Spatial Transcriptomics Prediction from Histology Images at Single-cell Resolution using RedeHist

Yunshan Zhong,Jiaxiang Zhang,Xianwen Ren

DOI: https://doi.org/10.1101/2024.06.17.599464

2024-06-22

Abstract:Spatial transcriptomics (ST) offers substantial promise in elucidating the tissue architecture of biological systems. However, its utility is frequently hindered by constraints such as high costs, time-intensive procedures, and incomplete gene readout. Here we introduce RedeHist, a novel deep learning approach integrating scRNA-seq data to predict ST from histology images at single-cell resolution. Application of RedeHist to both sequencing-based and imaging-based ST data demonstrated its outperformance in high-resolution and accurate prediction, whole-transcriptome gene imputation, and fine-grained cell annotation compared with the state-of-the-art algorithms.

Bioinformatics

What problem does this paper attempt to address?

The main problem that this paper attempts to solve is to use deep - learning methods to predict spatial transcriptomics (ST) at single - cell resolution from histological images. Specifically, the paper introduces a new method named RedeHist. By integrating single - cell RNA sequencing (scRNA - seq) data, this method can predict high - resolution and accurate spatial transcriptomics data from histological images and is superior to the existing state - of - the - art algorithms in gene imputation and fine - grained cell annotation.

### Main problems
1. **High cost and time - consuming**: Traditional spatial transcriptomics techniques, due to their high - cost and time - consuming characteristics, limit their wide use in clinical applications.
2. **Resolution and gene detection range**: Existing spatial transcriptomics techniques have limitations in resolution and gene detection range, especially in the prediction of single - cell resolution and the whole - transcriptome expression profile.
3. **Limitations of existing methods**:
- **Limited number of predicted genes**: Existing methods can only predict several hundred to several thousand highly variable genes and are unable to predict the whole - transcriptome expression profile.
- **Challenges of single - cell resolution**: Existing methods usually output block - level features rather than pixel - level features, making it difficult to achieve single - cell resolution.
- **Dependence on spatial coordinates**: Existing methods need to input spatial coordinates to establish the association between expression and spatial position, which limits their prediction ability on new images.
- **Lack of integration of single - cell RNA sequencing data**: Existing methods do not integrate single - cell RNA sequencing data, resulting in limitations in the accuracy of predicting spatial transcriptomics expression profiles.

### Solutions
RedeHist solves the above problems through the following methods:
1. **Cell segmentation**: Use cell segmentation methods (such as CellPose, VistoSeg, etc.) to identify cell or nuclear regions, enhance the prediction resolution and maximize the use of single - cell information.
2. **U - Net feature extraction**: Use U - Net to extract pixel - level features and combine nuclear masks to generate cell - level latent embeddings.
3. **Deconvolution algorithm**: Inspired by the deconvolution method, assume that the expression of each spot area can be represented as the sum of the corresponding single - cell expressions, and predict cell expressions through scRNA - seq reference data.
4. **Multi - task learning**: Simultaneously generate cell abundance matrices, perform gene imputation and cell annotation to improve the accuracy and resolution of prediction.

### Experimental results
The paper verifies the superior performance of RedeHist through experiments on sequencing - based spatial transcriptomics (such as 10x Visium) and imaging - based spatial transcriptomics (such as Xenium) datasets. The results show that RedeHist is superior to the existing state - of - the - art algorithms (such as iStar and Hist2ST) in prediction accuracy, resolution and gene imputation.

### Conclusion
RedeHist provides an effective tool for predicting spatial transcriptomics data at single - cell resolution from histological images and has important biological and clinical application potentials.