CMA-DTI aligns drug functional groups with protein binding regions to improve cold-start drug–target prediction
- A dual-channel encoder over pretrained drug and protein sequence-model embeddings captures local functional patterns and global context, and a bidirectional cross-modal alignment module links drug functional groups to protein binding regions.
- On four benchmark sets it was competitive in warm-start settings and generalized better in the target cold-start and blind-start scenarios.
- Ablations and structural visualizations are presented as support for the modules and for residue-level interpretability.
💡 Points Based on the abstract
DTI benchmarks are notorious for ligand and protein bias, so the cold-start and blind splits are where to look; I'd want them with similarity-clustered splits and a ligand-only baseline. Attention-style residue maps are suggestive, not evidence of binding mode, so compare them with known co-crystal contacts before using them as pocket hints. For a sparsely studied target family, the practical question is whether the language-model embeddings carry any signal at all.
Cross-Modal Semantic Alignment for Cold-Start Drug–Target Interaction Prediction ↗