Zero-shot prediction of mutation effects with multimodal deep representation learning guides protein engineering

Area of Science:

• Biotechnology and Bioinformatics
• Computational Biology
• Molecular Biology

Background:

• Predicting the functional impact of amino acid mutations is crucial for biotechnology and biomedicine.
• Accurate and unsupervised mutation effect prediction remains a significant challenge.
• Existing methods often require multiple sequence alignments, limiting their applicability.

Purpose of the Study:

• To introduce Protein Mutational Effect Predictor (ProMEP), a novel method for zero-shot prediction of mutation effects.
• To develop a multimodal deep representation learning model for comprehensive sequence and structure context learning.
• To demonstrate ProMEP's capability in guiding protein engineering for enhanced gene-editing tools.

Main Methods:

• Developed ProMEP, a general, multiple sequence alignment-free method for mutation effect prediction.
• Employed a multimodal deep representation learning model trained on ~160 million proteins.
• Utilized ProMEP to forecast consequences of mutations in TnpB and TadA gene-editing enzymes.

Main Results:

• ProMEP achieves state-of-the-art performance in mutation effect prediction with significant speed improvements.
• Engineered TnpB variants show enhanced gene-editing efficiency (e.g., 74.04% for a 5-site mutant vs. 24.66% wild type).
• Developed TadA-based base editors with high A-to-G conversion frequency (up to 77.27%) and reduced off-target effects.

Conclusions:

• ProMEP offers a powerful and efficient approach for predicting protein mutational effects.
• The method successfully guides the engineering of high-performance gene-editing tools.
• ProMEP facilitates exploration of protein space and practical protein design in synthetic biology and biomedicine.