Capturing single-molecule properties does not ensure accurate prediction of biomolecular phase diagrams

Intracellular liquid-liquid phase separation of proteins and nucleic acids represents a fundamental mechanism by which cells organise their components into biomolecular condensates that perform multiple biological tasks. Computer simulations provide powerful tools to investigate biomolecular phase separation, offering microscopic insights into the physicochemical principles that regulate these systems. In this study, we investigate the phase behaviour of the low-complexity domain (LCD) of hnRNPA1 and several mutants via Molecular Dynamics simulations. We systematically compare the performance of five state-of-the-art residue-resolution coarse-grained protein models: HPS, HPS-cation- π , CALVADOS2, Mpipi, and Mpipi-Recharged. Our evaluation focuses on how well these models reproduce experimental coexistence densities and single-protein radii of gyration for the LCD-hnRNPA1 set of mutants. While most models yield similar intramolecular contact maps and reasonable estimates of the single-protein radius of gyration compared to in vitro measurements, only Mpipi-Recharged, Mpipi, and CALVADOS2 accurately predict phase diagrams that align with experimental data. This suggests that force field parameterizations optimized solely to reproduce single-protein properties may not always capture the phase behaviour of protein solutions. Additionally, our findings reveal that some residue-resolution coarse-grained models can lead to significant discrepancies in predicting the roles of individual intermolecular interactions, even for relatively simple intrinsically disordered proteins like the low-complexity domain of hnRNPA1. Our work highlights the importance of balancing both single-molecule and collective properties of proteins to accurately predict condensate formation and material properties.