Our Technology
We learn a visual language of tissue directly from routine H&E slides, with no manual labels required, then map slides into reproducible histomorphological phenotype clusters (HPCs) that correlate with tumor biology and clinical outcomes.
How it works
- Tile & normalize: Whole‑slide images are split into fixed tiles and stain‑normalized to reduce site and scanner variation.
- Self‑supervised learning: A backbone model (e.g., Barlow Twins) learns color/scale‑robust representations of tissue without annotations.
- Phenotype discovery: A nearest‑neighbor graph over tile embeddings is clustered using our technology to yield distinct, interpretable HPCs.
- Slide & patient profiles: Each slide/patient is represented as a composition of HPCs, enabling interpretable models (e.g., logistic/Cox), our proprietary large language model, and multi‑omic associations.
Why it matters
- Interpretable phenotypes: HPCs capture tumour, stromal, immune and many other recognisable morphologies across all pathology.
- Prognostic signal: Phenotype compositions correlate with overall and recurrence‑free survival with interpretable hazard contributions.
- Biology‑aware: HPCs have demonstrated links to immune, proliferative, and stromal transcriptomic signatures and cell‑type enrichments.
- Generalisable: Unlabeled learning and stain‑robust embeddings support cross‑site reproducibility and multi‑cancer analyses.
From slides to insight
Discovery first, labels optional.
By learning directly from image data, our models surface phenotypes that mirror recognized growth patterns (e.g., lepidic, acinar, solid) and immune states, while also revealing mixed or transitional morphologies that are hard to annotate at scale. These phenotype profiles can feed simple linear models (logistic/Cox) for classification and risk, or be associated with RNA‑seq immune signatures and cell‑type densities for mechanistic context.
Label‑free learning
Scales across cohorts without manual region labels while remaining stain/scale robust.
Interpretable clustering
Leiden communities over tile embeddings yield discrete, reviewable phenotypes.
Clinical linkage
Phenotype mixtures align with survival risk and immune contexture across cancer types.