Genome-wide prediction of dominant and recessive neurodevelopmental disorder risk genes

Abstract

Despite great progress in the identification of neurodevelopmental disorder (NDD) risk genes, there are thousands that remain to be discovered. Computational tools that provide accurate gene-level predictions of NDD risk can significantly reduce the costs and time needed to prioritize and discover novel NDD risk genes. Here, we first demonstrate that machine learning models trained solely on single-cell RNA-sequencing data from the developing human cortex can robustly predict genes implicated in autism spectrum disorder (ASD), developmental and epileptic encephalopathy (DEE), and developmental delay (DD). Strikingly, we find differences in gene expression patterns of genes with monoallelic and biallelic inheritance patterns. We then integrate these expression data with 300 orthogonal features in a semi-supervised machine learning framework (mantis-ml) to train inheritance-specific models for ASD, DEE, and DD. The models have high predictive power (AUCs: 0.84 to 0.95) and top-ranked genes were up to two-fold (monoallelic models) and six-fold (biallelic models) more enriched for high-confidence NDD risk genes than genic intolerance metrics. Across all models, genes in the top decile of predicted risk genes were 60 to 130 times more likely to have publications strongly linking them to the phenotype of interest in PubMed compared to the bottom decile. Collectively, this work provides highly robust novel NDD risk gene predictions that can complement large-scale gene discovery efforts and underscores the importance of incorporating inheritance into gene risk prediction tools (https://nddgenes.com).

Phenome-wide identification of therapeutic genetic targets, leveraging knowledge graphs, graph neural networks, and UK Biobank data

“…To expose deeper relationships within the knowledge graph, we trained a graph neural network (GNN) ( 9 , 10 ) to propagate features to neighboring nodes such that the resulting feature vectors represent spatial proximity among nodes in the knowledge graph. We also now include additional single-cell transcriptomic data from a recent study ( 11 ).…”

“…Here, we use single-cell transcriptomics data as gene features ( 11 ). The single-cell transcriptomics is summarized as the average expression value of each gene across the available cell populations.…”

Essential genes: a cross-species perspective

“…Machine learning approaches integrating gene expression across development in relevant tissues among other features are providing prioritised sets of genes to be associated with developmental disorders (Dhindsa et al 2022 ). The Fetal Sequencing Consortium and specific sequencing programmes aimed at providing a molecular diagnosis for this type of disorders (e.g.…”

Lethal phenotypes in Mendelian disorders

“…Gene and variant prioritisation strategies leveraging this information have been successful in identifying novel neurodevelopmental disease genes 33,34 . Similarly, together with other metrics of intolerance to LoF, information on mouse essential genes is a highly predictive feature in machine learning implementations to predict disease risk genes 35 .…”