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

We propose an autoencoder-based framework for denoising networks estimated from Spatial Transcriptomics (ST) data for cell signaling analysis. Our method consists of an unsupervised encoder-decoder framework for denoising the network adjacency matrix and a supervised framework for cell signaling estimation. We validate our denoising component using the Frobenius norm metric for graphs simulated using the Barabasi-Albert (BA) and Erd[o]s-Renyi (ER) models against reconstructions generated using the singular value decomposition (SVD). We then validate the cell signaling estimates generated using the supervised component on real ST data for the Wnt3-Fzd1 and Ephb1-Efnb3 interactions. We report that our framework achieves better adjacency matrix reconstructions for superlinear BA and dense ER graphs and generates cell signaling estimates that are both regionally specific and biologically plausible. An important contribution of this work is the application of neural networks for network-based cell signaling estimation using ST data and the benchmarking of autoencoder versus SVD denoising for different graph models.