Clinically interpretable deep learning for breast cancer missense variant pathogenicity prediction.

Background: Missense variants in breast cancer remain diagnostically challenging due to their functional diversity and complex genomic contexts. Conventional laboratory assays for evaluating pathogenicity are labor-intensive, costly, and often impractical for large-scale screening, creating a pressing need for accurate, scalable, and clinically interpretable computational approaches. Methods: In this study, we present a novel deep learning framework for predicting the pathogenicity of breast cancer missense variants, integrating comprehensive preprocessing, advanced imputation, rigorous model benchmarking, and explainability. Genetic variants were curated from multiple genomic databases, annotated using the Ensembl Variant Effect Predictor (VEP), and processed with Variational Autoencoders (VAE) for missing-value imputation. Seven deep learning models, MLP, CNN, DNN, RNN, LSTM, GRU, and Transformer, were trained and evaluated across 11 performance metrics. To quantify performance stability, each model was trained across five random seeds; mean AUC ± SD across seeds is reported as the primary performance estimate, with the best-seed run used only for LIME and PMI interpretability analyses. Recursive feature elimination, permutation importance (PMI), and Local Interpretable Model-Agnostic Explanations (LIME) were employed to enhance transparency. Statistical analyses, including Z-tests, ANOVA, and calibration assessments, validated performance consistency and inter-model differences. Results: GRU achieved the highest internal AUC (0.9956 [95% CI 0.9936-0.9972]; mean across five seeds 0.9941 ± 0.0011), with precision 0.9967 and calibration ECE 0.0095. Externally, LSTM led with AUC 0.9457, exceeding all eleven standalone predictors benchmarked on the same set. Models showed strong alignment with conservation signals such as phyloP470way and Eigen-PC scores. Notably, the pipeline provides performance metrics with 95% confidence intervals and incorporates case-level LIME visualizations for true positive, true negative, false positive, and false negative predictions, bolstering interpretability and clinical relevance. Conclusion: This work delivers one of the most comprehensive evaluations of deep learning in breast cancer variant classification to date. By combining high-performance sequential models with interpretable AI tools, the proposed framework provides a reproducible, transparent benchmark for variant pathogenicity prediction and a foundation for future research use and translation in cancer genomics.