Induction of Comprehensible Models for Gene Expression Data Sets

1
Induction of Comprehensible Models for Gene
Expression Data Sets

• Dragan Gamberger, RBI Zagreb
• Nada Lavrac, IJS Ljubljana
• Filip elezný, CVUT Prague
• Jakub Tolar, UMN Minneapolis

2
Modeling Gene Expression Data

• Predictive classification task
• Input Gene Expression vector
• Output Disease Class
• To train a predictor
• Use examples of existing GE vectors
• Associated to known disease class
• Character
• Lots of Attributes (eg. 20,000 GE values)
• Few Examples (eg. 20 patients)

3
Modeling Gene Expression Data

• Domain prone to overfitting
• Due to abundance of possible patterns, many seem
  good by chance
• Poor prediction on unseen examples
• Mainstream solution
• Robustness large redundant numeric classifiers
• Usually 10s 1000s genes employed in
  classification
• Voting of informative genes
• Support vector machines

4
Modeling Gene Expression Data

• Problem
• Complex / numeric classifiers not appropriate for
  expert evaluation
• Difficult interpretation
• Single genes (disease markers) with high voting
  power can be extracted from the predictors
• But then
• prediction assessment results no longer valid
• logical connections are lost
• (such as G1 expressed AND G2 not expressed)

5
Modeling Gene Expression Data

• Challenge Can we induce predictors that are
• Logic rules (? easy to read)
• Simple (few employed attributes)
• Accurate (on test examples)
• Meaningful (for a biologist)
• ? Induction of Comprehensible Models , Jr.
  Biomed. Informatics (Elsevier)To appear in 2004

?
6
The Methodology
Gene Expression Data
Discretizereal expression values to Absent /
Marginal / Present
Search for Relevant Features
Search for Relevant Logic Rules
Assess Predictive Accuracyon test data
Assess Meaningfulnessby expert interpretation
7
Discretization

• Converting real expression values tothree
  values
• A (absent not expressed)
• M (marginal)
• P (present - expressed)
• Using Affymetrix discretization
• May not be ideal, but ready for improvement

8
Feature Construction

• Simple form
• g A
• g P
• Marginal values cannot build a feature
• Relevancy
• Absolute to small support on the target class or
  to large support on the non-target classes
• Relative features with a better coverage exists

9
Rule Construction

• Subgroup Discovery Gamberger, Lavrac JAIR
  17/2002
• Forms features into conjunctive rules (defining
  subgroups of the target class)
• Such as (g1002 P AND g211 A)
• Balances precision and support
• Precision / Support trade-off is the search
  heuristic
• May induce impure rules

10
Experimental Domains

• AML / ALL Distinguish between
• Acute Myeloid Leukemia
• Acute Lymphoblastic Leukemia
• 38 training samples, 34 testing samples, 7129
  attributes (genes measured)
• MultiClass Distinguish between
• 14 types of cancer
• 144 training samples, 54 testing samples,16063
  attributes

11
Impact of Relevancy Filter

• AML/ALL 7129 original attributes (genes)
• Multi-Class 16063 original attributes on
  average

2844 Absolutely Irrelevant
3633 Relatively Irrelevant
622Rele vant
72
Absolutely or Relatively Irrelevant
28 Relevant
12
Predictive Performance Assessment

• AML / ALL Classification

13
Predictive Performance Assessment

• Multi-Class
• Classification
• Quick Growth
• of Precision
• with of
• examples

14
Comparison to Previous Results

• 9 Chow et al., Physiol Genomics 200119
  Golub et al, Science 199945 Ramaswamy, Procs.
  NAS, 2001

15
Rule Examples (AML/ALL)

• Co-activity of Leptin and GST described in
  previous biological study Balasubramaniyan,
  Pharm. Research, 2003
• ? Plausible relevance to AML ?

16
Rule Examples (Multi-Class)

• Routinely used lymphoma marker
• Plausible co-factor

17
Conclusions

• It is feasible to induce simple, logic-based
  classifiers for some GE modeling problems
• Given with very few positive examples, we could
  not prevent overfitting
• Given reasonably few examples, we found well
  generalizing, plausible rules
• Optimism larger data sets can be expected in the
  near future