Surflex: Fully Automatic Flexible Molecular Docking Using a Molecular Similarity-Based Search Engine

1
Surflex Fully Automatic Flexible Molecular
Docking Using a Molecular Similarity-Based Search
Engine

• Ajay N. Jain
• UCSF Cancer Research Institute and
  Comprehensive Cancer Center, University of
  California
• Presentation by Susan Tang
• CS 379a
• January 23, 2006

2
Protein-Ligand Docking Overview

• Goal
• - To predict how well a given set of ligands will
  bind to a protein structure
• - To predict the structure of bound
  protein-ligand complexes
• Components
• - Search method explore different ways that
  ligand can interact/fit with protein
• - Scoring function assign a quantitative value
  to each ligand/protein fit

3
Protein-Ligand Docking Overview

• Criteria
• 1) Docking accuracy
• Measures ability to find a conformation
  alignment (pose) of a protein-ligand that is
  close to reality
• 2) Scoring accuracy
• Ability to rank a correct pose of a molecule
  higher than an incorrect one
• 3) Screening utility
• Ability to identify only true ligands in a set
  that contains false positives
• 4) Speed
• How fast the algorithm can screen a library of
  ligands

4
Surflex A new docking methodology

• Combines Hammerheads empirical scoring function
  with a molecular similarity method to generate
  putative poses of ligand fragments
• Like Hammerhead, Surflex has 1 mode that uses an
  incremental construction search approach. But
  Surflex also has another mode a whole molecule
  approach that is faster/more accurate
• Surflex is designed primarily as a screening tool
  for small molecule libraries

5
Surflex Computational Design

• Protomol Generation
• First create an ideal active site ligand from
  the protein structure of interest
• Input
• (a) protein structure (b) list of residues to
  identify protein active site
• Output
• A protomol, or target to which potential ligands
  or ligand fragments are aligned based on
  molecular similarity
• Procedure
• Molecular fragments are put into the protein
  binding site in multiple positions ? optimized
  for interaction with protein ? select
  high-scoring nonredundant fragments ? protomol
  formation

6
Surflex Computational Design

• Protomol for streptavidin compared with the
  native pose of biotin (green)
• The bond being pointed to is broken by Surflex to
  make fragments of biotin for docking.

7
Surflex Computational Design

• Docking
• Ligands are docked into the protein to optimize
  scoring function
• Input
• (a) protein structure, (b) protomol, (c)
  ligand(s)
• Output
• The optimized poses of docked ligands along with
  corresponding scores
• Procedure
• Divide input ligand into 1-10 molecular
  fragments ? search each fragment in terms of
  conformation ? each conformation of each
  fragment is aligned to protomol to get poses with
  maximum molecular similarity to protomol ? score
  aligned fragments and keep those with highest
  score and minimal protein interpenetration ?
  construct full ligand molecule from the aligned
  fragments using either an incremental
  construction approach or whole molecule approach
  ? highest scoring poses undergo further
  refinement of conformation and alignment

8
Surflex Computational Design

• Incremental Construction vs. Whole Molecule
  Algorithm
• Incremental Construction
• - Makes strong assumption that maximizing the
  similarity of tiny fragments to the protomol will
  generate good poses
• Whole Molecule Algorithm
• - bypasses the strong independence assumption
  made in incremental construction
• - dead pieces are carried with the live
  piece during conformation search
• - when creating putative poses to protomol, the
  dead pieces in their arbitrary initial
  conformation are carried into the molecular
  similarity computation ? eliminate those with
  worst protein interpenetration
• - for remaining poses, score on basis of
  individual fragments
• - recursive search yields whole molecules that
  consist of fragments selected from different
  docked poses
• - these whole molecules score well in total,
  over all fragments

9
Surflex Computational Design

• Illustrates the process of docking biotin to
  streptavidin (blue)
• Gray indicates the live fragment
• Magenta indicates the dead fragment
• Green lines show the result of merging the two
  well-docked fragments at the atoms indicated by
  yellow circles
• The merged pose closely follows the parent
  fragments original configurations

10
Surflex Evaluation

• Evaluation of reliability and accuracy of
  dockings
• - Comparison with experimental results on 81
  protein/ligand pairs
• - The pairs were selected to represent
  structural diversity
• Evaluation of Surflexs utility as a screening
  tool
• Performed on 2 protein targets (thymidine kinase
  and estrogen receptor)
• Competing docking methods were tested side by
  side using the same data set for comparison
  purposes (GOLD, Dock, FlexX)
• Evaluation of the Surflexs docking speed
• - Investigate relationship between docking
  time and of rotatable bonds

11
Surflex Evaluation Data Set Construction
134 protein-ligand Complexes
81 protein-ligand complexes
filter

• Filtering Criteria
• 15 or fewer rotatable bonds
• ? Most small molecules have lt 15 rotable bonds
• no covalent attachments between ligand and
  protein
• ? Since Surflexs scoring function was developed
  strictly on noncovalent complexes
• ligands with no obvious errors in structure
• ? Undesirable to modify an existing
  protein-ligand complex prior to testing
• data set used for GOLD docking program

12
Surflex EvaluationResults

• 1) Evaluation of reliability and accuracy of
  dockings
• Describes how thorough the search procedure is
  and to what extent scoring function can recognize
  good dockings
• Surflex returned a pose within 2.5 angstroms rmsd
  (94 of cases)
• Surflex returned a BEST scoring pose that was
  within 2.5 angstroms (86 of cases)
• With a single docking from a random initial pose,
  chances of finding a correct or nearly correct
  pose is averaged to be 70

13
Surflex EvaluationResults
14
Surflex EvaluationResults

• 2) Evaluation of Surflexs utility as a screening
  tool
• Tests ability of program to detect true
  positives against a background of random
  molecules (sensitivity vs. specificity)
• Surflex had a True Positive rate of gt 80 at a
  False Positive rate of lt 1
• Surflex had the best performance (lowest FP rate
  for a given TP rate) out of the different
  individual and combined methods assayed

15
Surflex EvaluationResults

• 3) Evaluation of the Surflexs docking speed
• Docking speed becomes very important in
  screening large compound libraries.
• Surflex demonstrated a docking time that was
  approx. linear in number of rotatable bonds
• Rigid molecules took a few seconds and each
  additional rotatable bond took an additional 10
  seconds
• Surflex yielded a mean running time of 44 seconds
  for the 81 protein-ligands in the test set used
  earlier
• Docking speed ranges from 50-100 seconds per
  molecule for FlexX, DOCK, and GOLD (Surflex speed
  is comparable to these times)
• Quantitative comparison across methods is
  difficult due to differences in hardware and
  methodology

16
Surflex EvaluationResults
17
Conclusions

• Surflex marks a step forward in flexible
  molecular docking programs
• Compared to the best docking methods available,
  Surflex is
• as fast
• as accurate in terms of docked ligand RMSD
• much more accurate in terms of scoring
• Assaying the top scoring 1 of compounds in the
  screening library should yield a large proportion
  of true positives
• Potential areas of improvement
• - scoring and penetration terms should be
  combined into a single score
• - scoring function should include training on
  non-binding ligands (negative examples)
• - effect of nonbonded self-interactions within
  ligands should be accounted for explicitly
• - allow a degree of protein flexibility (side
  chain movement)