Explaining the genetic causality for complex diseases via deep association kernel learning

Feng Bao, Yue Deng, Mulong Du, Zhiquan Ren, Sen Wan, Junyi Xin, Feng Chen, David C. Christiani, Meilin Wang, Qionghai DaibioRxiv 2019 · DOI

TL;DR

This paper introduces Deep Association Kernel learning (DAK), a novel deep learning framework for Genome-Wide Association Studies (GWAS) that moves beyond linear assumptions to capture complex genetic effects. DAK effectively identifies disease-associated pathways, outperforming existing methods, especially for rare variants and non-additive genetic models.

What Problem Does This Paper Address?

Genome-Wide Association Studies (GWAS) are crucial for understanding the genetic basis of complex diseases. However, traditional GWAS methods often rely on a linear assumption, meaning they assume a simple, additive relationship between genetic variants and disease risk. This limitation hinders their ability to detect more complex genetic effects, such as recessive or multiplicative interactions, which are common in real-world biological systems. Furthermore, existing pathway-based GWAS methods often require pre-defined genetic models, which may not accurately reflect the unknown biological reality. This paper addresses the need for a more sophisticated and flexible GWAS framework that can uncover complex genetic causality without prior assumptions about genetic models.

Methods

The core innovation is the Deep Association Kernel learning (DAK) framework. DAK employs a multi-stage deep learning architecture:

  • One-Hot Encoding: Genotypes for each Single Nucleotide Polymorphism (SNP) are encoded into a one-hot vector, allowing for flexible representation of different genotypes (e.g., homozygous, heterozygous).
  • Pathway Grouping: SNPs are grouped into biologically meaningful pathways (e.g., KEGG pathways).
  • Convolutional Layers: Convolutional layers are used to extract "deep features" from these pathway-level genotype representations. This is a key departure from traditional methods, as it allows the model to learn hierarchical representations of genetic information within pathways.
  • Kernel Machine Regression: The extracted deep features are then fed into a kernel machine regression layer. This layer is used to predict disease status and, importantly, to perform statistical significance tests for association.
  • Multiple Instance Learning: The entire framework is trained end-to-end using a multiple instance learning loss. This approach treats an individual as having the disease if any of their pathways are associated with the disease, which is a more realistic representation than assuming individual pathway labels.
  • End-to-End Optimization: All model parameters, including those in the convolutional and regression layers, are learned in a data-driven manner through backpropagation, eliminating the need for pre-specified genetic models.

Datasets

The paper utilizes both simulated and real-world datasets:

  • Simulated Datasets: Phenotypes were simulated under various genetic models (additive, dominant, recessive, multiplicative, heterozygous) and effect sizes, with varying sample sizes (3,000 and 5,000). These simulations were crucial for evaluating the performance of DAK against other methods under controlled conditions, particularly for rare variants and complex genetic effects.
  • Real-World Datasets: DAK was applied to four publicly available GWAS datasets:
  • Gastric Cancer
    • Colorectal Cancer
    • Lung Cancer
    • Schizophrenia

Key Insights and Ideas

  • Beyond Linearity: The most significant insight is DAK's ability to move beyond linear assumptions, which are a major bottleneck in current GWAS. This allows for the detection of complex genetic interactions that are often missed by traditional methods.
  • Data-Driven Model Discovery: By learning genotype encoding and association tests in an end-to-end fashion, DAK can discover novel genetic associations without relying on potentially inaccurate pre-defined biological models.
  • Superior Performance with Complex Models: DAK demonstrates a remarkable advantage in identifying disease pathways when genetic effects are non-additive (e.g., recessive, multiplicative) or when dealing with rare variants, areas where other methods struggle significantly.
  • Pathway-Level Interpretation: The framework's focus on pathway-level analysis provides a more interpretable view of genetic causality, linking individual genetic variants to broader biological processes.

Key Results

  • Robustness Across Genetic Models: DAK consistently outperformed seven other representative gene/pathway-based methods across a wide range of genetic models, including additive, multiplicative, dominant, recessive, and heterozygous effects.
  • Enhanced Power for Rare Variants: DAK showed significantly higher performance in detecting associations with rare variants, especially under recessive and multiplicative genetic models, where other methods often failed.
  • Real-World Pathway Discovery: When applied to real disease datasets, DAK identified several pathways with genome-wide significance. Notably, for gastric cancer, it highlighted "Terpenoid backbone biosynthesis" and "oxidative phosphorylation," and for colorectal cancer, "glyoxylate and dicarboxylate metabolism," many of which have supporting biological evidence or suggest novel avenues for research.
  • Type-I Error Control: In simulation studies under a null hypothesis (no causal loci), DAK exhibited low Type-I error rates, indicating good control over false positives.

Limitations and Future Work

  • Computational Cost: While not explicitly detailed as a limitation, deep learning models can be computationally intensive. The paper mentions using a Nvidia Titan X GPU for training, suggesting that significant computational resources are required.
  • Interpretability of Deep Features: While DAK provides pathway-level insights, the "deep features" learned by the convolutional layers might be challenging to interpret biologically in isolation.
  • Future Directions: The paper suggests that DAK's ability to discover new potential causal pathways opens doors for further biological investigation and validation. The framework's flexibility could also be extended to incorporate other types of biological data (e.g., epigenetics, transcriptomics) for a more comprehensive understanding of disease etiology.

Summary updated

Abstract

ABSTRACT The genetic effect explains the causality from genetic mutation to the development of complex diseases. Existing genome-wide association study (GWAS) approaches are always built under a linear assumption, restricting their generalization in dissecting complicated causality such as the recessive genetic effect. Therefore, a sophisticated and general GWAS model that can work with different types of genetic effects is highly desired. Here, we introduce a Deep Association Kernel learning (DAK) model to enable automatic causal genotype encoding for GWAS at pathway level. DAK can detect both common and rare variants with complicated genetic effects that existing approaches fail. When applied to real-world GWAS data, our approach discovered potential casual pathways that could be explained by alternative biological studies.
BibTeX
@misc{bao2019explaining,
  author = {Feng Bao and Yue Deng and Mulong Du and Zhiquan Ren and Sen Wan and Junyi Xin and Feng Chen and David C. Christiani and Meilin Wang and Qionghai Dai},
  title = {{Explaining the genetic causality for complex diseases <i>via</i> deep association kernel learning}},
  year = {2019},
  doi = {10.1101/2019.12.17.879866},
  howpublished = {bioRxiv},
}
Figure