TL;DR
POLARIS is a new computational method that tackles a major challenge in spatial transcriptomics: deconvoluting cell types within tissue spots and identifying genes that are specifically expressed in different tissue layers. It uniquely leverages both gene expression data and histology images to achieve high accuracy, even predicting cell composition from images alone.
What Problem Does This Paper Address?
Spatial transcriptomics (ST) technologies allow us to map gene expression within the physical context of tissues. However, a key limitation is that most ST platforms measure gene expression in "spots" that can contain many cells, making it difficult to know the exact cell type composition of each spot. This ambiguity can lead to inaccurate downstream analyses. While some ST methods use histology images (like H&E stains) to provide anatomical context, these images are often underutilized. Furthermore, gene expression can vary not just between cell types but also between different layers or regions within a tissue, a phenomenon known as layer-specific differential expression (LDE). Existing methods struggle to disentangle these factors and accurately identify LDE genes, especially when cell composition varies across layers. This paper introduces POLARIS to address these challenges by integrating gene expression, histology images, and layer information for more robust cell deconvolution and LDE gene identification.
Methods
POLARIS is a probabilistic-based inference method that models gene expression counts using a negative binomial distribution. Its core innovations include:
- Layer-Specific Modeling: Unlike many methods that assume gene expression profiles are invariant across a tissue, POLARIS explicitly models layer-specific gene expression patterns. This is achieved by introducing a layer-specific location parameter that allows gene expression profiles to vary across different tissue layers.
- Histology Image Integration: POLARIS can optionally incorporate co-registered histology images. It uses a Masked Autoencoder (MAE) to extract features from spot images and their neighborhoods. These features are then fed into a POLARIS image network, which can predict cell composition directly from images. This allows for "super-resolution" inference of cell composition in areas not covered by ST spots.
- Cell Type Deconvolution: POLARIS infers cell type composition at the spot level by combining information from a single-cell RNA-seq reference and the ST data.
- LDE Gene Identification: By leveraging the inferred layer-specific gene expression profiles and accounting for cell composition, POLARIS identifies layer-wise differentially expressed (LDE) genes using permutation tests.
Datasets
POLARIS was evaluated on several datasets:
- Simulated data: Used to benchmark deconvolution accuracy and LDE gene identification under controlled conditions.
- Mouse primary visual cortex (VISp) STARmap data: Single-cell resolution ST data used for deconvolution evaluation.
- Developing human heart ISS data: Single-cell resolution ST data with a DAPI-stained histology image, used to assess deconvolution accuracy with limited genes and image input.
- Mouse primary somatosensory cortex (SSp) 10x Visium data: Spot-level ST data with H&E-stained histology images, used to evaluate super-resolution cell composition inference from images alone.
- HER2+ breast cancer ST data: Spot-level ST data from multiple patients, with pathologist annotations, used to assess LDE gene identification and image-based inference.
Key Insights and Ideas
- Histology Images are Powerful: The paper strongly demonstrates that histology images, even single-channel ones like DAPI, contain significant information for inferring cell composition, a capability largely untapped by previous ST analysis tools.
- Layer Matters: Gene expression variation is not solely due to cell type differences; tissue layers play a crucial role. POLARIS's ability to model this layer-specific variation is key to its improved performance.
- Unsupervised Image-Based Inference: POLARIS's image network can infer cell composition without requiring manual annotations of the histology image, making it applicable to a wider range of datasets.
- Bridging the Resolution Gap: POLARIS can provide higher-resolution cell composition maps than the original ST spots by leveraging image information.
Key Results
- High Deconvolution Accuracy: POLARIS achieved state-of-the-art or near state-of-the-art deconvolution accuracy across multiple tissues, outperforming several leading methods in simulations and real datasets.
- Accurate Image-Based Prediction: POLARIS accurately predicted cell composition solely from histology images, even in areas without gene expression measurements.
- Meaningful LDE Genes: POLARIS successfully identified biologically relevant LDE genes in developing human heart and breast cancer samples, revealing complex tissue heterogeneity. For instance, it identified genes like TCF21 in the heart and ERBB2 in breast cancer, linking them to specific biological functions and disease characteristics.
- Super-Resolution Inference: POLARIS demonstrated the ability to generate super-resolution cell composition maps from histology images, revealing finer tissue structures.
Limitations and Future Work
- Layer Shift Assumption: POLARIS assumes that the layer-specific shift parameter is shared across all cell types. The authors acknowledge this as a simplification and a potential limitation, noting that a cell-type-specific layer shift might be challenging to estimate reliably due to model identifiability issues.
- Histology Image Quality: The effectiveness of the image-based inference relies on the quality of the co-registered histology images.
- Future Work: The authors suggest that future work could explore more sophisticated image integration models and potentially incorporate other modalities. The ability to perform unsupervised registration of histology images to anatomical/functional regions is highlighted as a significant future application.
Summary updated