Report

Resolution: Implications in Refinement Swanand Gore & Gerard Kleywegt May 6th 2010, 12-1 pm Macromolecular Crystallography Course Outline • Intuitive idea of resolution – why higher order diffraction is better. • Parameters, model, observations, refinement – more data is better. • Observations, parameters, over-fitting in crystallographic refinement. • Features that can be modeled at various resolutions. • Refinement practices at low and high resolution. Idealized diffraction in 1D h3 h2 h1 -h1 -h2 -h3 • Images scanned from David Blow’s book Idealized diffraction in 1D h3 h2 Assuming: •B = 0 •Occupancy = 1 •Uniform scattering power in all directions. •Phase angles = 0 h1 -h1 -h2 Increasing resolution -h3 • Images scanned from David Blow’s book Idealized diffraction in 1D • • • • • • Higher order diffraction Higher Fourier coefficients Higher frequency wave in real space Sharper signal Greater resolution Images scanned from David Blow’s book What separation can be resolved? • Nominal resolution O O – The h-th order diffracted wave samples the lattice at interval of a/h. – a/h is the crystallographic resolution which is routinely quoted. – In tetragonal cell abc, diffraction hkl comes from planes separated by • √[ (a/h)2 + (b/k)2 + (c/l)2 ] – For tetragonal cell 100, 95, 90, and highest order diffraction 50, 52, 48, resolution is ~3.29. • • For non-orthogonal axes, corrections apply. Resolution intuitively means the least distance between objects below which they cannot be distinguished apart. – For 3D crystallography, it is ~0.92*dmin, almost same as nominal. • • Images from B. Rupp’s book Image from Gerard’s ppt. blob Atomic scatterers in 1D Resolution Filter Peaks get sharper as higher resolution Fourier coefficients are included. Fourier Coefficients & Phases C • Images from B. Rupp’s book O C O C O Occupancy and B factors Peaks get broader due to larger B factors and shorter due to lower occupancy. • Images from B. Rupp’s book Data truncation Happens naturally due to B factors. Truncated data leads to incomplete reverse FT, causes ripples. Ripples around heavy atoms can ‘drown’ nearby lighter atoms. Ripples can seem to originate from real atoms. • Images from B. Rupp’s book N’s at 0.5 occupancy? O at 0.5 occupancy? Diffracting duck in 2D • • • Leaving out higher order diffraction data will reduce the detail retrieved through reverse transform. Leaving out lower resolution data will blur the boundaries. Randomly absent data is not too problematic for maps. – • Images from Kevin Cowtan’s website. Doesn’t matter if Rfree set is / not used in map calc? Make everything as simple as possible, but not simpler….. Fh exp(i φh) =V Σ fi oi exp(2πi h.x) exp (-Bi sin2θ/λ2) Noise. Errors in data collection. Static, dynamic disorder. Estimate phases. Model xyz, B, o. Model solvent. ….. ρx = 1/V Σ Fh exp (-2πi h.x + iφh) • Images from B. Rupp’s book Choosing resolution cutoff • B factors and scattering factors impose a natural cutoff on what can be observed. • Reliability of measurement is indicated by S/N ratio and completeness • Signal to noise ratio – – • Completeness – – – – • Images from B. Rupp’s book Low completeness in highest resolution shell does not confer a level of detail to the map as implied by nominal resolution Effective resolution = dmin . C-1/3 Randomly or systematically missing data creates undesirable effects in reverse FT. Completeness > 0.95 Number of reflections increases as cube of nominal resolution. – – • A low SNR does not matter too much if proper maximum likelihood target is used to weigh in error estimates. High <I/σ(I)> matters when collecting data for phasing. 2/3z π VUC / dmin3 Not unique due to centro-symmetry and spacegroup symmetry Model and refinement • • Model is defined as a set of parameters and a set of functions over parameters, designed to explain observations Observations Refinement – Is an algorithmic process of fitting a model to explain observations, by assigning optimal values to parameters. – Reduces the differences between observations and model-calculated values of observations • A linear model in 2D consists of 2 parameters – Y = mX + c • Some models are more accurate than others, depending on quality of refinement. • Refinement is necessary when observations contain errors and there are enough observations to refine the parameters. Well-refined model 1 c m Ill-refined model Model and refinement • A linear model in 2D – – consists of 2 parameters : Y = mX + c 1 observation, howsoever accurate, is not sufficient if model has 2 parameters • • – 2 distinct accurate observations are sufficient to determine the linear model – – 3 accurate observations over-determine the model But observations generally contain random error! Greater number of observations lead to error cancellation and more accurate model • • • Fitting to error too! Quality of modelling – – • Well-determined model Model with too few params can lead to under-fitting Model with too many params can lead to over-fitting – • Under-determined, over-fitted model Many models can be imagined Choice of model (linear, quadratic, higher polynomial?) Quality of refinement (R value) Images from B. Rupp’s book M1 Model and refinement • M2 In presence of errors, refinement quality does not indicate model quality – Well-refined model is of bad quality if it was fitted to erroneous observations. MC • Hence, observations not subject to refinement are required to assess the accuracy. – – – – • R and “free” R M1: 0.2, 0.3 M2: 0.2, 0.4 M1 better than M2 Free R and data/param ratio helps in comparing models with different number of parameters – MA: 0.2, 0.3. d/p = 15/2 = 7.5 • Under-fit – MB: 0.1, 0.2. d/p = 15/3 = 5 • optimal – MC: 0.01, 0.25. d/p = 15/10 = 1.5 • • Overfitting = Low d/p, high Rfree Images from B. Rupp’s book MA MB Occham’s valley A crystallographic model • Biochemical entities – Biopolymers • polypeptides, polynucleotides, carbohydrates – Small-molecule ligands (ions, organic) • • • Crystallographic additives, e.g. GOL, PEG Physiologically relevant, e.g. heme, ions Synthesized molecules, e.g. a drug candidate – Solvent • Coordinates, Displacement – – – – • Unique x,y,z Partial, multiple, absent (occupancy) Isotropic or anisotropic B factors TLS approximation Crystallographic etc. – Cell, symmetry, NCS – Bulk solvent correction (Ksol, Bsol) • • • • 3hbq images made with pymol. http://www.cgl.ucsf.edu/chimera/feature_highlights/ellipsoids.png B factor putty from Antonyuk et al. 10.1073/pnas.0809170106 www.ruppweb.org/xray/tutorial/Crystal_sym.htm Quick note on NCS, TLS • Non-crystallograpic symmetry – – – – – • Translation-libration-screw – – – – – – • Molecule/s -> ASU -> locally-related ASUs -> Unitcell -> Crystal Sometimes ASU can consist of multiple, nearly identical subunits. The transformation operator between subunits is local and distinct from space-group operators. Subunits need not be identical because they are in different environments, differences do not indicate problems! This additional symmetry can be used in refinement (restraints, constraints) and validation. Overall anisotropy = lattice disorder + inter-molecular motions + intramolecular rigid body motions within molecule + atomic anisotropy Paradigm shift from atom-level anisotropy modelling to anisotropic movements of rigid bodies 1d: a point (3) through which rotation axis (2) will pass + ratio (1) of rotation to translation on that axis = 6 2d: 2 points + 2 ratios + 2 orthogonal axes (3) + 2 more ratios = 13 3d: 3 points + 3 ratios + 3 orthogonal axes (3) + 6 more screws = 20(ish) TLS group granularity can range from full domain to sidechain Images from Rupp book and Martyn Winn ppt Counting parameters • Average-case parameters – Per atom 4 params • • 3 params for coordinates 1 param for isotropic B factor – No hydrogens, 1 water per residue – 8 atoms per residue – N * 8 * 4 = 32 N • Increasing the parameters – – – – • 6 params per atom for anisotropic B factor (>2x) Refining occupancy (1.25x) or multiple occupancy Hydrogens modeled explicitly (8 per residue) (2x) Multiple models (M x) Reducing parameters – 20 params per TLS group • • 5 groups: 20 * 5 groups of 40 res each = 100 => 32 * 200 to 100 (1/64 x for 200 res protein) – Strict NCS (1/n x for n-fold) Restraint counts taken from: http://ccp4wiki.org/~ccp4wiki/wiki/images/9/9f/Winn_prague09_data_parameters.pdf 1clm, calmodulin, 1.8Å 1132 protein atoms + 4 Ca + 71 waters = 4828 xyzB #unique reflections = 10610 Data / params = 2.2 1exr, calmodulin, 1Å 1467 protein atoms with alt conf + 5 Ca + 178 waters 9900 anisotropic B + 316 occupancy = 15166 params #unique reflections = 77150 Data / params = 4.6 1h6v , 3Å 6 TLS groups = 120 params 22514 protein atoms + 552 ligand atoms + 9 waters xyzB = 92300 (residual) #unique reflections = 69328 (5% free) d/p = 69328/92300 = 0.7 Data to parameters ratio • r = (number of unique reflections) / (number of parameters) – Graph for a calmodulin 1up5, ~2500 atoms, xyzB – r < 1, i.e. under-determined for dmin < 2.5Å – Reflections-based refinement is possible only for r > 10, i.e. resolution approaching 1Å! – But most PDB entries have r ~ 2-5 • There must be more observations provided to refinement than only the reflections – Reflections = observations specific to a particular MX experiment – But there are other more general observations applicable to any MX refinement – Covalent geometry, steric clashes, …. • Image from B. Rupp’s book (Graph by Konrad Hinsen, 2008) Observations to parameters ratio • Observations = reflections + constraints and restraints based on well-known features of macromolecules • o/p > d/p • Bonds, angles, planarity, chirality… – o/p = (22732 + 77150) / 15166 = 6.1 > 4.6 = d/p Hangman Bungee jumper RElaxation = REstraint CONvict = CONstraint Energy – Tricky to estimate the difference due to dependences, but generally sufficient to make refinement possible – 1exr: 1Å, 22732 restraints length • • Images from Gerard’s slides Restraint counts taken from: http://ccp4wiki.org/~ccp4wiki/wiki/images/9/9f/Winn_prague09_data_parameters.pdf Observations to parameters ratio • o/p > d/p for 1h6v at 3Å – Restraints (including NCS) = 209378 – o/p = (209378+69328)/92300 = 3 – d/p = 0.7 < 3 = o/p • 2 components of refinement residuals – Data-based • Changes model (xyzB..) to reduce Fo ~ Fc – Knowledge-based • Changes model (xyz) to take values of geometric features towards idealized values – Qtot = wx Qx + Qgeom – Small wx : greater stress on geometric correctness • Low resolution, low d/p – Large wx : model deviation from ideal geometry • High resolution, high d/p • Restraint counts taken from: http://ccp4wiki.org/~ccp4wiki/wiki/images/9/9f/Winn_prague09_data_parameters.pdf Greater d/p => more detail (given decent phases) 0.95Å • • Image from http://www.crystal.uwa.edu.au/px/alice/projects/SCOA_atomic.html, 1mxt Images from Rupp’s book Lower d/p => lower detail decent phases often not available 2g34, 5Å • Pics of 2g34, 1z56 with coot using EDS maps 1z56, 3.9Å Lower d/p => lower detail 2bf1, 4Å • Pics of 2bf1 with coot using EDS maps All resolutions not equal… • From Gerard’s slides and Phil Evans Levels of detail interpretable at various resolutions Protein Feature Resolution (Å) Nucleic Acid Feature Resolution (Å) Helix 9 Double helix 20 Sheet 4 Single strand 12 Main chain 3.7 Stacked base pairs 4 Aromatic sidechains 3.5 Phosphates 3.5 Small sidechains 3.2 Purine or pyrimidine? 3.2 Sidechain conformations 2.9 Individual bases 2.7 Carbonyl, peptide 2.7 Ribose pucker 2.4 Ordered waters 2.7 Individual atoms 1.5 Central dimple of aromatic ring 2.4 Correct stereochemistry at Ile CB 2.2 Proline pucker 2.0 Individual atoms 1.5 Orbitals and bonds (beyond 1Å)! • From David Blow’s book Rules of thumb at all resolutions for model-building and refinement • Start with few parameters and slowly enrich the model • Be very conservative till a majority of backbone is identified and produces stable refinement • Prioritize: Backbone > side-chains > small-mols > waters • Be aware of prevalent modeling practices at your resolution • Whole model contributes to quality of region of interest. • Use similar structures for comparison and copying. • Use quality criteria often. Low resolution refinement • Low resolution structures offer great biological insights. – Mainly for complexes e.g. 70S ribosome at 7Å, SIV gp120 envelope glycoprotein at 4Å – Large complexes generally diffract to lower resolution. • • Components may have physiologically relevant conformations only in complexed states. High impact – In absence of better resolution, low resolution data must be used. – Low resolution does not have to mean low quality! • Basic guidelines for model building and refinement. – – – – • Low d/p => Be cautious of biasing the model Make extensive use of information in addition to reflections Use as few parameters as possible Increase params only when confident Images from Karmali et al. Acta Cryst. 2009. Low resolution refinement • Build model with fewer parameters – Mainchain-only model – Constrain B factor values to be isotropic and constant. – Full occupancies only. – TLS to model anisotropic motions of rigid domains. – Strictly constrained or restrained NCS to reduce params many-fold – No waters or small molecules, use only ‘bulk solvent’ Low resolution refinement • Model cautiously – Initial tracing • Build regions that are likely to be seen clearly • Good packing, low B factors, bulky group, electron-rich groups • core, mainchain, helices, big sidechains, bases, phosphates – Sequence registry • Beware of register and topology errors • Guess sequence register from bulky sidechains • Extend the register by trial and error • Check sequence register with a homologous structure • Truncate to Gly wherever unsure of residue identity • From Gerard’s slides Low resolution refinement – Try copying fragments from other high resolution structures when there is clear homology – Treat ligands extra-carefully • Copy high-quality observed conformation or predicted low energy conformation • Restrain tightly unless there is density and other clues to deviate • Axel T. Brunger et al. 2009. Acta Cryst D 65 128–133 X-ray structure determination at low resolution. Low resolution refinement density modification tools • Expected solvent density – define solvent boundary – followed by solvent flattening / flipping, histogram matching • • • Images B. Rupp’s book and from Acta Cryst. (2003). D59, 1881-1890. The phase problem. G. Taylor Brunger 2006, Low resolution crystallography. Acta Cryst. https://wasatch.biochem.utah.edu/chris/tutorial/Density_Modification.pdf a ppt on DM Low resolution refinement density modification • Averaging maps of NCS-restrained copies • • • Image from B. Rupp’s Brook. unger 2006, Low resolution crystallography. Acta Cryst. https://wasatch.biochem.utah.edu/chris/tutorial/Density_Modification.pdf a ppt on DM Low resolution refinement density modification • B-factor sharpening – High-resolution reflections get attenuated most by B factors – Application of negative B factors can artificially up-weigh high-res terms to obtain greater detailed but possibly noisier map • • Brunger 2006, Low resolution crystallography. Acta Cryst. https://wasatch.biochem.utah.edu/chris/tutorial/Density_Modification.pdf a ppt on DM Low resolution refinement • Refinement techniques – Rigid body refinement • A fragment is constrained to be internally rigid, has only 6 degrees of freedom • B factor is isotropic and constant • Powerful first step of refinement needing only low resolution data • Arbitrary rigid fragments (high quality helices, highresolution domain structures) can be optimized for location and orientation relative to each other to yield better phases and maps – Torsion angle refinement • Bonds, angles, chirality, planarity not variables, only torsion angles are refined • Protein is divided into rigid subgroups to sample thoroughly a limited conformational space • Higher radius of convergence, reduced overfitting • • • Image from Schwieters, C.D. & Clore, G.M. (2001) Internal coordinates for molecular dynamics and minimization in structure determination and refinement. J. Magn. Reson. 152, 288-302 Nice tutorial at http://speedy.st-and.ac.uk/~naismith/workshop/torsion.pdf See Axel Brunger’s papers on torsion angle refinement Low resolution refinement • Solving multiple times – – – – • Gradual increase in number of parameters – – • Keep track of Ramachandran and sidechains rotamers Remove unlikely parts of mainchain and sidechain Do not restrain Rama distribution or sidechains to rotamers during refinement, it may give false validation results Read what others are doing for low resolution – • Mainchain -> bulky sidechains -> sequence register -> other sidechains Finally known small mol binders with known binding site can be modelled if reasonable density appears Validation – – – • Try to automate as much as possible the process of model building and refinement, and then repeat it Consensus substructures are more reliable, average them Regions with differences are unreliable, remove them Gives an idea of precision e.g. Axel Brunger’s literature, CCP4 & phenix tools, CCP4bb Images from wikipedia and Furnham et al. Structure 2006. High resolution refinement • High resolution structures provide atomic insights – – – – Packing, binding Flexibility Enzyme mechanisms Hydration • Basic guidelines for model building and refinement – High d/p => Be cautious of under-fitting! – Make greater use of data than in low res case – Make as detailed a model as possible, esp of interesting regions – Check all empty density critically High resolution refinement • Allow model to deviate from geometry when data is strong – Weight on xray term can be slowly increased to reveal any unusual geometry without risking model bias • Use automation to fit biopolymers – Trace secondary structure automatically, in coot or with phenix tools – Trace mainchain and build sidechains using programs, e.g. with buccaneer, warpNtrace, Rapper – Do this multiple times to identify regions requiring manual attention • • Validation tools: can they indicate the information content of macromolecular crystal structures? EJ Dodson et al. Volume 6, Issue 6, 1998, 685-690. Image from Terwillinger et al. papers in Acta Cryst D on automatic chain tracing. High resolution refinement • Explain all unoccupied density – Is it due to ligands? • Build expected ligands (including MX additives) • Search unexpected small-mols • E.g. coot or phenix ligand tools – Is it due to multi-conformer sidechains? – Is it water? • Images from B. Rupp’s book and Terwillinger et al. Acta Cryst. 2005. High resolution refinement • Build waters – Peak-pick semi-automatically to form a reasonable hydration network with sidechains • Model hydrogens – When difference density is visible • • Image from B. Rupp’s book Atomic resolution crystallography reveals how changes in pH shape the protein microenvironment. Lyubimov et al. Nature Chemical Biology 2, 259 - 264 (2006) High resolution refinement • Verify correct sidechain orientations of NQH – Manually or automatically flip NQH sidechains to improve hbonding – Model more sidechain conformations if necessary • Use non-standard atomic scattering models – At subatomic resolution, model electron density with nonspherical multipolar model, or model bonds as scatterers • • • Image from B. Rupp’s book Afonine et al. Acta Cryst. (2007). D63, 1194–1197 Jelsch et al. PNAS 2000 97 7 3171. High resolution refinement • Even in high res, maintain order of adding detail to avoid overfitting – bb > sc > ligand – Anisotropy, multiconformers, waters, hydrogens •Image from Antonyuk S V et al. PNAS 2005;102:12041-12046 •Image from David Blow’s book. • Invest more parameters around the regions of interest – – – – – multi-conformers Anisotropy waters near active site Possibility of multiple ligands Releasing constraints / restraints Summary • Resolution is the least distance between Bragg planes with observable reflection. Two atoms closer than resolution cannot be observed distinctly using data at that resolution. • Resolution dictates the detail revealed by electron density maps. – Low resolution => low detail – High resolution => high detail • Parameters in the model must be chosen to suit the resolution. • Over-fitting can be detected using Rfree and data to parameter ratio. • Knowledge-based constraints and restraints augment experimental data to make refinement possible. • Geometric target is weighted more than crystallographic data at low resolution. Model is allowed to diverge from ideal geometry at high resolution. • Greater detail should be modelled at higher resolution to make best use of data. Acknowledgements • Alejandro & IPMont MX organizers • Sameer Velankar, Jawahar Swaminathan (EBI) • Online resources – – – – – Kevin Cowtan Rupp web Randy Read’s course Various papers and images therefrom Martyn Winn’s ppt at on data to params • Books – David Blow – Alex McPherson – Bernhard Rupp