Discovering new amyloid-like peptides using all-atom simulations and artificial intelligence.

Establishing the fundamental relationships between peptide sequences and fibril formation is critical both for understanding protein misfolding processes and for guiding biomaterial design. Here, we combine all-atom molecular dynamics (MD) simulations with artificial intelligence (AI) to investigate how subtle variations in the arrangement of a short peptide sequence affect its propensity to form fibrils. Our results show that small shifts in the distribution of hydrophobic residues and charge clusters can significantly influence both the nucleation rate and the stability of cross-{beta} structures. To rapidly extend this analysis over a wide sequence space, we developed an active learning-enhanced framework--Machine Learning for Molecular Dynamics (ML4MD)--that iteratively refines its predictions based on MD-derived aggregation data. ML4MD efficiently screens numerous peptide permutations and guides the discovery of previously unrecognized fibril-prone sequences, achieving an area under the receiver operating characteristic (ROC) curve (AUC) of 0.939. Overall, ML4MD streamlines the rational design of amyloid-like peptides by integrating detailed atomistic simulations with rapid and high-accuracy ML predictions.