This makes it easier to determine whether the intellectual disability is caused by a genetic mutation, says researcher Christian Gilissen. “With this we can help parents in a number of areas in the care process for their children.”
As professor of genome bioinformatics at Radboudumc, Gilissen is involved in the analysis of genetic data. As a bioinformatician, he is active in the daily practice of patient diagnostics in the genetic center of the Nijmegen university hospital.
Rare Conditions
Some of the diagnoses concern rare disorders that lead to intellectual disabilities as a result of de novo mutations in the DNA. They arise spontaneously and are not passed on by one or both of a patient’s parents – usually a young child. It concerns about 80 to 100 mutations on approximately three billion base pairs, a small number.
Gilissen: “We know that these are often the cause of developmental delays or intellectual disabilities. We normally do this by comparing genetic data from parents and child, but this is very difficult. Often there is insufficient information, or there are errors in the technology we use. The result is a lot of false positives or missed mutations.”
Visual inspection genetic data
The diagnoses of the genetic center offer several advantages: access to care, more clarity about a condition and its cause, and possibilities for prevention. But couldn’t it be faster, and with fewer misdiagnoses? Two years ago, Gilissen and a few colleagues started working on these questions. They developed an algorithm – based on deep learning – that mimics a visual inspection of the genetic data.
Such visual material can be analyzed very well using deep learning, says Gilissen. To feed the algorithm, the genetics department of Radboudumc compiled cohorts of detected genetic mutations that were correct and those that were also incorrect on the basis of its own data, so that the algorithm could learn the difference.
“We also fed other, external datasets to the algorithm that were generated using other methods. This showed that our method generated fewer false positives. The algorithm also turned out to be robust enough to correctly analyze the data from other DNA readers or obtained through other enrichment techniques.”
Teething problems from algorithm
Gilissen and his team spent two years developing the algorithm. According to him, this was badly needed to eliminate all teething problems, to prevent biases from arising on the basis of the deep learning data, to add training data where the algorithm was not yet able to detect mutations very well. Currently, work is mainly being done on refinements and making other applications possible on the basis of fieldwork with other groups.
Knowing more? Read the entire interview with Christian Gilissen in ICT&health 4, which will be published on 26 August.