Régression en grande dimension et épistasie par blocs pour les études d association

Transcription

1 Régression en grande dimension et épistasie par blocs pour les études d association V. Stanislas, C. Dalmasso, C. Ambroise Laboratoire de Mathématiques et Modélisation d Évry "Statistique et Génome" 1 / 45

2 Summary 1 GWAS and Block of linkage desequilibrium Genome Wide Association Studies Blocks of linkage desiquilibrium Hierachical Clustering with Adjacency Constraints How to improve? Some computation times 2 Epistasis The Gene-Gene Eigen Epistasis Modeling approach Simulations 3 Application Ankylosing Spondylitis First results 2 / 45

3 Sommaire 1 GWAS and Block of linkage desequilibrium Genome Wide Association Studies Blocks of linkage desiquilibrium Hierachical Clustering with Adjacency Constraints How to improve? Some computation times 2 Epistasis The Gene-Gene Eigen Epistasis Modeling approach Simulations 3 Application Ankylosing Spondylitis First results 3 / 45

4 Single-Nucleotide Polymorphism Data 90 % of human genetic variation, In human genom, SNP with allelic frequency greater than 1 % are present every 300 base pairs (in average) 2 SNP among 3 substitute cytosine with thymine Figure: SNP (wikipedia) 2 4 / 45

5 SNP Data Four possible C/T configurations X ij {0, 1, 2} (Individual i at locus j) Father Mother C T C 0 1 T 1 2 High-dimension n individuals p markers X 11 X 1p.. X n1 X np up to few million of SNPs can be genotyped! number of variables number of individuals (p n) 5 / 45

6 Genome-Wide Association Studies GWAS characteristics : Objective : find associations between genetic markers (SNP i,j {0, 1, 2}) and a phenotypic trait (Y i {0, 1} or Y i R) Genetic markers SNP http :// c Pasieka, Science Photo Library 6 / 45

7 Genome-Wide Association Studies Generalized Linear Model g(e[y i x i ]) = β 0 + p β j x ij j=1, i = 1,..., n n : number of individuals p : number of covariates Y i : response for the individual i x.j : observations for covariate j (coded in 0, 1 or 2) 7 / 45

8 Genome-Wide Association Studies SNP analysis Differences between cases and controls at a specific SNP GWAS limits : Reproductibility Heritability Data particularities : Structuration High dimension (p» n) Small effects 8 / 45

9 The LD measures Linkage Desiquilibrium non-random association of alleles at two or more loci depends on the difference between observed allelic frequencies and those expected from a independent randomly distributed model. Computation Z j the indicator of the presence of minor allele for SNP j. Z j B(p j ) D(j, k) = p jk p j p k = E[Z j Zk] p j p k = cov(z j, Z k ) r 2 (j, k) = corr(z j, Z k ) ou D (j, k) = D(j, k)/dmax(j, k) 9 / 45

10 How to estimate LD? snp vv vv VV uu a b c uu d e f UU g h i snp v V u α β U γ δ Only the genotype data table is observed α, β, γ, δ are estimated a system of equations. e.g : α = 2a + b + d + pe with p the "probability" of the haplotype (uv, UV ). estimating p, then (α, β, γ, δ) and finally D = p UV p U p V. 10 / 45

11 The LD block structure the r 2 coefficients among the 50 first SNPs of the Chromosome 22 (Dalmasso et al. 2008) LD structured in blocks Row Column Dimensions: 50 x / 45

12 Hierachical Clustering with Adjacency Constraints Cluster Dendrogram Height corr LD / 45

13 Block-Wise Approach using Linkage Disequilibrium (BALD) 1 Hierarchical clustering of the SNPs with adjacency constraint and using the LD similarity. 2 Estimation of the optimal number of groups using the Gap statistic (Tibshirani et. al., 2001). 13 / 45

14 The h-band All coefficients outside the band h are null a p h similarity matrix a hierarchical clustering with adjacency constraint 14 / 45

15 A pseudocode Data: X {0, 1, 2} n p, Sim C {C i = {X.i }, i 1,..., p} /* clusters = singletons */ ; D {1 Sim(X.i, X.(i+1) ), i 1,..., p 1} ; for step = 1 to p 1 do i argmin i {1,...,p step} D(C i, C i+1 ) ; C C \ {C i, C i +1} {C i C i +1} ; d 1 D(C i 1, C i C i +1) ; d 2 D(C i C i +1, C i +2); D D \ {D(C i 1, C i ), D(C i, C i +1)} {d 1, d 2 } ; end 15 / 45

16 The Ward s distance Ward Constrained Hierarchical Clustering d(a, B) = n An B n A + n B ( 1 S na 2 A,A + 1 S nb 2 B,B 2 ) S A,B n A n B 16 / 45

17 The pencils trick : Calculating S AA and S AB 17 / 45

18 The pencils Assessing S AA, S BB and S AB requires the calculation of sums of LD measures within pencil-shaped areas defined by : direction : right or left depth : hloc end point : lim Two arrays of sizes p h for storing the pencils sums. 18 / 45

19 The binary min-heap All nodes are either less than or equal to each of its children. Uniquely represented by storing its level order traversal in an array. Given a position i : Parent(i) = i/2 Left(i) = 2i Right(i) = 2i / 45

20 DeleteMin Time complexity : O(log(p)) 20 / 45

21 InsertHeap Time complexity : O(log(p)) 21 / 45

22 BuildHeap Data: An array A Result: A min-heap H for i = length(a)/2 down to 1 do PercolDown(A, i); end Time complexity : O(plog(p)) 22 / 45

23 Time complexity of some operations findmin insert deletemin unordered array O(p) O(1) O(p) binary heap O(1) O(log(p)) O(log(p)) 23 / 45

24 cward in seconds... t (s) e+00 1e+05 2e+05 3e+05 4e+05 5e+05 p Figure: The mean computation time t versus the number of markers p for the cward algorithm applied to randomly sampled SNP matrices. N = 100, h = 30 and t is averaged across 50 simulation runs. 24 / 45

25 Compared to a former implementation t (s) pencils -heaps +pencils +heaps p Figure: The mean computation time t versus the number of markers p for the cward algorithm and an implementation without heaps. t is averaged across 20 simulation runs. 25 / 45

26 Scalable Hierarchical Clustering with pencils and binary heap To sum up : 1 A O(p 2 ) algorithm does not scale for GWA studies. 2 The Ward distance written in a simple way. 3 Space complexity of O(ph) by using the pencils trick. 4 Time complexity of : O(ph) }{{} + O(plog(p)) }{{} 2 p h arrays of pencils sums building the heap and insert/delete heaps operations within the loop Ongoing work : Currently implemented with a genotype matrix as input. can be generalized to any band similarity matrix. 26 / 45

27 Sommaire 1 GWAS and Block of linkage desequilibrium Genome Wide Association Studies Blocks of linkage desiquilibrium Hierachical Clustering with Adjacency Constraints How to improve? Some computation times 2 Epistasis The Gene-Gene Eigen Epistasis Modeling approach Simulations 3 Application Ankylosing Spondylitis First results 27 / 45

28 Epistasis Definition Interaction of alleles effects from different markers Existing methods mainly SNP x SNP some at the block (gene) scale Advantages of gene (or block) scale approaches results biologically interpretable genetic effects may be easier to detect reduce the number of variables 28 / 45

29 Epistasis - Gene scale methods Existing gene scale methods : Two or few genes PCA + logistic regression (He et al. 2011, Li et al. 2009, Zhang et al. 2008) PLS + logistic regression (Wang T et al. 2009) For a larger number of genes PCA + LASSO (D Angelo et al. 2009) PCA + pathway-guided penalized regression (Wang X et al. 2014) 29 / 45

30 Epistasis - Gene scale methods Existing gene scale methods : Two or few genes PCA + logistic regression (He et al. 2011, Li et al. 2009, Zhang et al. 2008) PLS + logistic regression (Wang T et al. 2009) For a larger number of genes PCA + LASSO (D Angelo et al. 2009) PCA + pathway-guided penalized regression (Wang X et al. 2014) Objectives : To develop a new gene scale method considers a more accurate definition of interaction variables, is applicable with with genes, takes into account the group structure 29 / 45

31 Group modeling approach SNP 1,1.. SNP 1,p1.. SNP G,1.. SNP G,pG Pheno Ind Ind Ind i }{{} gene 1 }{{} gene G We note SNP 1,1 = X 1,1 30 / 45

32 Group modeling approach SNP 1,1.. SNP 1,p1.. SNP G,1.. SNP G,pG Pheno Ind Ind Ind i }{{} gene 1 }{{} gene G We note SNP 1,1 = X 1,1 r, s two genes 30 / 45

33 Group modeling approach Model Y i = β 0 + g X g ik ) ( β g k C }{{} Main effects + ɛ i T β = β 1,1, β 1,2,, β 1,p1,, β G,1,, β G,p }{{} G }{{} gene 1 gene G 31 / 45

34 Group modeling approach Model Y i = β 0 + g X g ik ) ( β g k C }{{} Main effects + r,s γ r,s Z r,s i }{{} Interaction effects + ɛ i T β = β 1,1, β 1,2,, β 1,p1,, β G,1,, β G,p }{{} G }{{} gene 1 gene G couple γ = γ 12,... γ 1G }{{} γ 1G,1,,γ 1G,q..., γ (G 1)G q : # of interaction variables for a 31 / 45

35 Interaction variable : Gene-Gene Eigen Epistasis (G-GEE) We consider f u (X r i, X s i ) to represent the interaction between genes r, s. Eigen Epistasis û = arg max cor(y, f u(x r, X s )) u, u =1 f u (X r, X s ) = W rs u with W rs = {X r ij X ik s,pr ;k=1,,ps }j=1 i=1 n max cor[w ˆ u, y] 2 == max u T W rs T yy T W rs u u, u =1 u, u =1 u : only eigen vector associated of W rs T yy T W rs For each couple (r, s) Z rs = W rs u 32 / 45

36 Coefficients estimation Group LASSO regression (ˆβ, ˆγ) = argmin β,γ (y i X i β Z i γ) 2 + Limits of the grouplasso regression : i ( λ pg β g 2 + ) pr p s γ rs 2 g rs Difficult to compute p-value or confidence interval 33 / 45

37 Coefficients estimation Group LASSO regression (ˆβ, ˆγ) = argmin β,γ (y i X i β Z i γ) 2 + Limits of the grouplasso regression : i ( λ pg β g 2 + ) pr p s γ rs 2 g rs Difficult to compute p-value or confidence interval Adaptive-Ridge Cleaning Becu JM, 2015 Use of a specific penalty for group LASSO Permutation test based on Fisher test approach for each group P k = 1 B #{F k F k} 33 / 45

38 Interaction variable modeling approaches comparison methods criteria Z rs T i γ rs G-GEE cor(y, f u (X r, X s )) W rs uγ rs PCA var(g r v) and var(g s v) CCA cor(g r a, G s b) PLS cov(yg r c, G s w) q j=1 q k=1 γrs jk C r j C s k q j=1 γrs j A r j B s j q j=1 γrs j T rs j 34 / 45

39 Simulation design Genotype : X i N p (0, Σ) with Σ a block diagonal correlation matrix (ρ = 0.8 for two SNPs in the same gene) MAF j U[0.05, 0.5] with MAF j = 0.2 if j causal SNP Continuous phenotype simulated under two different schemes : from Wang X et al., 2014 : ) Y i = β 0 + g PCA model : β g ( k C Y i = β 0 + g X g ik β g ( k C + rs X g ik ) γ rs + rs Xij r Xik s (j,k) C 2 + ɛ i (1) γ rs C r i1c s i1 + ɛ i. (2) 35 / 45

40 Simulations design Scenarios : We consider 600 subjects and 6 SNPs by gene First scenario on 6 genes, two settings : same genes for main and interaction effects, different genes for main and interaction effects. Second scenario on 25 genes, one setting : different genes for main and interaction effects. 36 / 45

41 Simulations results - First scenario on 6 genes Gene Pair Power Wang X et al. model Methods G-GEE PLS CCA PCA Gene Pair Power PCA model Main effects : gene 1 gene 2 Interaction effects : gene 1 x gene r r 2 Wang X et al. model PCA model Gene Pair Power r 2 Gene Pair Power r 2 Main effects : gene 1 gene 2 Interaction effects : gene 3 x gene 4 37 / 45

42 Simulations results - First scenario on 6 genes Genes1 Genes2 Genes3 Wang X et al. model Genes1 Genes2 Genes3 PCA model Genes4 Genes5 Genes6 X.Genes1.Genes2 X.Genes1.Genes3 X.Genes1.Genes4 X.Genes1.Genes5 X.Genes1.Genes6 X.Genes2.Genes3 X.Genes2.Genes4 X.Genes2.Genes5 X.Genes2.Genes6 X.Genes3.Genes4 X.Genes3.Genes5 X.Genes3.Genes6 X.Genes4.Genes5 X.Genes4.Genes6 X.Genes5.Genes6 Genes4 Genes5 Genes6 X.Genes1.Genes2 X.Genes1.Genes3 X.Genes1.Genes4 X.Genes1.Genes5 X.Genes1.Genes6 X.Genes2.Genes3 X.Genes2.Genes4 X.Genes2.Genes5 X.Genes2.Genes6 X.Genes3.Genes4 X.Genes3.Genes5 X.Genes3.Genes6 X.Genes4.Genes5 X.Genes4.Genes6 X.Genes5.Genes6 Main effects : gene 1 gene 2 Interaction effects : gene 1 x gene 2 Genes1 Genes2 G-GEE CCA PCA PLS r 2 = 0.7 Wang X et al. model Genes1 Genes2 G-GEE CCA PCA PLS r 2 = 0.7 PCA model Genes3 Genes4 Genes5 Genes6 X.Genes1.Genes2 X.Genes1.Genes3 X.Genes1.Genes4 X.Genes1.Genes5 X.Genes1.Genes6 X.Genes2.Genes3 X.Genes2.Genes4 X.Genes2.Genes5 X.Genes2.Genes6 X.Genes3.Genes4 X.Genes3.Genes5 X.Genes3.Genes6 X.Genes4.Genes5 X.Genes4.Genes6 X.Genes5.Genes6 Genes3 Genes4 Genes5 Genes6 X.Genes1.Genes2 X.Genes1.Genes3 X.Genes1.Genes4 X.Genes1.Genes5 X.Genes1.Genes6 X.Genes2.Genes3 X.Genes2.Genes4 X.Genes2.Genes5 X.Genes2.Genes6 X.Genes3.Genes4 X.Genes3.Genes5 X.Genes3.Genes6 X.Genes4.Genes5 X.Genes4.Genes6 X.Genes5.Genes6 Main effects : gene 1 gene 2 Interaction effects : gene 3 x gene 4 G-GEE CCA PCA PLS r 2 = 0.6 G-GEE CCA PCA PLS r 2 = / 45

43 Simulations results - Second scenario on 25 genes Gene1 Gene10{ Gene1.Gene2 Gene3.Gene4 Main effects : gene 1 gene 2 Gene5.Gene6 Gene7.Gene8 Gene9.Gene10 Interaction effects : gene 3 x gene 4 gene 5 x gene 6 gene 7 x gene 8 gene 9 x gene 10 G-GEE CCA PCA PLS Figure: Wang X et al. model, r 2 = / 45

44 Sommaire 1 GWAS and Block of linkage desequilibrium Genome Wide Association Studies Blocks of linkage desiquilibrium Hierachical Clustering with Adjacency Constraints How to improve? Some computation times 2 Epistasis The Gene-Gene Eigen Epistasis Modeling approach Simulations 3 Application Ankylosing Spondylitis First results 40 / 45

45 Ankylosing Spondylitis Chronic inflammatory disease of the axial skeleton Epidemiology : Age at first symptoms : years Sexe : predominance for men (sex ratio 2M :1W) Prevalence : depend of populations (0.1% - 1.4%) Right etiology unknown : Environmental factors? Genetic factors? Importance of HLA complex HLA complex : Localized on chromosome 6 Regroup about 200 genes Coding the immunity system Antigen HLA-B27 : associated to SPA Effect from other gene in HLA group? Outside HLA group? 41 / 45

46 Known genes Tsui et al., 2014 : The genetic basis of ankylosing spondylitis : new insights into disease pathogenesis, The Application of Clinical Genetics : susceptibility genes identified by GWAS 42 / 45

47 Results Significant results G-GEE HLA-B x SULT1A1 IL23R x ERAP2 PLS HLA-B EOMES x BACH2 PCA HLA-B CCA - 43 / 45

48 Conclusions and perspectives The G-GEE method : Takes into account the gene structure of data Can be applied on a large number of genes Uses a specific interaction modeling approach Ankylosing Spondylitis : Identification of potential interactions to discuss with doctors HLA-B effect Perspectives : Explore new f u (X r i, X s i ) definition Additional simulations on larger data set New applications on other data set 44 / 45

49 Thank you for your attention! 45 / 45

50 Adaptive-Ridge Cleaning λ specific penalty for group LASSO : k(j) ˆθ m k(j) m 2