[Paper Review] A joint model of unpaired data from scRNA-seq and spatial transcriptomics for imputing missing gene expression measurements

Romain Lopez, Achille Nazaret+5 more|arXiv (Cornell University)|May 6, 2019

Single-cell and spatial transcriptomics63 citations

TL;DR

gimVI is a deep generative model that integrates unpaired scRNA-seq and spatial transcriptomics data to impute missing genes, using variational inference and an adaptable latent space with a controllable domain-adaptation term.

ABSTRACT

Spatial studies of transcriptome provide biologists with gene expression maps of heterogeneous and complex tissues. However, most experimental protocols for spatial transcriptomics suffer from the need to select beforehand a small fraction of genes to be quantified over the entire transcriptome. Standard single-cell RNA sequencing (scRNA-seq) is more prevalent, easier to implement and can in principle capture any gene but cannot recover the spatial location of the cells. In this manuscript, we focus on the problem of imputation of missing genes in spatial transcriptomic data based on (unpaired) standard scRNA-seq data from the same biological tissue. Building upon domain adaptation work, we propose gimVI, a deep generative model for the integration of spatial transcriptomic data and scRNA-seq data that can be used to impute missing genes. After describing our generative model and an inference procedure for it, we compare gimVI to alternative methods from computational biology or domain adaptation on real datasets and outperform Seurat Anchors, Liger and CORAL to impute held-out genes.

Motivation & Objective

• Integrate scRNA-seq and spatial transcriptomics data in a shared latent space to enable imputation of missing genes.
• Develop a generative model that accounts for technology-specific covariate shifts between modalities.
• Provide a probabilistic framework that quantifies uncertainty in imputations.
• Benchmark gimVI against existing methods (Liger, Seurat, CORAL, MAGAN) on real datasets.
• Explore how imputation quality relates to latent space alignment and gene observability.

Proposed method

• Extend scVI with a shared latent variable z and modality-specific variables to capture sequencing depth and protocol effects.
• Model observed counts with appropriate distributions (ZINB for scRNA-seq, Poisson/NB for spatial data) conditioned on z and nuisance variables.
• Use neural nets to map z,s to gene frequency expectations via a softmax layer, yielding rho for imputation.
• Perform variational inference with Gaussian posteriors q_phi(z, l | x,s) and q_psi(z | x',s) and optimize KL-regularized lower bounds.
• Impute unseen genes by sampling from the latent posterior and predicting counterfactual expressions with a g_eta network, informed by domain adaptation theory (H-divergence approximated by an adversarial loss).
• Benchmark via joint latent-space integration metrics (entropy of mixing, k-NN purity) and imputation performance (Spearman rho) across real datasets (mSMS, mPFC).

Experimental results

Research questions

• RQ1 Can unpaired scRNA-seq and spatial transcriptomics data be embedded in a common latent space to enable accurate imputation of missing genes?
• RQ2 Does gimVI balance dataset integration quality with preservation of biology, and can it quantify uncertainty in imputations?
• RQ3 How does the imputation performance of gimVI compare to Seurat, Liger, scVI, CORAL, and MAGAN across diverse spatial modalities?
• RQ4 What is the effect of the domain adaptation parameter kappa on integration and imputation performance?
• RQ5 Is imputation credible and spatially coherent for held-out genes when imputations are visualized in tissue context?

Key findings

• gimVI achieves a favorable trade-off between dataset integration and biological signal preservation, outperforming competitors on integration metrics across two dataset pairs.
• For imputation, gimVI substantially improves Spearman correlation of imputed genes over state-of-the-art methods, with improvements up to substantial relative gains in several settings.
• Varying the domain adaptation parameter kappa, gimVI can improve imputation over both kappa=0 and kappa=1 for some datasets (kappa* tuning).
• gimVI provides a probabilistic imputation with quantified uncertainty, unlike non-generative approaches which provide point estimates.
• Imputed genes show spatial coherence better than competing methods when examined in tissue context (e.g., Lamp5 example in mSMS).