Just like how Hansel and Gretel left behind a trail of breadcrumbs to guide them back home, cancer cells shed tiny fragments of DNA into the bloodstream. These fragments, known as circulating tumour DNA (ctDNA), offer doctors a way to detect cancer earlier, track treatment effectiveness and spot signs of relapse sooner—all from a simple blood test.
Yet current methods for measuring ctDNA often require complex and expensive DNA sequencing to screen for common cancer mutations. “Because cancer mutations vary between patients, test results can be inconsistent, making it difficult to track treatment response with blood tests effectively,” explained Anders Skanderup, a Group Leader at the A*STAR Genome Institute of Singapore (A*STAR GIS).
In collaboration with National Cancer Centre Singapore (NCCS), National University of Singapore and Singapore General Hospital, Skanderup and team developed ‘Fragle’, a machine learning model that can accurately quantify ctDNA levels. The model was trained on DNA sequencing data from hundreds of samples from patients with colon, breast, liver or ovarian cancers, as well as healthy volunteers.
As cancer DNA tends to be shorter and more fragmented than healthy DNA, Fragle uses an artificial intelligence model to profile the lengths of all DNA fragments and estimate the ctDNA fraction in a sample.
Across independent cohorts of different cancer types, the team showed that Fragle bested existing methods, offering both greater accuracy and an improved limit of detection, or the lowest ctDNA level that can be measured with confidence. The model effectively distinguished between healthy and ctDNA-bearing samples, even when the ctDNA made up a mere one percent of the total DNA fragments in the blood.
Unlike other methods that require whole genome sequencing, Fragle could also analyse targeted sequencing assays routinely collected in clinics. Testing the model on data from lung cancer patients post-surgery, the researchers found that Fragle was able to sub-classify samples flagged as ctDNA-negative by traditional assays into ‘ctDNA-low’ and ‘ctDNA-high’ groups. Patients in the ctDNA-high group were linked to significantly worse clinical outcomes, highlighting that Fragle could help stratify risk more accurately.
“Since the method is rapid and compatible with existing clinical workflows, Fragle has the potential to make ctDNA testing a routine part of patient management,” said Skanderup. “This could allow doctors to adjust treatments more quickly and precisely, ultimately improving outcomes while reducing the need for invasive diagnostic procedures.”
Skanderup is working with Daniel Tan, co-author of the study and a Senior Consultant at NCCS, to bring Fragle into the clinic. In an ongoing study, the team is using Fragle to regularly monitor ctDNA levels of over 100 patients who are undergoing treatment for lung cancer, hoping to detect signs of relapse earlier than routine scans and show the value of integrating such models in cancer care.
The A*STAR-affiliated researchers contributing to this research are from the A*STAR Genome Institute of Singapore (A*STAR GIS).
