Identifying genomic risk factors for neurodevelopmental disorders using machine learning

Neurodevelopmental disorders (NDDs) are a complex grouping of conditions arising in childhood relating to altered development and function of the brain. The primary conditions classified as NDDs include autism spectrum disorder (ASD), intellectual disability, attention deficit hyperactivity disorder, as well as motor, communication, and specific learning disorders. Many NDDs are known to have significant genetic risk, but the particular genes and molecular pathways controlling this genetic risk are still poorly understood. In addition to the genetic etiology of NDDs themselves, understanding the role of genetics in commonly associated comorbidities, such as sleep dysfunction or epilepsy in ASD, and how these insights might be leveraged to develop new therapeutics, remains a central goal of NDD genetic research. In ASD in particular, mutations in more than 100 genes have been significantly linked to increased risk for ASD. However, projections based on the frequency of mutations in these known risk genes has suggested that over 1000 genes may significantly increase risk for ASD when mutated. In response to this prediction, several machine learning approaches have been developed to use genome-wide data sources to predict which genes are the best candidates for ASD risk gene discovery. However, with different sources of data and training strategies used for each of these scores, there is not a clear consensus in the community on the most important predictors of genetic risk. My work develops a new ASD risk gene score that combines the benefits of all prior scores through a machine learning approach called ”ensemble learning”, unifying the previous scores while providing additional genome-wide data sources for model training. By comparing the previous scores with my work, I demonstrate the effectiveness of ensemble learning in this setting, and provide an ASD risk gene score that is enriched across a variety of ASD genetic data domains, such as common variant risk and gene expression data. While ASD as a whole has many known genetic associations, differences in medical issues experienced by those with ASD are highly variable, and the genetic factors underlying these comorbidities remain unclear. For instance, more than 70% of individuals with ASD have issues with sleep, but it is unknown whether genetic changes explain this difference seen between individuals with of ASD. Simply put, we know that genetics plays a large role in ASD, but we do not know the specifics of how genes map to subtypes of ASD. My work bridges this gap by studying the genetics of sleep dysfunction within individuals with ASD. To my knowledge, I am the first to be able to demonstrate and report that sleep dysfunction in ASD has a significant genetic component. Further, I find that genetic risk for ADHD, BMI, and several other conditions heightens an autistic individual’s risk for having issues with sleep. This work also uncovers associations between the type of sleep issue an individual has and the drugs that may be most effective for restoring normal sleep. Another major medical issue faced by individuals with ASD is epilepsy, with over 20% of individuals diagnosed with ASD having or going on to develop epilepsy later in life. Similar to sleep issues in ASD, treatment options in epilepsy are often effective but fall short in approximately 30% of cases. Finding treatments for these individuals who fail to find relief from the standard of care options is of critical importance. My work uses a bioinformatic technique called drug repositioning to computationally prioritize drugs that may be capable of reversing the transcriptional state induced by epilepsy. This approach yielded 184 potential therapeutic compounds, of which 4 were selected and tested in a zebrafish model of epilepsy. Three of the four compounds showed significant seizure suppression activity, including one with no previous literature surrounding its use in epilepsy (pyrantel tartrate). While a diverse set of work, the common thread is leveraging computational genetic techniques to better understand the causes, symptoms, and treatments of neurodevelopmental and associated disorders. By using ensemble learning, this work establishes a unified autism risk gene score that effectively summarizes a gene’s level of association with autism. Through studying sleep issues in ASD, I find a significant role for common variant risk and establish several genetic associations for poor sleep in ASD, such as ADHD and BMI genetic risk factors. Lastly, by using gene expression to model an effective therapeutic for epilepsy, this work reports on the first possible use of pyrantel tartrate in the treatment of epilepsy. Taken together, these findings demonstrate the power of leveraging big genetic datasets and innovative techniques in order to understand complex disease.