Methylation risk scores are associated with a collection of phenotypes within electronic health record systems

Abstract

Inference of clinical phenotypes is a fundamental task in precision medicine, and has therefore been heavily investigated in recent years in the context of electronic health records (EHR) using a large arsenal of machine learning techniques, as well as in the context of genetics using polygenic risk scores (PRS). In this work, we considered the epigenetic analog of PRS, methylation risk scores (MRS), a linear combination of methylation states. We measured methylation across a large cohort (n = 831) of diverse samples in the UCLA Health biobank, for which both genetic and complete EHR data are available. We constructed MRS for 607 phenotypes spanning diagnoses, clinical lab tests, and medication prescriptions. When added to a baseline set of predictive features, MRS significantly improved the imputation of 139 outcomes, whereas the PRS improved only 22 (median improvement for methylation 10.74%, 141.52%, and 15.46% in medications, labs, and diagnosis codes, respectively, whereas genotypes only improved the labs at a median increase of 18.42%). We added significant MRS to state-of-the-art EHR imputation methods that leverage the entire set of medical records, and found that including MRS as a medical feature in the algorithm significantly improves EHR imputation in 37% of lab tests examined (median R² increase 47.6%). Finally, we replicated several MRS in multiple external studies of methylation (minimum p-value of 2.72 × 10^-7) and replicated 22 of 30 tested MRS internally in two separate cohorts of different ethnicity. Our publicly available results and weights show promise for methylation risk scores as clinical and scientific tools.

Fig. 1. MRS increases imputation accuracy on a variety of outcomes.

a–c The performance of the PRS (blue) and MRS (green) imputations on the y-axis with the baseline model performance on the x-axis. The performance of binary phenotypes (Phecodes (a), medications (b)) is measured using area under the ROC curve (AUC) and the performance of continuous phenotypes (lab results (c)) is measured using proportion of variance explained (R²). Shown is the performance on the union of outcomes that were significantly improved over the baseline model by either the MRS or PRS and that were significantly imputed by their corresponding predictor (72 Phecodes, 59 medications, and 31 labs). d–f The disease incidence as a function of the PRS (blue) and MRS (green) binned by deciles (d, e); and the observed Urea Nitrogen lab result value plotted against its imputed value (f).

Fig. 2. Improvement in lab result imputation performance by including MRS.

For lab results that were significantly better imputed using a matrix completion imputation procedure that included the MRS values, we compare the quality of the imputed values (R²) using only the EHR data (SoftImpute) to the values generated when including the MRS values in addition to the EHR data (SoftImpute+MRS).

Fig. 3. Imputation accuracy may improve with additional samples.

We downsampled the number of individuals to evaluate the imputation performance as a function of sample size using a well-imputed Phecode (a), medication (b), and lab value (c). The performance is significantly affected by the number of individuals, suggesting that there is additional power to be gained with the addition of more methylation samples. Error bars indicate 95% confidence intervals.

Fig. 4. Labs as imputed by methylation, genotypes, and an externally-trained polygenic risk score.

The cross-validated R² between the true and imputed lab value on 541 unrelated patients of non-Hispanic-Latino white-identifying individuals using a baseline predictor as well as a baseline predictor with methylation, genotypes, and a PRS externally-trained from UKBiobank summary statistics. HDL corresponds to high-density lipoprotein cholesterol and HGBA1C to glycated hemoglobin. Error bars indicate 95% confidence intervals.

Fig. 5. Best methylation-imputed Phecodes within ancestral populations.

After training a model on the entire heterogeneous population of individuals, we evaluated the predictive performance within each population separately. We observed only 6 (of 60) significant differences between self-reported ancestral groupings. Error bars indicate 95% confidence intervals.