IC-VAE: A Novel Deep Learning Framework for Interpreting Multiplexed Tissue Imaging Data.

Abstract

Interpreting protein expression in multiplexed tissue imaging data presents a significant challenge due to the high dimensionality of the resulting images, the variety of intracellular structures, cell shapes resulting from 2-D tissue sectioning, and the presence of technological noise and imaging artifacts. Here, we introduce the Information-Controlled Variational Autoencoder (IC-VAE), a deep generative model designed to tackle this challenge. The contribution of IC-VAE to the VAE framework is the ability to control the shared information among latent subspaces. We use IC-VAE to factorize each cell9s image into its true protein expression, various cellular components, and background noise, while controlling the shared information among some of these components. Compared with other normalization methods, this approach leads to superior results in downstream analysis, such as analyzing the expression of biomarkers, classification for cell types, or visualizing cell clusters using t-SNE/UMAP techniques.