity scores were incorporated to enhance missense variant classification and reduce VUS uncertainty. Building on this resource, we trained an interpretable TabNet model to classify variant pathogenicity, whose step-wise sparse masks provide per-case rationales consistent with molecular tumor board review practices. On the held-out test set, the model achieved 89.9% accuracy with balanced class metrics, and the VEP correction yields an 6.5% absolute reduction in VUS.

Figures (2)

• Figure 1: Experimental setup and TabNet architecture overview. Input sources and leakage controls feed tiered features (VEP-corrected, core VEP, AlphaMissense, population, functional, clinical context, variant properties, prostate biology). TabNet executes six decision steps with sparse feature masks; step logits are aggregated into class probabilities. Masks are retained for per-variant explanations and tier-level analyses referenced in Section \ref{['subsec:attention-analysis']}
• Figure 2: Overview of results with feature–attention analysis. (A) Performance (balanced accuracy, Cohen’s $\kappa$, weighted F1, ROC–AUC) for TabNet, XGBoost, and logistic regression (mean$\pm95\%$ CI), (B) One-vs-rest ROC curves with macro-AUC, (C) Step-wise attention heat map across six decision steps highlighting top features (colors denote tiers), and (D) TabNet feature importance by hierarchical tiers.