DeepND

DeepND applies graph convolutional neural networks and multitask learning on gene coexpression networks of human brain development to discover and prioritize risk genes for Autism Spectrum Disorder (ASD) and Intellectual Disability (ID).

Key Features:

• Graph convolutional neural networks (GCNNs): Applies GCNNs to gene coexpression networks modeling human brain development to extract gene–gene associations.
• Mixture-of-experts model: Self-learns critical neurodevelopmental time windows and brain-region–specific contributions for interpretability.
• Multitask learning framework: Simultaneously models shared and disorder-specific genetic features to improve prediction in single-disorder and cross-disorder settings.
• Genomewide risk ranking: Produces genomewide risk rankings for each disorder.
• Spatiotemporal risk identification: Identifies the prefrontal cortex and primary motor-somatosensory cortex and neurodevelopmental windows from early fetal to mid-fetal and from early childhood to young adulthood as high-risk regions and periods.
• Copy number variation (CNV) analysis: Investigates frequent CNV regions and nominates susceptibility gene candidates.
• Generalizability: Applicable to analyses of other combinations of comorbid neurodevelopmental disorders.

Scientific Applications:

• Cross-disorder gene discovery: Prioritizes and ranks risk genes genomewide for ASD and ID.
• Spatiotemporal etiology mapping: Identifies neurodevelopmental time windows and brain regions implicated in disorder etiology, including prefrontal and primary motor-somatosensory cortices and specified fetal-to-adult windows.
• CNV candidate prioritization: Prioritizes candidate susceptibility genes within frequent CNV regions associated with ASD and ID.
• Comorbidity analysis: Enables joint analysis of shared and disorder-specific genetic architecture across comorbid neurodevelopmental disorders.

Methodology:

Implements graph convolutional neural networks (GCNNs) on gene coexpression networks of human brain development, integrates a mixture-of-experts model to learn neurodevelopmental time windows and brain-region contributions, employs multitask learning to jointly predict ASD and ID risk and produce genomewide risk rankings, and analyzes frequent copy number variation (CNV) regions to nominate candidate susceptibility genes.