Probing the Specificity of Fluorescent Deoxyribozymes Using Single-Step Selections and Machine Learning.

The ability of proteins and nucleic acids to form specific binding sites for ligands is critical for biological function, and methods to modulate biochemical specificity are important for fields such as enzyme engineering and drug design. Here, we systematically investigated the specificities of self-phosphorylating deoxyribozymes that convert the coumarin substrate 4-MUP into a fluorescent product using biochemical assays, single-step selections, and machine learning. Activity assays using a panel of 20 catalytic motifs and 10 substrates that generate different types of signals when they are dephosphorylated revealed that these deoxyribozymes are extremely specific for 4-MUP. To identify mutations that change specificity, we constructed a library based on a self-phosphorylating fluorescent deoxyribozyme called Aurora. A series of single-step selections yielded variants that react with 4-MUP and the structurally similar substrate diFMUP, but not with the more distinct substrates pNPP and ELF. Pairwise analysis of sequences in the 4-MUP and diFMUP data sets revealed four mutations that modulate Aurora specificity. The effects of these mutations were confirmed using biochemical assays and could be predicted using models developed by machine learning. Taken together, our results show how single-step selections can be used to identify mutations that change the specificity of a deoxyribozyme. They also highlight how machine learning can be used to model complex data sets from in vitro selection experiments.