CyTOF workflow: differential discovery in high-throughput high-dimensional cytometry datasets!

Transcription

1 Workshop!! CyTOF workflow: differential discovery in high-throughput high-dimensional cytometry datasets! Malgorzata Nowicka! University of Zurich! BioC 2017: Where Software and Biology Connect! Boston, 28 July 2017!

2 Have you analyzed CyTOF or flow cytometry data before?!

3 My approach to this workshop! This is just a shorter version of the original workflow! The text is basically the same but maybe try to not focus on it so much, as you can carefully read it later! Ideally, I would do the life coding but because of the time constrains, I will:! explain all the code in the vignette (html file)! copy-paste to R, to follow with you and have the opportunity to modify some things if needed!

4 Introduction! CyTOF (mass cytometry) experiment! Measuring ~40 parameters (metals)! Hundreds of cells per second Figure adapted from Bendall et al. Science 2011!

5 Introduction! High-dimensional cytometry because it is higher than before (from ~12 to ~40)! We refer to this differential approach as classic! Common strategy:! identify cell populations of interest by manual gating or automated clustering! determine which of the cell subpopulations or protein markers are associated with a phenotype of interest using statistical tests! Other approaches:! Citrus (Bruggner et al. 2014)! CellCnn (Arvaniti and Claassen 2016)! Hybrid approaches (thanks to the modularity)! This workflow can serve as a template!

6 Overview! 1. Data description! 2. Data pre-processing (CATALYST pckg)! 3. Data import flowcore pckg! 4. Data transformation - arcsinh with cofactor 5! 5. Spot checks! MDS plots (limma pckg)! 6. Cell clustering with FlowSOM and ClusterConsensusPlus pckgs! overclustering + manual merging! 7. Visualization of the clusters with! t-sne - Rtsne pckg! heatmaps - pheatmap pckg! ggplot pckg! 8. Differential analysis with linear mixed models (lme4 and multcomp pckgs)! diff. cell population abundance! diff. marker expression!

7 Data description! Subset of data from the Bodenmiller et al study; used mass cytometry to measure PBMC's! from 8 different patients! after 11 different stimulation conditions as well as an unstimulated "Reference" state! Here, we use the samples that correspond to the BCR-crosslinking and Reference! For each sample, 10 cell surface markers and 14 signaling markers were measured!

8 Data pre-processing! The pre-processing steps in the Bodenmiller data, as we can download it, included removal of debris and de-barcoding! In general, pre-processing steps may involve:! normalization using bead standards! de-barcoding! compensation! CATALYST pckg!

9 Data import! All the data is available at the Robinson Lab server cytofworkflow/! Downloading using the download.file()! Metadata! FCS files! Panel file! Files with cluster merging!

10 Data import: expr! S1 c1 c2 c3 c4 M1 M2 M3 S2 c5 c6 c7 S3 c8 c9 c10 c11 c12

11 Data transformation! Mass cytometry Linear Arcsinh cofactor 5 Skewed distributions! Hard to distinguish between positive and negative populations! Magic transformation - arcsinh (hyperbolic inverse sine) with cofactor 5! linear for the lower expression,! logarithmic for the higher expression,! works for the negative and zero values (they do not have to be excluded from the analysis)! Figure adapted from Bendall et al. Science 2011!

12 Diagnostic plots quick look on! Plot with per-sample marker expression distributions, colored by condition! Identify problematic samples or markers! Cell counts in samples! MDS plot (or PCA plot)! Standard usage in the RNA-seq analysis! Generated with the plotmds() function from the limma pckg! As we want to plot samples we need to summarize the information about cells to the sample level!

13 MDS plot: expr_median_sample_tbl! M1 M2 M3 c1 S1 c2 c3 * * * c4 M1 M2 M3 c5 S1 * * * S2 c6 c7 * * * S2 * * * c8 c9 S3 * * * S3 c10 * * * c11 c12 * - median!

14 MDS plot!

15 Cell population identification! Manual gating! Comparison of clustering algorithms by Lukas Weber et al., Cytometry Part A, 2016! FlowSOM (+ ClusterConsensusPlus) one of the best performing methods and super fast! 3 steps:! 1. Building of the self-organizing map (SOM) with the BuildSOM function - cells are assigned according to their similarities to 100 grid points (or, so-called codes) of the SOM! 2. Building of a minimal spanning tree, which is mainly used for graphical representation of the clusters! 3. Metaclustering of the SOM codes performed directly with the ConsensusClusterPlus function!

16 Over-clustering! Some level of over-clustering is necessary, in order to detect somewhat rare populations! In addition, merging can always follow an overclustering step, but splitting of existing clusters is generally not feasible!

17 Over-clustering! Over-clustering into 20 groups! Median marker expression of data normalized to 0-1! (3.85%) 5 (1.89%) 20 (0.34%) 15 (7.8%) 12 (24.25%) 18 (1.85%) 3 (14.19%) 4 (1.97%) 9 (16.62%) 2 (2.26%) 17 (0.64%) 19 (1.51%) 8 (8.82%) 14 (2.2%) 10 (2.08%) 6 (3.58%) 7 (4.37%) 11 (0.64%) 13 (0.42%) 16 (0.72%) CD14 CD45 CD33 CD123 IgM CD20 HLA_DR CD3 CD4 CD7 cluster Figures adapted from Nowicka et al. F1000Research 2017!

18 Merging by an expert! cluster cluster_merging Median marker expression of data normalized to 0-1! CD CD CD HLA_DR CD IgM CD CD CD CD14 1 (3.85%) 5 (1.89%) 20 (0.34%) 15 (7.8%) 12 (24.25%) 18 (1.85%) 3 (14.19%) 4 (1.97%) 9 (16.62%) 2 (2.26%) 17 (0.64%) 19 (1.51%) 8 (8.82%) 14 (2.2%) 10 (2.08%) 6 (3.58%) 7 (4.37%) 11 (0.64%) 13 (0.42%) 16 (0.72%) cluster_merging B cells IgM+ B cells IgM CD4 T cells CD8 T cells DC NK cells monocytes surface Figures adapted from Nowicka et al. F1000Research 2017!

19 Statistical testing differential abundance!

20 Statistical testing differential abundance! Ref BCRXL cluster B cells IgM+ B cells IgM CD4 T cells CD8 T cells DC NK cells monocytes surface 0 Ref1 Ref2 Ref3 Ref4 Ref5 Ref6 Ref7 Ref8 BCRXL1 BCRXL2 BCRXL3 BCRXL4 BCRXL5 BCRXL6 BCRXL7 BCRXL8 proportion sample_id

21 proportion Statistical testing differential abundance! Ref1 Ref2 Ref3 Ref Ref4 Ref5 proportion Ref6 Ref7 Ref B cells IgM+ BCRXL B cells IgM CD4 T cells CD8 T cells BCRXL1 BCRXL2 BCRXL3 B cells IgM+ B cells IgM CD4 T cells CD8 T cells BCRXL4 BCRXL5 BCRXL6 BCRXL7 BCRXL cluster 2.0 B cells IgM B cells IgM CD4 T cells CD8 T cells DC DC NK cells NK cells monocytes surface condition monocytes Patient surface Patient condition DC NK cells monocytes surface sample_id 2.0 Ref BCRXL The generalized linear mixed model (GLMM) for each cell 1.5 population! proportion Y number 1.5 of CD4 cells! π proportion of CD4 cells! 1.0 m total number of cells! 0.5 i = 1,,8 patient ID! j = 1, 2 condition! observation-level random effects!! random intercept for each patient! proportion! condition20 Y Bin(m, ) Ref BCRXL logit( 1.5 ij )= x ij + ij + i condition ij N(0, i N(0, 2 ) 2 ) patient_id Patient1 Patient2 Patient3 Patient4 Patient5 Patient6 p c

22 Statistical testing differential expression of signaling markers!

23 Statistical testing differential expression of signaling markers!

24 ! proportio DC NK cells monocytes surface 25 Statistical testing differential expression 3.0 of signaling Patient8 markers! B cells IgM+ 0 B cells IgM CD4 T cells CD8 T cells CD4 monocytes T cells condition patient_id CD8 surface T cells Ref Patient BCRXL Patient Patient condition Patient Patient Patient6 DC NK cells monocytes surface 1 Patient Patient DC 2.5 NK cells condition Ref antigen BCRXL The linear mixed model 2 (LMM) for each cell population! condition Y median expression 1 of 1 proportion marker in CD4 cells! median_expression Median expression! pnfkb pp38 pstat5 pakt pstat1 pshp2 pzap70 pstat3 pslp76 pbtk pplcg2 perk Y N(µ, plat ps6 2 ) 1 Patient6 Patient7 pnfkb pp38 pstat5 pakt pstat1 pshp2 pzap70 pstat3 pslp76 pbtk i = 1,,8 patient ID! j = 1, 2 condition! random intercept for each patient! Y ij = x ij + ij + i monocytes ij N(0, i N(0, 2 ) 2 ) surface