Chemical Data and
Computer-Aided Drug Discovery
Mike Gilson
School of Pharmacy
mgilson@ucsd.edu
2-0622
Outline
Overview of drug discovery
Structure-based computational methods
When we know the structure of the targeted protein
Ligand-based computational methods
When we don’t know the protein’s structure
What is a drug?
Small Molecule Drugs
Aspirin
Taxol
Sildenafil (Viagra)
Darunavir
Glipizide (Glucotrol)
Digoxin
Nanoparticles
(e.g., packaged small-molecule drugs)
Doxil
Abraxane
(liposome package,
extended circulation time,milder toxicity)
(albumin-packaged taxol)
http://www.doxil.com/about_doxil.html
http://www.abraxane.com/professional/nab-technology.aspx
Biopharmaceuticals
Erythropoietin (EPO)
Stabilized variant of a natural protein hormone
http://www.ganfyd.org/index.php?title=Erythropoietin_beta
Etanercept (Enbrel)
Protein with TNF receptor + Ab Fc domain
Scavenges TNF, diminishes inflammation
http://en.wikipedia.org/wiki/File:Enbrel.jpg
How are drugs discovered?
Natural Products
Aspirin
Digoxin
Taxol
Pacific Yew
Willow
Foxglove
How Aspirin Works
Aspirin
inflammation
platelet activation
platelet inactivation
Biomolecular Pathways and Target Selection
E.g. signaling pathways
Target protein
http://www.isys.uni-stuttgart.de/forschung/sysbio/insulin/index.html
Empirical Path to Ligand Discovery
Compound library
(commercial, in-house,
synthetic, natural)
High throughput screening
(HTS)
Hit confirmation
Lead compounds
(e.g., µM Kd)
Lead optimization
(Medicinal chemistry)
Animal and clinical
evaluation
Potent drug candidates
(nM Kd)
Compound Libraries
Commercial (also in-house pharma)
Academia
Government (NIH)
Computer-Aided Ligand Design
Aims to reduce number of compounds synthesized and assayed
Lower costs
Less chemical waste
Faster progress
Scenario 1
Structure of Targeted Protein Known: Structure-Based Drug Discovery
HIV Protease/KNI-272 complex
Protein-Ligand Docking
Structure-Based Ligand Design
Docking software
Potential function
Search for structure of lowest energy
Energy as function of structure
VDW
-
+
Screened Coulombic
Dihedral
Energy Determines Probability (Stability)
Boltzmann distribution
Probability
Energy
p( x)  e
x
 E ( x )/ RT
Structure-Based Virtual Screening
3D structure of target
Compound database
(crystallography, NMR, modeling)
Virtual screening
(e.g., computational docking)
Candidate ligands
Ligand optimization
Med chem, crystallography, modeling
Experimental assay
Ligands
Drug candidates
Fragmental Structure-Based Screening
3D structure of target
“Fragment” library
(crystallography, NMR, modeling)
Fragment docking
Compound design
Experimental assay and ligand optimization
Med chem, crystallography, modeling
http://www.beilstein-institut.de/bozen2002/proceedings/Jhoti/jhoti.html
Drug candidates
Potential Functions for Structure-Based Design
Energy as a function of structure
Physics-Based
Knowledge-Based
Physics-Based Potentials
Energy terms from physical theory
Van der Waals interactions (shape fitting)
Bonded interactions (shape and flexibility)
Coulombic interactions (charge-charge complementarity)
Hydrogen-bonding
Common Simplifications Used in
Physics-Based Docking
Quantum effects approximated classically
Protein often held rigid
Configurational entropy neglected
Influence of water treated crudely
Proteins and Ligand are Flexible
Protein
Ligand
Complex
+
DGo
Binding Energy and Entropy
2e EBound / RT
K   EFree / RT
6e
EFree
Unbound states
EBound
Bound states
DG  RT ln  K   Ebound  EFree  RT ln 3
Energy part
Entropy part
Structure-Based Discovery
Physics-oriented approaches
Weaknesses
Fully physical detail becomes computationally intractable
Approximations are unavoidable
Parameterization still required
Strengths
Interpretable, provides guides to design
Broadly applicable, in principle at least
Clear pathways to improving accuracy
Status
Useful, far from perfect
Multiple groups working on fewer, better approxs
Force fields, quantum
Flexibility, entropy
Water effects
Moore’s law: hardware improving
Knowledge-Based Docking Potentials
Ligand
carboxylate
Aromatic
stacking
Probability Energy
Boltzmann: p(r )  e
 E ( r )/ RT

Inverse Boltzmann: E(r )  RT ln p(r )

Example: ligand carboxylate O to protein histidine N
1.
2.
3.
4.
Find all protein-ligand structures in the PDB with a ligand carboxylate O
For each structure, histogram the distances from O to every histidine N
Sum the histograms over all structures to obtain p(rO-N)
Compute E(rO-N) from p(rO-N)
Knowledge-Based Docking Potentials
“PMF”, Muegge & Martin, J. Med. Chem. 42:791, 1999
A few types of atom pairs, out of several hundred total
Nitrogen+/Oxygen-
Aromatic carbons
Aliphatic carbons
Atom-atom distance (Angstroms)
Eprot lig  Evdw 

pairs (ij )
Etype(ij ) (rij )
Structure-Based Discovery
Knowledge-based potentials
Weaknesses
Accuracy limited by availability of data
Accuracy may also be limited by overall approach
Strengths
Relatively easy to implement
Computationally fast
Status
Useful, far from perfect
May be at point of diminishing returns
Limitations of Knowledge-Based Potentials
1. Statistical limitations
(e.g., to pairwise potentials)
100 bins for a histogram of O-N & O-C distances
10 bins for a histogram of O-N distances
r1
r2
…
r10
rO-C
rO-N
rO-N
2. Even if we had infinite statistics, would the results be accurate?
(Is inverse Boltzmann quite right? Where is entropy?)
Scenario 2
Structure of Targeted Protein Unknown: Ligand-Based Drug Discovery
e.g. MAP Kinase Inhibitors
Using knowledge of
existing inhibitors to
discover more
Why Look for Another Ligand if You Already Have Some?
Experimental screening generated some ligands, but they don’t bind tightly
A company wants to work around another company’s chemical patents
An high-affinyt ligand is toxic, is not well-absorbed, etc.
Ligand-Based Virtual Screening
Compound Library
Known Ligands
Molecular similarity
Machine-learning
Etc.
Candidate ligands
Optimization
Med chem, crystallography, modeling
Assay
Actives
Potent drug candidates
Sources of Data on Known Ligand
Journals, e.g., J. Med. Chem.
Some Binding and Chemical Activity Databases
PubChem (NIH) pubchem.ncbi.nlm.nih.gov
ChEMBL (EMBL) www.ebi.ac.uk/chembl
BindingDB (UCSD) www.bindingdb.org
BindingDB
www.bindingdb.org
Finding Protein-Ligand Data in BindingDB
e.g., by Name of Protein “Target”
e.g., by Ligand Draw  Search
Sample Query Results
BindingDB to PDB
PDB to BindingDB
Download data in
machine-readable
format
Sample Query Results
Machine-Readable Chemical Format
Structure-Data File (SDF)
SDF Format Defines Chemical Bonds
PDB Format Lacks Chemical Bonding
There are Many Other Chemical File Formats
Interconvert with Babel
Chemical Similarity
Ligand-Based Drug-Discovery
Compounds
(available/synthesizable)
Similar
Test experimentally
Don’t bother
Chemical Fingerprints
Binary Structure Keys
…
Molecule 1
Molecule 2
Chemical Similarity from Fingerprints
Tanimoto Similarity or Jaccard Index, T
NI
T
 0.25
NU
Intersection
NI=2
Union
NU=8
Molecule 1
Molecule 2
Hashed Chemical Fingerprints
Based upon paths in the chemical graph
1-atom paths: C
F
N
H
S
O
2-atom paths: F-C
C-C
C-N
C-S
S-O
C-H
3-atom paths: F-C-C C-C-N C-N-H
C-S-O
Each path sets a pseudo-random bit-pattern in a very long molecular fingerprint
C
S-O
etc.
Maximum Common Substructure
Ncommon=34
Potential Drawbacks of Plain Chemical Similarity
May miss good ligands by being overly conservative
Too much weight on irrelevant details
Scaffold Hopping
Identification of synthetic statins by scaffold hopping
Zhao, Drug Discovery Today 12:149, 2007
Abstraction and Identification of
Relevant Compound Features
Ligand shape
Pharmacophore models
Chemical descriptors
Statistics and machine learning
Pharmacophore Models
Φάρμακο (drug) + Φορά (carry)
A 3-point pharmacophore
Bulky
hydrophobe
3.2 ±0.4 Å
+1
Aromatic
Molecular Descriptors
More abstract than chemical fingerprints
Physical descriptors
molecular weight
charge
dipole moment
number of H-bond donors/acceptors
number of rotatable bonds
hydrophobicity (log P and clogP)
Topological
branching index
measures of linearity vs interconnectedness
Etc. etc.
Rotatable bonds
A High-Dimensional “Chemical Space”
Each compound is at a point in an n-dimensional space
Descriptor 3
Compounds with similar properties are near each other
Descriptor 2
Point representing a
compound in
descriptor space
Statistics and Machine Learning
Some examples
Partial least squares
Support vector machines
Genetic algorithms for descriptor-selection
Summary
Overview of drug discovery
Computer-aided methods
Structure-based
Ligand-based
Interaction potentials
Physics-based
Knowledge-based (data driven)
Ligand-protein databases, machine-readable chemical formats
Ligand similarity and beyond
Mike Gilson, School of Pharmacy, mgilson@ucsd.edu, 2-0622
Activities and Discussion Topics
BindingDB: Advil Machine-readable format, Binding activities
PDB/BindingDB
2ONY at PDB BindingDB Substructure search Related data
 Similarity search
Combined computational approaches
(physics + knowledge)-based docking potentials
(ligand + structure)-based computational discovery
Other data-driven methods where it may be hard to get enough statistics
Validation of computational methods
Protein-ligand databases: getting data and assessing data quality
Drug Discovery Pipeline
(One Model)
Target
identification
Target
validation
Assay
development
Phase I Clinical
(safety, metab, PK)
Lead
compound
(ligand)
discovery
Phase II Clinical
(efficacy)
Lead
optimization
Phase III Clinical
(comparison with
existing therapy)
Animal
Pharmacokinetics,
Toxicity
Updated Knowledge-Based PMF Potential
Muegge J. Med. Chem. 49: 5895, 2006