FarrSight® draws on knowledge built up from 25 billion data points
Portfolio companyUK company Concr’s FarrSight® platform made its predictions without seeing how patients fared. Researchers at University Health Network in Toronto then scored them against the outcomes of a randomised clinical trial.
An artificial intelligence platform that uses Bayesian methods adapted from cosmology to build ‘digital twins’ of cancer patients has accurately predicted which people with advanced pancreatic cancer were likely to have their disease brought under control by first-line chemotherapy, in an independent, blinded test run by researchers in Toronto.
The platform, FarrSight®, developed by Cambridge Enterprise Ventures portfolio company Concr, was applied to patients from PASS-01, a randomised trial comparing the two standard chemotherapy regimens for metastatic pancreatic cancer. Concr made its predictions without access to any patient outcomes.
The PASS-01 Challenge Team at University Health Network (UHN) then compared them with what actually happened. FarrSight® was the first external model to be evaluated in the PASS-01 Challenge. Of the 160 patients in the trial, 115 could be assessed. FarrSight® scored 74% on a standard accuracy measure (AUC, where 50% is no better than chance), rising to 78% among patients given the more intensive regimen, modified FOLFIRINOX. Patients it predicted would see their disease controlled, with tumours shrinking or holding stable, did so in 76% of cases, compared with 48% of those it predicted would not.
Pancreatic cancer has one of the lowest survival rates of any cancer, with around 13% of patients alive five years after diagnosis. Most are diagnosed at an advanced stage and face a choice between two chemotherapy regimens that perform similarly on average. No test has yet been validated in a randomised trial to show which will work for which patient.
Prof Andrew Biankin, Regius Professor of Surgery at the University of Glasgow and a key opinion leader in pancreatic cancer , said:
“For many patients with advanced pancreatic cancer, the first treatment is the only one they are likely to receive, so choosing it well matters enormously. In this disease, tumours often don’t shrink much even when a treatment is working, which makes disease control a crucial signal and is closely linked to how long patients live.”
Rather than learning from the trial itself, FarrSight® draws on knowledge built up from 25 billion data points spanning more than 190,000 patients, 17,600 compounds and over 1,200 cancer models. It builds a computer simulation, or ‘digital twin’, of each patient’s tumour biology from their genomic and clinical data, then predicts how that patient would respond to each treatment option.
“The ultimate test of any predictive model is how it performs on data it has never seen. FarrSight® picked out which patients the treatment was likely to work for, and the probabilities it gave were in line with what actually happened."
Dr Matthew Griffiths continued: “That’s what lets drug developers use digital twins to simulate a control group in a hard-to-recruit trial, or check whether an early signal will hold up before committing to a larger study.”
Because patients in PASS-01 were randomly assigned to treatment, some happened to receive the regimen FarrSight® favoured and some did not. Those who did had a 26% lower risk of death, an early sign that the platform could help choose between treatments.
The results build on Concr’s peer-reviewed study of the COMPASS pancreatic cancer cohort, published in Frontiers in Artificial Intelligence in August 2026. In that study, FarrSight®-derived biomarkers outperformed conventional tumour subtyping in the most aggressive form of the disease. A larger manuscript validating Concr’s digital twin approach across more than 5,000 patients is under editorial review.
Prof. Biankin added: “Results like these are a step in the right direction for how we develop drugs today, and how we will treat cancer patients in the future. Most approved treatments deliver only incremental benefit, largely because only a fraction of patients actually benefit. The rest bear the side effects without the gain, and health systems bear the cost. Predicting how an individual patient will respond, before treatment starts, would go a long way to addressing both.”
Drug developers already use FarrSight® to find the right cancer for a new drug and the right patients to administer it in clinical trials. It can also simulate control groups, so fewer patients are enrolled without receiving the investigational drug. The result is more ethical trials and more effective medicines for the patients they are meant for.
