Abstract/Summary

Pearson Correlation Coefficient and provides a value between -1 to 1, with -1 being a total negative correlation, 0 is no correlation and 1 is a total positive correlation based on the observed and predicted ligand-binding site residues. Scores of 0.40 to 0.69 are strong positive relationships and 0.70 and higher are strong positive relationships. The downside of MCC is that it does not take into consideration the overall 3D structure of the protein model. Therefore, BDT will also be utilised as this score, which is also scored from -1 to 1, to take into consideration the 3D structure. Both MCC and BDT are only possible to produce when there is an observed (actual) structure available with bound ligands to compare against the predicted structure and hence why MCC and BDT are objective measures of ligand-binding site prediction. The average MCC and BDT score from CASP11 was 0.42 and 0.51, respectively. CASP12 saw the prediction of ligands for low annotation level proteins with no known ligands, demonstrating the potential use of FunFOLD3 in novel protein prediction. The average MCC and BDT score from CASP13 was 0.47 and 0.53. CAFA3 showed FunFOLDQ can be used in the prediction of GO terms, however further refinements are needed to increase specificity of the term predictions. The development option this thesis has explored is the use of docking (preferred orientation of interacting partners) with AutoDock Vina to improve the accuracy of ligand-binding residues by FunFOLD3, as the problem with TBM methods can be that predicted ligand(s) from a similar template will be forced to fit within the ligand-binding pocket. However, with docking, the aim of this method is to predict the preferred orientation of the ligand within the ligand-binding space. Utilisation of docking has also added to the novelty of this research, as different grid box calculations around the ligand-binding space was explored, with varying degrees of success with each grid box calculation. Examples of two CASP targets which had improvements in MCC and BDT score following docking were CASP11 target T0783 (2-C-methyl-D-erythritol 4- phosphate cytidylyltransferase) the MCC and BDT scores by FunFOLD3 were 0.17 and 0.21, respectively. Following docking the MCC and BDT scores increased to 0.63 and 0.45, respectively. CASP13 target T1016 (alpha-ribazole-5'-P phosphatase) had MCC and BDT scores of 0.556 and 0.646 by FunFOLD3, respectively. Following docking the MCC and BDT increased to 0.85 and 0.91, respectively. Lastly, CASP_Commons, a community-wide experiment to find the consensus structures, explored the role of FunFOLD3 with predicting ligands and ligand-binding sites for the novel protein and proteins domains of SARS-CoV-2. The protein domains were non-structural proteins 2, 4 and 6, open reading frames 3a, 6, 7b, 8 and 10, membrane protein and papain-like protease. FunFOLD3 predicted ligands for ten of the protein domains, of which there were a total of 32 targets due to domains being split into smaller residues and subsequent rounds of 3D modelling improvement. Increased understanding of protein structures can provide further insight into a protein’s function, particularly if ligands are bound and identified, an example in this thesis is the prediction of chlorophyll A for non-structural protein 4 (nsp4). Chlorophyll A, like haemoglobin is a porphyrin ring and templates related to nsp4 show a role in blood clotting. Therefore, whilst chlorophyll A might not be the exact ligand, similarities between haemoglobin and chlorophyll A can clearly be determined and assist in understanding the role of nsp4 in the pathology of COVID-19. Identification of GO terms can provide more detailed understanding into the function or functions of proteins and, in proteins with limited annotation information this can assist with comprehending their role. This thesis has focused on improving and developing a function prediction method, FunFOLD3, to better understand the role and function of proteins. The new method of FunFOLD3 which utilises docking will be integrated into the McGuffin group prediction servers and will be benchmarked in subsequent CASP competitions, to critically assess the performance of the developed method.