1. CIS 630 - Term Project Proposal November 1, 2002 Sharon Diskin
Classification of Gene-Phenotype Co-Occurences in Biological Literature Using Maximum Entropy
CIS Term Project Proposal
November 1, 2002
Sharon Diskin

2. Motivation Numerous biological databases are manually curated
painstakingly slow process, curators review relevant literature
any reliable automation of this process would be of great help
Wealth of biological literature available
Medline currently contains over 12 Million journal articles
Perhaps we can use the manually curated data in the biological databases to help with the task of information extraction from biological literature
Automatic annotation of abstracts or complete articles

3. Pilot Study: Genes and Disease
Classification
OMIM
MorbidMap
Annotated Abstracts
Annotated Abstracts
Annotated Abstracts
(Automatic)
Does this co-occurrence of gene and phenotype belong to our “OMIM Relation”?
WordFreak
Pattern Matching
Annotated Abstracts
Medline
Abstracts
Annotated Abstracts
Annotated Abstracts
(Manual)
OMIM
overview of genes and genetic phenotypes (termed ‘disease’ from here on out)
started by Victor McKusik at Johns Hopkins in late 50’s
maintained by researchers at Johns Hopkins and around world - derived from biological literature
made avail to public through NCBI (National Center for Biotechnology Information) at NIH
approx diseases cataloged - simple mendelian as well as some complex
Medline
over 12 million jounal articles
Andy Schein’s Work
This Term Project

4. Some Examples of Automated Annotation
True Positive:
“Angiotensin converting-enzyme (ACE) inhibitors decrease mortality after myocardial infarction among patients with depressed left ventricular function.”
False Positive:
ACE - enzyme involved in blood pressure regulation
- involved in suseptability to myocardial infarction (heart attack)
NB - here NB is referring to cell lines and not a gene.
Want to be able to distinguish between these….
Classification task divided into 3 phases
feature selection
model building
evalutation
“In the present study, we screened four cell lines of human neuroblastoma (NB-1, NB-16, NB-19, and NH-6) for tumorigenicity and metastatic capacity in nude mice and found that NB-19 cells caused osteolytic lesions after s.c. injection into mice. “

5. Phase I – Feature Selection
Analysis of Corpus
Interested in : binary classification of gene-phenotype pairs that co-occur in a given sentence – in our relation or not?
Question: where are the meaningful words located?
Between gene and disease?
Is it sufficient to only look at a single sentence?
Vocabulary Selection
Simple bag of words with threshold
Top words based on mutual information
Word Counts as Features
Raw counts of words vs. scaled counts
Consider the use of positional information
If only look at sentence level, then perhaps no need to scale

6. Phase II – Maximum Entropy Model
Estimate the conditional distribution of the class label given an instance of gene-phenotype co-occurrence
co-occur in a sentence
labeled instance represented as set of word count features
Use the labeled training data to estimate the expected value of the word counts (features) for each class
training data used to set constraints on conditional probability
Use Improved Iterative Scaling (IIS) to find a classifier of an exponential form which satisfies the constraints represented by the training data
calculate parameters of maximum entropy model
Max Ent - should prefer most uniform

7. Phase III – Evaluation Cross Validation on Labeled Examples
Manual Annotation (Based on Andy’s review of automated annotation)
Automatic Annotation (Based on Andy’s pattern matching)
Some Questions we are interested in:
What is the accuracy?
What are the sources or error?
Poor Feature selection?
Have we oversimplified the problem? How can we improve?
How much does our accuracy suffer if use only automatic annotation?
Can we improve the automated annotation (and hence our classification accuracy?)
Does this method have potential for extracting information from the literature that does not yet exist in a structured database?

8. Potential Plans for the Future
Consideration of other classification methods
Potentially merge work with Ted’s Named Entity tagger for genes
Try to do some information extraction
how is this gene involved in a given phenotype?
Try with other databases
are issues similar?