Clarification on label normalization in ProteinGym supervised baselines

[yh79-ss]

Hi, thanks for releasing ProteinGym, it’s super helpful. Could you clarify how labels are normalized for the supervised baselines? Specifically, do you train directly on the provided DMS_score values (after unifying the direction), or do you apply any per-assay / per-fold normalization (e.g. z-score using training labels) before training and computing MSE?

[pascalnotin]

Hi @yh79-ss - thank you for the kind words! We do apply per-fold standard normalization, and MSE values are computed with respect to these normalized values. For more details, please refer to this issue.
Best,
Pascal