Presentation at the 5th International Conference on Bioinformatics and Biomedicine BIBM2011, Atlanta GA, November 12-15, 2011 (acceptance rate=19.4%)

1. 48th ACM Southeast Conference. ACMSE 2010.
Oxford, Mississippi. April 15-17, 2010.
Semantic Predications for Complex Information
Needs in Biomedical Literature
Delroy Cameron, RamakanthKavuluru, Pablo N. Mendes
AmitP. Sheth, KrishnaprasadThirunarayan
Ohio Center for Excellence in Knowledge-enabled Computing
kno.e.sisCenter, Wright State University
Dayton, OH 45435, USA
Olivier Bodenreider
National Library of Medicine
Bethesda, MD 20894, USA
2011 International Conference on Bioinformatics and Biomedicine (BIBM)
12-15 November, 2011 Atlanta, Georgia

2. MOTIVATION
• Information Retrieval
 Interaction Sequence
– Keyword Search
– Document Selection
– Document Inspection
– Query Reformulation Exploratory
Search
Document-Centric Model
– Hyperlink-driven Browsing
– Information is within Documents
 Limitations
– Query Reformulation
– Constrained Navigation
2

3. “How do mutations in the Presenilin-1 (PS1) gene affect Alzheimer’s disease (AD)?”
PS1 associated_withAlzheimer’s Disease
. . . mutations in PS1 lead to Alzheimer’s disease by increasing the
extracellular levels of [amyloid peptide 42] A42. (Source: PMID10652366)
Semantic Predications
. . . familial early onset Alzheimer’s disease is caused by point mutations in
the amyloid precursor protein gene on chromosome 21, in the presenilin
2(PS2)1 gene on chromosome 1, or, most frequently, in the presenilin
chromosome14 finding_site_ofpresenilin1
1(PS1) gene on chromosome14. . . (Source:PMID9013610)

4. COMPLEX INFO NEEDS
 Literature-Based Discovery (LBD)
 Don R. Swanson’s Hypotheses
o Raynaud’s Syndrome-Dietary Fish Oil
o Magnesium-Migraine
 Question Answering
 Text REtrieval Conference (TREC)
 2006, 28 questions
TREC Genomics Track - http://ir.ohsu.edu/genomics/
6

5. PROBLEM
 COMPLEX BIOMEDICAL QUESTION
 ANSWER DOCUMENTS
oLEGAL SPANS
 SEMANTIC PREDICATIONS-BASED RETRIEVAL
7

6. REACHABILITY
“the notion of being able to get from one vertex in a
directed graph to some other vertex”
Labeled Graph Calcium Channel Blocker
ISA a
INVERSE_ISA
Magnesium b
c
8

7. REACHABILITY-DOCS
“is the notion of being able to cover the documents in a
document set, usingthe vertices in a directed graph
from one vertex to some other vertex”
Predications Graph Plane
Document Plane
9

9. 11
Knowledge Abstraction
• Stopping Conditions C0007613 – Cell physiology
– No Reachable Docs in PG
– No Successors in PG +Reachable Docs in DP
– No Successors in PG + No Reachable Docs in DP
coexists_with
coexists_with
C1261468 - Cell fusion
C0040682 - cell transformation
Predications Graph Plane
Document Plane

10. 12
Reachability Framework
External Knowledge Base Graph Plane
Knowledge Abstraction
Predications Graph (PG) Plane
Provenance
Document Plane

11. DATASET
• TREC 2006 Corpus
 26 Questions
 162,259 full text documents
 12,641,116 text items
 Gold Standard
 1381 Gold Standard Documents
 3461 Text items
•Biomedical Knowledge Repository (BKR)
 13 million from UMLS Metathesaurus
 8 million from Literature using SemRep
TREC Genomics Track - http://ir.ohsu.edu/genomics/ 13

12. EXPERIMENTS
• Experiment I: Single-Graph
 2105 Vertices, 16942 Edges
> 13,000 unique predications
 240 no predications, 3 not processed
 121,162 text items
 Experiment II: Multiple-Graph
 26 predications graph
14

14. OBSERVATIONS
Absence of predications in text
 Predication extraction methods (SemRep)
 Absence of direct connections among text item
 Ambiguity in written language
Abstractions may lead to information overflow
 Quality of background knowledge
17

15. “How do mutations in the Pes gene affect cell growth?”
DNA Replication
G1_phase G2_phase

16. CONCLUSION
• Novel Knowledge-driven framework
• Semantic Predications to link scientific content
• Alternative to Query reformulation
• Background knowledge boost recall
• Effective at Coarse granularity
• Poor at fine granularity
19

17. FUTURE WORK
• Path/Predication Ranking
• Path/Predication Collapsing
• Knowledge Abstraction tuning
• Scalability
• Computational Complexity
• Replicate Swanson’s Hypotheses
20

18. ACKNOWLEDGEMENT
National Library of Medicine (NIH/NLM)
Human Performance & Cognition Ontology Project @knoesis
CarticRamakrishnan
 Michael Cooney
 Gary Alan Smith
 Paul Fultz II
 Jeffrey Ali Hyacinthe
Thomas C. Rindflesch
 Mohamed Cyclegar
Dongwook Shin
 John Nguyen
 May Cheh
21

19. QUESTIONS
22

20. TREC Genomics Track 2006 (Questions)
Topic ID Question
160 What is the role of PrnP in mad cow disease?
161 What is the role of IDE in Alzheimer's disease
163 What is the role of APC (adenomatouspolyposis coli) in colon cancer?
165 How do Cathepsin D (CTSD) and apolipoprotein E (ApoE) interactions contribute to Alzheimer's disease?
167 How does nucleoside diphosphatekinase (NM23) contribute to tumor progression?
168 How does BARD1 regulate BRCA1 activity?
169 How does APC (adenomatouspolyposis coli) protein affect actin assembly
170 How does COP2 contribute to CFTR export from the endoplasmic reticulum?
172 How does p53 affect apoptosis?
173 How do alpha7 nicotinic receptor subunits affect ethanol metabolism?
174 How does BRCA1 ubiquitinating activity contribute to cancer?
176 How does Sec61-mediated CFTR degradation contribute to cystic fibrosis?
177 How do Bop-Pes interactions affect cell growth?
178 How do interactions between insulin-like GFs and the insulin receptor affect skin biology?
179 How do interactions between HNF4 and COUP-TF1 suppress liver function?
180 How do Ret-GDNF interactions affect liver development?
184 How do mutations in the Pes gene affect cell growth?
185 How do mutations in the hypocretin receptor 2 gene affect narcolepsy?
186 How do mutations in the Presenilin-1 gene affect Alzheimer's disease?

21. 24
Most Relevant Subtree
a
Level-0
Level-1
Level-2
Level-3
Level-4