Example Poster

1
CLIP The Candidate Ligand Identification Program
 
Nicholas Rhodes1, Peter Willett1, Alain Calvet2
and Christine Humblet2

• Background
• Recent improvements in combinatorial synthesis
  techniques have resulted in the availability of
  very large numbers of molecules for
  high-throughput screening (HTS) systems.
  Although very efficient in operation,
  considerations of cost-effectiveness mean that
  screening should be restricted as far as possible
  to molecules that have a reasonable probability
  of being active. This has led to much interest
  in methods for virtual screening, i.e., the
  ranking of a set of compounds in decreasing
  probability of activity so that biological
  testing can be restricted to a (hopefully) small
  fraction of the total number of molecules
  available for consideration.
• One of the most important virtual screening
  techniques is ligand docking. This involves
  determining whether a molecule is complementary
  (in terms of its steric, hydrophobic and
  electrostatic characteristics) to the binding
  site of a protein for which the 3D structure is
  available (typically from X-ray crystallography).
  Several programs for ligand-docking are now
  widely available and although effective in
  operation they can be quite slow, especially when
  an attempt is made to explore the conformational
  space of the potential ligands for the chosen
  target.
• CLIP was designed to provide a fast alternative
  to docking methods, specifically, to meet the
  following criteria
• It should be based on the 3D structures of
  ligands, rather than the 2D structures used in
  conventional similarity searching.
• It should be able to utilise information about
  the binding site if a protein 3D structure is
  available
• It should be sufficiently fast in operation to
  permit the virtual screening of a million
  compounds in an overnight run.

• The programs
• CLIP takes as inputs modified MOL2 files that
  have been pre-classified to include information
  about donors, acceptors, electronegativity etc.
  This classification is done by a Python script
  (CAP.py) on a once and for all basis, using the
  classification scheme proposed by Pepperell et
  al.. One of these inputs is the query template,
  the others are candidate molecules in a database.
  The query is then successively matched against
  each element of the database and the results
  sorted and presented.

• Results
• The actives clustered into 4 groups (UNITY RNN
  clustering) consisting of one group of 36 and
  three singletons. When each of the actives was
  used as a template all of the structures from the
  major cluster retrieved a majority of the actives
  from the same cluster in the top-100 indeed, 11
  retrieved all 36 cluster members. As expected,
  the singletons retrieved only themselves, and in
  one case, one other active molecule, indicating
  that CLIP is highly discriminating.
• The data are presented as cumulative recall plots
  in Figures 2-4 (right)
• ideal situation (actives rated 1-39)
• average (random) or one hit every 5000/39
  structures
• ranking obtained by docking of 3D structures
  into the HIV protease binding site using GOLD in
  command-line mode with default parameters. The
  cavity was centred on atom 242 (D25 OD1) by GOLD
  flood-fill with a radius of 15?.
• rankings obtained from CLIP UNITY 2D searches
  for each of two templates
• Using the templates 0154385 (6 nodes) and 0162034
  (3 nodes), CLIP ranked respectively 12 and 61
  molecules with a coefficient of unity. Note that
  CLIP will not analyse any structures that have
  fewer nodes than the minimum clique size
  parameter. These analyses were performed with
  MINCLQ 3 and only compounds containing matching
  cliques were ranked, so for template 0154386 only
  1822 of a possible 4981 were ranked and only 1155
  for template 162034. Because CLIP ranked many
  structures with identical Simpsons coefficients,
  these were averaged and so the top-ranked
  molecules (Simpsons coefficient is 1) all
  received an equal ranking of 6.5. This technique
  results in some discontinuous jumps in the CLIP
  data series.
•  
• The same two molecules were also used as UNITY
  queries for a default 2D similarity search here,
  by reducing the minimum similarity to zero,
  UNITY effectively ranked the entire data set.
  UNITY is probably marginally more effective than
  CLIP and whilst GOLD is an improvement on random
  selection, it is considerably less effective at
  ranking actives in this data set. However it
  should be noted that GOLD was designed to
  identify the binding modes of small numbers of
  molecule and not for this type of approach.

Figure 2

• The experiments
• Using a set of 5k candidate HIV protease
  inhibitors containing 39 known actives, we
  present performance comparisons of CLIP against
• UNITY 2D fingerprints
• docking of 3D structures into the HIV protease
  binding site using GOLD (Jones et al.)
• In the first of these, though performing a 3-D
  match, CLIP is effectively acting as a similarity
  tool. Ranking was performed using a similarity
  metric based on Simpsons coefficient
• where a is the clique size, b the candidate
  molecule size and c the template size, and where
  the sizes are the number of vertices in that
  graph or subgraph. To remove bias towards large
  molecules, the coefficient was normalised using a
  correction based on the differences between the
  template and candidate intra-node distances, the
  aim being to increase the similarity for cases
  where there was a high measure of agreement in
  the matched distances from the template graph and
  a candidate graph.

Figure 3

• Theory
• CLIP is based on mapping the 3D arrangement of
  pharmacophore features, typically donors and
  acceptors, in a target molecule against either
• Corresponding donors and acceptors in other
  database structures, this being an example of 3D
  similarity searching
• Complementary acceptors and donors in a protein
  binding site, which we will refer to as
  complementary searching
• In both cases, the mapping is generated using a
  3D maximum common subgraph isomorphism algorithm,
  specifically the clique-detection algorithm of
  Bron and Kerbosch that has been used, both by us
  and by other workers, in several previous
  studies. This algorithm was chosen for two
  reasons it has been shown to both effective and
  efficient in operation (Brint, A.T. Willett,
  P. Gardiner, E.J., Artymiuk, P.J. Willett, P.)
  and it is also fairly easy to implement, in
  contrast to several of the other algorithms for
  MCS detection that have been described in the
  graph-theoretic literature.

• Test data
• A subset (5000) of the in-house HIV protease
  database (SMILES) with activities was supplied,
  this subset (candidates) contained a total of 39
  actives which were marked by renaming from
  xyz-0000 to xyz- to facilitate their
  identification by scripts processing result
  files. The SMILES were converted to 3D MOL2
  representations using CONCORD with default
  parameters. 19 molecules failed the conversion,
  none of them active. The resulting MOL2 file was
  then passed to the Python preprocessor, CAP.py,
  giving two files, both containing information on
  likely H-bond formers and one containing
  additional information on aromatic and
  hydrophobic moieties. The results described here
  were all obtained using the former, as aromatic
  and hydrophobic interactions in aspartyl protease
  binding sites were observed to be long-range and
  non-directional.
• Templates were constructed from 3D structures
  with bound inhibitors and also from each of the
  actives. These were used for similarity searches
  against the whole subset.

Figure 4

• Runtimes
• With regard to timings, GOLD processed between
  1.25 and 3.77 structures per hour, depending on
  processor speed and machine load. Assuming two
  structures per hour, the total GOLD runtime was
  around 100 CPU days. CLIP will rank about
  250,000 structures per hour for a 3-node
  structure, and about 150,000 for a 6-node one
  for the dataset in question CLIP took just under
  two minutes for the 6-node structure and 72
  seconds for the smaller one (both well within the
  design criterion of one million compounds in an
  overnight run). The runtimes for UNITY 2D
  similarity searching are comparable to those for
  CLIP (about 3 minutes per search) but the modus
  operandi makes it difficult to time UNITY
  searches accurately.
• It is difficult to compare CLIP against the SYBYL
  3-D searches, there seems to be little difference
  at all between the two approaches in terms of
  effectiveness there is, however, a substantial
  difference in terms of efficiency. Though
  difficult to time because of the way it operates,
  SYBYL takes around 4-5 minutes to search 4755
  compounds. CLIP is considerably faster, taking
  only 72 seconds for the same search. However, it
  is when taking into account the combinatorial
  problem of matching a larger template that the
  real advantage is seen. To search for all 3-point
  matches for a 6-entity template would take SYBYL
  approximately 80 to 100 minutes, CLIP takes
  around 150 seconds (2.5 minutes).

• Conclusions
• CLIP is capable of both similarity and
  complementary matches. In most cases, when doing
  a complementary match with a binding site, the
  sought-for positions of the entities are those of
  the bound ligand so the problem reduces to one of
  taking their positions and inverting the
  donor/acceptor status and is thus equivalent to a
  similarity search in 3D space. The program can,
  however, additionally be used when just the
  protein structure is available without a bound
  ligand. 
• CLIP proved comparable in retrieval performance
  and speed with the fingerprint search, and
  outperformed the docking search (which is, after
  all, designed for more exhaustive exploration of
  a much smaller number of ligands) in both
  respects. CLIP is capable of ranking between
  150k and 250k structures per hour and thus
  provides a fast 3D alternative to traditional 2D
  screening methods.

• Implementation
• Written in entirely in C, CLIP employs a
  user-supplied file of rules to determine whether
  or not two nodes are compatible and a match has
  been made. The current implementation supports 8
  types of node (donor, acceptor, donor-acceptor,
  electronegative, electropositive, ambivalent,
  hydrophobic and aromatic), thus there are 88
  possible matches. In addition, the user specifies
  that one of four match modes (rule sets) is to be
  used
• SIMPLE e.g. donors match with donors
  donor-acceptors
• IDENTITY e.g.donors match only with donors
• COMPLEMENTARY e.g. donors match with acceptors
  donor-acceptors
• FUZZY anything else the user might wish
• The four match modes in CLIP are all equally
  fast. They have been implemented systematically
  rather than specifically CLIP has not been
  programmed with rules relating DONORs, ACCEPTORs
  etc. but can only apply the following predicate
• if (entity1) is compatible with (entity2) then
• return (result)
• CLIP has three levels of configuration
  hard-coded defaults, configuration files and
  command-line parameters for running in script
  mode. For large datasets, writing to disk takes
  place at user-specified intervals (CHUNKSIZE).
  Results are summarised in the main output file
  and details of the cliques (the bulk of the
  output) are written to a separate file, and
  optionally gzip-compressed using the zlib library
  routines.

• References
• Brint, A.T. Willett, P. "Algorithms for the
  identification of three-dimensional maximal
  common substructures." JCICS 27, 1987, 152-158
• Bron, C. Kerbosch, J. Finding all cliques of
  an undirected graph. Communications of the ACM
  16, 1973, 575-577
• Gardiner, E.J., Artymiuk, P.J. Willett, P.
  Clique-detection algorithms for matching
  three-dimensional molecular structures. JMGM 15,
  1998, 245-253
• Jones, G., Willett, P., Glen, R.C., Leach, A.R.
  Taylor, R. "Development and validation of a
  genetic algorithm for flexible docking., JMB
  267, 1997, 727-748
• Pepperell, C.A., Poirrette, A.R., Willett, P.
  Taylor, R.., Development of an atom-mapping
  procedure for similarity searching in databases
  of three-dimensional chemical structures,
  Pestic. Sci. 33, 1991, 97-111

1 Department of Information Studies, University
of Sheffield, Western Bank, Sheffield, S10 2TN. 2
Pfizer Global Research and Development, Ann
Arbor, MI 48105, USA. Acknowledgements This work
was funded by Parke-Davis and Pfizer.
Computational facilities were provided by the
BBSRC.
Figure1 Template created from 1hvi showing the 9
nodes of the inhibitor A77003 and their
interaction nodes in HIV protease.