Using machine learning to predict the effects and consequences of mutations in proteins

Area of Science:

• Biochemistry
• Computational Biology
• Protein Engineering

Background:

• Massive datasets of protein sequences, structures, and mutational effects are increasingly available.
• Machine and deep learning (ML) models can leverage these datasets for variant prediction.
• Systematic benchmarking highlights data availability and quality as key constraints for ML model performance.

Purpose of the Study:

• To review the application of machine and deep learning in predicting protein variants with improved fitness.
• To discuss the impact of data availability and quality on predictive model performance.
• To explore strategies for leveraging limited experimental data in protein ML models.

Main Methods:

• Review of current machine and deep learning approaches for protein variant prediction.
• Analysis of benchmarking studies on ML model performance.
• Discussion of unsupervised, self-supervised, hybrid, and transfer learning strategies.

Main Results:

• ML and deep learning can predict protein variants with enhanced fitness.
• Data availability and quality are more critical than specific ML algorithm choices.
• Unsupervised/self-supervised pre-training on generic datasets can be effective with subsequent refinement.
• Recent progress in ML for protein science has been substantial.

Conclusions:

• Machine learning approaches are poised to significantly impact future breakthroughs in protein biochemistry and engineering.
• Effective strategies exist for utilizing limited experimental data through pre-training and transfer learning.
• Continued development and application of ML will accelerate protein design and understanding.