Nuclear magnetic resonance (NMR) spectroscopy technique is becoming exceedingly significant due to its capability of studying protein structures in solution. However, NMR protein structure determination has remained a laborious and costly process until now, even with the help of currently available computer programs. After the NMR spectra are collected, the main road blocks to the fully automated NMR protein structure determination are peak picking from noisy spectra, resonance assignment from imperfect peak lists, and structure calculation from incomplete assignment and ambiguous nuclear Overhauser enhancements (NOE) constraints. The goal of this dissertation is to propose error-tolerant and highly-efficient methods that work well on real and noisy data sets of NMR protein structure determination and the closely related protein structure prediction problems. One major contribution of this dissertation is to propose a fully automated NMR protein structure determination system, AMR, with emphasis on the parts that I contributed. AMR only requires an input set with six NMR spectra. We develop a novel peak picking method, PICKY, to solve the crucial but tricky peak picking problem. PICKY consists of a noise level estimation step, a component forming step, a singular value decomposition-based initial peak picking step, and a peak refinement step. The first systematic study on peak picking problem is conducted to test the performance of PICKY. An integer linear programming (ILP)-based resonance assignment method, IPASS, is then developed to handle the imperfect peak lists generated by PICKY. IPASS contains an error-tolerant spin system forming method and an ILP-based assignment method. The assignment generated by IPASS is fed into the structure calculation step, FALCON-NMR. FALCON-NMR has a threading module, an ab initio module, an all-atom refinement module, and an NOE constraints-based decoy selection module. The entire system, AMR, is successfully tested on four out of five real proteins with practical NMR spectra, and generates 1.25A, 1.49A, 0.67A, and 0.88A to the native reference structures, respectively. Another contribution of this dissertation is to propose novel ideas and methods to solve three protein structure prediction problems which are closely related to NMR protein structure determination. We develop a novel consensus contact prediction method, which is able to eliminate server correlations, to solve the protein inter-residue contact prediction problem. We also propose an ultra-fast side chain packing method, which only uses local backbone information, to solve the protein side chain packing problem. Finally, two complementary local quality assessment methods are proposed to solve the local quality prediction problem for comparative modeling-based protein structure prediction methods.

Nuclear magnetic resonance (NMR) spectroscopy technique is becoming exceedingly significant due to its capability of studying protein structures in solution. However, NMR protein structure determination has remained a laborious and costly process until now, even with the help of currently available computer programs. After the NMR spectra are collected, the main road blocks to the fully automated NMR protein structure determination are peak picking from noisy spectra, resonance assignment from imperfect peak lists, and structure calculation from incomplete assignment and ambiguous nuclear Overhauser enhancements (NOE) constraints. The goal of this dissertation is to propose error-tolerant and highly-efficient methods that work well on real and noisy data sets of NMR protein structure determination and the closely related protein structure prediction problems. One major contribution of this dissertation is to propose a fully automated NMR protein structure determination system, AMR, with emphasis on the parts that I contributed. AMR only requires an input set with six NMR spectra. We develop a novel peak picking method, PICKY, to solve the crucial but tricky peak picking problem. PICKY consists of a noise level estimation step, a component forming step, a singular value decomposition-based initial peak picking step, and a peak refinement step. The first systematic study on peak picking problem is conducted to test the performance of PICKY. An integer linear programming (ILP)-based resonance assignment method, IPASS, is then developed to handle the imperfect peak lists generated by PICKY. IPASS contains an error-tolerant spin system forming method and an ILP-based assignment method. The assignment generated by IPASS is fed into the structure calculation step, FALCON-NMR. FALCON-NMR has a threading module, an ab initio module, an all-atom refinement module, and an NOE constraints-based decoy selection module. The entire system, AMR, is successfully tested on four out of five real proteins with practical NMR spectra, and generates 1.25A, 1.49A, 0.67A, and 0.88A to the native reference structures, respectively. Another contribution of this dissertation is to propose novel ideas and methods to solve three protein structure prediction problems which are closely related to NMR protein structure determination. We develop a novel consensus contact prediction method, which is able to eliminate server correlations, to solve the protein inter-residue contact prediction problem. We also propose an ultra-fast side chain packing method, which only uses local backbone information, to solve the protein side chain packing problem. Finally, two complementary local quality assessment methods are proposed to solve the local quality prediction problem for comparative modeling-based protein structure prediction methods.

NMR spectroscopy is not only for protein structure determination, but also for drug screening and studies of dynamics and interactions. In both cases, one of the main bottleneck steps is backbone assignment. When a homologous structure is available, it can accelerate assignment. Such structure-based methods are the focus of this thesis. This thesis aims for fast and robust methods for NMR assignment problems; in particular, structure-based backbone assignment and chemical shift mapping. For speed, we identified situations where the number of 15N-labeled experiments for structure-based assignment can be reduced; in particular, when a homologous assignment or chemical shift mapping information is available. For robustness, we modeled and directly addressed the errors. Binary integer linear programming, a well-studied method in operations research, was used to model the problems and provide practically efficient solutions with optimality guarantees. Our approach improved on the most robust method for structure-based backbone assignment on 15N-labeled data by improving the accuracy by 10% on average on 9 proteins, and then by handling typing errors, which had previously been ignored. We show that such errors can have a large impact on the accuracy; decreasing the accuracy from 95% or greater to between 40% and 75%. On automatically picked peaks, which is much noisier than manually picked peaks, we achieved an accuracy of 97% on ubiquitin. In chemical shift mapping, the peak tracking is often done manually because the problem is inherently visual. We developed a computer vision approach for tracking the peak movements with average accuracy of over 95% on three proteins with less than 1.5 residues predicted per peak. One of the proteins tested is larger than any tested by existing automated methods, and it has more titration peak lists. We then combined peak tracking with backbone assignment to take into account contact information, which resulted in an average accuracy of 94% on one-to-one assignments for these three proteins. Finally, we applied peak tracking and backbone assignment to protein-ligand docking to illustrate the potential for fast 3D complex determination.

NMR spectroscopy is not only for protein structure determination, but also for drug screening and studies of dynamics and interactions. In both cases, one of the main bottleneck steps is backbone assignment. When a homologous structure is available, it can accelerate assignment. Such structure-based methods are the focus of this thesis. This thesis aims for fast and robust methods for NMR assignment problems; in particular, structure-based backbone assignment and chemical shift mapping. For speed, we identified situations where the number of 15N-labeled experiments for structure-based assignment can be reduced; in particular, when a homologous assignment or chemical shift mapping information is available. For robustness, we modeled and directly addressed the errors. Binary integer linear programming, a well-studied method in operations research, was used to model the problems and provide practically efficient solutions with optimality guarantees. Our approach improved on the most robust method for structure-based backbone assignment on 15N-labeled data by improving the accuracy by 10% on average on 9 proteins, and then by handling typing errors, which had previously been ignored. We show that such errors can have a large impact on the accuracy; decreasing the accuracy from 95% or greater to between 40% and 75%. On automatically picked peaks, which is much noisier than manually picked peaks, we achieved an accuracy of 97% on ubiquitin. In chemical shift mapping, the peak tracking is often done manually because the problem is inherently visual. We developed a computer vision approach for tracking the peak movements with average accuracy of over 95% on three proteins with less than 1.5 residues predicted per peak. One of the proteins tested is larger than any tested by existing automated methods, and it has more titration peak lists. We then combined peak tracking with backbone assignment to take into account contact information, which resulted in an average accuracy of 94% on one-to-one assignments for these three proteins. Finally, we applied peak tracking and backbone assignment to protein-ligand docking to illustrate the potential for fast 3D complex determination.

The enzyme nitroreductase (NR) catalyzes two-electron reduction of high explosives such as trinitrotoluene (TNT), a wide variety of other toxic nitroaromatic compounds, as well as riboflavin derivatives, using a tightly-bound flavin mononucleotide (FMN) cofactor. It has important environmental and clinical implications. Previous studies have focused on elucidating NRs catalytic mechanism and solving its crystal structure. In this dissertation work, we first develop and implement new strategies for labeling NR with stable isotopes, and report a completely re-designed protocol for NRs purification. Then we report the successful assignment of over half of NRs backbone resonances using 3d-NMR methods. The most significant observation is that we find a well-resolved 2d 1H-15N HSQC NMR spectrum is obtained at 37°C for NR, while the HSQC at 4°C is poorly-dispersed and comprised of sharp overlapped peaks. Thus, it would appear that NR denatures at 4°C. However, as we demonstrate, the non-covalently-bound FMN cofactor is not released at the lower temperature, based on retention of the native flavin visible-CD spectrum. Similarly, far-UV CD spectroscopy shows that the protein retains full secondary structural content at 4C. In addition, near-UV CD and Fluorine-19 NMR experiments demonstrate that under completely native conditions (neutral pH, no additives) NR maintains a high degree of tertiary structure and well-defined hydrophobic side-chain packing, ruling out the possibility of a molten-globule state. Thus, our studies report strong evidence that the dramatic low-temperature (low-T) transition near 20°C observed by NMR is not the result of protein structural changes, but rather, we propose that NR exists as an ensemble of rapidly inter-converting structures, at lower temperature, only adopting a well-defined unique structure above 20°C. Both saturation-transfer from water and solvent proton-exchange measurements support our proposed model in which the unique high-T structure is favored entropically, by release of water molecules; on the other hand, the fluxional low-T state incorporates greater water solvation at 4°C. In the latter part of this study, we present preliminary data suggesting that the flexibility implied by easy water-access to the loosely-structured state plays a role in accommodating binding of diverse substrates. Therefore, the fluxional low-T state may be functionally important. A possible direct link between the enzyme dynamics and its catalytic activity will be an area of future investigation.

Understanding biological phenomena at atomic resolution is one of the keys to modern drug design. In particular, knowledge of 3D structures of proteins and their interactions with other macromolecules are necessary for designing chemical compounds that modify biological processes. Conventional methods for protein structure determinations comprise X-ray crystallography and nuclear magnetic resonance (NMR) spectroscopy. These techniques can also determine the binding mode of chemical compounds. Either technique can be slow and costly, making it highly relevant to explore alternative strategies. Paramagnetic NMR spectroscopy is emerging as such an alternative technique. In order to measure the paramagnetic effects, two NMR spectra are compared that have been measured with and without a bound paramagnetic metal ion. In particular, pseudocontact shifts (PCS) of nuclear spins are easily measured as the difference (in ppm) of the chemical shifts between the two spectra. PCSs provide long range and orientation dependent restraints, allowing positioning of the spin with respect to the magnetic susceptibility tensor anisotropy (Δχ-tensor) of the metal ion. In this thesis, I used the PCS effect to computationally extract information from NMR spectra. I developed (i) a tool (called Possum) to automatically assign diamagnetic and paramagnetic spectra of the methyl groups of amino acid side chains, given structural information of the protein studied and prior knowledge of the Δχ-tensor; (ii) I designed a comprehensive software package (called Numbat) to extract Δχ-tensor parameters from assigned PCS values and the available 3D structure; and (iii) I incorporated PCS-based restraints into the protein structure prediction software CS-ROSETTA and demonstrated that this combination (PCS-ROSETTA) presents a significant improvement for de novo structure determination. The three projects serve different purposes at different stages of protein NMR studies. They could be combined in the following manner: Starting from assigned backbone PCSs, PCS-Rosetta could be used to determine the 3D structure of the protein. Possum can then be used to automatically assign the NMR resonances of the methyl groups using PCSs. Finally, Numbat can be used to fit improved Δχ-tensors to all the PCS data, analyze the quality of the Δχ-tensors and identify possible wrong assignments. Iterative repetition of this protocol would give a 3D structural model of the protein with a minimum of data. Alternatively, the Δχ-tensor parameters and PCSs could be used as input for a traditional software package such as Xplor-NIH to compute a 3D structure of the protein.

Understanding biological phenomena at atomic resolution is one of the keys to modern drug design. In particular, knowledge of 3D structures of proteins and their interactions with other macromolecules are necessary for designing chemical compounds that modify biological processes. Conventional methods for protein structure determinations comprise X-ray crystallography and nuclear magnetic resonance (NMR) spectroscopy. These techniques can also determine the binding mode of chemical compounds. Either technique can be slow and costly, making it highly relevant to explore alternative strategies. Paramagnetic NMR spectroscopy is emerging as such an alternative technique. In order to measure the paramagnetic effects, two NMR spectra are compared that have been measured with and without a bound paramagnetic metal ion. In particular, pseudocontact shifts (PCS) of nuclear spins are easily measured as the difference (in ppm) of the chemical shifts between the two spectra. PCSs provide long range and orientation dependent restraints, allowing positioning of the spin with respect to the magnetic susceptibility tensor anisotropy (Δχ-tensor) of the metal ion. In this thesis, I used the PCS effect to computationally extract information from NMR spectra. I developed (i) a tool (called Possum) to automatically assign diamagnetic and paramagnetic spectra of the methyl groups of amino acid side chains, given structural information of the protein studied and prior knowledge of the Δχ-tensor; (ii) I designed a comprehensive software package (called Numbat) to extract Δχ-tensor parameters from assigned PCS values and the available 3D structure; and (iii) I incorporated PCS-based restraints into the protein structure prediction software CS-ROSETTA and demonstrated that this combination (PCS-ROSETTA) presents a significant improvement for de novo structure determination. The three projects serve different purposes at different stages of protein NMR studies. They could be combined in the following manner: Starting from assigned backbone PCSs, PCS-Rosetta could be used to determine the 3D structure of the protein. Possum can then be used to automatically assign the NMR resonances of the methyl groups using PCSs. Finally, Numbat can be used to fit improved Δχ-tensors to all the PCS data, analyze the quality of the Δχ-tensors and identify possible wrong assignments. Iterative repetition of this protocol would give a 3D structural model of the protein with a minimum of data. Alternatively, the Δχ-tensor parameters and PCSs could be used as input for a traditional software package such as Xplor-NIH to compute a 3D structure of the protein.

La RMN à l’état solide est une méthode de choix pour l’étude des protéines insolubles et des complexes protéiques de haut poids moléculaire. L’insolubilité intrinsèque des protéines fibrillaires, ainsi que leur architecture complexe, rendent difficile leur caractérisation structurale par la cristallographie et par la RMN en solution. La RMN à l‘état solide n’est pas limitée par le poids moléculaire et constitue donc un outil puissant pour l’étude des protéines fibrillaires. L’attribution des résonances RMN est le prérequis pour obtenir des informations structurales à résolution atomique. La première partie de ce travail de thèse décrit le développement de méthodes en RMN à l’état solide pour l’attribution des résonances. Nous avons appliqué ces méthodes afin d’attribuer le domaine C-terminal du prion Ure2 (33 kDa), qui est à ce jour la plus grande protéine attribuée par RMN à l’état solide. Nos résultats fournissent les bases pour l’étude de protéines à haut poids moléculaire à l’échelle atomique. Ceci est démontré dans la seconde partie de ce travail de thèse avec les premières études RMN à l’état solide des fibrilles des prions Ure2 et Sup35. Nous avons caractérisé la structure de ces prions pour les fibrilles entières ainsi que pour les domaines isolés. La troisième fibrille étudiée est l’α- synuclein, fibrille associée à la maladie de Parkinson, pour laquelle nous présentons l’attribution des résonances RMN ainsi que la structure secondaire d’un nouveau polymorphe. Les études présentées ici fournissent de nouvelles clés pour comprendre la diversité des architectures de fibrilles, en considérant les fibrilles comme entités individuelles d’un point de vue structural / Solid-state NMR is the method of choice for studies on insoluble proteins and other high molecular weight protein complexes. The inherent insolubility of fibrillar proteins, as well as their complex architecture, makes the application of x-ray crystallography and solution state NMR difficult. Solid-state NMR is not limited by the molecular weight or by the absence of long-range structural order, and is thus a powerful tool for the 3D structural investigation of fibrillar proteins. The assignment of the NMR resonances is a prerequisite to obtain structural information at atomic level. The first part of this thesis describes the development of solid-state NMR methods to assign the resonances in large proteins. We apply these methods to assign the 33 kDa C-terminal domain of the Ure2p prion which is up to now the largest protein assigned by solid-state NMR. Our results provide the basis to study high molecular weight proteins at atomic level. This is demonstrated in the second part with the first high-resolution solid-state NMR study of Ure2 and Sup35 prion fibrils. We describe the conformation of the functional domains and prion domains in the full-length fibrils and in isolation. The third fibrillar protein addressed in this work is the Parkinson’s disease related α-synuclein whereof we demonstrate the NMR resonance assignment and the secondary structure determination of a new polymorph. Thus, the studies described here provide new insights in the structural diversity of fibril architectures, and plead to view fibrils as individuals from a structural point of view, rather than a homogenous protein family

No description available.