Abstract
Intrinsically disordered regions (IDRs) drive essential cellular functions but resist conventional structural-function annotation due to their dynamic conformations. Current computational methods struggle with cross-dataset generalization and functional subtype discrimination. We present IDPFunNet, a hybrid deep learning model integrating convolutional neural networks, bidirectional LSTM, residual MLP, and the protein language model ProtT5 to predict six IDR functional classes: five binding subtypes and disordered flexible linkers (DFLs). Its dual-path architecture decouples binding prediction from DFL identification. Leveraging ProtT5 evolutionary embeddings, which outperformed ESM-family models and AlphaFold2 structural features (by ≥1.3% average AUC and ≥ 12.7% average APS), IDPFunNet achieves state-of-the-art performance. Across six independent benchmarks, including CAID2/3 blind tests, it consistently surpasses existing general predictors DisoFLAG and DeepDISOBind in protein-binding prediction, with AUCs of 0.866 (TE210) and 0.832 (TE83), representing significant gains of 1.5%-8.1% in AUC and 13.5%-26.7% in APS (p-value < 0.05), while remaining competitive with specialized DFL predictors. Further analyses show multi-task learning enhances protein/lipid/small molecule-binding (3.1%-35.1% AUC gains), BiLSTMs are optimal for DFL identification, and self-attention shows potential for nucleic acid-binding (AUC 0.831). IDPFunNet thus provides an interpretable and generalizable framework for comprehensive IDR functional mapping. The webserver of IDPFunNet is freely available at https://yanglab.qd.sdu.edu.cn/IDPFunNet/ and the standalone package can be downloaded from https://github.com/IDRIDP/IDPFunNet/tree/master.