A machine learning framework for predicting and modulating condition-dependent protein phase separation

Protein phase separation is a fundamental process in organizing membraneless organelles and is implicated in a wide range of pathological conditions. Importantly, rather than being a static feature of specific proteins, phase separation is a condition-dependent phenomenon governed by environmental parameters, including protein concentration, temperature, and solvent composition. However, most existing machine learning models infer phase-separation propensity solely from amino-acid sequences, failing to capture these context-dependent behaviors. Here, we present LLPSense, a machine learning framework that integrates pre-trained protein language model embeddings with environmental parameters to achieve accurate, condition-aware predictions of protein phase separation. We demonstrate LLPSenses predictive power and utility through three key experimental demonstrations. First, the model revealed that SGTA, previously unrecognized as a phase-separating protein, exhibits complex, temperature-dependent reentrant phase behavior. Second, LLPSense accurately predicted mutations in -synuclein that either enhance or suppress phase separation, enabling systematic mapping of residues potentially relevant to Parkinsons disease. Third, using model-guided mutagenesis, we inverted the phase behavior of UBQLN4, shifting it from high-temperature to low-temperature separation. Collectively, LLPSense provides a robust computational tool for interrogating the condition-dependent landscape of protein phase separation, enabling mechanistic studies of disease-associated phase separation and the rational design of programmable condensates.