Source-linked AI summary
A joint model of unpaired data from scRNA-seq and spatial transcriptomics for imputing missing gene expression measurements
Romain Lopez, Achille Nazaret, Maxime Langevin, Jules Samaran, Jeffrey Regier, Michael I. Jordan, Nir Yosef
TL;DR
Spatial transcriptomics measures only a selected subset of genes, whereas scRNA-seq can capture more genes but lacks spatial locations, motivating imputation from unpaired data. gimVI integrates both modalities with a deep generative model and outperforms Seurat Anchors, Liger, and CORAL on held-out-gene imputation across real datasets.
Problem
Spatial transcriptomics often measures only a preselected subset of genes, while scRNA-seq can capture genes but does not recover cell locations.
Method
gimVI is a deep generative model that integrates spatial transcriptomics and scRNA-seq through a shared latent representation for imputing missing genes.
Results
gimVI outperforms Seurat Anchors, Liger, and CORAL for held-out-gene imputation, with more than 38% relative improvement in ρ for over half the imputed mSMS genes versus CORAL.
Takeaways & Limitations
The model uses unpaired scRNA-seq information to impute spatially missing genes and provides uncertainty estimates for its predictions.
Takeaways & Limitations
The spatial measurement model depends on assay-specific distributional assumptions, including a Poisson model for nearly perfectly sensitive osmFISH data.
Abstract
from arXiv · showhide
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.
1. The gimVI probabilistic model
gimVI models shared biology across scRNA-seq and spatial measurements with a latent representation, while accounting for modality-specific technical effects and assay-specific gene coverage.
- Shared biology: gimVI represents each cell with a low-dimensional latent variable interpreted as cell type or identity and usable for downstream analysis.The latent representation is shared across the model’s modalities.
- Shared biology: A neural network maps the latent representation and nuisance factors to normalized gene frequencies on the probability simplex.These frequencies support imputation and differential-expression analysis, although this manuscript focuses on imputation.
- Spatial measurements: scRNA-seq captures gene set G, assumed to contain the spatially measured subset G′, enabling modeling of genes absent from spatial assays.The spatial component observes only genes in G′ and re-normalizes measurements for that subset.
- scRNA-seq measurements: The scRNA-seq model uses an overdispersed count distribution, specifically a zero-inflated negative binomial in the experiments.Gene expression levels are modeled as conditionally independent given library size, latent biology, and modality variables.
- Spatial measurements: Spatial measurements use assay-dependent count distributions, with Poisson appropriate for nearly perfectly sensitive osmFISH data.The model uses observed transcript counts from genes included in the assay and assumes little technical variation from size effects in spatial data.
2. Posterior inference
gimVI uses variational inference with partially shared encoders to learn a joint latent space, then samples spatial posteriors to predict genes absent from spatial measurements. Domain-adaptation theory links accurate imputation to both scRNA-seq gene fit and good latent-space mixing.
- Variational inference: Shared encoder layers in the variational distributions help gimVI learn a joint latent space from scRNA-seq and spatial modalities.The encoders use Gaussian variational distributions with diagonal covariance matrices parameterized by neural networks.
- Variational inference: The model optimizes modality-specific variational lower bounds with stochastic backpropagation because exact posterior inference is intractable.The generative model parameters include the neural-network and gene-specific dispersion parameters.
- Missing gene imputation: gimVI imputes a missing spatial gene by sampling its latent posterior and applying the generative neural network to predict counterfactual expression.The latent posterior is conditioned on the observed spatial measurement and modality indicator.
- Domain adaptation: Accurate imputation requires the gene to be well fitted in the scRNA-seq model and the spatial and scRNA-seq samples to mix well in latent space.The latent-space discrepancy is represented by H-divergence, which can be approximated using an adversarial classifier.
3. Performance benchmarks
gimVI is benchmarked for integrating unpaired scRNA-seq and spatial datasets and imputing held-out genes. Across real mouse cortex datasets, it provides a tunable integration trade-off, improves imputation over alternatives, quantifies uncertainty, and produces more spatially coherent patterns.
- 3.1. Integrating cells into a joint latent space: gimVI with κ = 1 achieves the best entropy-of-mixing performance, while its tunable κ balances dataset merging against preservation of biological information.Across both dataset pairs, Liger has the lowest k-NN purity; gimVI’s purity is lower than scVI because it also uses latent variables to impute genes.
- 3.2. Imputing missing genes: More than 38% relative improvement in ρ for over half of mSMS imputed genes is achieved by gimVI with κ = 0 versus CORAL.All gimVI variants significantly improve over state-of-the-art methods across both datasets, while an intermediate κ can outperform κ = 0 and κ = 1 on mSMS.
- 3.2. Imputing missing genes: gimVI’s imputation uncertainty is higher for genes that are harder to predict and nearly absent for predictions on which the model is confident.Uncertainty is estimated from variance across fifty samples from the variational posterior and compared with scRNA-seq linear-regression residuals.
- 3.2. Imputing missing genes: For held-out Lamp5, gimVI produces spatially more coherent expression patterns than competitors on the mSMS spatial dataset.Competitors incorrectly predict high Lamp5 expression in regions such as layer 6.