Team:UChicago/Engineering
Description * Design * Engineering PoC
Engineering Success: a proof of concept for the algorithm
When we started designing Optizyme, we began by doing research on Michaelis-Menten kinetic modelling, as well as classes of optimization algorithms that we could use for our software. In fact, our optimization algorithm belongs to the same class of algorithms that artificial intelligence neural networks use to learn, and this inspiration came from our search through the literature. As we brainstormed how we were going to design our optimization algorithm, we had to reimagine Michaelis-Menten simulations not as an entire time-course for enzymes, but as a function that could take on the form f(x1, x2, … xn) so that we can use calculus within our algorithm. Combining everything we learned from the literature and our reformulation of Michaelis-Menten simulations, we were able to design, build, and eventually test our software on not only our novel plastic degradation pathway, but also on recently published papers including “A combined experimental and modelling approach for the Weimberg pathway optimisation” (Shen et al. 2020). This was the paper we used for proof of concept of our algorithm. The current version of Optizyme that exists is not the product of the UChicago iGEM team working in isolation. Rather, it is the result of the dynamic interchange that takes place between our team and our envisioned end users. After completing our initial version for Optizyme, we reached out to experts in industry to learn about what kind of improvements needed to be made to our software. In the process of implementing change based on their feedback, we continue through the entire engineering process again. The details of how we implemented feedback from experts in industry is described in more detail through our Human Practices.
We demonstrate a proof of concept for our optimization algorithm through working with the paper “A Combined Experimental and Modelling Approach for the Weimberg Pathway Optimization, by Shen et al.
Within this paper, the authors worked to optimize the Weimberg pathway, which is a five enzyme pathway that converts D-xylose to α-ketoglutarate, and is depicted below.
The authors performed initial-rate experiments to determine the kinetic constants and reaction mechanisms of each of the enzymes involved in the pathways, and then used these findings to construct a model based on the system of differential equations that represent the time evolution of their cell-free system. The researchers constructed their cell-free system in the lab, and compared the time evolution of the real system with the time evolution predicted by the model to confirm that their model accurately predicted the effect of the system. The lines in the graph below are the model predictions, and the dots are data points from the actual system.
The authors then used their model to computationally optimize the ratio of enzyme concentrations while holding the total enzyme concentration fixed at 22.6 micrograms/milliliter. The authors’ reported optimal ratio of enzymes results in a time to reach 99% yield of 63.07 minutes where the optimal enzyme concentrations were 2: 1.5: 5.7: 9.8: 3.6, where the concentrations are in micrograms/milliliter. We reconstructed the model presented in the paper, and used the optimization algorithm included in Optizyme to determine the optimal ratio of enzyme concentrations, which we determined to be 2.112440: 1.603529: 5.900251: 9.309787: 3.673993. Our optimal ratio of enzyme concentrations gives a time to reach 99% yield of 62.85 minutes, a slight improvement from the answer found in the paper. The enzyme ratio found by our algorithm giving an improved conversion time to 99% yield demonstrates the validity of the optimization algorithm provided in Optizyme.
We also attempted to demonstrate the validity of our modeling software on complex cell-free systems by using our generalized modelling function to recreate the model used in the paper. When the model was built using our generalized function, the resulting model predicted a final yield of 4.08296 millimolar of a-ketoglutarate after 63.07 minutes, while the model that we recreated from the paper predicted a final yield of 4.207643, indicating that our function’s prediction is off by a little under 3%. However, it is important to note that the cell-free system modelled in the paper includes a couple unusual reactions. Notably, the conversion of one of the compounds in the system occurs both spontaneously and through enzyme catalysis. Because our function was meant to write differential equations only for enzyme catalyzed reactions, we were unable to account for the rate equations of the spontaneous conversion of this compound. Furthermore, the cell-free system includes a mechanism that recycles NAD+ from NADH and pyruvate, and the machinery for this cofactor recycling was described using an experimentally determined rate equation, which our package can’t recreate but we approximated using an additional enzyme rate equation. Lastly, one of the enzymes in the system is competitively inhibited by NADH, but with differing inhibition constants based on if the NADH was bound to one of two different compounds. This unique kind of complex based inhibition is not included in our function and was unable to be modeled. With the three interactions that our model building function could not account for, it was still able to predict the system’s yield to 97% accuracy.
When there are no special interactions in a cell-free system, which is the case for our novel PET degradation pathway, the model function operates at 100% accuracy when compared to a model specifically made for the system. For example, if our PET system is set up with .25 micromolar of each enzyme, and 1 millimolar of PET, both the model generated from our function and the model we specifically built for the PET system predict that after 2000 seconds there will be .89 millimolar of catechol, showing that the model function does generate the correct differential equations when only enzyme rate equations are required in the model.
Works Referenced
1. Shen, L.; Kohlhaas, M.; Enoki, J.; Meier, R.; Schönenberger, B.; Wohlgemuth, R.; Kourist, R.; Niemeyer, F.; Niekerk, D. van; Bräsen, C.; Niemeyer, J.; Snoep, J.; Siebers, B. A combined experimental and modelling approach for the Weimberg pathway optimisation. https://www.nature.com/articles/s41467-020-14830-y (accessed Oct 27, 2020).