1. A Comprehensive Evaluation of Multicategory Classification Methods for Microarray Gene Expression Cancer Diagnosis Presented by: Renikko Alleyne

2. Outline Motivation Major Concerns Methods –SVMs –Non-SVMs –Ensemble Classification Datasets Experimental Design Gene Selection Performance Metrics Overall Design Results Discussion & Limitations Contributions Conclusions

3. Why? Clinical Applications of Gene Expression Microarray Technology Gene DiscoveryDisease Diagnosis CancerInfectious Diseases Drug Discovery Prediction of clinical outcomes in response to treatment

4. GEMS (Gene Expression Model Selector) Creation of powerful and reliable cancer diagnostic models Equip with best classifier, gene selection, and cross-validation methods Evaluation of major algorithms for multicategory classification, gene selection methods, ensemble classifier methods & 2 cross validation designs 11 datasets spanning 74 diagnostic categories & 41 cancer types & 12 normal tissue types Microarray data

5. Major Concerns The studies conducted limited experiments in terms of the number of classifiers, gene selection algorithms, number of datasets and types of cancer involved. Cannot determine which classifier performs best. It is poorly understood what are the best combinations of classification and gene selection algorithms across most array-based cancer datasets. Overfitting. Underfitting.

6. Goals for the Development of an Automated System that creates high-quality diagnostic models for use in clinical applications Investigate which classifier currently available for gene expression diagnosis performs the best across many cancer types How classifiers interact with existing gene selection methods in datasets with varying sample size, number of genes and cancer types Whether it is possible to increase diagnostic performance further using meta-learning in the form of ensemble classification How to parameterize the classifiers and gene selection procedures to avoid overfitting

7. Why use Support Vector Machines (SVMs)? Achieve superior classification performance compared to other learning algorithms Fairly insensitive to the curse of dimensionality Efficient enough to handle very large-scale classification in both sample and variables

8. How SVMs Work Objects in the input space are mapped using a set of mathematical functions (kernels). The mapped objects in the feature (transformed) space are linearly separable, and instead of drawing a complex curve, an optimal line (maximum-margin hyperplane) can be found to separate the two classes.

9. SVM Classification Methods SVMs Binary SVMs Multiclass SVMs OVROVODAGSVMWWSW

10. Binary SVMs Main idea is to identify the maximum-margin hyperplane that separates training instances. Selects a hyperplane that maximizes the width of the gap between the two classes. The hyperplane is specified by support vectors. New classes are classified depending on the side of the hyperplane they belong to. Support Vector Hyperplane

11. 1. Multiclass SVMs: one-versus-rest (OVR) Simplest MC-SVM Construct k binary SVM classifiers: –Each class (positive) vs all other classes (negatives). Computationally Expensive because there are k quadratic programming (QP) optimization problems of size n to solve.

12. 2. Multiclass SVMs: one-versus-one (OVO) Involves construction of binary SVM classifiers for all pairs of classes A decision function assigns an instance to a class that has the largest number of votes (Max Wins strategy) Computationally less expensive

13. 3. Multiclass SVMs: DAGSVM Constructs a decision tree Each node is a binary SVM for a pair of classes k leaves: k classification decisions Non-leaf (p, q): two edges –Left edge: not p decision –Right edge: not q decision

14. 4 & 5. Multiclass SVMs: Weston & Watkins (WW) and Crammer & Singer (CS) Constructs a single classifier by maximizing the margin between all the classes simultaneously Both require the solution of a single QP problem of size (k-1)n, but the CS MC-SVM uses less slack variables in the constraints of the optimization problem, thereby making it computationally less expensive

15. Non-SVM Classification Methods Non- SVMs KNNNNPNN

16. K-Nearest Neighbors (KNN) For each case to be classified, locate the k closest members of the training dataset. A Euclidean Distance measure is used to calculate the distance between the training dataset members and the target case. The weighted sum of the variable of interest is found for the k nearest neighbors. Repeat this procedure for the other target set cases. ??

17. Backpropagation Neural Networks (NN) & Probabilistic Neural Networks (PNNs) Back Propagation Neural Networks: –Feed forward neural networks with signals propagated forward through the layers of units. –The unit connections have weights which are adjusted when there is an error, by the backpropagation learning algorithm. Probabilistic Neural Networks: –Design similar to NNs except that the hidden layer is made up of a competitive layer and a pattern layer and the unit connections do not have weights.

18. Ensemble Classification Methods In order to improve performance: Classifier 1 Ensembled Classifiers Techniques: Major Voting, Decision Trees, MC-SVM (OVR, OVO, DAGSVM) Classifier 2Classifier N Output 1Output NOutput 2

19. Datasets & Data Preparatory Steps Nine multicategory cancer diagnosis datasets Two binary cancer diagnosis datasets All datasets were produced by oligonucleotide-based technology The oligonucleotides or genes with absent calls in all samples were excluded from analysis to reduce any noise.

20. Datasets

21. Experimental Designs Two Experimental Designs to obtain reliable performance estimates and avoid overfitting. Data split into mutually exclusive sets. Outer Loop estimates performance by: –Training on all splits but one (use for testing). Inner Loop determines the best parameter of the classifier.

22. Experimental Designs Design I uses stratified 10 fold cross-validation in both loops while Design II uses 10 fold cross-validation in its inner loop and leave-one-out-cross-validation in its outer loop. Building the final diagnostic model involves: –Finding the best parameters for the classification using a single loop of cross-validation –Building the classifier on all data using the previously found best parameters –Estimating a conservative bound on the classifier’s accuracy by using either Designs

23. Gene Selection Gene Selection Methods Ratio of genes between-categories to within-category sum of squares (BW) Signal-to-noise scores (S2N) S2N-OVRS2N-OVO Kruskal-Wallis non- parametric one-way ANOVA (KW)

24. Performance Metrics Accuracy –Easy to interpret –Simplifies statistical testing –Sensitive to prior class probabilities –Does not describe the actual difficulty of the decision problem for unbalanced distributions Relative classifier information (RCI) –Corrects for the differences in: Prior probabilities of the diagnostic categories Number of categories

25. Overall Research Design Stage 1:Conducted a Factorial design involving datasets & classifiers w/o gene selection Stage 2: Conducted a Factorial Design w/ gene selection using datasets for which the full gene sets yielded poor performance 2.6 million diagnostic models generated Selection of one model for each combination of algorithm and dataset

26. Statistical Comparison among classifiers To test that differences b/t the best method and the other methods are non-random Null Hypothesis: Classification algorithm X is as good as Y Obtain permutation distribution of XY ∆ by repeatedly rearranging the outcomes of X and Y at random Compute the p-value of XY ∆ being greater than or equal to observed difference XY ∆ over 10000 permutations If p < 0.05  Reject H0 Algorithm X is not as good as Y in terms of classification accuracy If p > 0.05  Accept H0 Algorithm X is as good as Y in terms of classification accuracy

27. Performance Results (Accuracies) without Gene Selection Using Design I

28. Performance Results (RCI) without Gene Selection Using Design I

29. Total Time of Classification Experiments w/o gene selection for all 11 datasets and two experimental designs Executed in a Matlab R13 environment on 8 dual-CPU workstations connected in a cluster. Fastest MC-SVMs: WW & CS Fastest overall algorithm: KNN Slowest MC-SVM: OVR Slowest overall algorithms: NN and PNN

30. Performance Results (Accuracies) with Gene Selection Using Design I Applied the 4 gene selection methods to the 4 most challenging datasets Improvement by gene selection

31. Performance Results (RCI) with Gene Selection Using Design I Applied the 4 gene selection methods to the 4 most challenging datasets Improvement by gene selection

32. Discussion & Limitations Limitations: –Use of the two performance metrics –Choice of KNN, PNN and NN classifiers Future Research: –Improve existing gene selection procedures with the selection of optimal number of genes by cross-validation –Applying multivariate Markov blanket and local neighborhood algorithms –Extend comparisons with more MC-SVMs as they become available –Updating GEMS system to make it more user-friendly.

33. Contributions of Study Conducted the most comprehensive systematic evaluation to date of multicategory diagnosis algorithms applied to the majority of multicategory cancer-related gene expression human datasets. Creation of the GEMS system that automates the experimental procedures in the study in order to: –Develop optimal classification models for the domain of cancer diagnosis with microarray gene expression data. –Estimate their performance in future patients.

34. Conclusions MSVMs are the best family of algorithms for these types of data and medical tasks. They outperform non-SVM machine learning techniques Among MC-SVM methods OVR, CS and WW are the best w.r.t classification performance Gene selection can improve the performance of MC and non- SVM methods Ensemble classification does not further improve the classification performance of the best MC-SVM methods