Model-driven promoter strength prediction based on a fine-tuned synthetic promoter library inEscherichia coli

Abstract

<jats:title>Abstract</jats:title><jats:p>Promoters are one of the most critical regulatory elements controlling metabolic pathways. However, in recent years, researchers have simply perfected promoter strength, but ignored the relationship between the internal sequences and promoter strength. In this context, we constructed and characterized a mutant promoter library of P<jats:sub>trc</jats:sub> through dozens of mutation-construction-screening-characterization engineering cycles. After excluding invalid mutation sites, we established a synthetic promoter library, which consisted of 3665 different variants, displaying an intensity range of more than two orders of magnitude. The strongest variant was 1.52-fold stronger than a 1 mM isopropyl-β-D-thiogalactoside driven P<jats:sub>T7</jats:sub> promoter. Our synthetic promoter library exhibited superior applicability when expressing different reporters, in both plasmids and the genome. Different machine learning models were built and optimized to explore relationships between the promoter sequences and transcriptional strength. Finally, our XgBoost model exhibited optimal performance, and we utilized this approach to precisely predict the strength of artificially designed promoter sequences. Our work provides a powerful platform that enables the predictable tuning of promoters to achieve the optimal transcriptional strength.</jats:p>

Exp2Ipynb: A general machine-learning workflow for the analysis of promoter libraries

“…The sequence extraction reduces the sample size because more redundant sequences are generated, and sequence diversity increased because nucleotide diversity covers a shorter sequence. Figure 3 shows that the partitioned dataset 'k' (from Zhao et al (2020)) maintained the quality of prediction from the original dataset: the F1-CoV and top three feature contribution increased proportionally to the original data (B, C and D). This proportional movement was not displayed by the extracted sequences ('h', 'i') from Meng et al (2013).…”

“…without detailed processing. For quality assessment, we generated three additionally data sets based on partitioned sequence regions from the Meng et al (2013) and the Zhao et al (2020) data sets. The original sequence from Meng et al (2013) is 224 nucleotides (nt) long and we extracted the starting 40 nt ('h') and last 40 nt ('i'), assuming that few predictive positions are contained in the new sequences.…”

“…The original sequence from Meng et al (2013) is 224 nucleotides (nt) long and we extracted the starting 40 nt ('h') and last 40 nt ('i'), assuming that few predictive positions are contained in the new sequences. Zhao et al (2020) tested 113 nt ranging from upstream regulating elements down to the coding sequence. Again the first 40 nt ('k') were extracted, corresponding to the upstream regulating element.…”

“…We were interested to examine how the performance is affected when subsequences were extracted from libraries. We chose the libraries from Meng et al (2013) and Zhao et al (2020) because they report high feature number with low sequence diversity. The starting ('h') and ending ('i') 40 nt were extracted from the promoter sequence in Meng et al (2013), which had low features with importance.…”

“…For example, Meng et al (2013Meng et al ( , 2017 analyzed 98  70 promoter sequences and fine-tuned heterologous expression with predicted synthetic promoters. The same system was analyzed by Zhao et al (2020) with over 3500 promoter sequences. In Bacillus subtilis Liu et al (2018) employed a synthetic promoter library with 214 sequences to fine-tune pathway activity for metabolite overproduction.…”