Genomic data based drug discovery/repurposing. Di Wu Oct 2015

Transcription

1 Genomic data based drug discovery/repurposing Di Wu Oct

2 Three aspects, three stories Identifying cell of origin of breast-cancer subtypes from normal mammary cell types. Help find relevant Drug targets Drug sensitivity across lung cancer subtypes Integrating GWAS risk SNPs with public drug databases for drug repurposing 2

3 Breast cancer is a heterogeneous disease month Basal-like luminalb luminala Are there different cells of origin for the different cancer subtypes? 3 Perou et al, 2000; Sorlie et al, 2001; Sotiriou et al, 2003, Sorlie et al, 2003;

4 Normal cell types in human mammary glands /basal Cross-section of duct Most breast cancers initiate in ductal tissue Stem cell Luminal progenitor Luminal 4

5 Microarray datasets Breast tumor subtypes Human normal mammary cell types Data from collaboration with Jane Visvader and Geoff Lindeman s lab in WEHI

6 4 cell types in normal human mammary gland ducts Homogeneous ML LP Mature Luminal Luminal Progenitor Heterogeneous Stroma Elgene Lim Stromal (mainly fibroblast) In vivo Mammary Stem Cell (MaSC)-enriched Xenotransplantation MS

7 stem cell luminal progenitor mature luminal stromal cell Illumina human beadchips

8 Normal cell types have distinct expression profiles Mammary gland Stem cell Luminal Progenitor Stromal cell Mature Luminal Multidimensional Scaling (MDS) plot 8

9 Breast cancer subtype data Expression profiles of human breast tumors from C Perou s group. 2-color Agilent Human Microarray Cancer subtype Number of samples Basal-like 33 Claudin-low 5 HER2+/ER- 14 Luminal A 23 Luminal B 14 Normal breast-like 5 Total 94 Herschkowitz et al Genome Biology (2007)

10 Do the tumors show traces of transcriptional signature from normal mammary cell types? Tumor subtypes Normal mammary cell types Which tumor subtype is most similar to which cell type? 10

11 Challenging to create a link between both datasets and find the traces statistically normal cell types tumor subtypes Heatmap was made based on the normalized expression value of 100 genes with largest variability across normalized combined data. 11

12 Signature gene sets for normal cell types

13 Signature genes for normal cell types Fold change>2, FDR p< up Stroma 397 down MS LP ML 358 LP MS ML 257 Significantly up (or down) in one cell type versus other types Discriminatory strength of each signature gene is average logfold change, versus the other cell types 13

14 Considering one positive LP signature gene expression in tumors CHI3L1 normal-like basal-like Expression CHI3L1 tumor subtypes

15 Considering another positive LP signature gene expression in tumors TFCB2L1 Expression TFCB2L1 tumor subtypes

16 logfc Combining two LP positive signature genes 2.18 CHI3L1 2.5 TFCB2L1 Combination Signature score Combination tumor subtype

17 Combining all LP signature genes Signature score in tumors = average log-expression of positive LP signature genes average log-expression of negative LP signature genes LP signature (weighted by size of log-fold-change between LP vs other cell types) higher scores more like luminal progenitor t test * p < 0.05, ** p < 0.01, *** p = Highest expression of luminal progenitor gene signature in basal-like breast cancer cell of origin for basal tumor

18 Score is the average log-expression of tumour samples weighted by log-fold-change between cell types Si, x y = x gigj,, g j g xg, For the gene g in gene set Gi for cell type i i is average log fold change between cell type i to other cell types g, i yg, j is log expression in tumour sample j Higher scores = high similarity between that cell type and tumor samples 18

19 MaSC enriched signature Stroma signature Mature Lum signature Lum Prog signature Similarities between normal mammary cell types and tumor subtypes 19 t test, * p < 0.05, ** p < 0.01, *** p = 0.001

20 Concordance Tumor subtypes Normal-like Claudin-low Basal-like Luminal A, B Normal mammary cell types MaSC enriched Luminal Progenitor Mature Luminal 20

21 Gene ranking gives expression barcodes Tumor subtype A Rank genes by differential expression (between 2 tumour types) Negative signature genes of a cell type Gene 1 Gene 2 Gene 3 Gene 4 Gene 5 Gene 6 Gene 7 Gene 8 Gene 9 Gene 10 Gene 11 Gene 12 Gene 13 Gene 14 Gene 15 Gene 16 Tumor subtype B Tumor subtype A Positive signature genes of a cell type Genomewide barcode plot Tumor subtype B 21

22 LP signature is higher in basal tumors Tumor Wilk P-values Basal-like tumor Normal < 10-6 like Claudin < 10-6 < 10-4 Low ERBB2 < 10-6 < 10-6 Lum B < 10-6 < 10-6 Lum A < 10-6 <

23 Permuting genes vs permuting arrays label label label Permuting array labels Tests significance of gene set in isolation: self-contained test label label label Goeman & Bühlmann 2007 Permuting gene labels Tests ranking of gene set relative to other genes: competitive test 23

24 Correlation Adjusted MEan RAnk gene set test (camera) Variance of mean-rank (r ) of G genes is determined only by the average correlation where Nucleic Acid Research, 2012 May 24

25 Basal-like tumor LP signature is higher in basal tumors P-values Tumor Wilk camera Normall ike < Claudin Low < 10-6 < ERBB2 < 10-6 < 10-6 < 10-3 < 10-4 Lum B < 10-6 < 10-6 < Lum A < 10-6 < 10-6 < 10-5 < 10-3 Correcting for inter-gene correlation 25

26 Permuting genes vs permuting arrays label label label Permuting array labels Tests significance of gene set in isolation: self-contained test label label label Goeman & Bühlmann 2007 Permuting gene labels Tests ranking of gene set relative to other genes: competitive test 26

27 Roast, rotation gene set test Used with linear models (allows complex designs) Replaces array permutation with random rotation of residuals (avoids granularity of Pvalues, handles small sample sizes) Assumes multivariate normality, but proves to be highly robust to deviations from normality Bioinformatics, July

28 Two steps: projection and rotation Project data onto the null-hypothesis residual space (eliminates coefficients not interested in) Random rotation of the remaining orthogonal space (coefficient-of-interest + residuals) Monte-Carlo p-value 28

29 Roast select genes Genes in set

30 Roast projection Project data onto space orthogonal to nuisance parameters in linear model X design matrix expression matrix

31 Roast projection Project data onto space orthogonal to nuisance parameters in linear model X independent design residuals test matrix parameter nuisance parameters

32 Roast projection Discard nuisance parameters X independent design residuals test matrix parameter nuisance parameters

33 Roast rotation Rotate effects to simulate null hypothesis Rotation matrix Rotation matrix T RRI= Parallel rotation preserves correlation structure

34 Roast Allows gene-wise weights and choice of test statistics Can combine positive and negative signature genes in one test, using genewise directional weights 34

35 P-values LP signature is higher in basal tumors Basal-like tumor Tumor Wilk camera roast Normall < 10-6 ike Claudin < 10-6 < 10-4 Low ERBB2 < 10-6 < 10-6 < 10-3 < Lum B < 10-6 < 10-6 < Lum A < 10-6 < 10-6 < 10-5 < Allowing for inter-gene correlation35

36 Summary Luminal progenitor might be the cell of origin for basal-like breast tumors. LP signature genes May be drug target of basal-like tumour Can be used for classification and risk prediction 36

37 Drug target in luminal progenitor signature for basal-like cancer c-kit: proto-oncogene, mutations associated with leukaemia and gastrointestinal stromal tumour. Inhibitors in clinical use. 37

38 Our other stories of relating gene expression datasets 38

39 Three aspects, three stories Identifying cell of origin of breast-cancer subtypes from normal mammary cell types. Help find relevant Drug targets Drug sensitivity across lung cancer subtypes Integrating GWAS risk SNPs with public drug databases for drug repurposing 39

40 SqCC Wilkerson et al., Clinical Cancer Research

41 Associate SqCC subtypes with airway cell culture time 41

42 42

43 43

44 Log expression Proliferation drug sensitivity drug sensitivity Less drug sensitivity in secretory SqCC cell line due to low proliferation Cell lines SqCC tumors Cell lines 44

45 Three aspects, three stories Identifying cell of origin of breast-cancer subtypes from normal mammary cell types. Help find relevant Drug targets Drug sensitivity across lung cancer subtypes Integrating GWAS risk SNPs with public drug databases for drug repurposing 45

46 Drug Repurposing Approved small molecules treating other diseases Two nice properties of drug repurpose Cost effectiveness Less safety issues

47 Two well-known examples Repurpose Thalidomide Morning sickness Viagra (sildenafil) (1994) 2006 multiple myeloma Repurpose Pulmonary hypertension 1998 erectile dysfunction 47

48 Why genetics-based drug repurposing? Published risk SNPs for 17 traits Publishe d G enome-widea ssociations th rough 12/ 2102 P< 5X10^-8 Published GWA at p 5X10 for 17rt ait c ategories -8 Genetics help better understand disease mechanism Previous drug repurposing is mostly not genetics based More identified risk SNPs based on GWAS in past 6-7 years HN GRIG WA Catal o g gov/ G WAStudie s f gpt/gwas/ 2008 (According to NIH GWAS catalog database, Use of genome-wide association studies for drug repositioning Sanseau et al, Nature Biotechonology

49 Our goals Genetics based drug repurposing How genetics guides drug discovery

50 Simple strategies to use genetics guiding drug repurposing Pathway/Network based risk SNPs from GWAS risk genes direct overlap Drug database one protein/gene drug GWAS: genome-wide association studies 50

51 Target Drug 1, Is a drug target also a risk gene? direct overlap method 2, Is a PPI neighbor of drug target also a risk gene? 1 way PPI method Risk gene in one disease PPI neighbors > 1,000 drugs > 500 target genes > 2,500 drug-gene pairs 3, are the neighbors are significantly overlapped? 2 way PPI method 51

52 52

53 53

54 Results 54

55 1/7 are indicated for RA direct overlap 92 SNPs 55 significant genes generic Name drug ID targ action snp Denileukin diftitox Intravenous Immunoglobulin DB00004 IL2RA binder DB00028 C5 binder rs rs , rs , rs Basiliximab DB00074 IL2RA antibody rs Muromonab DB00075 CD247 Daclizumab DB00111 IL2RA antibody Eculizumab DB01257 C5 antibody rs rs , rs , rs Ipilimumab DAP CTLA4 antibody rs231735, rs Target genes: IL2RA, C5, CTLA4 Direction? rs Rheumatoid Arthritis (RA) 55

56 4/10 of 1 way PPI expansion are for RA (top 10, 8/159 total) Generic Name Abatacept Antithymocyte globulin Abatacept Sulfacetamide Sulfoxone Sulfanilamide Rituximab Ibritumomab Tositumomab Intravenous Immunoglobulin drug ID DB01281 targ CD80 DB00098 DB01281 DAP DAP DAP DB00073 DB00078 DB00081 CD86 CD86 MPI MPI MPI MS4A1 MS4A1 MS4A1 DB00028 C5 action antagonist nppi opg 5 CTLA4 hyp.p snpid 0.08 rs231735, rs antagonist Inhibitor Inhibitor Inhibitor antibody antibody antibody CTLA4 CTLA4 PTPN2 PTPN2 PTPN2 CD40 CD40 CD rs231735, rs rs231735, rs rs , rs rs , rs rs , rs rs rs rs binder 13 C rs

57 57

58 Nature

59 ROC curves determine the best rank ipn hy p.p pr o p tpr tpr 0.12 Rheumat o id art h rit i s drugg ene 0.12 Rheumat o id art h rit i s ROC drugg ene p rf For RA p rf ROC: receiver operating characteristic tpr tpr 052 g urd ntop drugroc ntop 59

60 60

61 Drug and disease information not enough False negatives FN: disease s drugs not in repurpose table Drug and disease are not connected by genetics Type 1 Diabetes Anti-TNFa for RA Drug and diseases are connected by genetics, but not through identified risk genes PPI neighbors 61

62 Reasons to call false positives FP: drugs not known for the disease, but in repurpose table; can be rep drugs Threshold for risk genes maybe not high enough. Non- relevant genes were falsely considered as risk genes The drug action directions is not clear Drugs with the same target gene in the same action direction may be indicated for different diseases 62

63 63

64 Result summary We repurposed approved drugs for 286 disease-relevant phenotypes ROC and significant tests for possible combination of methods True disease drugs are enriched in repurposing table About 40% of risk SNPs are used for repurposing 64

65 Three aspects, three stories Identifying cell of origin of breast-cancer subtypes form normal mammary cell types. Help find relevant Drug targets Drug sensitivity across several lung cancer cell types Integrating GWAS risk SNPs with public drug databases for drug repurposing 65

66 66

67 Harvard Statistics Jun S Liu Ke Deng Ming Hu Yushu Pang Harvard Biostatistics Tianxi Cai Acknowledgments BWH, Harvard Medical School Robert Plenge (MERCK) Yukinori Okada Dorothee Diogo Bobby Reddy Gang Li Soumya Raychaudhuri Towfique Saj Gyan Srivastava Broad Institute Paul Clemons Vlado Dancik Bioinformatics Division WEHI Gordon Smyth Terry Speed Victoria Breast Cancer Research Consortium WEHI Jane Visvader Geoff Lindeman François Vaillant Marie-Liesse Asselin-Labat Dana-Farber Cancer Institute Elgene Lim Dana Farber Cancer Institute Broad Institute Matthew Meyerson Peter Hammerman Heidi Greulich University of North Carolina Chapel Hill Matthew D. Wilkerson Neil Hayes Texas A&M University Alan Dabney Monash University Bryan Williams Die Wang Beijing Jiaotong University Xuezhong zhou 67