Comparing Protein Structures Why? detect evolutionary relationships identify recurring motifs detect structure/function relationships predict function assess predicted structures classify structures - used for many purposes 7.91 Amy Keating Algorithms for detecting structure similarity Dynamic Programming - works on 1D strings - reduce problem to this - can’t accommodate topological changes - example: Secondary Structure Alignment Program (SSAP) 3D Comparison/Clustering - identify secondary structure elements or fragments - look for a similar arrangement of these between different structures - allows for different topology, large insertions - example: Vector Alignment Search Tool (VAST) Distance Matrix - identify contact patterns of groups that are close together - compare these for different structures - fast, insensitive to insertions - example: Distance ALIgnment Tool (DALI) Unit vector RMS - map structure to sphere of vectors - minimize the difference between spheres - fast, insensitive to outliers - example: Matching Molecular Models Obtained from Theory (MAMMOTH) SSAP - Structure and Sequence Alignment Program How about using dynamic programming? Any problems here? Taylor & Orengo JMB (1989) 208, 1-22 SSAP - Structure and Sequence Alignment Program How about using dynamic programming? Any problems here? 1. How will you evaluate if two positions are similar? Residue type expose to solvent secondary structure relationship to other atoms 2. Score that you give to an alignment of 2 residues depends on other residues ALIGNMENT depends on SUPERPOSITION but SUPERPOSITION depends on ALIGNMENT Taylor, WR, and CA Orengo. "Protein Structure Alignment." J Mol Biol. 208, no. 1 (5 July 1989): 1-22. SSAP - Structure and Sequence Alignment Program For each pair of residues, (i,j), assume their equivalence. How similar are their environments wrt other residues? i C A M G G k H S H R R V F E CV s ik = Σa/(|d ij -d kl | + b); so s is large if d ij and d kl are similar. Which j and l should you compare with each other? Images adapted from Taylor, WR, and CA Orengo. "Protein Structure Alignment." J Mol Biol. 208, no. 1 (5 July 1989): 1-22. Answer: use the j’s and l’s that give the best score Vectors from atom k to: i k H S E H R R V F C A M G G V Q H S E R R H V F 12 2 3 1 1 10 1 0 2 1 0 1 23 1 0 1 7 4 1 0 2 14 1 0 1 25 G Q V e ctors from atom i to: V G M A C NOTE: this gives an ALIGNMENT of how the residues of sequence A align with those of sequence B, when viewed from the perspective of i and k. BUT, which i’s and k’s should you compare? ALL OF THEM! Then combine the results and take a consensus via another round of dynamic programming = “double dynamic programming” Vectors from k = F Vectors from i = C Vectors from i = C 12 2 3 1 1 10 1 0 2 1 0 1 23 1 0 1 7 4 1 0 2 14 1 0 1 25 Protein A Protein B 28 21 10 4 27 12 15 14 25 2 5 Vectors from k = V 16 1 2 1 21 1 1 1 4 0 0 5 4 1 1 4 5 1 1 2 15 1 0 1 25 1 Instead of using distances, use vectors to include some directionality s ij = a/(|d ij -d kl | + b); s ij = a/(|V ij - V kl | + b); Can also include other information about residues i and k if desired (e.g. sequence or environment information) s ij = (a + F(i,k)/(|V ij - V kl | + b); It is important to assess whether detected similarities are SIGNIFICANT. Various statistical criteria have been used. General idea: How “surprising” is the discovery of a shared structure? Structural Classification of Proteins ? Structure vs. structure comparisons (e.g. using DALI) reveal related groups of proteins ? Structurally-similar proteins with detectable sequence homology are assumed to be evolutionarily related ? Similarities between non-homologous proteins suggest convergent evolution to a favorable or useful fold ? A number of different groups have proposed classification schemes – SCOP (by hand) – CATH (uses SSAP) –FSSP (uses Dali) Structural statistics from August, 2003 Classification Of Proteins 7 CLASSES (a,b,a/b,a+b…) 800 FOLDS domain structures 1,294 SUPERFAMILIES possible evolutionary relationship 2,327 FAMILIES strong sequence homology 54,745 DOMAINS Murzin, AG, SE Brenner, T Hubbard, and C Chothia. "SCOP: A Structural Classification of Proteins Database for the Investigation of Sequences and Structures." J Mol Biol. 247, no. 4 (7 April 1995): 536-40. Structural Classification Of Proteins 7 CLASSES (a,b,a/b,a+b…) 800 FOLDS domain structures 1,294 SUPERFAMILIES possible evolutionary relationship 2,327 FAMILIES strong sequence homology 54,745 DOMAINS all alpha all beta alpha/beta alpha + beta multi-domain membrane small coiled-coil low-resolution peptide designed Structural Classification Of Proteins 7 CLASSES (a,b,a/b,a+b…) 800 FOLDS same secondary structure elements, same order, same connectivity domain structures 1,294 SUPERFAMILIES possible evolutionary relationship 2,327 FAMILIES strong sequence homology 54,745 DOMAINS PDB Growth in New Folds structures submitted per year; new folds per year (note that PDB criteria for a new fold differ from SCOP) Structural Classification Of Proteins 7 CLASSES (a,b,a/b,a+b…) Low sequence identity, but probable evolutionary relationship (e.g. based on 800 FOLDS domain structures structure or function) 1,294 SUPERFAMILIES possible evolutionary relationship 2,327 FAMILIES strong sequence homology 54,745 DOMAINS Structural Classification Of Proteins 7 CLASSES (a,b,a/b,a+b…) 800 FOLDS domain structures 1,294 SUPERFAMILIES possible evolutionary relationship 2,327 FAMILIES Clear evolutionary relationship; strong sequence homology often sequence identity > 30% 54,745 DOMAINS Structural Classification Of Proteins 7 CLASSES (a,b,a/b,a+b…) 800 FOLDS domain structures 1,294 SUPERFAMILIES possible evolutionary relationship 2,327 FAMILIES strong sequence homology Autonomously-folding unit of 54,745 DOMAINS compact structure scop.mrc-lmb.cam.ac.uk/scop/index.html LCK kinase and p38 Map kinase in same family Wasn’t true last year! CATH classification CLASS ARCHITECTURE TOPOLOGY (fold) HOMOLOGY Courtesy of Christine Orengo. Used with permission. A few folds are highly-populated! Five folds in CATH contain 20% of all homologous superfamilies Courtesy of Christine Orengo. Used with permission. Some fold types are multi-functional “superfolds” with > 3 functions Courtesy of Christine Orengo. Used with permission. SCOP entry: Use RASMOL to view the structures for ubiquitin and ferredoxin… 11% sequence identity DALI superposition Ubiquitin [MEDLINE: 91274342], PUB00000768, PUB00005320 is a protein of seventy six amino acid residues, found in all eukaryotic cells and whose sequence is extremely well conserved from protozoan to vertebrates. It plays a key role in a variety of cellular processes, such as ATP-dependent selective degradation of cellular proteins, maintenance of chromatin structure, regulation of gene expression, stress response and ribosome biogenesis. Ubiquitin is a globular protein, the last four C-terminal residues (Leu-Arg-Gly-Gly) extending from the compact structure to form a 'tail', important for its function. The latter is mediated by the covalent conjugation of ubiquitin to target proteins, by an isopeptide linkage between the C-terminal glycine and the epsilon amino group of lysine residues in the target proteins. The ferredoxins are iron-sulfur proteins that transfer electrons in a widevariety of metabolic reactions. They have a cofactor which binds a 2FE-2S cluster. Ferredoxins can be divided into several subgroups depending upon the physiological nature of the iron sulfur cluster(s) and according to sequence similarities IPR000564. Pfam annotations Molecular Modeling: Methods & Applications Acknowledgement: The following materials were prepared by or with the help of Professor Bruce Tidor. How do we use computational methods to analyze , predict , or design protein sequences and structures? Theme: Methods based on physics vs. methods based on our accumulated empirical knowledge of protein properties Example: Design of Disulfide-Stabilized Proteins 2 wild-type 2 Cys 1 disulfide residues mutations bond Approach 1: learn from sequence If you only have a protein sequence, can you identify isolated Cys residues versus those that are involved in disulfide bonds? Training Set = with “correct outputs” Input Learning Algorithm Why did it work? (or not?) database of inputs Train a learning algorithm Correct Output Dissect trained learning algorithm tune method Muskal, SM, SR Holbrook, and SH Kim. "Prediction of The Disulfide-bonding State of Cysteine in Proteins." Protein Eng. 3, no. 8 (August 1990): 667-72. Results for Approach 1 ? Input: Protein sequence flanking Cys residues (±5) ? Learning algorithm: Neural network ? Predictive success: ~80% ? Implies that Cys-bond formation is largely influenced by local sequence ? Analysis of trained network weights ? Hydrophilic local sequence increases propensity for disulfide bonded structure ? Hydrophobic local sequence increases propensity for isolated sulfhydryl ? Shows interesting difference between Phe and Trp vs. Tyr ? Drawback: don’t learn which Cys residues are paired! Approach 2: Database Driven ? Start with database of known disulfide bond geometries from the PDB ? For target protein structure, search over all pairs of residues ? Try all disulfide bond geometries from database for compatibility with this pair of positions ? Record any compatible disulfides ? Report successful pairs of residues ? Result: successful introduction of S-S bond into l repressor -> more stable protein, still binds DNA Pabo, CO, and EG Suchanek. "Computer-aided Model-building Strategies for Protein Design." Biochemistry 25, no. 20 (7 October 1986): 5987-91. Approach 3: Energy-Function Based ? For our protein structure, search over all pairs of residues ? Build a model of the Cβ and Sγ atoms and determine if these are compatible with a disulfide bond in this geometry ? If so, build lowest energy disulfide between this pair of residues ? Evaluate energy of this disulfide with some energy function ? Report successful pairs of residues ? Succeeds in predicting the geometry of many known disulfide bonds Hazes, B, and BW Dijkstra. "Model Building of Disulfide Bonds in Proteins with known Three-dimensional Structure." Protein Eng. 2, no. 2 (July 1988): 119-25. Pros and Cons of the Different Approaches ? Machine-learning methods often don’t provide a clear understanding of why they worked ? There are obvious structural constraints on disulfide bonds, and sequence-based methods may not be able to capture these ? Structure data isn’t always available, so sequence-based methods can be valuable ? Databases of known disulfides may be incomplete ? Disulfides might not be transferable to a different context ? When using a database, you don’t need to have an accurate description of the physics ? Methods based on first principles can identify things never seen before ? Our ability to model proteins from first principles is limited Does the model include structural relaxation? And one more caveat… How do disulfide bonds stabilize proteins? What if you want to compute how much the disulfide bond stabilizes the protein? wt u G? u G? folded mutation G? unfolded mutation G? ox mut, ox mut, wt unfolded folded ?? G = ? G u ? ? G u = ? G mutation ? ? G mutationu Energy-based modeling of protein structure and function ? CONFORMATIONAL ANALYSIS - what are the low-energy structures a protein can adopt? ? DYNAMICS - how do proteins move? ? THERMODYNAMICS - can compute quantities that characterize the system (e.g. enthalpy, entropy, heat capacity, free energy differences) ? ENERGY COMPONENTS - which atoms or which forces contribute the most to protein stability? ? REACTIVITY - what are the mechanisms and rates of reactions? Typically requires quantum mechanics. For a molecular simulation or model you need: 1. A representation of the protein 2. An energy function 3. A search algorithm or optimizer Levels of Representation Electrons: Residues: on or off a lattice Atoms: H O H Cα Cβ N H H O C H H All-atom protein, DNA, solvent, ligands, ions O H + + - + - Solvent as a high dielectric, protein as Continuum: a low-dielectric “glob” with charges inside Quantum mechanics describes the energy of a molecule in terms of a wavefunction describing the location and motion of nuclei and electrons in the molecule HΨ(r,R) = E?Ψ(r,R) Ψ(r,R) = Ψ(r)?Ψ(R) Born-Oppenheimer This can only be solved exactly for a small number of systems even the helium atom is too complex for an exact solution! It is much too expensive to compute the energies of proteins and DNA using quantum methods. Instead, we use empirical approximations that capture the important effects. For the most part, this is ok for the description of biological macromolecules at room temperatures. NOTE: once we ignore the electronic part of the wavefunction we can no longer compute the energy of bonds breaking and forming. Potential Energy Using Molecular Mechanics Goal: Describe potential energy of any conformation of molecule Use molecular mechanics: based on physics, but uses simplified “ball and spring” model. Think Newton, not Schroedinger! Model is EMPIRICALLY adjusted to capture quantum effects that give rise to bonding. Covalent 3 )( UU N += bonds become “springs” + - covalent -Non R U Covalent Potential Energy Terms U Covalent = U bond + U angle bond + U dihedral improper + U torsion b k ? b 0 )U bond = ∑ 1 b ( 2 bonds 2 2 U angle bond = ∑ 1 k θ (θ ? θ ) angles bond 2 0 U dihedral improper = ∑ 1 dihedrals improper 2 k Φ ( Φ ? Φ 0 ) U torsion = ∑ 1 k φ [1 + cos(n φ ? δ )] torsions 2 Brooks et al., J. Comput. Chem. 4: 187-217 (1983) 2 +++= torsiondihedralimproper anglebondbondCovalent UUUUU Key to Symbols: covalent terms k b , k θ , and k Φ are harmonic force constants for bond, bond angle, and improper dihedral terms, respectively. b 0 , θ 0 , and Φ 0 are equilibrium bond U bond = ∑ 1 b k ? b 0 ) b ( 2 lengths, bond angles, and improper bonds 2 dihedrals, respectively. U angle bond = ∑ 1 2 angles bond 2 k θ (θ ? θ 0 ) b , θ , and Φ are actual values for bond lengths, bond angles, and improper dihedrals, respectively, in this particular structure. U dihedral improper = ∑ 1 = ∑ 1 dihedrals improper 2 k Φ ( Φ ? Φ 0 ) 2 U torsion k φ [1 + cos(n φ ? δ )] torsions 2 k φ is the barrier height for an individual torsion, n is its “periodicity” (2-fold, 3-fold, etc.), δ is the position of the maximum, and φ is the value of this torsion in this particular structure. Non-Covalent Potential Energy Terms U Non -covalent = U vdW + U elec Lennard-Jones potential 12 ij ij r B 6 ij ij r C ? ? B ij U vdW = ∑ ? 12 ? C ij ? ? inull j ? r ij r ij 6 ? “accurate” approximate q i q j U elec = ∑ Coulomb’s law inull j εr ij += UUU elecvdWcovalent-Non Key to Symbols: non-covalent ? ? ? ? ? ? B ij C r ij is the distance between atom i and j, B ij and C ij are parameters ij ∑ U ? vdW 12 6 atom r ij r ij ? ? inull j describing the vdW function q q i j q i & q are the partial atomic charges j r ij on atoms i & ε U elec j, and ε is the effective i jnull dielectric constant. = ∑ = Partial atomic charges are used in Coulomb’s Law 0.35 H 0.35 N N N H H O O H H 0.35 0.35 -0.45 -0.45 0.5 -0.4 0.1 -0.6 0.36 -0.6 -0.16 0.3 These charges come from higher-level quantum calculations. Parameterization of the Potential k b , b 0 , k θ , θ 0 , k Φ , Φ 0 , k φ , n, δ, B q ij , C iji , – Must develop set of transferable parameters – Parameters obtained from fits to both experimental and theoretical data ? Much of data is from small molecules ? Crystal structures (lengths & angles, non-bonded coeffs.) ? Vibrational spectroscopy & ab initio QM calculations (q’s, k’s) ? Calorimetric & thermodynamic measurements (q’s, k’s) – Test parameters in context of entire protein ? Overriding assumption: Parameters for fragments of proteins are appropriate for that fragment in different contexts. “Missing Terms” in the Potential Function ? No hydrogen-bond term – Treated as part of electrostatics ? No hydrophobic term – Is resultant from all other forces Adding either of these would result in an imbalance in the potential due to double-counting. What about the solvent? The preceding energy function will give you the energy in the GAS PHASE. Not so useful for studying biology… Aqueous solvent is troublesome for two reasons: 1. There are LOTS of solvent molecules 2. The water has a strong influence on the electrostatic interactions Calculations that provide an accurate description of proteins or DNA in solvent are computationally demanding. Alternative Electrostatic Treatments Microscopic treatment Macroscopic treatment Coulomb’s Law OK ?Must include all Coulomb’s Law Not OK ?But can use 80=ε 41 ?=ε solvent atoms in sum Poisson–Boltzmann equation Simpler Representation ? More Complex Physics Slide courtesy of B. Tidor. Slide courtesy of B. Tidor. Protein Boundary ?Defined by contact surface with water probe Interior of Protein ?Atoms represented as fixed point charges ?Low dielectric constant (usually 1, 2, 3, or 4) Exterior of Protein ?No explicit solvent atoms ?Solvent water represented by high dielectric constant (80) ?Ionic strength treated with Debye-Hückel-type model Continuum Electrostatics Numerically solve the Poisson-Boltzmann equation on a grid )(4)](sinh[)()()( 2 rrrrr πρφκφε ?=??? E(r) = dielectric, f(r) = electrostatic potential, r(r ) = charge density, k is related to the ionic strength Slide courtesy of B. Tidor. Empirical solvation models (crude!) “roll” solvent over 1. Solvent-accessible surface area model surface to get area polar atoms are rewarded for exposure to solvent hydrophobic atoms are penalized E solvation = ∑ s? SA i atoms _ i This model doesn’t account for the fact that water screens (weakens) electrostatic interactions. Often used in combination with: 2. Distance-dependent dielectric model U q i q j elec = ∑ inull j ε(r) r ij Properties of Potential 3 N (R U ) = U bond + U angle bond + U dihedral improper + U torsion + U vdW + U elec scales as N (number of atoms) scales as N 2 ? Often implement some type of cutoff function to smoothly turn off non-covalent interactions beyond some distance IMPORTANT: parameterized only to give differences in energy for conformations - does not give energy of folding or free energy of formation! Must formulate your problem (with an appropriate reference state) so that you are considering energy differences.