Super-resolved spatial transcriptomics by deep data fusion

Ludvig Bergenstråhle, Bryan He, Joseph Bergenstråhle, Alma Andersson, Joakim Lundeberg, James Zou, Jonas MaaskolabioRxiv 2020 · DOI

TL;DR

This paper introduces a novel deep learning method that fuses low-resolution spatial transcriptomics data with high-resolution histology images to generate super-resolved, full-transcriptome gene expression maps. This approach allows for unprecedented detail in understanding cellular heterogeneity within tissues and even enables predicting gene expression solely from histology, opening doors for cost-effective, large-scale spatial transcriptomics.

What Problem Does This Paper Address?

Understanding the spatial organization of gene expression within tissues is crucial for deciphering cellular function, development, and disease. Current spatial transcriptomics techniques face a trade-off: high-resolution methods can only profile a limited number of genes, while methods that capture a broader transcriptome (like in situ RNA capturing) suffer from lower resolution. This limits the ability to study complex, transcriptome-wide interactions at a fine anatomical scale. The paper aims to bridge this gap by creating a method that leverages the complementary strengths of histology images and gene expression data to achieve high-resolution, full-transcriptome insights.

Methods

The core of the method is a deep generative model that treats spatial gene expression and histology as observable effects of an underlying "latent tissue state." This state is modeled across multiple resolutions to capture both global and local anatomical features. The model employs a variational autoencoder (VAE) framework, where a recognition network learns to map histology images to the parameters of the latent state. Crucially, this allows for efficient inference and scalability to large datasets. The generator network then maps this latent state to both high-resolution gene expression and histology. The model is trained by minimizing the Kullback-Leibler divergence between the inferred posterior and the true posterior of the latent variables, effectively learning to "denoise" and "upscale" the gene expression data using the histology as a guide.

Datasets

The method was evaluated on two primary datasets:

  • Mouse Olfactory Bulb: A dataset of 12 sections used for in-sample and out-of-sample performance evaluation.
  • Human Breast Cancer (Ductal Carcinoma In Situ - DCIS): A dataset used to study spatial dynamics within a tumor.

Additionally, the inferred expression maps were compared against in situ hybridization data from the Allen Mouse Brain Atlas for validation.

Key Insights and Ideas

  • Histology as a powerful guide: The paper demonstrates that high-resolution histological images contain rich information about spatial gene expression patterns, even when the gene expression data itself is low-resolution.
  • Bridging the resolution gap: By fusing these two modalities, the method can generate gene expression maps at the resolution of the histology image, far exceeding the resolution of the original transcriptomics data.
  • "In silico" spatial transcriptomics: The method can predict spatial gene expression from histology alone, enabling "image-based in silico spatial transcriptomics" (ISST). This is a significant advancement for studying tissues where direct transcriptomic measurements are not feasible or cost-prohibitive.
  • Unraveling fine-grained heterogeneity: The super-resolved maps reveal transcriptional heterogeneity within micrometer-scale anatomical features, which are often blurred or indistinguishable in raw spatial transcriptomics data.

Key Results

  • Improved imputation: The method achieved a 24% lower median root-mean-square error (RMSE) for imputing missing expression data compared to pixel-wise interpolation.
  • Accurate out-of-sample prediction: The model could accurately predict expression patterns for entire sections not included in the training set, with performance approaching in-sample accuracy as more training data was used.
  • Validation with in situ hybridization: Inferred expression patterns closely matched reference in situ hybridization data, with specific genes like Ntng1 and Dusp14 being accurately replicated in their respective cell layers.
  • Discovery of spatial gradients: The method identified spatially resolved gene expression gradients in a breast cancer lesion, highlighting genes upregulated at the tumor border and providing insights into potential therapeutic targets.

Limitations and Future Work

  • Scalability for universal models: While the current method is scalable, the authors envision future work to enable ISST on a grander scale by developing universal models that can predict spatial gene expression across diverse tissues, given a sufficiently large and diverse training set.
  • Economic viability: ISST could make large-scale spatial transcriptomics projects more economically viable, allowing researchers to unlock spatial gene expression information from vast existing histology image databases.
  • Experimental data verification: The authors suggest ISST could also be used to verify the integrity of data from experimental spatial transcriptomics methods.

Summary updated

Abstract

In situ RNA capturing has made it possible to record histology and spatial gene expression from the same tissue section. Here, we introduce a method that combines data from both modalities to infer super-resolved full-transcriptome expression maps. Our method unravels transcriptional heterogeneity in micrometer-scale anatomical features and enables image-based in silico spatial transcriptomics without hybridization or sequencing.
BibTeX
@misc{bergenstrahle2020super,
  author = {Ludvig Bergenstråhle and Bryan He and Joseph Bergenstråhle and Alma Andersson and Joakim Lundeberg and James Zou and Jonas Maaskola},
  title = {{Super-resolved spatial transcriptomics by deep data fusion}},
  year = {2020},
  doi = {10.1101/2020.02.28.963413},
  howpublished = {bioRxiv},
}

Loading related content...

Figure