![Knowledge-Based Scoring Functions
Example: ligand carboxyl O to protein histidine N
Procedure:
1. Find all PDB structures with ligand carboxyl O
2. Compute all distances to protein histidine N’s
3. Plot histogram of all O-N distances: p(rO-N)
4. Calculate E(r) using inverse Boltzmann
Boltzmann: p(r) ~ exp[ -E(r)/(RT) ]
Inverse Boltzmann: E(r) = -RT ln[ p(r) ]
Muegge & Martin. J Med Chem 42, 791-804 (1999)](https://image.slidesharecdn.com/caddlecture-121015214843-phpapp02/85/CADD-Lecture-24-320.jpg)
CADD Lecture
- 1. COMPUTER-AIDED DRUG DESIGN Thursday, April 19, 2012 Kazi Shefaet Rahman k.s.rahman@gatech.edu
- 2. Outline • The pharmaceutical pipeline • Structure-based drug design: docking • Ligand-based drug design: pharmacophore modeling and QSAR
- 3. What is a drug? A substance that, when absorbed, alters normal bodily function In pharmacology: FDA-approved for the diagnosis, treatment, or prevention of disease.
- 4. Major Drug Classes Small Molecule Protein Vaccine Aspirin Insulin Lovastatin Trastuzumab Amprenavir Erythropoietin
- 5. Where do drugs come from? Vaccines Natural 4% 5% Biologics 14% Natural Derivatives 23% Synthetic Synthetic 40% Natural Mimics 14% Newnan & Cragg. J Nat Prod 70, 461–477 (2007)
- 6. Natural Sources Aspirin is derived Taxol was discovered Morphine was purified from willow bark in the Pacific yew tree from opium poppies
- 7. Success Rate in Pharma 10,000 1,000 100 10 1 marketable drug
- 8. Compound Libraries Commercial, government or 10,000 – 10,000,000 academic compounds
- 9. Target Identification Farooque & Lee. Annu Rev Physiol 71, 465–487 (2009)
- 10. Assay Development Stockwell. Nature 432, 846-854 (2004)
- 11. Ligand Design Pathway Identify High- in vitro Validate Compound throughput libraries screening in vivo Develop assay Target Discovery Lead Discovery Lead Optimization
- 13. Computer-Aided Drug Design • Enrich existing compound libraries • Reduce amount of chemical waste • Faster progress • Lower costs
- 14. Computers in the Ligand Design Pathway Identify High- in vitro Validate Compound throughput libraries screening in vivo Develop assay Target Discovery Lead Discovery Lead Optimization Bioinformatics Computer-Aided Drug Design
- 15. Structure-based and Ligand-based Receptor Structure? Known Unknown Ligand-Based Design Structure-Based Design e.g. pharmacophore modeling & e.g. Docking QSAR
- 16. Video: Molecular Docking using Glide
- 17. Protein-Ligand Docking • For a receptor-ligand complex, we want to predict: 1. Preferred orientation (pose) 2. Binding affinity (score) • A docking program has least two functional components 1. Search algorithm 2. Scoring function • Docking can be used for virtual screening, lead optimization, or de novo design of ligands
- 18. Molecular Representations in Docking 1. Atomic • Every atom represented • Usually used with a molecular mechanics scoring function • Computationally complex 2. Surface • Molecules represented by solvent excluded surface • Align points by minimizing surface angle • Commonly used in protein-protein docking 3. Grid • Receptor’s energetic contributions stored in grid points • van der Waals, electrostatic, H-bonding terms • May be combined with atomistic representation at binding site surface
- 19. Search Algorithms: Ligand Flexibility 1. Systematic • DOCK • FlexX • Cycle through values for each degree of freedom • Glide • Quickly leads to combinatorial explosion! • Hammerhead • Often implemented in anchor-and-grow algorithms • FLOG 2. Stochastic • AutoDock • MOE-Dock • Make random changes, and evaluate using Monte Carlo or genetic algorithms • GOLD • Tabu search algorithms minimize repetition of dead-ends • PRO_LEADS 3. Simulation • DOCK • Glide • Molecular dynamics • MOE-Dock • Energy minimization often used with other search • AutoDock technique • Hammerhead
- 20. Alchemical Free Energy Calculations ΔGI Laq + Paq ΔGbind Paq + Lg ΔGII ΔGbind = ΔGI – ΔGII (PL)aq
- 21. Scoring Functions 1. Force-field-based • D-Score • Quantify sum of receptor-ligand interaction energy and internal ligand • G-Score energy • GOLD • Ligand-receptor potential contains van der Waals and Coulomb • AutoDock electrostatic terms (and H-bonding, in some cases) • DOCK • Limited by lack of solvation and entropic terms 2. Empirical • LUDI • Binding energies as sums of uncorrelated terms, similar to but simpler • F-Score than force-field terms • ChemScore • Parameterized to fit regression analysis of experimental data • SCORE • Often contain terms to approximate (de)solvation and entropic penalties • Fresno • X-Score 1. Knowledge-based • PMF • Use potentials of mean force derived from libraries of protein-ligand complexes • DrugScore • Computationally simple • SMoG
- 22. Force-field-based Scoring Function Autodock v4.0 Kitchen et al. Nat Rev Drug Discov 3, 935–949 (2004)
- 23. Empirical Scoring Functions • LUDI Böhm. J Comput Aided Molec Des 8, 243-256 (1994) • ChemScore Eldridge et al. J Comput Aided Molec Des 11, 425-444 (1997)
- 24. Knowledge-Based Scoring Functions Example: ligand carboxyl O to protein histidine N Procedure: 1. Find all PDB structures with ligand carboxyl O 2. Compute all distances to protein histidine N’s 3. Plot histogram of all O-N distances: p(rO-N) 4. Calculate E(r) using inverse Boltzmann Boltzmann: p(r) ~ exp[ -E(r)/(RT) ] Inverse Boltzmann: E(r) = -RT ln[ p(r) ] Muegge & Martin. J Med Chem 42, 791-804 (1999)
- 25. Scoring: General Caveats • Ligand flexibility and size • For rigid molecules, correct pose predicted 90-100% of the time • Drops to 45-80% for molecules with more rotatable bonds and MW • Binding strength • Most strong binders (Kd <100 nM) correctly predicted • Difficult to predict weaker-binding ligands (Kd ~ 1μM) • Binding site • Hydrophobic binding sites yield better results than hydrophilic ones • Placement of water molecules play an important part • Active sites that require a conformational change (induced-fit) fair poorly in rigid protein models • Ideally want to start with holo- structure
- 26. Lead Optimization and de novo Design • Lead Optimization • Evaluate small changes in structure to distinguish between a μM and a nM compound • Need highly accurate docking and scoring functions • Typically implemented as an “anchored search” to reduce number of analogues • Can be used to prioritize sites for experimental modification • De novo Design: • Multiple-copy simultaneous search (MCSS): small fragments simultaneously docked, and preferred fragments combined • Difficult to predict synthetic availability of designed molecule
- 27. Docking and ADME Evaluation • Absorption, distribution, metabolism and excretion • Concentrate on drug interaction with models of cytochrome P450
- 28. Structure-based and Ligand-based Receptor Structure? Known Unknown Ligand-Based Design Structure-Based Design e.g. pharmacophore modeling & e.g. Docking QSAR
- 29. Lipinski’s Rule of Five A rule of thumb to evaluate the likely activity of an oral drug candidate 1. ≤5 hydrogen bond donors 2. ≤10 hydrogen-bond acceptors 3. MW < 500 Da 4. log P ≤ 5 (P = octanol-water partition coefficient)
- 30. Pharmacophore modeling • Pharmacophore: Set of features common to all known ligands of a particular target • Methodology: 1. Model conformational space of ligands in training set to simulate flexibility 2. Align generated conformations 3. Extract common features Richmond et al. J Comput Aided Mol Des 20, 567-587 (2006)
- 31. Pharmacophore-based Virtual Screening • Screen compound library for ligands containing pharmacophore of interest • Methodology 1. Generate ensemble of conformations for each ligand to be tested 2. Perform pharmacophore pattern matching (“substructure search”) on every conformer • Procedures from graph theory: Ullman, backtracking algorithm, GMA algorithm
- 32. Challenges in Pharmacophore Modelling • Modeling ligand flexibility • Conformers pre-enumerated or generated on-the-fly • Systematic torsional grids, genetic algorithms, Monte Carlo • Molecular alignment • Point-based: superimpose pairs of atoms. Anchor points need to be defined • Property-based: use molecular field descriptors to align • Choosing a training set • Choice of training set has big impact on generate pharmacophore model • New research: Pharmacophore-based de novo design
- 33. Quantitative Structure/Activity Relationships • A QSAR is a mathematical relationship between the geometrical and chemical characteristics of a molecule and its biological activity • Chemical descriptors are correlated with biological activity in terms of an equation • A valid QSAR should allow prediction of the biological activity of new ligands prior to synthesis and in vitro and in vivo assays
- 34. QSAR Requirements 1. Dataset • Experimental measurements of the biological activity of a group of chemicals 2. Descriptors • Numerical values that encode relevant structure and property data for this group of chemicals 3. Statistical methods • To find relationships between these two sets of data
- 35. Molecular Descriptors in QSAR 1. Constitutional • Total number of atoms, atoms of a certain type, number of bonds, number of rings 2. Topological • Molecular shape, degree of branching 3. Electronic • Partial atomic charges, dipole moments 4. Geometrical • van der Waals volume, molecular surface 5. Quantum Mechanical • Total energy, interaction energy between two atoms, nuclear repulsion between atoms 6. Physicochemical • Liquid solubility, log P, boiling point
- 36. QSAR Methodology • Thousands of descriptors can be generated for each molecule • Several descriptors will be correlated • E.g. MW and boiling point • Statistically analyze descriptors to isolate 3 to 5 independent descriptors that best correlate with biological activity • Use regression methods to express activity in terms of descriptors • BA = a + bX1 + cX2 + … • This model can now be used to predict activity in other test compounds
- 37. 3D QSAR • If the structure of ligands are known, one can map important chemical descriptors into 3D space • This will generate a 3D pattern of functionally significant regions of the ligand • Visual identification of regions responsible for (un)favorable interactions
- 38. Comparative Molecular Field Analysis • Principle: Differences in binding/activity often due to differences in the shape of the non-covalent fields surrounding the molecule • Methodology: 1. Align all test molecules 2. Place in a 3D grid (2 Å spacing) 3. Measure steric (van der Waals) and electrostatic (Coulomb) energy for each molecule with a probe atom 4. Correlate energies with activity to generate 3D-QSAR 5. Display QSAR as colors and/or contours around molecular structures
- 39. CoMFA Example 3D alignment CoMFA QSAR 44 compounds (37 training, 7 test) Shagufta et al. J Mol Model 13, 99-109 (2006)
Editor's Notes
- #4: People in the pharmaceutical industry are careful not to call drug candidates drugs.Potential small molecule binders are usu called “ligands”.
- #5: Small molecules: <1000 Da, chemical synthesisProteins: Bioengineered recombinant host expressionVaccine: Chicken eggs
- #6: Aspirin (irreversibly) inhibits cyclooxygenase (COX) activity.Lowers metabolic synthesis of prostaglandins 9which are implicated in inflammatory response, pain sensitization).Also reduces platelet activity by lower thrombaxane levels.Pushes balance towards lipooxygenase pathway, leading to leukotriene formation.
- #7: Atomic: Allows flexibility in both protein and ligandSurface: Both molecules rigidGrid: Interaction between receptor and probe atom/groups calculated once (assume protein rigid), stored in look-up table.
- #8: Systematic:1) Core (rigid) fragment selection and placement (DOCK: based on steric complementarity, FleXX: based on interactions between receptor and ligand groups), followed by growing of flexible side chains systematically.2) Hammerhead: Dock fragments, then rebuild ligandStochastic:Monte Carlo: Generate a random configuration, score it Generate another, score it and compare to previous using Metropolis criterion (if energy lower, use immediately, if higher, apply Boltzmann probability)Tabu: Compare generated structure to previously accepted structures, accept only if it is smaller than any other accepted conformation or is sufficiently different.To model protein flexiblity, MD, MC techniques can be used. An ensemble of protein structures can also be used to generate a grid.
- #10: Newer force-field-based scoring functions have torsional entropy terms.Empirical and knowledge-based functions attempt to capture implicitly what is difficult to capture explicitly
- #11: LUDI: Rotor term approximates entropic penalty with a weighted sum of the number of rotatable bondsChemScore: Rotor term is is treated with more sophistication – takes account of immediate chemical environment of rotatable bonds
- #12: Turns out that most of the errors in docking predictions are not due to incorrect pose predictions, but due to poor scoring – easier to deselect than to selectLarger, more flexible compounds can form many hypothetical interactions and generate better scores (especially if entropic penalties are not included).In hydrophobic binding sites, contact surface is a very good predictor of binding (and this problem has been worked on for a long time). This is not perfect either, since desolvation energies are not taken into account adequately.Electrostatic interactions pose a particularly difficult problem.Despite these problems, VS still works because accurate scores or binding energies are not needed to enrich the dataset.
- #13: H-bond donors: N or O atoms attached to HH-bond acceptors: N or O atoms