ESPE Abstracts

The majority of patients with Variations of Sex Development (VSD) lack a genetic diagnosis. Patients that are born with atypical chromosomal, gonadal, or phenotypical sex, present a wide spectrum of phenotypes that are often associated with ambiguous genitalia, infertility as well as increased susceptibility to testicular or ovarian cancer. Many different genetic causes of VSD have been reported [2], but for more than 50% [3] of all VSD patients, the molecular cause of their condition remains unknown, further complicating clinical management. The low rate of diagnosis is explained by the still incomplete knowledge of the molecular mechanisms of gonad morphogenesis and sex development. Recently, next-generation sequencing (NGS) enhanced clinical diagnosis but the value of NGS relies on the ability to analyze the gigantic amount of data in a reasonable and time-efficient manner. After the exclusion of known VSD causes and standard variant filtering methods, about 1000 variants per patient remain for further analysis. To accelerate and specify variant interpretation, we combine RNA single-cell sequencing (scRNA-seq) data of the developing human gonads [4] with known gene interaction networks (STRING) and deep learning methods to develop an encoding scheme for gene expression data. We use a convolutional neural network (CNN) structure to develop our model. To this end, we transformed single-cell 1D expression data into a gene interaction co-expression image-like 2D object [5]. With this data, we trained our deep learning structure to learn the gene expression network landscape of the developing gonads. Next, the model was retrained with the same data structure but specific for VSD. To evaluate the performance of the VSD trained model, we used 3-fold cross-validation and measured the overall area under the precision-recall curve. After training, we validated the accuracy of the prediction of known VSD genes which reached 85%. Applying our trained model to the remaining variants in VSD patients we were able to rank them from predicted to be “highly” to “less likely” VSD-related candidates. By selecting the top 10 genes from this ranked list, we were able to identify several genes that are tightly involved in sex development disorders and potentially new VSD candidate genes. This translational approach advances our knowledge of human sex development and potentially improves the diagnosis and management of its variations.