Active Learning–Guided Peptide Design for Modulating Condensate Properties upon Recruitment

The physical properties of biomolecular condensates, which form through phase separation, are central to their organisation and function and are increasingly implicated in disease mechanisms. To function correctly, condensates often recruit client biomolecules such as peptides and RNAs. This recruitment is not only essential to condensate biology but also offers an opportunity to engineer clients that tune the material properties of the condensate. Here we studied the previously characterised MUT-16 condensate, the scaffold of a membraneless organelle in C. elegans that recruits the N-terminal prion-like domain of MUT-8, a process essential for RNA silencing. We used coarse-grained molecular dynamics simulations to model and design peptide variants that interact with the scaffold protein MUT-16 and modulate its physical properties. To guide this design efficiently, we applied an active learning framework, which combines Bayesian optimization with neural networks to iteratively select the most informative peptide variants for simulation. This strategy reduced the number of simulations required while allowing the model to learn the sequence–property relationships that govern condensate behavior. Overall, this physics-based, learning-guided framework offered a computational approach to exploring how peptides can be engineered to influence condensate properties, and may inform the rational design of synthetic condensates.