Ligand

docking

CS/CME/BioE/Biophys/BMI 279
Oct. 25 and 27, 2016
Ron Dror

1
 Outline
• Goals of ligand docking
• Defining binding affinity (strength)
• Computing binding affinity: Simplifying the problem
• Ligand docking methodology
• How well does docking work?

2
Goals of ligand docking

3
 A drug binding to its target
(The great majority of drug targets are proteins)

Beta-blocker alprenolol binding to an adrenaline receptor

Dror et al., PNAS 2011
 Problem definition
• A ligand is any molecule that binds to a protein
– We’ll also use ligand to refer to any molecule that might bind to a protein
(e.g., any candidate drug)
• Ligand docking addresses two problems:
– Given a ligand known to bind a particular protein, what is its binding pose
(that is, the location, orientation, and internal conformation of the bound
ligand)
– How tightly does a ligand bind a given protein?

http://www.nih.gov/researchmatters/
october2012/images/structure_l.jpg 5
 Why is docking useful?

• Virtual screening: Identifying drug candidates by

considering large numbers of possible ligands
• Lead optimization: Modifying a drug candidate to
improve its properties
– If the binding pose of the candidate is unknown,
docking can help identify it (which helps envision how
modifying the ligand would affect its binding)
– Docking can predict binding strengths of related
compounds

6
Ligand docking: a graphical summary

7
http://www.slideshare.net/baoilleach/proteinligand-docking-13581869
Defining binding affinity (strength)

8
 How do we measure how tightly a
ligand binds to a protein?
• Binding affinity quantifies the binding strength of a ligand to a
protein (or other target)
– Conceptual definition: if we mix the protein and the ligand (with no
other ligands around), what fraction of the time will the protein have
a ligand bound?
• This depends on ligand concentration, so we assume that the ligand is
present at some standard concentration.
– Binding affinity is usually expressed as either:
• The difference ΔG in free energy of the bound state (all atomic
arrangements where the protein is ligand-bound) and the unbound
state (all atomic arrangements where the protein is not ligand-bound)
– Again, assume standard concentration of ligand
– From ΔG, one can compute the fraction of time the ligand will be bound
• A dissociation constant (Kd) which is (roughly) the ligand concentration
at which half the protein molecules will have a ligand bound 9
Computing binding affinity: 

Simplifying the problem

10
 Direct approach to computing
binding affinity
• Run a really long molecular dynamics (MD)
simulation in which a ligand binds to and unbinds
from a protein many times.
• Directly observe the fraction of time the ligand is
bound.

11
 The direct approach doesn’t work
• It is so computationally intensive that we usually
cannot do it for even a single ligand
– Drug molecules usually take seconds to hours to unbind
from their targets.
– Microsecond-timescale molecular dynamics simulations
usually take days.

12
 What can we do instead?
Option 1: Use alternative MD-based approaches
• It turns out that one can compute binding affinities by
MD in more efficient ways
– These methods, called free energy perturbation (FEP) and
thermodynamic integration (TI), are very clever
– They represent the most accurate way to determine binding
affinities computationally
– They are very expensive computationally and thus cannot be
used on large numbers of ligands
– They assume that one knows the binding pose
• There are also methods based on implicit solvent MD
simulation (water molecules not represented explicitly)
– These methods are faster, but still computationally intensive
– They are somewhat less accurate
– They again assume that one knows the binding pose 13

You are not responsible for any of the methods on this slide
 Option 2: Ligand docking
• Ligand docking is a fast, heuristic approach with two key
components
– A scoring function that very roughly approximates the binding affinity
of a ligand to a protein given a binding pose
– A search method that searches for the best-scoring binding pose for
a given ligand
• Most ligand docking methods assume that
– The protein is rigid
– The approximate binding site is known
• That is, one is looking for ligands that will bind to a particular site on the
target
• In reality, ligand mobility, protein mobility, and water molecules
all play a major role in determining binding affinity
– Docking is approximate but useful
– The term scoring function is used instead of energy function to 14
emphasize the highly approximate nature of the scoring function
 Docking software
Most popular
(based on citations
2001–2011):
!
AutoDock
GOLD
DOCK
FlexX
Glide
FTDOCK
QXP
Sousa et al., Current
Medicinical Chemistry
2013

http://en.wikipedia.org/wiki/
You are not responsible for the details on this slide Docking_(molecular)
Ligand docking methodology

16
 Scoring functions

• Scoring functions used for docking tend to be

empirical
– Capture chemists’ intuition about what makes a a ligand–
receptor interaction energetically favorable (e.g., hydrogen
bonding, or displacement of water from a hydrophobic
binding pocket)
– Parameters are often optimized based on known binding
affinities of many ligands for many receptors
– Some scoring functions borrow terms from molecular
mechanics force fields, but a molecular mechanics force
field is rarely used directly as a scoring function for docking
• The scoring function is an (extremely rough) attempt to
approximate the binding free energy. Molecular mechanics
force fields give potential energy associated with a particular17
arrangement of atoms.
 Example: Glide scoring function
• Glide (considered one of the most accurate docking software packages) uses the
following “GlideScore” function in SP (“standard precision”) mode:
!
! Friesner et al., Journal of
Medicinal Chemistry
! 47:1739-49 (2004)

!
!
– The first term rewards contacts between hydrophobic atoms of the ligand and protein, and is a
function of the distance between them
– The next several terms reward specific kinds of hydrogen bonds, and are a function of both
angle and distance
• The final ranking of ligands in Glide SP is determined by a combination of the
GlideScore, an interaction energy computed using a molecular mechanics force field
(OPLS-AA), and an estimate of the internal strain of the ligand in the bound pose
• Glide’s XP (“extra precision”) mode uses an even more complicated scoring function18
You are not responsible for the details on this slide
 Search methods
• Docking software needs to search for the best-scoring
pose for each ligand
• The search space is huge, because one needs to
consider all possible ligand positions and orientations,
and the ligand’s internal degrees of freedom
• To search this space efficiently, docking software
typically employs some combination of:
– Heuristic assumptions about what poses will/won’t work
– Monte Carlo methods
– Heirarchical methods in which one uses approximate
measures to identify promising groups of poses, then
evaluates them in more detail
19
 Example: Glide search
• Glide SP uses a hierarchical
search method
• It first identifies a discrete set
of “reasonable” conformations
for each ligand, by varying
internal torsion angles
• For each ligand, it scans
possible positions and
orientations, using a rough
metric of fit
• The most promising
approximate poses undergo
further “refinement” and Friesner et al., J Med Chem 47:1739, 2004
evaluation 20

You are not responsible for the details on this slide

How well does docking work?

21
 How well does docking work?

• The best docking protocols:

– Predict a reasonably accurate pose (for ligands that do
in fact bind the target protein) a little more than half the
time
• Usually one of the few top-ranked poses is close to the
correct one
– Provide useful, but far from perfect results, when
ranking ligands
• Tend to work best when comparing closely related
ligands
– Are not particularly useful when it comes to
quantitatively estimating binding free energies
22
Leach et al., J Med Chem 49:5851 (2006)
Warren et al., J Med Chem 49:5912 (2006)
 How well does docking work?
Example: Performance of Glide on ligand-ranking tests for multiple targets.

Good
performance
on these
targets
Different
Poor (near- target
random) proteins
performance
on these

Warren et al., J Med Chem 49:5912 (2006)

 How well does docking work?
Example: Correlation between docking scores and affinity for one target

Magenta points correspond

to ligands from one chemical
Docking score (FlexX)

family. Blue points

correspond to a second
chemical family.

Magenta points: decent

correlation between docking
score and affinity.
pAffinity = –log(Kd) !
Blue points: no correlation.

Warren et al., J Med Chem 49:5912 (2006)