Prediction of liquid-phase separation proteins using Siamese network with feature fusion

Liquid-liquid phase separation (LLPS) is a common and important phenomenon where biomolecules form dynamic, membrane-less condensates through multivalent interactions, spontaneously separating into distinct concentration-dense and dilute phases. Research has shown that LLPS is associated with a wide range of cellular functional regulation. In this work, we establish a feature fusion framework based on a Siamese network for the prediction of LLPS proteins, which can integrate automatically extracted features from the protein itself and the protein-protein interaction (PPI) networks, and achieve good accuracy even in small sample sets. We used two representative graph embedding methods, Node2vec and DeepNF, to extract the embedding features of PPI networks and compared the impact of the two methods on model performance at different feature lengths. Our work provides a way for integrating multivalent interactions between proteins that drive LLPS, as well as a flexible framework for the fusion of different types of protein features, not only for LLPS prediction but also for other downstream prediction tasks. All relevant materials can be found at https://github.com/ispotato/SiameseNetwork_LLPS.