Pro4S: Prediction of protein solubility by fusing sequence, structure, and surface

Area of Science:

• Biochemistry
• Computational Biology
• Drug Discovery

Background:

• Protein solubility is crucial for stability, efficacy, and developability in drug discovery.
• Experimental solubility assessment is time-consuming and resource-intensive.
• Existing AI models often focus on sequence data and underutilize structural and surface features.

Purpose of the Study:

• To develop a multimodal AI model, Pro4S, for accurate protein solubility prediction.
• To integrate sequence, structural, and surface features for enhanced predictive performance.
• To provide a robust tool for both qualitative and quantitative solubility assessment in drug discovery and protein engineering.

Main Methods:

• Developed Pro4S, a novel multimodal model integrating sequence, structural, and surface descriptors.
• Employed advanced contrastive learning techniques for feature integration.
• Benchmarked Pro4S against state-of-the-art predictors on diverse datasets.

Main Results:

• Pro4S demonstrated significant improvements in prediction accuracy, robustness, and generalizability for both classification and regression tasks.
• The model consistently outperformed existing state-of-the-art predictors.
• Application in de novo protein design showed a strong correlation between predicted solubility and experimental expression, reducing non-expressed proteins by 52.7%.

Conclusions:

• Pro4S offers a reliable computational approach for predicting protein solubility.
• The model effectively integrates multimodal data, outperforming existing methods.
• Pro4S can serve as an upfront screening tool to accelerate protein engineering and improve expression success rates.