stEnTrans: Transformer-based deep learning for spatial transcriptomics enhancement

TL;DR

This work tackles the challenge of incomplete and low-resolution spatial transcriptomics data by introducing stEnTrans, a Transformer-based framework that performs self-supervised pretraining on downsampled data and subsequent enhancement to high-resolution gene expression maps using only expression data and spatial coordinates. The method uses patch-based embedding and absolute positional encoding within a Transformer encoder to predict high-resolution expression and fill unmeasured regions, with a residual fusion that preserves original signals. Across six real datasets and additional simulations, stEnTrans achieves superior interpolation accuracy and reveals biologically meaningful spatial patterns and GO:BP enrichment, without relying on histology images. The approach advances spatial gene expression analysis by improving resolution, enabling more precise pattern discovery and downstream pathway insights, while future work aims to reach sub-spot resolution.

Abstract

The spatial location of cells within tissues and organs is crucial for the manifestation of their specific functions.Spatial transcriptomics technology enables comprehensive measurement of the gene expression patterns in tissues while retaining spatial information. However, current popular spatial transcriptomics techniques either have shallow sequencing depth or low resolution. We present stEnTrans, a deep learning method based on Transformer architecture that provides comprehensive predictions for gene expression in unmeasured areas or unexpectedly lost areas and enhances gene expression in original and inputed spots. Utilizing a self-supervised learning approach, stEnTrans establishes proxy tasks on gene expression profile without requiring additional data, mining intrinsic features of the tissues as supervisory information. We evaluate stEnTrans on six datasets and the results indicate superior performance in enhancing spots resolution and predicting gene expression in unmeasured areas compared to other deep learning and traditional interpolation methods. Additionally, Our method also can help the discovery of spatial patterns in Spatial Transcriptomics and enrich to more biologically significant pathways. Our source code is available at https://github.com/shuailinxue/stEnTrans.

Figures (4)

• Figure 1: The network architecture of stEnTrans. (A) The schematic diagrams of down-sampling and imputation on honeycomb-based (left) and matrix-based (right) data. Down-sample: From left to right and top to bottom, alternately extract every spots and Extracted spots are non-adjacent to each other in the original gene expression profiles. Impute: Interpolate between adjacent spots. There are three types of adjacency: left-right, top-bottom and diagonal adjacency. (B) The self-supervised learning process of a network architecture. Pretraining the stEnTrans using downsampled gene expression as input and original gene pression as labels. After training the stEnTrans, inputting the original gene expression to obtain the high-resolution gene expression. (C) Details of network architecture. The trainable components mainly consist of Transformer Encoder modules.
• Figure 2: The experiments indicate that stEnTrans exhibits superior interpolation performance. (A) and (B) Applying stEnTrans, DIST, NEDI, Linear, Cubic and NN to downsampled simulated data for predicting the ground truth of gene expression. This gene ZNF703 comes from 10X Visium IDC data. (B) A series of gene expression profiles are obtained through a consistent process with (A). This gene MUC1 also originates from 10X Visium IDC data. (C) and (D) apply stEnTrans and DIST to the ground truth of ZNF703 gene and MUC1 gene, resulting in HR gene expression maps. Due to inferior interpolation performance of other methods, we only compare with DIST. The gene-wise minimum and maximum of the ground truth map the color ranges consistently.
• Figure 3: PCCs evaluation between ground truth and each imputed expression in six ST datasets. The results indicate that stEnTrans exhibits outstanding precision in interpolation, as demonstrated through experiments on six datasets. (A) The human melanoma ST data (mel1 rep1). (B) Simulation created from STARmap mouse placenta. (C) Simulation created from Stereo-seq adult mouse hemi-brain. (D) Human breast cancer spatial transcriptomics data. (E) Visium mouse sagittal posterior brain. (F) Visium human IDC data. Gene-wise Pearson correlation correlations between ground truth and imputed expression using stEnTrans, DIST, Linear, Cubic, NN, NEDI on that six datasets. Boxes represent the middle 50% range of the data,which show 25th, 50th and 75th percentiles. Whiskers represent the overall distribution range of the data, providing information about the data's dispersion and potential outliers.
• Figure 4: stEnTrans can help find spatial patterns for ST data and enrich more biologically significant pathways. (A) The original histopathological annotations of hematoxylin- and eosin-stained tissue image, where black represents melanoma, red represents stroma, and yellow represents lymphatic tissue. (B) stEnTrans has endowed disease-related genes with distinct spatial patterns. These genes exhibit higher rankings in the imputed data, whereas they have lower rankings in the original data. We provided the rankings of genes at the top of each gene expression profile. (C) The number of genes and significant GO:BP terms in three mainly families of original data and imputed data, where orange represents original data and blue represents imputed data. (D) Top 10 significant GO:BP terms were only identified in the imputed data. above corresponds to the melanoma family, and below corresponds to the lymphoid family.