The Research Assistant for Biological Text Mining

1
The Research Assistant for Biological Text Mining

• Luc Dehaspe
• Other Members of the BioMinT Consortium

2
Text Mining in the biological domain

• Emerging field of research and development
• 40 articles in Bioinformatics 2004
• Dedicated workshops, competitions and interest
  groups
• Information retrieval and extraction to deal with
  information overflow
• 12 million citations in Medline from 4600
  journals
• Many more resources on the web
• Essential link in the semantic integration of the
  numerous biological resources.

3
Use of text mining for database annotation

• curated protein sequence database
• high level of annotation of proteins
• high level of integration with other databases

Swiss-Prot Entry Creation Flowchart
4
Use of database annotations for text mining

• Tools for information retrieval, filtering,
  classification, extraction rely on
• Corpora of examples used by machine learning
  methods
• Linguistic analysis and controlled vocabularies,
  (ontologies, thesauri, biological dictionaries).
• Databases provide semi-structured information
  that could be used
• for corpus elaboration
• as specific vocabulary resources

5

• 3 year FP5 European Project, started in January
  2003
• Official web site www.biomint.org
• Interdisciplinary consortium

6
The goals of BioMinT

• To develop a generic text mining tool that
• interprets different types of queries
• retrieves relevant documents from the biological
  literature
• extracts the required information
• outputs the result as a database slot filler or
  as a structured report
• The tool thus provides two essential research
  support services
• Curator's Assistant accelerate, by partially
  automating, the annotation and update of
  databases
• Researcher's Assistant generate readable reports
  in response to queries from biological
  researchers.

7
Curators Assistant forSwiss-Prot Annotation
8
Curators Assistant for PRINTS annotation

• PRINTS deals with groups of proteins
• Annotation of 3 types of protein fingerprints

Extracted Information
9
The Biological Research Assistant

• Overlap with Curators Assistant
• All biologists occasionally in the curators seat
• Keep ahead of Swiss-Prot in research area of
  interest
• Include private (confidential) document
  collections

10
Information retrieval and extraction modules
11
Information retrieval and extraction modules
G U I
IR
Query expansion
PubMed search
Document filtering/ranking
Document organisation
IE
Sentence extractor
NLP tools
Case frame generator
12
Information Retrieval

• A meta-query engine built round PubMed
• Expansion of the initial query with synonyms
  using a gene/protein synonym database (GPSDB)
• the goal being to retrieve an exhaustive set of
  documents containing information on a protein.
• Filtration and ranking of the retrieved documents
• Pre-classification according to information
  topics.

13
GPSDB

• Database for synonym expansion of gene and
  protein names
• Populated by the main resources on model
  organisms
• Contains 559294 synonyms referring to 292472
  proteins

14
GPSDB

• Cross-reference links are used to connect
  database entries that refer to a same
  gene/protein entity, thus pointing out the
  problem of homonymy when it occurs

15
GPSDB screenshot
lap2 is a synonym of three separate protein
entities
Erbin
HSP 86
Thymopoietin
16
GPSDB screenshot
17
GPSDB used for query expansion
lap2
Original user query
Query expansion based on GPSDB
18
Document filtering and ranking

• Interactive modules which permit a flexible
  selection of relevant documents for the IE
  process.
• Algorithmic approaches
• Query dependent
• Lucene Ranker java-based indexing engine giving
  a ranked output of queried documents
• Query independent
• Naive Bayes Ranker using pre-trained
  classification of relevant documents on specific
  topics

19
Document filtering and ranking
Output of query dependent ranking
20
Document filtering and ranking
Output of query independent ranking with respect
to topic Disease
21
Information retrieval and extraction modules
G U I
IR
Query expansion
PubMed search
Document filtering/ranking
Document organisation
IE
Sentence extractor
NLP tools
Case frame generator
22
Sentence extractor

• Goal extract sentences with information relevant
  for protein annotation
• Method machine learning from corpora with
  manually labeled sentences
• Data representation bag-of-words approach
• Best results with Support Vector Machines
  (linear/Radial Basis Function)

23
Sentence extractorSample output

• set of sentences extracted from the top 5 ranked
  papers
• query-terms are highlighted
• sentences classified according to topics
  (function, structure, disease)
• sentences linked to the PubMed abstract they
  originate from

24
Case frame generator
A protein containing the N-terminal domain with
the first transmembrane segment of MAN1 is
retained in the inner nuclear membrane.
TARGETED_TO X MAN1 Y inner nuclear membrane
25
Case frame generator

• Goal Automatic identification of selected types
  of entities, relations, or events in free text
• Methods
• Given a set of pre-labeled sentences, learn IE
  templates with Inductive Logic Programming (ILP)
• Background knowledge
• Syntactic semantic information from
  shallow-parser
• Ontologies providing entities in a given domain
• Text analysis tools
• Shallow Parser (MBSP) based on Machine Learning
  (TiMBL)
• Shallow parser adapted to biomedical field using
  Genia corpus

26
Case frame generatorSample output shallow parser

• The mouse lymphoma assay (MLA) utilizing the Tk
  gene is widely used to identify chemical mutagens.

Cell-line
The mouse lymphoma assay
MLA
DNA part
to identify
utilizing
chemical mutagens
the TK gene
27
Case frame generatorSample output

• Information extracted by the Case Frame
  Generator, which applied machine learned IE rules
  to output of the Shallow Parser

28
Summary

• The BioMinT prototype is a working unified system
  for Biological Text Mining
• Information Retrieval
• query expansion
• doc filtering/ranking
• Information extraction
• Extraction of sentences on user-specified topics
• Extraction of relationships between entities
  (Case frames)
• Based on variety of resources/technologies/experti
  ses
• Biological sciences corpus annotation, database
  annotation, fingerprints, ontologies,
• Artificial intelligence IR, machine learning
  (SVM, ILP, ), Natural Language Processing
  (Shallow Parser), Case Frames,
• Software development databases, web-server,
  GUI,

29
Future BioMinT developments

• Integration of BioMinT prototype in the future
  annotation environment of Swiss-Prot PRINTS
• Release Q4-2005
• Free web-based version, with restrictions on
• Simultaneous users
• Resources per user (computing storage)
• Customization services provided by PharmaDM
• Integration into researchers IT environment
  (E-mail alerts )
• Mining in-house document collections
• Combination with DMax data analysis software
• Incorporation of highly specialized background
  knowledge (ontologies, thesauri, biological
  dictionaries, etc)
• Custom reports and GUI, etc

30
WWW

• BioMinT home page http//www.biomint.org
• GPSDB synonyms database http//biomint.oefai.at
• BioMinT prototype Quick Tour
• http//biomint-server.pharmadm.com8080/xwiki/b
  in/view/BioMinT/ProtopQuickTour

31
Acknowledgements
Artificial Intelligence
Biological Sciences
Interested? Demo? Leave your card at POSTER 49