Noninvasive prenatal diagnosis using deep learning

Noninvasive prenatal diagnosis (NIPD) is a risk-free method to detect mutations in the fetus
during pregnancy. It is based on the presence of fetal-derived cell-free DNA (cfDNA) in
maternal plasma, and is already clinically used for diagnosing large chromosomal
abnormalities. However, other much smaller fetal genetic abnormities are extremely
challenging to detect. Recently, we have developed Hoobari, the first noninvasive fetal
variant caller, which enables accurate genotyping of fetal point mutations and small
insertions and deletions. Hoobari was inspired by common pipelines for small variant
calling, and hence, was based on a Bayesian algorithm followed by a machine learning step.
In the present study, we introduce DeepHoobari, the first deep learning-based algorithm for
NIPD. Using a dedicated convolutional neural network (CNN) that utilizes parental and cell-
free DNA, we managed to achieve higher accuracy than before. Our results demonstrate
that deep learning can be used for noninvasive fetal variant calling, and lead to improved
detection of fetal mutations. We believe our new approach would bring NIPD of any genetic
abnormities, large and small, closer to the clinic.