Molecular Docking in

Structure-based Drug Design

Ivan Tubert-Brohman

Presentation for Cooper Union, 2/22/2017

How do drugs work?
Drugs bind to a biological
target, usually a protein
or nucleic acid.

For example, binding to

an enzyme can disrupt or
modulate a metabolic
process.

Figure from Nature Reviews Microbiology 2003, 2, 704–720

Pharmacodynamics and pharmacokinetics

Pharmacodynamics: effect of the drug on the body; mechanism of action (e.g.,

binding to a biological target)

Pharmacokinetics: “how the body affects the drug”. ADME:

● Absorption
● Distribution
● Metabolism
● Excretion

Structure-based drug design principally focuses on pharmacodynamics, although

some structure-based approaches can also be applied to questions of
pharmacokinetics.
Drug discovery process

www.eupati.eu/non-clinical-studies/discovery-development-medicines/
 Cost of drug discovery

Estimates vary widely, but are

on the order of a gigadollar,
taking failures into account.

Tufts Center for the Study of

Drug Development: $2.5 B.

Doctors Without Borders:

$186 M.

www.scientificamerican.com/article/cost-to-develop-new-pharmaceutical-drug-now-exceeds-2-5b/
www.doctorswithoutborders.org/article/rd-cost-estimates-msf-response-tufts-csdd-study-cost-develop-new-drug
Structure-based drug design

Given a target for which the detailed 3D structure is known, how can we design a
ligand to bind to that target?

● Structure determination
● Model preparation
● Virtual screening
● Lead optimization

An alternative is ligand-based drug design, which doesn’t require a receptor

structure:

● Pharmacophore modeling
● Quantitative structure-activity relationships (QSAR)
Protein structure determination

Also NMR (10%), electron microscopy (1%).

www.slideshare.net/anurag_yadav/protein-chemistryiv-anu
Protein Data Bank

www.rcsb.org/pdb/statistics/contentGrowthChart.do?content=total&seqid=100
Structure preparation

Crystal structures aren’t perfect:

● Unresolved sidechains
● No hydrogens
○ Need to decide protonation states
○ Tautomers

Ligand structures also need

preparation.

● 2D → 3D
● Protonation, tautomers

J. Comput. Aided. Mol. Des. 2012, 26, 787–799

Molecular docking

Receptor Ligand Complex (pose)

Approximations in “basic docking”

● Rigid receptor: with a structure already close to the binding conformation.

(Like a “lock and key” model, but with a flexible key!)
● No explicit water
● No covalent interaction with receptor
● Binding site is known
● Ligand is a “small molecule”
● Molecular mechanics model

The problem can be described as a global optimization with 3N degrees of

freedom, where N is the number of ligand atoms.
Molecular mechanics

Energy is modeled using

equations from classical
physics, such as Coulomb’s
law, Hooke’s law, and
approximations such as the
Lennard-Jones potential
and Fourier series.

These equations and the

associated set of empirical
parameters constitute a
force field. Example: OPLS.

J. Chem. Theory Comput. 2016, 2, 281–296

Glide funnel
Scoring function Number of poses

~1000 conformers
Confgen

Conformer ensemble

Docking:
1. Site search
2. Orientation unlimited (may be > 106)
“Greedy” score

5000
“Rough” score 3. Discrete refinement

400
4. Optimize torsions
Emodel / GlideScore and orientation
5
5. Post-min.
Emodel / GlideScore

Pose selection 1
J. Med. Chem. 2004, 47, 1739-1749
Conformational search (“confgen”)
1. Scan each rotatable bond and identify local minima.
2. Enumerate combinations except for the last bond of
each chain, producing ensemble of “core
conformers”.
3. Sample terminal rotatable bonds.
4. Flexible rings are sampled using a pre-generated
template library.

Input ligand (9 rotatable bonds) 214 core conformers (1069 total)

Grids

Computing intermolecular
interactions is expensive.
For N ligand atoms and M
receptor atoms, need N×M
distances. (e.g., 100×10 K
= 1 M).

Trick: pre-compute the

potential due to the receptor
in a “grid” of points in space.
During docking, interpolate
between the nearest 8 grid
points. Then we only need
~N distances!

Grid generation is slow but

only done once. Trilinear interpolation
Docking

Systematic, discrete search of six degrees of

freedom: x, y, z, ɸ, θ, ψ.

For each core conformer:

● Identify ligand diameter and center

● Site search: x, y, z (~2000 sites)
● Diameter search: θ, ψ (302 directions)
● Rotate around diameter: ɸ (25 angles)

Up to 15 million possible poses per core

conformer! But early pruning reduces this by
~99.9%.

Once the core conformer is docked, sample

the terminal rotatable bonds.

After rough scoring, keep 5000 poses.

Post-docking steps

● Partial minimization
○ 400 poses, but only torsions
● Resampling
○ Try non-local torsional jumps on the
docked structure
● Full minimization
○ 5 poses, full xyz minimization
● Final scoring
● Pose selection
○ Glide uses a different scoring
functions just for pose selection.

Total CPU time: 15–30 s/lig (including

confgen, docking, and post-docking).
Scoring

Goal: free energy of binding.

Approximation: empirical scoring function.

- Several special reward and penalty

terms, each with an adjustable weight.

J. Med. Chem. 2006, 49, 6177–6196

Virtual screening (VS)

Goal: find actives (“hits”) in a large library.

Procedure: dock each ligand into the

receptor structure, and rank them by docking
score.

Then pick the top X% for further inspection:

- More accurate computational methods

- Expert visual inspection
- Experimental verification
−ΔG

(Any scoring function can be used, not just

docking. There is also ligand-based VS.)

Binding energy Some error added

distribution
Enrichment and ROC

● ROC: “receiver operating characteristic” (term from WWII radar!)

● Sensitivity: true positive rate (fraction of actives found)
● 1 − specificity: false positive rate (fraction of inactives examined)
● AUC: area under curve measures success at classification
Parallelization

Docking for virtual screening is

embarrassingly parallel: ligands can be
split among multiple CPU cores without any
communication between them.

Example: 1 million ligands, 200 CPUs:

Wallclock time = 106 / 200 × 30 s × (1 h /

3600 s) = 41.7 h
Lead optimization

The lead structure is modified to improve potency, selectivity, pharmacokinetic and

toxicological properties.

Design

Measure
Synthesize
Molecular interactions
Free energy perturbation

Two sets of MD simulations doing

“alchemical transformations”

thermodynamic cycle

⇒ relative binding free energies close

to “chemical accuracy” (1.0 kcal/mol).

Thousands of times slower than

docking; best for lead optimization.
Advanced docking

Constraints

Covalent docking

Induced fit docking

Ensemble docking

Water thermodynamics

Macrocycle docking

Peptide docking
Constraints

Sometimes we already have information about the binding mode (e.g., from a
close analog). We can help ensure that docking finds the right pose by applying
constraints:

Hydrogen bond

Positional / NOE

Metal coordination

Core constraints

Torsional constraints
Unconstrained docking With core constraints
Excluded volumes
Induced Fit Docking

J. Med. Chem. 2006, 49, 534-553

Covalent docking

Covalent ligands are tricky because

the ligand becomes, in effect, part of
the receptor.

Regular (non-covalent) docking can

be used for the initial step of a
covalent-docking workflow which also
includes protein refinement.

J. Chem. Inf. Model. 2014, 54, 1932−1940

Ensemble docking
● Dock each ligand into multiple
variants of the receptor
structure and pick the best one.
● Much faster than IFD (typically
use ~4 receptor structures),
but:
○ Ensemble selection
○ Protein reorganization energy

J. Comput. Aided Mol. Des. 2008, 22, 621–627

Water thermodynamics

WaterMap uses an MD simulation to estimate the

locations of water molecules in the binding site and
the free energy of each.

WScore uses that information to account for the

effect of water displacement on ligand binding.

A) WaterMap water ( ) kept.

B) Reward for displacing water in hydrophobic
environment.
C) Penalty for displacing water in hydrophilic
environment without compensating H-bonds.

J. Am. Chem. Soc. 2008, 130, 2817–2831

J. Med. Chem. 2016, 59, 4364–4384
Macrocycle docking

Macrocycles typically have 12 or more ring

atoms and are very flexible (hundreds of ring
conformations!).

Confgen’s ring-templating approach can’t

cover all possible macrocycles.

A workflow is needed which does a much

more intensive conformational search before
docking.

More conformers than usual need to be kept

during docking.
Peptide docking

Oligopeptides (~2-20 residues) are common

ligands, sized between small molecules and
biologics.

● Long, flexible chains.

● Not well suited for core-based confgen.
● Inaccurate orientation is amplified.

Glide peptide docking mode:

● Enhanced sampling; more conformers.

● More orientations (1002 instead of 302).
● Constraints can be used.
● MMGBSA rescoring.

Success rate went from 21% to 58%.

J. Chem. Inf. Model. 2013, 53, 1689−1699.
Summary

Docking can be useful in early

preclinical drug discovery

● Lead identification (virtual

screening)
● Lead optimization
○ Rationalize molecular
interactions
○ Rough prediction of
binding affinity of analogs
○ Precursor to FEP for more
accurate predictions

The Boston Globe, 04/04/2016

See also: Curr. Opin. Struct. Biol. 2017, 43, 38–44