from Nature Reviews Genetics by Liesbet Lieben
A systemic, minimally invasive diagnostic test to determine whether an individual has cancer could greatly improve cancer prognosis by facilitating early diagnosis and improving monitoring. However, the use of blood-based ‘liquid biopsies’ is challenging, because the biosources (such as plasma DNA, exosomes and circulating tumour cells) and analytical platforms typically have a suboptimal sensitivity. Now, Würdinger and colleagues have developed an alternative method based on their previous observation that blood platelets change their mRNA expression profile in response to contact with a tumour (so-called ‘tumour-educated platelets’); combining platelet RNA sequencing with a self-learning algorithm results in a highly sensitive, specific and accurate diagnostic test for cancer.
To establish their method, platelet samples from healthy controls (n=55), patients with early, localized tumours (n=39) and patients with advanced, metastatic tumours (n=189) were collected, covering 6 different tumour types (lung, colorectal, brain, pancreatic, hepatobiliary and breast cancer). SMARTer mRNA amplification followed by sequencing and selection of spliced RNA reads yielded ~5,000 differentially expressed protein-coding and non-coding RNAs between patients with tumours and healthy controls. Selection of a classifier-specific gene list of 1,072 RNAs combined with the use of a custom machine-learning algorithm allowed discrimination of the presence versus the absence of cancer with a sensitivity (proportion of correctly identified positives) of ~97%, a specificity (proportion of correctly identified negatives) of ~94% and an accuracy of ~95%. By contrast, the use of random classifiers had no predictive power. Thus, nearly all cancer patients — regardless of the type of tumour — have abnormal platelet RNA profiles and are identified by the algorithm.
“this diagnostic platform — based on sequencing RNA obtained from blood platelets and self-learning algorithms — can discriminate patients with cancer from healthy controls, predict the localization of the primary tumour and provide information on the molecular tumour phenotype”
Next, tumour-specific gene lists were selected by unsupervised hierarchical clustering of differentially expressed RNAs between the six different cancer types. Combined first and second ranked classification resulted in a 89% accuracy to classify the tumour types correctly. Finally, the authors assessed whether platelet RNA reflects the molecular profile of the tumour tissue. Selection of biomarker-specific gene lists allowed the discrimination of KRAS and EGFR mutations, HER2 amplification and MET overexpression compared to wild-type sequences in the platelet RNA profile, which matched the respective tumour DNA analysis.