#2 · Structural bio & protein design

Graph and language-model encoders advance drug-target and synergy prediction

The dominant thread is deep architectures that condition predictions on structured molecular representations. ResGIN-Att applies a residual graph isomorphism network with LSTM fusion and cross-attention to molecular graphs and cell-line features for drug synergy, and GTStrDTI applies graph transformer attention over molecular graphs and 3D protein contact graphs for binding affinity. A second thread uses pretrained encoders as backbones: DMAPLM pairs ChemBERTa-2 and BioBERT via contrastive alignment for drug-disease association, a dual-encoder model combines ESM embeddings with ChemBERTa and graph ligand encoders, and other work uses ProtT5-XL embeddings or AlphaFold2 Evoformer representations for sequence-based prediction. Newer elements include CellChem, which pretrains molecular embeddings on cellular transcriptional profiles, fragment-contribution contrastive pair construction with information bottleneck scoring, and DCI-SiteDTA's site-guided integration of binding-site detection with affinity prediction. A kinase benchmark across random, scaffold, and target-held-out splits makes evaluation strategy an explicit variable.