Machine learning reveals sequence-function relationships in family 7 glycoside hydrolases

Area of Science:

• Biochemistry
• Enzymology
• Computational Biology

Background:

• Family 7 glycoside hydrolases (GH7) are key enzymes in cellulose degradation, crucial for natural and industrial processes.
• GH7 enzymes are often bimodular, featuring catalytic and carbohydrate-binding domains (CBM), with active sites binding cello-oligomers.
• Understanding the sequence-structure-function relationships in GH7 enzymes, including cellobiohydrolases (CBH) and endoglucanases (EG), remains incomplete.

Purpose of the Study:

• To apply machine learning for data-driven insights into GH7 enzyme sequence, structure, and function.
• To identify key sequence features correlating with GH7 functional subtypes (CBH vs. EG) and CBM presence.
• To explore novel sequence positions potentially influencing GH7 functional variation.

Main Methods:

• Machine learning models were trained using active-site loop residue counts to classify GH7 subtypes.
• Classification rules were derived based on specific residue positions predicting functional subtype.
• A random forest model analyzed catalytic domain residues to predict CBM presence.

Main Results:

• Machine learning models accurately discriminated between GH7 CBHs and EGs (up to 99% accuracy) based on active-site loop lengths (A4, B2, B3, B4).
• Specific residues at 42 sequence positions predicted functional subtype with over 87% accuracy.
• A random forest model predicted CBM presence with 89.5% accuracy using 19 catalytic domain positions.

Conclusions:

• Active-site loop lengths and specific sequence positions are strong predictors of GH7 functional subtypes and CBM presence.
• Machine learning successfully recapitulates and expands upon experimentally identified key sequence positions for GH7 activity.
• These findings offer a foundation for understanding and engineering GH7 enzymes for improved cellulose degradation.