Extracting Latent Sequence-Function Information from Directed Evolution to Design Improved C-H Azidases

Abstract

Directed evolution is a powerful approach for enzyme engineering, but conventional workflows use only a fraction of the information generated during experimental screening. Typically, only top-performing variants are sequenced and propagated, while sequence-function information from the remainder of the library is discarded. Here, we develop a data-efficient strategy to recover and leverage this latent information for enzyme design. Screened variants are grouped into activity-defined bins, pooled, and subjected to long-read sequencing, enabling efficient acquisition of sequence-function relationships while preserving relative activity information. Ordinal machine learning models trained on these binned data are then used to predict beneficial mutations. We applied this strategy to SadX, an engineered Fe(II)- and α-ketoglutarate-dependent azidase previously evolved for site-selective C-H azidation. Using only data from a previous directed evolution campaign, the models identified mutations that significantly improved catalytic performance. Experimental validation showed that several predicted single mutations enhanced azidation activity by up to 3-fold relative to the parent enzyme. These results demonstrate that substantial gains in enzyme activity can be achieved by extracting additional information from existing directed evolution datasets, providing a general strategy for converting routinely generated screening data into actionable sequence-function information for enzyme design.