PCTPS: Accurate prediction of phase-separated proteins using protein language model embeddings and a CTN-KAN...

Phase-separated proteins play a critical role in understanding cellular mechanisms and functions. However, developing accurate and efficient computational methods for predicting phase-separated proteins remains challenging. Existing methods often rely on feature engineering, which limits their generalizability and performance, leaving substantial room for improvement. In this study, we propose PCTPS, a novel predictor that combines a pretrained protein language model (PPLM) with deep neural networks to predict phase-separated proteins. Specifically, PPLM is used to extract embedding features from protein sequences. These features are then processed using a Convolutional-Transformer Network to capture both local and global discriminative information. To further enhance classification performance, we introduce the Kolmogorov-Arnold Network. Notably, PCTPS not only predicts whether a protein is phase-separated but also distinguishes self-assembled phase-separated proteins and partner-dependent phase-separated proteins. Extensive experiments on human and non-human datasets demonstrate that PCTPS outperforms state-of-the-art methods. Our results validate the effectiveness of integrating PPLM and neural networks for phase-separated protein prediction.