Smart mining of drug discovery information using web service workflows

1
Smart mining of drug discovery information using
web service workflows

• David Wild
• Assistant Professor of Chemical Informatics
• Indiana University School of Informatics,
  Bloomington
• djwild _at_ indiana.edu
• March 2006

2
About Me

• B.Sc. in Computing Science
• Ph.D. 1994 in Willett group at University of
  Sheffield UK GAs and parallel processors for
  3D field similarity searching
• Postdocs at Sheffield and Parke-Davis (Ann Arbor)
• Senior Scientist at Parke-Davis / Pfizer
  Scientific Computing RD group
• 2002 Started scientific computing company,
  adjunct professorship in Pharm. Eng. at Michigan
• 2004 Joined SOI in part time visiting position

3
Overview

• Chemical Informatics what and why
• Challenges of diverse sources and large volumes
  of information
• Our research into using web services,
  workflowsand smart agents
• Careers in chemical informatics

4
Chemical informatics is

• More usually know as chemoinformatics or
  cheminformatics
• Very differently defined, reflecting its
  cross-disciplinary nature
• Librarian
• Chemist (synthetic, medicinal, theoretical)
• Biologist / Bioinformatician
• Molecular modeler
• Pharmaceutical or Chemical Engineer
• Computer Scientist / Informatician

5
My definition of Chemical Informatics

• Chemical Informatics (a.k.a. chemoinformatics) is
  the branch of informatics dealing with all
  aspects of the representation and use of chemical
  structures, and related information, on computer.
• It is an interdisciplinary field of that
  regularly pushes the boundaries of computer
  science, statistics, visualization methods,
  computing power and scientific technique. The
  subject covers a wide variety of applications and
  specialties, particularly in the pharmaceutical
  industry, where the rapid increase in new
  technologies in drug discovery puts chemical
  informatics at the forefront of drug design. It
  is foundational to such diverse applications as
  3D molecular modeling, artificial intelligence
  biological activity prediction methods, patent
  and chemical database searching, and high
  throughput screening data analysis.

6
More definitions

• Computational Chemistry The application of
  mathematical and computational methods to
  particularly to theoretical chemistry
• Molecular Modeling Using 3D graphics and
  optimization techniques to help understand the
  nature and action of compounds and proteins
• Computer-Aided Drug Design The discipline of
  using computational techniques (including
  chemical informatics) to assist in the discovery
  and design of drugs.

7
Chemoinformatics hits on Google
Dec 2005 348,100
April 2005 125,600
July 2000 723

Number of word occurrences on Google, Taken from
http//www.molinspiration.com/chemoinformatics.htm
l
8
Virtual screening predicting drug activity
Virtual chemistry learning about the way
compounds work
Analyzing and navigating large volumes of
chemical biological information
9
Example 1High-Throughput Screening
Testing perhaps millions of compounds in a
corporate collection to see if any show activity
against a certain disease protein
10
High-Throughput Screening

• Traditionally, small numbers of compounds were
  tested for a particular project or therapeutic
  area
• About 10 years ago, technology developed that
  enabled large numbers of compounds to be assayed
  quickly
• High-throughput screening can now test 100,000
  compounds a day for activity against a protein
  target
• Maybe tens of thousands of these compounds will
  show some activity for the protein
• The chemist needs to intelligently select the 2 -
  3 classes of compounds that show the most promise
  for being drugs to follow-up

11
Informatics Implications

• Need to be able to store chemical structure and
  biological data for millions of data points
• Computational representation of 2D structure
• Need to be able to organize thousands of active
  compounds into meaningful groups
• Group similar structures together and relate to
  activity cluster analysis
• Need to learn as much information as
  possible(data mining)
• Apply statistical methods to the structures and
  related information

12
Tools for mining the data
Tripos Benchware HTS Dataminer (formerly SAR
Navigator), www.tripos.com
13
Example 23D Visualization Docking

• 3D Visualization of interactions between
  compounds and proteins
• Docking compounds into proteins
  computationally

14
3D Visualization

• X-ray crystallography and NMR Spectroscopy can
  reveal 3D structure of protein and bound
  compounds
• Visualization of these complexes of proteins
  and potential drugs can help scientists
  understand the mechanism of action of the drug
  and to improve the design of a drug
• Visualization uses computational ball and stick
  model of atoms and bonds, as well as surfaces
• Stereoscopic visualization available

15
Visualization Demos (JMOL)

• jmol.sourceforge.net
• www-mslmb.niddk.nih.gov/prag/structures.html

16
Docking algorithms

• Require 3D atomic structure for protein, and 3D
  structure for compound (ligand)
• May require initial rough positioning for the
  ligand
• Will use an optimization method to try and find
  the best rotation and translation of the ligand
  in the protein, for optimal binding affinity

17
Genetic Algorithms

• Create a population of possible solutions,
  encoded as chromosomes
• Use fitness function to score solutions
• Good solutions are combined together
  (crossover) and altered (mutation) to provide
  new solutions
• The process repeats until the population
  converges on a solution

18
Sample GOLD output

• GMP into RNaseT1

19
Something fun

• Screensaver that docks molecules while your
  computer is idle at
• http//www.grid.org/projects/cancer/

20
Chemical Informatics Tools

• Databases chemical structure, biological
  activity, properties, genomic
• Computation docking, pharmacophore generation,
  property calculation, energy minimization,
  cluster analysis, format conversion, alignment,
  2D 3D predictive models (QSAR)
• Visualization analysis plotting, spreadsheet
  views, nonlinear maps, Kohonen maps, 3D molecular
  visualization, structure entry
• Mix of sources commercial vendors, open source,
  academic code
• Very complex to use together

21
Vast increase in quantity of information and
number of data sources

• Until last few years, main challenges in chemical
  informatics were about quality, but now its
  about quality, quantity and managing all the
  information effectively
• Traditionally an issue for the pharmaceutical
  industry but now an issue for academia too (NIH
  Roadmap, PubChem)
• High Throughput Screening can produce
  biological data points for 100,000 compounds per
  day
• Combinatorial Chemistry a single experiment can
  create thousands of new chemicals
• Microarray Assays can produce expression
  information on, e.g., 14,000 genes for a
  drug-treated tissue sample
• Computation techniques can dock hundreds of
  thousands of compounds into proteins or calculate
  millions of properties in a day
• Meta information go/no-go decisions, series
  decisions, patents, etc.

22
Observations about the problem

• Existing approaches do not scale up
• Scientists questions are not that complex, but
  finding the answers is currently very time
  consuming and/or complex (for a human)
• has anybody patented this chemical structure I
  just made?
• can I get hold of a compound that might bind to
  the active site of this protein I just resolved?
• which compounds in this series are least likely
  to exhibit toxic effects?
• Answers are often stale after a short period of
  time questions need to be re-answered as new
  information is generated
• Almost all available systems are passive, and
  follow the(web) browsing model
• There tends to be one interface for every data
  source(or encompassing just a few)

23
4 categories 72 advertisements 1,000 words X
50 newspapers accessed in different ways?
24
However large an array of facts, however rapidly
they accumulate, it is possible to keep them in
order and to extract from time to time digests
containing the most generally significant informat
ion, while indicating how to find those items
of specialized interest. To do so, however,
requires the will and the means we need to
get the best information in the minimum
quantity in the shortest time, from the people
who are producing the information to the people
who want it, whether they know they want it or
not J.D. Bernal, quoted in Murray-Rust et. al.,
Org. Biomol. Chem., 2004, 2, 3192-3203
25
The aim

• An open-source prototype that implements a new
  model of data mining that would, on request,
  push relevant information to pharmaceutical
  scientists in response to previously-defined
  straightforward expressions of needs, rather than
  relying on them stumbling upon the right
  information using traditional browsing models.

26
(No Transcript)
27
(No Transcript)
28
3-layer model
29
Web Services

• Semantic Web Next Big Thing
• Encode semantics in web pages (XML)
• Describes services as well as information (SOAP,
  WSDL, UDDI)
• Computation detached from interface
• Note seeping through to general web usage
• http//www.google.com/apis/
• http//www.amazon.com/webservices
• eScience (UK)
• 200m over 2001-2006 period
• http//www.rcuk.ac.uk/escience/
• Cyber Infrastructure / Grid (US)

30
Request from Human Interface
USE-CASE SCRIPT Invoke New Structure
Service Convert structures to 3D Dock results
protein file Extract any hits Return links for
visualization
AGENT / SMART CLIENT Parse request Select
appropriate use cases and/or web
service(s) Schedule as necessary
UDDI (?)
WSDL
SOAP
Online database (e.g. PubChem)
Local database
3D Docking Tool
2D-3D converter
3D visualizer
New Structure Service Search online
databases for recent structures Search local
databases for recent structures Merge Results
atomic services
aggregate services
31
Prototype development plan

• Develop a handful of use-cases based around
  industry/academia scientists
• Build 5-6 data / computation sources (e.g.
  enumeration, property calculation, structure
  database) that can fulfill the use cases
• Build WSDL and SOAP web services around the data
  sources that can be accessed from Taverna
• Develop workflows in Taverna (see
  taverna.sourceforge.net)
• Publish web services in UDDI
• Encode use-cases into scripts
• Build Intelligent Agent / Smart Client node that
  can match user needs with scripts web services
  using workflows
• Develop browser interface through Contextual
  Inquiry/Usability Studies
• Consider mapping to a Natural Language Interface

32
(No Transcript)
33
Technology

• Perl SOAPLite
• Will be used for initial web service development
• Doesnt really implement WSDL UDDI
• Apache Axis Tomcat
• Deploy WSDL for web services
• BPEL4WS Business Process Execution Language
• For aggregation of web services
• http//www-128.ibm.com/developerworks/library/spec
  ification/ws-bpel/
• Microsoft .NET C

34
A 2D structure is supplied for input into the
similarity search (in this case, the extracted
bound ligand from the PDB IY4 complex)
A protein implicated in tumor growth is supplied
to the docking program (in this case HSP90 taken
from the PDB 1Y4 complex)
Correlation of docking results and biological
fingerprints across the human tumor cell lines
can help identify potential mechanisms of action
of DTP compounds
The workflow employs our local NIH DTP database
service to search 200,000 compounds tested in
human tumor cellular assays for similar
structures to the ligand. Client portlets are
used to browse these structures
Once docking is complete, the user visualizes the
high-scoring docked structures in a portlet using
the JMOL applet.
Similar structures are filtered for drugability,
and are automatically passed to the OpenEye FRED
docking program for docking into the target
protein.
35
Search hits are docked into Acetylcholineesterase
and Visualized using JMOL
Hits returned from similarity search on Donepezil
in NIH DTP database using web service or SQL link
Information retrieved from PubChem using Expert
Query
Expert Query can be used to retrieve related
genomic Information from the web
36
Funding at Indiana

• 500,000 NIH grant to develop an exploratory
  Chemical Informatics Cyberinfrastructure
  Collaboratory over two years. Partnership with
  Informatics and Community Grids Lab. Developing
  web services and workflows particularly relating
  to PubChem and HTS analysis. May lead to 10m
  funding for full center. See www.chembiogrid.org.
  
• 49,000 Microsoft Research Smart Clients for
  eScience grant for development of this system
• For more information, see
• http//www.informatics.indiana.edu/djwild

37
PubChem

• pubchem.ncbi.nlm.nih.gov
• Currently contains 10,096,336 chemical structures
  (3/30/06) and growing rapidly
• Likely to be a (or the) major worldwide source of
  chemical information unless litigation
  restricts it (CAS)
• Includes the latest InChI representations
• Can be queried on chemical structure, exposed as
  web services (SOAP)
• Linked in with MLI project and biological data
  from HTS experiments

38
Supplemental slides
39
Cluster Analysis and Chemical Informatics

• Used for organizing datasets into chemical
  series, to build predictive models, or to select
  representative compounds
• Organizational usage has not been as well studies
  as the other two, but see
• Wild, D.J., Blankley, C.J. Comparison of 2D
  Fingerprint Types and Hierarchy Level Selection
  Methods for Structural Grouping using Wards
  Clustering, Journal of Chemical Information and
  Computer Sciences., 2000, 40, 155-162.
• Essentially helping large datasets become
  manageable
• Methods used
• Jarvis-Patrick and variants
• O(N2), single partition
• Wards method
• Hierarchical, regarded as best, but at least
  O(N2)
• K-means
• lt O(N2), requires set no of clusters, a little
  messy
• Sphere-exclusion (Butina)
• Fast, simple, similar to JP
• Kohonen network
• Clusters arranged in 2D grid, ideal for
  visualization

40
Limitations of Wards method forlarge datasets
(gt1m)

• Best algorithms have O(N2) time requirement (RNN)
• Requires random access to fingerprints
• hence substantial memory requirements (O(N))
• Problem of selection of best partition
• can select desired number of clusters
• Easily hit 4GB memory addressing limit on 32 bit
  machines
• Approximately 2m compounds

41
Scaling up clustering methods

• Parallelisation
• Clustering algorithms can be adapted for multiple
  processors
• Some algorithms more appropriate than others for
  particular architectures
• Wards has been parallelized for shared memory
  machines, but overhead considerable
• New methods and algorithms
• Divisive (bisecting) K-means method
• Hierarchical Divisive
• Approx. O(NlogN)

42
Divisive K-means Clustering

• New hierarchical divisive method
• Hierarchy built from top down, instead of bottom
  up
• Divide complete dataset into two clusters
• Continue dividing until all items are singletons
• Each binary division done using K-means method
• Originally proposed for document clustering
• Bisecting K-means
• Steinbach, Karypis and Kumar (Univ.
  Minnesota)http//www-users.cs.umn.edu/karypis/pu
  blications/Papers/PDF/doccluster.pdf
• Found to be more effective than agglomerative
  methods
• Forms more uniformly-sized clusters at given
  level

43
BCI Divkmeans

• Several options for detailed operation
• Selection of next cluster for division
• size, variance, diameter
• affects selection of partitions from hierarchy,
  not shape of hierarchy
• Options within each K-means division step
• distance measure
• choice of seeds
• batch-mode or continuous update of centroids
• termination criterion
• Have developed parallel version for Linux
  clusters / grids in conjunction with BCI
• For more information, see Barnard and Engels
  talks at http//cisrg.shef.ac.uk/shef2004/confere
  nce.htm

44
Comparative execution timesNCI subsets, 2.2 GHz
Intel Celeron processor
7h 27m
3h 06m
2h 25m
44m
45
Clustering a 1 million compound dataseton a 2.2
GHz Celeron Desktop Machine
Results from AVIDD clusters Teragrid coming
soon.
Time for a single run may vary due to
different selection of seeds. Runtimes can be
shortened e.g. by using a max. number of
iterations or a relocation cutoff.
46
Divisive Kmeans Conclusions

• Much faster than Wards, speed comparable to
  K-means, suitable for very large datasets
  (millions)
• Time requirements approximately O(N log N)
• Current implementation can cluster 1m compounds
  in under a week on a low-power desktop PC
• Cluster 1m compounds in a few hours with a 4-node
  parallel Linux cluster
• Better balance of cluster sizes than Wards or
  Kmeans
• Visual inspection of clusters suggests better
  assembly of compound series than other methods
• Better clustering of actives together than
  previously-studied methods
• Memory requirements minimal
• Experiments using AVIDD cluster and Teragrid
  forthcoming(50 nodes)

47
Visualization interface level tools

• No matter how clever the smarts underneath, the
  overriding factor in usefulness will be the
  quality of scientists interaction with the
  system
• Contextual Design, Interaction Design (Cooper)
  and Usability Studies have proven effective in
  designing the right interfaces for the right
  peoplein chemical informatics collaboration
  with HCI?
• Possibility of multiple interfaces for different
  people groups(Coopers primary personas)
• Dont assume the browser interface email / NLP
  ?
• Start with the basics
• 2D chemical structure drawing (input)
• Visualization of large numbers of chemical
  structures in 2D
• 3D chemical structure visualization
• Planning on evaluation of NLP, email, RSS, etc.
  as well asbrowser-based interfaces

48
Usability of 2D structure drawing tools

• Key difference between sequential and random
  drawers
• Huge difference in intuitiveness
• Key factor how badly you can mess things up
• Marvin Sketch JME gt ChemDraw gtgt ISIS Draw

49
Visualization methods for datasets clusters

• Partitions
• Spreadsheets
• Enhanced Spreadsheets
• 2D or 3D plots
• Hierarchies
• Dendograms
• Tree Maps
• Hyperbolic Maps

50
(No Transcript)
51
(No Transcript)
52
VisualiSAR with a nod to Edward Tufte. See
http//www.daylight.com/meetings/mug99/Wild/Mug99.
html
53
Tree Maps very Tufte-esque
54
3D Visualization - JMOL

• Open Source, very flexible, works in a web
  service environment jmol.sourceforge.net

55
Sentient - an alternative approachto managing
heterogenous data sources

• Collaboration with IO-Informatics (along with
  Cornell, and UCSD) for the investigation of
  service-oriented architectures in life sciences
  research using Sentient software
• Aim to integrate several sources of information
  relating to Alzheimers Disease (brain imaging,
  morphology, gene expression) so that
  cross-dataset biomarkers can be identified
• Sentient usies Intelligent Multidimensional
  Objects (IMOs) to define and query data sources
  and the tools used toaccess them
• Still a browsing approach, but with a layer of
  coherenceand intelligence
• Hope to expand to include chemistry data
• Can also be used as an interface-level tool

56
(No Transcript)
57
(No Transcript)