Molecular Docking Ugur Sezerman Sabanci University What is docking? Docking is finding the binding geometry of two interacting molecules with known structures The two molecules (“Receptor” and “Ligand”) can be: - two proteins - a protein and a drug - a nucleic acid and a drug Two types of docking: - local docking: the binding site in the receptor is known, and docking refers to finding the position of the ligand in that binding site - global docking: the binding site is unknown. The search for the binding site and the position of the ligand in the binding site can then be performed sequentially or simulaneously What Are Docking & Scoring? • To place a ligand (small molecule) into the binding site of a receptor in the manners appropriate for optimal interactions with a receptor. • To evaluate the ligand-receptor interactions in a way that may discriminate the experimentally observed mode from others and estimate the binding affinity. complex ligand docking scoring receptor X-ray structure & DG … etc Why Do We Do Docking? • Drug discovery costs are too high: ~$800 millions, 8~14 years, ~10,000 compounds (DiMasi et al. 2003; Dickson & Gagnon 2004) • Drugs interact with their receptors in a highly specific and complementary manner. • Core of the target-based structure-based drug design (SBDD) for lead generation and optimization. Lead is a compound that – shows biological activity, – is novel, and – has the potential of being structurally modified for improved bioactivity, selectivity, and drugeability. Docking Applications • Determine the lowest free energy structures for the receptorligand complex • Search database and rank hits for lead generation • Calculate the differential binding of a ligand to two different macromolecular receptors • Study the geometry of a particular complex • Propose modification of a lead molecules to optimize potency or other properties • de novo design for lead generation • Library design Key aspects of docking • Scoring Functions – What are they? – Which Scoring Functions are feasible? • Search Methods – How do they work? – Which search method should I use? • Which program should I use? Docking Challenge • Both molecules are flexible and may alter each other’s structure as they interact: – Hundreds to thousands of degrees of freedom – Total possible conformations are astronomical Formulation of Docking Problem • A scoring function that can discriminate correct (experimentally observed) docking complex structure from incorrect ones • A search algorithm that finds the docking complex structure measured by the scoring function Formulation of Docking Problem Factors Affecting ∆G0 Intramolecular Forces(covalent) • Bond lengths • Bond angles • Dihedral angles Intermolecular Forces (noncovalent) • Electrostatics • Dipolar interactions • Hydrogen bonding • Hydrophobicity • Van der Waals Types of Docking Problems • Docking – Bound docking : the goal is to reproduce a known complex – Unbound docking : complex structure not known • Protein-Small Molecule Docking – Rigid receptor, rigid ligand – Rigid receptor, flexible ligand – Flexible receptor, flexible ligand Types of Docking Problems Docking strategies require: 1) Protein representation 2) A search method 3) Final refinement and scoring 1. Protein Structure • A 3-D structure of the target protein at atomic resolution must be available – Crystal and solution structures (PDB) – Homology models – Pseudoreceptor models • Ideally, the atomic resolution of crystal structures should be below 2.5 A • Even small changes in structure can drastically alter the outcome Receptor Structures & Binding Site Descriptions • PDB (Protein Data Bank, www.rcsb.org/pdb/) containing proteins or enzymes: – X-ray crystal: >60,000 structures,~10 % have ≤ 1.5 Å, ~80% between 1.5-2.5 Å – NMR:, ensemble accuracy of 0.4-1 Å in the backbone region, 1.5 Å in average side chain position (Billeter 1992; Clore et al. 1993) – (and high quality homology models built from highly similar sequences) • Limitation of experimental structures (Davis et al. 2003): – Locations of hydrogen atoms, water molecules, and metal ions – Identities and locations of some heavy atoms (e.g., ~1/6 of N/O of Asn & Gln, and N/C of His incorrectly assigned in PDB; up to 0.5 Å uncertainty in position) – Conformational flexibility of proteins • Binding site descriptions: atomic coordinates, surface, volume, points & distances, bond vectors, grid and various properties such as electrostatic potential, hydrophobic moment, polar, nonpolar, atom types, etc. DOCK Drug, Chemical & Structural Space • Drug-like: MDDR (MDL Drug Data Report) >147,000 entries, CMC (Comprehensive Medicinal Chemistry) >8,600 entries • Non-drug-like: ACD (Available Chemicals Directory) ~3 million entries • Literatures and databases, Beilstein (>8 million compounds), CAS & SciFinder • CSD (Cambridge Structural Database, www.ccdc.cam.ac.uk): ~3 million X-ray crystal structures for >264,000 different compounds and >128,00 organic structures • Available compounds – Available without exclusivity: various vendors (& ACD) – Available with limited exclusivity: Maybridge, Array, ChemDiv, WuXi Pharma, ChemExplorer, etc. • Corporate databases: a few millions in large pharma companies 3D Structural Information & Ligand Descriptions • 2D->3D software: CORINA, OMEGA, CONCORD, MM2/3, WIZARD, COBRA. (reviewed by Robertson et al. 2001) • CSD: <0.1 Å for small molecules, but may not be the bound conformation in the receptor • PDB: ligand-bound protein structures ~6000 entries • Atoms associated with inter-atom distances, physical and chemical properties, types, charges, pharmacophore, etc • Flexibility: conformation ensemble, fragment-based Scoring Functions • A fast and simplified estimation of binding energies scores <-> DGbinding DGbinding RT ln K affinity -scores DGcomplex/ solv DGligand/ solv DG protein/ solv DGinteraction TDS D X-ray structure ? configurations of the complex 3. Scoring Functions • Factors Affecting ∆G0 Intramolecular Forces(covalent) • Bond lengths • Bond angles • Dihedral angles Intermolecular Forces • Electrostatics • Dipolar interactions • Hydrogen bonding • Hydrophobicity • Van der Waals Types of Scoring Functions • Force field based: nonbonded interaction terms as the score, sometimes in combination with solvation terms • Empirical: multivariate regression methods to fit coefficients of physically motivated structural functions by using a training set of ligand-receptor complexes with measured binding affinity • Knowledge-based: statistical atom pair potentials derived from structural databases as the score • Other: scores and/or filters based on chemical properties, pharmacophore, contact, shape complementary • Consensus scoring functions approach 3. Scoring Functions Force Field Based • CHARMM [Brooks83] • AMBER [Cornell95] Empirical methods: • ChemScore [Eldridge97] • GlideScore [Friesner04] • AutoDock [Morris98] • AutoDock Vina[Trott09] Knowledge-based methods • PMF [Muegge99] • Bleep [Mitchell99] • DrugScore [Gohlke00] Force Field Based Scoring Functions Aij Bij qi q j E a b 332 Drij rij i 1 j 1 rij lig rec e.g. AMBER FF in DOCK • Advantages – FF terms are well studied and have some physical basis – Transferable, and fast when used on a pre-computed grid • Disadvantages – Only parts of the relevant energies, i.e., potential energies & sometimes enhanced by solvation or entropy terms – Electrostatics often overestimated, leading to systematic problems in ranking complexes Molecular mechanics force fields • Usually quantify the sum of two energies – the receptor–ligand interaction energy – internal ligand energy (such as steric strain induced by binding) • Interactions between ligand and receptor are most often described by using van der Waals and electrostatic energy terms. • Advantages – FF terms are well studied and have some physical basis – Transferable, and fast when used on a pre-computed grid • Disadvantages – Only parts of the relevant energies, i.e., potential energies & sometimes enhanced by solvation or entropy terms – Electrostatics often overestimated, leading to systematic problems in ranking complexes Molecular mechanics force fields • CHARMM [Brooks83] Molecular mechanics force fields • AMBER: [Cornell95] FF Scoring: Implementations • AMBER FF: DOCK, FLOG, AutoDOCK • CHARMm FF: CDOCK, MC-approach (Caflisch et al. 1997) • Potential Grid: rigid receptor structure upon docking. The grid-based score interpolates from eight surrounding grid points only. 100-fold speed up. Examples: DOCK, CDOCK, and many other docking programs. • Soften VDW: A soft-core vdw potential is needed for the kinetic accessibility of the binding site (Vieth et al. 1998). FLOG: 6-9 Lennard-Jones function; GOLD: 4-8 vdw + H-bond, and intraligand energy. • Solvent Effect on Electrostatic: often approximated by rescaling the in vacuo coulomb interactions by 1/D, where D = 1-80 or = n*r, n = 1-4, r = distance. • Solvation and Entropy Terms: Solvation terms decomposed into nonpolar and electrostatic contributions (e.g., DOCK): Ebind E nonbond E solv,elec E solv,np Empirical Scoring Functions DG DG0 DGrot N rot DGHB neutral _ Hbonds f DR, D DGio ionic _ int f DR, D DGaro aro _ int f DR, D DGlipo lipo.cont f DR, D LUDI & FlexX (Boehm 1994) • Goals: reproduce the experimental values of binding energies and with its global minimum directed to the X-ray crystal structure • Advantages: fast & direct estimation of binding affinity • Disadvantages –Only a few complexes with both accurate structures & binding energies known –Discrepancy in the binding affinities measured from different labs –Heavy dependence on the placement of hydrogen atoms –Heavy dependence of transferability on the training set –No effective penalty term for bad structures Empirical Scoring: Implementations Mostly differ by what training set and how many parameters are used • Cerius2/Insight2000: LUDI, ChemScore, PLP, LigScore • SYBYL: FlexX, F-Score • Hammerhead: 17 parameters for hydrophobic, polar complementary, entropy, solvation. sLOO = 1.0 logK for 34 complexes • VALIDATE: 8 parameters for VDW and Coulomb interactions, surface complementarity, lipophilicity, conformational entropy and enthalpy, lipophilic and hydrophilic complementarity between receptor and ligand surfaces • PRO_LEADS: 5 coefficients for lipophilic, metal-binding, H-bond, and a flexibility penalty term. sLOO = 2 kcal/mol for 82 complexes • SCORE (Tao & Lai, 2001); ChemScore (GOLD) • Knowledge-based Potentials of Mean Force Scoring Functions (PMF) Assumptions – An observed crystallographic complex represents the optimum placement of the ligand atoms relative to the receptor atoms – The Boltzmann hypothesis converts the frequencies of finding atom A of the ligand at a distance r from atom B of the receptor into an effective interaction energy between A and B as a function of r • Advantages – Similar to empirical, but more general (much more distance data than binding energy data) • Disadvantages – The Boltzmann hypothesis originates from the statistics of a spatially uniform liquid, while receptor-ligand complex is a two-component nonuniform medium – PMF are typically pair-wise, while the probability to find atoms A and B at a distance r is non-pairwise and depends also on surrounding atoms PMF: Implementations • Verkhivker et al.(1995): 12 atom pairs, 30 complexes (HIV-1 and simian immunodeficiency virus). Test on 7 other HIV-1 protease complexes • Wallqvist et al. (1995): 38 complexes, 21 atom types (10 C, 5 O, 5 N, 1 S). Test on 8 complexes sd=1.5 kcal/mol, and 20 complexes rmsd=1.0 A. DG pred ij ln Pij • Muegge et al. (1999): 697 complexes, 16 atom types from receptor & 34 from ligand, 282 statistically significant PMF interactions. Test on 77 diverse compounds: sd=1.8 i j log Ki. The PMF was combined with a vdw term to account for short-range ij interactions for DOCK4 docking: j r seg PMF _ score Aij r where Aij r k B T ln f Vol _ corr r ij ij bulk kl , r rcu to ff • DrugScore (Gohlke et al, 2000), FlexX, BLEEP Two Kinds of Search Systematic ✽ Exhaustive ✽ Deterministic ✽ Outcome is dependent on granularity of sampling ✽ Feasible only for low dimensional problems ✽ e.g. DOT (6D) Stochastic ✽ Random ✽ Outcome varies ✽ Must repeat the search to improve chances of success ✽ Feasible for bigger problems ✽ e.g. AutoDock Searching Algorithms • • • • • • • Systematic search Molecular dynamics Monte Carlo Simulations Simulated annealing Genetic algorithms Lamarckian Genetic Algorithm Incremental construction Systematic Search • Uniform sampling of search space – Relative position (3) – Relative orientation (3) – Rotatable bonds in ligand (n) – Rotatable bonds in protein (m) FRED [Yang04] Systematic Search • Uniform sampling of search space • Exhaustive, deterministic • Quality dependent on granularity of sampling • Feasible only for low-dimensional problems Example: search all rotations FRED [Yang04] Molecular Mechanics • Energy minimization: • Start from a random or specific state (position, orientation, conformation) • Move in direction indicated by derivatives of energy function • Stop when reach local minimum Monte Carlo Simulations • Tries to dock the ligand inside the receptor site through many random positions and rotations • In ICM and MCDOCK, this method is used to make random moves of the ligand inside a receptor binding site. • After each random move, a force-field based energy minimization is applied. • To avoid trapping in local minima, Monte Carlo combine this procedure with other search methods, such as Simulated Annealing, Genetic Algorithm and Lamarckian GA Simulated Annealing • Global optimization technique based on the Monte Carlo method : • Start from a random or specific state (position, orientation, conformation) • Make random state changes, accepting up-hill moves with probability dictated by “temperature” • Reduce temperature after each move • Stop after temperature gets very small Genetic Algorithm (GA) • Genetic search of parameter space: • Start with a random population of states • Perform random crossovers and mutations to make children • Select children with highest scores to populate next generation • Repeat for a number of iterations Gold [Jones95], AutoDock [Morris98] Lamarckian Genetic Algorithm • LGA finds lowest fitness function (energy) values first, then maps these values to their respective genotypes • Each new child is allowed to create a new generation • Genetic algorithm plus Solis and Wets local search Better performance than either simulated annealing or genetic algorithm alone Incremental Extension • Used in DOCK, FLEXX, FLOG and Surflex • Greedy fragment-based construction: • Partition ligand into fragments Incremental Extension • Greedy fragment-based construction: • Partition ligand into fragments • Place base fragment (e.g., with geometric hashing) Incremental Extension • Greedy fragment-based construction: • Partition ligand into fragments • Place base fragment (e.g., with geometric hashing) • Incrementally extend ligand by attaching fragments Descriptor Matching Methods: DOCK • Distance-compatibility graph in DOCK (Ewing and Kuntz 1997): distances between sphere centers and distances between ligand heavy atoms Descriptor Matching Methods • Distance-compatibility graph in DOCK (Ewing and Kuntz 1997): • distances between sphere centers and distances between ligand heavy atoms Interaction site matching in LUDI (Boehm 1992): HBA<->HBD, HYP<>HYP • Pose clustering and triplet matching in FlexX (Rarey et al. 1996): HBA<->HBD, HYP<->HYP • • • • Shape-matching in FRED (Openeye www.eyesopen.com) Vector matching in CAVEAT (Lauri and Bartlett 1994) Steric effects-matching in CLIX (Lawrence and Davis 1992) Shape chemical complementarity in SANDOCK (Burkhard et al. 1998) • Surface complementarity in LIGIN: (Sobolev et al. 1996) • H-bond matching in ADAM (Mizutani et al. 1994) Fragment-based Methods • Flexibility and/or de novo design • Identification and placement of the base/anchor fragment are very important • Energy optimization (during or post-docking) is important • Examples –Incremental construction in FlexX with triplet matching and pose clustering to maximize the number of favorable interactions –Growing and/or joining in LUDI from pre-built fragment and linker libraries and maximize H-bond and hydrophobic interactions –Anchor-based fragment joining in DOCK Molecular Simulation: MD & MC • Two major components: – The description of the degrees of freedom – The energy evaluation • The local movement of the atoms is performed – Due to the forces present at each step in MD (Molecular Dynamics) – Randomly in MC (Monte Carlo) • Usually time consuming: – Search from a starting orientation to low-energy configuration – Several simulations with different starting orientation must be performed to get a statistically significant result • Grid for energy calculation. Larger steps or multiple starting poses are often used for speed and sampling coverage in MD: – Di Nola et al. 1994; Mangoni et al. 1999; Pak & Wang 2000; CDOCKER by Wu et al. 2003. MC-based Docking E ( B) E ( A) P exp k BT where T is reduced based on a so-called cooling schedule, and grid can be used for energy calculation. • An advantage of the MC technique compared with gradient-based methods (e.g. MD) is that a simple energy function can be used which does not require derivative information, and able to step over energy barrier. • AutoDOCK (Goodsell & Olson 1990). MCDOCK (Liu & Wang 1999), PRODOCK (Trosset & Scheraga 1999), ICM (Abagyan et al. 1994). • Simulated annealing is used in DockVision (Hart & Read 1992) and Affinity (Accelrys Inc., San Diego, CA) • Energy minimization is used in QXP (McMartin & Bohacek 1997). Genetic Algorithm Docking • A fitness function is used to decide which individuals (configurations) survive and produce offspring for the next iteration of optimization. Degrees of freedom are encoded into genes or binary strings. • The collection of genes (chromosome) is assigned a fitness based on a scoring function. There are three genetic operators: – mutation operator randomly changes the value of a gene; – crossover exchanges a set of genes from one parent chromosome to another; – migration moves individual genes from one sub-population to another. • Requires the generation of an initial population where conventional MC and MD require a single starting structure in their standard implementation. • GOLD (Jones et al. 1997); AutoDock 3.0 (Morris et al. 1998); DIVALI (Clark & Ajay 1995). DOCK (Kuntz, UCSF) Receptor Structure • X-ray crystal • NMR • homology Binding Site Molecular Surface of Binding Site Binding Mode Analysis for Lead Optimization: binding orientations and scores for each ligands Virtual Screening for MTS/HTS and Library Design: ligands in the order of their best scores Scoring Orientations 1. Energy scoring (vdw and electrostatic) 2. Contact scoring (shape complementarity) 3. Chemical scoring 4. Solvation terms Filters Spheres describing the shape of binding site and favorable locations of potential ligand atoms Ligands Matching heavy atoms of ligands to centers of spheres to generate thousands of binding orientations • 3D structure • atomic charges • potentials • labeling FlexX (Tripos/SYBYL) • Fragment-based, descriptor matching, empirical scoring (Rarey et al. 1996) • Procedures: – Select a small set of base fragment suitable for placement using a simple scoring function. – Place base fragments with the pose clustering algorithm: rigid, triplet matching of H-bond & hydrophobic interactions, Bohm's scoring function – Build up the remainder of the ligand incrementally from other fragments • Ligand conformations – MIMUMBA model with CSD derived low energy torsional angles for each rotatable bond and ring from CORINA. – Multiple conformations for each fragment in the ligand building steps • Other works: Explicit waters are placed into binding site during the docking procedure using pre-computed water positions(Rarey et al. 1999). Receptor flexibility using discrete alternative protein conformations (Claussen et al. 2001; Claussen & Hindle 2003) GOLD • GA method, H-bond matching, FF scoring (Jones et al. 1997) – A configuration is represented by two bit strings: 1. The conformation of the ligand and the protein defined by the torsions; 2. A mapping between H-bond partners in the protein and the ligand. – For fitness evaluation, a 3D structure is created from the chromosome representation. The H-bond atoms are then superimposed to H-bond site points in the receptor site. – Fitness (scoring) function: H-bond, the ligand internal energy, the protein-ligand van der Waals energy – Rotational flexibility for selected receptor hydrogens along with full ligand flexibility • Highlights: – Validation test set: 100 complexes, 66 with rmsd<2A. – The structure generation is biased towards inter-molecular H-bonds. – Hydrophobic fitting points was added (GOLD 1.2, CCDC, Cambridge, UK 2001). LUDI: Matching polar and hydrophobic groups • Calculate protein and ligand interaction sites (H-bond or hydrophobic), which are defined by centers and surface, from – non-bonded contact distributions based on a search through the CSD, – a set of geometric rules, – the output from the program GRID (Goodford 1985) which calculates binding energies for a given probe with a receptor molecule. • Fit fragments onto the interaction sites. – distance between interaction sites on the receptor – an RMSD superposition algorithm, – A hashing scheme to access and match surface triangles onto a triangle query of a ligand interaction center. – A list-merging algorithm creates all triangles based on lists of fitting triangle edges for two of the three query triangle edges. • Join/grow fragments using the databases of fragments and the same fitting algorithm. GLIDE (www.schrodinger.com) • Funnel: site point search -> diameter test -> subset test -> greedy score -> refinement -> grid-based energy optimization -> GlideScore. • Approximates a complete systematic search of the conformational, orientational, and positional space of the docked ligand. • Hierarchical filters, including a rough scoring function that recognizes hydrophobic and polar contacts, dramatically narrow the search space • Torsionally flexible energy optimization on an OPLS-AA nonbonded potential grid for a few hundred surviving candidate poses. • The very best candidates are further refined via a MC sampling of pose conformation. • A modified ChemScore (Eldridge et al. 1997) that combines empirical and force-field-based terms. • Validation: 282 complexes, new ligand conformation, the topranked pose: 50%<1 A, ~33% >2 A. Matrix of Accuracy & Success • • • • • • Drug <- Quality Novel Lead <- Active Reproduce binding mode (X-ray crystal structures) Predict binding affinity (free energies) Rank diverse set of compounds (by binding affinity) Enhance hit rate for database mining Reduce false positive (Nselected-Nhits) and false negative (Nall_hits-Nhits) Fast enough for iterative SBDD active inactive active TRUE FALSE inactive FALSE TRUE N hits N selected VS HVS EF H0 N all _ hits N all 0 Accuracy of Docking • Reality Boundary – Experimental errors: 0.1-0.25 kcal/mol (18-53%) with MSR (maximum significant ratio) as much as 3 fold (0.65 kcal/mol) – Free energy calculation accuracy: ~1 kcal/mol (5.4 fold) starting with an accurate geometric model & fully sampling – Entropy and solvation estimation need a sufficiently long simulation run with an accurate force field, an ensemble of explicit of water molecules, and fully sampling • Current – – – – – Reproduce X-ray structure with rmsd<2A: 50-90% achievable Binding affinity: 1.5~2 log unit (32-100 fold, 2.05-2.73 kcal/mol) Correlation between scores and affinities, r^2<0.3 Enthalpy ranking with minimization: ±5 kcal/mol Hit rate enhancement : 2~50 fold with hit rate 1-20% (and high false negative rate if 1~5% of total compounds selected) Background & Motivation • • • • • • Docking = process of starting with a set of coordinates for two distinct molecules and generating a model of the bound complex Numerous methods which perform protein- protein docking exist today Fourier correlation approach (Ritchie and Kemp, 2000) enabled the generation of billions of possible docked conformation via defined scoring functions Problem: Many false-positives (good surface complementarity) that are far from the native complex Motivation: Need to develop methods to filter and rank the docked conformations such that near-native complexes can be identified ClusPro: an automated, fast rigid-body docking and discrimination algorithm that: 1) Rapidly filters docked conformations 2) Ranks the conformations using clustering of computed pairwise RMSD values Input and Method Outline CAPRI Receptor-Ligand Pairs 2,000 docked conformations for 48 receptor-ligand pairs Free Energy Filtering 2,000 conformations w/ low desolvation or electrostatic energies Discrimination Via Clustering Top 10 Clusters (Centers) Compare with Native Structure (RMSD) Part I: Free-Energy Filtering • • • • • • • Goal: to identify docked conformations having good surface complementarity by selecting those w/ lowest desolvation and electrostatic energies Surface complementarity is an important criteria due to the observation that proteins tend to bury large surface areas after complex formation Electrostatic and desolvation potentials (capturing the free energy of association) are used independently since different binding mechanisms are governed by different ratios of electrostatic/desolvation contributions 500 structures w/ lowest values of desolvation free energy retained 1500 structures w/lowest electrostatic energy retained Electrostatics more sensitive to small coordinate perturbations noisy Cannot combine desolvation and electrostatics due to the noisy behavior of electrostatics potential Part II: Clustering based on Pairwise RMSD • • • By examining free energy landscapes of partially solvated receptor-ligand complexes: native binding site is expected to be characterized by a local minima having greatest width In other words, the most probable conformation is expected to be surrounded by lots of other low-energy conformations Goal: to use a hierarchical clustering method to select and rank docked conformations having the most “neighbors” given a defined cluster radius (in terms of C-alpha RMSD) Procedure: 1) Need to define fixed molecule (receptor) and flexible molecule (ligand) 2) Define a set of relevant ligand residues to be within 10 Angs of any atom in receptor 3) For each docked conformation X, calculate its pairwise ligand RMSD with 1999 other conformations Pairwise ligand RMSD = deviations between coordinates of X’s defined set of ligand residues and corresponding coordinates of another conformation 4) Cluster the set of 2000 docked conformations using a 2000 by 2000 matrix of RMSD values, and a cluster radius constraint of 9 Angs RMSD from the center 5) Pick largest cluster rank cluster center remove conformations within this cluster from matrix 6) Pick next largest cluster -> rank cluster center remove conformations within this cluster from matrix keep iterating until matrix is empty Results Result I: • Tested the discrimination step of the method on a benchmark set of 48 interacting protein pairs (2000 docked conformations each) • In 31/48 protein pairs, top 10 predictions include at least one near-native complex (average RMSD of 5 angs from native structure) Result II: - Tested method in the CAPRI (Critical Assessment of Predictions of Interactions) experiment and generated predictions for 9 target complexes - Round 3 (automated server): ClusPro prediction ranked as #3 for Target 8 ClusPro Web Server • • • • • User Input: PDB files of the 2 protein structures that user would like to analyze in terms complex formation Output: 10 (default) top predictions of docked conformations closest to native structure First, docking of the 2 proteins is performed using 2 established FFT-based docking programs (DOT and ZDOCK) Then, filtering and discrimination is performed Server allows for customization of parameters: – Clustering radius Smaller protein smaller radius maybe more suitable – Relative number of desolvation and electrostatic best hits used during filtering – Number of predictions to generate (1-30) Protein Drug Discovery • • • • • • • • Although small molecule drugs are more prevalent therapeutics in current drug discovery, protein drugs is a rapidly growing area in pharmaceuticals It is true that protein therapeutics can be much more costly (in terms of R&D and synthesis) than small-molecule therapeutics, but protein therapeutics can deliver biological mechanisms that are not possible with small-molecule therapeutics Multiple blockbuster protein drugs are currently on the market Conservative estimation: there exist between 3,000 and 10,000 possible drug targets Many of these new targets offer great opportunities for the development of protein drugs In 2002, drug companies sold nearly $33 billion in protein drugs Rising at an average annual growth rate (AAGR) of 12.2%, this market is expected to reach $71 billion in 2008. Examples of popular classes of drug targets: 1) G-protein-coupled receptors Compounds will be screened for their ability to inhibit (antagonist) or stimulate (agonist) the receptor 2) Protein kinases Compounds will be screened for their ability to inhibit the kinase Application to Protein Drug Discovery • • Ideal Drug: demonstrate high specificity and high affinity for the target protein In order to evaluate the affinity of the potential drug with the target, you must first predict what the binding interface looks like, and the relative positions of the potential drug and target • ClusPro is the first integrated automated server that incorporates both docking and discrimination steps for structural predictions of protein-protein complexes Using ClusPro, one can generate many relative orientation/conformations of the 2 proteins filter using desolvation + electrostatics potentials discriminate via clustering find the best fit (closest to native structure from x-ray crystallography results) between the 2 proteins Top ranked predictions of ClusPro further manual refinement and discrimination using existing biochemical constraints and analysis to eliminate false positives test binding affinity of promising protein pairs in vitro lead compounds used as starting points for drug development/optimization • • • • Can use ClusPro to screen databases of various existing, recombinant, or de novo proteins for their interaction to a protein target of interest ClusPro can be used to predict either: – How a protein drug may bind (either inhibit or stimulate) a receptor – How 2 proteins bind, and based on the structural details of the interaction design/screen for a drug that can inhibit that interaction 2.1 Rigid Docking • Protein and ligand fixed. • Search for the relative orientation of the two molecules with lowest energy • Fastest way to perform an initial screening of a smallmolecule database -> virtual-screening initiative Rotamer Libraries • Rigid docking of many conformations: • Precompute all low-energy conformations • Dock each precomputed conformations as rigid bodies Glide [Friesner04] Rigid Docking Methods • All rigid-body docking methods have in common that superposition of point sets is a fundamental sub-problem that has to be solved efficiently: • Geometric hashing • Pose clustering • Clique detection Geometric Hashing • Originates from computer vision technology for recognizing partially occluded objects in camera scenes • Given a picture of a scene and a set of objects within the picture, both represented by points in 2d space, the goal is to recognize some of the models in the scene • Objects with certain geometric features can be accessed very fast through a geometric hashing table Pose-Clustering • Originally developed to detect objects in 2-D scenes with unknown camera location • For each triangle of receptor compute the transformation to each ligand matching triangle. • Cluster transformations. • If a cluster grows large, a location with a high number of matching features is found eg. The FlexX Method • The base fragment (the ligand core) is automatically selected and is placed into the active site using a pattern recognition technique called pose clustering • Next, the remainder of the ligand is built up incrementally from other fragments. Clique-Detection •Nodes comprise of matches between protein and ligand •Edges connect distance compatible pairs of nodes •In a clique all pair of nodes are connected Eg. DOCK 6 • The rigid body orienting code is written as a direct implementation of the isomorphous subgraph matching method of Crippen and Kuhl • Conceptually, the algorithm matchings the centers of the ligand heavy atom to the centers of the receptor site spheres. DOCK 6 • The algorithm follows the steps below: 1) Generate node 2) Label as match if atom and sphere edges are equivalent 3) Extend match by adding more nodes 4) Exhaustively generate set of nondegenerate matches 5) Use matches to create transformation matrices to move the entire molecule node = pairing of one heavy atom and one sphere center edge length = Euclidean distance between atom or sphere centers • Once an orientation has been generated, the interaction between the ligand and the receptor can be energetically optimized (ligand is allowed to be flexible in optimization) 2.2. Rigid Receptor, Flexible Ligand Multiple steps in the receptor – ligand interaction: • Approach • Desolvation of the ligand and the binding site of a protein • Penetration into the protein cavity • Change of the ligand orientation • Adoption of the correct “active” conformation • Establishing of new H-bonds, electrostatic and hydrophobic contacts Free energy function : Challenges • Predicting energetics of protein-ligand binding • Searching space of possible poses & conformations – Relative position (3 degrees of freedom) – Relative orientation (3 degrees of freedom) – Rotatable bonds in ligand (n degrees of freedom) – Rotatable bonds in protein (m degrees of freedom) 2.3. Flexible Receptor, Flexible Ligand • Protein flexibility can be introduced through Monte Carlo or Molecular Dynamics – Protein can be divided into rigid and flexible parts -> only flexible receptor site atoms are free to move – The procedure is still very slow • Leach* developed a docking algorithm that sequentially fixes the degrees of freedom of the protein side-chain atoms • Broughton** reported the use of conformational samples from short protein MD simulation runs+ *Leach AR. Ligand docking to proteins with discrete side-chain flexibility. J Mol Biol1994; 235:345–356 **Broughton HB. A method for including protein flexibility in protein–ligand docking: Improving tools for database mining and virtual screening. J Mol Graph Model 2000;18:247–257 AutoDock 4 • AMBER FF-based energy grid, flexible ligands, rigid protein as represented in a grid • GA as a global optimizer combined with energy minimization as a local search method • The fitness function: a Lennard-Jones 12-6 dispersion/repulsion term a directional 12-10 hydrogen bond term a coulombic electrostatic potential a term proportional to the number of sp3 bonds in the ligand to represent unfavorable entropy of ligand binding a desolvation term Comparison of Two Recent Versions Autodock 4 • Scoring Function is based on AMBER FF – FF includes electrostatic interactions, hydrogen bonds, desolvation energy. • “Torsion Tree” for Ligand Flexibility • Protein Flexibility by sidechain rotations • Too many torsions are problematic Autodock Vina • Faster than AutoDock 4 • More accurate than AutoDock 4 • More User-friendly than AutoDock 4 in case of calculation of grid maps and clusters Our Case: Triacylglyceride Docking into Lipase • Lipase: Geobacillus thermocatenulatus Lipase (BTL2) • Crystal Structure in 2009 – 2.2 Å Resolution (Carrasco-López C et al, J Biol Chem. 2009, PMID: 19056729) • 2 Triton X-100 Molecule found in the crystal allows identification of putative binding pockets for the acyl chains (sn-1, sn-2, sn-3) of triglyceride. Tributyrin (4 carbons in chain) Tricaprylin (8 carbons in chain) BTL2 (Apo-enzyme in open-conformation) 79 /26 Work-Flow of Docking Study Separating bound molecules from active site cleft Apo-enzyme Ligand Definition of flexible/rigid bonds Autodock 4.2 and Vina Assesment of Docking Outcomes Poses, Scores Selection of Best Binding Modes Preparation of Input Structures: Protein (BTL2) S114 S114 F17 Open-Lid Conformation displaying catalytic residues for ligand binding F17 Locating search space (grid-box) for triglyceride binding Preparation of Input Structures: Ligand (tricaprylin) Preparation of Input Structures: Ligand (tributyrin) Results and Evaluation of Poses: Tricaprylin (8C) The predicted binding affinity is in kcal/mol. rmsd/lb(c1, c2) = max(rmsd'(c1, c2), rmsd'(c2, c1)) This score matches each atom in one conformation with itself in the other conformation, ignoring any symmetry Results and Evaluation of Poses: Tricaprylin (8C) S114 _OH TCPN _O Mode_1 F17 Results and Evaluation of Poses: Tricaprylin (8C) S114 _OH TCPN _O _2 F17 _1 Results and Evaluation of Poses: Tricaprylin (8C) TCPN _O S114 _OH F17 _7 Results and Evaluation of Poses: Tricaprylin (8C) S114 _OH S114 _OH TCPN TCPN _O _O _8 F17F17 _7 Results and Evaluation of Poses: Tributyrin (4C) S114 TBTN _OH _O _1 F17 _3 Results and Evaluation of Poses: Tributyrin (4C) S114 _OH _1 TBTN _O _6 F17 Results and Evaluation of Poses: Tributyrin (4C) S114 _OH _1 F17 TBTN _O _7 VINA Outcome After 1ns 93 /26 3.1. Force field-based scoring functions • The parameters of the Lennard–Jones potential vary depending on the desired ‘hardness’ of the potential. • D-Score: Higher terms, 12–6 Lennard–Jones potential,result in increasingly repulsive potentials and will be less forgiving of close contacts between receptor and ligand atoms • G-score: Lower terms, 8–4 Lennard–Jones potential, make the potential softer 3.1. Force field-based scoring functions 3.2. Empirical methods • Goals: reproduce the experimental values of binding energies and with its global minimum directed to the X-ray crystal structure • Advantages: fast & direct estimation of binding affinity • Disadvantages –Only a few complexes with both accurate structures & binding energies known –Discrepancy in the binding affinities measured from different labs –Heavy dependence on the placement of hydrogen atoms –Heavy dependence of transferability on the training set –No effective penalty term for bad structures 3.2. Empirical methods • ChemScore: 3.2. Empirical methods • GlideScore: 3.2. Empirical methods • Autodock 4.0 3.2. Empirical methods • Autodock Vina: – Combines advantages of empirical methods and knowledge-based potentials – AutoDock Vina can be several orders of magnitude faster than AutoDock 4 3.3. Knowledge-based methods • Designed to reproduce experimental structures rather than binding energies. • Protein–ligand complexes are modelled using relatively simple atomic interaction-pair potentials. • Advantages – Similar to empirical, but more general (much more distance data than binding energy data) • Disadvantages – The Boltzmann hypothesis originates from the statistics of a spatially uniform liquid, while receptor-ligand complex is a two-component non-uniform medium – PMF are typically pair-wise, while the probability to find atoms A and B at a distance r is non-pairwise and depends also on surrounding atoms 3.3. Knowledge-based methods • Parametrized Pairwise Potential (PMF) score: Boltzmann constant Ligand volume correction factor Radial distribution function for a protein atom i and a ligand atom j 3.3. Knowledge-based methods • DrugScore: [Gohlke00] Multiple Method Approach systematic search conformations initial poses filters rigid DOCK minimization finer docking MD/SA (Wang et al. 1999) final scoring (FRED, GLIDE, DOCK) • Similarity-guided MD simulated annealing to improve accuracy (Wu & Vieth 2004). • Shape similarity & clustering to speed up conformational search in docking (Makino & Kuntz 1998). Better input or constrains for the existing docking engines Computing Scoring Functions • Point-based calculation: • Sum terms computed at positions of ligand atoms (this will be slow) Computing Scoring Functions • Grid-based calculation: • Precompute “force field” for each term of scoring function for each conformation of protein (usually only one) • Sample force fields at positions of ligand atoms -> Accelerate calculation of scoring function by 100X [Huey & Morris] Consensus Scoring • Typically evaluate the ranking of binding modes measured with different scoring functions and favor those that rank consistently high in several of them • Reduces false positive rate • Examples – – – – SYBYL Cscore (Tripos) : FlexX, PMF, DOCK energy, GOLD score C2 (Accelrys) : LigScore2, PLP, PMF, Ludi, Jain FRED (OpenEye) : ChemScore, PB-SA, ChemGauss, PLP, ScreenScore DOCK: AMBER FF, PMF, contact scores, ChemScore Flexible ligand-search methods Random/stochastic • AutoDock (MC) • MOE-Dock (MC,TS) • GOLD (GA) • PRO_LEADS (TS) Systematic • DOCK (incremental)24 • FlexX (incremental)50 • Glide (incremental)134 • Hammerhead (incremental)28 • FLOG (database) Simulation • DOCK • Glide • MOE-Dock • AutoDock • Hammerhead Docking Software: Important Factors • Sensitivity on and transferability of the parameters, including the starting conformation • Adaptability to additional scoring functions, pre- and/or post- docking processing and filters • Ability for iteratively refining docking parameter/protocol based on new results • • • • Design, components, and results of validation studies Speed, user interface & control, I/O, structural file formats User learning curve, customer supports, and cost Code availability and upgrading possibility Docking Softwares DOCK 6.0 (Ewing & Kuntz 1997) AutoDOCK 4.0 (Morris et al. 1998) GOLD (Jones et al. 1997) FlexX: (Rarey et al. 1996) GLIDE: (Friesner et al. 2004) ADAM (Mizutani et al. 1994) CDOCKER (Wu et al. 2003) CombiDOCK (Sun et al. 1998) DIVALI (Clark & Ajay 1995) DockVision (Hart & Read 1992) FLOG (Miller et al. 1994) GEMDOCK (Yang & Chen 2004) Hammerhead (Welch et al. 1996) LIBDOCK (Diller & Merz 2001) MCDOCK (Liu & Wang 1999) SDOCKER (Wu et al. 2004) de novo design tools LUDI (Boehm 1992), BUILDER (Roe & Kuntz 1995) SMOG (DeWitte et al. 1997) CONCEPTS (Pearlman & Murcko 1996) DLD/MCSS (Stultz & Karplus 2000) Genstar (Rotstein & Murcko 1993) Group-Build (Rotstein & Murcko 1993) Grow (Moon & Howe 1991) HOOK (Eisen et al. 1994) Legend (Nishibata & Itai 1993) MCDNLG (Gehlhaar et al. 1995) SPROUT (Gillet et al. 1993) FRED (OpenEye www.eyesopen.com) • • • • • Systematic, nonstochastic, docking Multiple active site comparisons Multiple simultaneous scoring functions and hit lists RMS clustering of hit-lists Algorithm: 1. Exhaustive Docking (a) Enumerate all possible poses of the ligand around the active site by rigidly rotating and translating each conformer within the site. (b) Filter the resulting pose ensemble by rejecting poses that do not fit within the larger of the two volumes specified by the receptor file’s shape potential grid and a contour level. 2. Systematic solid body optimization by Shapegauss, PLP, Chemgauss2, Chemgauss3, CGO, CGT, Chemscore, OEChemscore or Screenscore 3. Rank poses via the Consensus Structure method and discard all but the top ranked poses DOCK 6.4 • Generates many possible orientations/conformations of a putative ligand within a user-selected region of a receptor structure • Orientations may be scored using several schemes designed to measure steric and/or chemical complementarity of the receptor-ligand complex • Evaluate likely orientations of a single ligand, or to rank molecules from a database • Search databases for DNA-binding compounds • Examine possible binding orientations of proteinprotein and protein-DNA complexes • Design combinatorial libraries GOLD • GA method, H-bond matching, FF scoring (Jones et al. 1997) – A configuration is represented by two bit strings: 1. The conformation of the ligand and the protein defined by the torsions; 2. A mapping between H-bond partners in the protein and the ligand. – For fitness evaluation, a 3D structure is created from the chromosome representation. The H-bond atoms are then superimposed to H-bond site points in the receptor site. – Fitness (scoring) function: H-bond, the ligand internal energy, the proteinligand van der Waals energy • Highlights: – Full ligand flexibility – Partial protein flexibility, including protein side chain and backbone flexibility for up to ten user-defined residues – A choice of GoldScore, ChemScore, Astex Statistical Potential (ASP) or Piecewise Linear Potential (PLP) scoring functions – GOLD's genetic algorithm parameters are optimised for virtual screening applications Hammerhead • Focus on screening large databases of small molecules • The algorithm is fast enough to allow screening of a library of roughly 100.000 small organic compounds in a few days • Empirical scoring function • Start with automatic pocket finder • Breaking ligands into fragments, and aligning each of these onto the protein. • At each stage of the fragment alignment computation, gradientdescent pose optimization improves the conformation and alignment of the growing ligand – Relaxing van der waals surface interpenetrations – Improving hydrogen bond and hydrophobic surface contact geometries. LUDI: Matching polar and hydrophobic groups • Calculate protein and ligand interaction sites (H-bond or hydrophobic), which are defined by centers and surface, from – non-bonded contact distributions based on a search through the CSD, – a set of geometric rules, – the output from the program GRID (Goodford 1985) which calculates binding energies for a given probe with a receptor molecule. • Fit fragments onto the interaction sites. – distance between interaction sites on the receptor – an RMSD superposition algorithm, – A hashing scheme to access and match surface triangles onto a triangle query of a ligand interaction center. – A list-merging algorithm creates all triangles based on lists of fitting triangle edges for two of the three query triangle edges. • Join/grow fragments using the databases of fragments and the same fitting algorithm. GLIDE (www.schrodinger.com) • Funnel: site point search -> diameter test -> subset test -> greedy score -> refinement -> grid-based energy optimization -> GlideScore. • Approximates a complete systematic search of the conformational, orientational, and positional space of the docked ligand. • Hierarchical filters, including a rough scoring function that recognizes hydrophobic and polar contacts, dramatically narrow the search space • Torsionally flexible energy optimization on an OPLS-AA nonbonded potential grid for a few hundred surviving candidate poses. • The very best candidates are further refined via a MC sampling of pose conformation. • A modified ChemScore (Eldridge et al. 1997) that combines empirical and force-field-based terms. • Validation: 282 complexes, new ligand conformation, the top-ranked pose: 50%<1 A, ~33% >2 A. GRAMM v1.03 • Protein-Protein Docking and Protein-Ligand Docking • exhaustive 6-dimensional search through the relative translations and rotations of the molecules. • Empirical approach to smoothing the intermolecular energy function. • The quality of the prediction depends on the accuracy of the structures. CDOCKER & SDOCKER Randomly generate ligand seeds in the binding site High temperature MD using a modified version of CHARMM Locate minima from all of the MD simulations Fully minimization Cluster on position and geometry Rank by energy (interaction + ligand conformation) SDOCKER: X-ray structure of complex as templates to guide docking Wu et al. 2003; Wu et al. 2004. Docking Webservers Assessment of CAPRI Predictions 2009 ClusPro Webserver • Fast rigid-body docking • Ligand-Protein, Protein-Protein Docking • Use FFT-based docking programs (DOT and ZDOCK) 1) Rapidly filters docked conformations 2) Ranks the conformations using clustering of computed pairwise RMSD values • Desolvation and Electrostatic energies are calculated Haddock • Driven by experimental knowledge (e.g., from mutagenesis, mass spectrometry or a variety of NMR experiments) • Protein-Protein Docking server • Supports nucleic acids • Algorithm: 1. 2. 3. Rigid-body Energy Minimization, Semi-flexible Refinement In Torsion Angle Space Final refinement in explicit solvent. • The HADDOCK score : van der Waals, electrostatic, desolvation and restraint violation energies together with buried surface area GRAMM-X • Protein-Protein Docking server • Use FFT for the global search of the best rigid body conformations. • Use a smoothed Lennard-Jones potential on a fine grid • Ability to smooth the protein surface to account for possible conformational change • The smoothing of the intermolecular energy landscape is achieved by increasing potential range and lowering the value of the repulsion part Softened Lennard-Jones potential function: PatchDock and SymmDock Server • Based on a rigid-body geometric hashing algorithm • Aim: Good molecular shape complementarity yield • Algorithm divides the Connolly dot surface representation of the molecules into concave, convex and flat patches. • Then, complementary patches are matched in order to generate candidate transformations. • Each candidate transformation is further evaluated by a scoring function that considers both geometric fit and atomic desolvation energy. PatchDock detects transformations with high shape complementarity SymmDock Server SymmDock restricts its search to symmetric cyclic transformations of a given order n. FireDock server • Fast rigid-body docking algorithms • Protein-protein docking RosettaDock protein-protein docking server • • • Computationally intensive approach incorporating models flexibility Multi-start, multi-scale Monte Carlo based algorithm Start with 1000 independent structures, and the server returns pictures, coordinate files and detailed scoring information for the 10 top-scoring models • The low-resolution phase: – Random rigid-body perturbations – Scoring : residue–residue contacts and bumps, knowledge-based terms for residue environment and residue–residue pair propensities and for antibody-antigen targets, a score to favor interactions with antibody complementarity determining regions. • The high-resolution (all-atom, including hydrogens) phase – Smaller rigid-body perturbations, sidechain optimization via rotamer packing and continuous minimization, and explicit gradient-based minimization of the rigid-body displacement. – Scoring: the energy is dominated by van der Waals energies , orientation-dependent hydrogen bonding , implicit Gaussian solvation, side-chain rotamer probabilities and a low-weighted electrostatics energy. ZDOCK HexServer • In order to address the main limitations of the Cartesian • FFT approaches, we developed the ‘Hex’ spherical polar • Fourier (SPF) approach which uses rotational correlations • (10), and which reduces execution times to a matter of • minutes Bold entries in the first column correspond to programmes that can be run on a web server. (a) Refined with SMOOTHDOCK. (b) Uses DOT or ZDOCK as search methods; (c) Refined with RDOCK Virtual Screening • Drug discovery costs are too high: ~$800 millions, 8~14 years, ~10,000 compounds (DiMasi et al. 2003; Dickson & Gagnon 2004) • Drugs interact with their receptors in a highly specific and complementary manner. • Core of the target-based structure-based drug design (SBDD) for lead generation and optimization. Lead is a compound that – shows biological activity, – is novel, and – has the potential of being structurally modified for improved bioactivity, selectivity, and drugeability. Drug, Chemical & Structural Space • Drug-like: MDDR (MDL Drug Data Report) >147,000 entries, CMC (Comprehensive Medicinal Chemistry) >8,600 entries • Non-drug-like: ACD (Available Chemicals Directory) ~3 million entries • Literatures and databases, Beilstein (>8 million compounds), CAS & SciFinder • CSD (Cambridge Structural Database, www.ccdc.cam.ac.uk): ~3 million X-ray crystal structures for >264,000 different compounds and >128,00 organic structures • Available compounds – Available without exclusivity: various vendors (& ACD) – Available with limited exclusivity: Maybridge, Array, ChemDiv, WuXi Pharma, ChemExplorer, etc. • Corporate databases: a few millions in large pharma companies Docking to Nucleic Acid Targets • RNA and DNA as potential drug targets – Ribosome RNA structures (Agalarov et al. 2000; Ban et al. 2000; Filikov et al. 2000; Nissen et al. 2000; Wimberly et al. 2000) • Highly charged environments, well-defined binding pocket • DOCK identified compounds selectively bind to RNA duplexes or DNA qudraplexes (Chen et al. 1996; Chen et al. 1997). The portions in the DOCK suite that calculate electrostatics, including solvation, partial charges, and scoring function were recently optimized for RNA targets (Downing et al. 2003; Kang et al. 2004). • A MC minimization and an empirical scoring function which accounts for solvation, isomerization free energy, and changes in conformational entropy were used to rank compounds (Hermann & Westhof 1999).