Interpretable deep neural network identifies robust biomarkers for diseases with mechanistic insights from omics data.

Identifying essential biomarkers remains a core challenge in elucidating the pathogenic mechanisms and achieving precise diagnosis of complex diseases. Deep neural networks offer immense predictive power, yet their lack of interpretability severely limits downstream biological insight. Here, we introduce DeepVaris, an explainable deep learning framework that reframes feature selection as the interpretation of a pretrained convolutional neural network via surrogate modeling. In extensive simulations and real-world datasets, DeepVaris successfully identifies important features and reveals deeper insight into different diseases. Specifically, it overcomes extreme feature sparsity to identify crucial microbial biomarkers in preterm birth pregnancies. In single-cell RNA sequencing data, it reveals key transcriptional drivers governing myelin regeneration in neurodegenerative diseases missed by traditional differential expression analysis. Furthermore, in complex breast cancer cohorts, DeepVaris moves beyond generic pan-cancer signals to precise subtype-specific microenvironmental targets. In summary, we believe that DeepVaris will serve as a robust tool for biomarker discovery.