# A multi-view graph convolutional network framework based on adaptive adjacency matrix and multi-strategy fusion mechanism for identifying spatial domains

## TL;DR

This paper introduces a new deep learning framework for analyzing spatial transcriptomics data to better identify spatial domains in tissues.

## Contribution

The novel STMGAMF model uses an adaptive adjacency matrix and multi-strategy fusion to improve spatial domain identification in noisy ST data.

## Key findings

- STMGAMF outperforms existing methods in spatial domain identification and visualization.
- The model demonstrates strong generalization across multiple spatial transcriptomics datasets.
- The adaptive adjacency matrix enhances the capture of complex spatial structures.

## Abstract

Spatial transcriptomics (ST) addresses the loss of spatial context in single-cell RNA-sequencing by simultaneously capturing gene expression and spatial location information. A critical task of ST is the identification of spatial domains. However, challenges such as high noise levels and data sparsity make the identification process more difficult.

To tackle these challenges, STMGAMF, a multi-view graph convolutional network model that employs an adaptive adjacency matrix and a multi-strategy fusion mechanism is proposed. STMGAMF dynamically adjusts the edge weights to capture complex spatial structures during training by implementing the adaptive adjacency matrix and optimizes the embedded features through the multi-strategy fusion mechanism. STMGAMF is evaluated on multiple ST datasets and outperforms existing algorithms in tasks like spatial domain identification, visualization, and spatial trajectory inference. Its robust performance in spatial domain identification and strong generalization capability position STMGAMF as a valuable tool for unraveling the complexity of tissue structures and underlying biological processes.

Source code is available at Github (https://github.com/Fuyh0628/STMGAMF) and Zenodo (https://zenodo.org/records/15103358).