Visible neural networks for multi-omics integration: a critical review

Biomarker discovery and drug response prediction are central to personalized medicine, driving demand for predictive models that also offer biological insights. Biologically informed neural networks (BINNs), also referred to as visible neural networks (VNNs), have recently emerged as a solution to this goal. BINNs or VNNs are neural networks whose inter-layer connections are constrained based on prior knowledge from gene ontologies and pathway databases. These sparse models enhance interpretability by embedding prior knowledge into their architecture, ideally reducing the space of learnable functions to those that are biologically meaningful.

This systematic review-the first of its kind-identified 86 recent papers implementing BINNs/VNNs. We analyzed these papers to highlight key trends in architectural design, data sources and evaluation methodologies.

Our analysis reveals a growing adoption of BINNs/VNNs. However, this growth is apparently juxtaposed with a lack of standardized, terminology, computational tools and benchmarks.

BINNs/VNNs represent a promising approach for integrating biological knowledge into predictive models for personalized medicine. Addressing the current deficiencies in standardization and tooling is important for widespread adoption and further progress in the field.