Deep Learning for RNA Synthetic Biology

Abstract

Engineered RNA elements are programmable tools capable of detecting small molecules, proteins, and nucleic acids. Predicting the behavior of these tools remains a challenge, a situation that could be addressed through enhanced pattern recognition from deep 30 learning. Thus, we investigate Deep Neural Networks (DNN) to predict toehold switch function as a canonical riboswitch model in synthetic biology. To facilitate DNN training, we synthesized and characterized in vivo a dataset of 91,534 toehold switches spanning 23 viral genomes and 906 human transcription factors. DNNs trained on nucleotide sequences outperformed (R 2 =0.43-0.70) previous state-of-the-art thermodynamic and kinetic models (R 2 =0.04-0.15) and allowed 35for human-understandable attention-visualizations (VIS4Map) to identify success and failure modes. This deep learning approach constitutes a major step forward in engineering and understanding of RNA synthetic biology. 40One Sentence Summary: Deep neural networks are used to improve functionality prediction and provide insights on toehold switches as a model for RNA synthetic biology tools.

Sequence-to-function deep learning frameworks for synthetic biology

“…Aside from the observed conservation of the Shine-Dalgarno (SD) and start codon sequences at positions 31-41 and 48-50, respectively, the sequence logo constructed from all 91,534 toeholds confirms that each of the four nucleotides are relatively evenly distributed at each position 23 .…”

“…A dataset of 244,000 toehold switches ( Fig. 1A), including sequences tiled from viruses and the human genome as well as random sequences (see Methods), have been tested experimentally by Angenent-Mari et al 23 , with 91,534 switches meeting well-defined quality control criteria. Each toehold sequence is 59 nucleotides in length, with the first 30 nucleotides distinguishing the unique unstructured region followed by part of the hairpin; the remaining 29 nucleotides can be inferred by hairpin complementarity as well as Shine-Dalgarno sequence and start codon conservation (Table S1).…”

“…In order to define the sequences to be tested (see Angenent-Mari et al 23 ), we perform a sequence tiling on a variety of prokaryotic genomes, as well as the complete set of human transcriptional regulators. Briefly, each chosen sequence was tiled with a sequence length of 30 nucleotides and a stride of 5 nucleotides.…”

“…244,000 toehold sequences were tested by Angenent-Mari et al 23 and the experimental data was obtained as logarithm-transformed GFP fluorescence measured at both the modified ON state (with trigger present, fused to the switch sequence) and OFF state (without trigger).…”

“…Machine learning approaches have been applied successfully to synthetic biology 1,19 , as exemplified in a recent study by Yang et al 19 using a 'white box' approach to extract antibiotic mechanisms of action, and in motif finding and DNA sequence prediction tasks [20][21][22] . As typical deep learning approaches require large amounts of training examples, we used a dataset from Angenent-Mari et al (2019) 23 that experimentally characterized 91,534 toehold switches, in both active and repressed states. We fed these switch sequences as input to a multi-layer convolutional neural network that can predict the ON and OFF state performances simultaneously.…”