Predictive design of sigma factor-specific promoters

Maarten Van Brempt, Jim Clauwaert, Friederike Mey, Michiel Stock, Jo Maertens, Willem Waegeman, Marjan De MeyNature Communications 2020 · DOI

TL;DR

This paper introduces ProD, an online tool for designing synthetic promoters with predictable transcription initiation frequencies (TIFs) for E. coli and B. subtilis sigma factors. By leveraging high-throughput sequencing and a novel convolutional neural network for ordinal regression, ProD enables precise control over gene expression, a critical need for complex synthetic biology circuits.

What Problem Does This Paper Address?

Engineering sophisticated synthetic gene circuits requires precise control over gene expression levels. Promoters, the DNA sequences that initiate transcription, are key regulatory elements. While methods exist to tune promoter strength, a standardized, automated tool for de novo promoter design with predictable TIFs has been lacking. This gap hinders the ability to build complex genetic systems where multiple components need to be finely tuned to work together without interference. Furthermore, ensuring that promoters are specific to certain sigma factors (orthogonality) is crucial for preventing unintended gene expression in mixed genetic systems, especially when using heterologous sigma factors.

Methods

The authors developed a workflow centered around creating and analyzing large promoter libraries. They engineered libraries of E. coli σ<sup>70</sup> and B. subtilis σ<sup>B</sup>, σ<sup>F</sup>, and σ<sup>W</sup>-dependent promoters by randomizing the "spacer" region between the conserved -35 and -10 recognition sites. This randomization allows for TIF variation while maintaining sigma factor specificity.

Key techniques include:

  • Fluorescence-Activated Cell Sorting (FACS): Cells containing these promoter libraries were sorted into 12 bins based on their fluorescent protein expression levels, effectively categorizing promoters by their TIF.
  • High-Throughput DNA Sequencing: DNA from each sorted bin was sequenced to identify the specific promoter sequences within each TIF category. This generated massive datasets linking DNA sequences to their expression levels.
  • Convolutional Neural Network (CNN) for Ordinal Regression: A custom CNN architecture was trained on this data. The "ordinal regression" aspect is crucial, as it accounts for the inherent order of TIF classes (e.g., class 0 is lower TIF than class 1). This approach is more suitable than standard classification for predicting a continuous or ordered variable. The network learns complex sequence-function relationships within the promoter spacer.
  • ProD Tool: The trained models form the basis of an online tool, ProD, which allows users to input desired TIF levels and obtain predicted promoter sequences.

Datasets

The primary datasets were generated in-house through the described experimental workflow.

  • Promoter Libraries: Four distinct promoter libraries were created: E. coli σ<sup>70</sup>, B. subtilis σ<sup>B</sup>, σ<sup>F</sup>, and σ<sup>W</sup>.
  • High-Throughput Sequencing Data: For each library, millions of DNA sequences were sequenced, linked to specific TIF bins. For the E. coli σ<sup>70</sup> library alone, over 1.3 million reads were obtained, covering over 320,000 unique sequences.
  • Orthogonality Data: Additional data was collected for libraries expressed in non-cognate sigma factor backgrounds to assess promoter orthogonality.

These datasets represent a significant expansion in the scale of sequence-function data for promoter design, differing from previous studies that often relied on smaller, less comprehensive datasets.

Key Insights and Ideas

  • The Spacer Region is Key for TIF Tuning: While the -35 and -10 regions are critical for sigma factor recognition, the spacer region between them offers significant flexibility for fine-tuning promoter TIF without compromising orthogonality.
  • Ordinal Regression is Essential for Promoter TIF Prediction: Treating TIF prediction as an ordinal regression problem, rather than a simple classification task, significantly improves the model's ability to capture the continuous nature of promoter strength and its ordered classes.
  • Deep Learning Uncovers Complex Sequence-Function Relationships: The use of CNNs demonstrates that promoter TIF is not solely determined by simple sequence motifs but by more complex, higher-order interactions within the spacer DNA.
  • In Vivo Validation is Crucial: The study highlights a notable difference between test set performance and in vivo validation. While models perform well on test data, in vivo validation of the E. coli σ<sup>70</sup> model showed a much stronger correlation, suggesting that accounting for real-world biological noise (both intrinsic and extrinsic) is vital for accurate prediction.
  • Orthogonality Prediction is Challenging but Promising: Predicting promoter orthogonality proved more difficult, potentially due to noise in the low-expression bins. However, the models showed promising trends, suggesting that further refinement could lead to tools for designing orthogonal promoters.

Key Results

  • ProD Tool Developed: An online tool, ProD, was created, enabling the design of E. coli σ<sup>70</sup> promoters with predictable TIFs.
  • High In Vivo Validation Correlation: The E. coli σ<sup>70</sup> model achieved a Spearman correlation of 0.91 in in vivo validation, indicating a strong ordinal relationship between predicted and measured promoter TIFs.
  • Model Performance Varies by Sigma Factor: The models for E. coli σ<sup>70</sup> and B. subtilis σ<sup>B</sup> showed better performance (e.g., higher ROC AUC, lower MAE) compared to those for σ<sup>F</sup> and σ<sup>W</sup>. This is attributed to potential differences in how spacer engineering affects TIF for different sigma factors.
  • Large-Scale Data Generation: The study generated unprecedentedly large datasets of promoter sequences and their corresponding TIFs, crucial for training robust deep learning models.

Limitations and Future Work

  • Orthogonality Prediction Challenges: The models for predicting promoter orthogonality showed lower performance, likely due to noise in the low-expression data bins. Further work is needed to improve data quality and model robustness for orthogonality prediction.
  • Sigma Factor Specificity: The performance of the models varied across different sigma factors, with σ<sup>F</sup> and σ<sup>W</sup> showing lower predictive power. This suggests that the impact of spacer engineering on TIF might be more complex or less pronounced for these sigma factors, requiring further investigation.
  • Extrinsic Noise Impact: While the in vivo validation accounted for some noise, the authors acknowledge that factors beyond the promoter sequence (e.g., genetic context, cellular environment) can still influence promoter performance.
  • Future Directions: The authors suggest extending ProD to include predictions for other sigma factors and to incorporate upstream regulatory elements (like UP elements) for even greater design flexibility. They also propose exploring the structural properties of DNA within the spacer region, as simple sequence motifs do not fully explain TIF. The developed methodology could also be applied to other sequence-dependent biological problems.

Summary updated

Abstract

To engineer synthetic gene circuits, molecular building blocks are developed which can modulate gene expression without interference, mutually or with the host's cell machinery. As the complexity of gene circuits increases, automated design tools and tailored building blocks to ensure perfect tuning of all components in the network are required. Despite the efforts to develop prediction tools that allow forward engineering of promoter transcription initiation frequency (TIF), such a tool is still lacking. Here, we use promoter libraries of E. coli sigma factor 70 (σ<sup>70</sup>)- and B. subtilis σ<sup>B</sup>-, σ<sup>F</sup>- and σ<sup>W</sup>-dependent promoters to construct prediction models, capable of both predicting promoter TIF and orthogonality of the σ-specific promoters. This is achieved by training a convolutional neural network with high-throughput DNA sequencing data from fluorescence-activated cell sorted promoter libraries. This model functions as the base of the online promoter design tool (ProD), providing tailored promoters for tailored genetic systems.
BibTeX
@inproceedings{brempt2020predictive,
  author = {Maarten Van Brempt and Jim Clauwaert and Friederike Mey and Michiel Stock and Jo Maertens and Willem Waegeman and Marjan De Mey},
  title = {{Predictive design of sigma factor-specific promoters}},
  booktitle = {Nature Communications},
  year = {2020},
  doi = {10.1038/s41467-020-19446-w},
}
Figure