1. Information Extraction from Literature Yue Lu BeeSpace Seminar Oct 24, 2007

2. Outline Overview of BeeSpace v4 Entity Recognition Relation Extraction

3. Overview BeeSpace V4  deeper semantic base than the current v3 system  entities and relations VS mutual information Four levels  Level1: Entity Recognition  Level2: Entity Association Mining  Level3: Relation Extraction  Level4: Inference and Hypothesis Generation

4. Overview Level1: Entity Recognition (detailed later) Level2 Entity Association Mining  Suppose entities are properly tagged  Utilize the co-occurrence patterns of entities to extract semantics  e.g. a bee biologist may want to know which genes are important for foraging behavior.  Similar to TREC Genomics 2007 task

5. TREC Genomics 2007 e.g. “Which [PATHWAYS] are possibly involved in the disease ADPKD?” currently only retrieval techniques  Gene synonym expansion  Conjunctive query interpretation  User relevance feedback tagged Entities definitely would help

6. Overview Level3: Relation Extraction  Goal is to extract the relations between entities  Generally requires entities to be properly tagged first  Detailed later Level4: Inference and Hypothesis Generation  Inference on knowledge base  Graph mining

7. Outline Overview of BeeSpace v4 Entity Recognition Relation Extraction

8. Entity Recognition Gene Example:  Although mxp and Pb display very similar expression patterns, pb null embryos develop normally

9. Entity Recognition Anatomy Example:  In normal embryos, mxp is expressed in the maxillary and labial segments, whereas ectopic expression is observed in some GOF variants.

10. Entity Recognition Biological process Example:  Amongst these are the Bicoid, the Nanos, and the terminal class gene products, some of which are oncoproteins involved in signal transduction for the formation of terminal structures in the embryo.

11. Entity Recognition Pathways Example:  Several signal transduction pathways have been described in Drosophila, and this review explores the potential of oncogene studies using one of those pathways - the terminal class signal transduction pathway - to better understand the cellular mechanisms of proto- oncogenes that mediate cellular responses in vertebrates including humans

12. Entity Recognition Protein family Example:  While non-arthropod orthologs have been found for many Drosophila eye developmental genes, this has not been the case for the glass (gl) gene, which encodes a zinc finger transcription factor required for photoreceptor cell specification, differentiation, and survival.

13. Entity Recognition CRE (cis-regulatory elements) Example:  A synthetic, 23-bp ecdysterone regulatory element (EcRE), derived from the upstream region of the Drosophila melanogaster hsp27 gene, was inserted adjacent to the herpes simplex virus thymidine kinase promoter fused to a bacterial gene for chloramphenicol acetyltransferase (CAT).

14. Entity Recognition Phenotype Definition:  a set of observable physical characteristics of an individual organism Example:  Fog, dumpy

15. Entity Recognition Class1: Small Variation (Dictionary/Ontology)  Organism, Anatomy, Biological Process, Pathway, Protein Family Class2: Medium Variation  Gene, cis Regulatory Element Class3: Large Variation  Phenotype, Behavior

16. Entity Recognition Generally can be defined as a classification problem Boils down to feature definition  Class1: matching a word in the Dictionary/Ontology  Class2: prefix/suffix of the word, POS tags, …  Class3:?

17. Entity Recognition Firstly focus on Class1  Relatively simple Class2 and Class3 need training examples  Useful in entity association mining  Useful in facilitating extraction of many interesting relations Related work: Textpresso

18. Textpresso Input: full text C. elegans literature Output: tagged XML format Defined a Textpresso ontology  First category is biological entities manually curated a lexicon of names Implemented by PERL regular expressions We could reuse some of the regular expressions

20. Entity Recognition OrganismEntrez gene table, Textpresso, BeeSpace DB AnatomyFlyBase Biological Process, Cellular Component, Molecular Function Textpresso PathwayKEGG Protein FamilyPDB, NCBI Resources:

21. Outline Overview of BeeSpace v4 Entity Recognition Relation Extraction

22. Expression Location  the expression of a gene in some location (tissues, body parts) Homology/Orthology  one gene is homologous to another gene

23. Relation Extraction Biological process  one gene has some role in a biological process Genetic/Physical/Regulatory Interaction  one gene interacts with another gene in a certain fashion (3 types of relations)  a simple case: Protein-Protein Interaction (PPI)

24. Relation Extraction Generally can be defined as a classification problem, which requires training data Domain adaptation?  an example of PPI

25. PPI Problem Definition:  Gene/protein names are already tagged  A known list of interaction words 133 words  classify each tuple (p1, p2, interWord) in one single sentence

26. PPI Methods  Learning algorithm: Maximum Entropy  Context features “Extracting protein-protein interactions using simple contextual features training data” BioNLP Workshop on HLT-NAACL 06 e.g. lexical forms, POS tags … Less dependent on domain

27. PPI Training/Testing data:  BioCreative  1000 hand labeled sentences, 3964 tuples  5-fold cross validation Performance  avgpr = 47.14624  avgre = 43.97337  avgf1 = 45.35523

28. PPI Training data:  BioCreative  1000 hand labeled sentences, 3964 tuples Testing Data (different domain)  Bee collection Performance (Judged by Moushumi)  Total number of tuples extracted as PPI instances: 92  Precision: 63%

29. PPI Misclassification examples Type1: No interaction Sentence: Pretreatment of platelet suspension with phospholipase A2 from N. naja atra or A. mellifera venom (50.mu.g/ml) inhibited platelet aggregation induced by sodium arachidonate or collagen, but not induced by thrombin or ionophore A-23187. False: (collagen, thrombin, induced) True: relation between protein and platelet aggregation; no PPI

30. PPI Misclassification examples Type2: Incorrect interaction word Sentence: IgG antibody was able to inhibit binding of IgE antibody in the PLA radioallergsorbent test (RAST) from 10-40% at a molar excess of 10- to 1000-fold. False: (IgG antibody, IgE antibody, binding) True: (IgG antibody, IgE antibody, inhibit)

31. PPI Misclassification examples Type3: Incorrect protein involved Sentence: AChE exhibits a butyrylcholinesterase (BuChE) activity that represents about 14% of AChE activity. False: (AChE, AChE, exhibits) True: (AChE, BuChE, exhibits )

32. PPI Possible Improvement  syntactic patterns: “Optimizing syntax-patterns for discovering protein-protein interactions” In Proc ACM Symposium on Applied Computing, SAC, Bioinformatics Track,  parse tree  dependency parsing  …

33. The End