SurfSol: A Multimodal Surface-Based Deep Learning Framework for Protein Solubility Prediction.

Area of Science:

• Computational biology
• Biotechnology
• Protein science

Background:

• Protein solubility is crucial for recombinant protein applications in pharmaceuticals and biotechnology.
• Predicting protein solubility accurately remains a significant challenge despite advancements in experimental methods.

Purpose of the Study:

• To introduce SurfSol, a novel computational approach for predicting protein solubility.
• To leverage explicit surface representations, integrating geometry and physicochemical properties for improved prediction.

Main Methods:

• SurfSol utilizes an E(3)-equivariant graph neural network incorporating surface geometry and properties (electrostatics, hydropathy, hydrogen-bonding).
• It combines surface data with sequence embeddings (ESM-2) and structural features (TransformerConv).
• The model was evaluated on the eSOL dataset.

Main Results:

• SurfSol achieved an R² of 0.555 and an AUC of 0.895 on the eSOL dataset.
• The model outperformed existing protein solubility predictors.
• Ablation studies confirmed the importance of surface, sequence, and structural data.

Conclusions:

• Explicit surface modeling is vital for accurate protein solubility prediction.
• SurfSol offers a generalizable framework for predicting other protein properties.