Compared with the current standard genetic testing method, deep learning germline genetic testing has a higher sensitivity and specificity for detecting pathogenic variants in patients with prostate cancer and those with melanoma who are at a high risk for recurrence, according to study results published in JAMA.
Pathogenic variants are inherited genetic changes that can be associated with increased risk of cancer. Researchers have increasingly turned to germline genetic testing to identify these variants, most commonly using the Genome Analysis Toolkit as the standard.
DeepVariant genetic testing uses deep learning and neural network methods similar to those used in image recognition software. It is an alternative to germline testing, but few studies have documented its value, according to wrote Saud H. AlDubayan, MD, of Harvard University, Boston, MA, and fellow researchers.
“Computational methods that use deep learning neural networks, which incorporate layers of networks to learn and analyze complex patterns in data, have demonstrated superior performance compared with standard methods for disease recognition, pathological and radiological image analysis, and natural language processing. Deep learning methods also have shown enhanced germline variant detection compared with standard methods in laboratory samples with known genetic variation,” they explained.
In this cross-sectional study, AlDubayan and colleagues compared standard germline detection of pathogenic variants with DeepVariant testing in two convenience samples, comprised of 1,072 patients with prostate cancer (mean age at diagnosis: 63.7 years; 79.9% with European ancestry) and 1,295 patients with melanoma (mean age at diagnosis 59.8 years; 37.7% women; 81.9% with European ancestry).
The deep learning method, they found, identified more patients with pathogenic variants in 118 cancer susceptibility genes compared with the standard method among patients with prostate cancer (198 versus 182, respectively) and in those with melanoma (93 versus 74).
Other findings in their comparison of the deep learning method versus the standard method favored the deep learning method as well, in the following metrics:
- Sensitivity for prostate cancer: 94.7% versus 87.1%, respectively (difference: 7.6%; 95% CI: 2.2%-13.1%).
- Sensitivity for melanoma: 74.4% versus 59.2% (difference: 15.2%; 95% CI: 3.7%-26.7%).
- Specificity for prostate cancer: 64.0% versus 36.0% (difference: 28.0%; 95% CI: 1.4%-54.6%).
- Specificity for melanoma: 63.4% versus 36.6% (difference: 26.8%; 95% CI: 17.6%-35.9%).
- Positive predictive value (PPV) for prostate cancer: 95.7% versus 91.9% (difference: 3.8%; 95% CI: -1.0%-8.4%).
- PPV for melanoma: 54.4% versus 35.4% (difference: 19.0%; 95% CI: 9.1%-28.9%).
- Negative predictive value (NPV) for prostate cancer: 59.3% versus 25.0% (difference: 34.3%; 95% CI: 10.9%-57.6%).
- NPV for melanoma: 80.8% versus 60.5% (difference: 20.3%; 95% CI: 10.0% versus 30.7%).
No between-test differences in sensitivity were found for American College of Medical Genetics and Genomics gene set (94.9% versus 90.6% for deep learning versus standard method, respectively; difference: 4.3%; 95% CI: -2.3%-10.9%). The deep learning method, however, had a higher sensitivity in melanoma (71.6% versus 53.7%; difference: 17.9%; 95% CI: 1.82%-34.0%), as well as in identifying mendelian genes in both prostate cancer (99.7% versus 95.1%; difference: 4.6%; 95% CI: 3.0%-6.3%) and melanoma (91.7% versus 86.2%; difference: 5.5%; 95% CI: 2.2%-8.8%).
In an accompanying editorial, W. Gregory Feero, MD, PhD, of Dartmouth Geisel School of Medicine, Hanover, NH, voiced his agreement with the value of DeepVariant genetic testing.
“Neural network approaches for variant detection are potentially more sensitive for rare variant detection because such variations can be filtered out algorithmically by methods that are dependent on comparisons of detected sequence to population data sets. In some regards this report is a marriage of 2 major scientific advances: genetics and artificial intelligence,” he wrote.
According to Feero, the choice of algorithms is of paramount importance in genome sequencing. He also cautioned that there is more work to be done before a reference standard for clinical variant detection is established.
“In this study the deep learning approach was associated with detection of a higher number of manually validated deleterious variants vs the standard method. However, the spectrum of variants detected by the 2 methods was not wholly overlapping, eg, while the deep learning method found variants that the standard method did not, the converse was also true. If the true pathogenic variant for an affected individual resided in the set of variants detected by the standard method but was not detected by the deep learning approach, the deep learning approach was not an improvement. The authors acknowledged that sequential use of several algorithms would provide the highest sensitivity,” he wrote.
“For the foreseeable future, nongeneticist clinicians should be familiar with the quality of their chosen genome-sequencing laboratory and engage expert advice before changing patient management based on a test result,” concluded Feero.
Study limitations include the inclusion of only patients with a diagnosis of cancer who were primarily of European ancestry, limited data on clinical outcomes, the lack of independent validation of genomic positions, the lack of comparisons between deep learning and alternatives to the standard method, failure to evaluate these methods on genetic data generated by anything but the paired-end, short-read Illumina platform, and the nongeneralizability of the calculated PPV and NPV in prostate cancer and melanoma patients.
-
Automated deep learning technology may have a higher sensitivity and specificity in detecting pathogenic genetic variants in patients with prostate cancer and melanoma who are at a high risk of disease recurrence compared with current standard testing methods.
-
In patients with prostate cancer and melanoma, the number of cancer susceptibility genes identified with germline genetic testing may partially depend on the automated approach used to analyze sequence data.
E.C. Meszaros, Contributing Writer, BreakingMED™
This study was supported by awards or grants from the American Society of Clinical Oncology, the Prostate Cancer Foundation, the PCF-V Foundation the Mark foundation, the National Institutes of Health, and King Abdulaziz City.
AlDubayan and Feero reported no conflicts of interest.
Cat ID: 25
Topic ID: 78,25,496,497,25,26,192,925
