DeepND: Deep multitask learning of gene risk for comorbid neurodevelopmental disorders

Area of Science:

• Neurodevelopmental disorders
• Genetics
• Computational biology

Background:

• Autism spectrum disorder (ASD) and intellectual disability (ID) are common neurodevelopmental disorders with shared genetic underpinnings.
• Current genetic studies have identified only a subset of risk genes for ASD and ID, necessitating advanced analytical approaches.

Purpose of the Study:

• To develop and validate DeepND, a novel network-based algorithm for prioritizing gene risk in comorbid neurodevelopmental disorders.
• To leverage multitask learning and graph convolutional networks for cross-disorder analysis of ASD and ID.

Main Methods:

• Utilized human brain gene co-expression networks and graph convolutional networks within a multitask learning framework.
• Developed a network-based gene risk prioritization algorithm named DeepND.
• Analyzed copy-number variation (CNV) regions associated with ASD and ID.

Main Results:

• DeepND improved state-of-the-art prediction for both single and cross-disorder analyses.
• Identified the prefrontal and motor-somatosensory cortex (PFC-MFC) as key brain regions implicated in ASD and ID.
• Highlighted specific neurodevelopmental windows, including early- to mid-fetal stages and childhood to young adulthood, as high-risk periods.

Conclusions:

• DeepND effectively prioritizes genetic risk for comorbid neurodevelopmental disorders by integrating network information and multitask learning.
• The algorithm provides insights into critical spatiotemporal windows and brain regions relevant to ASD and ID etiology.
• DeepND is a generalizable tool applicable to the analysis of various comorbid disorder combinations.