#5 · ML / AI methods

Biological priors make multi-omics deep learning interpretable

The dominant thread is biological structure supplied as inductive bias or pretrained representation to make multimodal omics prediction interpretable. Expanded P-NET builds pathway hierarchies into its architecture and is competitive with traditional machine learning, while PathMoG and bioGraph apply pathway-module and heterogeneous graph learning over omics data. SpaMosaic, INSPIRE and SCIGMA apply contrastive or graph-based integration to spatial multi-omics, with SCIGMA adding uncertainty awareness across up to five modalities. CMGL estimates per-patient modality confidence via evidential deep learning before fusion. scpFormer and H2O extend this toward pretrained representations, embedding antibody panels and inferring spatial omics from H&E images. What is new is the joint treatment of cross-modal integration, built-in priors, and explicit uncertainty or interpretability within one framework, though not every item addresses all three.