#16 · ML / AI methods

Fusing sequence, structure and graph features for multimodal interaction prediction

The dominant thread is fusing heterogeneous representations under task-specific objectives. MMM-PPI combines sequence, structure and function residue features hierarchically for protein-protein interaction, and FunBind aligns sequences, text, domains, structures and GO terms for zero-shot and supervised function prediction. In drug-target and molecular settings, CM-MTL-DTI couples 1D CNN sequence encoders and a GIN graph encoder through bidirectional cross-modal attention with multi-task learning, HKD-CPI adds LLM-aligned graph tokens, hypergraph modeling and knowledge distillation, and DMPKformer combines Transformer, GAT and MLP branches for ADMET prediction with latent OOD confidence. MPMFMol pairs multitask pretraining with fusion of graph, functional-group and SMILES inputs, and GraFSyn uses graphlet fingerprints for interpretable synergy prediction. Several items pair fusion with interpretability or uncertainty estimation, while others emphasize pretraining or knowledge transfer. The P-glycoprotein study fine-tunes a single pretrained GIN encoder and does not share this fusion design. The set shares a multimodal conditioning strategy rather than a common architecture, dataset or benchmark.