UCL Digital Twins Model Breast Cancer Treatment

Scientists at UCL have built computer copies of 31 patients’ breast tumours and used them to suggest treatments that might have worked when chemotherapy did not. The “digital twins”, described in a paper in Cell Reports Medicine published on 26 October, reproduced the patients’ real trial outcomes in 74% of cases. Every proposed treatment is a computer prediction. None has been given to these patients or tested in a new trial.

The work, done with the University of Cambridge, targets triple-negative breast cancer (TNBC), an aggressive form that accounts for about 15% of breast cancers, according to UCL’s release. TNBC lacks the three receptors that hormone therapies and HER2 drugs act on, so options are limited. The team’s pitch is that its model, unlike a “black box” AI, shows how it reaches each answer.

In brief

UCL built digital twins of 31 tumours from the UK PARTNER trial; they matched real outcomes 74% of the time.
For 12 of 13 patients who did not respond fully, the model proposed at least one extra targeted drug.
The authors say the ideas need prospective functional and clinical validation; the cohort is small.

How the virtual tumours were built

The model is what the authors call a mechanistic qualitative network. Each node stands for a gene or protein involved in cancer, and each link is an interaction between them, drawn from published research. UCL describes this as “symbolic AI”, which “relies on logic rather than prediction and where its workings-out can be clearly seen”. The network was built in BioModelAnalyzer, an open-source tool, and checked against 160 results from 33 earlier lab studies that had nothing to do with the patients.

To make a twin, the team loaded each patient’s tumour mutations and copy-number changes, taken from whole-genome sequencing, into the network. The 31 patients had taken part in PARTNER, a Cambridge-led trial that gave patients with TNBC chemotherapy with or without olaparib, a PARP inhibitor, before surgery. The researchers then reran the trial on the computer: chemotherapy alone, chemotherapy plus olaparib, and no treatment as a control.

What the twins got right, and what they proposed

According to the paper, the model correctly called 10 of 14 patients who had a complete response and 13 of 17 who did not, an overall accuracy of 74%. That is 71% sensitivity and 76% specificity. Rerunning the model while leaving out one patient at a time gave a similar average, 74.3%.

The team then asked the twins of the 13 patients who did not respond fully what else might help. For 12 of them, the model found at least one added targeted drug that improved the predicted response. For 10, it predicted a complete response, meaning no invasive cancer would likely be left in the breast and lymph nodes. Eight of those needed one extra drug and two needed two. One patient’s twin found nothing that worked.

The most broadly useful target was ATM, a DNA-repair protein. An ATM inhibitor was predicted to produce a complete response in seven non-responders and a better response in two more. Other suggested targets included Chk1, ATR, MEK, mTOR and Wee1. The paper says drugs aimed at those have passed phase 2 trials, but that ATM inhibitors are still only in phase 1.

“We see our computer model as an aid for existing cancer clinical practice, not an alternative to it,” said Professor Jasmin Fisher of the UCL Cancer Institute, the paper’s senior author and lead contact; Pedro Victori is first author. She said a clinician might reject a suggested drug as too toxic, “But it may also lead them to explore, and eventually settle on, a third option which they might otherwise not have arrived at.”

The study was funded by Wellcome Leap, Cancer Research UK, the Mark Foundation and the NIHR UCLH Biomedical Research Centre, among others. The paper discloses that Fisher sat on a company’s scientific advisory board and that Professor Jean Abraham, who led PARTNER, has received honoraria from AstraZeneca, which funded the trial.

Olaparib’s mixed record is part of the test

UCL’s release describes olaparib as “known to be effective in treating TNBC in some cases”. The PARTNER trial itself, published in Nature in 2024, found that adding it before surgery did not help patients without an inherited BRCA mutation: 51% had a complete response with olaparib against 52% without, and event-free and overall survival were no better. All 31 twins were run both with and without the drug.

That makes PARTNER a useful benchmark. A model that predicted olaparib would rescue most patients would have been wrong; what the paper reports instead is that response depended on each tumour’s genetic background.

Predictions that still need patients

The strongest case for this work is not the 74% figure, which on 31 patients carries wide uncertainty. It is that every prediction can be traced. When the model says an ATM inhibitor should work for a patient, a doctor can see which mutations and which pathway produced that call, and argue with it. That is a different proposition from the risk scores in FDA-cleared imaging tools or models trained on routine records, which are judged mainly on how well they rank people.

The limits are spelled out by the authors. “Therapeutic hypotheses are derived from in silico simulations and require prospective functional and clinical validation,” the paper says. “The clinical cohort is modest in size.” The model uses only genetic changes, leaves out the immune system and surrounding tissue, and covers only the drugs used in PARTNER, so it says nothing about immunotherapy or antibody-drug conjugates, both now used in TNBC.

There is also a tension in the headline target. ATM came out as the most broadly useful drug, yet it is the least developed, with inhibitors only in early trials. The suggestions that could reach patients soonest are the ones the twins rated less often.

Like Stanford’s recent immune digital twins for preterm labour, this is a hypothesis generator. It sits alongside other UK work, such as KCL’s AI for spotting DNA-repair defects, that tries to match drugs to tumour biology. The next proof will come only when a twin’s suggestion is given to a patient in a trial and the tumour responds as predicted.