DeepSpot2Cell

Spatial transcriptomics reveals where genes are expressed within intact tissue, but the assays are costly and, at the popular Visium spot resolution, each measured spot mixes signal from many cells. DeepSpot2Cell tackles both limitations by predicting virtual single-cell spatial gene expression directly from routine hematoxylin-and-eosin (H&E) histology images. Developed by the Rätsch and Koelzer groups at ETH Zurich and released as a preprint in 2025, it extends the DeepSpot lineage from spot-level to single-cell-resolution prediction.

The model reuses frozen pathology foundation models—UNI, H-optimus-0, and Phikon-v2—as feature extractors and layers a DeepSet neural network on top. For each cell it aggregates multi-level visual context: the segmented cell tile, the full Visium spot (~55 µm) containing it, and neighboring spot tiles. Crucially, it is trained with spot-level supervision—the only labels available from standard spatial transcriptomics—yet learns to output expression profiles at individual-cell resolution.

Because the pathology backbones stay frozen, a trained DeepSpot2Cell model can be applied to new H&E slides without retraining, positioning it as an inexpensive way to enrich histology with predicted single-cell expression.

#Key Features

• Single-cell expression from H&E: Predicts virtual single-cell spatial transcriptomic profiles from standard histology, moving beyond spot-level prediction.
• Frozen pathology foundation models: Uses UNI, H-optimus-0, or Phikon-v2 as fixed encoders, inheriting large-scale pathology pretraining without fine-tuning them.
• Multi-level spatial context: A DeepSet head integrates the cell tile, its containing spot, and neighboring spots to inform each prediction.
• Spot-supervised, cell-resolved: Trained only on spot-level labels yet outputs cell-resolution expression, sidestepping the lack of single-cell ground truth.
• Reusable without retraining: Runs on new H&E images out of the box and is released under a permissive MIT license.

#Technical Details

DeepSpot2Cell couples frozen pathology foundation models with a permutation-invariant DeepSet regression head that pools features across the cell, its spot, and neighboring spots. Training draws on paired H&E and spatial transcriptomics data from the HEST-1k resource. Evaluated on 29 cancer tissue samples spanning lung, breast, and pancreatic tumors, the method reports substantial gains in expression-correlation metrics over baselines—improvements of 46%, 65%, and 38% for the three cancer types, respectively. The approach is agnostic to which pathology foundation model supplies the features, allowing users to swap encoders as stronger backbones become available.

#Applications

DeepSpot2Cell is aimed at digital pathology and cancer-research workflows where spatial transcriptomics is unavailable or too expensive to run at scale. From archival or freshly scanned H&E slides, it can generate predicted single-cell expression maps to support tumor microenvironment characterization, cell-state analysis, and hypothesis generation, complementing—rather than replacing—experimental spatial assays. Because inference needs only histology images and released code, it fits readily into existing pathology pipelines.

#Impact

By showing that spot-level supervision is sufficient to train a model that predicts expression at single-cell resolution, DeepSpot2Cell advances the goal of extracting molecular information from ubiquitous H&E images. Its reliance on frozen, interchangeable pathology foundation models and its MIT-licensed release make it a practical, extensible tool for the community. As a preprint awaiting peer review, its reported improvements are based on in-silico evaluation across the tested cancer cohorts.