Fully Bayesian analysis of allele-specific RNA-seq data … · Var ∆ g|y) = Var(β g2|y ......

1
Fully Bayesian analysis of allele-specific RNA-seq data using a hierarchical, overdispersed, count regression model Ignacio Alvarez Jarad Niemi Dan Nettleton Department of Statistics, Iowa State University Allele-specific gene expression Diploid orgnisms have two copies of each genes (alleles) that can be separately transcribed. The RNA abundance of any particular allele is known as allele-specific expres- sion (ASE). In plant breeding, hybrids benefits from heterosis (hybrid vigor). ASE is relevant for the study of this phenomenon at the molecular level. Goals of the study: We present statistical methods for modelling ASE and detecting genes where differential allele expression. We propose a hierarchical overdispersed Poisson model to deal with ASE counts. Maize experimental dataset Dataset : Data from Paschold et al. (2012). Design : Hybrid genotype (B73xMo17), 2 flow cell blocks, 4 replicate plants per block, 3 measures per plant allele_b allele_m total B73 B73xMo17 Mo17 1 2 3 4 1 2 3 4 1 2 3 4 0 5 10 15 0 5 10 15 0 5 10 15 replicate log(cnt + 1, 2) replicate 1 2 3 4 Poisson-lognormal hierarchical model Data model Y gn : ASE count of gene g , in sub-sample n Y gn ind Po(e h n +x n β g + gn ) gn ind N (0g ) (1) h n : normalization factor p regression coefficients (p 2): gn : overdispersion with gene-specific variance Gene-specific layer Regression coefficients β gk ind N (θ k 2 k ξ gk ) ξ gk p(η ) Shrinkage distributions: Student-t, Laplace, horseshoe, normal Overdispersion variances γ g ind IG( ν 2 , ντ 2 ) γ g are shrunk around τ , ν controls amount of shrinkage. Allele effect: Δ g we set β g 2 as the half allele difference Δ g = β g 2 - θ 2 . Bias correction θ 2 : systematic difference among alleles. Differential expression {|Δ g | >c} Credible interval Use posterior mean and variance for normal approximation. E g |y ) = E (β g 2 |y ) - E (θ 2 |y ) V ar g |y )= V ar (β g 2 |y )+ V ar (θ 2 |y ) - 2Cov (β g 2 2 |y ) V ar (β g 2 |y ) Full Bayesian inference using GPU Hyperparameter prior will not have a large impact in the gene-specific parameters (Ghosh et al. 2006) θ k ind N(0,c k ) ν Unif(0,d) σ k ind Unif(0,s k ) τ Ga(a, b) Obtain fully Bayesian inference with fbseq package (Landau and Niemi 2016) Parallel MCMC algorithm using GPU 2 hours for single hybrid data, 10 when we include 3 varieties. fbseq output: posterior summaries for gene-specific parameters Results from Simulation Study Simulate 24 scenarios Sparsity (w ): proportion of genes with NO effect. Strengh (s): enlarge factor for DE genes Bias (p): proportion of non-reference allele lost due to bias Overdispersion (T ): multiplicative factor of overdispersion Table 1: Simulation study design parameter values Description Sparsity Strength Bias Overdispersion Parameter w s p T Values (.5,.95) (1, 1.8) (1,.5) (0.25, 1, 4) ROC curves from simulations s = 1 w = 0.5 s = 1 w = 0.95 s = 1.8 w = 0.5 s = 1.8 w = 0.95 T = 0.25 T = 1 T = 4 0.05 0.15 0.25 0.05 0.15 0.25 0.05 0.15 0.25 0.05 0.15 0.25 0.00 0.10 0.25 0.50 0.75 0.90 1.00 0.00 0.10 0.25 0.50 0.75 0.90 1.00 0.00 0.10 0.25 0.50 0.75 0.90 1.00 True positive fraction Hierarchical non-Hierarchical t 1 (Cauchy29 Laplace t 6 normal non-hierarchical model does not capture bias Cauchy slightly better when w =0.95 or T =4 Bayesian analysis of maize data β kg Cauchy (θ k k ) is based on the results from simulation study. Non-Diff Exp. Diff Exp. 0 2 4 6 0 2 4 6 -2.5 0.0 2.5 Expected expression (log) Allele effect (95% CI) Non-Diff Exp. Diff Exp. |Δ g | >c in 17% of genes, c = log(1.25)/2 (25% fold change) Higher expression = narrower CI Some DE genes with low expression Few genes with high overdispersion variances Bias: E (θ 2 |y )=0.126, suggest 1 out 5 reads from Mo17 is lost results suggest σ 4 10σ 5 mean CI 95% ν 3.6 (3, 4.3) τ 0.0023 (0.0019, 0.0028) θ 2 0.12 (0.12, 0.13) σ 2 4 0.0011 (0.00094, 0.0012) σ 2 5 0.000015 (0.0000095, 0.000023) 0.01 0.10 0 2 4 6 Expected expression (log) Posterior expecation γ g 1 10 100 Frequence

Transcript of Fully Bayesian analysis of allele-specific RNA-seq data … · Var ∆ g|y) = Var(β g2|y ......

Fully Bayesian analysis of allele-specific RNA-seq data using ahierarchical, overdispersed, count regression model

Ignacio Alvarez Jarad Niemi Dan NettletonDepartment of Statistics, Iowa State University

Allele-specific gene expression

Diploid orgnisms have two copies of each genes (alleles)that can be separately transcribed. The RNA abundanceof any particular allele is known as allele-specific expres-sion (ASE).

• In plant breeding, hybrids benefits from heterosis(hybrid vigor).

• ASE is relevant for the study of this phenomenonat the molecular level.

Goals of the study:We present statistical methods for modellingASE and detecting genes where differentialallele expression. We propose a hierarchicaloverdispersed Poisson model to deal with ASEcounts.

Maize experimental dataset

Dataset : Data from Paschold et al. (2012).Design : Hybrid genotype (B73xMo17), 2 flow cell

blocks, 4 replicate plants per block, 3 measuresper plant

●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●● ●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●●●●●●● ●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●

●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●

●●

●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●

●●●●●●●●●●●●●●●●●

●●●●●●●●

●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●●

●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●

●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●

●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●

●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●

●●

●●●●●

●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●

●●●

●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●

●●

●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●

●●●●●

●●●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●

●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●

●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●

●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●

●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●

●●●

●●●●●●●●●●●●●●●●

●●●●●●●

●●●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●

●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●

●●●●●●●●●●

●●●●●●

●●●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●

●●●●●●●●

●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●

●●●●●●

●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●

●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●

●●●●●●

●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●

●●●●●●●

●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●● ●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●

●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●●●●●●●●●●

●●●●●

●●

●●

●●●●●

●●●●

●●●

●●●

●●●●●●●

●●●

●●●●

●●

●●●●●

●●●

●●●●●

●●●

●●

●●

●●●●●

●●●●

●●●

●●

●●

●●●●●●

●●●●

●●

●●

●●●

●●

●●

●●

●●●●●●●

●●●●●●●●●●●

●●

●●

●●

●●

●●●●

●●●●●

●●●●●●

●●

●●●

●●●●●

●●

●●

●●

●●

●●

●●●

●●

●●●

●●

●●

●●

●●●

●●●

●●●●

●●●●●

●●

●●●

●●

●●●

●●

●●

●●●●●●●●●

●●

●●

●●

●●●●

●●●

●●●

●●●

●●

●●●●●●●

●●

●●

●●

●●●●●●●

●●

●●

●●●

●●●

●●●●

●●

●●

●●

●●

●●●●●●

●●●

●●

●●

●●●●●●●

●●●

●●

●●●

●●

●●●●●

●●●●

●●●

●●●

●●●●●

●●●●●

●●

●●

●●●●●●●

●●

●●●●●●

●●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●

●●●●●●●●●●

●●

●●●●●●●●●

●●

●●●●

●●

●●

●●●

●●●

●●●●

●●

●●●●●

●●●●●

●●

●●●

●●●

●●

●●

●●

●●●●●●

●●

●●

●●●

●●●●●

●●

●●

●●●

●●

●●

●●

●●

●●

●●●

●●●● ●

●●●●●

●●●

●●

●●

●●●

●●●●

●●

●●●

●●●●●●●●●

●●●

●●

●●●●

●●●

●●●

●●

●●●

●●

●●

●●●●●●●●●●●●●

●●

●●●●●●

●●

●●●

●●●

●●

●●●●●●

●●

●●

●●●●

●●●●

●●

●●●●●●●●●●●●●

●●

●●●

●●

●●

●●●

●●●●●

●●

●●●●●●●●●

●●

●●

●●●●●

●●●●

●●

●●

●●

●●●

●●

●●●●

●●

●●

●●

●●●●●●●●●●●

●●

●●

●●●

●●●

●●●●

●●●●●●●

●●●●● ●

●●●

●●●

●●●●●●●●●

●●●●●●●●●

●●●●●●

●●●

●●

●●

●●●●●●●●●●●●●●●●●●●

●● ●●●●●

●●●●

●●●●●●●●●●●●●

●●●●

●●●●●

●●●●●●●

●●●

●●●

●●●●●

●●

●●●

●●●●

●●●

●●

●●●●●●●●●●●●●●●

●●●

●●

●●●●●

●●

●●

●●●

●●

●●

●●

●●●

●●●

●●●●●●●●●●●

●●

●●●

●●●●●●

●●

●●

●●●●

●●●●●

●●●●●●

●●●●●

●●●●●

●●

●●

●●●●●●●●

●●

●●

●●●●●●●

●●●●●●●●●●●●●●●●●

●●●●●●●●●●●●●●●●●

●●●

●●●●

●●●

●●●

●●●

●●●●

●●●●

●●●

●●

●●●●

●●

●●

●●●●

●●

●●

●●

●●●

●●●●●

●●●●●●

●●

●●

●●●●●●

●●●

●●●●

●●●

●●

●●

●●

●●

●●

●●

●●●●

●●

●●

●●

●●

●●

●●●●●

●●

●●●

●●●●

●●●●●●●●●●●●●●●●●●●●

●●●

●●●●●

●●

●●

●●●●

●●

●●

●●●

●●

●●

●●●

●●

●●●●

●●

●●●●

●●

●●

●●●●

●●

●●●●●●

●●●●●●

●●●●●●

●●●

●●●●●●●

●●●●●●●●

●●●

●●

●●●

●●●●●●●●●

●●

●●●●●●●●●●●●●●

●●●

●●●

●●

●●

●●

●●●

●●●●●

●●●●

●●

●●

●●

●●●●

●●●●

●●●●●●●

●●

●●●

●●●●●●●●

●●

●●●●●●

●●●●●

●●

●●●

●●

●●

●●●●●

●●●

●●●●●●

●●

●●●●●

●●●●●●

●●

●●●●●●●●●●●●●

●●●●

●●●●●

●●

●●●

●●●

●●●●●

●●

●●●

●●●●

●●

●●●●

●●

●●●●●●●

●●●●●●

●●

●●

●●

●●

●●●

●●

●●

●●

●●●

●●●

allele_b allele_m total

B73

B73xM

o17M

o17

1 2 3 4 1 2 3 4 1 2 3 4

0

5

10

15

0

5

10

15

0

5

10

15

replicate

log(

cnt +

1, 2

) replicate

1

2

3

4

Poisson-lognormal hierarchical model

Data modelYgn : ASE count of gene g, in sub-sample n

Ygnind∼ Po(ehn+x>nβg+εgn)

εgnind∼ N(0, γg)

(1)

• hn: normalization factor• p regression coefficients (p ≥ 2):• εgn: overdispersion with gene-specific variance

Gene-specific layerRegression coefficients βgk ind∼ N(θk, σ2

kξgk) ξgk ∼ p(η)• Shrinkage distributions: Student-t, Laplace, horseshoe,

normal

Overdispersion variances γg ind∼ IG(ν2,ντ2 )

• γg are shrunk around τ , ν controls amount ofshrinkage.

Allele effect: ∆g

we set βg2 as the half allele difference∆g = βg2 − θ2.

Bias correction θ2: systematic difference among alleles.Differential expression {|∆g| > c}

Credible interval Use posterior mean and variance fornormal approximation.

E(∆g|y) = E(βg2|y)− E(θ2|y)V ar(∆g|y) = V ar(βg2|y) + V ar(θ2|y)− 2Cov(βg2, θ2|y)

≈ V ar(βg2|y)

Full Bayesian inference using GPU

• Hyperparameter prior will not have a large impact in thegene-specific parameters (Ghosh et al. 2006)

θkind∼ N(0, ck) ν ∼ Unif(0, d)

σkind∼ Unif(0, sk) τ ∼ Ga(a, b)

Obtain fully Bayesian inference with fbseq package (Landauand Niemi 2016)

• Parallel MCMC algorithm using GPU• 2 hours for single hybrid data, 10 when we include 3

varieties.• fbseq output: posterior summaries for gene-specific

parameters

Results from Simulation Study

Simulate 24 scenarios• Sparsity (w): proportion of genes with NO effect.• Strengh (s): enlarge factor for DE genes• Bias (p): proportion of non-reference allele lost due to

bias• Overdispersion (T ): multiplicative factor of

overdispersion

Table 1: Simulation study design parameter valuesDescription Sparsity Strength Bias OverdispersionParameter w s p TValues (.5, .95) (1, 1.8) (1, .5) (0.25, 1, 4)

ROC curves from simulationss = 1

w = 0.5

s = 1

w = 0.95

s = 1.8

w = 0.5

s = 1.8

w = 0.95

T=

0.25T

=1

T=

4

0.05 0.15 0.25 0.05 0.15 0.25 0.05 0.15 0.25 0.05 0.15 0.25

0.000.100.25

0.50

0.750.901.00

0.000.100.25

0.50

0.750.901.00

0.000.100.25

0.50

0.750.901.00

False positive fraction

True

pos

itive

frac

tion

Hierarchical

non−Hierarchical

t1 (Cauchy)Laplace

t6

normal

• non-hierarchical model does not capture bias• Cauchy slightly better when w = 0.95 or T = 4

Bayesian analysis of maize data

• βkg ∼ Cauchy(θk, σk) is based on the results from simulation study.

Non−Diff Exp. Diff Exp.

0 2 4 6 0 2 4 6

−2.5

0.0

2.5

Expected expression (log)

Alle

le e

ffect

(95

% C

I)

Non−Diff Exp.

Diff Exp.

• |∆g| > c in 17% of genes, c = log(1.25)/2 (25% fold change)• Higher expression = narrower CI• Some DE genes with low expression• Few genes with high overdispersion variances• Bias: E(θ2|y) = 0.126, suggest 1 out 5 reads from Mo17 is lost• results suggest σ4 ≈ 10σ5

mean CI 95%ν 3.6 (3, 4.3)τ 0.0023 (0.0019, 0.0028)θ2 0.12 (0.12, 0.13)σ2

4 0.0011 (0.00094, 0.0012)σ2

5 0.000015 (0.0000095, 0.000023)

0.01

0.10

0 2 4 6

Expected expression (log)

Pos

terio

r exp

ecat

ion

γ g

1 10 100Frequence