Capturing α-synuclein polymorphism with AlphaFold 3 and rivals

Amyloid fibrils formed by a single protein can adopt many distinct folds, or polymorphs, which differ between diseases and shape the surfaces available to therapeutics and diagnostics. Whether deep-learning structure prediction can capture this diversity is unclear, and the field lacks an objective definition of a polymorph from which to judge this. Here we cluster the -synuclein structures in the Protein Data Bank by TM-score, giving an objective reference set of known polymorphs. Against this reference we assess AlphaFold 3, Boltz-2 and OpenFold 3 run through ABCFold. All three produced fibril-like assemblies, though they differed in which polymorphs they reached. With templates included as default the methods over-represented the dominant fold. When untemplated the methods produced a substantial number of models matching no known structure, a subset of which have the stacking geometry of genuine fibrils. Supplying a polymorph as a template recovered 13 of 16 target folds, but only when alignment depth and the number of modelled chains were varied together. These results establish that current methods sample real -synuclein structural space and can be steered through it, and provide a framework for assessing prediction methods on amyloid targets.