1. Clustering and Classification – Introduction to Machine Learning BMI 730 Kun Huang Department of Biomedical Informatics Ohio State University

2. How do we use microarray? Profiling Clustering Cluster to detect patient subgroups Cluster to detect gene clusters and regulatory networks

3. Clustering and Classification Preprocessing Distance measures Popular algorithms (not necessarily the best ones) More sophisticated ones Evaluation Data mining

5. - Clustering or classification? - Is training data available? - What domain specific knowledge can be applied? - What preprocessing of data is needed? - Log / data scale and numerical stability - Filtering / denoising - Nonlinear kernel - Feature selection (do I need to use all the data?) - Is the dimensionality of the data too high?

6. How do we process microarray data (clustering)? - Feature selection – genes, transformations of expression levels. - Genes discovered in the class comparison (t-test). Risk: missing genes. - Iterative approach : select genes under different p- value cutoff, then select the one with good performance using cross-validation. - Principal components (pro and con). - Discriminant analysis (e.g., LDA).

7. - Dimensionality Reduction - Principal component analysis (PCA) - Singular value decomposition (SVD) - Karhunen-Loeve transform (KLT) Basis for P SVD

8. - Principal Component Analysis (PCA) - Other things to consider - Numerical balance/data normalization - Noisy direction - Continuous vs. discrete data - Principal components are orthogonal to each other, however, biological data are not - Principal components are linear combinations of original data - Prior knowledge is important - PCA is not clustering!

9. - Dimensionality reduction: linear discriminant analysis (LDA) B. 2.0 1.5 1.0 0.5 0.5 1.0 1.5 2.0............. A w. (From S. Wu’s website)

10. Linear Discriminant Analysis B. 2.0 1.5 1.0 0.5 0.5 1.0 1.5 2.0............. A w. (From S. Wu’s website)

11. Visualization of Microarray Data Multidimensional scaling (MDS) High-dimensional coordinates unknown Distances between the points are known The distance may not be Euclidean, but the embedding maintains the distance in a Euclidean space Try different dimensions (from one to ???) At each dimension, perform optimal embedding to minimize embedding error Plot embedding error (residue) vs. dimension Pick the knee point

12. Visualization of Microarray Data Multidimensional scaling (MDS)

13. Clustering and Classification Preprocessing Distance measures Popular algorithms (not necessarily the best ones) More sophisticated ones Evaluation Data mining

14. Distance Measure (Metric?) -What do you mean by “similar”? -Euclidean -Uncentered correlation -Pearson correlation

15. Distance Metric -Euclidean 102123_atLip11596.0002040.9001277.0004090.5001357.6001039.2001387.300 3189.0001321.3002164.400868.600185.300266.4002527.800 160552_atAp1s14144.4003986.9003083.1006105.9003245.8004468.4007295.000 5410.9003162.1004100.9004603.2006066.2005505.8005702.700 d E (Lip1, Ap1s1) = 12883

16. Distance Metric -Pearson Correlation 102123_atLip11596.0002040.9001277.0004090.5001357.6001039.2001387.300 3189.0001321.3002164.400868.600185.300266.4002527.800 160552_atAp1s14144.4003986.9003083.1006105.9003245.8004468.4007295.000 5410.9003162.1004100.9004603.2006066.2005505.8005702.700 d P (Lip1, Ap1s1) = 0.904

17. Distance Metric -Pearson Correlation r = 1r = -1 Ranges from 1 to -1.

18. Distance Metric -Uncentered Correlation 102123_atLip11596.0002040.9001277.0004090.5001357.6001039.2001387.300 3189.0001321.3002164.400868.600185.300266.4002527.800 160552_atAp1s14144.4003986.9003083.1006105.9003245.8004468.4007295.000 5410.9003162.1004100.9004603.2006066.2005505.8005702.700 d u (Lip1, Ap1s1) = 0.835  About 33.4 o

19. Distance Metric -Difference between Pearson correlation and uncentered correlation 102123_atLip11596.0002040.9001277.0004090.5001357.6001039.2001387.300 3189.0001321.3002164.400868.600185.300266.4002527.800 160552_atAp1s14144.4003986.9003083.1006105.9003245.8004468.4007295.000 5410.9003162.1004100.9004603.2006066.2005505.8005702.700 Pearson correlation Baseline expression possible Uncentered correlation All are considered signals

20. Distance Metric -Difference between Euclidean and correlation

21. Distance Metric -Missing: negative correlation may also mean “close” in signal pathway (1-|PCC|, 1-PCC^2)

22. Clustering and Classification Preprocessing Distance measures Popular algorithms (not necessarily the best ones) More sophisticated ones Evaluation Data mining

23. How do we process microarray data (clustering)? - Unsupervised Learning – Hierarchical Clustering

24. How do we process microarray data (clustering)? -Unsupervised Learning – Hierarchical Clustering Single linkage: The linking distance is the minimum distance between two clusters.

25. How do we process microarray data (clustering)? -Unsupervised Learning – Hierarchical Clustering Complete linkage: The linking distance is the maximum distance between two clusters.

26. How do we process microarray data (clustering)? -Unsupervised Learning – Hierarchical Clustering Average linkage/UPGMA: The linking distance is the average of all pair-wise distances between members of the two clusters. Since all genes and samples carry equal weight, the linkage is an Unweighted Pair Group Method with Arithmetic Means (UPGMA).

27. How do we process microarray data (clustering)? -Unsupervised Learning – Hierarchical Clustering Single linkage – Prone to chaining and sensitive to noise Complete linkage – Tends to produce compact clusters Average linkage – Sensitive to distance metric

28. -Unsupervised Learning – Hierarchical Clustering

29. Dendrograms Distance – the height each horizontal line represents the distance between the two groups it merges. Order – Opensource R uses the convention that the tighter clusters are on the left. Others proposed to use expression values, loci on chromosomes, and other ranking criteria.

30. -Unsupervised Learning - K-means -Vector quantization -K-D trees -Need to try different K, sensitive to initialization

31. -Unsupervised Learning - K-means [cidx, ctrs] = kmeans(yeastvalueshighexp, 4, 'dist', 'corr', 'rep',20); K Metric

32. -Unsupervised Learning - K-means -Number of class K needs to be specified -Does not always converge -Sensitive to initialization

33. -Unsupervised Learning - K-means

34. -Unsupervised Learning -Self-organized maps (SOM) -Neural network based method -Originally used as a visualization method for visualize (embedding) high-dimensional data -Also related vector quantization -The idea is to map close data points to the same discrete level

35. -Issues -Lack of consistency or representative features (5.3 TP53 + 0.8 PTEN doesn’t make sense) -Data structure is missing -Not robust to outliers and noise D’Haeseleer 2005 Nat. Biotechnol 23(12):1499-501

36. Review of Microarray and Gene Discovery Clustering and Classification Preprocessing Distance measures Popular algorithms (not necessarily the best ones) More sophisticated ones Evaluation Data mining

37. -Model-based clustering methods (Han) http://www.cs.umd.edu/~bhhan/research2.html Pan et al. Genome Biology 2002 3:research0009.1 doi:10.1186/gb-2002-3-2-research0009

38. -Structure-based clustering methods

39. -Supervised Learning -Support vector machines (SVM) and Kernels -Only (binary) classifier, no data model

40. -Supervised Learning - Support vector machines (SVM) and Kernels -Kernel – nonlinear mapping

41. -Supervised Learning - Naïve Bayesian classifier -Bayes rule -Maximum a posterior (MAP) Prior prob. Conditional prob.

42. Review of Microarray and Gene Discovery Clustering and Classification Preprocessing Distance measures Popular algorithms (not necessarily the best ones) More sophisticated ones Evaluation Data mining

43. -Accuracy vs. generality -Overfitting -Model selection Model complexity Prediction error Training sample Testing sample (reproduced from Hastie et.al.)

44. -Assessing the Validity of Clusters -Most clustering algorithms do not assume any structure or a prior relationship among the genes. However, the found clusters should more or less reflect the structures (e.g., pathways). (An interesting research problem is to develop new algorithms that can accommodate such relationships.) -If different patients are grouped into clusters, it implies that there are subtypes for the disease, which is a big claim and must be validated using other methods (e.g., pathology). -Relationship with external variables is important. E.g., clustering on cells from different tissue types may correspond to the relationship among the tissues.

45. -Assessing the Validity of Clusters -Where should we cut the dendrograms? -Which clustering results should we believe, i.e., different (or even the same) clustering algorithms may find different clustering results? -Many tests are flawed, e.g., circular reasoning: using genes with significant different between two classes as features for clustering, then use the clusters to detect signatures which are genes significantly changed.

46. -Assessing the Validity of Clusters -Most clustering algorithms can find clusters even from random data. -The clusters found by clustering algorithms should exhibit greater intra- cluster similarity (homogeneity) and larger inter-cluster distance (separation). -How to be sure that the clustering is not from random data? -How to find good partition among any possible partitions of the data? -How to assess the reproducibility of the partitioning?

47. -Assessing the Validity of Clusters -Global tests of clustering (meaningful cluster vs. random cluster) -Check the distribution of the nearest neighbor distances (NN) and pairwise distances, uniform distribution and multiple distribution are very different NN Pairwise

48. -Assessing the Validity of Clusters -Reproducibility of clustering -Global perturbation methods (McShane et al, Bioinformatics, 2002, 1462-1469 -Using only the first three principal components (the observation is that they convey the clustering information well enough -Adding Gaussian noise and check if the clustering relationship is still preserved -Indices R and D. R – the ratio of same cluster data pairs that are preserved after the perturbation. D - discrepancy between best-matched clusters

49. How do we process microarray data (clustering)? - Cross-validation: assessment of the classifier. Note the key thing is to strike the balance between accurate classification on training data and the prediction power. - Training vs. testing (10%) - Leave-one-out bootstraping: for small sample size, ratio on the correct prediction of the left-out sample.

50. Validation cDNA or Affymetrix chips measure mRNA levels, which may not reflect final protein concentrations Various splice variants exist, the expressed protein may not be active Post-translational modification Quantitative real-time PCR (RT-PCR) is widely used for this purpose Other high-level consideration – correlation does not mean causation

51. Review of Microarray and Gene Discovery Clustering and Classification Preprocessing Distance measures Popular algorithms (not necessarily the best ones) More sophisticated ones Evaluation Data mining

52. – Data Mining is searching for knowledge in data –Knowledge mining from databases –Knowledge extraction –Data/pattern analysis –Data dredging –Knowledge Discovery in Databases (KDD)

53. −The process of discovery Interactive + Iterative  Scalable approaches

54. Popular Data Mining Techniques – Clustering: Most dominant technique in use for gene expression analysis in particular and bioinformatics in general. –Partition data into groups of similarity – Classification: –Supervised version of clustering  technique to model class membership  can subsequently classify unseen data. – Frequent Pattern Analysis – A method for identifying frequently re-curring patterns (structural and transactional). – Temporal/Sequence Analysis –Model temporal data  wavelets, FFT etc. – Statistical Methods –Regression, Discriminant analysis

55. Summary −A good clustering method will produce high quality clusters with −high intra-class similarity −low inter-class similarity −The quality of a clustering result depends on both the similarity measure used by the method and its implementation. −Other metrics include: density, information entropy, statistical variance, radius/diameter −The quality of a clustering method is also measured by its ability to discover some or all of the hidden patterns.

56. Recommended Literature 1. Bioinformatics – The Machine Learning Approach by P. Baldi & S. Brunak, 2 nd edition, The MIT Press, 2001 2. Data Mining – Concepts and Techniques by J. Han & M. Kamber, Morgan Kaufmann Publishers, 2001 3. Pattern Classification by R. Duda, P. Hart and D. Stork, 2 nd edition, John Wiley & Sons, 2001 4. The Elements of Statistical Learning by T. Hastie, R. Tibshirani, J. Friedman, Springer-Verlag, 2001