OctopathAI's £1.5m launch puts a new question to cancer pathology: what does every cell reveal?

On a conventional cancer slide, colour and shape offer clues about disease. The new venture announced by University College London on 2 November proposes to turn those clues into a countable map of cells: not merely where a tumour is, but how densely its cells are arranged, where immune cells gather and how diagnostic stains vary across a specimen. OctopathAI has launched with £1.5 million in seed investment and grant support to develop tools that make such measurements useful to pathologists and cancer researchers. The premise is compelling precisely because the company is not claiming to replace the microscope. It is trying to measure what a human reviewer cannot feasibly count cell by cell.[1]

The technology rests on two models, Kraken and Nemo, built using 14.1 million individually labelled cells spanning 14 cancer types. Kraken reads standard haematoxylin-and-eosin slides, the familiar pink and purple tissue images. Nemo works with immunohistochemistry, in which stains indicate the presence and intensity of selected proteins. According to UCL, the platform performs analysis at single-cell resolution in under two minutes. That processing-time statement describes the platform's claimed capability, not the total turnaround time of a patient's biopsy from collection to clinical decision.[1]

Why a cell count can alter the interpretation of a trial

The most interesting evidence predates the company launch. In one study reported by UCL as published in eBioMedicine, researchers examined tumour-cell density in 414 biopsies from the phase 3 ARISTOTLE rectal cancer trial. The AI-derived measurements identified a subgroup whose outcomes varied depending on the chemoradiotherapy regimen they received, a pattern that had not emerged from the trial's overall result. Another study associated AI-measured immune-cell density with survival and reported improved prognostic accuracy when the measurement was added to conventional clinical variables, including in cohorts from University College London Hospitals and The Cancer Genome Atlas.[1]

Those are potentially important distinctions. A clinical trial can find little average advantage between treatments while still containing groups who respond differently. The difficulty is identifying those groups reliably, proving that the effect is not a chance discovery and showing that changing treatment based on the new measurement benefits future patients. A retrospective or secondary subgroup result is a route to a clinical hypothesis, not by itself proof of a better treatment policy. OctopathAI's work is interesting because it makes the hypothesis measurable at a finer level of tissue detail.[1]

Spatial-omics information helped train the models. This is molecular evidence linked to where molecules sit inside tissue, rather than a chemical inventory detached from its original location. A conventional slide does not directly sequence RNA or measure every protein; the machine learns associations between rich molecular references and patterns in an image. That can make inexpensive, widely used slide formats more informative. But the distinction matters: an image-derived estimate remains an estimate, with calibration and validation requirements for the particular clinical task.[1]

A deployed research platform is not an approved diagnostic

UCL says the technology is operating at five sites, including UCLH, and processing about 200 anonymised slides a day. It is also being used to analyse material from four further clinical trials. These facts demonstrate a working research and deployment footprint beyond a laboratory demo. They do not mean that routine clinical decisions are already authorised to rely on every proposed model output. The company is participating in the UK's Medicines and Healthcare products Regulatory Agency AI Airlock Phase 2, a regulatory sandbox, while working towards approval.[1]

The money also needs careful reading. The announced £1.5 million combines seed investment with matched grant funding; a £500,000 Innovate UK Growth Catalyst grant is unlocked by the seed round. It should not be described as a £1.5 million standalone venture-capital cheque. The capital is intended to pay for the unglamorous parts of clinical translation: evaluation, quality systems, regulatory work and building commercially usable products.[1]

The pressure on pathology explains the interest. UCL cites a severe staffing shortage in histopathology and more than 55,000 people in England in May 2026 who exceeded the four-week cancer diagnosis standard. Yet quicker slide computation does not automatically remove bottlenecks in specimen preparation, specialist review or treatment scheduling. The decisive comparison for hospitals will be whether this specific system reduces work and disagreement without adding errors, and whether its new measurements improve patient outcomes when acted upon.[1]

A useful next milestone would therefore be prospective evidence: independently tested accuracy by cancer type and staining protocol, agreement with expert assessment, and a trial of decisions made with and without the added cell map. Underneath the launch is a significant scientific possibility. Rather than trying to make AI write the final diagnosis, OctopathAI is asking whether routine tissue contains more quantifiable, clinically relevant information than current workflows have the capacity to extract.[1]

Report an error or suggest a correction