Machine Learning-Guided Engineering of Cre-lox Recombination for Comprehensive Analysis of Neural Networks

• Published in: The FASEB journal Vol. 36 Suppl 1
• Main Authors: Yamauchi, Yuji, Ueda, Mitsuyoshi, Aoki, Wataru
• Format: Journal Article
• Language: English
• Published: United States 01.05.2022
• ISSN: 1530-6860
• DOI: 10.1096/fasebj.2022.36.S1.R3325

Neural networks generate complex behaviors. However, the relationships between them have not been understood well. Thus, we have developed 'functional cellomics', which enables functional annotation of neural networks in a hypothesis-free manner. The key is the stochastic labeling of opsin achieved by Cre-mediated stochastic recombination. We designed a gene cassette, which contained two sets of lox variants, lox2272 and loxP, alternately. A transcription factor, QF2w, is interposed between loxP. Cre excises DNA with an exclusive choice of between lox2272 or loxP. When Cre excises lox2272, expressed QF2w induces opsin. Constructing a C. elegans strain carrying this gene cassette, we successfully implemented the stochastic labeling of opsin dependent on Cre induction [1]. However, the labeling rate of opsin was about 30 %. If a large subset of neurons is labeled by opsin, it becomes difficult to distinguish the function of each neuron. In this study, we demonstrated a strategy to precisely control the labeling rate of opsin by machine learning-guided engineering of Cre-lox recombination. In our genetic circuit, the opsin coding gene is expressed in neurons where Cre-lox recombination occurs between lox2272 sequences. Therefore, we hypothesized that the labeling rate of opsin would be reduced by using lox2272 variants, which is difficult to be excised by Cre. Thus, we generated a library of randomized lox sequences by PCR and introduced them in Saccharomyces cerevisiae. We then induced Cre in the S. cerevisiae and extracted DNA from S. cerevisiae. We used NGS to assess the excision rate of the lox2272 variants. To develop a machine learning model to predict the excision rate of the lox2272 variants, we trained a Gaussian process (GP) model using data containing lox2272 sequences and their excision rates by Cre. Approximately 1000 lox2272 variants were randomly selected as training data and 100 variants were selected to test the prediction accuracy. We quantified the excision rates of over 2000 lox2272 variants by NGS and successfully found lox2272 variants with excision efficiencies ranging from 0.05% to 100%. The results of the NGS analysis were confirmed by qPCR. Next, we created a GP model and observed a high correlation between the actual and predicted cleavage rates of the test data. Using the GP model, we predicted the efficiency of the unevaluated lox2272 sequence and successfully identified lox2272 variants that had various cleavage rates. These results demonstrate the feasibility of precise control of opsin labeling rates in functional cellomics [2]. [1] Aoki et al, Sci Rep, 8, 10380, 2018 [2] Yamauchi et al, submitted.