MPSol: A Multimodal Prompt Learning Framework for Protein Solubility Prediction.

Area of Science:

• Computational protein engineering
• Biotechnology
• Artificial intelligence in drug discovery

Background:

• Protein solubility is vital for biologic drug developability, stability, and efficacy.
• Accurate prediction of protein solubility is a significant challenge in computational protein engineering.
• Existing methods struggle with the complexity of protein sequences for solubility prediction.

Purpose of the Study:

• To develop an advanced computational framework for accurate protein solubility prediction.
• To integrate diverse protein data representations for enhanced predictive power.
• To leverage large language models (LLMs) and prompt learning for protein property prediction.

Main Methods:

• Proposed a multimodal prompt learning framework (MPSol) integrating primary sequences, structural proxies, and LLM-generated text.
• Utilized a unified multimodal backbone with a cross-modal fusion module for inter-modal interaction.
• Designed label-aware prompts for semantic supervision and alignment of fused protein representations.

Main Results:

• MPSol achieved state-of-the-art performance on the PDBSol test set (Accuracy: 0.815, AUC: 0.867, MCC: 0.642).
• Demonstrated strong generalization to an external out-of-distribution test dataset (Accuracy: 0.632, AUC: 0.653, MCC: 0.332).
• Indicated the effectiveness of prompt-driven multimodal learning for protein property prediction.

Conclusions:

• MPSol offers an interpretable and effective approach for protein solubility prediction.
• Multimodal learning integrating LLM-derived features significantly enhances predictive accuracy.
• The framework shows promise for accelerating the development of stable and effective biologic drug candidates.