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

Directed evolution is a powerful strategy for tailoring enzyme substrate specificity, but it often requires negative screening or selection to eliminate variants with residual activity toward undesired substrates. Here, we present a machine learning-assisted workflow that replaces experimental counter-selection with an in silico negative filter trained on positive selection data for an undesired substrate. Using the promiscuous TEM-52 β-lactamase as a template, we randomized six active-site residues to construct a ~ 4 × 10⁸-member library. Data from two rounds of ampicillin (AMP) selection coupled with next-generation sequencing were used to train a convolutional neural network that predicts the activity toward AMP from sequence. We then performed selection on the cephalosporin ceftazidime (CAZ) and applied the AMP-trained model to triage target-enriched pools, prioritizing variants predicted to be AMP-inactive. This process yielded variants that supported cellular growth in CAZ but not in AMP and hydrolyzed CAZ with no detectable AMP activity in kinetic assays. Extending the approach to cefotaxime (CTX) and cephalothin (CET), which are structurally more similar to AMP, produced fewer AMP-inactive candidates and revealed partial convergence of sequence solutions. Cross-substrate profiling further partitioned specificity: CAZ-selected variants were largely inactive on CTX and CET, whereas CTX/CET-selected variants often showed overlapping activity within this cephalosporin subgroup. More broadly, the strategy proposed in this study could generalize to other proteins whenever sequence-to-activity data can be generated for undesired targets, enabling scalable engineering of orthogonality.