A new evaluation in Scientific Reports finds that AlphaFold 3 frequently misjudges whether a protein forms disease-linked amyloid fibrils.
Researchers tested AlphaFold 3 (AF3) on three sets of proteins. Among 153 proteins and peptides known to form fibrils but lacking experimental structures, AF3 predicted a fibrillar structure for only 34%. Among 56 peptides known not to aggregate, it wrongly predicted fibrillar structures for 54%. On a third set of seven proteins with experimentally known fibril structures, AF3 got five right.
The authors compared AF3 against two other modelling tools, Boltz and RibbonFold, across the same cases and checked how closely predicted structures matched known amyloid folds, including CsgA, alpha-synuclein and Abeta-42.
The authors state the result plainly: the "amyloid modelling quality of AF3 seems underwhelming," and they attribute this mainly to the small number of amyloid structures in the Protein Data Bank compared with globular forms, plus the fact that many amyloids adopt multiple different fibril shapes (polymorphism), which current training data do not capture well.
For researchers using AF3 on amyloid-forming sequences, the paper's findings, reported as preliminary case-study results rather than a general benchmark, suggest treating any fibrillar prediction as a hypothesis to test experimentally, not as a final answer. Shorter fragments appear more likely to be modelled correctly than full-length proteins.
