Answers are grounded in this paper and cite their sources.
Accurate drug–target affinity (DTA) prediction benefits from complementary molecular and protein information, but fused inputs may carry overlapping signals. Drug-GCPR integrates drug graphs, pharmacophore descriptors, ESM-2 embeddings and residue lookup features. A zero-initialized atom–residue cross-attention branch adds a local residual. With stop-gradient conditioning and grouped inner selection, Drug-GCPR attenuates the pharmacophore component predicted from graph features. Formal validation retains correction only when it lowers mean squared error (MSE) relative to the uncorrected residual predictor. Across five predefined seeds, mean MSE was 0.2495 ± 0.0043 on Davis and 0.1609 ± 0.0087 on KIBA. Relative reductions were 7.94% and 3.04% against GraphDTA-GIN and 4.56% and 3.34% against GIN+ESM Pooled, respectively. Entity-aware resampling supports both Davis baseline comparisons and the KIBA correction relative to Plain Residual. Intervals for KIBA external-baseline and Davis correction effects include zero. On KIBA, Drug-GCPR reduced mean MSE by 1.50% relative to Plain Residual in three cold-drug splits. On KIBA, normalized conformal prediction reduced interval width and score at both nominal levels with empirical coverage near target. On Davis, the mean interval-score change at 95% nominal coverage was negligible. These gains varied across evaluation settings on the two kinase-focused benchmarks.
No takes yet. Share an insight, caveat, or question.
Zhu et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: