Cold-start evaluation: identities before architectures

The decision: define what must be unseen

Cold-start generalization is an identity question. S2 holds out drugs; S3 holds out targets; S4 holds out both. Merely withholding individual drug–target pairs does not establish that either identity is new to the model.

The lab’s six-by-six grid makes the distinction visible. In S4, the training block uses only old identities and the test block uses only new identities. Cross-block pairs are excluded in this schematic fold. The drawing explains the split rule; it is not the study’s actual fold manifest.

The consequence: richer representations need careful attribution

The published topology combines classical drug and protein representations, bidirectional cross-attention, simulated quantum residual paths, and gated fusion. Looking at the classical backbone and residual paths separately helps inspect the architecture, but hiding a path in a diagram is not a numerical ablation.

Similarly, the attention lens uses invented logits. Lower temperature concentrates their normalized weights; it does not add biological evidence. The model’s state-vector simulations are not a quantum-hardware run, and neither attention sharpness nor architectural complexity establishes quantum advantage.

Paired evidence, four contrasts, one corrected result

The complete paired evidence shown here covers ten DAVIS folds for S2 and ten for S4. Deltas are defined as QVGAT-DPI minus Q-BAFNet: positive favors QVGAT-DPI for concordance index, while negative favors it for RMSE. The sign must be read with the metric.

The small check recomputes means from the published delta columns, enumerates all 1,024 sign assignments per contrast for exact two-sided sign-flip p-values, and applies Holm correction across the four S2/S4 by CI/RMSE contrasts. Only S2 RMSE remains below 0.05 after correction. Neither protocol establishes a corrected dual-endpoint win.

• Use reported deltas rather than subtracting independently rounded model scores.
• Keep all four tested endpoints in the stated correction family.
• Distinguish arithmetic agreement from validation of sign-flip assumptions or fold independence.

Missing experiments must remain visible

DAVIS S3 has only three completed folds. KIBA S2 has two; KIBA S3 and S4 were not run. The evidence atlas leaves these cells incomplete or blank, without constructing aggregate results from missing experiments.

The downloadable package verifies reported summary arithmetic, not training, learned representations, bootstrap intervals, or missing cells. Extending the experimental grid is a separate research task. A larger set of valid paired comparisons could change the conclusion; the partial pilot should not be presented as a universal model ranking.