Machine learning optimized polygenic scores for blood cell traits identify sex-specific trajectories and genetic correlations with disease

Figures

• 

  Figure 1. PGS construction of blood cell traits using 6 different methods Six PGS methods were evaluated in this study: pruning and thresholding (P + T) and 5 learning methods: LDpred2, elastic net (EN), Bayesian ridge (BR), multilayer preceptron (MLP), and convolutional neural network (CNN).

• 

  Figure 2. Performance comparison of 5 learning methods with the P + T method Pearson r score performance of the P + T method for PGS construction of 26 blood cell traits are presented in testing on UKB or INTERVAL. Relative to the P + T method, performance of the 5 learning methods: EN, BR, LDpred2, MLP, and CNN, are presented for each blood cell trait in descending order, left to right, according to EN (largest Pearson r increases on left). Given a particular method, a trait and a cohort, the averaged r performance of the 5 trained models, corresponding to the 5 different training-testing data partitions, is shown. Detailed comparison between variant effect sizes estimated using EN/BR and P + T are presented in Figure S1.

• 

  Figure 3. Performance of P + T, EN, and LDpred2 methods on different variant sets in INTERVAL Using conditional analysis variants as a base set, we added in the selected variant sets with LD thinning and p value thresholding to form different sizes of expanded variant sets for each trait. We used the CA variant set as the starting point and then observed the performance of P + T, EN, and LDpred2 on these expanded variant sets. Note that in this figure, P + T refers to the method that directly applies the weighted sum on a given variant set with effect sizes from GWAS. See Figure S2 for similar performance comparison in UKB.

• 

  Figure 4. Trait levels by quintiles of EN-trained trait PGS in men and women for traits MCV, WBC#, and neutrophil count (NEUT#) in INTERVAL The y axis is the observed measurements adjusted only for technical artifacts and season for each blood cell trait. The generalized additive model (GAM) was used to fit the data across INTERVAL samples, and the shaded areas represent 95% confidence intervals. See Figure S3 for results of all other traits.

• 

  Figure 5. Correlation between PGS for blood cell traits and PGS for 6 common diseases in INTERVAL PGSs for blood cell traits; diseases were adjusted for the first 10 genetic principal components before the correlation analysis. Pearson r correlation analysis was performed between the blood cell trait PGSs and disease PGSs across INTERVAL samples, and the correlation tests with the p value passing the threshold of p = 10−4 (Bonferroni adjusted for all trait-disease tests) were deemed significant.