This thesis presents two papers addressing important biochemical prediction challenges. The first paper focuses on accurate protein distance predictions and introduces updates to the ProSPr network. We evaluate its performance in the Critical Assessment of techniques for Protein Structure Prediction (CASP14) competition, investigating its accuracy dependence on sequence length and multiple sequence alignment depth. The ProSPr network, an ensemble of three convolutional neural networks (CNNs), demonstrates superior performance compared to individual networks. The second paper addresses the issue of accurate ligand ranking in virtual screening for drug discovery. We propose MILCDock, a machine learning consensus docking tool that leverages predictions from five traditional molecular docking tools. MILCDock, an ensemble of eight neural networks, outperforms single-network approaches and other consensus docking methods on the DUD-E dataset. However, we find that LIT-PCBA targets remain challenging for all methods tested. Furthermore, we explore the effectiveness of training machine learning tools on the biased DUD-E dataset, emphasizing the importance of mitigating its biases during training. Collectively, this work emphasizes the power of ensembling in deep learning-based biochemical prediction problems, highlighting improved performance through the combination of multiple models. Our findings contribute to the development of robust protein distance prediction tools and more accurate virtual screening methods for drug discovery.

No description available.

Doctor of Philosophy / In my research, I developed protein model quality estimation methods aimed at evaluating the reliability of computationally predicted protein models in the absence of experimentally solved ground truth structures. These methods specifically focus on estimating errors within the protein models to quantify their structural accuracy. Recognizing that even the most advanced protein structure prediction techniques may produce models with errors, I also developed a complementary protein model refinement method. This refinement method iteratively optimizes the weakly modeled regions, guided by the error estimation module of my quality estimation approach. The development of these model quality estimation methods, therefore, not only offers valuable insights into the structural reliability of protein models but also contributes to optimizing the overall reliability of protein models generated by state-of-the-art computational methods.

No / In the 2007 blind test of crystal structure prediction hosted by the Cambridge Crystallographic Data Centre (CCDC), a hybrid DFT/MM method correctly ranked each of the four experimental structures as having the lowest lattice energy of all the crystal structures predicted for each molecule. The work presented here further validates this hybrid method by optimizing the crystal structures (experimental and submitted) of the first three CCDC blind tests held in 1999, 2001, and 2004. Except for the crystal structures of compound IX, all structures were reminimized and ranked according to their lattice energies. The hybrid method computes the lattice energy of a crystal structure as the sum of the DFT total energy and a van der Waals (dispersion) energy correction. Considering all four blind tests, the crystal structure with the lowest lattice energy corresponds to the experimentally observed structure for 12 out of 14 molecules. Moreover, good geometrical agreement is observed between the structures determined by the hybrid method and those measured experimentally. In comparison with the correct submissions made by the blind test participants, all hybrid optimized crystal structures (apart from compound II) have the smallest calculated root mean squared deviations from the experimentally observed structures. It is predicted that a new polymorph of compound V exists under pressure.

No / We report on the organization and outcome of the fourth blind test of crystal structure prediction, an international collaborative project organized to evaluate the present state in computational methods of predicting the crystal structures of small organic molecules. There were 14 research groups which took part, using a variety of methods to generate and rank the most likely crystal structures for four target systems: three single-component crystal structures and a 1:1 cocrystal. Participants were challenged to predict the crystal structures of the four systems, given only their molecular diagrams, while the recently determined but as-yet unpublished crystal structures were withheld by an independent referee. Three predictions were allowed for each system. The results demonstrate a dramatic improvement in rates of success over previous blind tests; in total, there were 13 successful predictions and, for each of the four targets, at least two groups correctly predicted the observed crystal structure. The successes include one participating group who correctly predicted all four crystal structures as their first ranked choice, albeit at a considerable computational expense. The results reflect important improvements in modelling methods and suggest that, at least for the small and fairly rigid types of molecules included in this blind test, such calculations can be constructively applied to help understand crystallization and polymorphism of organic molecules.

Yes / The sixth blind test of organic crystal-structure prediction (CSP) methods has been held, with five target systems: a small nearly rigid molecule, a polymorphic former drug candidate, a chloride salt hydrate, a co-crystal, and a bulky exible molecule. This blind test has seen substantial growth in the number of submissions, with the broad range of prediction methods giving a unique insight into the state of the art in the field. Significant progress has been seen in treating flexible molecules, usage of hierarchical approaches to ranking structures, the application of density-functional approximations, and the establishment of new workflows and best practices for performing CSP calculations. All of the targets, apart from a single potentially disordered Z0 = 2 polymorph of the drug candidate, were predicted by at least one submission. Despite many remaining challenges, it is clear that CSP methods are becoming more applicable to a wider range of real systems, including salts, hydrates and larger flexible molecules. The results also highlight the potential for CSP calculations to complement and augment experimental studies of organic solid forms. / EPSRC (grants EP/J01110X/1 and EP/K018132/1) and the European Research Council under the European Union’s Seventh Framework Programme (FP/2007-2013)/ERC through grant agreements n. 307358 (ERC-stG- 2012-ANGLE) and n. 321156 (ERC-AG-PE5-ROBOT). The Russian Foundation (14-03-01091). GlaxoSmithKline, Merck, and Vertex. VIDI Research Program 700.10.427, which is financed by The Netherlands Organisation for Scientific Research (NWO), and the European Research Council (ERC-2010-StG, grant agreement n. 259510-KISMOL). Foundation for Fundamental Research on Matter (FOM). NSF grant number ACI-1053575. University of Buenos Aires and the Argentinian Research Council. Conflex Corp. and Ministry of Education, Culture, Sports, Science and Technology. Louisiana Board of Regents Award # LEQSF(2014-17)-RD-A-10 “Toward Crystal Engineering from First Principles”, the NSF award # EPS-1003897 “The Louisiana Alliance for Simulation-Guided Materials Applications (LA-SiGMA)”, and by the Tulane Committee on Research Summer Fellowship. Solar Technologies Go Hybrid initiative of the State of Bavaria, Germany. U.S. Department of Energy under contract DE-AC02-06CH11357. EPSRC (EP/J003840/1, EP/J014958/1) and [EP/J017639/1]. Leadership Fellowship Grant [EP/K013688/1]. Robinson College, Cambridge, and the Cambridge Philosophical Society for a Henslow Research Fellowship. Army Research Office under Grant W911NF-13-1- 0387 and by the National Science Foundation Grant CHE-1152899. Polish National Science Centre Grant No. DEC-2012/05/B/ST4/00086. EPSRC, though grant ESPRC EP/K039229/1, and Eli Lilly. The PhD students support: RKH by a joint UCL Max-Planck Society Magdeburg Impact studentship, REW by a UCL Impact studentship; LI by the Cambridge Crystallographic Data Centre and the M3S Centre for Doctoral Training (EPSRC EP/G036675/1). Army Research Office Grant W911NF-13-1-0387 and by the National Science Foundation Grant CHE-1152899. U.S. Army Research Laboratory and the U.S. Army Research Office under contract/grant number W911NF-13-1-0387 (MET and LV) and MRSEC program of the National Science Foundation under Award Number DMR-1420073 (MET and ES). U.S. Army Research Laboratory and the U.S. Army Research Office contract/grant number W911NF-13-1- 0387 and by the National Science Foundation Grant CHE-1152899. National Science Foundation (DMR-1231586), the Government of Russian Federation (Grant No. 14.A12.31.0003), the Foreign Talents Introduction and Academic Exchange Program (No. B08040) and the Russian Science Foundation, project no. 14-43-00052, base organization Photochemistry Center of the Russian Academy of Sciences. Army Research Office Grant W911NF-13-1-0387 and the National Science Foundation Grant CHE-1152899. Deutsche Forschungsgemeinschaft program DFG-SPP 1807. Department of Energy (DOE) Grant Nos. DE-SC0008626. Office of Science of the U.S. Department of Energy Contract No. DE-AC02-06CH11357. Office of Science of the U.S. Department of Energy contract No. DEAC02-05CH11231.

This thesis investigates the correlations between short protein peptide sequences and local tertiary structures. In particular, it introduces a novel algorithm for partitioning short protein segments into clusters of local sequence-structure motifs, and demonstrates that these motif clusters contain useful structural information via two applications to structural prediction. The first application utilizes motif clusters to predict local protein tertiary structures. A novel dynamic programming algorithm that performs comparably with some of the best existing algorithms is described. The second application exploits the capability of motif clusters in recognizing regular secondary structures to improve the performance of secondary structure prediction based on Support Vector Machines. Empirical results show significant improvement in overall prediction accuracy with no performance degradation in any specific aspect being measured. The encouraging results obtained illustrate the great potential of using local sequence-structure motifs to tackle protein structure predictions and possibly other important problems in computational biology.

This thesis investigates the correlations between short protein peptide sequences and local tertiary structures. In particular, it introduces a novel algorithm for partitioning short protein segments into clusters of local sequence-structure motifs, and demonstrates that these motif clusters contain useful structural information via two applications to structural prediction. The first application utilizes motif clusters to predict local protein tertiary structures. A novel dynamic programming algorithm that performs comparably with some of the best existing algorithms is described. The second application exploits the capability of motif clusters in recognizing regular secondary structures to improve the performance of secondary structure prediction based on Support Vector Machines. Empirical results show significant improvement in overall prediction accuracy with no performance degradation in any specific aspect being measured. The encouraging results obtained illustrate the great potential of using local sequence-structure motifs to tackle protein structure predictions and possibly other important problems in computational biology.

This dissertation concerns the study and prediction of loops in protein structures. Proteins perform crucial functions in living organisms. Despite their importance, we are currently unable to predict their three dimensional structure accurately. Loops are segments that connect regular secondary structures of proteins. They tend to be located on the surface of proteins and often interact with other biological agents. As loops are generally subject to more frequent mutations than the rest of the protein, their sequences and structural conformations can vary significantly even within the same protein family. Although homology modelling is the most accurate computational method for protein structure prediction, difficulties still arise in predicting protein loops. Protein loop structure prediction is therefore a bottleneck in solving the protein structure prediction problem. Reflecting on the success of homology modelling, I implement an improved version of a database search method, FREAD. I show how sequence similarity as quantified by environment specific substitution scores can be used to significantly improve loop prediction. FREAD performs appreciably better for an identifiable subset of loops (two thirds of shorter loops and half of the longer loops tested) than ab initio methods; FREAD's predictive ability is length independent. In general, it produces results within 2Å root mean square deviation (RMSD) from the native conformations, compared to an average of over 10Å for loop length 20 for any of the other tested ab initio methods. I then examine FREAD’s predictive ability on a specific type of loops called complementarity determining regions (CDRs) in antibodies. CDRs consist of six hypervariable loops and form the majority of the antigen binding site. I examine CDR loop structure prediction as a general case of loop structure prediction problem. FREAD achieves accuracy similar to specific CDR predictors. However, it fails to accurately predict CDR-H3, which is known to be the most challenging CDR. Various FREAD versions including FREAD with contact information (ConFREAD) are examined. The FREAD variants improve predictions for CDR-H3 on homology models and docked structures. Lastly, I focus on the local properties of protein loops and demonstrate that the protein loop structure prediction problem is a local protein folding problem. The end-to-end distance of loops (loop span) follows a distinctive frequency distribution, regardless of secondary structure elements connected or the number of residues in the loop. I show that the loop span distribution follows a Maxwell-Boltzmann distribution. Based on my research, I propose future directions in protein loop structure prediction including estimating experimentally undetermined local structures using FREAD, multiple loop structure prediction using contact information and a novel ab initio method which makes use of loop stretch.

The purpose of the research contained in this thesis is to allow for the prediction of new structures and properties of crystalline structures due to the application of external pressure by using first-principles numerical computations. The body of the thesis is separated into two primary research projects. The properties of cupric oxide (CuO) have been studied at pressures below 70 GPa, and it has been suggested that it may show room-temperature multiferroics at pressure of 20 to 40 GPa. However, at pressures above these ranges, the properties of CuO have yet to be examined thoroughly. The changes in crystal structure of CuO were examined in these high-pressure ranges. It was predicted that the ambient pressure monoclinic structure changes to a rocksalt structure and CsCl structure at high pressure. Changes in the magnetic ordering were also suggested to occur due to superexchange interactions and Jahn-Teller instabilities arising from the d-orbital electrons. Barium chloride (BaCl) has also been observed, which undergoes a similar structural change due to an s – d transition, and whose structural changes can offer further insight into the transitions observed in CuO. Ammonia borane (NH3BH3) is known to have a crystal structure which contains the molecules in staggered conformation at low pressure. The crystalline structure of NH3BH3 was examined at high pressure, which revealed that the staggered configuration transforms to an eclipsed conformation stabilized by homopolar B–Hδ-∙∙∙ δ-H–B dihydrogen bonds. These bonds are shown to be covalent in nature, comparable in bond strength to conventional hydrogen bonds, and may allow for easier molecular hydrogen formation in hydrogen fuel storage.