UCL scientists design symbolic AI digital twins to treat cancer

In a new paper published in Cell Reports Medicine, the UCL team outline how they, in collaboration with a University of Cambridge team, developed a new type of treatment-predicting computer program that uses an individual's detailed genetic fingerprint to explore which treatments, if any, may kill their cancer – especially when standard treatments have failed.

Computer models have long been used in cancer medicine and more recently artificial intelligence has opened up a whole new avenue in this regard.

But the "black box" nature of 'traditional' AI associated with Large Language Models – where how they arrive at their conclusions is unclear – has prompted a degree of scepticism among the medical profession, the researchers said.

In developing their model, specifically to recommend treatment options for Triple-Negative Breast Cancer (TNBC), a hard-to-treat cancer, the UCL team instead used a form of AI called 'symbolic AI' which relies on logic rather than prediction and where its workings-out can be clearly seen.

TNBC is an aggressive, fast-growing form of breast cancer that accounts for about 15% of all breast cancer cases.

To test the effectiveness of their approach, the team built computer-generated tumors – or 'digital twins' – that aped the real-life tumors of 31 patients involved in the Platinum and PolyAdenosine 5'-Diphosphoribose Polymerization Inhibitor for Research and Treatment of Neo-Adjuvant Early and Relapse-Risk Breast Cancer (PARTNER) clinical trial using their genetic information.

They then copied the process the researchers had carried out in the trial, but as a computer simulation: on each digital twin, they saw what happened when they treated the tumor with either chemotherapy alone or with chemotherapy and a drug called Olaparib, known to be effective in treating TNBC in some cases. They also did a control simulation where no treatment was carried out.

They then compared the results of the computer simulations with those from the clinical trial and found the computer program correctly predicted whether treatment would work in 74% of cases.

The digital twins also suggested ways to help patients whose tumors had not responded well to treatment.

In 12 of the 13 patients who did not respond fully, the models identified at least one extra targeted treatment that might have improved the outcome.

For 10 of these patients, the models predicted that the additional treatment could prompt such a high level of response that no invasive cancer would likely be found in the patient's breast and lymph nodes afterwards.

Lead author Professor Jasmin Fisher (UCL Cancer Institute) said: "We have shown the broad and powerful range of insights that personalized virtual tumors could bring to cancer medicine through our symbolic AI-based computer model.

"By identifying treatment approaches for patients who did not respond to standard treatments, our approach paves the way for digital twins to be used by clinicians as a tool to guide their therapeutic decisions, particularly in hard-to-treat cancers like TNBC.

"We see our computer model as an aid for existing cancer clinical practice, not an alternative to it: a way of generating new ideas particularly in relation to proteins or genes that could be targeted in ways that, as it were, restore a patient's sensitivity to existing drugs and make those treatments effective.

"In the past, some cancer specialists have understandably regarded computer models with a degree of scepticism, especially where they are reliant on AI platforms where the inner workings are not visible – so-called black box models.

"We have got round this problem by using symbolic AI which is based on logical reasoning and where how the platform arrive at its answers can be seen easily and interrogated by doctors.

"In practice, the treatment option recommended by our model may, under closer scrutiny, be discounted by the clinician – for example, if they view the 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."

TNBC is a specific type of breast cancer where the cancer cells test negative for oestrogen receptors, progesterone receptors and the HER2 protein. Because these three common markers are absent, standard hormone therapies and HER2-targeted drugs do not work.

It is also often characterized by a condition known as Homologous Recombination Deficiency, affecting roughly 15% to 20% or more of cases.

This condition means it is able to mutate very quickly and very dangerously, but also creates a major biological blind spot and means drugs used to block a particular enzyme – the poly-ADP ribose polymerase (PARP) enzyme which normally helps damaged cells repair single-strand DNA breaks – can kill it.

These drugs include Olaparib, which was used in the PARTNER trial led by Professor Jean Abraham and scientists at the University of Cambridge and Addenbrooke's Hospital, Cambridge, between 2016 and 2024. The UCL team worked with Professor Abraham's team on this new study.