Liquid-liquid phase separation (LLPS) produces membraneless biomolecular condensates that regulate gene expression, stress response, and signal transduction. Distinguishing RNA-dependent from RNA-independent LLPS proteins is critical for understanding neurodegenerative and other human diseases, yet most sequence-based predictors cannot make this distinction. Here, we present LiPs—a hierarchical deep-learning framework to identify proteins with RNA-dependent LLPS propensity. Our method integrates ProtBERT transformer embeddings with 68 curated physicochemical descriptors. These features are first evaluated by an ensemble of machine-learning classifiers, and their outputs are combined through a stacked artificial neural network. This architecture captures both local sequence signals and long-range contextual patterns. Model interpretability analyses using SHAP and feature-correlation metrics show that the most influential ProtBERT dimensions correspond to known LLPS-promoting sequence features. On an independent test set, the framework achieves 96.54% accuracy, an F1-score of 0.965, and Matthews correlation coefficient of 0.932, surpassing existing LLPS predictors and outperforming the only previously reported method for RNA-dependent LLPS classification. These results demonstrate that transformer-based protein embeddings encode latent patterns predictive of RNA-mediated phase separation. Our model provides a robust, interpretable tool for distinguishing RNA-dependent LLPS proteins and establishes a foundation for broader applications in protein function prediction.
Chatterjee et al. (Sun,) studied this question.