Machine learning guided optimization of an artificial carbon dioxide fixation cycle and its extension towards value-added compounds

The global carbon cycle is a highly balanced exchange system of carbon between the geo-, hydro- and atmosphere. Since the industrial revolution, the combustion of fossil fuels is one of the main reasons for the shift of carbon levels towards higher concentrations in the hydro- and atmosphere. The amount of carbon dioxide (CO2) in the air is comparibly low but it is a highly potent greenhouse gas due to its structural properties. Its capability of absorbing infrared light, which would otherwise escape into space, leads to an increase in atmospheric temperatures. Because CO2 molecules are very stable, the conversion into multi-carbon-molecules is energy demanding and mainly done by organisms which use light as an energy source. Therefore, the main workhorses of capturing CO2 from the atmosphere are plants and algae, which incorporate the carbon for the production of biomass. In these complex organisms the Ribulose-1,5-bisphosphate carboxylase oxygenase (RuBisCO) is responsible for the carboxylation reaction. RuBisCO is considered a slow catalyst with a high error rate in accepting oxygen (O2) instead of CO2. In 2016, Schwander et al. published a new-to-nature pathway for the fixation of CO2. The so-called crotonyl-coenzyme A (CoA) /ethylmalonyl-CoA/hydroxybutyryl-CoA (CETCH) cycle was designed to circumvent RubisCO. In contrast to the Calvin cycle, the CETCH cycle is based on a highly efficient crotonyl-CoA carboxylase/reductase (Ccr), which does not display any side reaction with oxygen such as RuBisCO. Due to the overall pathway design, including Ccr as the key catalyst, the CETCH cycle has a higher net efficiency than natural aerobic CO2-fixation pathways and thus harbors the potential to play an important role in the reduction of atmospheric CO2 levels. To gain insights into this complex in vitro assay consisting of more than 25 components, we established a high-throughput workflow that enabled us to test hundreds of reaction conditions simultaneously. Prerequisite was the implementation of an acoustic liquid handling robot with a minimal pipetting volume of 25 nl. This enabled a fast reaction assembly and a drastic reduction in assay volume while maintaining a high pipetting accuracy. The acquired data was used to subsequently train an XGBoost-based machine learning algorithm aiming to optimize the CETCH cycle reaction parameters. After five rounds of optimization, the final model predicts reaction parameters for assay conditions with a ten-fold improvement on the manually optimized pathway version published in 2016. In addition, the algorithm identified the important components of the pathway and revealed one enzyme as a potential bottleneck of the current assay. Follow up experiments showed that the loss of intermediates by side reactions or hydrolysis is the limiting factor of the assay. Furthermore, we sought to extend the product portfolio of the CETCH cycle beyond its primary product glyoxylate. To this end, we first coupled the CETCH cycle to the β-hydroxyaspartate cycle. This enabled the production of oxaloacetate from two molecules glyoxylate. Adding only three additional enzymes from the serine cycle leads to the formation of acetyl-CoA, which we used to produce different terpenes via the mevalonate pathway. Despite the wide range of products, their synthesis has remained limited to the use of molecules produced downstream of the CETCH cycles primary product glyoxylate. Intermediates of the cycle were inaccessible, as their removal would lead to stalling of the pathway and a pre-mature arrest of CO2–fixation. To enable the utilization of CETCH cycle intermediates, we implemented anaplerotic routes that use the fixed CO2 to replenish drained cycle intermediates. We successfully reconstituted three anaplerotic pathways that enabled the production of the polyketide 6-deoxyerythronolide B (6-dEB). The biosynthesis of 6-dEB (C21) requires one molecule of propionyl-CoA and six molecules methylmalonyl-CoA, both intermediates of the CETCH cycle. The biosynthesis of complex molecules from CO2 in context of different highly convoluted pathways with up to 50 reactions highlights the robustness and versatility of the CETCH cycle