No-Code Bayesian Optimisation Workflow
During our experiments, we noted that having an easy to use optimisation model would be useful for various aspects of our experiments, such as media composition, incubation times, and bioreactor and shake flask maintenance. Through a comprehensive review of scientific literature and past iGEM projects, we realised that a no-code batch Bayesian optimisation workflow that was generalisable to many different aspects of an experiment would be a novel contribution to iGEM. Thus, we developed BioKernel
To validate this workflow, we designed an experiment that would test the effectiveness of our optimisation algorithm by attempting to have it optimise inducer concentrations for a metabolic pathway controlled by orthogonal inducers.
We came across the core paper for our experiment through an initial literature review, which contributed a “Marionette” strain of E. coli with twelve highly orthogonal optimised inducers for the fine-tuning of inducible pathways in E. coli[1]. We planned an experiment around this strain that would produce astaxanthin, a carotenoid that can be quantified by 96-well plate assays without the need for cell lysis, which substantially simplifies our experimental workflow and allows for higher throughput[2].
Astaxanthin as a Final Target for Quantification
The carotenoid pathway is one of the most thoroughly characterised pathways that yields an end product whose concentration can be colorimetrically determined. Extensive research has been taken into improving the efficiency of carotenoid production, with particular interest in beta-carotene. Yang first reported a high yield of beta-carotene accomplished by co-expressing the mevalonate (MVA) and methylerythritol 4-phosphate (MEP) pathways in E. coli BL21[3]. More literature from 2016 points towards the high efficiency of isoprene and carotenoid production, with Yang reporting a yield of 24.0 g/L of isoprene[4].
However, we decided not to pursue this dual-pathway co-expression approach since it would take our total enzyme count over 12, the maximum that the Marionette strains of E. coli could handle.
We settled on using only the MEP pathway, assuming that although the pathway was endogenous to E. coli, Bayesian Optimization’s black-box nature and negative feedback from the relatively high expression of each intermediate would minimise the effect this would have on our pathway as a unit.
While looking into ways to quantify beta-carotene from 96-well plates, we realised that astaxanthin may be a better alternative for quantifying a carotenoid. Zhang reported a R-squared value of 0.9548 when quantifying the extracellular concentration of astaxanthin against the intracellular concentration, in contrast to much poorer values for other carotenoids[9]. This is because astaxanthin’s chemical structure allows is to diffuse through the plasma membrane more easily. This prompted us to add CrtW and CrtZ to our enzymatic pathway to produce astaxanthin instead of beta-carotene, enabling the high-throughput quantification of carotenoids more feasible on a tight time frame.
Further Optimisation of Experimental Plan
While looking into genetic circuit design and the Marionette strain of E. coli, we realised that adding insulators to our transcriptional units would give us more confidence that whatever change in expression results from a change in the inducer concentrations is due to the inducer concentration only.
Insulators are genetic elements used to maintain the heterologous and orthogonal nature of the expression patterns of all enzymes used in our experiment. This is because there is the potential for enhancement or silencing of the sequences randomly by sequences we did not consider. Insulators provide an extra level of safety to limit the effects of these faraway sequences.
We came across a 2016 article that describes a set of orthogonal insulators of which we chose to use eleven for our experiment[5].
While looking into ways to further optimise the experimental design, we realised that we could optimise the RBS sequence to standardise the translational initiation strength of our proteins. We chose to use the RBS Calculator from De Novo DNA[6] in order to generate the optimal RBS for our sequences.
References
- Meyer, A. J., Segall-Shapiro, T. H., Glassey, E. et al. Escherichia coli “Marionette” strains with 12 highly optimized small-molecule sensors. Nat. Chem. Biol. 15, 196–204 (2019). https://doi.org/10.1038/s41589-018-0168-3
- Zhang, C., Seow, V. Y., Chen, X. et al. Multidimensional heuristic process for high-yield production of astaxanthin and fragrance molecules in Escherichia coli. Nat. Commun. 9, 1858 (2018). https://doi.org/10.1038/s41467-018-04211-x
- Yang, J., Guo, L. Biosynthesis of β-carotene in engineered E. coli using the MEP and MVA pathways. Microb. Cell Fact. 13, 160 (2014). https://doi.org/10.1186/s12934-014-0160-x
- Yang, C., Gao, X., Jiang, Y., Sun, B., Gao, F., Yang, S. Synergy between methylerythritol phosphate pathway and mevalonate pathway for isoprene production in Escherichia coli. Metab. Eng. 37, 79–91 (2016). https://doi.org/10.1016/j.ymben.2016.05.003
- Nielsen, A. A. K. et al. Genetic circuit design automation. Science 352, aac7341 (2016). https://doi.org/10.1126/science.aac7341
- Reis, A. C. & Salis, H. M. An automated model test system for systematic development and improvement of gene expression models. ACS Synth. Biol. 9(11), 3145–3156 (2020).