Determine the causes of retraining gains on PocketWatch

Determine how much of the performance improvement obtained by retraining VN-EGNN and UniSite-1D on the PocketWatch dataset is attributable to the new PocketWatch data rather than to differences in training procedures or the closer match between the training and test distributions.

Background

Retraining VN-EGNN and UniSite-1D on the PocketWatch train/validation split improves their per-target AUPRC relative to their published weights. However, the retraining protocols differ from the procedures used to produce the published checkpoints, and PocketWatch training data are more closely matched to the PocketWatch test set than the original training distributions. Consequently, the observed gains cannot be cleanly assigned to the dataset itself without additional controlled ablations.

References

However, we acknowledge that some differences in training procedures (Appendix~\ref{app:baselines}) may also affect results, so it's not possible to entirely attribute the gains to the new data without further ablations.

— zPocket: Better Pocket Finding with Frozen Co-Folding Features  (2610.17696 - Portnoi et al., 13 Oct 2026) in Section 4.1, subsection “Retraining on PocketWatch improves baselines”