Hybrid deep learning with protein language models and dual-path architecture for predicting IDP functions

Area of Science:

• Computational Biology
• Bioinformatics
• Structural Biology

Background:

• Intrinsically disordered regions (IDRs) are crucial for cellular functions but challenging to annotate due to their dynamic nature.
• Existing computational tools face limitations in generalizing across datasets and distinguishing functional subtypes of IDRs.

Purpose of the Study:

• To develop an advanced deep learning model, IDPFunNet, for accurate prediction of six IDR functional classes.
• To improve the functional annotation of IDRs by overcoming limitations of current computational methods.

Main Methods:

• IDPFunNet integrates convolutional neural networks, bidirectional LSTM, residual MLP, and the ProtT5 protein language model.
• A dual-path architecture separates binding prediction from disordered flexible linker (DFL) identification.
• ProtT5 evolutionary embeddings were leveraged, outperforming other models and structural features.

Main Results:

• IDPFunNet achieved state-of-the-art performance, surpassing general predictors like DisoFLAG and DeepDISOBind.
• Significant improvements in Area Under the Curve (AUC) and Accuracy in Prediction Score (APS) were observed in benchmark tests, including CAID2/3.
• Multi-task learning enhanced binding predictions, BiLSTMs proved optimal for DFL identification, and self-attention showed promise for nucleic acid-binding prediction.

Conclusions:

• IDPFunNet offers a robust, interpretable, and generalizable framework for comprehensive IDR functional mapping.
• The model demonstrates superior performance in predicting various IDR functions, including binding subtypes and DFLs.
• IDPFunNet provides a valuable tool for advancing our understanding of IDR roles in biological processes.