Blood-based multivariate methylation risk score for cognitive impairment and dementia

Abstract

Introduction: The established link between DNA methylation and pathophysiology of dementia, along with its potential role as a molecular mediator of lifestyle and environmental influences, positions blood-derived DNA methylation as a promising tool for early dementia risk detection.

Methods: In conjunction with an extensive array of machine learning techniques, we employed whole blood genome-wide DNA methylation data as a surrogate for 14 modifiable and non-modifiable factors in the assessment of dementia risk in independent dementia cohorts.

Results: We established a multivariate methylation risk score (MMRS) for identifying mild cognitive impairment cross-sectionally, independent of age and sex (P = 2.0 × 10^-3). This score significantly predicted the prospective development of cognitive impairments in independent studies of Alzheimer's disease (hazard ratio for Rey's Auditory Verbal Learning Test (RAVLT)-Learning = 2.47) and Parkinson's disease (hazard ratio for MCI/dementia = 2.59).

Discussion: Our work shows the potential of employing blood-derived DNA methylation data in the assessment of dementia risk.

Highlights: We used whole blood DNA methylation as a surrogate for 14 dementia risk factors. Created a multivariate methylation risk score for predicting cognitive impairment. Emphasized the role of machine learning and omics data in predicting dementia. The score predicts cognitive impairment development at the population level.

Keywords: Alzheimer's disease; DNA methylation; Parkinson's disease; aging; dementia; epigenetics; machine learning; mild cognitive impairments; risk prediction.

FIGURE 1

Overview of the model generation pipeline. The model generation workflow consists of model training in the EXTEND and EMIF‐AD cohorts using two different approaches. In the first approach, models for the prediction of CAIDE and LIBRA were trained in the EXTEND cohort (n = 1076). Furthermore, in the second approach, the DNA methylation data obtained from the EXTEND cohort was used to predict 14 known dementia risk factors. The 14 predicted risk scores (i.e., methylation profile scores; MPSs) were next used as variables for the prediction of AD and MCI status in the training set of the EMIF‐AD cohort (n = 436). The resulting multivariate methylation risk scores for AD (MMRS‐AD) and MCI (MMRS‐MCI) were evaluated in terms of AD and MCI classification performance in the independent test set in the EMIF‐AD cohort (n = 187). The model with the best performance was also validated in the ADNI (n = 223), PPMI (n = 129), and BASE‐II cohorts (n = 1017).

FIGURE 2

ROC curves of cross‐sectional AD and MCI status prediction in the independent test set of the EMIF‐AD MBD study. The MMRS models (red) are trained on the 14 MPSs for the prediction of AD (A) and MCI (B). The epi‐LIBRA and epi‐CAIDE scores (blue) are both predicted by the Random Forest model with a correlation‐based feature selection method (i.e., the best‐performing model) from the EXTEND data. The 95% confidence intervals of the AUROC values are indicated between brackets. EN, ElasticNet; sPLS‐DA, sparse partial least squares‐discriminant analysis; RF‐RFE, random Forest with recursive feature elimination.

FIGURE 3

Kaplan‐Meier curves of cognitive impairments in the ADNI and PPMI cohorts. The risk categories were defined based on the baseline score predicted by the MMRS‐MCI (RF‐RFE) model. The shaded area around the line indicates the 95% confidence interval. ADAS, Alzheimer's Disease Assessment Scale; RAVLT, Rey's Auditory Verbal Learning Test; TMT, Trail Making Test Part B Time; MMSE, Mini‐Mental State Examination.

FIGURE 4

Radar chart of the scaled mean absolute SHAP values. The scaled mean absolute SHAP values indicate the variable importance of the 14 MPSs in the three MMRS‐MCI models. The values for each of the models are scaled such that the sum equals one. EN, ElasticNet; sPLS‐DA, sparse partial least squares‐discriminant analysis; RF‐RFE, random Forest with recursive feature elimination.

FIGURE 5

ROC curves of MCI prediction. MCI status was predicted in the independent test set of the EMIF‐AD MBD study using the PGSs (top right), CSF biomarkers (bottom left), and both (top left) with and without the MPSs as additional variables. EN, ElasticNet; sPLS‐DA, sparse partial least squares‐discriminant analysis; RF‐RFE, random Forest with recursive feature elimination.