Dissecting Autism Genetic Risk Using Single-cell RNA-seq Data

Siying Chen, Xueya Zhou, Eve Byington, Samuel L. Bruce, Haicang Zhang, Yufeng ShenbioRxiv 2020 · DOI

TL;DR

This paper introduces "A-risk," a novel machine learning approach that leverages single-cell RNA sequencing data from developing human brains to predict autism risk genes. A-risk significantly improves the identification of potential autism risk genes, particularly those with less severe genetic implications, and sheds light on the critical role of deep-layer excitatory neurons in autism etiology.

What Problem Does This Paper Address?

Autism spectrum disorder (ASD) has a strong genetic component, but current genetic studies have only identified a fraction of the genes contributing to ASD risk. Many identified de novo variants (new mutations) in individuals with ASD are not explained by known risk genes, suggesting a vast number of unknown risk genes. The challenge lies in efficiently and accurately identifying these novel risk genes from the massive amount of genetic data being generated. Previous approaches often lacked the resolution to pinpoint specific cell types or developmental stages crucial for understanding ASD pathogenesis.

Methods

The core of this work is the development of a supervised machine learning model called "A-risk" (Autism risk). A-risk uses a gradient boosting tree algorithm trained on single-cell RNA sequencing (scRNA-seq) data from human fetal midbrain and prefrontal cortex. The model learns the cell-type-specific expression patterns of known autism risk genes and then predicts the "plausibility" of any gene being an autism risk gene based on its own expression profile.

Key methodological aspects include:

  • scRNA-seq Integration: Combining data from multiple scRNA-seq studies to create a comprehensive feature set of gene expression across various cell types and developmental time points.
  • Gradient Boosting: A powerful ensemble learning technique used for its ability to handle complex relationships and provide feature importance scores.
  • Empirical Bayesian Framework (extTADA): A-risk scores are integrated into an existing Bayesian framework (extTADA) to refine prior probability estimates for genes, improving the identification of candidate risk genes. This stratification by both A-risk and gene constraint (pLI) is a novel application.

Datasets

  • Training Data:
    • Known Autism Risk Genes: 88 genes with SFARI gene scores of 1 or 2.
  • Negative Control Genes: 154 genes with at least one de novo likely gene-disrupting (LGD) variant in control individuals from exome-sequencing studies.
  • Evaluation Data:
    • De novo LGD and deleterious missense (D-mis) variants from 8838 trios from multiple published exome sequencing studies.
  • Single-nucleus RNA-seq data from adult human cortex (middle temporal gyrus) to investigate expression patterns in mature neurons.

Key Insights and Ideas

  • Cell-Type Specificity is Key: The study strongly supports the hypothesis that autism risk genes share common cell-type-specific expression patterns during brain development. This is a significant departure from approaches that might overlook such fine-grained biological context.
  • Deep-Layer Excitatory Neurons are Critical: The research highlights deep-layer excitatory neurons in the adult human cortex as a particularly important cell type implicated in autism etiology. This finding offers a more precise target for understanding the biological mechanisms of ASD.
  • A-risk Complements Gene Constraint: Combining A-risk with mutation intolerance metrics (like pLI) provides a more robust way to estimate prior probabilities for genes, leading to the discovery of more candidate risk genes than using either metric alone.
  • Prioritizing "Less Constrained" Genes: A-risk is particularly effective at prioritizing genes that are less intolerant to loss-of-function variants, suggesting it can uncover risk genes that might be missed by methods solely focused on highly constrained genes.

Key Results

  • Improved Prioritization of de novo Variants: A-risk achieved a significantly higher enrichment rate (2.1) for de novo LGD variants in candidate risk genes compared to a baseline of 1.4 for all genes excluding known risk genes.
  • Identification of Novel Risk Genes: The stratified analysis using A-risk and pLI, combined with extTADA, identified 71 novel candidate risk genes, an increase of 27 genes compared to methods without stratification.
  • Superior Performance: A-risk demonstrated superior performance in enrichment and precision compared to mouse brain expression levels and other existing metrics (D-score, Krishnan 2016) when prioritizing de novo variants, especially within the top 2000-4000 ranked genes.
  • Specific Cell Type Contributions: Feature importance analysis revealed that GABAergic neurons in the midbrain (week 9), dopaminergic neurons (week 10), and prefrontal cortex excitatory neurons (week 12) were the most significant contributors to A-risk prediction.

Limitations and Future Work

  • Data Availability: The study acknowledges limitations due to the availability of comprehensive single-cell expression profiles across all critical brain regions and developmental stages. Information from extended neuronal projections and axons can be lost in current scRNA-seq techniques.
  • Supervised Learning Bias: As a supervised learning method, A-risk is inherently biased towards genes with expression patterns similar to the known risk genes used in training.
  • Expression vs. Risk: Abundant and specific gene expression alone is not sufficient to define a gene as a risk gene; other factors like functional redundancy and protein complex formation also play a role.

Future work suggested by the authors includes:

  • New Profiling Techniques: Leveraging new single-nucleus RNA-seq and subcellular transcriptomic profiling techniques to capture more comprehensive cellular information.
  • Unsupervised Approaches: Exploring unsupervised methods to mitigate the bias inherent in supervised learning.
  • Broader Phenotypic Data: Incorporating more comprehensive neurodevelopmental disorder (NDD) phenotyping in autism cohorts to better understand pleiotropy.
  • Integrating Other Factors: Considering functional redundancy and protein complex formation in future risk gene prediction models.

Summary updated

Abstract

Abstract Autism spectrum disorder (autism) is a condition with strong but heterogenous genetic contribution. Recent exome and genome sequencing studies have uncovered many new risk genes through de novo variants. However, a large fraction of enrichment of de novo variants observed in cases are not accounted for by known or candidate risk genes, suggesting that the majority of risk genes are still unknown. Here we hypothesize that autism risk genes share a few common cell-type specific gene expression patterns during brain development, and such information can be quantified to improve statistical power of detecting new risk genes. We obtained large-scale single-cell RNA-seq data from human fetal brain collected through a range of developmental stages, and developed a supervised machine-learning approach “A-risk” (Autism risk), to predict the plausibility of autism risk genes across the genome. Using data from recent exome sequencing studies of autism, A-risk achieves better performance in prioritizing de novo variants than other methods, especially for genes that are less intolerant of loss of function variants. We stratified genes based on A-risk and mutation intolerance metrics to improve estimation of priors in extTADA and identified 71 candidate risk genes. In particular, CLCN4, PRKAR1B, and NR2F1 are potentially new risk genes with further support from neurodevelopmental disorders. Expression patterns of both known and candidate risk genes reveals the important role of deep-layer excitatory neurons from adult human cortex in autism etiology. With the unprecedented revolution of single-cell transcriptomics and expanding autism cohorts with exome or genome sequencing, our method will facilitate systematic discovery of novel risk genes and understanding of biological pathogenesis in autism.
BibTeX
@misc{chen2020dissecting,
  author = {Siying Chen and Xueya Zhou and Eve Byington and Samuel L. Bruce and Haicang Zhang and Yufeng Shen},
  title = {{Dissecting Autism Genetic Risk Using Single-cell RNA-seq Data}},
  year = {2020},
  doi = {10.1101/2020.06.15.153031},
  howpublished = {bioRxiv},
}
Figure