Using structured libraries, selection, and machine learning to rapidly explore the sequence space of a fluorescent deoxyribozyme

Finding ways to more comprehensively explore the sequence space of complex functional motifs is an important and unresolved question in nucleic acid engineering. Standard approaches use libraries in which a single variant of a motif is randomly mutagenized at a low level. This provides comprehensive coverage of sequence space over short mutational distances, but only limited information about more distant variants. Here we describe a new approach that uses libraries made up of sequences consistent with the multiple constraints of a desired target motif. Functional variants are rapidly identified in a single round of selection followed by high-throughput sequencing, and rules relating sequence to function are elucidated using machine learning. This method was tested using a fluorescent deoxyribozyme recently discovered in our group called Aurora. Single-step selections showed that a secondary structure library based on Aurora contained ~40-fold more unique catalytic sequences than one generated by random mutagenesis. Furthermore, models developed by machine learning could quantitatively predict read numbers and identify the most active variants using small subsets of sequences as training sets. By combining secondary structure libraries, selection, and machine learning in this way, sequence space can be explored far more quickly and efficiently than in standard approaches.