Cross-omics interpretable neural network for discovery of molecular markers in prostate cancer

Determining molecular markers that mediate clinically aggressive phenotypes in prostate cancer is a significant challenge. While traditional linear models offer some interpretability, they often lack the precision needed for complex multi-omics data. Conversely, conventional deep learning methods provide robust predictions but typically remain opaque, hindering the identification of impactful molecular markers and biological mechanisms. To address this, we propose the Cross-omics Interpretable Neural Network (CINN), a biomimetic framework designed to predict prostate cancer states and identify key molecular markers by integrating diverse omics data. CINN innovatively leverages prior biological knowledge from either pathway or protein-protein interaction (PPI) networks, combined with a novel trainable mask layer. This mask dynamically optimizes the strength of pre-defined biological connections, thereby enhancing both knowledge representation and model interpretability. The framework effectively integrates multi-omics data, including gene expression, somatic mutations, and copy number variations, to provide a holistic view of the disease. Extensive experiments on a prostate cancer dataset demonstrate that CINN achieves substantial and statistically significant performance enhancements over a strong baseline (P-NET). Specifically, our best-performing variant, CINN-pw with a trainable mask, improved F1 scores by 13.1% to 0.843, Accuracy by 8.3% to 0.894, and AUC by 2.3% to 0.949. These gains were consistently statistically significant (p<0.0001 for most key metrics), underscoring the robustness of our approach. Crucially, CINN's inherent interpretability facilitated the identification of pivotal molecular candidates, including TBP and TAF2, which are implicated in prostate cancer progression. These findings are supported by existing literature and provide valuable insights into the underlying mechanisms of prostate cancer, offering potential avenues for targeted therapeutic interventions and precision medicine.