5

MACROMOLECULAR STRUCTURE DETERMINATION BY NMR SPECTROSCOPY

John L. Markley, Arash Bahrami, Hamid R. Eghbalnia, Francis C. Peterson, Robert C. Tyler, Eldon L. Ulrich, William M. Westler, and Brian F. Volkman

Bioinformatics plays an important role in biomolecular structure determination by NMR spectroscopy. Increasingly, structural biology is being driven by information gained from gene sequencing and hypotheses derived from bioinformatics approaches. This is particularly so for structural proteomics efforts. A major rationale for structural proteomics is that a more complete mapping of peptide sequence space onto conformational space will lead to efficiencies in determining structure-function relationships1, 2, 3, 4. Longer range scientific goals are the prediction of structure and function from sequence and simulations of the functions of a living cell. Structural proteomics is part of a wider functional genomics effort, which promises to assign functions to proteins within complex biological pathways and to enlarge the understanding and appreciation of complex biological phenomena5. Because of its potential to greatly broaden the targets for new pharmaceuticals, structural proteomics is expected to join combinatorial chemistry and screening as an integral approach to modern drug discovery6. It is clear that much larger databases of structures, dynamic properties of biomolecules, biochemical mechanisms, and biological functions are needed to approach these goals. Because of its ability to provide atomic-resolution structural and chemical information about proteins, NMR spectroscopy is positioned to play an important role in this endeavor. Already, NMR spectroscopy contributes about 15% of the protein structures deposited at the Protein Data Bank (PDB). In addition, NMR spectroscopy is used routinely in high-throughput screens to determine protein-ligand interactions7, 8. NMR also is a key tool in mechanistic enzymology and in studies of protein folding and stability. As discussed here, advances in key technologies promise rapid increases in the efficiency and scope of NMR applications to structural and functional genomics. Bioinformatics figures prominently in these advances.

This Chapter discusses the current status and prospects of macromolecular structure determination by solution state NMR spectroscopy. Although the focus is on proteins, the approaches can be generalized to other classes of biological macromolecules. Solid-state NMR spectroscopy shows great promise for structural studies of proteins that may not be amenable to investigation in solution, such as membrane proteins, and for functional investigations of protein in the solid state. Solid-state NMR strategies for biomolecular sample preparation, data collection, and analysis are developing rapidly and are expected to assume prominence in the next few years. Further discussion of this highly specialized field is beyond the scope of this chapter, and interested readers are directed to recent reviews9, 10, 11.

This chapter starts with an overview of the physical basis for NMR spectroscopy and a discussion of NMR experiments and instrumentation. Then, the basic approaches used in determining structures and dynamic properties of proteins are presented. Sample requirements and sample preparation, including labeling methods, are discussed. Next a protocol for NMR structure determination covering all steps from protein characterization to data deposition is detailed. This is followed by a section on evolving bioinformatics technology for achieving a more automated, probabilistic approach to protein NMR structure determination. The final section summarizes various bioinformatics resources that are available to support biomolecular NMR spectroscopy.

Comparison of NMR Spectroscopy and X-ray Crystallography

Current methods for high-throughput protein structure determination are single-crystal diffraction and solution-state NMR spectroscopy. The two methods have complementary features, as X-ray crystallography represents a mature and rapid approach for proteins that form suitable crystals while NMR has advantages for structural studies of small proteins that are partially disordered, exist in multiple stable conformations in solution, show weak interactions with important cofactors or ligands, or do not crystallize readily. NMR spectroscopy is an incremental method that can rapidly provide useful information concerning overall protein folding, local dynamics, existence of multiple folded conformations, or protein-ligand or protein-protein interactions in advance of a three-dimensional structure. This information can be useful in designing strategies for 3D structure determinations by either NMR or X-ray crystallography. Several ongoing structural proteomics pilot projects are employing a combination of X-ray crystallography and NMR spectroscopy.

Physical Basis for Biomolecular NMR Spectroscopy

As summarized schematically in Figure 5.1, NMR spectroscopy investigates transitions between spin states of magnetically active nuclei in a magnetic field. The most important magnetically active nuclei for proteins are the proton (1H), carbon-13 (13C), nitrogen-15 (15N), and phosphorus-31 (31P). Each of these stable (nonradioactive) isotopes has a nuclear spin of one half and, as a consequence, two spin states: one at lower energy, with the magnetic spin paired with the external field, and one at higher energy, with the magnetic spin opposing the external field. The magnetic moment of each nucleus precesses about the external magnetic field and is influenced by other fields. Influences on a given spin are generated by neighboring spins in the molecule, giving rise to intrinsic NMR parameters, or by radiofrequency pulses and/or pulsedfield gradients as programmedby theNMRspectroscopist. In an NMR experiment, a time sequence of radiofrequency pulses and pulsed field gradients are applied to the spins present in the molecule studied. The excited spins are allowed to interact with one another and with the externalmagnetic field. Then, the state of the system is read out by detection of the current induced by the nuclear spins of the sample in the receiver coil of the NMR spectrometer. The physical basis for NMR is well understood, and the spectroscopic consequences of a given pulse sequence applied to a particular molecule can be simulated to a good level of precision12, 13.

Figure 5.1. Schematic illustrating several basic principles of biomolecular NMR spectroscopy. (Clockwise from upper left) The NMR sample solution is placed in an NMR tube, which is then inserted into the static magnetic field of the NMR spectrometer. NMR-active nuclei (e.g., 1H, 13C,and 15N) within the sample precess (like a spinning gyroscope) with their net magnetization aligned with the direction of the magnetic field; the precession frequency is governed by the type of nucleus but is modulated by the chemical environment of the atom in the macromolecule. Application of a radiofrequency pulse excites the nucleus by flipping its spin to the excited energy level opposing the external magnetic field. A “90 pulse” flips the rotating magnetization into the plane perpendicular to the axis of the magnetic field (the “x-y plane”), and the rotating magnetization induces an oscillating current (the “free induction decay”) into the coil that detects the signal. Fourier transformation of the time-domain signal (free induction decay) converts it to a frequency-domain signal (the “NMR peak”). More complicated pulse sequences are used in experiments that create two- or three-dimensional NMR spectra; the type of 3D spectrum illustrated is the result of a “triple resonance” experiment in which the three orthogonal frequency axes display the peak positions of combinations of 1H, 13C, and 15N nuclei that are related by spectral connectivities created by the pulse sequence.

images

NMR Experiments

The NMR spectrometer can be programmed to operate on sets of nuclei of single or multiple atom types and to accommodate particular spectral features (chemical shift range, spin-spin coupling range, etc.). NMR spectroscopy is unique in that the Hamiltonian ofthe system under study can be manipulated easily by the application of radiofrequency pulses and/or pulsed field gradients. The versatility in creating sequences of complex “spin gymnastics” makes possible a myriad of NMR experiments by which particular parameters can be investigated. A huge variety of NMR experiments are at one’s disposal when collecting data for a structure determination or functional investigation. The NMR field is still young and dynamic, and these experiments continue to evolve; thus the optimal set of experiments for a structure determination or structure-function study is a matter of exploration and individual taste. The approaches described here are ones that we have found to be effective.

Structure from NMR

The classical approach to structure determination is to first use multidimensional and multinuclear NMR methods to determine “sequence-specific assignments,” that is, resolve signals from the 1H, 15N, and 13C nuclei of a protein and to assign them to specific nuclei in the covalent structure of the molecule. The assigned chemical shifts (cs) themselves provide reliable information about the secondary structure of the protein14, 15, 16, 17, and the oxidation states of cysteine residues18, and can be used to test or validate structural models. Additional structural restraints are obtained from an interpretation of data from one or more different classes of NMR experiments: nuclear overhauser effect (NOE) spectra, which provide 1H — 1H distance constraints; empirical torsion angle constraints based on chemical shifts19; and (3) residual dipolar couplings from partially oriented proteins20, 21, 22, 23, which provide spatial constraints (with respect to the orientation axes) for pairs of coupled nuclei. Additional hydrogen bond constraints are determined from hydrogen exchange experiments, chemical shifts, and/or trans-hydrogen bond couplings24, 25.

Molecular Size Limitations

Protein NMR methods have advanced to the point where small to medium sized protein domain structures can be determined in a routine manner. Key milestones in the development of this technique include NOE-based resonance assignment and structure refinement26, efficient stable isotope labeling methods27, detection of multinuclear correlations28, 29, 30, 31, 32, 33 coupled with indirect detection and multinuclear NMR spectroscopy34, 35. More recent advances have included the use of pulsed field gradients for coherence selection and line-narrowing by transverse relaxation-optimized spectroscopy (TROSY)36, 37, 38, 39, 40, 41, 42. Theimprovements incommerciallyavailableNMRhardware and remarkable advances in desktop computing capabilities have also been important.

Protein domains of up to roughly 25 kDa generally are amenable to high-throughput NMR structure determination, provided that soluble, stable, U — 13C- and U — 15N-labeled samples can be obtained. This isotope enrichment enables the use of sensitive multinuclear correlation experiments for the task of matching all chemical shifts in the protein with specific amino acid residues and atoms. Once all 1H, 15N, and 13C resonances have been assigned, the correlations observed in nuclear Overhauser enhancement experiments can be interpreted in terms of short (< 5Å) inter-proton distances. NMR structures are obtained from constrained molecular dynamics calculations, with these NOE-derived 1H--1H distances as the primary experimental constraints. As a consequence of chemical shift degeneracy, many NOE correlations may have multiple assignment possibilities, and the results of preliminary structure calculations are used to eliminate unlikely candidates on the basis of inter-proton distances. Refinement continues in an iterative manner until a self-consistent set of experimental constraints produces an ensemble of structures that also satisfies standard covalent geometry and steric overlap considerations. Recent advances in software have automated many of the steps in structure determination43. Structures are validated and reported in the Protein Data Bank in a manner analogous to those obtained from X-ray crystallographic methods (see Chapters 14 and 15).

Stable Isotope Labeling

Structures of smaller proteins can be determined with samples at natural abundance. If the protein is reasonably soluble (≥ 1 mM), and particularly if a high-sensitivity cryogenic probe is available, it is possible to obtain 1H-13C and 1H-15N correlations at natural abundance. This additional information can be useful for signal assignments and to provide additional chemical shifts for secondary structure analysis. These structures are generally of lower quality than those determined from labeled proteins because of the greater difficulty in obtaining complete assignments. High-throughput strategies for protein structure determination generally are designed for proteins ’double-labeled’ with 13C and 15N (Mr >~25,000 Da) or “triple-labeled” with 13C, 15N, and 2H(Mr <~25,000 Da). NMR spectra of larger molecules are characterized by increasingly rapid decay of the signal to be detected leading to increased line widths and lowered sensitivity (signal-to-noise ratio) and by greater overlap and signal degeneracy. Even with triple-labeled proteins, these problems make it increasingly difficult to solve NMR structures of proteins larger than 40 kDa with current methodology. Using higher magnetic fields and TROSY methodology suggests that these limitations can be overcome to some extent. Kay and coworkers have developed a labeling strategy for large molecules by 15N, 13C, and 2H labeling while the methyl groups of isoleucine, leucine, and valine are protonated44, 45.

Selective labeling approaches make it possible to obtain information about selected regions of proteins as large as 150 kDa. These methods include labeling by single residue type27 and segmental labeling by peptide semisynthesis, including intein technology46, 47. In the designer labeling scheme, known as stereo-array isotope labeling (SAIL)48, a complete stereospecific and regiospecific pattern of stable isotopes is incorporated into the protein using cell-free protein synthesis. SAIL, by significantly reducing the proton density and thus reducing the resonance overlap and line width, has great potential in extending the molecular weight range for structure determination by NMR.

Data Collection and Instrumentation

Even with protein sample concentrations in the millimolar range (> 10 mg/mL), coaddition of 8 or 16 transients may be required at each time point in some two- or three-dimensional experiments to achieve sufficient signal-to-noise (s/n). Cryogenic probe technology can reduce the data acquisition period by up to a factor of 10. This results from the three- to fourfold enhancement in sensitivity achieved by cooling the probe and preamplifier circuitry to very low temperatures49. Because of the square-root relationship between time and sensitivity in signal averaging, an inherent sensitivity improvement of threefold is equivalent to a ninefold increase in signal averaging. In practical terms, the s/n of an experiment acquired with one or two transients at each time increment on a cryogenic probe should be equivalent (if not superior) to one acquired with 16 transients using a conventional (noncooled) probe. For this example, the cryogenic probe results in an eightfold reduction in experiment time with no negative consequences. This is especially important for NOE experiments. There are some caveats associated with the use of cryogenic probes. First, the three- to fourfold sensitivity increases generally are realized only for simple 1D experiments in nonionic solutions. The RF field inhomogeneity of the 13C and 15N coils decreases the efficiency of the probes in multipulse triple-resonance experiments. Practically, for a protein solution at low ionic strength, the sensitivity increase is closer to twofold. However, this increase yields either a fourfold saving in time or, more importantly for systems of poor solubility, the ability to lower the sample concentration by a factor of two. The most severe limitation of the cryogenic probe is that they lose their sensitivity advantage in solutions having high ionic strength. Recent studies have shown that by changing the shape of the sample tube to limit the penetration of the electric field into the salty solution, the sensitivity can be recovered even in solutions of high ionic strength50.

Approaches for decreasing the overall experimental time by collecting sparse data sets have been developed, such as HIFI51, GFT52, and reconstruction of 3D from 2D planes53. These methods decrease the collection time simply by collecting projected 2D planes out of a 3D spectrum and can reduce the data collection time by nearly an order of magnitude. The ultimate limit for rapid data collection is the longitudinal relaxation rate of the nucleus to be detected. Data collection schemes, such as “SOFAST,” that use optimal recycle rates have produced 2D spectra in less than 1 min54, and these strategies have been developed into a suite of pulse sequences for triple-resonance data collection55.

Protein Dynamics from NMR

NMR signals report on the chemical properties of nuclei including their relative motions with respect to the magnet and molecular frames. Consequently, NMR spectroscopy is uniquely suited for investigations of the kinetics and thermodynamics of molecular motions, segmental motions, dynamic conformational equilibria, and ligand binding equilibria— as reviewed in greater detail elsewhere56, 57, 58 (Palmer III, Kroenke, and Loria, 2001). Whereas regions of proteins that are statically or dynamically disordered may fail to produce resolvable electron density, NMR signals from such regions generally are observed and can be investigated59.

Dynamic processes observable by NMR spectroscopy cover a vast timescale: from picoseconds to seconds, days, and even months. Nuclear spin relaxation rates are sensitive to protein motions ranging from picosecond to milliseconds. Overall molecular tumbling on the nanosecond timescale is the dominant factor in relaxation of most nuclei in folded proteins. Internal protein motions on the picosecond timescale cause a slight averaging of the dipolar interactions that in turn reduces the overall spectral density. These variations can be measured for the backbone of each residue in a protein by determining the time constants for 15N longitudinal (T1) and transverse (T2) relaxation rates as well as the 1H(15N) heteronuclear NOE60. Analogous measurements of 2H or 13C relaxation parameters can yield information on side chain dynamics61.

The motional effect on relaxation rates can be expressed as an order parameter (S2) in a model-free formalism as originally expounded by Lipari and Szabo62, 63. Order parameters range from 0 to 1, with smaller values reflecting increased internal motion. Often the simple single exponential model is insufficient to fit the relaxation data, and additional exponentials describing other rapid motions must be added to extend the basic model-free approach64, 65. However, it should be noted that as parameters are added to the analysis, the available experimental data points may not suffice to constrain the system to a unique neighborhood of solutions66. Outliers in the transverse relaxation rate often arise from motions in the microsecond-to-millisecond time regime that report on chemical or confor-mational exchange phenomena67. Relaxation dispersion methods detect these slower motions directly68, and exchange rates have been correlated with protein binding and folding69, rate-limiting steps in enzyme catalysis70, and even conformational dynamics of the 20S proteasome— a 670 kDa complex71. For conformational dynamics involving millisecond-to-second interconversion between two or more states giving rise to distinct NMR signals, traditional 2D exchange spectroscopy provides another method for quantification of exchange rates, and ultrafast 2D NMR pulse schemes enable real-time monitoring of kinetic processes on the 1-10 s timescale72.

Dynamics on longer timescales can be interrogated through hydrogen exchange kinetics73. In this experiment, a protein that has the natural abundance of 1H atoms on the amide groups is dissolved in a solution containing ~100% 2H2O and then quickly transferred to the spectrometer. By observing the loss of amide 1H signals due to exchange with the 2H atoms, the kinetics can be followed. Protons with larger protection factors are interpreted to be in more highly structured regions with more stable hydrogen bonds. Highly labile protons may exchange completely in the interval between the addition of 2H2O and acquisition of the first NMR spectrum. A complementary method (CLEANEX) for rapidly exchanging amides uses solvent saturation transfer to quantify exchange rates74. In combination, saturation transfer and H/D exchange measurements can monitor local backbone fluctuations from milliseconds to months. This broad repertoire of NMR methods enables quantitative, site-specific measurement of fast motions and slower conformational exchange process associated with catalysis, allostery, folding, and signaling—central features of most biological processes.

Approaches to Protein Production and Labeling

A variety of protein expression systems have been investigated in support of structural proteomics75. By far the most widely used and most successful platform is heterologous overexpression of proteins in Escherichia coli cells. It is a general experience, however, that a certain percentage of proteins does not express abundantly in soluble form in bacteria. Insolubility arises either from an intrinsic property of the protein (for example, aggregation due to an exposed hydrophobic surface) or because inappropriate folding mechanisms in the expression host permit aggregation of folding intermediates76, 77, 78. Expression can be particularly difficult for eukaryotic proteins that consist of multiple domains, that require cofactors or protein partners for proper folding, or that normally undergo extensive post-translational modification. In industry, the next most widely used approach is protein production from insect cells by a baculovirus system. However, this approach is not widely used for NMR spectroscopy, because labeled media for insect cells are highly expensive. Recently, a more economical medium for Sf9 insect cells has been described, which promises to improve the practicality of this approach79. Yeast cells offer another alternative, and approaches for large-scale production of labeled proteins from the yeast Pichia pastoris have been described80, 81.

Cell-free biological systems are capable of synthesizing proteins with high speed and accuracy approaching in vivo rates82, 83. Commercial implementations of this technology incorporate recent advances as described below84, 85, 86, 87, 88. E. coli cell-free systems give high yields89, and active extracts are easily prepared90, 91. Membrane proteins can be produced in soluble form by incorporating detergents into cell-free systems92, 93, 94. Two groups have made extensive use of cell-free (in vitro) protein production for NMR spectroscopy: the RIKEN Structural Genomics Group95 and the Center for Eukaryotic Structural Genomics96.

In our experience, a combination of protein production in E. coli cells and in a wheat germ cell-free system provides an economical approach for producing eukaryotic proteins for NMR structural analysis. The two methods are complementary in the sense that many proteins that fail to be produced in soluble, folded state from E. coli can be made from the wheat germ cell-free system, whereas, when production is successful from E. coli, the yields can be very high97.

Cell-Based Methods

As noted above, protein production in E. coli has an established record of being the most successful approach for providing protein targets for structural proteomics. Suitable expression vectors are readily available, and the method is more economical than production in eukaryotic cells. Furthermore, promising proteins can be subsequently labeled metaboli-cally with heavy atom-labeled amino acids for X-ray or stable isotopes for NMR98. 13C—, 15N—, and 2H-metabolic labeling for NMR work was pioneered with E. coli expression systems27. A recent important advance has been the development of self-induction media containing glucose, glycerol, and lactose for protein production from E. coli, which support very high cell densities while removing the need to follow cell growth to initiate induction and associated centrifugation steps. This approach has been fine-tuned recently to improve yields99. The approach can be used to produce stable isotope labeled samples for NMR spectroscopy100, although the costs for isotopes are higher, because 13C -glycerol must be added in addition to13C-glucose and 15N-ammonia. Additional cost saving can be achieved with E. coli production systems through the use of disposable polyethylene terephthalate (PET) bottles as growth vessels, which do not require cleaning or sterilization101.

Cell-Free Methods

We have published detailed descriptions of the protocols, which have been developed for the production of labeled proteins for NMR spectroscopy from a wheat germ cell-free system102, 103, 104. Cell-free protein production methods have several advantages: (1) volumes can be kept manageable—in favorable cases, sufficient protein for a structural investigation can be produced from a reaction volume of under 6-12 mL; (2) proteins can be produced which would be toxic to cells; (3) the ratio of desired protein to unwanted protein is low, simplifying purification; (4) in a cell-free system, one has the potential for introducing metal ions, cofactors, or other agents that promote correct folding; (5) enzymes and substrates can be added to promote post-translational modifications; and (6) labeled amino acids can be introduced without scrambling of the label— this enables clean selective labeling and newer labeling strategies, such as stereo-array isotope labeling48.

The production of labeled proteins from cell-free systems requires that labeled amino acids be supplied in the reaction mixture. The optimal concentration of each amino acid in the cell-free reaction mixture is about 1 mM. At currently achievable protein yields (0.4-1.1 mg protein/mL of reaction mixture), a reaction volume of 6-12 mL is needed to produce enough protein for an NMR sample. [U-15N]-and [U-15N, U-13C] -amino acids are commercially available at prices that make cell-free protein production of labeled proteins competitive with production from E. coli cells when relative labor costs are taken into account. Figure 5.2 illustrates the automated cell-free production and purification of three protein targets incorporating 15N-amino acids for screening for suitability as structural targets.

Sample Concentration, Volume, and Stability

The higher the protein concentration the faster NMR data can be collected, provided that the protein does not aggregate. Practical lower limit protein concentrations are about 200 μM with ordinary probes and about 60 μM with cryogenic probes. The decreased concentration requirement of the cryogenic probe may make it possible to collect data from proteins that aggregate at higher concentrations. With ordinary NMR tubes, a sample volume of 300-500 μL is required, depending on the length of the detection coil in the probe. Susceptibility-matched cells (Shigemi Inc.) enable the use of smaller volumes: 200 μL in a 5 mm OD tube or 60 μL in a 3 mm OD tube. In cases where the protein is sufficiently soluble (> 1.5 mM), it is possible to use a 1 mm microcoil NMR probe to greatly reduce the amount of protein required105, 106.

Figure 5.2. Examples of 15N-labeled His-tagged proteins (two Arabidopsis and one human target) produced by a wheat germ cell-free robotic system (CellFree Sciences DT-II). The robot starts with a DNAplasmid and carries out automated transcription and translation and metal ion affinity protein purification. The SDS gel scans show molecular weight markers on either side, and for each target, the total reaction mixture, column flow-through, and purified soluble fraction seem incomplete. The two bands, at higher molecular weight, correspond to proteins present in the wheat germ extract that bind to the affinity column. They do not show up in the NMR spectra because they are not labeled with 15N. The only manual steps required in preparing the NMR samples were buffer exchange and concentration of the purified protein. The three targets illustrate how 1H-15N HSQC NMR is used to screen targets for their suitability for structure determination. HSQC+ indicates a spectrum with the expected number of well-resolved peaks. HSQC+/- indicates a protein that is partially unstructured and may be salvageable as a structural target by either solvent optimization or by trimming the protein expressed. HSQC indicates a natively unfolded or aggregated protein unsuitable for NMR structure determination (Loushin Newman, Markley, Song, and Vinarov, unpublished results).

images

The sample must remain stable over the data collection period. If the stability of the sample is a limiting factor, it is possible to use separate samples for different NMR experiments. Ideally, all data sets are collected with the same sample, because slight differences in solution conditions can lead to chemical shift differences that make it difficult to compare results from different experiments. One of the advantages of cryogenic probes and higher fields is that the overall data collection time for each experiment is shortened. This makes it possible to investigate systems that are less stable over time.

As noted above, NMR methodology is evolving rapidly and competing approaches are being explored in various structural proteomics centers. The following sections outline a protocol for semiautomated protein structure determination by NMR spectroscopy. Proteins (≤25 kDa) labeled with 15N emerging from the cell-based or cell-free protein production pipeline are subjected to 1D and 2D NMR screening to determine their folding and aggregation states under a variety of solution conditions (Figure 5.2). Ideally, an NMR spectrometer equipped with a sample preparation robot, sample changer, and cryogenic probe is used for these screens (a 600 MHz system is adequate for this purpose). NMR screens are used to determine aggregation state (through measurement of transverse diffusion by gradient NMR), assess folding state (from analysis of the amide region of 1D 1HNMR spectra and the pattern of1H-15N correlation spectra), monitor ligand binding (by detecting chemical shift perturbations), and survey proteins for the presence of unstructured regions. The results of the NMR screens are used in deciding whether to proceed with NMR structure determination, transfer the protein to the crystallization trials pipeline, or re-engineer the protein to improve its suitability for NMR analysis. This may be accomplished by removing unstructured regions to simplify the 2D and 3D NMR spectra or by isolating individual domains to achieve greater stability or solubility and discourage aggregation107, 108. Proteins that pass these screens as candidates for NMR structure determination are double-labeled with 15N and 13C as described above.

A 600 MHz NMR spectrometer equipped with a cryogenic probe is well suited to the collection of protein triple-resonance spectra. If available, spectrometers with cryogenic probes operating at or above 750 MHz may be employed in collecting multidimensional NOE and isotope-filtered/selected NOE spectra with enhanced resolution. Measurement of residual dipolar couplings (RDCs) for structure refinement and validation109, 110, 111, 112, 113 requires that samples be partially oriented in the magnetic field. The orientation can be achieved by dissolving the protein in an anisotropic medium, such as phospholipid bicelles114, 115 (Sanders, Schaff, and Prestegard, 1993), phage particles116, or stretched polyacrylamide gels117. Identification of asuitable alignment medium may require additional screening, based on the ranges of pH and temperature that are compatible with the target protein and the degree of association between the protein and medium. In the absence of a robust, “universal” alignment medium, RDC analysis is treated as a supplemental approach to this protocol, and structures are solved using primarily distance constraints derived from NOESY spectra and dihedral angle constraints predicted from chemical shifts. In the absence of RDCs, validation is performed by using a variety of software tools that analyze the coordinates and constraint data.

Standard Data Collection Protocols

The improved sensitivity from cryogenic probes permits data collection sufficient for a protein structure in 7-10 days, rather than 1 or 2 months. If the protein concentration is sufficiently high that data can be recorded without signal averaging, the total data collection period can be as short as 30 h. The 2D and 3D pulse schemes listed in Table 5.1 have been in routine use in many laboratories for several years.

Updated versions of these pulse sequences that are compatible with cryogenic probes are provided as part of the standard pulse sequence libraries by NMR instrument manufacturers and are available from the NMRFAM web site (<http://www.nmrfam.wisc.edu>) and the BioMagResBank (BMRB) web site (http://www.bmrb.wisc.edu/tools/choose_pulse_info.php). Reduced dimensionality versions of these pulse sequences have also been developed and should yield further reductions in the time required for data acquisition51, 118, 119. Acquisition of all spectra on a single sample of doubly labeled [U-15N, U-13C] protein is preferred, since it minimizes sample-dependent variations between NMR experiments.

TABLE 5.1. Typical Series of NMR Experiments Collected for a Full Structure Determination

images

In cases where the protein of interest is a homo-/heterodimer or protein-ligand complex, we prepare a differentially labeled sample by mixing equal amounts of unlabeled protein and uniformly 15N/13C labeled protein. This sample allows us to obtain 3D NOESY data that are F1-filtered and F3-edited for either 15N or 13C. The resulting spectra will only contain NOEs that arise between a proton directly bound to an NMR active heteronucleus and a proton bound to any NMR inactive nucleus while suppressing all other cross-peaks by isotope filtering120. These experiments are essential for correctly defining the dimer interfaces.

Processing and Analyzing NMR Datasets: The Stepwise Approach

The usual approach in a biomolecular NMR study is to first convert time-domain data to frequency-domain spectra by Fourier transformation. Then peaks are picked from each spectrum and analyzed. Methods have been developed for automated peak picking or global analysis of spectra to yield models consisting of peaks with known intensity, frequency, phase, and decay rate or linewidth (in each dimension)121, 122. Our current protocols for processing, peak picking, and assignment of NMR spectra primarily use the programs NMRPipe (Fourier transformation)123, XEASY (peak picking and semiautomated assignment)124, SPSCAN (automated peak picking; available at http://www.personal.uni-jena.de/—b1glra/spscan/manual/index.html.), and GARANT (fully automated assignment)125 on Linux and Mac OS X computing platforms. More recently, for automated backbone and side chain assignments, we have been using PINE (http://miranda.nmrfam.wisc.edu/PINE/. The iterative process of NOE assignment and structure calculations relies primarily on XEASY and CYANA (automated NOE assignment and torsion angle dynamics calculations)126, 127.

Convert and Process Raw Data. Before processing, time-domain data files acquired on Bruker spectrometers must be converted to the proper format for NMRPipe123. This is accomplished using the bruk2pipe program supplied with NMRPipe. The corresponding bruk 2 pipe and NMR Pipe scripts for each experiment type are stored in the database at the time of the initial setup. Postconversion to XEASY format is also easily performed with the pipe2xeasy program, available for download from http://www.personal.uni-jena.de/—b1glra/spscan/download/.

Peak Picking of All Spectra. Standard methods for obtaining chemical shift assignments begin with triple-resonance experiments that correlate various combinations of backbone and side chain 13C signals with the amide N and 1H signals. A 2D 1H-15N HSQC serves as the initial reference spectrum for directing the identification of signals in the 3D HNCO spectrum, which has the highest resolution and sensitivity of all the triple-resonance experiments. 2D HSQC peaks are picked and then transferred to the 3D HNCO using the strip plot features of XEASY. This array of HNCO strips is manually inspected for completeness and to eliminate spurious noise or artifact peaks. Peak picking of all other 1HN — correlated 3D spectra (HN(CA)CO, HNCA, HN(CO)CA, C(CO)NH, CBCA (CO)NH, etc.) is performed by SPSCAN in an automated fashion with the HNCO peaklist as a reference by execution of a single input script (available from BMRB). After inspection in XEASY and acceptance of the final peak lists, the data tabulated from 3D triple-resonance spectra are ready for analysis by automated methods for obtaining sequence-specific chemical shift assignments.

Sequence-Specific Assignments

A number of automated assignment strategies have been described in the literature. These methods have been implemented in programs like AUTOASSIGN128, CONTRAST129, GARANT125, and others. GARANT, which has been used extensively at CESG (Center for Eukaryotic Structural Genomics), shares the same file formats with XEASY, SPSCAN, and CYANA, simplifying the data pathway by eliminating file conversion problems, and it has the ability to provide partial side chain assignments, along with backbone assignments. PINE, a new system developed at NMRFAM (Bahrami et al., manuscript in preparation) (discussed below) accepts a flexible range of experimental peaklists in XEAS Y and other formats and provides backbone assignments via a web-based server (http://pine.nmrfam.wisc.edu/).

Depending on the quality of the NMR spectra and the resulting peaklists, chemical shift assignments generated by automated software are typically ~75-90% accurate and complete. Manual correction and verification of chemical shift assignments to a level of at least 90% completeness is required before proceeding to automated structure determination130. Side chain 1H and 13C chemical shift assignments are completed using the 3D HBHA(CO)NH, H(CCO)NH, C(CO)NH, and HCCH-TOCSY spectra (for aliphatic side chains) and the 13C — edited 3D (aromatic) spectrum for aromatic side chains. A list of verified chemical shift assignments serves as the input for TALOS, a program that generates backbone images and ψ dihedral angle constraints from a combination of secondary 1H, 15N, and 13C shifts and pattern matching to a database of tripeptide conformations from experimental structures for which chemical shift values are also known19. Tabular chemical shift data in the TALOS input format is easily generated using the CYANA macro “taloslist.” We have streamlined this process with two scripts. One converts chemical shift assignments into the proper format and executes TALOS. After inspection of images/ψ predictions in the graphical interface provided with TALOS, the table of results is converted into a CYANA-readable list of dihedral angle constraints with a second script (available from BMRB).

NOE Assignment and Structure Calculation

Semiautomated assignment of NOEs using simple chemical shift filters to suggest possible assignment combinations is an error-prone iterative process incompatible with a high-throughput structural proteomics pipeline. The NOEASSIGN module of the torsion angle dynamics (TAD) program CYANA provides a fully automated approach for the iterative assignment of NOEs126, 127, 131 that has performed robustly in our experience. Over the past 3 years, we have used NOEASSIGN to solve 14 unique structures by providing unassigned NOESY peak lists, dihedral angle constraints generated by TALOS19, and a chemical shift list as input102, 107, 132, 133, 134, 135, 136, 137, 138, 139. NOEASSIGN uses the concepts of network anchoring and constraint combination126, 131, 140 in performing seven rounds of iterative NOE assignment with the result of the previous round used as a structural model (Figure 5.3). This automated routine enables the user to go from unassigned peak lists to folded structure in as little as 10-20 min, depending on the size of the protein and the available computational power (a 16 node/32 processor cluster was used here). NOESY peak assignments from NOEASSIGN are then verified or corrected manually using XEASY, and refined with additional CYANA calculations to generate a final structural model that meets basic acceptance criteria for agreement with experimental constraints, Ramachandran statistics, and coordinate precision.

Resolving Symmetric Dimers. A significant proportion of structural proteomics targets form stable oligomers in solution. Dimers and other oligomers with total molecular weight less than 30 kDa are amenable to high-through put NMR methods, and this highlights the need for an automated approach to solving structures of symmetric assemblies. CYANA supports automated structure calculations on symmetric homodimers and higher order oligomers. Dimeric proteins do not typically present a special problem for X-ray crystallography, but they offer a challenge for NMR methods because inter- and intramolecular NOEs are indistinguishable in uniformly 15N/13C-labeled samples due to chemical shift degeneracy between symmetry-related subunits. Thus, NOEs that define the dimer interface are not readily apparent in standard isotope-edited NOESY spectra. To identify intermolecular NOEs at the dimer interface, we prepare a differentially labeled sample by mixing equal amounts of unlabeled protein and uniformly 15N/13C-labeled protein and acquire a 3D F1-13C/15N-filtered, F3-13C-edited NOESY spectrum120. This spectrumcontains onlyNOEsarising between a proton directly bound to 13C nucleus and a proton bound to any NMR inactive nucleus while suppressing all other cross-peaks by isotope filtering. Constraints generated from the filtered NOESY spectrum can therefore be unambiguously assigned to protons on opposing faces of the dimmer interface.

Figure 5.3. Automated structure calculations using the NOEASSIGN module of CYANA. The structure of the second PDZ domain (residues 450–558) from Par-3 (PDZ2) was determined using the NOEASSIGN module of CYANA. Six files, including the sequence file, a chemical shift list, three unassigned NOESY peak lists with intensity values, and a list of dihedral angle constraints from TALOS were provided as input. The structural model in cycle 1 is generated based on chemical shift mapping, network-anchoring, and constraint combination. Continual improvements in the structural model are observed in the progression from cycle 1 to the final calculation as NOE assignments are verified and filtered against the previous structural model starting in cycle 2. The final structural model was defined by a total of 1,487NOE (759 short range, 199mediumrange, and 487 long range) and 117 dihedral angle constraints, and was completed in ~10 min using a 32 node cluster.

images

In theory, CYANA should be able to resolve the correct monomeric fold and dimmer interface from a combination of isotope edited and filtered NOE peaks when the MOLECULES DEFINE command is utilized in an NOEASSIGN run. MOLECULES DEFINE specifies the residue range for each monomer and works in conjunction with the MOLECULES IDENTITY and MOLECULES SYMDIST commands to minimize differences between torsion angles and distances, respectively, between corresponding atoms in the two monomers. Calculation of a symmetric homodimer requires the user to provide a sequence and chemical shift lists that contain explicit values for both subunits but only a single set of (nonduplicated) peak lists. Duplication of the symmetry-related NOESY distance constraints is performed automatically in NOEASSIGN when the MOLECULES DEFINE command is implemented. In our experience, NOEASSIGN fails to correctly define a homodimer unless ~20–30 intermolecular NOEs are manually assigned and maintained during the course of the iterative structure refinement. For example, 26 preassigned intermolecular NOEs were required for NOEASSIGN to converge to the correct structure forMZF1136. Once a structural model is obtained from NOEASSIGN, it can be used to seed a second NOEAS-SIGN run with all peak assignments reset at the beginning of the calculation. The second NOEASSIGN calculation typically generates more NOE constraints and a structural ensemble with lower root mean squared deviation (RMSD) and target function values when compared with the initial calculation. This ensemble is usually a suitable starting point for manual refinement in CYANA, final refinement in explicit water, or deposition in the PDB.

NMR Molecular Replacement Using CYANA/NOEASSIGN. As structural proteomics efforts increase the number of known folds and their representation in sequence space, the likelihood of finding a structural homologue for a particular protein increases. A 3D model of the target protein of interest may be generated from the structure of a likely homologue using MODELLER141 or other homology modeling programs. Inclusion of a homology model as a reference structure for the NOEASSIGN algorithm often improves the quality of the final model when compared with the results of a de novo calculation. In the early stages of iterative NOE assignment, knowledge of the global fold can provide a valuable distance filter that allows the selection of many unique distance constraints that would otherwise be treated ambiguously due to chemical shift degeneracy. We have applied this method to monomeric proteins with outstanding results, but the impact is even greater when applied to symmetric homodimers.

To solve the NMR structure of the homodimeric SCAN domain from ZNF24, we constructed a ZNF24 model (Figure 5.4a) based on the structure of tumor suppressor protein myeloid zinc finger 1 (MZF1; PDB ID 2FI2)136. The model was included as a reference structure in an NOEASSIGN calculation performed on completely unassigned NOESY peaklists. The ensembles generated by NOEASSIGN converged to a domain swapped dimeric assembly (Figure 5.4b) similar to that observed for other members of the SCAN family136, 142. Consequently, the ZNF24 dimer structure was solved in much less time than the MZF1 structure.

Protein crystal structures solved by molecular replacement methods are susceptible to systematic error due to model bias, and it is important to assess the extent to which NMR refinement may be biased by areference structure. In contrast to crystallography, NOE-based refinement appears relatively insensitive to and unconstrained by the orientation of different structural elements in a seed model. Model-based filtering of NOE assignments relies on coarse estimates of interatomic distances for the initial round of refinement only, and is insensitive to long-range differences between the reference model and the true structure of the target protein. Consequently, the NMR analogue of molecular replacement may be more robust than the crystallographic version. To illustrate this point, we compared the NMR structure of the ZNF24 homodimer calculated using NOEASSIGN with the reference model created with MODELLER from the MZF1 structure (Figure 5.4d) and with an X-ray crystal structure of ZNF24 (Figure 5.4e). The reference model and the structure produced by NOEASSIGN exhibited clear differences in the overall arrangement of structural elements, suggesting that this method does not introduce a significant structural bias. Moreover, the NMR structure agrees very closely with the X-ray structure (Figure 5.4c), which was solved independently using phases obtained from single-wavelength anomalous dispersion, providing additional evidence that the automated NOEASSIGN algorithm converged on the correct structure for ZNF24.

Figure 5.4. Automated structure calculation of the ZNF24 SCAN homodimer using a seed model. (a) Seed model (green) of ZNF24 constructed using MODELLER software and theMZF1SCAN domain structure (PDB 2FI2). (b) NMR structure (magenta) of ZNF24 generated using the NOEASSIGN module of CYANA and the ZNF24 seed model shown in panel a. The NMR structure is defined by 2595 NOEs and 134 dihedral angles per monomer. (c) X-ray structure (cyan) of ZNF24 solved by single-wavelength anomalous diffraction using a mercury derivative. (d) Overlay of the ZNF24 model (green) and the NMR structure determined using the NOEASSIGN module of CYANA. (e) Comparison of the NMR and the X-ray structures. (f) Side view of all three models showing the change in the angle between helix 5 from each monomer as the seed model of ZNF24 is refined by NOEASSIGN to fit the experimental data. The more than 40° difference in the helix 5 angle between the seed model and the finalNMRstructure of ZNF24 suggests that the seed model does not bias the structure calculation. This is supported by the close agreement in helix angle when the NMR and Xray structures are compared. The figure also appears in the Color Figure section.

images

Final Refinement

The final stages of refinement often require inspection of NOE assignments that result in consistently violated distance constraints. NOEs that have overestimated intensities due to peak overlap can produce too restrictive constraints and require manual adjustment. A fully automated refinement scheme may be able to resolve many such conflicts, but manual intervention is typically required at some point to achieve the desired goals for precision (backbone RMSD below ~0.6Å), agreement with experimental constraints (none violated by more than 0.5 or 5Å), and good torsion angle geometry (no residues in disallowed regions of images/ψ space). CYANA provides diagnostic output with concise summaries of all the necessary data for each ensemble of structures calculated, often simplifying the search for problematic restraints or assignments.

Average constraint density is an informative parameter that typically correlates with coordinate precision. NMR structures defined by more than 15-20 nontrivial constraints/ residue typically exhibit high coordinate precision, with RMSD values of 0.5-0.7 A for backbone atoms and 1.0-1.2 for all nonhydrogen atoms. However, it is also important to assess the local mobility of the polypeptide backbone using 15N relaxation measurements, since flexible regions will often be devoid of long-range NOEs. For example, 1H-15N heteronuclear NOE values show that most residues of a PDZ domain are ordered but an internal loop and both termini are flexible on picosecond-nanosecond timescales (Figure 5.5a). Consequently, the local NOE constraint density is expected to be low (Figure 5.5b), and no extra effort should be made to identify long-range NOE constraints for those residues. Automated NOE assignment algorithms may incorrectly identify long-range distance constraints for flexible residues, and these assignments can be rejected based on the evidence of local disorder provided by 15N relaxation measurements. Atomic coordinate precision, as reflected in local RMSD values (Figure 5.5c), mirrors the local constraint density, and this local flexibility is illustrated by the structural ensemble (Figure 5.5d).

When a complete and consistent set of experimental constraints has been obtained, and the resulting structural ensemble meets the basic acceptance criteria, a final refinement calculation is performed in explicit water with physical force field parameters. A method for Cartesian space molecular dynamics in explicit water using XPLOR-NIH was shown previously to improve the stereochemical quality of NMR structural ensembles generated with simple force fields consisting only of terms for covalent geometry, experimental constraints, and a simple repulsive potential143. This conclusion was reinforced through the recalculation of structures for over 500 proteins from standardized sets of restraints144. Improvements in a variety of validation criteria were observed upon water refinement of NMR ensembles, particularly for the percentage of residues in the most favored region of the Ramachandran (images /ψ) plot. Systematically poor Ramachandran statistics for NMR structures calculated with simplistic force fields can be attributed to the unrealistic constraining of peptide bond planarity (w angle). In CYANA, for example, ω is fixed at 0° or 180°, corresponding to the cis and trans peptide bond isomers. However, examination of high-resolution crystal structures in the PDB revealed a strong correlation between the standard deviation of w and the proportion of residues in the most favored images/ψ regions. Consequently, relaxation of peptide bond rigidity during water refinement in Cartesian space is accompanied by changes in images and ψ, which improve overall quality of the backbone conformation.

To simplify the water refinement routine, we have developed a script (available from BMRB) that converts the output of a CYANA calculation into input for water refinement and initiates the water refinement routine that utilizes a variety of software packages including XPLOR-NIH, CNS-SOLVE, PROFIT, PROCHECK-NMR, WHAT IF, and AQUA. While this routine is not fully integrated into CYANA calculations, the user can initiate a water refinement run with this single script once a complete and consistent set of experimental constraints has been achieved. Starting from the CYANA calculation directory, the script first creates a validation directory with a standardized structure and copies the required output files from CYANA to the validation directory. The CYANA files are then modified to ensure that residue naming conventions and numbering are compatible with the water refinement routine. Specialized input files that specify disulfide bonds, metal centers or cis-peptide bonds are also copied to the appropriate location during the initial setup. The script then initiates the validation run by parsing the input data and converting the distance and dihedral angle constraints into the XPLOR format. The initial structural model from CYANA is analyzed against the distance and dihedral angle constraints using WHAT IF and PRO-CHECK-NMR. The structural model is then refined using restrained molecular dynamics in explicit water using XPLOR-NIH as described above. Finally, the refined structural model is verified against the distance and dihedral angle constraints as previously described. A comparison of the input and refinement statistics shows that the Ramachandran statistics of the torsion angle dynamics ensemble from CYANA and the WHAT IF RMS Z scores are systematically improved by refinement in explicit water (Table 5.2).

Figure 5.5. Refinement and validation of structural models. (a) Heteronuclear 15N-1 H NOE values for the second PDZ domain from Par-3. NOE values below 0.7 are shown in light gray. (b) NOE distance constraint density plotted by residue number. Distance constraints used in the final structure calculation were derived from 3000 assigned peaks observed in three 3D NOESY spectra. Elimination of redundant NOEs and meaningless constraints (distances that impose no conformational restriction or are held fixed by covalent geometry) yielded a final set of 1222 NOE distance constraints. Including dihedral angle constraints from TALOS, a total of 1339 NOE and dihedral angle constraints were used in the structure calculation corresponding to ~12 restraints per residue. However, if residues with 15N- 1H NOE values below 0.7 are excluded from consideration, the restraint density increases to ~17 restraints per residue. Structures defined by <10, 10-15, and >15 constraints per residue typically correspond to low, medium, and high-quality structures, respectively. (c)The root mean squared deviation for the structural ensembles before (green line) and after (orange line) water refinement in explicit water plotted as a function of residue number. (d) Structural ensembles produced in CYANA using NOEASSIGN before (left panel) and after water refinement (right panel) show differences in conformational sampling for flexible loop regions as reflected by increased RMSD values in panel c. The helix, strand and loop elements in each ensemble are colored cyan, magenta, and salmon, respectively. The figure also appears in the Color Figure section.

images

Validation and Assessment of Structural Models

The optimal validation approach to be taken depends on the experimental data collected and how they were used in deriving the family of conformers that represents the structure. Two approaches are used for the validation of biomolecular NMR structures. In the first, the agreement between experimental data and coordinates is assessed. The second measures the match between many aspects of the geometry and the expected range of permissible values for each geometric parameter. Ranges of expected values can be derived theoretically or can be obtained from databases of protein structures (such as PDB), nucleic acid structures (such as NDB), and small molecules (such as Cambridge Structural Database, CSD). These types of structure validation provide complementary checks on the quality of structures and are described below along with various existing software packages.

TABLE 5.2. Structure Statistics for PDZ 2 of Mouse Par-3 (Residues 450-558)

images

As with X-ray structures, the stereochemical quality of protein models can be checked with the software tools PROCHECK145 and WHAT IF146. NMR specific tools include AQUA and PROCHECK-NMR147. Recently, WHAT IF has been extended to include a check on hydrogen geometry148. Structural studies on peptides149 and nucleic acids150 solved at high resolution and present in the CSD resulted in a set of reference values and standard deviations of bond lengths and angles for heavy atoms that are taken as reference values for larger biomolecules. In most common force fields, the force constants for geometry constraints are chosen to reflect the variation in these geometric parameters. The check on bond lengths and angles is one among the various checks that should be performed before accepting a refined model. Severe deviations from the commonly accepted standard geometry signal the problematic regions in a structure.

Software packages used to solve and refine NMR structures (X-PLOR, CNS, DYANA, AMBER, and others) provide a rich set of diagnostic tools and a summary of the violations between the experimental data and the coordinates. Structures uploaded at deposition through the ADIT-NMR system can be checked using PROCHECK145 and MAXit151. A number of additional software tools are available for evaluating the quality of derived NMR structures. Many of these can be downloaded from the World Wide Web or are available as web servers. These include Verify3D152, 153, ProSA154, 155, MolProbity156, 157, 158, VADAR159. WHAT IF (Vriend,) and AQUA147, 148. In general, NMR structures of acceptable quality meet the following validation criteria: (1) an RMSD for NOE violations less than ~0.05Å and no persistent NOE violation greater than 0.5Å across the ensemble of structures160; (2) an NOE completeness of at least 50% for all NMR observable proton contacts within 4Å161; (3) all torsion angles within the range of the restraints; and (4) chemical shift values within acceptable ranges, unless verified independently. At the time the structures are released, full documentation of the structure and the experimental data, as described in the IUPAC publication “Recommendations for the Presentation of NMR Structures of Proteins and Nucleic Acids”162, should be provided, as well as the reports generated by the software packages mentioned above.

A number of approaches to validate NMR structures against NOE restraints or peak volumes and against residual dipolar couplings have been proposed163. Kalbitzer and coworkers have developed an improved software package for back-calculating NOESY spectra from structures that take into account relaxation and scalar coupling effects; this approach and an associated “R” factor offer a promising way of validating NMR structures164, 165. Brunger and coworkers have developed a complete cross-validation technique, in analogy to the free R-factor used in X-ray crystallography, which provides an independent quality assessment of NMR structures based on the NOE violations166. The free R-factors of different NMR structures are comparable only if the NOE intensities have been translated into distance restraints in exactly the same way. Montelione and coworkers have proposed a combined scoring system for assessing the quality of NMR structures167. The groups of Clore168 and Bax115 have developed a quality factor for residual dipolar couplings. In cases where dipolar coupling data can be obtained easily, this provides an attractive approach. A chemical shift validation quality factor also has been proposed169.

Validation of structural models against assigned chemical shifts is another promising technique for checking structures. The measurement of chemical shifts is easy and precise but the back-calculation of the expected chemical shift from a structure is less trivial170. Williamson and coworkers have related measured proton chemical shifts of proteins to values calculated on the basis of NMR and X-ray structures171. The chemical shifts derived from NMR structures showed the same degree of agreement with the measured proton chemical shifts as X-ray structures with a resolution between 2 and 3 A(o of 0.35 ppm). Methods for back-calculating chemical shifts from structure are improving rapidly172, 173, 174, 175. In addition to the AVS software tools176, BMRB uses the linear analysis of chemical shift (LACS)177 software application and chemical shift back-calculation routines in NMRPipe (F. Delaglio, unpublished) to validate assigned chemical shift depositions.

Deposition of Completed NMR Structures

Deposition of coordinates, chemical shifts, and structural constraints in public databases requires extensive data formatting and the entry of multiple kinds of information. In an effort to accelerate the deposition of new structures and improve the informational content of the entries, the PDB developed the AutoDep Input Tool (ADIT), a tool that accepts formatted input files using an expanded form of the macromolecular Crystallographic Information File (mmCIF) dictionary178, 179 or files in the PDB format. An mmCIF contains all the required structural data and experimental information needed to complete the deposition and automatically populates the query fields when parsed using ADIT. The BMRB in collaboration with the PDB used this system in developing a one-stop deposition system based on the NMR-STAR dictionary (ADIT-NMR; http://deposit.bmrb.wisc.edu/bmrb-adit/). This system is now available for submitting NMR-derived atomic coordinates, restraints, and all other experimental data to the PDB and BMRB. This interface, which can be accessed from the Research Collaboratory for Structural Bioinfor-matics (RCSB), Protein Data Bank (PDB), PDBj, or BMRB, minimizes the time required to complete the deposition to both data banks by allowing the user to upload and answer all queries once. The PDB takes responsibility for annotating the coordinates, and the BMRB takes responsibility for annotating the NMR data and restraints. The EBI is developing a new deposition system, but currently uses the older AutoDep system for structure depositions.

Generation of the mmCIF file can be conducted in either a semiautomated or manual fashion using the PDB_extract software available from the PDB website180. This software extracts critical structural and experimental information from the output and log files generated by the programs commonly used in NMR and X-ray structure determination and combines it with the structural coordinates to generate a single file for upload. The resulting file can then be verified and/or edited manually to ensure that all information has been captured correctly prior to upload through the ADIT-NMR system. Alternatively, a template file generated by PDB_extract for NMR structures can be edited manually to include all the relevant information for the deposition of a particular target and passed to PDB_extract with the structural coordinates to assemble the mmCIF required for deposition. Our protocol calls for deposition of completed structures using a manually edited template file in conjunction with a script (available from BMRB) that automates the preparation of the final structural ensemble, distance and dihedral angle constraints, and chemical shift list for deposition. The script generates a single file containing both the distance and the dihedral angle restraints and runs PDB_extract to generate an mmCIF that can be uploaded to the PDB and BMRB data banks using the ADIT-NMR interface after a final review for accuracy. After the mm CIF, constraints file, and chemical shift list are uploaded to ADIT-NMR, the depositor must answer only few queries to complete the deposition to both databases. In our experience, this approach is considerably faster and more reliable than performing each structure deposition by completing all fields of the ADIT-NMR form via the web interface.

Since the 1980s, the field of automated NMR analysis has been making strides toward a sound algorithmic foundation. A great deal of thought has gone into modeling the “analysis pipeline” (Figure 5.6). The principal aim of automated analysis is to ensure that the performance is satisfactory, or nearly optimal, even when there is uncertainty about the underlying data. The general approach has been to assume that the protein structure can be described as a bundle of conformers that satisfy an empirical force field and structural constraints. Variation in the conformers arises from uncertainty in the data, with large variations possibly indicating errors in prior steps in the pipeline (such as peak assignments orambiguity inNOEassignments) thatcouldbe “corrected” so as to lead to atighterbundle ofconformers. Overtime, itbecame clearthatsuchan approach can lead to overconstrained bundles. The reason is that in this model of uncertainty, every resonance assignment, experimentally derived restraint, bundle of structures, or protein residue in the sphere of uncertainty is deemed to be equally likely at each stage, and the decision is reduced to a binary accept/reject of observed data for use in the subsequent stage. Under such conditions, and combined with the incompleteness of the NMR experimental data, the algorithm is therefore obliged to guarantee equally satisfactory performance for every residue of every bundle within this sphere of uncertainty, irrespective of the level of uncertainty that may exist within the context of a specific NMR experiment and the specific protein under consideration.

Figure 5.6. The standard pipeline for NMR structure determination consists of a number of discrete tools that, in reality, are intertwined by iterative cycles involving the detection and correction of errors. The front-end of NMR analysis is mostly concerned with “labeling of frequencies” derived from spectra while the back-end focuses on “spatial labeling.” Normally, the front-end and back-end are part of separate analyses; but ideally they should become part of a single analytic process.

images

Inherent uncertainties and incompleteness in experimental data as well as errors in modeling approximations (e.g., force field definitions) can lead to errors in attaining the correct fold or bundle of conformers. In short, NMR structure determination is a noisy, inverse, and ill-posed problem. In automating protein structure determination, it is particularly essential to carefully consider issues arising from uncertainties, including ways to explicitly measure and report them, because their impact can be remarkably difficult to detect otherwise. Errors can be broadly classified as either random or systematic. Random errors, often referred to as noise, lead to increased variance in results. In the language of robust analysis, noise is generally thought to have its largest impact on the reliability of results, but not necessarily on their accuracy. Systematic errors are usually more difficult to detect by computational methods, and they typically affect the accuracy of the results. To deal with sources of error and to address challenges faced in the traditional design of the pipeline, a recent approach has been adopted to include some notion of weight to the uncertainty in the databy assigning aprobability to results from each discrete step and to each subset of data. Thus, instead of assuring equally satisfactory performance at every possible residue or for every nucleus, the aim becomes to maximize the expected value of the performance of the algorithm. With this reformulation, there is reason to believe that the resulting model structures will better reflect the uncertainties in data than those based on deterministic uncertainty models.

The prospects for carrying out a fully probabilistic, automated program for NMR structure determination have been bolstered by advances in theoretical areas, such as algorithms that exploit relationships between local statistical potentials and global characteristics. It is now understood that global phenomena that obey a certain (large) family of probability distributions can often be described, or reasonably approximated, in terms of local potentials. Conversely, a set of local potentials could lead to a global description given by the same (large family) of probability distributions. These linkages are based on the Hammersly-Clifford theorem181, 182, 183, which implies that the efficacy of approximations given by statistical potentials is intimately tied to the accuracy of the description of the local potential. To understand the behavior of a certain global model, its sensitivity to estimates, and the reliability and accuracy of solutions, we can use an approach that combines elements of statistical mechanics and statistical decision theory while relying on local potential descriptions184, 185, 186, which seems promising.

In the following sections, we first discuss how probabilistic approaches to discrete steps in the NMR structure determination pipeline (data collection, peak assignments, secondary structure determination, and outlier detection) can achieve remarkably effective outcomes. Then, we demonstrate how iterative combination of information from these discrete steps can lead to increased reliability and improved accuracy. The final section discusses how probabilistic tools can be used to address the ill-posed, inverse problem of NMR structure determination.

Adaptive, Probabilistic Data Collection

One of the challenges in protein NMR spectroscopy is to minimize the time required for multidimensional data collection. An innovative approach to this problem, “high-resolution iterative frequency identification (HIFI) for NMR” (HIFI-NMR)51, combines the reduced dimensionality approach for collecting n-dimensional data as a set of tilted, two-dimensional (2D) planes with features that take advantage of the probabilistic paradigm. For 3D triple-resonance experiments of the kind used to assign protein backbone and side chain resonances, the probabilistic algorithm used by HIFI-NMR automatically extracts the positions (chemical shifts) of peaks with considerable time saving compared to conventional approaches in data collection, processing, and peak picking.

The HIFI-NMR algorithm starts with the collection of two-orthogonal 2D planes. In cases where1H-15N and1H-13C 2D planes are common to multiple experiments, they need only be collected once. The algorithm then moves to the list of experiments specified for data collection and starts with the first experiment. For this experiment, tilted planes are selected adaptively, one-by-one, by use of probabilistic predictions. A robust statistical algorithm extracts information from each plane, and results are incorporated into the online algorithm that maintains a probabilistic representation of peak positions in 3D space. The updated probabilistic representation of the peak positions is used to derive the optimal angle for the next plane to be collected. The online software predicts whether the collection of this additional data plane will improve the model of peak positions. If the answer is “yes,” the next plane is collected and the results incorporated into the evolving model; if the answer is “no,” data collection is terminated for the experiment. The algorithm then proceeds to next NMR experiment in the list to be done. Once all experiments have been run, the peak lists with associated probabilities generated directly for each experiment, without total reconstruction of the three-dimensional (3D) spectrum, are ready for use in subsequent assignment or structure determination steps.

In achieving the steps outlined above, the probabilistic algorithm in HIFI-NMR divides the 3D spectral space into discrete voxels on the basis of prior information: the number of expected peaks; the nuclei involved and their expected chemical shift ranges; and the data collection parameters. Next, an empirical estimate for the prior probability distribution of peak locations in the multidimensional spectral space is obtained. The resulting joint probability distribution estimated in this step signifies the probability of a peak’s existence in a specific voxel of the spectrum. Let Ppr(X) denote the prior probability for the presence of apeak at position X, where the elements of the vector X =(x1, x2, x3) represent the voxels in three-dimensional space. The initial step in the HIFI approach is to collect data for the two orthogonal planes (0o and 90o planes). For experiments involving 1H, 15N, and 13C dimensions (e.g., HNCO, HNCACB), the 0° (13C - 1H), and the 90° (15N - 1H) planes are collected and peak-picked. Because the 1H dimension is common to the two planes, the possible candidates for peaks in 3D space can be generated by considering all combinatorial possibilities for peaks in the two planes that have the same 1H chemical shift within a given tolerance. This generates many more candidates than actual peaks, but will include most, not all, real peaks. These candidate peaks restrain the prior probability space generated in the earlier step. At the end of this stage, P0(x) designates the new probability distribution that includes information gained about the probability space consisting of candidate peaks.

The next best tilted plane is determined by assuming that the candidate peaks have been observed on a plane at angle θ and by evaluating the information theoretic impact of this observation on the probability of the peaks. The measure of this influence is the divergence Sθbetween the initial probability P0(x) (probability at the initial or the 0th step) and the predicted probability after the impact of the projection at angle θ has been taken into account Qθ(x):

(5.1)images

where the sum is taken over all voxels in the spectral space.

Next, each candidate peakx is examined in turn in the light of the data from the new tilted plane and prior data, and Bayes’ rule is used to update its probability P0(x) of being a “real peak,”

(5.2)images

or

(5.3)images

in which x_not_observed_in_θ is the complement event for x_observed_in_θ. The term in the denominator, P(x_observed_in_θ)orP(x_not_observed_in_θ) represents the probability of a voxel x in plane θ to be considered a peak (or not a peak), regardless of any other consideration. The probabilities P(x_observed_in_θ| x) and P(x_not_observed_in_θ| x)do not experience much angular variation and can be reasonably approximated by the empirical data. This process is continued until no significant information is obtained further. At this stage, n peaks with probabilities above a threshold are reported as the peak list. The value of n is estimated automatically from the number of amino acid residues in the protein, the type of experiment, and the probability distribution P(x). The value of n can be adjusted in a subsequent post processing step to include additional expert information regarding the data observed.

Probabilistic approaches are useful in a variety of situations where the robustness of analysis results needs to be validated. This includes the experimental determination of NMR coupling constants. HIFI-C187 represents the extension HIFI-NMR’s adaptive and intelligent data collection approach to the rapid and robust collection of coupling data. In this case, data collected from one or more optimally tilted 2D planes are used to determine couplings with high resolution and precision in less time than required by a 3D experiment. A further benefit of the approach is that data from independent planes provide a statistical measure of reliability for each measured coupling. Fast data collection can enable measurements in cases where sample stability is a limiting factor, as frequently happens with proteins in the orienting media used for determining reduced dipolar coupling data.

Probabilistic Approach to NMR Peak Assignment

Resonance assignment programs can be categorized by the methods they use in the mapping steps. These methods include stochastic approaches, such as simulated annealing/Monte Carlo algorithms188, 189, 190, genetic algorithms191, exhaustive search algorithms192, 193, 194, 195, heuristic comparison-based algorithms that predict chemical shifts derived from homologous proteins196, and heuristic best-first algorithms197, 198, 199. For example, AutoAssign200 is a constraint-based expert system that uses a heuristic best-first mapping algorithm. Among these algorithms, those that rely on stochastic approaches have the best potential for satisfactorily addressing issues arising from noise.

In the PISTACHIO assignment algorithm201, the cost function is recast as a measure of “system energy” and the optimization problem is restated in physical terms as that of finding a “ground state.” This restatement enables one to ask questions that turn out to have tractable solutions with practical applications. For example, the approach makes it possible to ask for a “typical low energy configuration” and to find a computationally feasible (polynomial time) solution that gives a great deal of information about the “ground state.” This is pertinent to the NMR assignment problem because the available data may not support a unique set of assignments. An additional advantage to this approach is that it yields a set of assignment configurations with associated probabilities that can be explored and refined further as additional data becomes available. The local state of the statistical system is constructed in three steps. First, the peak list data are parsed into the set of all possible tripeptide spin systems and their associated scores. Next, tripeptides are assembled to match the sequence and to achieve the optimal probabilities for correct assignments. Each of the three steps serves to restrict the size of the combinatorial search space and to minimize the impact of noise on the analysis. The impact of noise is further ameliorated because the global ground state of the model corresponds to a state that is “statistically closest” to reported observations.

The starting point is to define a “local cost model” J(Y) in terms of a set of cost functions JN and their configurations YN that represent the spin system for a given tripeptide:

images

To minimize the overall cost J, one can maximize the following expression:

images

To treat this as a probability, a normalization factor is introduced,

images

where the cost J is stated on the basis of pair-wise matching of tripeptides:

(5.4)images

Upon dividing by the normalization factor to obtain probabilities, and by denoting the set of all permutations as S, the following expression is obtained:

(5.5)images

In Eq. 5.5, the second term measures the merit of assigning a tripeptide spin system Φ to a given tripeptide, whereas in the first term, ψ, provides a measure of the compatibility of the spin systems of adjacent, overlapped tripeptides. To achieve the goal of deriving probabilities for pairing spin systems with residues given by the Φ term, the marginal probability rather than the joint probability given by Eq. 5.5 must be calculated. One major practical advantage of this formulation is its generality because it does not rely on a specific experiment or set of experiments; rather, it enables an easy and efficient incorporation of data from multiple experiments into the framework. PISTACHIO is available as part of the PINE web server described below.

Probabilistic Secondary Structure Determination

Protein secondary structure plays an important role in classifying proteins202 and in analyzing their functional properties203. Several algorithms have been developed for identifying secondary structure from NMR chemical shifts alone; these include the ∆δ method204, the chemical shift index method205, the database approach by TALOS19, and the probability-based method206. A supervised machine-learning approachhas been used to combine data on sequence chemical shifts207.

PECAN (protein energetic conformational analysis from NMR chemical shifts) is a probabilistic approach that utilizes a local statistical potential to combine the separate predictive potential of the sequence and chemical shifts in determining protein secondary structure17. Probabilistic and energetic models share some common ground through the Boltzman-Gibbs distribution. PECAN derives a residue-specific statistical energy function that is used to derive probabilistic secondary structure determinations. Suppose that wn is the number of observations of state n and En is the energy of state n. Further, assume that wn is approximately proportional to the probability for state n. Then, for a system in equilibrium in one of N possible states, the energy and probability are related as

(5.6)images

where k is the Boltzmann constant, T is the temperature, Z is the normalization factor called the partition function, and p(En) is the probability of state En. The constant β sets a monotonic scale for units of energy and can be set equal to a convenient constant. The energy term E is given by

(5.7)images

Eq. 5.7 represents a finite model of length l, where the term Vi, represents the bias potential vector, which is determined by constructing optimized residue-specific density estimates that are related to energy through Eq. 5.6. Hj; j + 1 is a transition matrix for each state and represents the propensities of residues j and j + 1 to be in each of the possible state combinations. The values for the term Hj; j + 1 are derived from a database of experimental values. The two products Hj; j + j Vj and Hj— 1; j Vj represent two different stochastic mixings of the initial vector Vj. The parameter cj which can be replaced by a constant c, represents the proportional influence of the two terms and is optimized by reference to the database of experimental data.

Each site can be viewed as being in one of the three geometrically defined states: helix (H), strand (extended) (E), or “random coil” (R) (Figure 5.7). R is defined simply as neither H nor E. Each site has an associated energy that is coherent in the finite neighborhood l. For NMR chemical shifts of proteins collected under varying conditions of pH and temperature, PECAN yielded an average of 90% accuracy (93.5% helix, 83.5% strand, and 89.5% random coil) on a large data set (more than 85,000 residues). PECAN achieved a similar rate of success with a second set of data not previously considered (97 proteins; 12,000 residues). For 99% of the proteins, the identification accuracy exceeded 75%, and for 96% of the proteins, the accuracy exceeded 80%. PECAN is available from a web server at http://bija.nmrfam.wisc.edu/PECAN/ and also is incorporated into the PINE server described below.

Figure 5.7. Example of the use of the PECAN software to determine secondary structure from NMR chemical shifts and protein sequence17. The data are from BMRB entry 4083. The horizontal axis represents the sequence number. The vertical axis represents the probability of helix (gray), represented by positive values, and extended structure (black) represented by negative values. Values near zero indicate random coil residues. The white bars indicate the identification of regions without a discrete structural designation. Because a 3D structure of this protein is known (PDB 1CEY), the PECAN results can be compared with those derived from analysis of the protein conformation by DSSP208, as shown at the bottom of the plot in cartoon representation where strands are arrows and helices are waves.

images

Validation

One validation approach that we have found valuable takes advantage of the finding that, for a correctly referenced protein data set, linear regression plots of images versus images pass through the origin from two directions, the helix-to-coil and strand-to-coil directions (Figure 5.8). Thus, linear analysis of chemical shifts177 can be used to detect referencing errors and to recalibrate the 1H and 13C chemical shift scales, if needed, and to recognize outliers that may indicate a possible misassignment or abnormal backbone chemical shift. In the context of statistical potentials, the model of LACS assumes a trimodal Gaussian distribution. LACS is available from a web server at http://bija.nmrfam.wisc.edu/MANI-LACS/.

Figure 5.8. Example showing the analysis of assigned backbone 13Cαand 13Cβchemical shifts by the LACS algorithm177. Data shown are for a single protein (BMRB entry 4998). The left-hand line represents the statistically linear trend of chemical shift values in α-helical regions and the right-hand line represents those for residues in β-sheet (extended) regions of the protein. Residues in nonhelical, nonextended regions are represented by points near the intersection of the two lines. This probabilistic analysis is useful for detecting possible misreferencing (as indicated by the deviation of the crossing point of the two lines from (0,0) expected for properly referenced 13C data) and for correcting the data so that the chemical shifts better represent secondary structure. The analysis can also detect outlier values that may represent misassignments or, as in the case shown (histidine, residue 56, indicated by arrow), a residue whose backbone carbon chemical shifts are pH dependent and thus not a reliable indicator of secondary structure. The probabilistic detection of outliers is depicted in the inset, which shows the normal distribution of points about the right-hand line.

images

Integration of Probabilistic Steps within the Pipeline

The integration of individual probabilistic steps into a more global iterative analysis can be considered as the process of assigning a finite set of tags or labels to a collection of observations, subject to side conditions. This challenge is notable for its computational complexity and has theoretical and practical significance to a range of applications in bioinformatics. The combinatorial labeling paradigm is particularly problematic in modeling the early stages of NMR structure determination in which there is the need to associate amino acid labels, secondary structure labels, and other discrete attributes, to the observed data (NMR peaks with associated intensities and positions). The observed data often are incomplete and contain false positives and negatives. These characteristics render the robust assignment of labels remarkably challenging. The front-end labeling process associates one or more NMR parameters with a physical entity (e.g., nucleus, residue, tripeptide, helix, and chain); the back-end labeling process associates NMR parameters with constraints that define conformational states (e.g., dihedral angle, distance, and dipole orientation) (Figure 5.6).

One approach to carry out labeling in the presence of noise and incomplete data is cost function. For example, in the most straightforward case of using a chemical shift index to assign secondary structure, an entirely local three-state cost function was utilized15. Numerous cost functions of empirical or statistical nature have been proposed for problems ranging from assignment, secondary structure determination, dihedral angle restraints, to molecular dynamics force fields. Other approaches have used training algorithms, neural networks, or genetic algorithms. However, none of these approaches provides a principled answer to the questions of how many residues or resonances should be assigned, or how reliable and consistent such assignments need to be. To assign the best labels, the algorithmic process often optimizes additional parameters or introduces auxiliary criteria. As a result, it becomes necessary to introduce a variety of parameter setting rules, stopping rules, or other inspection methods. Often, validation tools are not utilized to test for systematic errors or outliers in the data, and no decision is tested against subsequent downstream decisions derived from experimental data (to ensure, for example, that the label from one step of data analysis is not in conflict with a label from another step).

A different view of labeling can be formulated within the context of information theory that explicitly implements the idea that NMR data analysis is motivated primarily by interest in some derived quantity (e.g., the protein structure) and that the solution strategy should focus on preserving information relevant to this goal, rather than trying to derive an optimal similarity metric in the local space of data. We imagine that each frequency label occurs together with a corresponding set of labels for the nuclei of a residue, along with additional labels indicating other information such as secondary structure state. In this formulation, there is no need to define a similarity measure; this measure arises from the iterative optimization principle itself. This approach is attractive because it frees us from the need to specify in advance what it means for the labels to be optimal. In addition, the basic idea can be extended with similar benefits to the entire NMR structure determination pipeline so as to identify in advance what it means for an NMR-derived structure to be most consistent with the data. Importantly, the approach enables the unbiased comparison of alternative algorithms for structure determination and analysis.

We have developed an algorithm called PINE that implements this approach to labeling and takes into account the interconnectedness of different stages of analysis (Figure 5.9). PINE begins with a set of local statistical potentials. It then proceeds iteratively until a stationary state for a consistent global similarity measure is achieved. The resulting software enables a seamless and robust integration of multiple steps in the NMR structure determination pipeline. PINE accepts combinations of NMR data specified by the user from an extensive list of NMR experiments, and provides as output a probabilistic assignment of backbone and side chain signals and the secondary structure of the protein. It also identifies, verifies, and rectifies problems related to referencing, assignment, or outlying data. PINE can make use of prior information supplied from selective labeling or spin system assignments derived independently by other means. PINE currently incorporates, and improves upon, the capabilities of PISTACIO, PECAN, and LACS. Extension of this core of algorithms can, in principle, support the additional statistical potentials involved in the back-end process of NMR structure determination.

Figure 5.9. Modeled relationships linking independent stages of evidence in protein NMR spectroscopy. The shaded ovals and solid lines indicate parts of the scheme currently implemented in the software package PINE (Bahrami, Markley, Assadi, and Eghbalnia, manuscript in preparation). PINE takes, as input, peak lists from different experiments, the protein sequence, information about labeling, and any prior assignments (used to override those determined). The analysis starts with an initial generation of an “almost exhaustive” list of spin systems in relation to the input data. The output currently gives backbone and side chain assignments and secondary structure. Future versions of PINE will work with longer range connectivities (from TOCSY and NOESY data) and, in addition, provide structural constraints (unshaded ovals and dotted lines).

images

The input to PINE consists of the amino acid sequence and peak lists from one or more specified NMR experiments; the peak lists can be either probabilistic (e.g., peak lists generated by HIFI-NMR) or traditional types generated by popular NMR analysis tools. Starting with the spectral alignment step, which consists of a global and a local alignment, a probabilistic model is invoked. This can be viewed as assigning a distance probability for a peak i to belong to a spin system (SSj) as follows:

(5.8)images

(5.9)images

where d(ij) is the distance of the peak i and the reference peak in spin system j, and C is a constant that can be set to the resolution of the spectra189. In the spin system generation, nuclei of the spin system may have multiple chemical shift candidates with associated probabilities. Those probabilities can be derived from the distance probabilities described above along with the probability of existence of the peak itself (e.g., from HIFI-NMR-generated peak lists).

(5.10)images

The total number of states of the spin system includes the combinatorial set of all nuclei choices, plus an extra state called null state. For the case of nonprobabilistic peak picking (traditional NMR experiments), the volume of the peak may be used to derive an estimate for significance of the peak to be used instead of Phifi. The preservation of all information at this step is particularly important for larger proteins where noise peaks and real peaks are closely interspersed in the spectral data.

The amino acid typing (scoring) step is the key in the integration of subnetworks. This process makes use of chemical shift probability density functions calculated from BMRB-PDB data for each atom of every amino acid in three states: alpha-helix, beta strand, and random coil. Amino acid typing is achieved by computing the probability of a chemical shift belonging to each residue i by

(5.11) images

where images, are the chemical shift probabilities of the related atom in amino acid i and phelix(i), pstrand(i), and pcoil(i) are the secondary structure probability of the amino acid i with the condition

(5.12)images

The above method connects amino acid typing and secondary structure states through a conditional dependency model. Notice that for the first iteration, where no assignment is available, secondary structure information follows a uniform distribution.

Deriving the backbone and side chain assignments from amino acid typing and other experimental data (connectivity experiments) is a computationally challenging element of the whole network. The approach in this step follows the model described in subsection, “Sequence-Specific Assignments.” First, the probabilistic assignments are analyzed by LACS for possible outlier detection. For each nucleus i and its assignment candidates j, one assigns a probability of being an outlier Py (outlier). Outliers need to be re-evaluated at each iteration. LACS also detects and corrects any possible referencing error in the spectra at this stage. Next, the PECAN algorithm17 is used to derive a probabilistic secondary structure determination. As the assignments are probabilistic (not unique), the combinatorial set of possible assignments and their corresponding secondary structures must be considered. After computing the probability of each residue i to be in each of three states (x = H, S, C) for different assignment configurations (c1, c2, cn), the overall probability can be calculated by

(5.13)images

in which p(x|cj) is the probability of residue i to be in state x with the assignment configuration Cj, and p(cj) is the probability of the configuration Cj.

Posterior probabilities derived in the assignment process are used as prior probabilities in the next round of assignment, provided that (1) LACS has not flagged the assignment as an outlier, (2) the assignment is fully consistent with the predicted secondary structure, and (3) the assignment is consistent with established connectivities. If any of the above conditions does not hold, scores are re-evaluated in the next iteration. The iteration process continues until a stationary or quasistationary state is reached. The iteration process leads to “self-correction” through appropriate adjustments to the similarity metric to preserve maximum information. The integration of individual steps within PINE leads to results that surpass those achieved by stepwise application of the separate steps in the pipeline (Figure 5.10). Results from multiple automated analyses by PINE for a range of proteins demonstrate average assignment accuracies well above the level of 90% needed for subsequent stages of automation. PINE is available from a web server at http://pine. nmrfam.wisc.edu/.

Inferential Structure Determination

The back-end stages of structure determination involve the specification of three-dimensional coordinates for atoms of the macromolecule as derived from indirect observations in the earlier stages. In application domains where the forward model is very accurate and the numerical predictions of the model are in good agreement with the measurements, the procedure for dealing with noise and the sparsity of data is well understood. However, NMR data are noisy and sparse, and the forward model is insufficiently accurate to determine aunique structure. Thus, errors resulting from model misspecification can be significant. A common approach is to model these errors by including some “model-noise” and by regularizing the forward model to include restraint terms that enforce agreement with the observed data.

Figure 5.10. Representation of automated assignments determined for the protein ubiquitin from a set of input data generated by HIFI-NMR. The heights of the vertical bars indicate assignment probabilities. Correct assignments are indicated by green (high probability) and yellow (low probability) bars; incorrect assignments, as determined from the known structure, are indicated by red bars. (α) Results achieved by the PISTACHIO software package alone. (β) Results from the PINE software package. The comparison demonstrates the improvement in performance achieved by the multistep iterative probabilistic approach embodied in PINE (A. Bahrami, A. H. Assadi, J. L. Markley, and H. R. Eghbalnia, manuscript in preparation). Figure also appears in the Color Figure section.

images

The current paradigm in NMR structure calculation is analogous to the process involved in inverting a noisy forward model. A typical scenario involves the minimization of a hybrid energy composed of an energy term representing the physical energy and a “regularization” term that measures the agreement of the model with the data through empirical measures. The components of the physical energy term describe the physical properties of the macromolecule, such as interactions between the atoms, and comprise the “force field.” The specification of the force field, however, does not sufficiently remove the degeneracy, and a regularization term is necessary to control the noise through agreement with the data. The degree to which this strategy works crucially depends on the quality of data. Specifically, when NMR structure determination is viewed as an ill-posed inverse problem, it becomes clear that the inversion strategy is successful if and only if the hybrid energy function does not describe a very rough energy surface. In cases where the hybrid energy represents a rough landscape, the determination of the weights for the regularization term becomes an expert’s task. In the absence of guiding principles, it is difficult to objectively evaluate the quality and reliability of determined structures. When structure ensembles depend on user-specific parameter settings and on the minimization protocol, statistically meaningful error estimates for atomic coordinates become practically infeasible.

Bayesian and maximum likelihood approaches have already proven useful for data analysis in NMR spectroscopy and X-ray crystallography. Intuitively, the uncertainties can be modeled by associating distributions to the parameter values that define the physical model of the protein. By replacing parameter values with distributions of values, and subsequently solving for a distribution of structure coordinates, the sensitivity of the structure coordinates to the parameters, and therefore the reliability of the resulting structure bundle, can be examined. This can allow the time (and space) evolution of the physical system to depart from deterministic evolution, thereby accommodating the effect of insufficiently modeled background events. For example, it is common practice to include the variance of the model-noise process in the collection of the parameters we wish to estimate from the data. It has been pointed out that allowing “model-noise” results in a good fit to the data, even if the deterministic part of the model is qualitatively wrong209.

One approach to address the information insufficiency in structure determination is by amending the experimental data with some prior knowledge. Inferential structure determination views structure determination as an inference problem that requires reasoning from noisy and incomplete data210. This approach does not lead to a unique structure; rather, it provides weights Pi for each conformation Xi —indicating the degree to which each conformation is supported by the experimental data. This inferential approach provides a principled way for handling the calculus of the inference problem210, 211. In this setting, the distribution of the probabilities Pi is interpreted as the information content of the data. A very sharp distribution of Pi values indicates a practically unique structure, while a flat distribution indicates lack of sufficient information for deciding among the conformational alternatives. The discrete setting can be readily generalized to the continuous model for the Cartesian coordinates of a macromolecule where probability densities now play the role of the probability mass. This Bayesian inferential approach demands that the probabilities be objective in the sense that they only depend on data D, and on relevant prior information I. The formal construction, through Bayes’ theorem, sets the posterior density to be proportional to the product of the likelihood of data given the model and the prior probability of a given macromolecular structure. This dependence on the prior is reflected in, for example, the use of the forward model; although other priors could be included. Other parameters that are required but cannot be measured, often referred to as nuisance parameters, are added to the list of parameters to be estimated. Therefore, the estimated density is formally a conditional probability that depends on the data D and the prior information I. This is a departure from the frequency view of probability.

In practice, the approach involves simulations of the posterior and sampling from areas of high probability mass. Once the terms in the Bayesian model have been established, the simulation provides an objective measure of agreement between data and the resulting structural model within numerical and simulation precision. Increased sampling of the conformational space and the suppression of entropically unfavorable states can provide a more realistic model of the macromolecule. The influence of parameters on the resulting model is transparent. For example, if the forward model does not incorporate any explicit information about protein dynamics, the formulation clearly shows that motion and imprecision due to experimental factors are indistinguishable. In addition, the sensitivity of the atomic positions to the parameters and the functional form of the force field can be explored in an objective manner.

It is often argued that the plausible specification of a prior is difficult or impossible. This difficulty can be addressed. Methods under development can achieve excellent results with little or no prior information. The robustified versions of these methods can guard against unexpected inferences resulting from prior selections. The current inferential approach focuses on the final stage of structure determination and does not rely on the propagation of uncertainties along the structure determination pipeline212, 213, 214. Although it is conceptually straightforward to include additional nuisance parameters, the increasing computational complexity and the presence of additional uncertainty models may lead to solutions that must be independently tested for stability and reliability. In the future, a comprehensive NMR pipeline based on inferential principles and capable of modeling other macromolecular features, for example, dynamics, should be capable of providing an objective, detailed, and feature-rich insight regarding the internal consistency, robustness, and accuracy of the resulting structural model.

The major international repository for biomolecular NMR data is Ulrich, et al., (2008), which is affiliated with the Research Collaboratory for Structural Bioinformatics (www.rcsb.org/) and is a member of the Worldwide Protein Data Bank (ww PDB, http://wwpdb.org/)215. Coordinates for structural models derived from NMR data, as from crystallography, are archived at the Protein Data Bank (PDB, http://www.rcsb.org/pdb)216.

Table 5.3 summarizes NMR data classes that are specific to biomolecular structure-function relationships. The primary information to be archived in a protein NMR investigation includes; (1) a complete description of the protein system studied (specification of each constituent and the stoichiometry of interacting constituents); (2) the solution conditions for each protein sample investigated (solvent, pH, temperature, pressure, and the concentration and isotopic labeling of each constituent); (3) a full description of each NMR experiment used in collecting data for the sample; (4) the NMR parameters and derived information obtained in the investigation; and (5) the unprocessed, time-domain NMR spectra collected. The present goal of the databases (PDB and BMRB) is to accommodate the inclusion of additional information on the preparation of the sample and the methods for structure determination at the level of detail provided in the methods section of a research journal. BMRB also accepts full data sets (time-domain spectra plus intermediate results) contributed by authors associated with a structure determination and now has archived more than 90 such sets to date.

In the interest of maintaining a common chemical shift standard for biomolecules, the international community has adopted the methyl signal of internal 2,2-dimethylsilapentane-5-sulfonic acid (DSS) at low concentration as the1H chemical shift standard. Chemical shifts of nuclei other than1H are referenced indirectly to the1H standard through the application of a conversion factor for each nucleus derived from ratios of NMR frequencies162. The BMRB217, 218 contains over 3.1 million experimental chemical shifts for proteins, nucleic acids, and small molecules as well as more than 14,000 experimental coupling constants. As of 2008, there were 2357 BMRB entries with associated PDB entries; of these, 1404 BMRB were available with chemical shifts referenced consistently using the IUPAC recommendations.

NMR structures in the PDB are represented by a description of the input data used for structure determination and refinement, the methods used to determine the structure, a statistical analysis of the results, and the structure itself, which usually is represented by a family of conformers that best satisfy the input constraints along with geometric and energetic criteria. One of these conformers is specified as being the single representative structure.

BMRB archives additional information about the chemical properties and dynamics of proteins derived from NMR spectroscopy, including hydrogen exchange rates at specified sites and pH titration parameters (pKa values, Hill coefficients, and pH-dependent spectroscopic shifts) for titratable groups and relaxation parameters (heteronuclear NOE, T1, T1rho, T2, and order parameters). Figure 5.11 illustrates the full range of information at BMRB.

TABLE 5.3. Summary of the NMR Data Classes That are Specific to Structural and Functional Genomics

Raw NMR Data
Time-domain data: free induction decays from a particular experiment with a particular sample under
defined conditions. The multiplicity of the spectrum (1D, 2D, 3D, 4D) results from the number of
time domains sampled
Processed NMR Data
Frequency-domain data: spectra derived from time-domain data by Fourier transformation and/or other
signal processing methods
NMR Parameters Extracted from NMR Data
Peak lists derived from individual data sets
Chemical shifts
1H--1H NOE
J-couplings
Residual dipolar couplings
NMR relaxation rates
Derived Information
Percentage of expected peaks observed in a data set
NMR peak assignments
Percentage of theoretical peaks assigned
Covalent structure
Bond hybridizations
Secondary structural elements
Interatomic distances
Torsion angles
Hydrogen bonds
Order parameters and other dynamic information
Solvent exposure
Three-dimensional structure (coordinates of the family of conformers that best correspond to the
experimental data; coordinates of the conformer that is designated as “representative”)
Binding constants
pH titration parameters (pKa values, Hill coefficients, titration shifts)
Hydrogen exchange rates
Delocalization of unpaired electrons
Thermodynamics and kinetics of structural interconversions
Disordered regions

NMR-STAR is the data model and tag-value data format for biomolecular NMR developed by BMRB in collaboration with the PDB, the Collaborative Computing Project for NMR (CCPN)219, and a number of contributors from the NMR community. NMR-STAR is an implementation of the STAR format developed for the communication of information in a compact ASCII format220, 221. The mmCIF data dictionary and data format are STAR implementations used in biomolecular crystallography (see Chapter 10). BMRB has developed software tools for interconverting NMR-STAR and mmCIF. NMR-STAR is used by BMRB as its data input format and by many biomolecular NMR software packages as a data exchange format. Specification of the NMR-STAR format and software tools for operating on NMR-STAR files are available from the BMRB web site (http://www.bmrb.wisc.edu). BMRB data are exported in theASCIINMR-STAR format and in a format that can beloaded easily into a relational database.

Figure 5.11. Potential information content of aBMRBentry for a particular target (study) showing the interrelatedness of the data to be collected.

images

BMRB established two new databases of structural restraints used to calculate NMR structures: DOCR (Database of Converted Restraints) and FRED (Filtered Restraints Database)222. Both databases were constructed in collaboration with EBI and CCPN groups from restraint data available from the PDB. The DOCR database contains restraint files with atom nomenclature consistent with the nomenclature in the corresponding atomic coordinate files. The restraints data in the FRED database have been filtered to remove redundant values and other restraints that do not impact the structure calculations. Over 500 protein structures were recalculated using restraints from these databases and the resulting structures analyzed for their quality144. The resulting database of recalculated structures (RECOORD) is available on the web (http://www.ebi.ac.uk/msd-srv/docs/NMR/recoord/main.html). The DOCR and FRED databases now contain over 4000 PDB NMR entries. The data are available in NMR-STAR format, X- PLOR, or an XML format.

As BMRB expands, the database of publicly accessible biomolecular NMR data also expands, creating a highly valuable resource for data mining and the development of statistical correlations between NMR observables and the structural and chemical properties of biomolecules. Examples include SHIFTY223, RefDB175, TALOS19, AVS176, PISTACHIO201, PECAN17, and other probabilistic studies224. BMRB routinely uses the AVS software package to inform depositors of suspicious or anomalous assigned chemical shifts within their depositions.

We thank the members of the CESG, NMRFAM, and BMRB teams for their contributions. Support from NIH grants P41 LM05799, 1U54 GM074901, and P41 RR02301 is gratefully acknowledged.

  1. Burley SK (2000): An overview of structural genomics. Nat Struct Biol 7:932-934.
  2. Heinemann U (2000): Structural genomics in Europe: slow start, strong finish?. Nat Struct Biol 7:940-942.
  3. Terwilliger TC (2000): Structural genomics in North America. Nat Struct Biol 7:935-939.
  4. Yokoyama S (2000): RIKEN Structural Genomics Initiative. International Conference on Structural Genomics 2000, November 2-5, Yokohama, Japan. p 40.
  5. Thornton JM, Todd AE, Milburn D, Borkakoti N, Orengo CA (2000): From structure to function: approaches and limitations. Nat Struct Biol, 7:(Suppl.): 991-994. [In process citation.]
  6. Dry S, McCarthy S, Harris T (2000): Structural genomics in the biotechnology sector. Nat Struct Biol 7:946-949.
  7. Mercier KA, Baran M, Ramanathan V, Revesz P, Xiao R, Montelione GT, Powers R (2006): FAST- NMR: functional annotation screening technology using NMR spectroscopy. J Am Chem Soc 128:15292-15299.
  8. Hajduk PJ, Greer J (2007): A decade of fragment-based drug design: strategic advances and lessons learned. Nature Rev 6:211-219.
  9. McDermott AE (2004): Structural and dynamic studies of proteins by solid-state NMR spectroscopy: rapid movement forward. Curr Opin Struct Biol 14:554-561.
  10. Baldus M (2006): Molecular interactions investigated by multidimensional solid-state NMR. Curr Opin Struct Biol 16:618-623.
  11. Hong M (2006): Oligomeric structure, dynamics, and orientation of membrane proteins from solid-state NMR. Structure 14:1731-1740.
  12. Cavanagh J, Fairbrother W, Palmer III AG, Skelton NJ Rance M (2006): Protein NMR Spectroscopy:Principles and Practice, 2nd ed. San Diego: Academic Press.
  13. Ernst RR, Bodenhausen G, Wokaun A (1987): Principles of Nuclear Magnetic Resonance in One and Two Dimensions. Oxford: Oxford University Press.
  14. Wishart DS, Sykes BD, Richards FM (1992): The chemical shift index: a fast and simple method for the assignment of protein secondary structure through NMR spectroscopy. Biochemistry 31:1647-1651.
  15. Wishart DS, Sykes BD (1994a): The13 C chemical shift index: a simple method for the identification of protein secondary structure using 13C chemical shifts. J Biomol NMR 4:171-180.
  16. Wishart DS, Case DA (2001): Use of chemical shifts in macromolecular structure determination. Methods Enzymol 338:3-34.
  17. Eghbalnia HR, Wang L, Bahrami A, Assadi A, Markley JL (2005c): Protein energetic conformational analysis from NMR chemical shifts (PECAN) and its use in determining secondary structural elements. J Biomol NMR 32:71-81.
  18. Sharma D, Rajarathnam K (2000): 13C NMR chemical shifts can predict disulfide bond formation. J Biomol NMR 18:165-171.
  19. Cornilescu G, Delaglio F, Bax A (1999): Protein backbone angle restraints from searching a database for chemical shift and sequence homology. J Biomol NMR 13:289-302.
  20. Bax A, Kontaxis G, Tjandra N (2001): Dipolar couplings in macromolecular structure determination. Methods Enzymol 339:127-174.
  21. Lipsitz RS, Tjandra N (2004): Residual dipolar couplings in NMR structure analysis. Annu Rev Biophys Biomol Struct 33:387-413.
  22. Prestegard JH, Bougault CM, Kishore AI (2004): Residual dipolar couplings in structure determination of biomolecules. Chem Rev 104:3519-3540.
  23. Bax A, Grishaev A (2005): Weak alignment NMR: a hawk-eyed view of biomolecular structure. Curr Opin Struct Biol 15:563-570.
  24. Cordier F, Grzesiek S (1999): Direct observation of hydrogenbonds in proteins by interresidue 3H J NC’ scalar couplings. J Am Chem Soc 121:1601-1602.
  25. Cordier F, Rogowski M, Grzesiek S, Bax A (1999): Observation of through-hydrogen-bond (2h)J(HC’) in a perdeuterated protein. J Magn Reson 140:510-512.
  26. Wuuthrich K (1986): NMR ofProteins and Nucleic Acids. New York, NY: Wiley-Interscience.
  27. Markley JL, Kainosho M (1993): Stable isotope labeling and resonance assignments in larger proteins. In: Roberts G.C.K. editor. NMR of Biological Macromolecules: A Practical Approach. Oxford, England: Oxford University Press, pp 101-152.
  28. Westler WM, Ortiz-Polo G, Markley JL (1984): Two-dimensional 1 H-13C chemical-shift correlated spectroscopy of a protein at natural abundance. J Magn Reson 58:354-357.
  29. Ortiz-Polo G, Krishnamoorthi R, Markley JL, Live DH, Davis DG, Cowburn D (1986): Natural- abundance 15N NMR studies of Turkey ovomucoid third domain. Assignment of peptide 15N resonances to the residues at the reactive site region via proton-detected multiple-quantum coherence. J Magn Reson 68:303-310.
  30. Oh BH, Westler WM, Darba P, Markley JL Protein carbon-13 spin systems by a single two-dimensional nuclear magnetic resonance experiment. Science 240:(1988): 908-911.
  31. Westler WM, Kainosho M, Nagao H, Tomonaga N, Markley JL (1988a): Two-dimensional NMR strategies for carbon-carbon correlations and sequence-specific assignments in carbon-13 labeled proteins. J Am Chem Soc 110:4093-4095.
  32. Westler WM, Stockman BJ, Hosoya Y, Miyake Y, Kainosho M, Markley JL (1988b): Correlation of carbon-13 and Nitrogen-15 chemical shifts in uniformly labeled proteins by heteronuclear two-dimensional NMR spectroscopy. J Am Chem Soc 110:6265-6258.
  33. Oh BH, Westler WM, Markley JL (1989): Carbon-13 spin system directed strategy for assigning cross peaks in the COSY fingerprint region of a protein. J Am Chem Soc 111:3083-3085.
  34. Kay LE, Marion D, Bax A (1989): Practical aspects of 3D heteronuclear NMR of proteins. J Magn Reson 84:72-84.
  35. Kay LE, Ikura M, Tschudin R, Bax A (1990): 3-Dimensional triple-resonance NMR-spectroscopy of isotopically enriched proteins. J Magn Reson 89:496-514.
  36. Czisch M, Boelens R (1998): Sensitivity enhancement in the TROSY experiment. J Magn Reson 134:158-160.
  37. Pervushin KV, Riek R, Wider G, Wuuthrich K (1998): Transverse relaxation-optimized spectroscopy (TROSY) for NMR studies of aromatic spin systems in 13 C-labeled proteins. J Am Chem Soc 120:6394-6400.
  38. Salzmann M, Pervushin KV, Wider G, Senn H, Wuthrich K (1998): TROSY in triple-resonance experiments: new perspectives for sequential NMR assignment of large proteins. Proc Natl Acad Sci USA 95:13585-13590.
  39. Zhu G, Kong XM, Sze KH (1999): Gradient and sensitivity enhancement of 2D TROSY with water flip-back, 3D NOESY-TROSY and TOCSY-TROSY experiments. J Biomol NMR 13:77-81.
  40. Zhu G, Xia YL, Sze KH, Yan XZ (1999): 2D and 3D TROSY-enhanced NOESY of N-15 labeled proteins. J Biomol NMR 14:377-381.
  41. Fernandez C, Wider G (2003): TROSY in NMR studies of the structure and function of large biological macromolecules. Curr Opin Struct Biol 13:570-580.
  42. Hu K, Eletsky A, Pervushin K (2003): Backbone resonance assignment in large protonated proteins using a combination of new 3D TROSY-HN(CA)HA, 4D TROSY-HACANH and 13C-detected HACACO experiments. J Biomol NMR 26:69-77.
  43. Lopez-Mendez B, Guuntert P (2006): Automated protein structure determination from NMR spectra. JAm Chem Soc 128:13112-13122.
  44. Gardner KH, Kay LE (1998): The use of 2H, 13C, 15N multidimensional NMR to study the structure and dynamics of proteins. Annu Rev Biophys Biomol Struct 27:357-1-406.
  45. Tugarinov V, Kanelis V, Kay LE (2006): Isotope labeling strategies for the study of high-molecular-weight proteins by solution NMR spectroscopy. Nat Protoc 1:749-754.
  46. Yamazaki T, Otomo T, Oda N, Kyogoku Y, Uegaki K, Ito N, Ishino Y, Nakamura H (1998): Segmental isotope labeling for protein NMR using peptide splicing. J Am Chem Soc 120:5591-5592.
  47. Xu R, Ayers B, Cowburn D, Muir TW (1999): Chemical ligation of folded recombinant proteins: segmental isotopic labeling of domains for NMR studies. Proc Natl Acad Sci USA 96:388-393.
  48. Kainosho M, Torizawa T, Iwashita Y, Terauchi T, Mei Ono A, Guntert P (2006): Optimal isotope labelling for NMR protein structure determinations. Nature 440:52-57.
  49. Styles P, Soffe NF, Scott CA, Cragg DA, Row F, White DJ, White PCJ (1984): A high-resolution NMR probe in which the coil and preamplifier are cooled with liquid helium. J Magn Reson 60:397-404.
  50. de Swiet TM (2005): Optimal electric fields for different sample shapes in high resolution NMR spectroscopy. J Magn Reson 174:331-334.
  51. Eghbalnia HR, Bahrami A, Tonelli M, Hallenga K, Markley JL (2005a): High-resolution iterative frequency identification for NMR as a general strategy for multidimensional data collection. Jam Chem Soc 127:12528-12536.
  52. Kim S, Szyperski T (2003): GFT NMR, a new approach to rapidly obtain precise high-dimensional NMR spectral information. JAm Chem Soc 125:1385-1393.
  53. Kupce E, Freeman R (2003): Reconstruction of the three-dimensional NMR spectrum of a protein from a set of plane projections. J Biomol NMR 27:383-387.
  54. Schanda P, Brutscher B (2005): Very fast two-dimensional NMR spectroscopy for real-time investigation of dynamic events in proteins on the time scale of seconds. J Am Chem Soc 127:8014-8015.
  55. Lescop E, Schanda P, Brutscher B (2007): A set of BEST triple-resonance experiments for time- optimized protein resonance assignment. J Magn Reson 187:163-169.
  56. Wand AJ (2001): Dynamic activation of protein function: a view emerging from NMR spectroscopy. Nat Struct Biol 8:926-931.
  57. Kern D, Zuiderweg ER (2003): The role of dynamics in allosteric regulation. Curr Opin Struct Biol 13:748-757.
  58. Mittermaier A, Kay LE (2006): New tools provide new insights in NMR studies of protein dynamics. Science 312:224-228.
  59. Dyson HJ, Wright PE(2005): Intrinsically unstructured proteins and theirfunctions. Nat RevMolCell Biol 6:197-208.
  60. Palmer III AG (1997): Probing molecular motion by NMR. Curr Opin Struct Biol 7:732-737.
  61. Igumenova TI, Frederick KK, Wand AJ (2006): Characterization of the fast dynamics of protein amino acid side chains using NMR relaxation in solution. Chem Rev 106:1672-1699.
  62. Lipari G, Szabo A (1982a): Model-free approach to the interpretation of nuclear magnetic resonance relaxation in macromolecules. 1. Theory and range of validity. JAmChemSoc 104:4546-4559.
  63. Lipari G, Szabo A (1982b): Model-free approach to the interpretation of nuclear magnetic resonance relaxation in macromolecules. 2. Analysis of experimental results. J Am Chem Soc 104:4559-4570.
  64. Clore GM, Szabo A, Bax A, Kay LE, Driscoll PC, Gronenborn AM (1990): Deviations from the simple two-parameter model-free approach to the interpretation of nitrogen-15 nuclear magnetic relaxation of proteins. J Am Chem Soc 112:4989-4991.
  65. Mandel AM, Akke M, Palmer III AG (1995): Backbone dynamics of Escherichia coli ribonuclease HI: correlations with structure and function in an active enzyme. J Mol Biol 246:144-163.
  66. Andrec M, Montelione GT, Levy RM(2000): Lipari-Szabomapping: AgraphicalapproachtoLipari-Szabo analysis of NMR relaxation data using reduced spectral density mapping. J Biomol NMR 18:83-100.
  67. Volkman BF, Lipson D, Wemmer DE, Kern D (2001): Two-state allosteric behavior in a single-domain signaling protein. Science 291:2429-2433.
  68. Akke M, Palmer III, AG (1996): Monitoring macromolecular motions on microsecond-millisecond time scales by R1R-R1 constant relaxation time NMR spectroscopy. J Am Chem Soc 118:911-912.
  69. Sugase K, Dyson HJ, Wright PE (2007): Mechanism of coupled folding and binding of an intrinsically disordered protein. Nature 447:1021-1025.
  70. Eisenmesser EZ, Millet O, Labeikovsky W, Korzhnev DM, Wolf-Watz M, Bosco DA, Skalicky JJ, Kay LE, Kern D (2005): Intrinsic dynamics of an enzyme underlies catalysis. Nature 438:117-121.
  71. Sprangers R, Kay LE (2007): Quantitative dynamics and binding studies of the 20S proteasome by NMR. Nature 445:618-622.
  72. Schanda P, Forge V, Brutscher B (2007): Protein folding and unfolding studied at atomic resolution by fast two-dimensional NMR spectroscopy. Proc Natl Acad Sci USA 104:11257-11262.
  73. Huyghues-Despointes BM, Pace CN, Englander SW, Scholtz JM (2001): Measuring the conforma- tional stability of a protein by hydrogen exchange. Methods Mol Biol 168:69-92.
  74. Hwang TL, van Zijl PCM, Mori S (1998): Accurate quantitation of water-amide proton exchange rates using the phase-modulated CLEAN chemical EXchange (CLEANEX-PM) approach with a Fast-HSQC (FHSQC) detection scheme. J Biomol NMR 11:221-226.
  75. Yokoyama S (2003): Protein expression systems for structural genomics and proteomics. Curr Opin Chem Biol 7:39-43.
  76. Henrich B, Lubitz W, Plapp R (1982): Lysis of Escherichia coli by induction of cloned Phi-X174 genes. Mol Gen Genet 185:493-497.
  77. Goff SA, Goldberg AL (1987): An increased content of protease la, the Ion gene-product, increases protein-degradation and blocks growth in Escherichia coli. JBiolChem262: 4508-4515.
  78. Chrunyk BA, Evans J, Lillquist J, Young P, Wetzel R (1993): Inclusion-body formation and protein stability in sequence variants of interleukin-1-beta. J Biol Chem 268:18053-18061.
  79. Walton WJ, Kasprzak AJ, Hare JT, Logan TM (2006): An economic approach to isotopic enrichment of glycoproteins expressed from Sf9 insect cells. J Biomol NMR 36:225-233.
  80. Wood MJ, Komives EA (1999): Production of large quantities of isotopically labeled protein in Pichia pastoris by fermentation. J Biomol NMR 13:149-159.
  81. Pickford AR, O’Leary JM (2004): Isotopic labeling of recombinant proteins from the methylotrophic yeast Pichia pastoris. Methods Mol Biol 278:17-33.
  82. Kurland CG (1982): Translational accuracy in vitro. Cell 28:201-202.
  83. Pavlov MY, Ehrenberg M (1996): Rate of translation of natural mRNAs in an optimized in vitro system. Arch Biochem Biophys 328:9-16.
  84. Spirin AS, Baranov VI, Ryabova LA, Ovodov SY, Alakhov YB (1988): A continuous cell-free translation system capable of producing polypeptides in high-yield. Science 242:1162-1164.
  85. Baranov VI, Morozov IY, Ortlepp SA, Spirin AS (1989): Gene-expression in a cell-free system on the preparative scale. Gene 84:463-466.
  86. Endo Y, Otsuzuki S, Ito K, Miura K (1992): Production of an enzymatic active protein using a continuous-flow cell-free translation system. J Biotechnol 25:221-230.
  87. Sawasaki T, Hasegawa Y, Tsuchimochi M, Kamura N, Ogasawara T, Kuroita T, Endo Y (2002a): A bilayer cell-free protein synthesis system for high-throughput screening of gene products. FEBS Lett 514:102-105.
  88. Sawasaki T, Ogasawara T, Morishita R, Endo Y (2002b): A cell-free protein synthesis system for high-throughput proteomics. Proc Natl Acad Sci USA 99:14652-14657.
  89. Kigawa T, Yabuki T, Yoshida Y, Tsutsui M, Ito Y, Shibata T, Yokoyama S (1999): Cell-free production and stable-isotope labeling of milligram quantities of proteins. FEBS Lett 442:15-19.
  90. Torizawa T, Shimizu M, Taoka M, Miyano H, Kainosho M (2004): Efficient production of isotopically labeled proteins by cell-free synthesis: a practical protocol. J Biomol NMR 30:311-325.
  91. Matsuda T, Koshiba S, Tochio N, Seki E, Iwasaki N, Yabuki T, Inoue M, Yokoyama S, Kigawa T (2007): Improving cell-free protein synthesis for stable-isotope labeling. J Biomol NMR 37:225-229.
  92. Klammt C, Lohr F, Schafer B, Haase W, Dotsch V, Ruterjans H, Glaubitz C, Bernhard F (2004): High level cell-free expression and specific labeling of integral membrane proteins. Eur J Biochem 271:568-580.
  93. Klammt C, Schwarz D, Lohr F, Schneider B, Dotsch V, Bernhard F (2006): Cell-free expression as an emerging technique for the large scale production of integral membrane protein. FEBS J 273:4141-4153.
  94. Liguori L, Marques B, Villegas-Mendez A, Rothe R, Lenormand JL (2007): Production ofmembrane proteins using cell-free expression systems. Expert Rev Proteomics 4:79-90.
  95. Kigawa T, Muto Y, Yokoyama S (1995): Cell-free synthesis and amino acid-selective stable isotope labeling of proteins for NMR analysis. J Biomol NMR 6:129-134.
  96. Vinarov DA, Loushin Newman CL, Markley JL (2006): Wheat germ cell-free platform for eukaryotic protein production. FEBS J 273:4160-4169.
  97. Tyler RC, Aceti DJ, Bingman CA, Cornilescu CC, Fox BG, Frederick RO, Jeon WB, Lee MS, Newman CS, Peterson FC, et al, 2005a Comparison of cell-based and cell-free protocols for producing target proteins from the Arabidopsis thaliana genome for structural studies. Proteins 59:633-643.
  98. Edwards AE, Arrowsmith CH, Christendat D, Dharamsi A, Friesen JD, Greenblatt JF, Vedadi M (2000): Protein production: feeding the crystallographers and NMR spectroscopists. Nat Struct Biol 7:970-972.
  99. Blommel PG, Becker KJ, Duvnjak P, Fox BG (2007): Enhanced bacterial protein expression during auto-induction obtained by alteration of lac repressor dosage and medium composition. Biotechnol Prog 23:585-598.
  100. Tyler RC, Sreenath HK, Singh S, Aceti DJ, Bingman CA, Markley JL, Fox BG (2005b): Auto- induction medium for the production of [U-15N]- and [U-13C, U-15N]-labeled proteins for NMR screening and structure determination. Prot Expr Purif 40:268-278.
  101. Sanville MC, Stols L, Quartey P, Kim Y, Dementieva I, Donnelly MI (2003): A less laborious approach to the high-throughput production of recombinant proteins in Escherichia coli using 2-liter plastic bottles. Prot Expr Purif 29:311-320.
  102. Vinarov DA, Lytle BL, Peterson FC, Tyler EM, Volkman BF, Markley JL (2004): Cell-free protein production and labeling protocol forNMR-based structural proteomics. Nat Methods 1:149-153.
  103. Vinarov DA, Markley JL (2005): High-throughput automated platform for nuclear magnetic resonance-based structural proteomics. Expert Rev Proteomics 2:49-55.
  104. Vinarov DA, Loushin Newman CL, Tyler EM, Markley JL (2006): Protein production using the wheat germ cell-free expression system. In: Wingfield PT, editor. Current Protocols in Protein Science. John Wiley & Sons, Chapter 5.
  105. Peti W, Page R, Moy K, O’Neil-Johnson M, Wilson IA, Stevens RC, Wuuthrich K (2005): Towards miniaturization of a structural genomics pipeline using micro-expression and microcoil NMR. J Struct Funct Genomics 6:259-267.
  106. Aramini JM, Rossi P, Anklin C, Xiao R, Montelione GT (2007): Microgram-scale protein structure determination by NMR. Nat Methods 4:491-493.
  107. Lytle BL, Peterson FC, Tyler EM, Newman CL, Vinarov DA, Markley JL, Volkman BF (2006): Solution structure of Arabidopsis thaliana protein At5g39720.1, a member of the AIG2-like protein family. Acta Crystallogr F 62:490-493.
  108. Peterson FC, Hayes PL, Waltner JK, Heisner AK, Jensen DR, Sander TL, Volkman BF (2007b): Erratum to “Structure of the SCAN domain from the tumor suppressor protein MZF1” [J Mol Biol 363 (2006) 137-147]. J Mol Biol 366:346.
  109. Tolman JR, Flanagan JM, Kennedy MA, Prestegard JH (1995): Nuclear magnetic dipole interactions in field-oriented proteins: information for structure determination in solution. Proc Natl Acad Sci USA 92:9279-9283.
  110. Tjandra N, Bax A (1997): Direct measurement of distances and angles in biomolecules by NMR in a dilute liquid crystalline medium. Science 278:1111-1114.
  111. Tjandra N, Omichinski JG, Gronenborn AM, Clore GM, Bax A (1997): Use of dipolar 1 H-15 N and 1 H-13 C couplings in the structure determination of magnetically oriented macromolecules in solution. Nat Struct Biol 4:732-738.
  112. Prestegard JH (1998): New techniques in structural NMR: anisotropic interactions. Nat Struct Biol 5:517-522.
  113. Tian J, Valafar H, Prestegard JH (2001): A dipolar coupling based strategy for simultaneous resonance assignment and structure determination of protein backbones. JAmChemSoc 123:11791-11796.
  114. Bax A, Tjandra N (1997): High-resolution heteronuclear NMR of human ubiquitin in an aqueous liquid crystalline medium. J Biomol NMR 10:289-292.
  115. Ottiger M, Bax A (1999): Bicelle-basedliquid crystals forNMR-measurementofdipolarcouplings at acidic and basic pH values. J Biomol NMR 13:187-191.
  116. Hansen MR, Mueller L, Pardi A (1998): Tunable alignment of macromolecules by filamentous phage yields dipolar coupling interactions. Nat Struct Biol 5:1065-1074.
  117. Chou JJ, Gaemers S, Howder B, Louis JM, Bax A (2001): A simple apparatus for generating stretched polyacrylamide gels, yielding uniform alignment of proteins and detergent micelles. J Biomol NMR 21:377-382.
  118. Montelione GT, Zheng D, Huang YJ, Gunsalus KC, Szyperski T (2000): Protein NMR spectroscopy in structural genomics. Nat Struct Biol 7:982-985.
  119. Szyperski T, Yeh DC, Sukumaran DK, Moseley HN, Montelione GT (2002): Reduced-dimensionality NMR spectroscopy for high-throughput protein resonance assignment. Proc Natl Acad Sci USA 99:8009-8014.
  120. Stuart AC, Borzilleri KA, Withka JM, Palmer III AG (1999): Compensating for variations in 1H-13C scalar coupling constants in isotope-filtered NMR experiments. J Am Chem Soc 121:5346-5347.
  121. Chylla RA, Markley JL (1995): Theory and application of the maximum likelihood principle to NMR parameter estimation of multidimensional NMR data. J Biomol NMR 5:245-258.
  122. Chylla RA, Volkman BF, Markley JL (1998): Practical model fitting approaches to the direct extraction of NMR parameters simultaneously from all dimensions of multidimensional NMR spectra. J Biomol NMR 12:277-297.
  123. Delaglio F, Grzesiek S, Vuister GW, Zhu G, Pfeifer J, Bax A (1995): NMRPIPE: a multidimensional spectral processing system based on UNIX pipes. J Biomol NMR 6:277-293.
  124. Bartels C, Xia TH, Billeter M, Guntert P, Wuthrich K (1995): The program XEASY for computer- supported NMR spectral-analysis of biological macromolecules. J Biomol NMR 5:1-10.
  125. Bartels C, Billeter M, Guntert P, Wuthrich K (1996): Automated sequence-specific NMR assignment of homologous proteins using the program GARANT. J Biomol NMR 7:207-213.
  126. Guntert P, Mumenthaler C, Wuthrich K (1997): Torsion angle dynamics for NMR structure calculation with the new program DYANA. J Mol Biol 273:283-298.
  127. Herrmann T, Guntert P, Wuuthrich K (2002b): Protein NMR structure determination with automated NOE assignment using the new software CANDID and the torsion angle dynamics algorithm DYANA. J Mol Biol 319:209-227.
  128. Zimmerman DE, Kulikowski CA, Huang Y, Feng W, Tashiro M, Shimotakahara S, Chien C, Powers R, Montelione GT (1997): Automated analysis of protein NMR assignments using methods from artificial intelligence. J Mol Biol 269:592-610.
  129. Olson Jr JB , Markley JL (1994): Evaluation of an algorithm for the automated sequential assignment of protein backbone resonances: a demonstration of the connectivity tracing assignment tools (CONTRAST) software package. J Biomol NMR 4:385-410.
  130. Jee J, Guntert P (2003): Influence of the completeness of chemical shift assignments on NMR structures obtained with automated NOE assignment. J Struct Funct Genomics 4:179-189.
  131. Guuntert P (2004): Automated NMR structure calculation with CYANA. Methods Mol Biol 278:353-378.
  132. Lytle BL, Peterson FC, Kjer KL, Frederick RO, Zhao Q, Thao S, Bingman C, Johnson KA, Phillips GN Jr, Volkman BF (2004a): Structure of the hypothetical protein At3g17210 from Arabidopsis thaliana. J Biomol NMR 28:397-400.
  133. Lytle BL, Peterson FC, Qiu SH, Luo M, Zhao Q, Markley JL, Volkman BF (2004b): Solution structure of a ubiquitin-like domain from tubulin-binding cofactor B. J Biol Chem 279:46787-46793.
  134. Peterson FC, Lytle BL, Sampath S, Vinarov D, Tyler E, Shahan M, Markley JL, Volkman BF (2005): Solution structure of thioredoxin h1 from Arabidopsis thaliana. Prot Sci 14:2195-2200.
  135. Waltner JK, Peterson FC, Lytle BL, Volkman BF (2005): Structure of the B3 domain from Arabidopsis thaliana protein At1g16640. Prot Sci 14:2478-2483.
  136. Peterson FC, Hayes PL, Waltner JK, Heisner AK, Jensen DR, Sander TL, Volkman BF (2006a): Structure of the SCAN domain from the tumor suppressor protein MZF1. J Mol Biol 363:137-147.
  137. Peterson FC, Thorpe JA, Harder AG, Volkman BF, Schwarze SR (2006b): Structural determinants involved in the regulation of CXCL14/BRAK expression by the 26 S proteasome. J Mol Biol 363:813-822.
  138. Peterson FC, Deng Q, Zettl M, Prehoda KE, Lim WA, Way M, Volkman BF (2007a): Multiple WASP-interacting protein recognition motifs are required for a functional interaction with N-WASP. JBiol Chem 282:8446-8453.
  139. Tuinstra RL, Peterson FC, Elgin ES, Pelzek AJ, Volkman BF (2007): An engineered second disulfide bond restricts lymphotactin/XCL1 to a chemokine-like conformation with XCR1 agonist activity. Biochemistry 46:2564-2573.
  140. Herrmann T, Guntert P, Wuuthrich K (2002a): Protein NMR structure determination with automated NOE-identification in the NOESY spectra using the new software ATNOS. J Biomol NMR 24:171-189.
  141. Sali A, Blundell TL (1993): Comparative protein modelling by satisfaction of spatial restraints. J Mol Biol 234:779-815.
  142. Ivanov D, Stone JR, Maki JL, Collins T, Wagner G (2005): Mammalian SCAN domain dimer is a domain-swapped homolog of the HIV capsid C-terminal domain. Mol Cell 17:137-143.
  143. Linge JP, Williams MA, Spronk CA, Bonvin AM, Nilges M (2003): Refinement of protein structures in explicit solvent. Proteins 50:496-506.
  144. Nederveen AJ, Doreleijers JF, Vranken W, Miller Z, Spronk CA, Nabuurs SB, Guntert P, Livny M, Markley JL, Nilges M, et al, (2005): RECOORD: a recalculated coordinate database of 500 + proteins from the PDB using restraints from the BioMagResBank. Proteins 59:662-672.
  145. Laskowski RA, MacArthur MW, Moss DS, Thornton JM (1993): PROCHECK: a program to check the stereochemical quality of protein structures. J Appl Crystallogr 26:283-291.
  146. Vriend G (1990): WHAT IF: a molecular modeling and drug design program. J Mol Graph 8:52-56.
  147. Laskowski RA, Rullmann J.A.C, MacArthur MW, Kaptein R, Thornton JM (1996): AQUA and PROCHECK-NMR: programs for checking the quality of protein structures solved by NMR. J Biomol NMR 8:477-486.
  148. Doreleijers JF, Vriend G, Raves ML, Kaptein R (1999b): Validation of nuclear magnetic resonance structures of proteins and nucleic acids: hydrogen geometry and nomenclature. Proteins 37:404-416.
  149. Engh RA, Huber R (1991): Accurate bond and angle parameters for X-ray protein structure refinement. Acta Cryst A47:392-400.
  150. Parkinson G, Voitechovsky J, Clowney L, Brunger AT, Berman HM (1996): Bond lengths and angles, DNA/RNA. Acta Cryst D 52:57-64.
  151. Berman HM, Bhat TN, Bourne PE, Feng Z, Gilliland G, Weissig H, Westbrook J (2000a): The Protein Data Bank and the challenge of structural genomics. Nat Struct Biol 7(Suppl):957-959.
  152. Bowie JU, Luthy R, Eisenberg D (1991): A method to identify protein sequences that fold into a known three-dimensional structure. Science 253:164-170.
  153. Luuthy R, Bowie JU, Eisenberg D (1992): Assessment of protein models with three-dimensional profiles. Nature 356:83-85.
  154. Sippl MJ (1993): Recognition of errors in three-dimensional structures of proteins. Proteins 17: 355-362.
  155. Wiederstein M, Sippl MJ (2007): ProSA-web: interactive web service for the recognition of errors in three-dimensional structures of proteins. Nucl Acids Res 35:W407-410.
  156. Lovell SC, Davis IW, Arendall III WB de Bakker PI, Word JM, Prisant MG, Richardson JS, Richardson DC (2003): Structure validation by C-alpha geometry: phi, psi and C-beta deviation. Proteins 50:437-450.
  157. Davis IW, Murray LW, Richardson JS, Richardson DC (2004): MOLPROBITY: structure validation and all-atom contact analysis for nucleic acids and their complexes. Nucl Acids Res 32: W615-W619.
  158. Davis IW, Leaver-Fay A, Chen VB, Block JN, Kapral GJ, Wang X, Murray LW, Arendall III WB Snoeyink J, Richardson JS, Richardson DC (2007): MolProbity: all-atom contacts and structure validation for proteins and nucleic acids. Nucl Acids Res 35:W375-383.
  159. Willard L, Ranjan A, Zhang H, Monzavi H, Boyko RF, Sykes BD, Wishart DS (2003): VADAR: a web server for quantitative evaluation of protein structure quality. Nucl Acids Res 31:3316-3319.
  160. Doreleijers JF, Rullmann J.A.C, Kaptein R (1998): Quality assessment of NMR structures: a statistical survey. J Mol Biol 281:149-164.
  161. Doreleijers JF, Raves ML, Rullmann T, Kaptein R (1999a): Completeness of NOEs in protein structure: a statistical analysis of NMR data. J Biomol NMR 14:123-132.
  162. Markley JL, Bax A, Arata Y, Hilbers CW, Kaptein R, Sykes BD, Wright PE, Wuuthrich K (1998): Recommendations for the presentation of NMR structures of proteins and nucleic acids. J Mol Biol 280:933-952.
  163. Gronwald W, Kalbitzer HR (2004): Automated structure determination of proteins by NMR spectroscopy. Prog NuclMagn Reson Spectrosc 44:33-96.
  164. Gronwald W, Kirchhofer R, Gorler A, Kremer W, Ganslmeier B, Neidig KP, Kalbitzer HR (2000): RFAC, a program for automated NMR R-factor estimation. J Biomol NMR 17:137-151.
  165. Gronwald W, Brunner K, Kirchhofer R, Trenner J, Neidig KP, Kalbitzer HR (2007): AUREMOL-RFAC-3D, combination of R-factors and their use for automated quality assessment of protein solution structures. J Biomol NMR 37:15-30.
  166. Brunger AT (1992): Free R-value: a novel statistical quanity for assessing the accuracy of crystal- structures. Nature 355:472-475.
  167. Huang YJ, Moseley HN, Baran MC, Arrowsmith C, Powers R, Tejero R, Szyperski T, Montelione GT (2005): An integrated platform for automated analysis of protein NMR structures. Methods Enzymol 394:111-141.
  168. Clore GM, Gronenborn AM, Tjandra N (1998): Direct structure refinement against residual dipolar couplings in the presence of rhombicity of unknown magnitude. J Magn Reson 131:159-162.
  169. Cornilescu G, Marquardt JL, Ottinger M, Bax A (1998): Validation of protein structure from anisotropic carbonyl chemical shifts in a dilute liquid crystalline phase. JAm Chem Soc 120:6836-6837.
  170. Case DA (1998): The use of chemical shifts and their anisotropies in biomolecular structure determination. Curr Opin Struct Biol 8:624-630.
  171. Williamson MP, Kikuchi J, Asakura T (1995): Application of 1 H NMR chemical shifts to measure the quality of protein structures. J Mol Biol 247:541-546.
  172. Case DA (2000): Interpretation of chemical shifts and coupling constants in macromolecules. Curr Opin Struct Biol 10:197-203.
  173. Xu XP, Case DA (2001): Automated prediction of N-15, C-13(alpha), C-13(beta) and C-13’ chemical shifts in proteins using a density functional database. J Biomol NMR 21:321-333.
  174. Neal S, Nip AM, Zhang H, Wishart DS (2003): Rapid and accurate calculation of protein 1H, 13Cand 15N chemical shifts. J Biomol NMR 26:215-240.
  175. Zhang H, Neal S, Wishart DS (2003): RefDB: a database of uniformly referenced protein chemical shifts. J Biomol NMR 25:173-195.
  176. Moseley HN, Sahota G, Montelione GT (2004): Assignment validation software suite for the evaluation and presentation of protein resonance assignment data. J Biomol NMR 28:341-355.
  177. Wang L, Eghbalnia HR, Bahrami A, Markley JL (2005): Linear analysis of carbon-13 chemical shift differences and its application to the detection and correction of errors in referencing and spin system identifications. J Biomol NMR 32:13-22.
  178. Bourne PE, Berman HM, McMahon B, Watenpaugh KD, Westbrook JD, Fitzgerald PMD (1997): The macromolecular crystallographic information file (mmCIF). Methods Enzymol 277: 571-590.
  179. Berman HM, Westbrook J, Feng Z, Gilliland G, Bhat TN, Weissig H, Shindyalov IN, Bourne PE (2000b): The Protein Data Bank. Nucl Acids Res 28:235-242.
  180. Yang H, Guranovic V, Dutta S, Feng Z, Berman HM, Westbrook JD (2004): Automated and accurate deposition of structures solved by X-ray diffraction to the Protein Data Bank. Acta Crystallogr D 60:1833-1839.
  181. Hammersley JM, Clifford P (1971): Markov field on finite graphs and lattices. (Unpublished manuscript.).
  182. Spitzer F (1971): Markov random fields and Gibbs ensembles. Amer Math Monthly 78:142-154.
  183. Besag J (1974): Spatial interaction and statistical-analysis of lattice systems. J Royal Stat Soc Ser B 36:192-236.
  184. Chentsov NN (1982): Statistical Decision Rules and Optimal Inference. Providence, RI: American Mathematical Society.
  185. Janyszek H, Mrugala R (1989): Riemannian geometry and the thermodynamics of model magnetic systems. Phys Rev A 39:6515-6523.
  186. Ruppeiner G (1995): Riemannian geometry in thermodynamic fluctuation theory. Rev Modern Phys 67:605-659.
  187. Cornilescu G, Bahrami A, Tonelli M, Markley JL, Eghbalnia HR (2007): HIFI-C: a robust and fast method for determining NMR couplings from adaptive 3D to 2D projections. J Biomol NMR 38:341-351.
  188. Buchler NE, Zuiderweg ER, Wang H, Goldstein RA (1997): Protein heteronuclear NMR assignments using mean-field simulated annealing. J Magn Reson 125:34-42.
  189. Lukin JA, Gove AP, Talukdar SN, Ho C (1997): Automated probabilistic method for assigning backbone resonances of (13C, 15N)-labeled proteins. J Biomol NMR 9:151-166.
  190. Leutner M, Gschwind RM, Liermann J, Schwarz C, Gemmecker G, Kessler H (1998): Automated backbone assignment of labeled proteins using the threshold accepting algorithm. J Biomol NMR 11:31-43.
  191. Bartels C, Guuntert P, Billeter M, Wuuthrich K (1997): GARANT: a general algorithm for resonance assignment of multidimensional nuclear magnetic resonance spectra. J Comput Chem 18:139-149.
  192. Atreya HS, Sahu SC, Chary KV, Govil G(2000): AtrackedapproachforautomatedNMRassignments in proteins (TATAPRO). J Biomol NMR 17:125-136.
  193. Andrec M, Levy RM (2002): Protein sequential resonance assignments by combinatorial enumeration using 13C alpha chemical shifts and their (i, sequential connectivities. J Biomol NMR 23:263-270.
  194. Coggins BE, Zhou P (2003): PACES: Protein sequential assignment by computer-assisted exhaustive search. J Biomol NMR 26:93-111.
  195. Jung Y.-S, Zweckstetter M (2004): Mars: robust automatic backbone assignment of proteins. J Biomol NMR 30:11-23.
  196. Gronwald W, Willard L, Jellard T, Boyko RF, Rajarathnam K, Wishart DS, Sonnichsen FD, Sykes BD (1998): CAMRA: chemical shift based computer aided protein NMR assignments. J Biomol NMR 12:395-405.
  197. Zimmerman DE, Kulikowski CA, Wang L, Lyons BA, Montelione GT (1994): Automated sequencing of amino acid spin systems in proteins using multidimensional HCC(CO)NH-TOCSY spectros- copy and constraint propagation methods from artificial intelligence. J Biomol NMR 4:241-256.
  198. Li KB, Sanctuary BC (1997): Automated resonance assignment of proteins using heteronuclear 3D NMR. 2. Side chain and sequence-specific assignment. J Chem Inf Comput Sci 37:467-477.
  199. Hyberts SG, Wagner G (2003): IBIS: a tool for automated sequential assignment of protein spectra from triple resonance experiments. J Biomol NMR 26:335-344.
  200. Moseley HN, Monleon D, Montelione GT (2001): Automatic determination of protein backbone resonance assignments from triple resonance nuclear magnetic resonance data. Methods Enzymol 339:91-108.
  201. Eghbalnia HR, Bahrami A, Wang L, Assadi A, Markley JL (2005b): Probabilistic identification of spin systems and their assignments including coil-helix inference as output (PISTACHIO). J Biomol NMR 32:219-233.
  202. Lesk AM, Rose GD (1981): Folding units in globular proteins. Proc Natl Acad Sci USA 78:4304-4308.
  203. Przytycka T, Aurora R, Rose GD (1999): A protein taxonomy based on secondary structure. Nat Struct Biol 6:672-682.
  204. Reily MD, Thanabal V, Omecinsky DO (1992): Structure-induced carbon-13 chemical shifts: a sensitive measure of transient localized secondary structure in peptides. J Am Chem Soc 114:6251-6252.
  205. Wishart DS, Sykes BD (1994b): The 13C chemical-shift index: a simple method for the identification of protein secondary structure using 13C chemical-shift data. J Biomol NMR 4:171-180.
  206. Wang Y, Jardetzky O (2002): Probability-based protein secondary structure identification using combined NMR chemical-shift data. Prot Sci 11:852-861.
  207. Hung LH, Samudrala R (2003): Accurate and automated classification of protein secondary structure with PsiCSI. Prot Sci 12:288-295.
  208. Kabsch W, Sander C (1983): Dictionary of protein secondary structure: pattern recognition of hydrogen-bonded and geometrical features. Biopolymers 22:2577-2637.
  209. Judd K (2003): Chaotic-time-series reconstruction by the Bayesian paradigm: right results by wrong methods. Phys Rev E 67:026212.
  210. Cox RT (1946): Probability, frequency and reasonable expectation. Am J Phys 14:1-13.
  211. Jaynes ET (2003): Probability Theory: The Logic of Science. Cambridge: Cambridge University Press.
  212. Habeck M, Nilges M, Rieping W (2005): Bayesian inference applied to macromolecular structure determination. Phys Rev E 72:031912.
  213. Rieping W, Habeck M, Nilges M (2005): Inferential structure determination. Science 309:303-306.
  214. Nilges M, Habeck M, O’Donoghue SI, Rieping W (2006): Error distribution derived NOE distance restraints. Proteins 64:652-664.
  215. Berman HM, Henrick K, Nakamura H (2003): Announcing the worldwide protein data bank. Nature Structural Biology 10:980.
  216. Berman HM, Westbrook J, Feng Z, Gilliland G, Bhat TN, Weissig H, Shindyalov IN, Bourne PE (2000): The protein data bank. Nucleic Acids Res 28:235-242.
  217. Ulrich EL, Markley JL, Kyogoku Y (1989): Creation of a nuclear magnetic resonance data repository and literature database. Prot Seq Data Anal 2:23-37.
  218. Seavey BR, Farr EA, Westler WM, Markley JL (1991): A relational database for sequence-specific protein NMR data. J Biomol NMR 1:217-236.
  219. Fogh R, Ionides J, Ulrich E, Boucher W, Vranken W, Linge JP, Habeck M, Rieping W, Bhat TN, Westbrook J, et al., (2002): The CCPN project: an interim report on a data model for the NMR community. Nat Struct Biol 9:416-418.
  220. Hall SR (1991): The STAR file: a new format for electronic data transfer and archiving. J Chem Inform Comput Sci 31:326-333.
  221. Hall SR, Spadaccini N (1994): The STAR file: detailed specifications. J Chem Inform Comput Sci 34:505-508.
  222. Doreleijers JF, Nederveen AJ, Vranken W, Lin J, Bonvin AM, Kaptein R, Markley JL, Ulrich EL (2005): BioMagResBank databases DOCR and FRED containing converted and filtered sets of experimental NMR restraints and coordinates from over 500 protein PDB structures. J Biomol NMR 32:1-12.
  223. Wishart DS, Watson MS, Boyko RF, Sykes BD (1997): Automated 1H and 13C chemical shift prediction using the BioMagResBank. J Biomol NMR 10:329-336.
  224. Wang L, Eghbalnia HR, Markley JL (2006): Probabilistic approach to determining unbiased random- coil carbon-13 chemical shift values from the protein chemical shift database. J Biomol NMR 35:155-165.
  225. Palmer III AG Kroenke CD, Loria JP (2001): Nuclear magnetic resonance methods for quantifying microsecond-to-millisecond motions in biological macromolecules. Methods Enzymol 339: 204-238.
  226. Sanders II CR Schaff JE, Prestegard JH (1993): Orientational behavior of phosphatidylcholine bilayers in the presence of aromatic amphiphiles and a magnetic field. Biophys J 64:10691080.
  227. Ulrich EL, Akutsu H, Doreleijers JF, Harano Y, Ioannidis YE, Lin J, Livny M, Mading S, Maziuk D, Miller Z, Nakatani E, Schulte CF, Tolmie DE, Wenger RK, Yao H, Markley J (2008): BioMagResBank. Nucleic Acids Res 36:D402-D408.