Protein multi-level structure feature-integrated deep learning method for mutational effect prediction.

Area of Science:

• Computational biology
• Protein engineering
• Bioinformatics

Background:

• Protein engineering relies on directed evolution, but identifying optimal mutation sites is difficult due to vast sequence spaces and complex mutational interactions.
• Predicting the functional impact of mutations is crucial for enhancing protein properties but remains a significant challenge.

Purpose of the Study:

• To develop a deep learning-based approach, MLSmut, for predicting the effects of protein mutations.
• To leverage multi-level structural features for improved prediction accuracy in protein engineering.

Main Methods:

• MLSmut utilizes multi-level structural features, including protein co-evolution, sequence semantics, and geometric information.
• A two-stage training strategy involves coarse-tuning on unlabeled protein data and fine-tuning on experimental measurements.
• The model was evaluated on 10 single-site and two multi-site deep mutation scanning datasets.

Main Results:

• MLSmut significantly outperforms existing methods in predicting mutational outcomes across benchmark datasets.
• The two-stage training strategy effectively addresses limited training data availability.
• The model demonstrates satisfactory performance on downstream protein prediction tasks.

Conclusions:

• MLSmut offers a powerful computational tool to predict mutation effects, accelerating protein engineering efforts.
• This approach can reduce the need for laborious wet lab experiments.
• The findings enhance our understanding of genotype-phenotype relationships in proteins.