MTPSol: Multimodal Twin Protein Solubility Prediction Architecture Based on Pretrained Models

Area of Science:

• Biochemistry
• Computational Biology
• Protein Engineering

Background:

• Protein solubility is a critical factor for efficient functional expression in enzyme mining and de novo design.
• Accurate solubility prediction algorithms can reduce experimental costs and improve protein engineering success rates.
• Existing models often lack accuracy and generalization due to limited use of protein structural information.

Purpose of the Study:

• To develop an advanced protein solubility prediction architecture named MTPSol.
• To enhance prediction accuracy by integrating multimodal protein features, including structural information.
• To improve the efficiency of screening natural enzymes and facilitate de novo protein design.

Main Methods:

• Developed MTPSol, a novel architecture utilizing pretrained models for feature extraction from multimodal protein inputs.
• Implemented cross-modal twin attention and multiscale feature networks to integrate diverse protein features.
• Validated the architecture using public benchmark datasets and a custom transaminase dataset.

Main Results:

• MTPSol achieved competitive predictive performance on public benchmark datasets.
• The architecture demonstrated superior performance compared to state-of-the-art models on a validated transaminase dataset.
• Results indicate strong generalization capabilities across different protein families.

Conclusions:

• MTPSol offers a more efficient method for screening natural enzymes.
• The architecture shows significant potential for advancing de novo protein design.
• Integrating multimodal features significantly enhances protein solubility prediction accuracy and applicability.