Prediction of Physical Characteristics of Disordered Proteins Using Molecular Simulation and Physics-Informed...

Area of Science:

• Computational Biology
• Biophysics
• Machine Learning

Background:

• Intrinsically disordered proteins (IDPs) lack stable 3D structures, posing challenges for traditional protein modeling.
• Predicting IDP conformational properties is crucial for understanding their biological functions.

Purpose of the Study:

• To develop a novel hybrid machine learning (ML) framework for accurate prediction of IDP conformational properties, such as radius of gyration.
• To integrate sequence information with physical features using an attention mechanism for enhanced predictive power.

Main Methods:

• A hybrid ML framework combining sequence-based models (e.g., GRU, biGRU) with 23 physical features.
• An attention mechanism to weigh residue importance and a shared latent space for feature fusion.
• Training and evaluation on Brownian dynamics (BD) simulation data for ~7000 IDPs from the MobiDB database.

Main Results:

• The hybrid biGRU model achieved the best predictive performance, outperforming sequence-only and feature-only models.
• Attention-guided fusion significantly improved accuracy metrics (e.g., mean absolute percentage error, mean squared error).
• SHAP and integrated gradient analyses identified key features (e.g., sequence charge, hydropathy, asymmetry) and protein length influencing predictions.

Conclusions:

• The developed ML framework offers a fast, interpretable, and scalable tool for predicting IDP behavior.
• The model enables efficient initial screening of IDPs, reducing the need for extensive molecular simulations.
• Feature importance analysis aids in understanding IDP conformational determinants and guides feature selection for improved generalization.