Data for "When Can We Trust Strict Cold-Start Drug–Target Affinity Predictions?"

This dataset supports the manuscript “When Can We Trust Strict Cold-Start Drug–Target Affinity Predictions? A Systematic Study of Dual-Entity Distance, Uncertainty Signals, and Cross-Dataset Transferability of Reliability Boundaries”. It contains processed strict cold-start splits, GraphDTA-GINConvNet ensemble predictions, risk scores, drug/protein distance files, entity-level analysis outputs, threshold-transfer outputs, supplementary tables, and supplementary DeepDTA robustness results. The archived protein-distance files are provided to reproduce the post-hoc reliability analyses; the upstream protein sequence identity generation script/tool settings were not fully recovered and are documented as an implementation-provenance limitation.