Biological literature mining

1. Biological literature mining • Information retrieval (IR): retrieve papers relevant to specific keywords • Entity recognition (ER): specific biological entities (e.g., genes) identified in papers • Information extraction (IE): enable specific facts to be automatically pulled out of papers

2. Example sentence “Mitotic cyclin (Clb2)-bound Cdc28 (Cdk1 homolog) directly phosphorylated Swe1 and this modification served as a priming step to promote subsequent Cdc5-dependent Swe1 hyperphosphorylation and degradation” Its context is the cell cycle of the yeast Saccharomyces cerevisiae and it allows us to demonstrate the powers and pitfalls of current literature-mining approaches.

3. Information Retrieval: finding the papers • Aim is to identify text segments pertaining to a particular topic (here, “yeast cell cycle”) • Topic may be a user provided query • ad hoc IR • Topic may be a set of papers • text categorization

4. Ad hoc IR • Pubmed is an example • Supports “boolean model” as well as “vector model” • Boolean model: combination of terms using logical operations (OR, AND) • Vector model: We’ll see more of this later

5. Ad hoc IR: tricks • Lessons learned from regular IR also applicable to biomedical literature • Removal of “stop words” such as the, it, etc. • Truncating common word endings such as -ing, -s • Use of thesaurus to automatically “expand” query • e.g., “yeast AND cell cycle” => “(yeast OR Saccharomyces cerevisae) AND cell cycle”

6. Ad hoc IR “Even with these improvements, current ad hoc IR systems are not able to retrieve our example sentence when they are given the query ‘yeast cell cycle’. Instead, this could be achieved by realizing that ‘yeast’ is a synonym for S. cerevisiae, that ‘cell cycle’ is a Gene Ontology term, that the word ‘Cdc28’ refers to an S. cerevisiae protein and finally, by looking up the Gene Ontology terms that relate to Cdc28 to connect it to the yeast cell cycle.”

7. Entity recognition (ER) • Goal: to identify biological entities (e.g., genes, proteins) in text • Two sub-goals: • recognition of the words in text that represent these entities • unique identification of these entities (the synonym problem)

8. ER goals • In our example, Clb2, Cdc28, Cdk1, Swe1, Cdc5 should be recognized as gene or protein names • Additionally, they should be identified by their respective “Saccharomyces Genome Database” accession numbers • Perhaps the most difficult task in biomedical text mining

9. ER approaches: rule based • Manually built rules that look for typical features of names, e.g., names followed by numbers, the ending “-ase”, occurrences of word “gene”, “receptor” etc in proximity • Automatically built rules using machine learning techniques

10. ER approaches: dictionary based • Comprehensive list of gene names and their synonyms • Matching algorithms that allow variations in those names, e.g., ‘CDC28’, ‘Cdc28’, ‘Cdc28p’ or ‘cdc-28. • Advantage: they can also associated the recognized entity with its unique identifier

11. Why ER is difficult • Each gene has several names and abbreviations, e.g., ‘Cdc28’ is also called ‘Cyclin-dependent kinase 1’ or ‘Cdk1’ • Gene names may also be • common english names, e.g., hairy • biological terms, e.g., SDS • names of other genes, e.g., ‘Cdc2’ refers to two different genes in budding yeast and in fission yeast

12. Information Extraction (IE) • IR extracts texts on particular topics • IE extracts facts about relationship between biological entities • e.g., deduce that • Cdc28 binds Clb2, • Swe1 is phosphorylated by the Cdc28–Clb2 complex • Cdc5 is involved in Swe1 phosphorylation

13. IE approaches: co-occurrence • Identify entities that co-occur in a sentence, abstract, etc. • Two co-occuring entities may be unrelated, but if they co-occur repeatedly, then likely related. Therefore, some statistical analysis used • Finds related entities but not necessarily the type of relationship

14. IE approaches: NLP • Natural Language Processing (NLP) • Tokenize text and identify word and sentence boundaries • Part of speech tag (e.g., noun/verb) for each word • Syntax tree for each sentence, delineating noun phrases and their interrelationships • ER used to assign semantic tags for biological entities (e.g., gene/protein names) • Rules applied to syntax tree and semantic labels to extract relationships between entities

15. Summary • Information retrieval: getting the texts • Entity recognition: identifying genes, proteins etc. • Information extraction: recovering reported relationships between entities

16. Automatically Generating Gene Summaries from Biomedical Literature (Ling et al. PSB 2006) CS 466

17. Outline • Introduction • Motivation • System • Keyword Retrieval Module • Information Extraction Module • Experiments and Evaluations • Conclusion and Future Work

18. Motivation • Finding all the information we know about a gene from the literature is a critical task in biology research • Reading all the relevant articles about a gene is time consuming • A summary of what we know about a gene would help biologists to access the already-discovered knowledge

19. An Ideal Gene Summary • http://flybase.org/reports/FBgn0000017.html GP EL SI GI MP WFPI Above summary is from ca. 2006

20. Problem with Manual Procedure • Labor-intensive • Hard to keep updated with the rapid growth of the literature information How can we generate such summaries automatically?

21. Structured summary on 6 aspects Gene products (GP) Expression location (EL) Sequence information (SI) Wild-type function and phenotypic information (WFPI) Mutant phenotype (MP) Genetic interaction (GI) 2-stage summarization Retrieve relevant articles by keyword match Extract most informative and relevant sentences for 6 aspects. The solution

22. Outline • Introduction • Motivation • System • Keyword Retrieval Module • Information Extraction Module • Experiments and Evaluations • Conclusion and Future Work

23. System Overview: 2-stage IE = Information Extraction; KR = Keyword Retrieval

24. Keyword Retrieval Module (IR) • Dictionary-based keyword retrieval: to retrieve all documents containing any synonyms of the target gene. • Input: gene name • Output: relevant documents for that gene • Gene SynSet Construction • Keyword-based retrieval

25. KR module

26. Gene SynSet Construction &Keyword Retrieval • Gene SynSet: a set of synonyms of the target gene • Issues in constructing SynSet • Variation in gene name spelling • gene cAMP dependent protein kinase 2: PKA C2, Pka C2, Pka-C2,… • normalized to “pka c 2” • Short names are sometimes ambiguous, e.g., gene name “PKA” is also a chemical term • Require retrieved document to have at least one synonym that is >= 5 characters long • Retrieving documents based on keywords: Enforce the exact match of the token sequence

27. Information Extraction Module • Takes a set of documents returned from the KR module, and extracts sentences that contain useful factual information about the target gene. • Input: relevant documents • Output: gene summary • Training data generation • Sentence extraction

28. IE module

29. Training Data Generation • Construct a training data set consisting of “typical” sentences for describing a category (e.g., sequence information) • Training data is not about the gene to be summarized. It is about a “type” of information in general. • These sentences come from a manually curated database • e.g., Flybase has separate sections for each category.

30. Sentence Extraction • Extract sentences from the documents related to our gene • Then try to identify key sentences talking about a certain aspect of the gene (“category”) • In determining the importance of a sentence, consider 3 factors • Relevance to the specified category (aspect) • Relevance to its source document • Sentence location in its source abstract

31. Scoring strategies • Category relevance score (Sc): • “Vector space model” • Construct “category term vector” Vcfor each category c • Weight of term tiin this vector is wij=TFij*IDFi • TFij is frequency of ti in all training sentences of category j • IDFi is “inverse document frequency” = 1+log(N/ni), N = total # documents, ni = number of documents containing ti. • TF measures how relevant the term is, IDF measures how rare it is • Similarly, vector Vs for each sentence s • Category relevant score Sc = cosine(Vc, Vs )

32. Scoring strategies • Document relevance score (Sd): • Sentence should also be related to this document. • Vd for each document, Sd = cos(Vd, Vs ) • Location score (Sl): • News: early sentences are more useful for summarization • Scientific literature: last sentence of abstract • Sl = 1 for the last sentence of an abstract, 0 otherwise. • Sentence Ranking: S=0.5Sc+0.3Sd+0.2Sl

33. Summary generation • Keep only 2 top-ranked categories for each sentence. • Generate a paragraph-long summary by combining the top sentence of each category

34. Outline • Introduction • Motivation • Related Work • System • Keyword Retrieval Module • Information Extraction Module • Experiments and Evaluations • Conclusion and Future Work

35. Experiments • 22092 PubMed abstracts on “Drosophila” • Implementation on top of Lemur Toolkit • Variety of information retrieval functions • 10 genes are randomly selected from Flybase for evaluation

36. Evaluation • Precision of the top k sentences for a category evaluated • Three different methods evaluated: • Baseline run (BL): randomly select k sentences • CatRel: use Category Relevance Score to rank sentences and select the top-k • Comb: Combine three scores to rank sentences • Ask two annotators with domain knowledge to judge the relevance for each category • Criterion: A sentence is considered to be relevant to a category if and only if it contains information on this aspect, regardless of its extra information, if any.

38. Precision of the top-k sentences

39. Discussion • Improvements over the baseline are most pronounced for EL, SI, MP, GI categories. • These four categories are more specific and thus easier to detect than the other two GP, WFPI. • Problem of predefined categories • Not all genes fit into this framework. E.g., gene Amy-d, as an enzyme involved in carbohydrate metabolism, is not typically studied by genetic means, thus low precision of MP, GI. • Not a major problem: low precision in some occasions is probably caused by the fact that there is little research on this aspect.

40. Summary example (Abl)

41. Summary example (Camo|Sod)

43. Outline • Introduction • Motivation • Related work • System • Keyword Retrieval Module • Information Extraction Module • Experiments and evaluations • Conclusion and future work

44. Conclusion and future work • Proposed a novel problem in biomedical text mining: automatic structured gene summarization • Developed a system using IR techniques to automatically summarize information about genes from PubMed abstracts • Dependency on the high-quality training data in FlyBase • Incorporate more training data from other model organisms database and resources such as GeneRIF in Entrez Gene • Mixture of data from different resources will reduce the domain bias and help to build a general tool for gene summarization.

45. References • L. Hirschman, J. C. Park, J. Tsujii, L. Wong, C. H. Wu, (2002) Accomplishments and challenges in literature data mining for biology. Bioinformatics 18(12):1553-1561. • H. Shatkay, R. Feldman, (2003) Mining the Biomedical Literature in the Genomic Era: An Overview. JCB, 10(6):821-856. • D. Marcu, (2003) Automatic Abstracting. Encyclopedia of Library and Information Science, 245-256.

46. Vector Space Model • Term vector: reflects the use of different words • wi,j: weight of term ti in vactor j