1. Canadian Bioinformatics Workshops www.bioinformatics.ca

2. 2Module #: Title of Module

3. Module 5 Metabolomic Data Analysis Using MetaboAnalyst

4. Learning Objectives To become familiar with the standard metabolomics data analysis workflow To become aware of key elements such as: data integrity checking, outlier detection, quality control, normalization, scaling, etc. To learn how to use MetaboAnalyst to facilitate data analysis

5. A Typical Metabolomics Experiment

6. 2 Routes to Metabolomics 1234567ppm Quantitative (Targeted) Methods Chemometric (Profiling) Methods

7. Metabolomics Data Workflow Data Integrity Check Spectral alignment or binning Data normalization Data QC/outlier removal Data reduction & analysis Compound ID Data Integrity Check Compound ID and quantification Data normalization Data QC/outlier removal Data reduction & analysis Chemometric Methods Targeted Methods

8. Data Integrity/Quality LC-MS and GC-MS have high number of false positive peaks Problems with adducts (LC), extra derivatization products (GC), isotopes, breakdown products (ionization issues), etc. Not usually a problem with NMR Check using replicates and adduct calculators MZedDB http://maltese.dbs.aber.ac.uk:8888/hrmet/index.html HMDB http://www.hmdb.ca/search/spectra?type=ms_search

9. Data/Spectral Alignment Important for LC-MS and GC-MS studies Not so important for NMR (pH variation) Many programs available (XCMS, ChromA, Mzmine) Most based on time warping algorithms http://mzmine.sourceforge.net/ http://bibiserv.techfak.uni-bielefeld.de/chroma http://metlin.scripps.edu/xcms/

10. bin1 bin2 bin3 bin4 bin5 bin6 bin7 bin8... x i,y i x = 232.1 (AOC) y = 10 (bin #) Binning (3000 pts to 14 bins)

11. Data Normalization/Scaling Can scale to sample or scale to feature Scaling to whole sample controls for dilution Normalize to integrated area, probabilistic quotient method, internal standard, sample specific (weight or volume of sample) Choice depends on sample & circumstances Same or different?

12. Data Normalization/Scaling Can scale to sample or scale to feature Scaling to feature(s) helps manage outliers Several feature scaling options available: log transformation, auto- scaling, Pareto scaling, probabilistic quotient, and range scaling MetaboAnalyst http://www.metaboanalyst.ca Dieterle F et al. Anal Chem. 2006 Jul 1;78(13):4281-90.

13. Data QC, Outlier Removal & Data Reduction Data filtering (remove solvent peaks, noise filtering, false positives, outlier removal -- needs justification) Dimensional reduction or feature selection to reduce number of features or factors to consider (PCA or PLS-DA) Clustering to find similarity

14. MetaboAnalyst http://www.metaboanalyst.ca A comprehensive web server designed to process & analyze LC-MS, GC-MS or NMR-based metabolomic data

15. MetaboAnalyst History 2009 v1.0 - Supports both univariate and multivariate data processing, including t- tests, ANOVA, PCA, PLS-DA, colorful plots, with detailed explanations & summaries 2012 v2.0 - Identifies significantly altered functions & pathways 2015 v3.0 – Better performance, better graphical interactivity, biomarker analysis, power analysis, integration with gene expression data …

16. MetaboAnalyst Overview Raw data processing Data reduction & statistical analysis Functional enrichment analysis Metabolic pathway analysis Power analysis and sample size estimation Biomarker analysis Integrative analysis

17. MetaboAnalyst Modules 17 Data pre- processing Data normalization Data analysis Data interpretation

18. MetaboAnalyst Modules

19. Example Datasets

21. Metabolomic Data Processing

22. Common Tasks Purpose: to convert various raw data forms into data matrices suitable for statistical analysis Supported data formats –Concentration tables (Targeted Analysis) –Peak lists (Untargeted) –Spectral bins (Untargeted) –Raw spectra (Untargeted)

23. Select a Module (Statistical Analysis)

24. Data Upload

25. Alternatively …

26. Data Set Selected Here we have selected a data set from dairy cattle fed different proportions of cereal grains (0%, 15%, 30%, 45%) The rumen was analyzed using NMR spectroscopy using quantitative metabolomic techniques High grain diets are thought to be stressful on cows

27. Data Integrity Check

28. Data Normalization Samples = rows Compounds = columns

29. Data Normalization At this point, the data has been transformed to a matrix with the samples in rows and the variables (compounds/peaks/bins) in columns MetaboAnalyst offers three types of normalization, row-wise normalization, column-wise normalization and combined normalization Row-wise normalization aims to make each sample (row) comparable to each other (i.e. urine samples with different dilution effects)

30. Data Normalization Column-wise normalization aims to make each variable (column) comparable in scale to each other, thereby generating a “normal” distribution This procedure is useful when variables are of very different orders of magnitude Four methods have been implemented for this purpose – log transformation, autoscaling, Pareto scaling and range scaling

31. Normalization Result

32. Data Normalization You cannot know a priori what the best normalization protocol will be MetaboAnalyst allows you to interactively explore different normalization protocols and to visually inspect the degree of “normality” or Gaussian behavior This example is nicely normalized

33. Next Steps After normalization has been completed it is a good idea to look at your data a little further to identify outliers or noise that could/should be removed

34. Quality Control Dealing with outliers –Detected mainly by visual inspection –May be corrected by normalization –May be excluded Noise reduction –More of a concern for spectral bins/ peak lists –Usually improves downstream results

35. Visual Inspection What does an outlier look like? Finding outliers via PCAFinding outliers via Heatmap

36. Outlier Removal (Data Editor)

37. Noise Reduction (Data Filtering)

38. Noise Reduction (cont.) Characteristics of noise & uninformative features –Low intensities –Low variances (default)

39. Data Reduction and Statistical Analysis

40. Common Tasks To identify important features To detect interesting patterns To assess difference between the phenotypes To facilitate classification or prediction We will look at ANOVA, Multivariate Analysis (PCA, PLS-DA) and Clustering

42. ANOVA Looking at 4 different dairy cow populations –0% grain in diet –15% grain in diet –30% grain in diet –45% grain in diet Try to identify those metabolites that are different between all groups or just between 0% and everything else

43. ANOVA Click this spot and the 3-PP graph pops up Click this to view the table

44. View Individual Compounds Click this to see the uracil graphs

45. What’s Next? Click and compare different compounds to see which ones are most different or most similar between the 4 groups Click on the Correlation link (under the ANOVA link) to generate a heat map that displays the pairwise compound correlations and compound clusters

46. Overall Correlation Pattern Click this to save a high res. image

47. High Resolution Image

48. What’s Next? When looking at >2 groups it is often useful to look for patterns or trends within particular metabolites Use Pattern Hunter to find these trends

49. Pattern Matching Looking for compounds showing interesting patterns of change Essentially a method to look for linear trends or periodic trends in the data Best for data that has 3 or more groups

50. Pattern Matching (cont.) Strong linear + correlation to grain % Strong linear - correlation to grain %

52. Multivariate Analysis Use PCA option to view the separation (if any) in the 4 groups Look at the 2D PCA Score Plot –2 most significant principal components Look at the 2D PCA Loading Plot Look at the PCA Plot in 3D –3 most significant principal components Options for viewing are located in the top tabs

53. PCA Scores Plot

54. PCA Loading Plot Compounds most responsible for separation Click on a point to view

55. 3D Score Plot 55 Drag to rotate Mouse over to see sample names

57. Multivariate Analysis Use PLS-DA option to view the separation of the 4 (labeled) groups PLS-DA “rotates” the PCA axes to maximize separation Look at the 2D PLS Scores Plot Look at the Q 2 and R 2 Values (Cross Validation) Use the VIP plot to ID important metabolites (VIP > 1.2)

58. PLS-DA Score Plot

59. Evaluation of PLS-DA Model PLS-DA Model evaluated by cross validation of Q 2 and R 2 Using too many components can over-fit 3 component model seems to be a good compromise here Good R 2 /Q 2 (>0.7)

60. Important Compounds

61. Model Validation Note, permutation is computationally intensive. It is not performed by default. Users need to set the permutation number and press the submit button

63. Hierarchical Clustering (Heat Maps) An alternative way of viewing or clustering multivariate data Allows one to look at the behavior of individual metabolites Can ask questions such as: which compounds have a low concentration in group 0, 15 but increase in the group 35 and 45? or which compound is the only one significantly increased in group 45?

64. Heatmap Visualization Note that the Heatmap is not being clustered on Rows. It is ordered by the class labels

65. Heatmap Visualization (cont.)

66. What’s Next? Most of the multivariate analysis is now done MetaboAnalyst has been keeping track of the plots or graphs you have generated Now its time to generate a printed report that summarizes what you’ve done and what you’ve found

67. Download Results

68. Analysis Report

69. Select a Module (Enrichment Analysis)

70. Metabolite Set Enrichment Analysis (MSEA) http://www.msea.cahttp://www.msea.ca Now part of Metaboanalyst Designed to handle lists of metabolites (with or without concentration data) Modeled after Gene Set Enrichment Analysis (GSEA) Supports over representation analysis (ORA), single sample profiling (SSP) and quantitative enrichment analysis (QEA) Contains a library of 6300 pre-defined metabolite sets including 85 pathway sets & 850 disease sets

71. Enrichment Analysis Purpose: To test if there are biologically meaningful groups of metabolites that are significantly enriched in your data Biological meaningful in terms of: –Pathways –Disease –Localization Currently, MSEA only supports human metabolomic data

72. MSEA Accepts 3 kinds of input files –list of metabolite names only (ORA – over representation analysis) –list of metabolite names + concentration data from a single sample (SSP – single sample profiling) –a concentration table with a list of metabolite names + concentrations for multiple samples/patients (QEA – quantitative enrichment analysis)

73. The MSEA Approach 73 Assess metabolite sets directly Compound concentrations Compound selection (t-tests, clustering) Abnormal compounds Compound concentrations Metabolite set libraries Over Representation Analysis Quantitative Enrichment Analysis Biological interpretation Find enriched biological themes Compound concentrations Important compound lists Single Sample Profiling Compare to normal references ORA input For MSEA ORA SSP QEA

74. Data Set Selected Here we are using a collection of metabolites identified by NMR (compound list + concentrations) from the urine from 77 lung and colon cancer patients, some of whom were suffering from cachexia (muscle wasting)

75. Start with a Compound List for ORA

76. Upload Compound List Normally GSEA would require a list of all known genes for the given platform. Here we just use the list of metabolites found in KEGG. ORA is a “weak” analysis in MSEA

77. Perform Compound Name Standardization

78. Name Standardization (cont.)

79. Select a Metabolite Set Library

80. Result

81. Result (cont.) Click on details to see more

82. The Matched Metabolite Set Click on SMPDB to see more information

83. Phenylalanine and Tyrosine Metabolism in SMPDB

84. Single Sample Profiling (SSP) (Basically used by a physician to analyze a patient)

85. Concentration Comparison

86. Concentration Comparison (cont.)

87. Quantitative Enrichment Analysis (QEA)

88. Result Click on details to see more

89. The Matched Metabolite Set

90. Select a Module (Pathway Analysis)

91. Pathway Analysis Purpose: to extend and enhance metabolite set enrichment analysis for pathways by –Considering pathway structures –Supporting pathway visualization Currently supports analysis for 21 diverse (model) organisms such as humans, mouse, drosophila, arabadopsis, E. coli, yeast, etc. (KEGG pathways only)

92. Data Set Selected Here we are using a collection of metabolites identified by NMR (compound list + concentrations) from the urine from 77 lung and colon cancer patients, some of whom were suffering from cachexia (muscle wasting)

93. Pathway Analysis Module

94. Data Upload

95. Perform Data Normalization

96. Select Pathway Libraries

97. Perform Network Topology Analysis

98. Pathway Position Matters 98 Junker et al. BMC Bioinformatics 2006  Which positions are important?  Hubs  Nodes that are highly connected (red ones)  Bottlenecks  Nodes on many shortest paths between other nodes (blue ones)  Graph theory  Degree centrality  Betweenness centrality

99. Which Node is More Important? High betweenness centrality High degree centrality

100. Pathway Visualization

101. Pathway Visualization (cont.)

102. Pathway Impact Incorporates parameters such as the log fold-change of the DE metabolites, the statistical significance of the set of pathway genes and the topology of the signaling pathway Combines the pathway topology with the over-representation evidence

103. Result

104. Select a Module (Biomarker Analysis)

105. Biomarker Analysis Purpose is to find biomarkers using ROC (receiver operator characteristic) curves with high sensitivity and specificity Maximize AUC under ROC curve while minimizing the number of metabolites used in the biomarker panel 3 different modules (univariate – single marker at a time, multivariate – many combinations of biomarkers, manual – user choice)

106. Select Test Data Set 1 Click Here

107. Data Set Selected 90 patients (expectant mothers) at 3 months pregnancy Serum samples 45 patients went on to develop pre- eclampsia at 6-7 months 45 patients had normal pregancies Trying to find biomarkers for predicting early pre-eclampsia

108. Perform Data Integrity Check Click Here

109. Perform Log Normalization Click Here

110. Check That It’s Normally Distributed Click Here beforeafter

111. Select Multivariate Option Click Here

112. View ROC Curve

113. Choose a Model (95% conf.) Select model

114. 95% Confidence Interval

115. Select Sig. Features Tab Click Here

116. View VIP Plot

117. Select a Module (Power Analysis)

118. Statistical Power Statistical power is the ability of a test to detect an effect, if the effect actually exists –A power of 0.8 in a clinical trial means that the study has a 80% chance of ending up with a statistically significant treatment effect if there really was an important difference between treatments. To answer research questions: –How powerful is my study? –How many samples do I need to have for what I want to get from the study?

119. Statistical Power (cont.) The statistical power of a test depends: 1.Sample size, 2.Significance criterion (alpha) 3.Effect size Increase power Effect size Sample size Decrease Power Significance criterion

120. The Approach How do we get these values? –Effect size can be estimated from a pilot data; –Significance criteria Single metabolite - p value cutoff (i.e. 0.05, 0.01) Metabolomics data – FDR (i.e. 0.1) –Sample size is our interest –Power value is our interest You need to upload a pilot data, and set the criteria, MetaboAnalyst will compute a power vs. sample size curve by computing power values for a range of sample sizes [3, 1000]

121. Power vs. Sample size At least 60 samples/group will needed to get a power of 0.8

122. Not Everything Was Covered Clustering (K-means, SOM) Classification (SVM, randomForests) Time-series data analysis Two factor data analysis Integrative pathway analysis (gene and metabolite)

123. Time Series Analysis in MetaboAnalyst 123

124. Integrative Pathway Analysis