Comparative study of hybrid quantum-classical models for cold-start drug–target affinity prediction: a partial pilot analysis

Abstract

Cold-start drug–target affinity (DTA) prediction requires evaluation on genuinely unseen chemical and protein entities. We considered QVGAT-DPI (Quantum Variational Graph Attention Network for Drug–Protein Interaction) as a residual hybrid model combining full molecular graphs, permutation-invariant Breaking of Retro-synthetically Interesting Chemical Substructures (BRICS) fragment queries, ESM-2 and ChemBERTa representations, fingerprints, cross-attention, and two simulated six-qubit residual modules. QVGAT-DPI was compared with specification-based implementation of Q-BAFNet, the closest structured hybrid architecture, under identical local fold manifests. Because of the limitation of the computational resource, the planned ten paired folds were completed only for DAVIS S2 (unseen drugs) and S4 (unseen drugs and targets). For S2, paired differences were + 0.060 CI (bootstrap 95% interval − 0.000 to 0.123; exact p = 0.0996) and − 0.094 RMSE (− 0.145 to − 0.046; p = 0.0098). For S4, differences were + 0.087 CI (0.032–0.143; p = 0.0195) and − 0.135 RMSE (− 0.294 to − 0.005; p = 0.1191). After Holm adjustment, only S2 RMSE remained below 0.05, and no protocol supported both endpoints. The prespecified aggregate S2–S4 conclusion was unevaluable because S3 was incomplete. These results support further controlled evaluation, not system-wide superiority or quantum advantage.