Machine learning-assisted engineering of substrate-specific β-lactamases.

Area of Science:

• Biochemistry and Molecular Biology
• Enzyme Engineering
• Computational Biology

Background:

• Directed evolution is key for tailoring enzyme specificity but often needs laborious negative selection against undesired substrates.
• Existing methods for enzyme engineering struggle with efficiently eliminating residual activity toward off-target molecules.

Purpose of the Study:

• To develop and validate a machine learning-assisted workflow to replace experimental counter-selection in enzyme engineering.
• To engineer enzymes with enhanced specificity by computationally predicting and filtering out undesired substrate activity.

Main Methods:

• Constructed a large library (~4x10^8 members) by randomizing active-site residues of TEM-52 beta-lactamase.
• Used next-generation sequencing data from ampicillin (AMP) selection to train a convolutional neural network (CNN) for predicting AMP activity.
• Applied the in silico AMP-inactivity filter to triage variants selected for activity against ceftazidime (CAZ), cefotaxime (CTX), and cephalothin (CET).

Main Results:

• Successfully identified enzyme variants with high activity on CAZ but no detectable activity on AMP.
• Demonstrated the workflow's applicability to structurally similar substrates (CTX, CET), yielding fewer inactive candidates but showing partial sequence convergence.
• Cross-substrate profiling revealed distinct specificity profiles for variants selected against different cephalosporins.

Conclusions:

• The machine learning-based in silico negative filter effectively replaces experimental counter-selection for engineering enzyme specificity.
• This strategy is generalizable to other proteins and targets, enabling scalable engineering of functional orthogonality.
• The workflow accelerates the development of enzymes with precisely tailored substrate specificities.