SRPM-Sol: A Structure Robust Protein Multimodal Model for Solubility Prediction.

Area of Science:

• Computational biology
• Protein science
• Bioinformatics

Background:

• Protein solubility is crucial for expression and purification, impacting functional studies and experimental costs.
• Current deep learning models for protein solubility often rely on limited multimodal inputs (sequences, structures) and struggle with prediction errors.
• Existing multimodal approaches overlook key factors influencing protein solubility and lack robustness against structural inaccuracies.

Purpose of the Study:

• To develop a novel multimodal protein solubility prediction model (SRPM-Sol) that enhances accuracy and robustness.
• To address limitations of existing models by incorporating a wider range of input features.
• To validate the model's performance using a newly constructed hierarchical dataset.

Main Methods:

• Developed SRPM-Sol, a multimodal model leveraging the ESM3 framework.
• Integrated diverse data types: amino acid sequences, predicted structure information, secondary structure sequences, and physicochemical properties.
• Created the first hierarchical dataset, PDE-Sol, categorized by Predicted Local Distance Difference Test (pLDDT) scores.

Main Results:

• SRPM-Sol demonstrated superior accuracy and robustness compared to baseline methods.
• The model performed effectively across different levels of the PDE-Sol dataset, even with uncertain structural data.
• The integration of diverse multimodal inputs proved beneficial for prediction performance.

Conclusions:

• SRPM-Sol represents a significant advancement in multimodal protein solubility prediction.
• The model's diverse input strategy enhances prediction accuracy and robustness, particularly in the presence of structural uncertainties.
• The PDE-Sol dataset provides a valuable resource for evaluating protein solubility prediction models.