1. Claudio Lottaz and Rainer Spang
Decomposing Complex Clinical Phenotypes by Biologically Structured Microarray Analysis
Claudio Lottaz and Rainer Spang
Berlin Center for Genome Based Bioinformatics, Berlin (Germany)
Computational Diagnostics, Max Planck Institute for Molecular Genetics, Berlin (Germany)

2. Overview Introduction
Using functional annotation for semi-supervised classification
Heterogeneity vs. performance
Evaluation on cancer related data
Concluasions
22-Oct-19

3. Tumor Classification Setting: More formally:
Introduction
Tumor Classification
Patients
Genes
D C
Setting:
Data: gene expression profiles
Goal: prediction/classification of outcome/sub-type
More formally:
Many expression levels measured
Samples labelled as disease and control
Train classifier
22-Oct-19

4. State-of-the-Art Various powerful methods:
Introduction
State-of-the-Art
Various powerful methods:
Support vector machines
Shrunken centroids...
Regularization to fight overfitting:
Feature selection
Large margins...
Common hypothesis: Generate a single molecular signature
22-Oct-19

5. Introduction
Complex Phenotypes
A single clinical phenotype may be caused by different molecular mechanisms
Our approach: discover several sub-classes in disease group
Each sub-class has a homogeneous molecular signature
22-Oct-19

6. Molecular Symptoms Classical signatures are globally optimal
Introduction
Molecular Symptoms
Classical signatures are globally optimal
They have no biological focus
Genes are corregulated thus correlated  in a global signature genes can be replaced with little loss
Molecular Symptom:
A functionally focused signature to identify a disease sub-class
High specificity – sub-optimal sensitivity
22-Oct-19

7. Molecular Patient Stratification
Introduction
Molecular Patient Stratification
Patterns of molecular symptoms define a molecular patient stratification
Control
Subclass
Control
Subclass
Molecular Symptom
Control
Another Molecular Symptom
Diagnostic signature
22-Oct-19

8. Using Functionl Annotations: A Priori vs. A Posteriori
Using Functional Annotations
Using Functionl Annotations: A Priori vs. A Posteriori
Common procedure
Data
Functional
Annotations
Statistical
Analysis
Data
Functional
Annotations
Statistical
Analysis
Our suggestion
22-Oct-19

9. Gene Ontology Biological terms in a directed graph
Using Functional Annotations
Gene Ontology
Biological terms in a directed graph
Genes annotated to terms
Levels represent specificity of terms
22-Oct-19

10. Structured Analysis of Microarrays
Using Functional Annotations
Structured Analysis of Microarrays
Classification in leaf nodes
Regularized multivariate classifier
Local signatures
Diagnosis propagation
Combine child diagnoses in inner nodes
Generate more general diagnoses
Regularization
Shrink the classifier graph
Remove uninformative branches
22-Oct-19

11. Leaf Node Classification
Using Functional Annotations
Leaf Node Classification
Shrunken centroid classification (Tibshirani et al. 2002)
Classificatino according to distance to centroids
Regularization via gene shrinkage
Determine probability-like values as classification results
22-Oct-19

12. Propagation of Classification
Using Functional Annotations
Propagation of Classification
Weighted averages
Weight according to child performance
Weights are normalized per inner node Pa w1 w3 w2 C1 C2 C3
22-Oct-19

13. Graph Shrinkage Weights of nodes are shrunken by a constant
Using Functional Annotations
Graph Shrinkage
Weights of nodes are shrunken by a constant
Negative weights are set to zero  uninformative branches vanish
Best shrinkage level chosen in cross-validation
22-Oct-19

14. Biased Classifier Evaluation
Heterogeneity vs. Performance
Biased Classifier Evaluation
Calibration of Sensitivity and Specificity
Shrinkage Parameter
Worst Performance in Leaf Node
Cj = DCi ( j Dj )-1
22-Oct-19

15. Classifier Heterogeneity
Heterogeneity vs. Performance
Classifier Heterogeneity
Difference between two classifiers: measures inconsistency of classifications
Node‘s redundancy:
Graph‘s redundancy (K nodes of the shrunken graph)
22-Oct-19

16. Calibration Sensitivity vs. Specificity:
Heterogeneity vs. Performance
Calibration
Sensitivity vs. Specificity:
Best classifiers: set to control prevalence
More molecular symptoms: set  higher than control prevalence
Heterogeneity vs. Performance: 
Molecular symptoms are heterogeneous
Thus high  eliminates them
22-Oct-19

17. Leukemia Data Set Data set by Yeoh et al. 2002 Task for illustration
Evaluation on Cancer Related Data
Leukemia Data Set
Data set by Yeoh et al. 2002
Acute lymphocytic leukemia
327 patients of 7 clinical sub-types
Expression profiles by HG-U95Av2
Task for illustration
Detect MLL sub-type
20 MLL samples
109 test set / 218 training set
22-Oct-19

18. Functional Annotations
Evaluation on Cancer Related Data
Functional Annotations
Focus on GO‘s Biological Process branch (8‘173 terms)
12‘625 probesets on the chip
8‘679 genes (68.7% of probesets)
In 1‘359 leaf nodes
845 inner nodes (total 2‘204 nodes)
22-Oct-19

19. MLL Classifier 2‘796 genes accessible through 32 nodes
Evaluation on Cancer Related Data
MLL Classifier
2‘796 genes accessible through 32 nodes
22-Oct-19

20. Evaluation on Cancer Related Data
MLL Stratification
22-Oct-19

21. Conclusions Semi-supervised classification Functional annotation
Conclustions
Conclusions
Semi-supervised classification
Datect sub-classes
In labelled disease groups
Functional annotation
Use in an a priori fashion
To find biologically focused signatures  molecular symptoms
Resolve complex clinical phenotypes (stratification through molecular symptoms)
22-Oct-19