AmyloPick: A New Feature Selection Method and Proper Evaluation for Amyloid Hexapeptides and Aggregation-Prone...

Given the critical importance of preventing protein aggregation in neurodegenerative diseases, aggregation prediction tools are essential. Amyloid predictors would facilitate the understanding and exploitation of the amyloid state of proteins, providing an alternative to costly and slow laboratory tests. In recent years, hexapeptides have become a model for studying amyloid formation. Hexapeptides can also be used to identify aggregation-prone regions in proteins, particularly those involved in amyloid formation. While numerous computational methods using sophisticated feature sets and architectures have been developed for classifying hexapeptides and predicting amyloidogenic regions in proteins, predictive performance remains limited; for instance, BAP achieves only 84% accuracy. Here, we designed a novel feature selection method for hexapeptides, resulting in an easy to interpret four-feature representation called the AmyloPick model. A classifier based on this representation outperforms existing state-of-the-art methods. When extended to detect aggregation-prone regions (APRs) in full proteins, it performs comparably to established tools. A key contribution of this study is the statistical methodology that enables a rigorous performance assessment and direct comparison with other classifiers. This is particularly important because differing methodologies in the literature often hinder the comparability of the proposed methods. Our AmyloPick classifier significantly outperformed the state-of-the-art Budapest Amyloid Predictor (BAP) across all metrics, particularly in adjusted geometric mean (AGM) (0.7808 vs. 0.7649 for BAP) and accuracy (0.8089 vs. 0.7955 for BAP). For APR identification, APR-AmyloPick was comparable to ANuPP overall but significantly outperformed it in the ${\mathrm{SOV}}_{\text{Non}-\mathrm{APR}}$ metric. We have also developed a web server for the AmyloPick classifier.