Deciphering tissue heterogeneity from spatially resolved transcriptomics by the autoencoder-assisted graph...

Area of Science:

• Computational Biology
• Genomics
• Bioinformatics

Background:

• Spatially resolved transcriptomics (SRT) enables detailed tissue organization studies.
• Integrating spatial context within and across samples is a significant computational challenge.
• Existing models struggle to capture the full complexity of spatial transcriptomic data.

Purpose of the Study:

• To develop a novel ensemble model for accurate spatial domain identification.
• To improve the representation learning for complex and heterogeneous tissue data.
• To leverage both autoencoder and graph convolutional neural network strengths.

Main Methods:

• Developed AE-GCN (autoencoder-assisted graph convolutional neural network), an ensemble model.
• Integrated autoencoder (AE) and graph convolutional neural network (GCN) components.
• Employed a clustering-aware contrastive mechanism for unified deep learning.

Main Results:

• AE-GCN demonstrated effectiveness in spatial domain identification and data denoising across multiple SRT platforms (ST, 10x Visium, Slide-seqV2).
• In cancer datasets, AE-GCN identified disease-related spatial domains with greater heterogeneity than histological annotations.
• The model facilitated the discovery of novel differentially expressed genes with high prognostic relevance.

Conclusions:

• AE-GCN successfully integrates AE and GCN for robust spatial transcriptomic data analysis.
• The model's ability to identify fine-grained spatial domains and prognostic genes highlights its potential in biological discovery.
• AE-GCN offers a powerful tool for unveiling complex spatial patterns in heterogeneous tissues.