Illustration of a diverse primary tumour whose genome has been sequenced to identify 23 key genes that can predict survival chance. The mass on the right represents a secondary metastatic tumour seeded by cancer cells in the bloodstream. Credit: Jeroen Claus, Phospho Biomedical Animation
A new genetic test to distinguish between high- and low-risk tumours, which could help doctors and patients to make crucial treatment decisions, has been developed by a London-led international consortium.
The new approach, published in Nature Medicine, was developed and tested in lung cancer patients but this same method could be applied to other cancer types in future.
In lung cancer, Stage I tumours are typically surgically removed without chemotherapy, while patients with Stage II cancer are usually given chemotherapy after surgery. However, around a quarter of Stage I cancers return after surgery, and chemotherapy may cause unnecessary harm to patients with low-risk Stage II cancers.
Previous genetic approaches have failed to reliably predict the risk of relapse, as there are significant genetic differences between tumours from patient to patient, and even within different parts of the same tumour within an individual patient. The new technique, called ORACLE (Outcome Risk Associated Clonal Lung Expression), combines machine learning approaches with an understanding of cancer evolution to overcome these problems.
"Knowing how aggressive a cancer is at an early stage could help us to make better treatment decisions, and we're hoping to test this in future trials," explained senior author Professor Charlie Swanton, Group Leader at the Francis Crick Institute and UCL Cancer Institute, and Chief Clinician of Cancer Research UK. " If we could identify high-risk cancers that may seem low-risk by traditional measures, we could potentially nip them in the bud with additional treatment and save lives. More accurate prognostic information would also help patients to make more informed decisions, which we hope would make a meaningful difference to survival."
When tested on data from international cancer studies covering 904 patients, the ORACLE risk score was significantly associated with patient survival.
To compare the risk score with detailed clinical data, the team applied ORACLE to 103 lung cancer patients treated at Uppsala University in Sweden. Having a high ORACLE risk score was associated with a three-fold increased risk of dying in the five years following diagnosis. This survival association was independent of known risk factors such as clinical stage and smoking history, suggesting that ORACLE offers a new objective measure of tumour aggressiveness at a molecular level.
For the 60 Uppsala patients with Stage I cancer, ORACLE was able to identify those with a significantly higher chance of survival. By contrast, existing methods were unable to predict any significant survival difference between the two groups.
The ORACLE approach builds on the group's previous research into cancer evolution, using extensive genetic data from lung tumours. The team analysed gene activity across 156 different tumour regions from 48 patients from TRACERx, a major Cancer Research UK-funded study to track the evolution of lung tumours over time. They focused on genes that tend to be expressed across the whole tumour, to avoid sampling bias.
"In clinical practice, it's not generally practical to take more than one tumour sample for genetic testing and could be prohibitively expensive," explained co-lead author Dhruva Biswas from the Francis Crick Institute and UCL Cancer Institute. "However, with existing tests, you could sample two regions of the same tumour from the same patient and get two completely different results. We set out to develop a test that would avoid this bias by looking at genes that are consistently expressed across tumours and not affected by where you place the biopsy needle."
Looking across all 20,000 expressed genes, the team identified a pool of 1,080 candidate genes that were stable across each tumour and unaffected by biopsy location. They then used machine learning algorithms to find the genes that were most associated with survival. They narrowed this down to a smaller subset of 23 genes, as these were predictive without requiring too many genes to be feasible in a clinical setting.
"Our next step will involve further validating the ORACLE gene signature with samples from even larger clinical data sets," said Dhruva. "If successful, we hope to conduct clinical trials in the coming years to test whether using ORACLE to inform treatment options can improve cancer survival."