AI combines multiple DNA clues to improve cancer detection from blood

The method, developed by a team led by scientists at Queen Mary University of London’s Barts Cancer Institute, and the University of Cambridge, uses a comparatively inexpensive form of genome sequencing to detect signals shared by a range of cancer types.

When tested on samples it had not encountered before, the system performed better than several established approaches. The findings provide a proof-of-principle and offer a roadmap for the development of more accurate, cheaper and more accessible cancer screening and monitoring tools in the future.

The study, published in Science Advances, describes a new machine learning approach called UNITE. It was led by Dr Haichao Wang, a postdoctoral researcher at Barts Cancer Institute, with Institute Director Professor Nitzan Rosenfeld and Group Leader Dr Hui Zhao as senior authors.

UNITE analyses cell-free DNA: small fragments of genetic material that are released into the blood as cells break down. Some of these fragments can come from cancer cells and carry information about the disease. However, tumour DNA makes up only a very small share of all the cell-free DNA in a blood sample, especially when a cancer is at an early stage. This makes it difficult to detect.

The new method tackles this challenge by looking for several different clues carried by DNA fragments rather than relying on a single sign. These include the length of the fragments, how their lengths vary across the genome, the DNA letters found at their ends and changes in the number of copies of different sections of DNA.

Dr Haichao Wang, first author of the paper, says:

“Each feature lets us look at the data from a different angle. When we combine these signals, the model becomes more accurate and better able to recognise the subtle patterns linked with cancer.”

Searching for signals shared across cancers

Many existing liquid biopsy tests – which look for signs of cancer in blood, urine and other bodily fluids – search for mutations known to be present in a person’s tumour. These approaches can be useful for monitoring someone after they have been diagnosed and treated, but they are less suited to cancer screening, where there is no tumour sample to analyse in advance.

By contrast, UNITE does not need prior information about an individual tumour. Instead, it looks for wider patterns in cell-free DNA that can occur in a range of different cancers. This could make approaches like UNITE more suitable for testing people who do not have a cancer diagnosis.

The team developed and evaluated their model using more than 2,000 blood plasma samples from people with 26 different cancer types and people without cancer. Training the model using a wide range of cancers allowed it to learn broader patterns linked with the disease, rather than focusing on markers associated with one type.

The researchers tested UNITE on a completely independent set of samples from clinical studies, which the model had not encountered before. At a threshold that correctly classified 95% of samples from people without cancer, the best-performing version of UNITE detected the presence of cancer in 47% of samples from people with cancers across all stages. This included 31% of samples from people with stage I and II cancers, 60% of samples from people with stage III cancers and 59% of samples from people with stage IV cancers, outperforming two established models for assessing cell-free DNA.

An additional advantage of UNITE is its use of shallow whole-genome sequencing, in which a limited amount of sequencing data is used to study signals across the whole genome. Because shallow sequencing generates less data than standard whole-genome sequencing, it could make the approach cheaper and easier to scale compared with standard deep sequencing. This may be important if blood-based detection methods are eventually used to test large numbers of people.

A roadmap towards better screening tools

UNITE provides a proof-of-concept, and further development will be required to translate these findings into better clinical tools for cancer screening and monitoring. The current method still misses many cancers, particularly at an early stage, and the study did not test UNITE’s ability to track changes over time.

However, the researchers suggest that approaches like UNITE could one day be used to complement established screening methods. Liquid biopsies might help identify people who could benefit from further investigation using imaging, endoscopy or other diagnostic tests.

Dr Wang says: “Our longer-term vision is that approaches like this could help us look for signs of cancer using a simple blood sample. It would not replace existing screening and diagnostic methods, but it could help guide who should be referred for further investigation.”

Dr Hui Zhao says: "We believe liquid biopsy will play an increasingly important role in the future of cancer detection, and our hope is that this work will contribute to that progress. Rather than developing a single model, we created UNITE as a flexible framework that can continue to evolve as new biological features, analytical approaches and larger datasets become available. As we integrate more complementary sources of information, we expect the model to become increasingly sensitive while remaining scalable and practical. We hope this framework will support future innovation in blood-based cancer detection and help move more effective liquid biopsy technologies closer to clinical use."

The team now plans to add other types of biological data to the method to improve its accuracy and sensitivity. One area of interest is DNA methylation – chemical marks on DNA that help control how genes are switched on and off as part of a process known as epigenetics. Methylation patterns differ between cells from different organs and can also change in cancer cells. Adding this information could therefore help the model detect subtle cancer signals and may also provide clues about where in the body they came from.

The researchers are also exploring ways to make blood-sample collection more accessible, including the use of dried blood spots that could be easier to store and transport.

Professor Nitzan Rosenfeld says: “We are grateful for the support of patients and their families who take part in this research, and for cancer and research charities and individuals who contribute to them. Without this support, such research would not be possible. I am very excited about the results we publish today, which show what we can achieve with relatively small amounts and low-cost sequencing data, once this is collected and compared across many individuals. These findings help demonstrate the types of approaches that could become useful for clinical use in the foreseeable future, to help detect cancer earlier and to reduce the burden of this disease.”

Hayley Brown, Research Information Manager at Cancer Research UK, says:

“Detecting cancer earlier gives people the best chance of successful treatment, so developing better blood tests that do so is an important area of research.

“Rather than relying on a single clue, this study shows how artificial intelligence can combine lots of subtle signals from DNA in the blood to help identify cancers that might otherwise be difficult to detect.

“Although this approach is still in the early stages and will need further testing before it could be used in healthcare, it’s an encouraging example of how advances in technology could help make earlier cancer detection more accurate and accessible in the future.”

This research was primarily funded by Cancer Research UK, with additional support from Barts Charity, the National Institute for Health and Care Research and others. A full list of funders is available in the published paper.