Prediction and design of IDR co-phase separation

Multiple biomolecular condensates with distinct compositions can coexist in cells to house diverse biochemical processes. Although biomolecular condensation is in many cases known to be driven by multivalent interactions between proteins with intrinsically disordered regions (IDRs), the molecular language that dictates the compositional specificity of multicomponent condensates remains poorly understood. Here, we develop a theoretical approach that integrates high-throughput molecular dynamics (MD) simulations and machine learning (ML) methods to decipher the molecular determinants of multicomponent IDR interactions. The central feature of our ML model is a learned common latent space, which maps pairs of IDR sequences to the concentration-dependent thermodynamic properties of corresponding IDR mixtures. We show that this approach can predict the tendency of IDR pairs to co-phase separate and can accurately reconstruct complete concentration-dependent phase diagrams. Despite the vast size of IDR sequence space, the learned latent space identifies a small number of essential IDR features that dictate the phase behavior of IDR mixtures in an unbiased manner. Taken together, our approach establishes a rational tool to predict and design IDR co-phase separation that can easily be extended to systems with many components, laying the groundwork for systematic studies of condensate compositional specificity in the complex cellular milieu.