E-Chemistry and Web 2.0
Marlon Pierce
mpierce@cs.indiana.edu Community Grids Lab
One Talk, Two Projects
£ NIH funded Chemical Informatics and
Cyberinfrastructure
Collaboratory (CICC) @ IU.
¤ Geoffrey Fox
¤ Gary Wiggins
¤ Rajarshi Guha
¤ David Wild
¤ Mookie Baik
¤ Kevin Gilbert
¤ And others
£ Proposed Microsoft-Funded Project: E-Chemistry
¤ Carl Lagoze (Cornell),
¤ Lee Giles (PSU),
¤ Steve Bryant (NIH),
¤ Jeremy Frey (Soton),
¤ Peter Murray-Rust (Cambridge),
¤ Herbert Van de Sompel (Los Alamos),
¤ Geoffrey Fox (Indiana)
¤ And others
CICC Infrastructure Vision
£ Chemical Informatics: drug discovery and other academic chemistry,
pharmacology, and bioinformatics research will be aided by powerful,
modern, open, information technology.
¤ NIH PubChem and PubMed provide unprecedented open, free data and
information.
¤ We need a corresponding open service architecture (i.e. avoid stove-piped
applications)
¤ CICC set up as distributed cyberinfrastructure in eScience model
£ Web clients (user interfaces) to distributed databases, results of high
throughput screening instruments, results of computational chemical simulations and other analyses.
¤ Composed of clients to open service APIs (mash-ups)
¤ Aggregated into portals
£ Web services manipulate this data and are combined into workflows.
CICC Databases
£
Most of our databases aim to add value to
PubChem or link into PubChem
¤ 1D (SMILES) and 2D structures
£
3D structures (MMFF94)
¤ Searchable by CID, SMARTS, 3D similarity
£
Docked ligands (FRED, Autodock)
¤ 906K drug-like compounds into 7 ligands
¤ Will eventually cover ~2000 targets
Building Up the Infrastructure
£
Our SOA philosophy: use standard Web services.
¤ Mostly stateless
¤ Some cluster, HPC work needed but these populate databases
£
Services are aggregate-able into different
workflows.
¤ Taverna, Pipeline Pilot, …
£
You can also build lots of Web clients.
£
See
http:/
/www.chembiogrid.org/wiki/index.php/CICC_
Web_Resources for li
nks and details.
Sample Services
More Clients…
Example: PubDock
£ Database of approximately 1 million PubChem structures (the most drug-like) docked into
proteins taken from the PDB
£ Available as a web service, so structures can be accessed in your own programs, or using workflow tools like Pipeline Polit
£ Several interfaces developed, including one based on Chimera (right) which integrates the
database with the PDB to allow browsing of compounds in
different targets, or different compounds in the same target
£ Can be used as a tool to help understand molecular basis of activity in cellular or image based assays
Example: R Statistics applied to
PubChem data
£ By exposing the R statistical package, and the Chemistry Development Kit
(CDK) toolkit as web services and integrating them with PubChem, we can quickly and easily perform statistical analysis and virtual screening of
PubChem assay data
£ Predictive models for particular screens are exposed as web services, and can be used either as simple web tools or integrated into other applications
Example assay screening
workflow: finding cell-protein
relationships
A protein implicated in tumor growth with known ligand is selected (in this case HSP90 taken from the PDB 1Y4 complex)
Similar structures to the ligand can be browsed using client
portlets.
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, are converted to 3D, and are automatically passed to the OpenEye FRED docking program for docking into the target protein. The screening data from
a cellular HTS assay is similarity searched for compounds with similar 2D structures to the ligand.
Docking results and activity patterns fed into R services for building of activity models and correlations Leas Squares Regression Rando
Forests NeuraNets
Relevance to Web 2.0
£
Some Web 2.0 Key Features
¤ REST Services
¤ Use of RSS/Atom feeds
¤ Client interfaces are “mashups”
¤ Gadgets, widgets for portals aggregate clients
£
So…
¤ We provide RSS as an alternative WS format.
¤ We have experimented with RSS feeds, using Yahoo Pipes to manipulate multiple feeds.
¤ CICC Web interfaces can be easily wrapped as universal gadgets in iGoogle, Netvibes.
RSS Feeds/REST Services
£
Provide access to DB's via RSS feeds
£
Feeds include 2D/3D structures in CML
£
Viewable in Bioclipse, Jmol as well as Sage etc.
£
Two feeds currently available
¤ SynSearch – get structures based on full or partial chemical names
¤ DockSearch – get best N structures for a target
£
Really hampered by size of DB and Postgres
Tools and mashups based on web
service infrastructure
Mining information from journal
articles
£ Until now SciFinder / CAS only chemistry-aware portal into journal information
£ We can access full text of journal articles online (with subscription)
£ ACS does not make full text available … but there are ways round that!
£ RSC is now marking up with SMILES and GO/Goldbook terms!
¤ www.projectprospect.org
£ Having SMILES or InChI means that we can build a
similarity/structure searchable database of papers: e.g. “find me all the papers published since 2000 which
contain a structure with >90% similarity to this one”
£ In the absence of full text, we can at least use the abstract
Text Mining: OSCAR
£ A tool for shallow, chemistry-specific natural
language parsing of chemical documents (e.g. journal articles).
£ It identifies (or attempts to identify):
¤ Chemical names: singular nouns, plurals, verbs etc., also formulae and acronyms.
¤ Chemical data: Spectra, melting/boiling point, yield etc. in experimental sections.
¤ Other entities: Things like N(5)-C(3) and so on.
£ Part of the larger SciBorg effort
¤ See
http://www.cl.cam.ac.uk/~aac10/escience/sciborg.html)
Mash-Up: What published compounds might bind to this protein? Create a database containing th
text of all recent PubMed abstract (2006-2007 = ~500,000)
Convert molecules to 3D and dock into a protein of interest
Visualize top docked molecules in a
Google-like interface
Use OSCAR to extract all of the chemical names referred to in the abstracts and covert to SMILES
DATABASE SERVICE
DOCKING SERVICE +
E-Chemistry and Digital
Libraries
E-Chemistry and Digital Libraries
£
Key problem with our SOA-based e-Science is
information management.
¤ Where is the service that I need?
¤ What does it do?
£
We may consider our data-centric services to be
digital libraries.
£
Data is diverse
¤ Documents
¤ Not just computational information like structures.
£
Another point of view: how can I link together
publications, results, workflows, etc?
¤ That is, I need to manage digital documents.
Digital Libraries
£ Open Archives Initiative Object Reuse and Exchange Project (OAI-ORE)
£ Developing standardized, interoperable, and machine-readable mechanisms to express information about compound information objects on the web.
£ Graph-based representations of connected digital objects.
£ Objects may be encoded in (for example) RDF or XML,
£ Retrievable via repositories with REST service interfaces (c.f. Atom Publishing Protocal)
Challenges for E-Chemistry
£
Can digital library principals be applied to data as
well as documents?
¤ Can you link your workflow to your conference paper?
£
Can we engineer a publishing framework and
message formats around Web 2.0 principals?
¤ REST, Atom Publishing Protocol, Atom Syndication Format, JSON, Microformats
£
Can we do this securely?
¤ Access control, provenance, identify federation are key problems.
More Information
£
Project Web Site:
www.chembiogrid.org
£
P
roject Wiki: w
ww.chembiogrid.org/wiki
£
Co
ntact me: mpierce@cs.indiana.edu
CICC Combines Grid Computing with Chemical Informatics
CICC
Chemical Informatics and Cyberinfrastucture CollaboratoryFunded by the National Institutes of HealthCICC
www.chembiogrid.org
Indiana University Department of Chemistry, School of Informatics, and Pervasive Technology Laboratories
Science and Cyberinfrastructure
.
Large Scale Computing Challenges
Chemical Informatics is non-traditional area of high performance computing, but many new, challenging problems may be investigated.
CICC is an NIH funded project to support chemical informatics needs of High Throughput Cancer
Screening Centers. The NIH is creating a data deluge of publicly available data on potential new drugs.
CICC supports the NIH mission by combining state of the art chemical informatics techniques with
• World class high performance computing • National-scale computing resources (TeraGrid) • Internet-standard web services
• International activities for service orchestration
• Open distributed computing infrastructure for scientists world wide NIH PubMed DataBas e OSCAR Text Analysis POVRay Parallel Renderin g Initial 3D Structure Calculatio n Toxicity Filtering Cluster Groupin g Docking Molecular Mechanic s Calculatio ns Quantum Mechanics Calculatio ns IU’s Varuna DataBase NIH PubChe m DataBase Chemical informatics text analysis programs can process 100,000’s of abstracts of online journal articles to extract chemical signatures of potential drugs.
OSCAR-mined molecular signatures can be clustered, filtered for toxicity, and docked onto larger proteins. These are classic “pleasingly parallel” tasks. Top-ranking docked molecules can be further examined for drug potential.
Big Red (and the TeraGrid) will also enable us to perform time consuming, multi-stepped Quantum Chemistry
MLSCN Post-HTS Biology Decision
Support
Percent Inhibition or IC50 data is
retrieved from HTS
Question: Was this screen successful?
Question: What should the active/inactive cutoffs be?
Question: What can we learn about the target protein or cell line from this screen?
Compounds submitted to PubChem
Workflows encoding distribution analysis of screening results
Grids can link data
analysis ( e.g image
processing developed in existing Grids), traditional
Chem-informatics tools, as
well as annotation
tools (Semantic Web, del.icio.us) and
enhance lead ID and
SAR analysis
A Grid of Grids linking
collections of services a
PubChem ECCR centers
MLSCN centers
Workflows encoding plate & control well statistics, distribution analysis, etc
Workflows encoding statistical comparison of results to similar screens, docking of compounds into proteins to correlate binding, with activity, literature search of active compounds, etc
CHEMINFORMATICS
Why?
£
Need access to math and stat
functionality
£
Did not want to recode algorithms
£
Wanted latest methods
£
Needed a distributed approach to
computation
¤
Keep computation on a powerful machine
Why R?
£
Free, open-source
£
Many cutting edge methods avilable
£
Flexible programming language
£
Interfaces with many languages
¤
Python
¤
Perl
¤
Java
¤
C
The R Server
£
R can be run as a remote compute
server
¤
Requires the
rserve
package
£
Allows authenticated access over
TCP/IP
£
Connections can maintain state
R as a Web Service
£
On its own the R server is not a web
service
£
We provide Java frontends to specific
functionalities
£
The frontend classes are hosted in a
Tomcat web container
£
Accessible via SOAP
£
Full Javadocs for all available WS’s
Functionality
£
Two classes of functionality
¤
General functions
¥ Allows you to supply data and build a predictive model
¥ Sample from various distributions
¥ Obtain scatter plots and hisotgram
¥ Model development functions use a Java front-end to encapsulate model specific information
Functionality
£
Two classes of functionality
¤
Model deployment
¥ Allows you to build a model outside of the infrastructure
¥ Place the final model in the infrastructure
¥ Becomes available as a web service
¥ Each model deployed requires its own front end class
Available Functionality
£
Predictive models - OLS, RF, CNN,
LDA
£
Clustering - k-means
£
Statistical distributions
£
XY plot and scatter plots
£
Model deployment for single model
types and ensemble model types
Deployed Models
£
Since deployed models are visible as
web services we can build a simple web
front end for them
£
Examples
¤
NC
I anti-cancer predictions
Applications
£
The R WS is not restricted to ‘atomic’
functionality
£
Can write a whole R program
¤ Load it on the R compute server
¤ Provide a Java WS frontend
£
Examples
¤ Feature selection
¤ Automated model generation
¤ Pharmacokinetic parameter calculation
Data Input/Output
£
Most modeling applications require data
matrices
£
Depending on client language we can
use
¤
SOAP array of arrays (2D matrices)
¤
SOAP array (1D vector form of a 2D
matrix)
Data Input/Output
£
Some R web services can take a URL
to a VOTables document
¤
Conversion to R or Java matrices is done
by a local VOTables Java library
£
R also has basic support for VOTables
directly
¤
Ignores binary data streams
Interacting With R WS’s
£
Traditional WS’s do not maintain state
£
Predictive models are different
¤
A model is built at one time
¤
May be used for prediction at another time
¤
Need to maintain state
£
State is maintained by serialization to R
binary files on the compute server
Interacting with R WS’s
£
Protocol
¤
Send data to model WS
¤
Get back model ID
¤
Get various information via model ID
¥ Fitted values
¥ Training statistics
¥ New predictions
Cheminformatics at Indiana
University School of Informatics
David J. Wild
djwild@indiana.edu
Associate Director of Chemical Informatics & Assistant Professor
Indiana University School of Informatics, Bloomington
Cheminformatics education at
Indiana
£ M.S. in Chemical Informatics
¤ 2 years, 36 semester hours
¤ Includes a 6-hour capstone / research project
¤ Opportunity to work in Laboratory Informatics (IUPUI) or closely with Bioinformatics (IUB)
¤ Currently 9 students enrolled
£ Ph.D. in Informatics, Cheminformatics Specialty
¤ 90 credit hours, including 30 hours dissertation research. Usually 4 years.
¤ Research rotations expose students to research in related areas
¤ Currently 4 students enrolled
£ Graduate Certificate
¤ 4 courses, all available by Distance Education
¥ I571 Chemical Information Technology
¥ I572 Computational Chemistry & Molecular Modeling
¥ I573 Programming for Science Informatics
¥ I553 Independent Study in Chemical Informatics
¤ D.E. students pay in-state fees! (~$800 per class)
£ See http://cheminfo.informatics.indiana.edu for more information, or a general review of cheminformatics education in Drug Discovery Today 11, 9&10 (May 2006), pp436-439
Distance Education for
Cheminformatics
£
Uses Breeze + teleconference for live sharing
of classes: all that is required is a P.C. and a
telephone. Optional Polycom
videoconferencing.
£
Lectures are recorded for easy playback
through a web browser
£
Wiki or similar webpage for dissemination of
course materials
£
Also participate in CIC courseshare to give
class at University of Michigan
£
Of 75 students taking our courses since fall
2005, 39 have been D.E. students
£
See JCIM 2006; 46(2) pp 495 - 502 for more
details
Current research in the Wild
lab
£
Integration of cheminformatics tools and data
sources
¤ A web service infrastructure for cheminformatics
¤ Compound information & aggregation web service and interface (“by the way box”)
¤ An enhanced chatbot for exploting chemical information & web services
¤ A semantically-aware workflow tools for cheminformatics
¤ Data mining the NIH DTP tumor cell line database
¤ PubDock: a docking database for PubChem
£
Aggregating life science information from web
and journal documents
¤ Data mining semantically rich chemistry journal articles
¤ Document similarity based on chemical structure similarity
¤ Evaluating semantic markup of chemistry journal articles
£
Integrating cheminformatics into the
chemistry lab
¤ Integrating cheminformatics with the Second Life virtual world
¤ Integrating cheminformatics tools with electronic lab notebooks
¤ Usability of cheminformatics tools
Current research in the Guha
lab
£
Predictive Modeling
¤ Interpretation, validation, domain applicability
¤ Generalization to other ‘models’ such as docking, pharmacophore etc
¤ Integration of multiple data types
¤ Addressing imbalanced and noisy datasets
£
Analysis of Chemical Spaces
¤ Quantify distributions in spaces
¤ Investigation of density approaches
¤ Applications to lead hopping, model domains
£
Methods to summarize & compare data
¤ Applications to HTS and smaller lead series type datasets
£
Network models combining chemical
structures and biological systems
£
Software and infrastructure
¤ Model exchange and annotation
¤ Pharmacophore representations, matching
¤ Toolkit development (CDK)
Cheminformatics web service
infrastructure
Database Services
PostgreSQL + gNova
PubChem mirror (augmented)
Pub3D - 3D structures for PubChem
PubDock - Bound 3D structures
Compound-indexed journal article DB
NIH Human Tumor Cell Line
Local PubChem mirror VARUNA quantum
chemistry database
Statistics (based on R)
Regression, LDA
Neural Nets, Random Forest
K-means clustering Plotting
T-test and distribution sampling
Cheminformatics services
Docking (FRED)
3D structure generation (OMEGA)
Filtering (FRED, etc) OSCAR3
Fingerprints (BCI, CDK) Clustering (BCI)
Toxicity prediction (ToxTree)
R-based predictive models Similarity calculations
(CDK)
Descriptor calculation (CDK)
2D structure diagrams (CDK)
Xiao Dong, Kevin E. Gilbert, Rajarshi Guha, Randy Heiland, Jungkee Kim, Marlon E. Pierce, Geoffrey C. Fox and David J. Wild, Web service infrastructure for chemoinformatics, Journal of Chemical Information and Modeling, 2007;
47(4) pp 1303-1307
RSC Project Prospect - what
can we do with the
information?
£
www.projectprospect.org
£
>
100 papers marked up with SMILES/InChI
(using OSCAR3), plus Gene Ontology and
Goldbook Ontology terms
£
Created similarity searchable PostgreSQL /
gNova database with paper DOIs, SMILES,
and ontology terms
£
Web service and simple HTML interfaces for
searching … “which papers reference
compounds similar to this one in the scope of
these ontological terms?”
£
Applying statistics to look at co-occurrence of
compounds, structural features (MACCS
keys) and ontological terms in papers
Greasemonkey / OSCAR
script
http://cheminfo.informatics.indiana.edu:8080/ChemGM/index.jsp
By the way… annotation
(mock-up!)
By the way…
This compounds is very similar to a prescription drug,Tamoxifen.
This compound is referenced in20 journal articlespublished in the last 5 years Similar compounds are associated with the words “toxic” and “death” in280 web pages
It appears to be covered under3 patents
It has been shown to be active in5 screens
Computer models predict it to show some activity against8 protein targets
Here are some comments on this compound:
Some useful chemical reactions
IodoacetateaIodoacetamideI-CH4COO- ICH2CONH2
This may also react, chem favored byalkaline pH
….
Cheminformatics aware
simple lab notebook (mock
up!)
Free text input can be converted to machine
readable form by electrovaya
Automatic detection o data fields (yield, etc)
Where possible Plug-in allows structures
to be drawn with the pen and cleaned up
Web service interfac provides access to computation and searching.
Page is marked up by what is possible
FIND INFO ABOUT THIS REACTION
Automatic workflow
generation and natural
language queries
£
Develop service ontology using OWL-S or
similar language
¤ Allows service interoperability, replacement and input/outut compatibility
£
We can then use generic reasoning and
network analysis tools to find paths from
inputs to desired outputs
£
Natural language can be parsed to inputs and
desired outputs
£
Smart Clients <--> Agents <--> Services
£
Possible “supercharged life science Google?”
- e.g. type in “what compounds might bind to
the enclosed protein?”
2D -> 3D 2 struct ur crawl er dock 3D searc h P’pho r searc h 2 simila rity 2D structures
2D structures 3D structures
3D structures 3D structures & complexes
dock = bind
3D protei structure
result
3D structures are compounds 2D structures are
compounds
3D structures are compounds