Discovery of unconventional and nonintuitive self-assembling peptide materials using experiment-driven machine learning

Abstract The synthesis of secondary structures such as β-sheet assembly in peptides poses significant challenges, as their formation is governed by complexity of molecular interactions, sequence-specific behavior, and influence of various environmental factors such as solvent, pH, and temperature. While rational design strategies have been attempted to predict peptide sequences that form desired nanostructures, these methods are often biased and ineffective, particularly when dealing with the complexities associated with β-sheet formation. Computational models also struggle to reliably predict β-peptide formation due to the limited understanding of the vast chemical space and the lack of comprehensive experimental data. Furthermore, existing models often fail to capture the diversity and non-intuitive nature of sequences capable of forming β-sheets, leading to a constrained discovery process. To address these issues, we developed an integrated high-throughput experimental workflow and an active learning-based AI framework to better predict the β-sheet formation propensity of peptide sequences. By focusing on sequences where ML predictions deviate from conventional design strategies, we synthesized and tested 268 pentapeptides, discovering 96 that successfully formed β-sheet assemblies. The ML models were found to surpass the commonly used β-sheet propensity table in identifying β-sheet forming sequences, especially for cases that do not contain valine, a high propensity β-sheet conformer. Furthermore, several diverse, non-intuitive and unusual β-sheet forming pentapeptides (for e.g., ILFSM, LMISI, MITIY, MISIW, and WKIYI) were identified by the ML models. Novel chemical trends were mined by analyzing the ML models and a user-friendly web-interface is provided to make these models accessible to the community. Overall, this work underscores the need for AI driven experimental exploration of diverse peptide chemistries and provides a robust strategy to overcome the biases and limitations inherent in current rational design methods to form secondary structures like β-sheets.