Estimating heritability and genetic correlations from large health datasets in the absence of genetic data

Gengjie Jia, Yu Li, Hanxin Zhang, Ishanu Chattopadhyay, Anders Boeck Jensen, David R. Blair, Lea Davis, Peter N. Robinson, Torsten Dahlén, Søren Brunak, Mikael Benson, Gustaf Edgren, Nancy J. Cox, Xin Gao, Andrey RzhetskyNature Communications 2019 · OpenReview

Abstract

Typically, estimating genetic parameters, such as disease heritability and between-disease genetic correlations, demands large datasets containing all relevant phenotypic measures and detailed knowledge of family relationships or, alternatively, genotypic and phenotypic data for numerous unrelated individuals. Here, we suggest an alternative, efficient estimation approach through the construction of two disease metrics from large health datasets: temporal disease prevalence curves and low-dimensional disease embeddings. We present eleven thousand heritability estimates corresponding to five study types: twins, traditional family studies, health records-based family studies, single nucleotide polymorphisms, and polygenic risk scores. We also compute over six hundred thousand estimates of genetic, environmental and phenotypic correlations. Furthermore, we find that: (1) disease curve shapes cluster into five general patterns; (2) early-onset diseases tend to have lower prevalence than late-onset diseases; and (3) the disease onset age and heritability are negatively correlated. Disease heritability and genetic correlations between traits depend on genetics, the environment and their interaction. Here, Jia et al. compute disease prevalence curves and disease embeddings from electronic health records and impute heritability for hundreds of diseases and genetic correlations for thousands of disease pairs.

Introduction: The Bottleneck of Genetic Data

Estimating genetic parameters—like heritability ($h^2$) and genetic correlation ($corr$)—is the bedrock of understanding disease risk. However, traditional "gold standard" methods are often bottlenecked by data availability. Twin and family studies require rare datasets with detailed kinship information, while GWAS and SNP-based studies require massive cohorts of genotyped individuals, which are expensive and often lack diverse representation.

Meanwhile, researchers are sitting on a gold mine of "phenotypic" data in the form of Electronic Health Records (EHR). The challenge is that EHR data is inherently noisy, designed for insurance billing rather than scientific inquiry. This study, published in Nature Communications, asks a provocative question: Can we extract enough biological signal from the "shape" of a disease's manifestation in a population to infer its underlying genetic architecture?

Methodology: EHR as a Biological Proxy

The authors proposed a "proxy" approach using two novel disease metrics derived from massive national-scale health registries in the United States, Denmark, and Sweden:

  1. Temporal Prevalence Curves: They mapped how the prevalence of over 500 diseases changes across age and sex. By measuring the similarity between these curves, they identified distinct patterns of disease manifestation.
  2. Disease Embeddings: Borrowing from Natural Language Processing (NLP), they used the word2vec algorithm to treat a patient’s chronological sequence of diagnoses as a "sentence." This created a 20-dimensional vector for each disease, where diseases that frequently co-occur (comorbidities) are positioned closer together in mathematical space.
Temporal disease prevalence curves and clustering into five distinct shape archetypes.

Using these metrics as features, the team trained a Gradient Boosting Regression model on a "gold standard" training set of ~3,000 previously published heritability and correlation estimates mined from decades of literature.

Key Insights: The Five Archetypes and Evolutionary Selection

A profound insight from this research is that the way a disease manifests over time and its "social circle" of comorbidities are incredibly strong predictors of its genetic roots. Prevalence curves alone accounted for approximately 45% of the model's predictive power.

As shown in the figure above, disease trajectories cluster into five patterns (archetypes), such as "L-shaped" (early-onset, like congenital issues) or "reversed L-shaped" (late-onset, like neurodegeneration). Furthermore, the authors found a significant negative correlation ($\rho = -0.46$) between onset age and heritability. Essentially, diseases that strike early in life tend to be more heritable but less prevalent—likely because evolution aggressively selects against high-prevalence, highly heritable traits that impair reproduction.

Results: Unprecedented Scale and Accuracy

The model’s predictions showed a Pearson correlation of 0.87 with previously published "gold standard" estimates for both heritability and genetic correlations. This high accuracy allowed the researchers to fill in the "missing pieces" of the human pathology map, generating 11,230 heritability estimates and 621,434 correlation estimates—orders of magnitude more than currently exist in the literature.

Despite the massive differences between the US, Danish, and Swedish healthcare systems, the heritability estimates remained remarkably consistent. This suggests the model is capturing fundamental human biology rather than administrative or regional quirks.

Limitations and Future Outlook

While groundbreaking, the study has limitations. The US dataset was limited to patients under 65, which may skew results for geriatric-specific conditions. Additionally, while the model predicts genetic correlation, it doesn't explicitly disentangle shared environmental factors from shared genetic factors as cleanly as controlled molecular studies. Future work is needed to see if these "prevalence archetypes" hold true in the Global South or in populations with different genetic backgrounds. To aid further research, the authors have provided a searchable web application for exploring these prevalence curves.

Summary updated

Abstract

Typically, estimating genetic parameters, such as disease heritability and between-disease genetic correlations, demands large datasets containing all relevant phenotypic measures and detailed knowledge of family relationships or, alternatively, genotypic and phenotypic data for numerous unrelated individuals. Here, we suggest an alternative, efficient estimation approach through the construction of two disease metrics from large health datasets: temporal disease prevalence curves and low-dimensional disease embeddings. We present eleven thousand heritability estimates corresponding to five study types: twins, traditional family studies, health records-based family studies, single nucleotide polymorphisms, and polygenic risk scores. We also compute over six hundred thousand estimates of genetic, environmental and phenotypic correlations. Furthermore, we find that: (1) disease curve shapes cluster into five general patterns; (2) early-onset diseases tend to have lower prevalence than late-onset diseases (Spearman’s ρ = 0.32, p < 10–16); and (3) the disease onset age and heritability are negatively correlated (ρ = −0.46, p < 10–16). Disease heritability and genetic correlations between traits depend on genetics, the environment and their interaction. Here, Jia et al. compute disease prevalence curves and disease embeddings from electronic health records and impute heritability for hundreds of diseases and genetic correlations for thousands of disease pairs.
BibTeX
@inproceedings{jia2019estimating,
  author = {Gengjie Jia and Yu Li and Hanxin Zhang and Ishanu Chattopadhyay and Anders Boeck Jensen and David R. Blair and Lea Davis and Peter N. Robinson and Torsten Dahlén and Søren Brunak and Mikael Benson and Gustaf Edgren and Nancy J. Cox and Xin Gao and Andrey Rzhetsky},
  title = {{Estimating heritability and genetic correlations from large health datasets in the absence of genetic data}},
  booktitle = {Nature Communications},
  year = {2019},
  url = {https://openreview.net/forum?id=3HkGyx6vbu},
}
Figure