Deep biomarkers of human aging: Application of deep neural networks to biomarker development

Abstract

One of the major impediments in human aging research is the absence of a comprehensive and actionable set of biomarkers that may be targeted and measured to track the effectiveness of therapeutic interventions. In this study, we designed a modular ensemble of 21 deep neural networks (DNNs) of varying depth, structure and optimization to predict human chronological age using a basic blood test. To train the DNNs, we used over 60,000 samples from common blood biochemistry and cell count tests from routine health exams performed by a single laboratory and linked to chronological age and sex. The best performing DNN in the ensemble demonstrated 81.5 % epsilon-accuracy r = 0.90 with R(2) = 0.80 and MAE = 6.07 years in predicting chronological age within a 10 year frame, while the entire ensemble achieved 83.5% epsilon-accuracy r = 0.91 with R(2) = 0.82 and MAE = 5.55 years. The ensemble also identified the 5 most important markers for predicting human chronological age: albumin, glucose, alkaline phosphatase, urea and erythrocytes. To allow for public testing and evaluate real-life performance of the predictor, we developed an online system available at http://www.aging.ai. The ensemble approach may facilitate integration of multi-modal data linked to chronological age and sex that may lead to simple, minimally invasive, and affordable methods of tracking integrated biomarkers of aging in humans and performing cross-species feature importance analysis.

Keywords: aging biomarkers; biomarker development; deep learning; deep neural networks; human aging; machine learning.

Figure 1. Project pipeline

Laboratory blood biochemistry data sets were normalized and cleaned of outliers and some abnormal markers. For biological age prediction, 21 different DNNs with different parameters were combined in ensemble based on ElasticNet model. For biological sex prediction, single DNN were trained.

Figure 2. Analysis of best DNN model in the ensemble and the whole ensemble

(A) Correlation between actual and predicted age values by the best DNN in the ensemble. (B) Biological age epsilon-prediction accuracy plot for the best DNN. (C) Biological age marker Importance, performed using FPI method. (D) Correlation between actual and predicted age values by whole ensemble based on ElasticNet model. (E) Biological age epsilon-prediction accuracy plot for the ensemble. (F) Heat map for Pearson's correlation coefficients between 40 DNNs. Scale bar colors indicate the sign and magnitude of Pearson's correlation coefficient between predictions of DNNs.

Figure 3. DNNs outperform baseline ML approaches in terms of R² statistics

DNN were compared with 7 ML techniques: GBM (Gradient Boosting Machine), RF (Random Forests), DT (Decision Trees), LR (Linear Regression), kNN (k-Nearest Neighbors), ElasticNet, SVM (Support Vector Machines). (A) GBM shows the higher 0,72 R² among ML models for biological age prediction. (B) All ML models have comparable high R² for biological sex prediction.

Figure 4. Comparison of sub-models for stacking ensemble and evaluation of filling strategies

(A) ElasticNet model has the higher epsilon-prediction accuracy among the stacking models. (B) ElasticNet is the best model for stacking from the point of R² statistics. (C) Median filling strategy has higher epsilon-prediction accuracy than other strategies. Median filling strategy shows 64,5 % epsilon accuracy within 10 years frame. (D) Median filling strategy is better from the point of R² statistics.

Figure 5. Top features analysis

(A) Dependence of the epsilon-prediction accuracy from the number of features. (B) Dependence of R² statistics from the number of features.