Predicting novel substrates for enzymes with minimal experimental effort with active learning.

Area of Science:

• Biochemistry
• Computational Biology
• Enzymology

Background:

• Enzyme promiscuity is widespread, impacting metabolism and biocatalysis.
• Current methods for characterizing enzyme promiscuity use limited, potentially unrepresentative compound sets.
• Efficiently exploring enzyme versatility requires improved substrate selection strategies.

Purpose of the Study:

• To develop and validate a computational approach for efficient characterization of enzymatic substrate promiscuity.
• To improve the selection of informative substrates for experimental testing.
• To enhance the understanding of enzyme versatility for biocatalysis and metabolic engineering.

Main Methods:

• Utilized existing experimental data and tested additional compounds for four enzymes.
• Developed support vector machine (SVM) models to predict enzymatic activity.
• Employed an active learning strategy to select novel compounds for experimental validation.

Main Results:

• SVM models trained on diverse compound sets achieved ~80% accuracy using 33% fewer compounds compared to traditional methods.
• Active learning effectively resolved data conflicts and introduced chemical diversity.
• The approach demonstrated a more efficient exploration of enzyme substrate versatility.

Conclusions:

• Computational active learning significantly enhances the efficiency and thoroughness of characterizing enzyme promiscuity.
• This method offers a valuable tool for designing novel metabolic pathways and optimizing biocatalysis.
• The developed SVM models can predict high-probability promiscuous enzymatic reactions, aiding future research.