can be higher or lower depending on the number and frequency of screenings and the type of mammogram performed.
The test was able to predict the tissue in which the cancer originated in 96% of samples, and it was accurate in 93%.
Tumours shed DNA into the blood, and this contributes to what is known as cell-free DNA (cfDNA). However, as the cfDNA can come from other types of cells as well, it can be difficult to pinpoint cfDNA that comes from tumours. The blood test reported in this study analyses chemical changes to the DNA called "methylation" that usually control gene expression. Abnormal methylation patterns and the resulting changes in gene expression can contribute to tumour growth, so these signals in cfDNA have the potential to detect and localise cancer.
The blood test targets approximately one million of the 30 million methylation sites in the human genome. A machine learning classifier (an algorithm) was used to predict the presence of cancer and the type of cancer based on the patterns of methylation in the cfDNA shed by tumours. The classifier was trained using a methylation database of cancer and non-cancer signals in cfDNA. The database is believed to be the largest in the world and is owned by the company involved in this research, GRAIL, Inc. (California, USA).
Senior author of the paper, Dr. Michael Seiden (MD, PhD), President of US Oncology (Texas, USA), said: "Our earlier research showed that the methylation approach outperformed both whole genome and targeted sequencing in the detection of multiple deadly cancer types across all clinical stages, and in identifying the tissue of origin. It also allowed us to identify the most informative regions of the genome, which are now targeted by the refined methylation test that is reported in this paper."
In the part of the Circulating Cell-free Genome Atlas (CCGA) study reported today, blood samples from 6,689 participants with previously untreated cancer (2482 patients) and without cancer (4207 patients) from North America were divided into a training set and a validation set. Of these, results from 4316 participants were available for analysis: 3052 in the training set (1531 with cancer, 1521 without cancer) and 1264 in the validation set (654 with cancer and 610 without cancer). Over 50 types of cancer were included.
The machine learning classifier analysed blood samples from the participants to identify methylation changes and to classify the samples as cancer or non-cancer, and to identify the tissue of origin.
The researchers found that the classifier's performance was consistent in both the training and validation sets, with a false positive rate of 0.7% in the validation set.
The classifier's ability to correctly identify when cancer was present (the true positive rate) was also consistent between the two sets. In 12 types of cancer that are often the most deadly (anal, bladder, bowel, oesophageal, stomach, head and neck, liver and bile duct, lung, ovarian and pancreatic cancers, lymphoma, and cancers of white blood cells such as multiple myeloma), the true positive rate was 67.3% across clinical stages I, II and III. These 12 cancers account for about 63% of cancer deaths each year in the USA and, at present, there is no way of screening for the majority of them before symptoms show. The true positive rate was 43.9% for all cancer types in the study across the three clinical stages.
MEDICA-tradefair.com; Source: European Society for Medical Oncology
