Discovering efficient enzymes for newly designed biosynthetic pathways remains a major challenge in synthetic biology. Large enzyme families often contain thousands of poorly characterized sequences, while conventional homology-based searches provide limited information about catalytic performance. Exhaustive experimental screening, on the other hand, is labor-intensive and costly. This challenge is particularly relevant to the biosynthesis of noncanonical amino acids, where key pathway steps often lack well-characterized and efficient biocatalysts.
Recently, a research team led by Prof. LUO Xiaozhou and Associate Prof.WANG Xinran from the Shenzhen Institutes of Advanced Technology (SIAT), Chinese Academy of Sciences, developed an AI-guided framework for the discovery and engineering of sulfotransferases for intracellular sulfotyrosine (sTyr) production. The study was published in Chemical Engineering Journal.
The researchers used UniKP, a deep-learning model for predicting enzyme kinetic parameters, to prioritize sulfotransferases with the potential to catalyze tyrosine sulfation. They also established a genetic-code-expansion-based fluorescent biosensor that links intracellular sTyr production to sfGFP fluorescence, enabling rapid experimental validation. Among the top four candidates ranked by UniKP, two showed clear activity, corresponding to a 50% hit rate.
To rigorously evaluate whether AI-guided prioritization outperformed conventional strategies, the researchers compared UniKP with BLAST- and sequence-similarity-network-based approaches. From 1,000 homologous sequences identified by BLAST, UniKP re-ranked the candidates, and testing only the top five was sufficient to identify EfST1A1, whose in vitro activity was approximately 6.5-fold higher than that of the initial hit. In contrast, none of the top BLAST-ranked candidates outperformed the starting enzyme, while improvements observed for SSN-selected candidates were not statistically significant.
The AI-prioritized EfST1A1 was then further improved through directed evolution, generating evo_EfST1A1, which showed a 7.25-fold increase in cellular fluorescence and an approximately sixfold increase in specific activity while retaining broad substrate promiscuity.
Following metabolic optimization in Escherichia coli, the engineered strain produced approximately 1,578 μM sTyr and efficiently supported multi-site sTyr incorporation into proteins. Together, these findings show that AI can move beyond in silico prediction to serve as a practical tool for experimental decision-making in enzyme discovery, protein engineering, and synthetic pathway development.
File Download：