Abstract
Variability of response to medication is a well-known phenomenon determined by both environmental and genetic factors. Understanding the heritable component of drug response is challenging due to small study cohorts and computational limitations. This study leverages 17 years of electronic health records (EHR) from Israel’s largest healthcare provider, encompassing over five million patients. By constructing a six-generation population-size pedigree and applying a Linear Mixed Model-based framework, we estimated the heritability ($h^2$) of glycated hemoglobin (HbA1c%) reduction after metformin treatment. We found heritability measures of 15.9% (95% CI, 1.2%–30.5%) in males and 20.9% (95% CI, 7.5%–34.3%) in females. This work demonstrates that while metformin response has a heritable component, the majority of variation is likely due to other factors, further motivating non-genetic analyses aimed at unraveling metformin’s mechanism of action.
Introduction
Metformin is the gold-standard first-line treatment for Type 2 Diabetes (T2D), yet clinical response varies significantly among patients. Quantifying the heritable component of this variation is crucial for personalized medicine but has traditionally required expensive genome-wide association studies (GWAS) or twin studies, which are often limited by sample size and the difficulty of separating genetic effects from shared environmental factors. This study investigates whether "real-world" data—specifically EHRs combined with national genealogical records—can provide a high-powered, low-cost alternative for estimating heritability.
Methods
Data Source and Cohort Selection
The study utilized the Clalit Health Services database in Israel, covering over five million people. From an initial pool of 721,123 T2D patients, a final cohort of 74,871 patients was selected. These patients met strict criteria, including no prior anti-diabetic medication, sufficient HbA1c measurements, and valid renal function. The dataset included over 1.3 million HbA1c measurements and 323,260 metformin prescriptions.
Pedigree-Based Heritability Estimation
Instead of direct DNA sequencing, the researchers constructed a six-generation population-sized pedigree using national registries. They used the Nadiv R package to calculate an Additive Relationship Matrix (ARM), which estimates the probability of shared alleles based on family relationships. To handle the computational burden of this massive pedigree, the team employed the Sci-LMM framework, which utilizes sparse Cholesky factorization to fit models in hours rather than days.
Phenotyping and Validation
The primary phenotype was the absolute reduction in HbA1c% following metformin initiation, adjusted for covariates like age, BMI, and dosage. As a "sanity check," the model was first used to estimate the heritability of adult height using 11.6 million measurements. Since height is a well-studied heritable trait, it served as a positive control to validate the pedigree-based approach.
Results
Validation via Height
The model estimated the heritability of height at 80.7%, which aligns perfectly with established literature. This result confirmed that the pedigree-based approach was robust enough to capture genetic signals from clinical data without actual DNA samples.
Metformin Response Heritability
The study identified a significant but modest heritable component for metformin response:
- Males: $h^2 = 15.9\%$ (95% CI: 1.2% – 30.5%)
- Females: $h^2 = 20.9\%$ (95% CI: 7.5% – 34.3%)
Results remained consistent after adjusting for pre-treatment HbA1c% in females and when analyzing proportional rather than absolute reduction in HbA1c%. The average absolute reduction in HbA1c% was 0.86%, consistent with clinical expectations. By using a cohort of nearly 75,000 patients, the researchers achieved greater statistical power and tighter confidence intervals than previous genetic-based studies.
Discussion
This research represents the first estimation of drug response heritability using EHR data. The findings suggest that while genetics play a role, approximately 80% of the variation in metformin response is driven by non-genetic factors, such as the gut microbiome, lifestyle, and medication adherence.
The study highlights the potential of EHRs as a powerful resource for biological research. By leveraging population-scale pedigrees, researchers can extract deep biological insights from "noisy" clinical data. Future work should focus on integrating family medical history into clinical prediction tools to help clinicians identify patients who may require earlier transition to second-line treatments.
Summary updated