LONDON, UK, 04 SEPTEMBER 2026 – Concr, the techbio company using Bayesian AI for precision oncology, today announces the publication of new peer-reviewed evidence in Frontiers in Artificial Intelligence showing that Concr’s proprietary molecular digital twins of individual patients outperform conventional biomarker stratification in pancreatic ductal adenocarcinoma (PDAC).
The study applied Concr's FarrSight® Bayesian foundation model to participants in the COMPASS trial (NCT02750657), and compared its predictions head-to-head against the established Moffitt classification of PDAC.
Why PDAC?
Pancreatic cancer has a five-year survival of around 13%. Its most aggressive form, the basal-like subtype, is the group most in need of better stratification, and where conventional approaches consistently fail. Cohort-derived biomarkers rely on population averages and large sample sizes; in the small cohorts typical of early-phase oncology trials, they are statistically underpowered before they begin. That directly contributes to trial failure, and to patients receiving toxic regimens they will not benefit from.
The FarrSight® approach
Rather than fitting a single fixed set of feature weightings across a cohort, FarrSight® builds a digital twin of each individual patient — a simulation of their molecular biology — and determines which features drive response for that person. These individual predictions are then aggregated into a response signature.
This is a departure from prevailing approaches:
- Transcriptomic subtyping assigns a patient to a subgroup and predicts from that subgroup's average.
- Statistical biomarker discovery derives its feature weightings from the study cohort itself, and applies one fixed set of weights to everyone in it.
- Traditional digital twins trained on historical outcome data learn expected trajectories from the population they were trained on.
Each is bounded by the same constraint: performance depends on how well the patient in front of you is represented in the data available, which is why they struggle in the small, heterogeneous cohorts typical of early-phase oncology.
FarrSight® arrives carrying Bayesian priors learned from 25 billion data points spanning 190,005 patients, 17,633 compounds, 890 cell lines and 335 patient-derived xenograft models (as at June 2026), and uses transfer learning to bring pre-clinical signals into clinical predictions. It builds a digital twin of each individual patient across clinico-pathological, expression, mutation and copy-number data, where the comparator biomarker in this study drew on gene expression alone, and reports which features drove that patient's prediction.
The results
Across the 206 patient COMPASS cohort, comprehensive whole-genome, RNA-Seq and clinico-pathological data were available. In the 38 patient basal-like subtype:
AUC of 72.3% for the FarrSight®-derived biomarker, against 44.8% for the conventional biomarker, which performed no better than a coin toss
Overall accuracy of 65.8% vs 47.4%
The FarrSight® biomarker significantly enriched for disease control, where the conventional biomarker failed to stratify at all
At the individual level, FarrSight® correctly identified all three partial responders in the basal-like cohort; the conventional approach misclassified two of the three
FarrSight®-based stratification translated into a significant difference in overall survival
Feature importance analysis showed substantial patient-to-patient variation in what drove each prediction and independently recovered known biological hallmarks of the basal-like subtype, alongside mechanisms less prominent in standard subtyping.
Dr Irina Babina, CEO of Concr, commented: "Biomarkers built on population averages tell you what happens to a group. Patients aren't groups. What this paper shows is that when you model the individual, you can find signal in cohorts that are far too small for conventional statistics — exactly the situation drug developers face in early-phase trials, and exactly where assets are being abandoned that might still work in the right patients."
Dr Babina added: “Our model worked on the same data that conventional approaches failed because it arrives carrying prior biological knowledge rather than trying to learn everything from a handful of cases — and because it can show you why, patient by patient. HTA bodies and regulators demand mechanistic justification of therapeutic value, and they want the biological assumptions behind it. That is exactly what a Bayesian digital twin produces."
What's next
A full manuscript, extending this approach with validation across a cohort of more than 5,000 patients, is currently under editorial review.
Concr is inviting biopharma partners — particularly those seeking to reposition oncology assets following an unsuccessful trial readout, or to define patient selection strategies before committing to pivotal studies — to get in touch.
Read the full paper here.
About Concr
Concr is a London-based TechBio company that uses astrophysics-derived Bayesian models to predict which cancer treatments will work for individual patients. Its FarrSight® platform creates digital twins, simulations of a patient's molecular biology, to match them with effective therapies and help drug developers design better clinical trials. Founded in 2018, the company is led by CEO Dr Irina Babina. Concr is a VC-backed venture with investors including Cambridge Enterprise, R42 Group, Oncology Ventures, and Debiopharm Innovation Fund, and has had two Innovate UK-funded clinical studies. Concr works with partners including the NHS, Debiopharm, Roche, Step Pharma, and the Institute of Cancer Research.
For Media Enquiries
Emre Misirlioglu — emre@concr.co
Release Notes
This study was conducted with the support of the Ontario Institute for Cancer Research through funding provided by the Government of Ontario. The views expressed in the publication are the views of the authors and do not necessarily reflect those of the Government of Ontario.
Concr products and services are for research use only. They are intended to support research and strategic decision-making in therapeutic and diagnostic development. They are not a medical device and are not intended to diagnose, treat, cure, or prevent any disease.
Back
LinkTree