NAND Hybrid Riboswtich Design by Deep Batch Bayesian Optimization

The design of large genetic circuits requires genetic regulatory devices capable of performing complex logic operations. Hybrid riboswitches, synthetically enhanced compact RNA elements (<100 nucleotides) that form a tertiary structure with the ability to specifically bind two different target molecules, can be used to design genetic regulators that emulate Boolean logic. When inserted into the 5' UTR of an mRNA, these devices can regulate translation initiation upon specific binding of one or both ligands. The goal of this study is to design hybrid riboswitches that emulate Boolean NAND logic in yeast. We propose a novel machine learning-based design framework combining high-throughput in vivo screening and deep Bayesian optimization. Through an initial screening, we discovered a hybrid riboswitch with NAND behavior. Using batch Bayesian optimization with an ensemble neural network as surrogate, we further improve the NAND functionality of our hybrid riboswitch with respect to a performance score, thereby achieving near digital NAND behavior. With its focus on model-based and score-driven design, our proposed method can complement experiment driven approaches by allowing fine grained adaptation of functionality, including constructs sensitive to single nucleotide changes. Overall design: During the first screening a communication module (connecting stem between the two binding pockets of the hybrid ribowitch) conferring NAND behavior was discovered using a massive parallel reporter assay. The fluorescence range of GFP regulated by the construct was divded into 8 gates using controls, with the areas occupied by the controls being gated as well. This sorting grid was used for every ligand condition of the hybrid riboswitch (no ligand(wo), tetracycline(Tc), neomycin(Neo), both(TcNeo)). The communication module sequence was randomized and a yeast culture containing the pool sequence upstream of a gfp gene was inoculated for each ligand condition. Each culture was separatly sorted using the same grid and the plasmid DNA from every subpopulation was extracted afterwards. Construct sequences were amplified from the plasmid DNA and sequenced using the AMPLICON-EZ service form Azenta/GENEWIZ. The NGS data of each subpopulation under every ligand condition was then used to approximate the switching pattern of each sequence contained within the sequence pool. The entire screening process was repeated once (replicate). For the second screening the base stem of a construct containing a NAND communication module discovered during the first screening was randomized and the sequence pool was sorted for low GFP fluorescence in the presence of both ligands and different levels of GFP fluorescence in the absence of both ligands (P7: 20-40%; P8: 40-60%; P9: 60-80%; P10: 80-100%). Plasmid DNA was again extracted from each subpopulation and the randomized area was amplified to be analyzed via AMPLICON-EZ by Azenta/GENEWIZ.