#23 · ML / AI methods

Structure-aware architectures couple molecular substructures with biological networks

The common thread is that structural information enters these models as architectural components, governing pairing, relational, or fusion operations rather than serving only as static input descriptors. GMJRL fuses network-level representations with atomic-geometry features for drug-target interaction prediction, while engGNN couples external biological networks with tree-ensemble-derived feature graphs for omics disease classification and feature selection. DiSPA applies differential cross-attention between chemical substructures and pathway gene expression to model bidirectional interactions for drug response, and GenRel-DDI recasts drug-drug interaction prediction as relation-centric learning that generalizes to unseen drugs on entity-disjoint benchmarks. What is new is the joint modeling of atomic, molecular, and interaction-network scales within a single architecture. The tasks and data modalities differ across studies, so the shared principle is architectural rather than uniform in implementation.