Artificial intelligence-driven approaches for the rational design of peptides with predictable aggregation propensity

The rapid development of artificial intelligence has enabled accurate and efficient de novo design of protein and peptide structures. However, applying AI to design sequences with specific aggregation tendencies remains challenging due to the need to span multiple spatial-temporal scales. Here, we combined deep learning strategies—including genetic algorithms and reinforcement learning—to generate decapeptides with tunable aggregation propensities. Coarse-grained molecular dynamics simulations were used to evaluate solvent-accessible surface area and define aggregation propensity (AP). A Transformer-based prediction model with self-attention achieved high accuracy in AP prediction with only a 6% error rate. Furthermore, Monte Carlo Tree Search enabled targeted optimization of peptide sequences while preserving desired functional features. This study demonstrates how integrating AI with molecular modeling can guide the rational design of peptides with controlled assembly behavior, providing a scalable strategy for applications in biotechnology and medicine.