6 September 2018. Medical researchers and data scientists developed a system for analyzing genomic data from cancer patients that accurately detects more cancer-causing mutations than other current techniques. An academic-industry team reports its findings in yesterday’s issue of the journal Science Translational Medicine (paid subscription required).
Researchers from the lab of cancer researcher Victor Velculescu at Johns Hopkins University and the company Personal Genome Diagnostics, both in Baltimore, are seeking better methods for detecting mutations in the genomes of cancer patients that develop and change as tumors progress. Accurately detecting and characterizing these non-inherited, or somatic, mutations better define the targets for more precise therapies, but according to the authors, today’s analytical techniques — even those using next-generation or high-throughput genomic sequencing — can vary in accuracy. In addition, some of these methods require high-quality specimen samples that may not always be available.
The study evaluated a technology being developed by Personal Genome Diagnostics called Cerebro that applies machine learning, a form of artificial intelligence, to genomic analysis of suspected tumor specimens and blood samples. Cerebro uses a random-forest algorithm that builds data into decision-trees, then merges the data from the trees together, thus the “forest” reference, for more accurate and stable predictions. The team led by Velculescu and Samuel Angiuoli, chief information officer at Personal Genome Diagnostics, trained the algorithm with next-generation sequencing data from blood samples. The training data had some 30,000 known genomic variations associated with tumors, as well as 2 million more errors and artifacts from genomic sequencing that could erroneously be considered as mutations. Once trained, Cerebro uses about 1,000 decision trees to classify each mutation.
The team assessed Cerebro’s ability to accurately detect and classify each genomic variation as a cancer-causing somatic mutation. When compared to data from 1,368 cases in the Cancer Genome Atlas, a collection and catalogue of cancer-related mutations, Cerebro matched 74 percent of the variations, but also highlighted false negatives and positives in the data, including those in actionable genes. And Cerebro’s analysis of simulated low-purity tumor data, similar to those taken from many patients, was more accurate and had higher predictive value than 6 other current mutation-detection techniques.
In addition, the researchers coupled Cerebo’s analysis with next-generation sequencing of tumors from 22 lung cancer patients, to determine their likelihood to respond to immunotherapies. In this analysis, the team looked particularly for tumor mutation burden, a quantitative measure of acquired somatic mutations in tumors and an indicator of response to treatment. The results show that adding Cerebro to 3 other next-generation sequencing methods more accurately classifies tumors by their likelihood to respond to immunotherapy.
“Our data showed improved classification of patients when using Cerebro and highlights the importance of accurate mutation detection on treatment decisions,” says Angiuoli in a company statement. Personal Genome Diagnostics is a spin-off enterprise from Johns Hopkins University, founded in 2010 by Velculescu and Luis Diaz, now head of solid tumor oncology at Memorial Sloan Kettering Cancer Center in New York, but previously a postdoctoral researcher at Johns Hopkins.
More from Science & Enterprise:
- A.I., Imaging Shown to Predict Immunotherapy Success
- Precision Medicine Technique Devised for Brain Tumors
- Blood Tests Shown Able to Identify Early Lung Cancer
- Data Tools Designed for Genomics-Based Precision Cancer Care
- Virtual Biopsy in Development to Detect Melanoma
* * *
You must be logged in to post a comment.