Evaluating integrative strategies for incorporating phenotypic features in spatial transcriptomics.

Area of Science:

• Spatial transcriptomics
• Computational biology
• Bioimaging

Background:

• Spatial transcriptomics (ST) enables spatially informed interrogation of biological samples and integration with imaging modalities.
• Exploiting spatial context and integrating ST with morphological imaging data remains a challenge, especially under experimental constraints like limited data and class imbalance.

Purpose of the Study:

• To evaluate variational autoencoders (VAEs) for extracting informative low-dimensional representations from cell crops in spatial transcriptomics data.
• To assess the utility of VAE-derived latent spaces (LSs) for biological variation capture and multi-modal integration, particularly under constrained experimental conditions.

Main Methods:

• Utilized a murine ileum Multiplexed Error-Robust Fluorescence In Situ Hybridisation (MERFISH) dataset.
• Extracted low-dimensional representations using a minimally tuned VAE from spot counts, nuclear/membrane stains, or combinations.
• Assessed VAE embeddings via PERMANOVA, cross-validated classification, unsupervised clustering, and compared them to CellProfiler features.

Main Results:

• VAE-derived latent spaces captured meaningful biological variation and improved cell type label recovery.
• Morphology-trained VAEs showed predictive power for gene expression.
• Combining transcript counts with VAE latent spaces enhanced clustering outcomes, improving homogeneity or label recovery depending on the clustering approach.

Conclusions:

• Variational autoencoders can extract biologically meaningful signals from imaging data, even under constrained conditions.
• Learned representations from VAEs outperform handcrafted features from tools like CellProfiler for spatial transcriptomics analysis.
• VAEs offer a promising strategy for multi-modal integration in spatial transcriptomics.