A blood test that can ‘detect over 50 cancer types’ is big news this week.
There’s a lot of excitement around the latest research, published in Annals of Oncology. And it’s easy to see why.
Scientists have used machine learning to help identify if someone has cancer based on tiny bits of tumour DNA floating in their blood. Which could open the door to a blood test that can detect and identify multiple types of cancer.
But it’s not there yet. And in the blood test buzz, some news articles have missed out crucial details.
The new method
The team looked for differences in the DNA shed from cancer cells and healthy cells into the blood.
They focused on differences in a chemical ‘tag’ that sit on top of DNA in cells, called methyl groups. These groups are usually spread evenly across the DNA in cells, but in cancer cells they tend to cluster at different points. And it’s this distinction scientists wanted to exploit.
They trained a machine learning algorithm – a type of artificial intelligence that pick up patterns and signals – to detect differences between methylation patterns in DNA from cancer and non-cancer cells.
The algorithm was trained on 3,052 samples from people with and without cancer from two large databases.
And once the program was fired up and ready to go, the team tested its cancer-spotting ability on a different set of 1,264 samples f, half of which were from people with cancer.
What did they find?
Any test with the goal of being able to detect cancers at their earliest stages in people without symptoms must strike the right balance between picking up cancer (sensitivity) and not giving false positives (specificity). We’ve blogged before about what makes a good cancer test, as well as the efforts to develop a cancer blood test.
How do you assess a cancer test?
Researchers look at 3 main things when assessing a new diagnostic test.
- Sensitivity – the probability that you test positive if you have the disease.
- Specificity – the probability that you test negative if you do not have the disease.
- Accuracy – the proportion of samples correctly classified by the test. It’s a combined measure of sensitivity and specificity. It’s dependent on how experiments are run, and so can be misleading in news stories.
Firstly the good news: fewer than 1% of people without cancer were wrongly identified as having the disease. Which is a good sign for the specificity of this test.
And when it came to detecting cancer, across all types of cancer, the test correctly identified the disease in 55% of cases. This is a measure of the test’s sensitivity.
But there was a huge variation in sensitivity depending on the type of cancer and how advanced the disease was. The test was better at picking up more advanced cancer, which makes sense – more advanced cancers typically shed more DNA into the bloodstream.
If we look at the numbers, across all cancer types the test correctly detected the disease in 93% of those with stage 4 cancer, but only 18% of early, stage 1 cancers.
An important consideration is that the study was only testing if the algorithm could detect cancer in patients who were already known to have cancer. According to the researchers, these figures may change if the test was used on a wider, general population.
Encouragingly for a multicancer test, when the researchers looked at a smaller number of samples to explore if the test helped them identify where the cancer was growing, the algorithm was able to predict the location in 96% of samples, and it was accurate in 93%.
First things first, although the samples numbers are big, they become a lot smaller when you break them down by cancer type and cancer stage. Some cancer types were particularly poorly represented, with only 1 or 2 samples included in the final analysis – so there’s more work to do there. Based on this, it’s a bit too soon to say that the test can pick up 50 cancer types.
And if the plan is to use this as a screening tool, then the researchers will need to do more to study people who didn’t have symptoms when they were diagnosed. The current study included people who were symptomatic as well as people without symptoms.
And the participant data lacked variation in age, race and ethnicity. Between 83 and 87% of all the samples used to train and test the algorithm were Caucasian.
The big conclusion is that these results are encouraging and should be taken forward into bigger studies. But it’s important to put the results in context – they’re a step in the right direction. There are a lot of steps between this study and a fully-fledged cancer test.
According to the research team, they plan to validate the results using samples from US and UK studies, and well as to begin to examine if the test could be used to screen for cancer. We look forward to seeing the results.
Our head of early detection research, Dr David Crosby, sums it up nicely: “Although this test is still at an early stage of development, the initial results are encouraging. And if the test can be fine-tuned to be more efficient at catching cancers in their earliest stages, it could become a tool for early detection.
“But more research is needed to improve the test’s ability to catch early cancers and we still need to explore how it might work in a real cancer screening scenario.”