Sophont's pan-cancer model that predicts spatially resolved gene and pathway activity from routine H&E whole-slide images, for research use.
No providers recorded yet. Browse all providers
Holon-S1 is a pan-cancer model from Sophont that predicts spatially resolved gene expression, and from it the activity of signalling pathways and biological programs, directly from routine hematoxylin and eosin (H&E) whole-slide images. It is described in a technical report by Hassan Muhammad, Benjamin Warner, Paul S. Scotti and Tanishq Mathew Abraham, released on 20 October 2026.
The problem it addresses is a gap between two views of the same tissue. An H&E slide is cheap, archived for nearly every cancer patient, and rich in morphology, but it says nothing directly about which genes are active. Spatial transcriptomics measures exactly that while keeping tissue location, but it is expensive, technically demanding and rarely run in routine care. Pathology foundation models turn slides into powerful embeddings, but the biology those embeddings encode stays opaque. Holon-S1 is framed as a translator from morphology into a readable molecular map, so that a researcher sees which genes are predicted to be active and where, rather than a risk score with no stated mechanism.
A second idea runs through the design: not every gene leaves a visible trace in tissue morphology, so a model that claims the whole transcriptome will predict some genes well and many poorly. Holon-S1 therefore ships with a confidence-filtered gene "vocabulary" per organ, limited to genes whose predictions held up against paired spatial transcriptomics data.
The technical report does not disclose the network architecture, the image encoder, the parameter count, or the composition and size of the training corpus beyond its platform mix and pan-cancer scope. Validation on held-out slides paired with spatial transcriptomics found that accuracy tracked the generation of the measuring platform: predictions agreed more closely with newer assays, and Atera samples gave a higher median per-gene Pearson correlation than any Xenium sample. On Atera whole-transcriptome slides, Holon-S1 achieved a higher median per-gene correlation than DeepSpot-M at matched regions. On the Human Tumor Atlas Network breast Visium cohort it exceeded the published Path2Space figure on the quality-curated subset, performed comparably on slides Path2Space had excluded, and showed no significant difference between FFPE and fresh-frozen tissue (p = 0.184). Evaluations covered 16 organ types and more than 15 scanner models; on prostate biopsies rescanned on three scanners, its predicted expression distributions were more consistent than DeepSpot-M's in nine of ten specimens. Against multiplex immunofluorescence on the OrionCRC dataset, it separated B-cell- from T-cell-dominant regions with 88.5% balanced accuracy and an AUROC of 0.943. All benchmarks are reported by the developers and have not been independently reproduced or peer reviewed.
Holon-S1 is aimed at translational and biopharma research on existing slide archives. In a retrospective osteosarcoma case built on one patient's publicly shared data, it recovered a fibroblast program including FAP, the target of the therapy the patient received, elevated MDM2, and increased immune infiltration after checkpoint inhibition, and it predicted B-cell markers absent from the patient's targeted Xenium panel. In triple-negative breast cancer, the spatial gradient of predicted EREG was associated with pathologic complete response to neoadjuvant chemotherapy in the IMPRESS discovery cohort and the held-out TransNEO cohort. In non-metastatic colorectal cancer, segregation of proliferation and stroma/EMT programs was associated with better disease-specific survival in TCGA and in the held-out SurGen cohort. The model is for research use only and is not approved for diagnosis.
Holon-S1 joins a fast-growing group of histology-to-expression models, including Path2Space and Path2Omics, and its distinctive contribution is methodological candour about which genes morphology can actually recover. The per-organ vocabulary and the observation that older assays may understate model accuracy are useful framing for anyone comparing virtual transcriptomics methods. Its reach is limited by access: unlike Sophont's openly released CortexMAE, Holon-S1 has no public weights, code or training data, and is offered to outside groups through research collaborations, so outside groups cannot yet run the independent evaluations its own report invites.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.