Julia Salas

1

• Julia Salas
• CS379a
• 1-24-06

2
Aim of the Study

• To survey the docking and scoring algorithms
  available today
• Evaluate protocols for three tasks
• 1. Prediction of the conformation of ligand
  bound to protein target
• 2. Virtual screening of database to identify
  leads
• 3. Prediction of binding affinities
• General Methods
• Investigate several docking programs using a
  variety of different target types
• Use a large set of closely related compounds
  (compound set) for each target type

3
Target Types/Targets Used

• Target Types 7 protein classes represented
• Targets 8 proteins of interest to GSK
• Variety Diversity of mechanisms, binding site
  shape, binding site chemical environment

4
Compound Sets Used (Ligands)

• Goal Represent a typical pharmaceutical compound
  collection

• Compound/Ligand Sets 1303 compounds
• 150-200 closely related compounds
• Compounds have experimentally determined
  affinities
• Affinities of compounds in a single set span a
  min of 4 orders of magnitude
• Each set has shown biological activity towards
  target protein
• Each set has a max of 20 inactive and 20
  extremely active compounds
• Each set has published (2-54) cocrystal
  structures with the target protein

5
Compound Sets Used (Ligands)

• zdc

6
Docking and Scoring Algorithms

• Docking Algorithms
• Evaluated 10 programs with different algorithms
  and scoring functions
• 19 protocols total
• Procedure
• Each method evaluated by an expert, no time
  restrictions or other constraints
• Evaluators did not have cocrystal structures,
  only ligand structure and protein active site
  residues

• Same ligand starting structure
• Optimized to a (local) min
• Reasonable bond distances/angles
• Correct atom hybridization
• 4 structures provided (differ in ionization)
• SMILES (text-based) structure description

7
Analysis of Docking Programs and Scoring Functions

• 19 protocols evaluated on three tasks
• 1. Prediction of the conformation of ligand
  bound to protein target
• 2. Virtual screening of database to identify
  leads
• 3. Prediction of binding affinities

8
Prediction of Ligand Conformation Bound to
Protein Target

• Compare predictions to (136) cocrystal structures
  using
• 1. rmsd for heavy atoms
• 2. Volume overlap Tanimoto similarity index

• Two standards for success rmsd within
• 2Å (correct orientation) Black Bars
• 4Å (within binding site) Gray Bars
• Can evaluate both the scoring function and the
  overall methods

IX, ID Vol overlap integrals for crystal and
docked structure OX,DVol overlap between crystal
and docked pose 0 Tvol 1
9
Prediction of Ligand Conformation Bound to
Target Conclusions

• The good
• Docking programs could generate crystal
  conformations
• For all (-HCVP) targets, at least one program
  could dock 40 of ligands within 2
• 90 of ligands could be docked with 4Å with 100
  docked in correct location

• The bad
• Program with best performance changes target to
  target
• Scoring function lead to consistently incorrect
  predictions
• HCVP had very weak predictions

10
Virtual Screening of Database to Identify Leads

• Ability to identify the active compounds
• Enrichment How quickly did the protocol identify
  the active compound vs. random chance?
• Success Identify at least 50 of the active
  compounds within the top 10 of the score-ordered
  list ? halfway between random and max.

• Lead Identification Cost analysishow many
  compounds do you need to screen to find at least
  one active compound from each class?
• All active compound classes IDd within top 10
• Percent actives vs. percent compounds
  screened measured

11
Prediction of Binding Affinities

• Calculated docking scores compared to measured
  affinity
• Docking scores were autoscaled and then compared
• Conclusions
• No statistically significant correlation between
  scoring function and measured affinity

12
Conclusions and Discussion Questions

• Docking programs were able to generate poses that
  resemble cocrystal structures
• Largest difficulties were in determining the
  small molecule structure, not placing ligand in
  binding site
• Scoring functions were not successful in
  predicting the best structures
• Active compounds could be identified in a pool of
  decoys
• Docking scores could not be correlated to
  affinity
• Question 1 What factors may have contributed to
  the failure of these programs to predict small
  molecule conformation?
• Question 2 The failure of the programs to
  predict HCVP structures was attributed to the
  enzymes large active site. Why? Additionally,
  should flexibility/dynamics be considered?
• Question 3 Compound classes were defined by
  similar backbone structure. Although all
  compounds in a class had measured affinities, can
  we assume they all have the same binding mode?