#14 · ML / AI methods

Graph neural networks fuse multimodal omics and network data for biological prediction

The dominant thread is graph-structured deep learning that fuses heterogeneous biological inputs into shared representations for prediction. Mahi combines chromatin, TF binding, histone and structure features in a tissue-aware graph neural network to model context-specific gene regulation, while DVMMHGNN integrates microbiome, multimodal features and meta-paths in heterogeneous graph contrastive learning for drug-disease associations. CellAwareGNN injects cell-type-specific expression from OneK1K into a knowledge graph foundation model for drug indication prediction, and the multiplex-network framework pairs foundation-model embeddings with a topology-aware design for zero-shot interaction prediction. The omics graph study encodes probability distributions on nodes and edges across five cancer types. GraSPNet extends the paradigm to molecules through hierarchical fragment-level self-supervision. What is new across these preprints is the combination of self-supervised or foundation-model pretraining with multimodal graph inputs spanning regulatory, network and molecular scales.