PhaseNet: A computational framework for identifying phase-separating proteins based on protein language model

doi:10.1016/j.ijbiomac.2025.149044

PhaseNet: A computational framework for identifying phase-separating proteins based on protein language model

2025-11-13

0

PRE

AI

DOI：10.1016/j.ijbiomac.2025.149044

Original

Original request for help

Save

Abstract

En 中文

Biomolecular condensates formed via liquid–liquid phase separation (LLPS) play vital roles in diverse cellular processes, and their dysregulation has been linked to numerous diseases. Accurate identification of phase-separating proteins is therefore essential for elucidating the molecular mechanisms underlying these condensates. We present PhaseNet, a dual-task computational framework designed to both distinguish LLPS proteins from non-LLPS proteins and classify LLPS proteins into self-assembling (PS-Self) and partner-dependent (PS-Part) categories. In the first task, PhaseNet integrates features from pretrained protein language models (ESM), sequence encodings (ZSCALE and BLOSUM), and a modular architecture combining CNN-BiGRU and multi-head attention mechanisms. These heterogeneous features are fused using an attention-guided strategy and optimized with HSIC-based regularization to enhance discriminative performance. For the second task, PhaseNet employs Lasso-based feature selection on ESM-derived embeddings, followed by a stacking ensemble of five classifiers: Random Forest, Extra Trees, GBDT, XGBoost, and MLP. Comprehensive benchmarking across multiple independent test sets demonstrates that PhaseNet surpasses existing LLPS predictors in both general identification and fine-grained classification. This modular and interpretable framework offers a robust tool for the systematic discovery and annotation of phase-separating proteins.

PhaseNet: A computational framework for identifying phase-separating proteins based on protein language model

Abstract

Journal

Organization