Multiple crystallographic dataset analysis to explore crystallographic occupancy and the conformational states of proteins

Protein crystallography is a powerful technique, to elucidate binding poses of ligands, which when sufficiently high-resolution data are available, gives atomic detail. Different conformations of the protein can also be revealed, potentially identifying which states are relevant to function and thus informing structure-based design. Recent advances in interpreting electron density from multiple datasets, such as Pan Dataset Density Analysis (PanDDA) have improved ligand detectability. This work focuses on producing higher quality models of protein-ligand complexes and alternate conformations of proteins, by utilising electron density information from multiple datasets. In chapter 2, a recently developed modelling strategy for protein structures was used to determine conformational differences between datasets. The structural model used is a superposition of the ground state and an altered state. The ground state is modelled from the average electron density of datasets which are sufficiently similar, and do not have significant regions of electron density different to other datasets. The altered state corresponds to local perturbations to the ground state, such as ligand binding. To improve the modelling of regions which undergo large conformational changes, the real space correlation coefficient was used to select the structural model which best explained the electron density. The superposition of a ground and altered state was for the first time used to model ligands binding near a metal ion site, for models of T. cruzi farnesyl pyrophosphate synthase. In order to achieve a well-behaved model external distance restraints which relate the superposed states, had to be augmented with distance restraints between the metal ions, the coordinating protein residues, and the coordinating water molecules. Furthermore, the occupancy of the metal ions needed to vary between unoccupied (0) and fully occupied (1) and not inherit the refined value of occupancy for the nearby ligand. Understanding what proportion of each dataset is well represented by the ground state, and thus the crystallographic occupancy of the altered state, is important in determining whether these superposed models appropriately explain the electron density without overfitting. To explore crystallographic occupancy of weakly binding ligands, repeated crystallographic datasets of protein ligand complexes were collected. Surprisingly a wide distribution of occupancies for the same ligand was observed (chapter 3). This was confirmed across different refinement programs, with both conventional models of protein-ligand complexes, and by modelling the ligand binding as a perturbation to the ground state. To determine whether this represented variation in the crystals, a brute force search of ligand occupancy and B factor values, which minimised difference density in the ligand binding region was developed. The exhaustive search method had a smoothly varying function near the minima of difference density. The method led to similar distributions of ligand occupancy to that of the refinement programs. Also note that occupancies below 0.4 were not refined to, or found in the brute force search. This suggests that the wide distribution of occupancies was most likely due to the poor parametrisation of the model rather than a underlying variability between crystals of the same protein ligand complex. Chapter 4 applied the same modelling strategies, and refinement programs to approximately two thousand protein ligand complexes found by XChem fragment screening. These refinements also showed that very few models had ligand occupancies below 0.4. As in chapter 3, refinement programs resulted in different occupancy and B factor values when comparing the same dataset. To calibrate the refinement of crystallographic occupancy an orthogonal method to measure the amount of a ligand present in a protein-ligand complex is required. In chapter 5 the development of such a method is discussed by measuring the proportion of a covalently labelled protein (NUDT7) by intact mass spectrometry of crystals. The proportion is determined by mixing two batches of protein, one labelled with a covalent electrophile, one not. The proportion can be accurately measured well below 0.4, and corresponds with the expected amount of labelled compound. Crystals grown with the same proportion of covalent compound were used to calibrate the crystallographic occupancy. It was found that for crystallographic models to sample the occupancy equivalent to say a 0.3 proportion of protein with the covalently bound ligand, the B factors of the ligand needed to be fixed, so that the model did not compensate for low occupancy by increasing the B factors. Lastly, in chapter 6 multiple datasets were sampled to identify new alternate side chain conformations, with the aim to determine whether the new conformations were correlated with one another. Such a relationship between two distant conformations could indicate allosteric signal propagation across the protein. In a case study of T. cruzi farnesyl pyrophosphate synthase, alternate conformations were detectable by sampling electron density around the dihedral angles of side chains, using a tool called Ringer. These appeared at both the allosteric ligand binding site and near the metal ions in the active site. Methods to automate the detection of such related events were attempted but did not isolate such examples. Overall this thesis presents evidence to show that the use and careful analysis of multiple datasets improves protein-ligand complex models, and can help detect large conformational changes in the protein. By orthogonal measurements of the proportion of a ligand present in a crystal, it was seen that the current parameterisation of occupancy is challenged when modelling weakly binding fragments.