EZPro-Multi: Contrastive Learning-Enhanced Multi-property Prediction for Enzyme Engineering.

Area of Science:

• Biochemistry and Molecular Biology
• Computational Biology and Bioinformatics
• Enzyme Engineering and Biocatalysis

Background:

• Accurate prediction of enzyme mutant functional attributes is vital for enzyme engineering and biocatalysis.
• Existing methods often neglect crucial enzyme mutant-substrate interactions.
• A unified approach is needed to predict multiple biochemical properties simultaneously.

Purpose of the Study:

• To develop EZPro-Multi, a deep learning framework for predicting multiple enzyme mutant properties: catalytic efficiency (kcat), stability (ΔΔG), and solubility (ΔSol).
• To capture complex mutant-substrate interactions by integrating protein and substrate representations.
• To enhance prediction accuracy and discriminability through supervised contrastive learning and auxiliary classification.

Main Methods:

• Utilized ProtT5 for protein representations and Molformer for substrate representations.
• Implemented a cross-attention module to model mutant-substrate interactions.
• Incorporated supervised contrastive learning and an auxiliary classification head for improved feature learning and regression performance.

Main Results:

• Achieved state-of-the-art results on a curated kcat dataset, outperforming existing methods in regression accuracy and classification consistency.
• Demonstrated strong performance in predicting ΔΔG and ΔSol across multiple benchmark datasets.
• Showcased significant improvement in identifying high-activity mutants on deep mutational scanning data when integrating kcat, ΔΔG, and ΔSol.

Conclusions:

• EZPro-Multi offers a unified computational framework for multi-property assessment of enzyme variants.
• The framework provides practical value for prioritizing enzyme candidates in enzyme engineering.
• Integrating multiple properties enhances the identification of superior enzyme variants.