Telescoping bimodal latent Dirichlet allocation to identify expression QTLs across tissues

Abstract

Expression quantitative trait loci (eQTLs), or single nucleotide polymorphisms (SNPs) that affect average gene expression levels, provide important insights into context-specific gene regulation. Classic eQTL analyses use one-to-one association tests, which test gene-variant pairs individually and ignore correlations induced by gene regulatory networks and linkage disequilibrium. Probabilistic topic models, such as latent Dirichlet allocation, estimate latent topics for a collection of count observations. Prior multi-modal frameworks that bridge genotype and expression data assume matched sample numbers between modalities. However, many data sets have a nested structure where one individual has several associated gene expression samples and a single germline genotype vector. Here, we build a telescoping bimodal latent Dirichlet allocation (TBLDA) framework to learn shared topics across gene expression and genotype data that allows multiple RNA-sequencing samples to correspond to a single individual’s genotype. By using raw count data, our model avoids possible adulteration via normalization procedures. Ancestral structure is captured in a genotype-specific latent space, effectively removing it from shared components. Using GTEx v8 expression data across ten tissues and genotype data, we show that the estimated topics capture meaningful and robust biological signal in both modalities, and identify associations within and across tissue types. We identify 53,358 cis-eQTLs and 1,173 trans-eQTLs by conducting eQTL mapping between the most informative features in each topic. Our TBLDA model is able to identify associations using raw sequencing count data when the samples in two separate data modalities are matched one-to-many, as is often the case in biological data.

Expression QTLs in single-cell sequencing data

“…Notably, 11/15 (73%) of those trans-eGenes showed heritable enrichment associated with the active enhancer marker H3K27ac. This result echoes earlier findings between trans-eGenes and enhancer markers [1,17].…”

“…To enable adaptation to single-cell sequencing data, we developed three functionally-important modifications to the TBLDA model presented in prior work [17]: GPU-compatibility, a sparsity-inducing prior for the loading matrices, and amortization during inference. We refer to the new model as single-cell TBLDA (scTBLDA).…”

“…Due to the encouraged sparsity in the posterior, we modified how we select the top features in each factor from the previous 2-Wasserstein-based metric [17]. To correct for the correlation between feature levels and factor weights, we conducted quantile regression on the factor loadings to adjust for allele frequency and average expression level (Fig.…”

“…We include all of these ideas in our probabilistic framework telescoping bimodal latent Dirichlet analysis (TBLDA), using both data modalities to jointly determine a set of latent factors that each correspond to a collection of covarying genes and SNPs. In prior work on TBLDA, the estimated factors contain sets of associated genes from multiple tissues using bulk RNA-seq (instead of many cells from the same individual) and SNPs measured in a single donor [17]. In this paper, we expand the TBLDA framework to enable efficient inference across large-scale scRNA-seq data sets.…”

A Poisson reduced-rank regression model for association mapping in sequencing data

“…Recently, a latent variable model based on latent Dirichlet allocation (Blei et al, 2003; Pritchard et al, 2000) for jointly modeling gene expression and genotype was proposed (Gewirtz et al, 2021). This model projected both genotype data — using an equivalent of the Structure model (Pritchard et al, 2000) — and count-based gene expression data — using a telescoping LDA model (Blei et al, 2003) — onto a shared latent subspace; we may then identify covarying genes and gentoypes in a nonnegative latent representation.…”