DeepBINN: A tailored biologically-informed neural network for robust biomarker identification

Biologically-informed neural networks (BINNs) have recently been proposed to aid in identifying disease-relevant biomarkers and biological pathways using deep neural networks, a class of machine learning models typically lacking necessary interpretability. By incorporating a priori knowledge of the relationships between biomarkers (e.g., proteins) and biological pathways into sparse neural networks, BINNs offer competitive performance across a range of biologically relevant tasks. Nevertheless, the robustness of the biomarker signatures found by such models is often insufficiently studied. To address this, we introduce a new BINN architecture and rigorously evaluate its ability to identify stable biomarker signatures. Our architecture, named DeepBINN, is designed to learn the importance of pathways during training, without depending on post-hoc explainability methods. This marks a significant advance in the interpretability of "black box" neural networks. In a preliminary study, we assess DeepBINN’s capability to stratify patients into two sub-phenotypes of septic acute kidney injury. Early tests indicate substantial improvements in both model interpretability and robustness, while maintaining state-of-the-art performance.