Unlocking the potential of photosynthesis : in silico identification of enhanced rubisco enzymes

Abstract

The global population is expected to exceed 9 billion people by 2050 while at the same time atmospheric CO2 levels and global temperature is projected to rise. Unless strategies are adopted to reduce global warming, crop yields will suffer and not keep pace with global demand. Multiple efforts to improve global food security are therefore underway based on the principle that improvements in photosynthesis, the process by which plants convert CO2 into usable energy, leads to increases in yield. For some time, Ribulose-1,5-bisphosphate carboxylase/oxygenase (Rubisco) has been considered a major limiting step of plant photosynthesis. What is perplexing is that Rubisco is claimed to be the most abundant enzyme on Earth. However, it is also the slowest carboxylase in nature, the only carboxylase that reacts with both CO2 and O2 and one of the slowest evolving proteins on the planet. Studies on the kinetics of Rubisco indicate that there is some natural diversity among the Rubisco superfamily. This opens up the possibility of enhancing the kinetics of Rubisco in crops in order to boost yields. The aim of this thesis was, to develop in silico tools to improve our understanding of Rubisco by a) forecasting the outcomes of transplanting catalytically more efficient Rubisco isoforms in crops, and b) developing a machine learning (ML) pipeline for predicting uncharacterised Rubiscos. The first part of this thesis describes the development of a sunlit/shaded canopy photosynthesis model inspired by land surface models. The canopy models predicted Rubisco from the C4 grasses could substantially enhance the carbon uptake of wheat, sugar beet and maize cropland sites. However, it is important to know that these simulations were initially done under optimal conditions i.e. all enzymes of the Calvin-Benson-Bassham cycle (CBBC) and electron transport chain were perfectly balanced with the foreign Rubiscos otherwise benefits would dwindle. In the middle of this thesis, we performed multi-site wheat simulations, which revealed that similar improvements can be achieved across a range of environments. A yield model also revealed that similar improvements in photosynthesis may translate to similar improvements in yield but several areas were explored that may have overestimated the benefits. This will warrant further investigation in future studies. The final part of this thesis details the creation of a machine learning (ML) pipeline that utilises Rubisco sequence information and kinetic data for training. The basis of the ML pipelines were gaussian processes (GPs), a family of non-parametric, Bayesian models. Multiple representations of Rubisco sequences, known as protein encodings, were tested with the GPs. iv The best protein encoding type and model settings had given remarkable accuracy in conditions the models would be used in. In summary, this thesis opens up the possibility of using ML to accelerate directed evolution and screening of Rubiscos from a wide range of plant species.