BPS2026 – Physics-guided design of intrinsically disordered protein binding peptides

Recent advances in machine learning have enabled the design of proteins that bind disordered regions capable of adopting transient folded states. However, designing binders for highly dynamic intrinsically disordered regions (IDRs), which lack transient structure, remains a major challenge, as most AI-based frameworks rely on static representations of molecular interactions. Yet these regions still play key roles in transcriptional regulation and disease, motivating new approaches for targeting their conformational ensembles. To address this, we developed a physics-based genetic algorithm that integrates coarse-grained simulation forcefields with sequence-based disorder predictors. This framework evolves peptide sequences predicted to engage IDRs while allowing both partners to remain dynamic in the bound state. We focused on the activation domain of the PAX3-FOXO1 fusion oncoprotein, a highly disordered driver of alveolar rhabdomyosarcoma, and generated diverse peptide populations predicted to form sequence-specific, dynamic interactions. To experimentally test these predictions, we are establishing a high-throughput single-molecule binding assay capable of measuring affinities and dynamic behaviors for hundreds of peptides in parallel. We have identified generated peptides that bind with nM binding affinities. By linking computational design with single-molecule measurements, we are assembling a comprehensive dataset that connects sequence composition, disorder, and binding propensity. This foundation will enable training of predictive models for designing dynamic binders to disordered targets. Together, our approach establishes a generalizable framework for exploring sequence-ensemble-binding relationships within IDRs, advancing our ability to target the disordered proteome.