Literature Based Discovery Dimitar Hristovski Institute of Biomedical Informatics, Faculty of Medicine, University of Ljubljana, Ljubljana, Slovenia
Let me introduce myself … Research and Development BS – Biomedicina Slovenica database Research Evaluation Decision Support System Medical Information Systems –Surgical clinics –Genetic laboratory –Biochemical laboratory Web User Behaviour Analysis Data warehousing and OLAP
Motivation Overspecialization Information overload Large databases For many diseases the chromosomal region known, but not the exact gene
Background Literature-based discovery (Swanson): Concept X (Disease) Concepts Y (Pathologycal or Cell Function, …) Concepts Z (Genes) New Relation?
Biomedical Discovery Support System (BITOLA) Goal: –discover potentially new relations (knowledge) between biomedical concepts –to be used as research idea generator and/or as –an alternative way to search Medline System user (researcher or intermediary): –interactively guides the discovery process –evaluates the proposed relations
Extending and Enhancing Literature Based Discovery Goal: –Make literature based discovery more suitable for disease candidate gene discovery –Decrease the number of candidate relations Method: –Integrate background knowledge: Chromosomal location of diseases and genes Gene expression location Disease manifestation location
Usage Scenarios For a disease with known chromosomal location, find a candidate gene For a gene, find a disease that might be influenced For a disease and gene found to be related by linkage study, find the mechanism of the relation (intermediate concepts should help)
System Overview Knowledge Base Concepts Association Rules Background Knowledge (Chromosomal Locations, …) Discovery Algorithm User Interface Databases (Medline, LocusLink, HUGO, OMIM, …) Knowledge Extraction
Databases Medline: source of known relationships between biomedical concepts Set of concepts: –MeSH (Medical Subject Headings): Controlled dictionary and thesaurus used for indexing and searching the Medline database –HUGO: official gene symbols, names and aliases –LocusLink: gene symbols, aliases and chr.locations –OMIM: genetic diseases UMLS (Unified Medical Language System) Entrez: used to search PubMed, GenBank,... UniGene: gene expression
Knowledge Extraction Build master set of concepts (MeSH terms and gene symbols) Extract occurrence of concepts from each Medline record (MeSH terms from MH field, gene symbols from Title and Abstract) Association rule mining (concept co-occurrence) Chromosomal location extraction (from LocusLink and HUGO) Load into knowledge base
Terminology Problems during Knowledge Extraction Gene names Gene symbols MeSH and genetic diseases
Detected Gene Symbols by Frequency type| II| III| component| CT| AT| ATP| IV| CD4|99657 p53|89357 MR|88682 SD|85889 GH|84797 LPS| |67272 E2| |63521 AMP|61862 TNF|59343 RA|58818 CD8|57324 O2|56847 ACTH|54933 CO2|53171 PKC|51057 EGF|50483 T3|49632 MS|46813 A2|44896 ER|43212 upstream|41820 PRL|41599
Gene Symbol Disambiguation Find MEDLINE docs in which we can expect to find gene symbols JD indexing (Susanne Humphrey) as possible solution: –Identifies the semantic context of docs –If semantic context not genetic, then gene symbol probably false positive Example of false positive: –Ethics in a twist: "Life Support", BBC1. BMJ 1999 Aug 7;319(7206):390 –breast basic conserved 1 (BBC1) gene, v.s. BBC1 television station featuring new drama series Life Support
JD Indexing JDs are 127 Journal Descriptors (e.g., JDs for journal Hum Mol Genet: Cytogenetics; Genetics, Medical) Training set docs (435,000) inherit JDs from journals Training set provides co-occurrence data between inherited JDs and: –indexing terms assigned to docs directly –words in docs Docs having indexing terms/words occurring often with genetics JDs in tr. set assumed to have genetics context Extended to indexing by 134 UMLS semantic types (e.g. Gene or Genome, Gene Function,…)
System Overview Knowledge Base Concepts Association Rules Background Knowledge (Chromosomal Locations, …) Discovery Algorithm User Interface Databases (Medline, LocusLink, HUGO, OMIM, …) Knowledge Extraction
Binary Association Rules X Y (confidence, support) If X Then Y (confidence, support) Confidence = % of docs containing Y within the X docs Support = number (or %) of docs containing both X and Y The relation between X and Y not known. Examples: –Multiple Sclerosis Optic Neuritis (2.02, 117) –Multiple Sclerosis Interferon-beta (5.17, 300)
Discovery Algorithm Concept X (Disease) Concepts Y (Pathologycal or Cell Function, …) Concepts Z (Genes) Chromosomal Region Chromosomal Location Candidate Gene? Match Manifestation Location Expression Location Match
Discovery Algorithm Let X be starting concept of interest. Find all Y for which X Y. Find all Z for which Y Z. Eliminate those Z for which X->Z already exists. Eliminate those Z that do not match the chromosomal region of X Eliminate those Z that do not match the expression location of X Remaining Z are candidates for new relation between X and Z. In general: X Y 1 … Y n Z, but not X Z Example: X = disease Y = (pato)physiology of X Z = (de)regulators of Y (drugs, proteins, genes) New relation example: Z is candidate gene for disease X
Ranking Concepts Z X Y1Y1 Y2Y2 Y3Y3 YiYi YjYj … … Z1Z1 Z2Z2 Z3Z3 ZkZk ZnZn
Results: Concepts in Medline Full Medline (end 2001) analyzed (11,226,520 recs) Looking for 19,781 MeSH terms and 22,252 human genes (14,659 from HUGO and 7,593 from LocusLink). 24,613 alias gene symbols added Gene symbols found in 2,689,958 Medline recs. Most frequent ambiguous symbols (CT, MR, CO2,…) or format errors
Results: Co-occurring Concepts in Medline 29,851,448 distinct pairs of co-occurring concepts: –In 7,106,099 at least one gene symbol appeared –In 679,159 pairs both concepts are gene symbols Total co-occurrence frequency: 798,366,684 59,702,986 association rules calculated and stored
Bilateral Perisylvian Polymicrogiria - BPP (OMIM: ) Polymicrogyria of the cerebral cortex is a developmental abnormality characterized by excessive surface convolution Clinical characteristics: –Mental retardation –Epilepsy –Pseudobulbar palsy (paralysis of the face, throat, tongue and the chewing process) X linked dominant inheritance
It is considered a disorder of neuronal migration (unlayered type) or a consequence of intrauterine ischemia (layered type) BPP - pathogenesis
Finding Candidate Genes for Polymicrogyria, bilateral perisylvan
18 gene candidates 15 gene candidates Tissue specific expression 2 gene candidates: L1CAM and FLNA relation between semantic types Cell Movement and Gene or gene products Sublocalisation in the Xq genes in Xq28
User Interface “cgi-bin” version
Automatically search for supporting Medline Citations
Cleft Palate – Predicting Candidate Genes
Summary and Conclusions We extend and enhance an existing discovery support system (BITOLA) The system can be used as: –Research idea generator, or –Alternative method of searching Medline Genetic knowledge about the chromosomal locations of diseases and genes included to make BITOLA more suitable for disease candidate gene discovery
Further Work Increase the number of concepts Gene symbol disambiguation Semantic relations extraction System evaluation Improve the Web version of the system
System Availability URL:
Related work: SemGen Tom Rindflesch et al Extract semantic predications on genetic basis of disease “Deletions of INK4 occur in malignant tumors” –INK4|ASSOCIATED_WITH|Malignant Tumors Evaluation and visualization of SemGen output
Semantic Structures CAUSEPREDISPOSEASSOCIATED_WITH ETIOLOGY_OF cause determine result in control underlie transmit responsible predispose lead to promote susceptibility risk associate involve link implicate influence related
Statistical Evaluation Assoc. rule base divided into 2 segments: older ( ) and newer ( ) The system predicts new relations based on the older segment Predictions compared with actual new relations in the newer segment
Summary Statistical Evaluation Results
Statistical Evaluation Results With no assoc. rules constraints: – predicts almost all new relations, but too many candidate relations With constraints: –predicts new relations 6.9 times better than random predictions –tighter the constraints, better (correct / all predictions) ratio (6.5%)