Enzyme function prediction with interpretable models

Area of Science:

• Biochemistry
• Bioinformatics
• Computational Biology

Background:

• Enzymes are crucial for metabolic pathways, with gene assignment to reactions being key for genome analysis.
• Traditional sequence similarity methods struggle to identify enzymes with similar functions but different sequences, limiting metabolome mapping.
• A new approach is needed to predict enzyme function beyond sequence homology.

Purpose of the Study:

• To develop a tool for predicting enzyme function, specifically Enzyme Commission (EC) numbers, using basic protein properties.
• To complement existing sequence and structure-based prediction techniques.
• To establish a direct relationship between protein properties and their enzymatic functions.

Main Methods:

• Collected 453 features and properties related to protein structure and function.
• Employed a mixture model of stochastic decision trees to learn complex feature-function relationships.
• Tested the model on Pfam (sequence-based) and EC (function-based) classifications.

Main Results:

• The model effectively learns relationships in highly diverged protein families, including those not defined by sequence.
• Identified key protein properties strongly correlated with structural and functional aspects.
• Demonstrated the model's utility in complementing sequence-based enzyme identification.

Conclusions:

• Protein properties can be effectively used to predict enzyme function, overcoming limitations of sequence-based methods.
• The developed model aids in mapping the metabolome and understanding protein family definitions.
• This approach offers a valuable tool for functional genomics and enzyme discovery.