Abstract
Spatial transcriptomics (ST) has revolutionized our ability to understand biological processes by providing spatially resolved transcriptional profiles. However, sequencing-based ST technologies often lack single-cell resolution, resulting in data "spots" that contain mixtures of multiple cell types. This lack of resolution complicates downstream analysis. We present POLARIS (Probabilistic-based cell cOmposition inference with LAyer infoRmatIon Strategy), a versatile framework that integrates gene expression data with high-resolution histology images. POLARIS performs accurate cell type deconvolution, identifies layer-wise differentially expressed (LDE) genes, and enables cell composition inference directly from images. Our results across multiple tissues demonstrate that POLARIS achieves high accuracy and provides biologically meaningful insights into anatomical and functional tissue structures.
Introduction
Spatial transcriptomics (ST) allows researchers to map gene activity within the physical context of tissue architecture. Despite its potential, popular platforms like 10x Genomics Visium face a significant resolution challenge: each capture spot (typically 55–100 $\mu$m in diameter) represents a "soup" of diverse cell types rather than a single cell. To interpret this data, researchers rely on deconvolution methods to estimate cell type proportions.
Existing tools, however, suffer from two primary limitations. First, they often ignore the "layer" effect—the fact that a cell's transcriptional profile changes based on its anatomical neighborhood. Second, they under-utilize the high-resolution histology images (such as H&E stains) that accompany ST data. POLARIS addresses these gaps by combining probabilistic modeling of gene counts with deep learning-based image analysis.
The POLARIS Framework
POLARIS is built upon three methodological pillars designed to bridge the gap between image morphology and molecular signatures:
- Probabilistic Modeling: The framework utilizes a Negative Binomial distribution to model gene counts, incorporating single-cell RNA-seq (scRNA-seq) references to guide the deconvolution process.
- Layer-Specific Parameters: Unlike standard models that assume cell types are transcriptomically static, POLARIS introduces a "shift parameter" for different anatomical layers. This allows the model to identify Layer-wise Differentially Expressed (LDE) genes—genes whose expression is modulated by the cell's specific location within a tissue structure.
- Deep Learning Image Network: POLARIS employs a Masked Autoencoder (MAE) to extract features from histology images. By learning the relationship between visual morphology and cellular composition, the model can perform "super-resolution" inference, predicting cell types in the gaps between sequencing spots or on entirely new tissue sections where sequencing was not performed.
Experimental Validation
The researchers validated POLARIS across several complex biological systems:
- Simulated Data: Established a ground truth to benchmark deconvolution accuracy.
- Mouse Brain (VISp & SSp): Used STARmap and 10x Visium data to test the recovery of cortical layers.
- Developing Human Heart: Applied In Situ Sequencing (ISS) data to evaluate performance on limited gene sets (65 genes).
- HER2+ Breast Cancer: Utilized ST platform data to distinguish between invasive and in situ cancerous regions.
Key Insights and Results
Superior Deconvolution Accuracy
In comparative tests against state-of-the-art methods such as RCTD, CARD, and DestVI, POLARIS consistently achieved the lowest Root Mean Square Error (RMSE). This indicates that its predicted cell mixtures are more aligned with biological reality than existing probabilistic or regression-based models.
Decoupling Composition from Expression
A common pitfall in ST analysis is "composition bias," where a gene appears spatially variable simply because a specific cell type is more abundant in one area. POLARIS successfully distinguishes between changes in cell density and true changes in cellular behavior (LDE), allowing for a more nuanced understanding of tissue heterogeneity.
Image-Based Composition Inference
One of the most significant findings is the model's ability to generalize across slides. After training, the POLARIS image network can predict cell composition on a new slide using only the histology image. In mouse brain tissues, this image-only approach achieved a Pearson correlation of 0.8 for certain neuron types, proving that morphological features are deeply coupled with transcriptional identity.
Biological Discovery
POLARIS identified TCF21 as a key LDE gene in fibroblast-like cells within the human heart, a finding supported by literature on cardiac development. In breast cancer samples, it accurately mapped the upregulation of ERBB2 (HER2) within specific pathological regions, demonstrating its potential utility in clinical diagnostics and pathology.
Discussion and Limitations
While POLARIS represents a significant advancement, it requires "layer information" as an input, which often necessitates a pre-processing step using spatial clustering tools. Furthermore, the accuracy of image-only predictions is sensitive to the quality and consistency of H&E or DAPI staining. Future developments aim to automate layer identification within the pipeline and extend the model's application to complex disease states like neurodegeneration, where subtle layer-specific changes are critical to understanding disease progression.
Summary updated