December 20th, 2013 | Category: Mathematics, Metabolic engineering, Papers | Leave a comment

Multi-enzyme pathway engineering

Engineering metabolic pathways, whether for pharmaceuticals or fuels, involves many trade-offs. Flux imbalances (for example where flux through the first step exceeds flux through the second) can lead to the accumulation of toxic intermediates. Moreover, expressing heterologous proteins beyond that which is necessary can place a burden upon the cell for protein, rather than product, synthesis. Sometimes lowering expression levels of heterologously expressed proteins improves product titers (e.g. here and here).

This recent paper (online Sep 2013) from John Dueber’s and Claire Tomlin’s labs tackles the thorny issue of how to optimise expression levels of proteins along a multi-step branching pathway where no high-throughput assays are available (at least not for all end products). This latter limitation means that whilst a full combinatorial approach might be desirable to fully explore the design space, it is unrealistic to assess all of the constructs and strains. Time and money does not permit this. Finally, information about the pathway biochemistry (and indeed enzyme regulation) is limited, making explicit modelling predictions unfeasible.

The authors chose to use the pathway for violacein biosynthesis. This pathway comprises five enzymatic steps, one of which, VioC, also produces an unwanted side-product (VioC catalyses both the conversion of protoviolaceinic acid to violacein (which is good) but also the conversion of an early pathway intermediate, protodeoxyviolaceinic acid to deoxyviolacein (which is not so good) – see Fig. 1 (Fig. 3 in their paper)). In addition, two further compounds are produced via chemical conversion within the cell, meaning that they have one target molecule and three off-target molecules. The former can be assayed in a high-throughput manner, but the unwanted side reactions cannot. In order to test expression at multiple levels (and the authors wished to use five different expression levels) the total number of combinations that exist is 55 (allowing for the five enzymes to be expressed at five different levels). In other words, 3125 combinations. That’s a lot of constructs to test, especially when you are reliant on a method such as HPLC to assay your products.

Fig. 1. Production of violacein from tryptophan, and the unwanted output molecules.

The authors first of all characterised the expression behaviour of a number of different yeast (S. cerevisiae) promoters: they identified five promoters whose constitutive activity spanned a range of expression levels (using the reporter YFP). These reporters were also confirmed to be independent of the sequence context (specifically the downstream coding sequence). In other words, these particular promoters behave in a highly predictable manner.

Secondly they exploited a standardized assembly procedure – based on isothermal assembly (Gibson Assembly) – to generate a combinatorial library. However, they did not test all the combinations. Instead they randomly sampled just 3% of the 3125 possible combinations – 91 constructs (Fig. 2, Supplementary Fig. 9 in the paper) and tested those for output of their target and off-target molecules. At this stage they had the output data, but did not know which combinations of constructs resulted in a particular outputs. To determine this, they sequenced the constructs (handily barcoded) using TRAC – a rapid genotyping assay in order to identify which genes were present behind which promoter. They could now determine which genotype gives rise to a particular phenotype.

Fig. 2. A sub-sample of all the different pathway combinations for violacein biosynthesis in yeast.

This data was in turn used to train a linear regression for each of the four output molecules which allowed the researchers to predict the best combination for production of each product (Fig. 3. (4 in  the paper)).  The five top predictions were summarily analysed and the predictions held true. In so doing they were able to direct flux towards whichever branch of the metabolic pathway they chose.

Fig. 3. Model predictions for each of the four outputs. These models were successfully used to predict the behaviour of untested combinations.

The beauty of this approach is its simplicity – using the high throughput techniques where they are available (at the molecular and statistical level) to lift the load on onerous analytical techniques and, importantly, circumventing the need for in-depth knowledge of the cellular biology of the system. Importantly, they also found that the library size used to train the model needed only to be between 1 and 2% of all the combinations (rather than the initial 3%). In other words, increasing the sample size had limited effect on model improvement. Such approaches will not only become much more commonplace for engineering biology, but also for understanding (and quantifying) how particular metabolic pathway components interact to regulate flux.


Leave a Reply




You can use these HTML tags

<a href="" title=""> <abbr title=""> <acronym title=""> <b> <blockquote cite=""> <cite> <code> <del datetime=""> <em> <i> <q cite=""> <s> <strike> <strong>