Systematic investigation of synthetic operon designs enables prediction and control of expression levels of multiple proteins

Daniel Gerngross, Niko Beerenwinkel, Sven PankebioRxiv 2022 · DOI

TL;DR

This paper tackles the challenge of precisely controlling the expression levels of multiple proteins within synthetic biological systems by systematically investigating synthetic operon designs. They developed a machine learning model that can predict protein expression levels based on operon sequence features with high accuracy, offering a powerful tool for rational design of complex biological circuits.

What Problem Does This Paper Address?

Designing synthetic biological systems, whether for producing chemicals or creating therapeutic agents, often requires orchestrating the expression of multiple proteins. Achieving optimal performance hinges on fine-tuning the relative abundance of these proteins, a task complicated by the numerous DNA-encoded factors influencing gene expression from transcription to translation. While encoding multiple genes in operons (transcribed as a single mRNA) is an efficient strategy, the sequence-level requirements for precise expression control within these operons remain poorly understood. This lack of understanding leads to ad-hoc design choices, hindering predictable and robust system performance. The authors aim to bridge this gap by developing a predictive model for protein expression in synthetic operons based on DNA sequence features.

Methods

The researchers employed a multi-pronged approach:

  • Synthetic Operon Construction: They engineered 88 synthetic operons, ranging from 2 to 7 genes, using a Ligase Cycling Reaction (LCR) assembly method. This allowed for systematic variation of gene order and length.
  • Reporter System: Seven fluorescent proteins (FPs) were used as reporters to simultaneously measure protein expression levels (PE) via flow cytometry. To enable direct comparison across different constructs, they defined Differential Protein Expression (DPE) as the PE of a gene in an operon relative to its PE in a monocistronic construct.
  • Machine Learning Model: A random forest regression model was trained using sequence-derived features (e.g., RBS strength, gene position, RNA folding energy) to predict DPE.
  • Feature Importance Analysis: Shapley Additive Explanation (SHAP) values were used to identify and quantify the impact of different sequence features on protein expression.

Datasets

The study generated a novel dataset comprising 2772 protein expression measurements from 88 synthetic operons. This dataset, along with sequence-derived features and the trained model, is made publicly available to facilitate future research. This dataset represents a significant expansion over previous studies, particularly in its systematic exploration of gene order and operon length for operons exceeding three genes.

Key Insights and Ideas

  • Operon Design Matters Significantly: Gene order and operon length have a profound and predictable impact on protein expression levels, far beyond what might be assumed from ad-hoc design.
  • Position is Key: Genes located closer to the 3' end of an operon generally exhibit higher expression levels, a trend that holds across various operon designs.
  • Sequence Features are Predictive: A surprisingly high percentage of protein expression variation can be explained by sequence-level features, making a predictive model feasible.
  • Junctions are Critical: The structural and energetic properties of the RNA sequences at the junctions between genes within an operon are crucial determinants of protein expression.
  • Inducer Concentration is Dominant: As expected, the inducer concentration has the most significant impact on protein expression, but its effect is modulated by other sequence features.

Key Results

  • A random forest model was developed that explains 83% of the variation in protein expression data.
  • The model achieved a mean absolute error (MAE) of 9% relative to reference constructs, indicating robust predictive power.
  • Protein expression levels varied by up to 600-fold across different operon designs and genes.
  • Operon length and gene position were identified as major drivers of protein expression, with longer operons generally leading to lower expression and increased distance from the promoter leading to higher expression.
  • Features related to RNA secondary structure at gene junctions (folding energy and paired/unpaired nucleotide stretches) were found to be highly influential.

Limitations and Future Work

  • Limited Gene Repertoire: While the model performs well, its predictive power for entirely new genes not represented in the training set is reduced, though still acceptable. Future work should involve expanding the training set with a more diverse range of genes.
  • Expression System Specificity: The current model is trained on a T7 RNA polymerase expression system. The authors acknowledge that the importance of certain features might change in different expression systems, suggesting future work to generalize the model to other systems.
  • Feature Coverage Gaps: The SHAP analysis revealed gaps in the coverage of certain feature values in the training data, which could limit the generalization and interpretation of those features. Systematically closing these gaps is a direction for future improvement.
  • Proxy Features: Many identified features may act as proxies for multiple underlying molecular mechanisms, highlighting the complexity of biological systems and the need for sophisticated models.

The authors have also developed an online tool, Ocapy, to predict operon expression levels, making their findings directly accessible to the synthetic biology community.

Summary updated

Abstract

Abstract Controlling the expression levels of multiple recombinant proteins for optimal performance is crucial for synthetic biosystems but remains difficult given the large number of DNA-encoded factors that influence the process of gene expression from transcription to translation. In bacterial hosts, biosystems can be economically encoded as operons, but the sequence requirements for exact tuning of expression levels in an operon remain unclear. Here, we demonstrate the extent and predictability of protein-level variation using diverse arrangements of twelve genes to generate 88 synthetic operons with up to seven genes at varying inducer concentrations. The resulting 2772 protein expression measurements allowed the training of a sequence-based machine learning model that explains 83% of the variation in the data with a mean absolute error of 9% relative to reference constructs, making it a useful tool for protein expression prediction. Feature importance analysis indicates that operon length, gene position and gene junction structure are of major importance for protein expression.
BibTeX
@misc{gerngross2022systematic,
  author = {Daniel Gerngross and Niko Beerenwinkel and Sven Panke},
  title = {{Systematic investigation of synthetic operon designs enables prediction and control of expression levels of multiple proteins}},
  year = {2022},
  doi = {10.1101/2022.06.10.495604},
  howpublished = {bioRxiv},
}
Figure