Accelerating the Understanding and Design of Intracellular Biosensors by Massively Multiplexed Experimentation and Machine Learning

Abstract

Recent progress in large-scale DNA synthesis and next-generation DNA sequencing technology have enabled studies of biological processes at a massive scale. These studies can be further coupled to advanced computational methods using machine learning to explore and reveal essential elements of biological function. Two separate factors controlling gene expression are studied here using such a paradigm: the lac repressor protein which can regulate transcription of DNA, and a riboregulatory toehold switch that can control translation. The function of the lac repressor is interesting since it intrinsically couples the binding of a small molecule to the binding of DNA and has emerged as a useful tool in synthetic biology as an intracellular biosensor. A deep neural network was developed to predict transcriptional repression mediated by the lac repressor, using 43,669 experimental measurements of variant function. When validated across ten separate training and testing splits of single mutations in the lac repressor, our best performing model achieved a median Pearson correlation of 0.79, exceeding any previous model. Deep representation learning approaches, first trained in an unsupervised manner across millions of diverse proteins, can be fine-tuned in a supervised fashion using lac repressor experimental datasets to more effectively predict a variant effect on repression. Separately, engineered RNA elements are programmable tools capable of detecting small molecules, proteins, and nucleic acids. To facilitate understanding and design of one such RNA element, the toehold switch, we synthesized and characterized in vivo a dataset of 91,534 toehold switches spanning 23 viral genomes and 906 human transcription factors. Deep neural networks trained on nucleotide sequences outperform (R2=0.43-0.70) previous state-of-the-art thermodynamic and kinetic models (R2=0.04-0.15) and allow for human-understandable attention-visualizations to identify success and failure modes. Thus, two factors controlling gene expression are studied here in the context of large-scale mutational scanning and machine learning in order to understand and design such factors toward effective intracellular biosensing.