Autoencoder Denoising for Network-Based Spatial Transcriptomics Data with Applications for Cell Signaling Estimation

Area of Science:

• Computational Biology
• Bioinformatics
• Systems Biology

Background:

• Spatial Transcriptomics (ST) data offers insights into cellular organization and function.
• Accurate network reconstruction is crucial for cell signaling analysis from ST data.
• Existing denoising methods may not optimally handle the complexity of ST-derived networks.

Purpose of the Study:

• To develop and validate an autoencoder-based framework for denoising networks from ST data.
• To enhance cell signaling estimation using denoised networks.
• To compare autoencoder denoising with Singular Value Decomposition (SVD) for different graph models.

Main Methods:

• An unsupervised encoder-decoder framework was employed for denoising network adjacency matrices.
• A supervised framework was utilized for cell signaling estimation.
• Denoising performance was validated using the Frobenius norm on simulated Barabási-Albert (BA) and Erdős-Rényi (ER) graphs.
• Cell signaling estimates were validated on real ST data for specific ligand-receptor interactions.

Main Results:

• The autoencoder framework demonstrated superior adjacency matrix reconstructions for superlinear BA and dense ER graphs compared to SVD.
• Cell signaling estimates derived from the denoised networks were regionally specific.
• The estimated cell signaling interactions were biologically plausible.
• The study provides a novel application of neural networks for network-based cell signaling analysis in ST data.

Conclusions:

• The proposed autoencoder-based framework effectively denoises networks from ST data, improving cell signaling analysis.
• This approach offers a powerful tool for uncovering cell-cell communication in complex biological tissues.
• Benchmarking highlights the advantages of autoencoder denoising over SVD for specific graph structures in ST data analysis.