AmyloCore-ML predicts amyloid fibril core regions from sequence with protein language models, reaching AUROC about 0.88 and outperforming CrossBeta and AggrescanAI

The work introduces AmyloCore-ML, a framework that predicts amyloid fibril core regions at residue resolution directly from sequence, targeting the residues actually incorporated into fibril structure rather than aggregation-prone regions in general. Many earlier predictors estimate aggregation-prone regions, which may not correspond to the residues found in actual disease-associated fibril structures; this work shifts the prediction target to structurally incorporated core regions. Labels come from experimentally determined fibril structures in the Amyloid Atlas, with ordered residues designated as core and unresolved residues sampled from the same proteins as matched controls, giving a supervised setting grounded in experimental structures.

Protein language model embeddings consistently outperformed physicochemical descriptors for core-region identification, with the best models reaching AUROC about 0.88 and AUPRC about 0.85. The comparison evaluates physicochemical descriptors alongside three embedding families, ESM-2, ANKH and ProtT5, within one evaluation framework, giving a relative picture across representations. Evaluation used protein-grouped cross-validation, keeping residues from the same protein out of both training and test splits; the abstract reports AUROC and AUPRC metrics.

In locked full-length protein scans the model localized experimental cores, with ESM-2/ExtraTrees W21 achieving AUROC 0.833, AUPRC 0.751 and a mean peak distance of 12.4 residues. The scan moves evaluation from residue classification to core-region localization across full-length sequences and reports a distance measure between predicted peaks and true cores. Results come from locked full-length protein scans, reporting AUROC, AUPRC and mean peak distance.

In comparative benchmarking the predictor showed better performance metrics than CrossBeta and AggrescanAI, and it is provided as an interactive Google Colab notebook requiring no local installation or dedicated computing infrastructure. Beyond the metric comparison, the work treats usability as part of the deliverable, lowering the barrier to sequence-based fibril-core prediction. The comparison rests on the performance metrics described in the abstract; usability is evidenced by the public Colab notebook.

The framework is aimed at researchers who need to judge amyloid fibril core regions from sequence, applies to protein systems referenced against experimentally determined fibril structures, and outputs residue-level core propensity with core-region positions available in full-length protein scans. Because it is delivered as a Google Colab notebook, users can run it without local installation or dedicated computing infrastructure, making it suitable as a screening and hypothesis-generation tool for proposing core-region candidates in sequences beyond those with solved structures.

The visible text is the abstract and does not include figures or data tables, so the number of proteins used for training and testing, the class balance between core and non-core residues, the full comparison table across embeddings and classifiers, and the specific settings used in the CrossBeta and AggrescanAI comparison cannot be confirmed from the text. How the mean peak distance of 12.4 residues varies for longer or shorter core regions, and how the predictor behaves on proteins without experimental structures, remain open questions worth watching.