Researchers tested next-amino-acid predictions on 668 little-studied proteins from 167 families. Claude Fable 5.2 scored about 54%, versus 9% on shuffled sequences. The authors interpret the gap as evidence of substantial exposure to protein sequences, not proof of a particular data source or training stage. Their published code and data acknowledge that limitation. This is a sequence-prediction result, not validation of protein structure or clinical efficacy.
Back to Tech
Claude protein probe points to sequence exposure
Strong next-amino-acid predictions do not establish training provenance or biological efficacy.