Genetic association studies using disease liabilities from deep neural networks

Lu Yang, Marie C. Sadler, Russ B. AltmanmedRxiv 2023 · DOI

TL;DR

Traditional genetic studies often treat disease as a simple "yes/no" binary, but health is a dynamic spectrum. By using deep learning to transform high-dimensional patient data into continuous "disease liability" scores, researchers uncovered significantly more genetic risk factors and improved the prediction of future diagnoses compared to standard methods.

What Problem Does This Paper Address?

The standard tool for finding genetic variants associated with disease is the Case-Control Genome-Wide Association Study (CC-GWAS). You take a group of people with a condition (cases) and a group without it (controls) and look for DNA differences.

However, this binary approach has a major flaw: it’s a snapshot in time. In long-term studies like the UK Biobank, a "control" today might become a "case" tomorrow. This "misclassification" of pre-symptomatic individuals as healthy controls weakens the statistical power of the study. Furthermore, binary labels ignore the complexity of disease progression. The authors aim to bridge this gap by treating disease risk as a continuous probability—a "liability"—rather than a toggle switch.

Methods

The researchers leveraged a deep learning framework called POPDx.

  1. AI Phenotyping: POPDx was trained on a massive array of data from the UK Biobank, including electronic medical records (ICD codes), laboratory results, physical measurements, lifestyle questionnaires, and family history.
  2. Liability Scoring: Instead of a binary label, the model outputs a continuous probability (from 0 to 1) for 1,538 different "Phecodes" (standardized disease groupings).
  3. GWAS Comparison: The team conducted GWAS using three different phenotype definitions for 44 common traits:
  4. Standard CC: Binary (0 or 1).
  5. AI-based Liability: The raw probability score from the neural network.
  6. Combined CC-Liability: A hybrid where known cases are 1, and controls are assigned their AI-calculated probability.
  7. Validation: They estimated SNP-based heritability and calculated Polygenic Risk Scores (PRS) to see which method better predicted "new" cases that appeared in the 2022 UK Biobank update (who were listed as controls in the 2019 version).

Datasets

  • UK Biobank (UKBB): The primary dataset, containing genetic and phenotypic data for ~500,000 participants. The final analysis focused on 261,807 unrelated individuals of white British ancestry.
  • Temporal Split: The authors used the 2019 release for initial modeling and the 2022 release to identify "emerging cases"—individuals who were healthy in 2019 but diagnosed by 2022—providing a unique real-world test for their predictive models.

Key Insights and Ideas

  • Capturing the "Hidden" Signal: The AI model can "see" risk in laboratory values and lifestyle factors before a clinical diagnosis is ever recorded. By using these probabilities in a GWAS, researchers can extract genetic signals from individuals who are technically "controls" but genetically and phenotypically predisposed to the disease.
  • Boosting Statistical Power: Moving from a binary scale to a continuous one effectively increases the "resolution" of the data. This led to a significant increase in the number of independent genetic loci discovered.
  • Heritability Recovery: One of the biggest mysteries in genetics is "missing heritability"—the gap between known genetic influences and what we can actually measure. This study found that AI-based liabilities resulted in higher heritability estimates, suggesting that traditional binary labels have been "masking" the true genetic contribution to disease.

Key Results

  • More Loci Discovered: For essential hypertension, the AI-based method identified 772 significant SNPs across 114 loci, compared to only 299 SNPs in 100 loci using the traditional CC method.
  • Higher Heritability: The average SNP-based heritability ($h^2_{SNP}$) across 44 traits rose from 0.030 (CC) to 0.040 (AI-based).
  • Better Future Prediction: Polygenic Risk Scores (PRS) derived from AI liabilities were significantly better at identifying individuals who would go on to develop a disease. On average, there was a 6.2% increase in percentile rank for newly diagnosed cases when using the AI-informed PRS versus the standard version.
  • Novel Gene Associations: The study identified several putative susceptibility genes missed by standard GWAS, such as ABCG8 for angina pectoris and ZC3H4 for Type 2 Diabetes.

Limitations and Future Work

  • Noise in Phenotyping: AI models are not perfect; if the underlying phenotyping is "noisy" or incorrect, it can lead to inaccurate liability scores that impair genetic analysis.
  • Data Requirements: The POPDx framework requires rich, multi-modal data (labs, lifestyle, etc.). This makes the method highly effective for "rich" biobanks like the UKBB, but it may be difficult to apply to smaller clinical studies that only have basic diagnostic codes.
  • Computational Constraints: This study limited its analysis to directly genotyped SNPs. The authors suggest that future work should include imputed variants to provide an even more comprehensive genetic map.
  • Generalization: The pipeline needs to be validated in other large-scale cohorts, such as the "All of Us" program or the Million Veteran Program, to ensure the findings hold across diverse populations.

Summary updated

Abstract

The case-control study is a widely used method for investigating the genetic underpinnings of binary traits. However, long-term, prospective cohort studies often grapple with absent or evolving health-related outcomes. Here, we propose two methods, <i>liability</i> and <i>meta</i>, for conducting genome-wide association study (GWAS) that leverage disease liabilities calculated from deep patient phenotyping. Analyzing 38 common traits in ~300,000 UK Biobank participants, we identified an increased number of loci compared to the conventional case-control approach, with high replication rates in larger external GWAS. Further analyses confirmed the disease-specificity of the genetic architecture with the meta method demonstrating higher robustness when phenotypes were imputed with low accuracy. Additionally, polygenic risk scores based on disease liabilities more effectively predicted newly diagnosed cases in the 2022 dataset, which were controls in the earlier 2019 dataset. Our findings demonstrate that integrating high-dimensional phenotypic data into deep neural networks enhances genetic association studies while capturing disease-relevant genetic architecture.
BibTeX
@misc{yang2023genetic,
  author = {Lu Yang and Marie C. Sadler and Russ B. Altman},
  title = {{Genetic association studies using disease liabilities from deep neural networks}},
  year = {2023},
  doi = {10.1101/2023.01.18.23284383},
  howpublished = {medRxiv},
}
Figure