#5 · Systems biology

Graph neural networks model molecular interactions, drug synergy and disease drivers

The dominant thread is graph-structured learning over molecular entities, with biological relations such as protein-protein interactions, drug-target-disease links, and multi-omics networks encoded as graph topology rather than flattened into feature vectors. Several items use attention or directed message passing: PPI-Net embeds patient RNA-seq in STRING and Reactome hierarchies, TopoFuseNet adds group-centrality features for drug synergy prediction, and MNDGNN applies directed GNNs to multiplex networks for cancer driver genes. Others inject explicit priors, as in L3-PPI's L3 path rule and ProtDiS's biologically grounded embedding dimensions. Structure-learning work (PACER, RSNet, PFS) extends the thread to causal and local network inference. The newer element is explicit biological constraint or topological feature engineering paired with benchmarked evaluation on public datasets, though the items do not share a single modality or method.