Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Masked Graph Autoencoders with Contrastive Augmentation for Spatially Resolved Transcriptomics Data

Donghai Fang, Fangfang Zhu, Dongting Xie, Wenwen Min

TL;DR

STMGAC tackles spatial transcriptomics clustering and denoising by integrating a masked graph autoencoder with contrastive augmentation. It learns robust latent embeddings through masked reconstruction, a momentum-encoder-based latent supervision signal, and triplet-based anchor learning to sharpen spatial domain separation. Across five datasets from multiple platforms, STMGAC achieves state-of-the-art clustering accuracy and enhanced gene expression denoising, with ablation studies validating the contribution of each component. The approach provides a scalable, plug-in framework for precise tissue structure delineation and more accurate downstream spatial analyses in SRT data.

Abstract

With the rapid advancement of Spatial Resolved Transcriptomics (SRT) technology, it is now possible to comprehensively measure gene transcription while preserving the spatial context of tissues. Spatial domain identification and gene denoising are key objectives in SRT data analysis. We propose a Contrastively Augmented Masked Graph Autoencoder (STMGAC) to learn low-dimensional latent representations for domain identification. In the latent space, persistent signals for representations are obtained through self-distillation to guide self-supervised matching. At the same time, positive and negative anchor pairs are constructed using triplet learning to augment the discriminative ability. We evaluated the performance of STMGAC on five datasets, achieving results superior to those of existing baseline methods. All code and public datasets used in this paper are available at https://github.com/wenwenmin/STMGAC and https://zenodo.org/records/13253801.

Masked Graph Autoencoders with Contrastive Augmentation for Spatially Resolved Transcriptomics Data

TL;DR

STMGAC tackles spatial transcriptomics clustering and denoising by integrating a masked graph autoencoder with contrastive augmentation. It learns robust latent embeddings through masked reconstruction, a momentum-encoder-based latent supervision signal, and triplet-based anchor learning to sharpen spatial domain separation. Across five datasets from multiple platforms, STMGAC achieves state-of-the-art clustering accuracy and enhanced gene expression denoising, with ablation studies validating the contribution of each component. The approach provides a scalable, plug-in framework for precise tissue structure delineation and more accurate downstream spatial analyses in SRT data.

Abstract

With the rapid advancement of Spatial Resolved Transcriptomics (SRT) technology, it is now possible to comprehensively measure gene transcription while preserving the spatial context of tissues. Spatial domain identification and gene denoising are key objectives in SRT data analysis. We propose a Contrastively Augmented Masked Graph Autoencoder (STMGAC) to learn low-dimensional latent representations for domain identification. In the latent space, persistent signals for representations are obtained through self-distillation to guide self-supervised matching. At the same time, positive and negative anchor pairs are constructed using triplet learning to augment the discriminative ability. We evaluated the performance of STMGAC on five datasets, achieving results superior to those of existing baseline methods. All code and public datasets used in this paper are available at https://github.com/wenwenmin/STMGAC and https://zenodo.org/records/13253801.
Paper Structure (32 sections, 13 equations, 3 figures, 3 tables)

This paper contains 32 sections, 13 equations, 3 figures, 3 tables.

Table of Contents

  1. Introduction
  2. PROPOSED METHODS
  3. Overview of the proposed STMGAC
  4. Data preprocessing and spatial graph construction
  5. Data augmentation with spot masking
  6. Latent representation learning with masked reconstruction
  7. Graph encoding
  8. Latent representation predicting
  9. Feature decoding
  10. Reconstruction loss in the raw data space
  11. Latent representation matching for robustness enhancement
  12. Momentum graph encoding
  13. Matching loss in the latent representation space
  14. Spot triplet learning with positive and negative anchors
  15. Selecting positive and negative anchor pairs
  16. ...and 17 more sections

Figures (3)

  • Figure 1: Overview of STMGAC. (A) In data processing, raw gene expression is split into a mask matrix and a visible matrix. (B) Reconstruct raw data using a mask GAE, obtain latent space supervision signals through a momentum encoder for latent representation matching, and then utilize selected anchor spots for triplet learning. (C) The learned latent representations from STMGAC will be utilized in downstream task analysis, including clustering and visualization. Additionally, the reconstructed gene expression is considered as the denoised outcome.
  • Figure 2: STMGAC enables the identification of tissue structures. (A) Schematic of DLPFC data. (B) The evaluation of STMGAC and existing methods on the DLPFC dataset was conducted based on ARI and ACC. (C) Manual annotation for slice 151675. (D) We conduct a comparison of spatial domain identification on slice 151675. (E) The UMAP plot of slice embeddings colored cortical layers. (F) Manual annotation for HM. (G) Spatial domain identification on HM. (H) Here are the top 10 significant GO:BP terms for Cluster 2 (Lymphoid) and Cluster 3 (Melanoma). (I) Manual annotation for MBA. (J) Spatial domain identification on MBA.
  • Figure 3: STMGAC denoises gene expressions. (A) Manual annotation for BRCA. (B) Spatial domain identification using the STMGAC, GraphST, and STAGATE methods. (C) Bar chart of ARI and ACC metrics for various methods on BRCA. (D) Raw spatial expression and STMGAC denoised expression of IGLC2 and AQP1 genes. (E) Volcano plot of differentially expressed genes between Cluster 4 (health) and Cluster 5 (tumor edge). (F) Manual annotation for ME9.5. (G) Spatial domain identification. (H) Bar chart of ARI and ACC metrics for various methods on ME. (I) Spatial domains of major tissues identified by STMGAC. (J) Raw spatial expression and denoised expression by STMGAC. (K) Violin plots of the denoised expression of domain-specific marker genes by STMGAC.