MMSol: Predicting Protein Solubility with an Antinoise Multimodal Deep Model

Area of Science:

• Biochemistry and Molecular Biology
• Computational Biology
• Protein Science

Background:

• Protein solubility is crucial for biological function, cellular processes, and therapeutic applications.
• Accurate computational prediction of protein solubility aids in developing therapeutic proteins and industrial enzymes.
• Current models are limited by their inability to integrate multimodal information and by label noise in experimental data.

Purpose of the Study:

• To develop a novel computational model, MMSol, for enhanced protein solubility prediction.
• To address limitations of existing models by incorporating multimodal protein information (sequence, structure, function).
• To mitigate the impact of label noise in experimental data through an integrated antinoise algorithm.

Main Methods:

• Developed the MMSol model integrating sequence, structure, and function modalities for enriched protein representation.
• Incorporated an antinoise algorithm during the training phase to handle noisy solubility data.
• Empirically evaluated the model's performance on both noise-free and noisy datasets.

Main Results:

• MMSol demonstrated superior performance in protein solubility prediction across both noisy and noise-free datasets.
• The integration of multimodal protein information significantly enhanced predictive capabilities.
• The antinoise algorithm effectively reduced the negative impact of label noise on model accuracy.

Conclusions:

• The MMSol model offers a robust approach to protein solubility prediction by leveraging multimodal data and noise-robust training.
• This advancement facilitates more accurate and efficient development of proteins for therapeutic and industrial applications.
• The findings highlight the importance of integrating diverse protein information and addressing data noise for improved computational predictions.