Protein representations: Encoding biological information for machine learning in biocatalysis

Area of Science:

• Biocatalysis and enzyme engineering
• Computational biology and machine learning

Background:

• Enzymes offer sustainable alternatives to conventional chemistry but require engineering for industrial applications.
• Machine learning (ML) can accelerate enzyme engineering by creating predictive models.
• Accurate ML models depend on effective conversion of biological data into numerical protein representations.

Purpose of the Study:

• To review critical methods for encoding protein information into numerical representations for ML.
• To explore requirements and inductive biases of primary sequence, 3D structure, and dynamics representations.
• To guide the selection of optimal protein representations for ML in biocatalysis.

Main Methods:

• Examination of established and emergent protein representation strategies.
• Categorization of representations into fixed (rule-based) and learned (neural network-derived) types.
• Introduction of combined protein-substrate representations for biocatalysis.

Main Results:

• Analysis of encoding approaches for primary sequence, 3D structure, and dynamics.
• Proposal of fixed and learned representation categories.
• Identification of model setup (dataset size, architecture) and objectives (property, mutant prediction, explainability) as key selection factors.

Conclusions:

• Effective protein representation is crucial for accurate ML models in enzyme engineering.
• The choice of representation depends on specific ML model parameters and research goals.
• This review provides a framework for selecting appropriate protein representations in biocatalysis research.