Interpreting Microarray Expression Data Using Text Annotating the Genes

1
Interpreting Microarray Expression DataUsing
Text Annotating the Genes

• Michael Molla, Peter Andreae, Jeremy Glasner,
  Frederick Blattner, Jude Shavlik
• University of Wisconsin Madison

2
The Basic Task

• Given
• Microarray Expression Data
• Text Annotations of Genes
• Generate
• Model of Expression

3
Motivation

• Lots of Data Available on the Internet
• Microarray Expression Data
• Text Annotations of Genes
• Maybe we can Make the Scientists Job Easier
• Generate a Model of Expression Automatically
• Easier First Step for the Human

4
Microarray Expression Data

• Each spot represents a gene in E. coli
• Colors Indicate Up- or Down-Regulation Under
  Antibiotic Shock
• Four our Purpose 3 Classes
• Up-Regulated
• Down-Regulated
• No-Change

5
Microarray Expression Data
From Genome-Wide Expression in Escheria Coli
K-12, Blattner et al., 1999
6
Our Microarray Experiment

• 4290 genes
• 574 up-regulated
• 333 down-regulated
• 2747 un-regulated
• 636 non enough signal

7
Text Annotations of Genes

• The text from a sample SwissProt entry (b1382)
• The description field
• HYPOTHETICAL 6.8 KDA PROTEIN IN LDHA-FEAR
  INTERGENIC REGION
• The keyword field
• HYPOTHETICAL PROTEIN

8
Sample Rules From a Model for Up-Regulation

• IF
• The annotation contains FLAGELLAR AND does NOT
  contain HYPOTHETICAL
• OR
• The annotation contains BIOSYNTHESIS
• THEN
• The gene is up-regulated

9
Why use Machine Learning?

• Concerned with machines learning from available
  data
• Informed by text data, the leaner can make
  first-pass model for the scientist

10
Desired Properties of a Model

• Accurate
• Measure with cross validation
• Comprehensible
• Measure with model size
• Stable to Small Changes in the Data
• Measure with random subsampling

11
Approaches

• Naïve Bayes
• Statistical method
• Uses all of the words (present or absent)
• PFOIL
• Covering algorithm
• Chooses words to use one at a time

12
Naïve Bayes

• For each word wi, there are two likelihood ratios
  (lr)
• lr (wi present) p(wi present up) / p(wi
  present down)
• lr (wi absent) p(wi absent up) / p(wi
  absent down)
• For each annotation, the lrs are combined to form
  a lr for a gene
• where X is either present or absent.

13
PFOIL

• Learn rules from data
• Produces multiple if-then rules from data
• Builds rules by adding one word at a time
• Easy to interpret models

14
Accuracy/Comprehensibility Tradeoff
15
Stabilized PFOIL

• Repeatedly run PFOIL on randomly sampled subsets
• For each word, count the number of models it
  appears in
• Restrict PFOIL to only those words that appear in
  a minimum of m models
• Rerun PFOIL with only those words

16
Stability Measure

• After running the algorithm N times to generate
  N rule sets
• Where
• U the set of words appearing in any rule set
• count(wi) number of rule sets containing word
  wi

17
Accuracy/Stability Tradeoff
18
Discussion

• Not very severe tradeoffs in Accuracy
• vs. stability
• vs. comprehensibility
• PFOIL not as good at characterizing data
• suggests not many dependencies
• need for softer rules

19
Future Directions

• M of N rules
• Permutation Test
• More Sources of Text Data

20
Take-Home Message

• This is just a first step toward an aid for
  understanding expression data
• Make expression models based on text in stead of
  DNA sequence.

21
Acknowledgements

• This research was funded by the following grants
• NLM 1 R01 LM07050-01,
• NSF IRI-9502990,
• NIH 2 P30 CA14520-29, and
• NIH 5 T32 GM08349.