Phoenix

Phoenix is a generative foundation model that predicts pan-cancer, spatially-resolved single-cell gene expression directly from routine hematoxylin and eosin (H&E) histology images. Introduced in the bioRxiv preprint "Pan-cancer virtual spatial transcriptomics from routine histology with Phoenix" (posted April 2026) by Manuel Tran and colleagues at the Computational Health Center of Helmholtz Munich, with the lead author also affiliated with the Technical University of Munich, the model addresses a central bottleneck in spatial biology: real spatial transcriptomics assays are expensive, slow, and not part of standard clinical workflows, whereas H&E slides are produced for nearly every cancer patient.

Rather than running a wet-lab spatial assay, Phoenix infers the molecular profile of each cell from morphology alone, producing "virtual" spatial transcriptomics maps. The model is pretrained once and then applied to new patient slides without any per-slide retraining or fine-tuning, positioning it as a foundation model for inferring molecular state from tissue appearance. This places Phoenix in the rapidly growing intersection of computational pathology and spatial omics, alongside efforts to bridge cheap, ubiquitous histology with costly molecular readouts.

#Key Features

• Histology-to-expression prediction: Maps routine H&E whole-slide images to spatially-resolved, single-cell gene expression, turning standard pathology slides into virtual spatial transcriptomics data.
• Latent flow-matching generative model: Uses a flow-matching objective in a learned latent space to model the distribution of single-cell expression conditioned on local tissue morphology, rather than regressing a single point estimate.
• Pretrain-once, apply-anywhere: After pretraining, Phoenix generalizes to new patient slides without re-training, supporting use as a reusable foundation model.
• Pan-cancer scope: Evaluated across more than 10,000 patients spanning multiple cancer types and a murine model, demonstrating breadth beyond a single tissue or cohort.

#Technical Details

Phoenix is built as a latent flow-matching generative model: histology features are encoded into a latent representation, and a flow-matching network learns to transform a base distribution into single-cell gene-expression profiles conditioned on the local morphological context. This generative formulation lets the model capture cell-to-cell variability in expression rather than collapsing predictions to an average. The preprint reports evaluation across more than 10,000 patients, including head-and-neck cancer (n=763), breast cancer (n=84), ovarian cancer (n=157), sarcoma (802 tissue-microarray cores), and a murine pancreatic cancer cohort, testing generalization across organs, cancer types, and species. The work is released under a CC BY-NC 4.0 license.

#Applications

Phoenix is aimed at researchers and translational pathology teams who want molecular, spatially-resolved readouts without commissioning costly spatial transcriptomics experiments. Because it operates on routine H&E slides, it can in principle be applied retrospectively to large archived cohorts and biobanks, enabling cell-type mapping, tumor-microenvironment characterization, and hypothesis generation across cancer types. Such virtual spatial transcriptomics could support biomarker discovery and large-scale molecular profiling where physical spatial assays are impractical.

#Impact

By demonstrating pan-cancer virtual spatial transcriptomics from histology alone across more than 10,000 patients and multiple species, Phoenix contributes to the case that ubiquitous H&E imaging can serve as a proxy for expensive molecular assays. As a preprint, its results await peer review and independent validation. Notably, at the time of writing there is no public code repository, released weights, or HuggingFace artifact, and although the preprint is licensed CC BY-NC 4.0, the licensing of any future model weights is unspecified, which currently limits independent reproduction and downstream reuse.