# Evaluating Integrative Strategies for Incorporating Phenotypic Features in Spatial Transcriptomics

**Authors:** Levin M Moser, Ahmad Kamal Hamid, Esteban Miglietta, Nodar Gogoberidze, Beth A Cimini  
**arXiv:** [2507.22212](https://arxiv.org/abs/2507.22212)  
**Date:** 2025-07-29  
**Summary page:** https://www.summarizepaper.com/en/paper/2507.22212v1/

## Key points

- Spatial transcriptomics (ST) technologies allow analysis of intact biological samples in a spatially informed manner
- Integration of ST with imaging is a current challenge
- Researchers used murine ileum MERFISH dataset to explore the potential of a minimally tuned variational autoencoder (VAE)
- VAE successfully extracted informative low-dimensional representations from cell crops
- Evaluation methods included PERMANOVA, cross-validated classification, and unsupervised Leiden clustering
- VAE-derived latent spaces (LSs) captured meaningful biological variation and improved label recovery for specific cell types
- LS2, trained on morphological input, showed moderate predictive power for select genes
- Combining transcript counts (TC) with LSs through multiplex clustering enhanced cluster homogeneity
- Features derived from CellProfiler underperformed compared to LSs, emphasizing the advantage of learned representations
- VAEs can extract biologically relevant signals from imaging data even under constrained conditions
- VAEs offer promise for multi-modal integration in spatial transcriptomics research

## Summary

Spatial transcriptomics (ST) technologies provide a unique opportunity to analyze intact biological samples in a spatially informed manner and enable integration with other imaging-based modalities. However, the optimal way to leverage spatial context and integrate ST with imaging remains an ongoing challenge. In this study, researchers utilized a murine ileum MERFISH dataset to investigate the potential of a minimally tuned variational autoencoder (VAE) in extracting informative low-dimensional representations from cell crops of spot counts, nuclear stain, membrane stain, or their combinations. The resulting embeddings were evaluated through various methods such as PERMANOVA, cross-validated classification, and unsupervised Leiden clustering. These embeddings were compared with classical image-based feature vectors obtained using CellProfiler. While transcript counts (TC) generally performed better than other feature spaces, the VAE-derived latent spaces (LSs) successfully captured meaningful biological variation and improved label recovery for specific cell types. Particularly noteworthy was LS2, trained solely on morphological input, which exhibited moderate predictive power for select genes in a ridge regression model. Combining TC with LSs through multiplex clustering consistently enhanced cluster homogeneity. This trend persisted even when augmenting only subsets of TC with stain-derived LS2. In contrast, features derived from CellProfiler underperformed relative to LSs, underscoring the advantage of learned representations over manually crafted features. Overall, these findings highlight the capability of VAEs to extract biologically relevant signals from imaging data even under constrained conditions. The study demonstrates that VAEs represent a promising strategy for multi-modal integration in spatial transcriptomics research. The refined analysis sheds light on the importance of leveraging advanced computational techniques to enhance our understanding of complex biological systems at a spatial level.