Abstract
The causes of depression are complex, and the current diagnosis methods rely solely on psychiatric evaluations with no incorporation of laboratory biomarkers in clinical practices. We investigated the stability of blood DNA methylation depression signatures in six different populations using six public and two domestic cohorts (n = 1942) conducting mega-analysis and meta-analysis of the individual studies. We evaluated 12 machine learning and deep learning strategies for depression classification both in cross-validation (CV) and in hold-out tests using merged data from 8 separate batches, constructing models with both biased and unbiased feature selection. We found 1987 CpG sites related to depression in both mega- and meta-analysis at the nominal level, and the associated genes were nominally related to axon guidance and immune pathways based on enrichment analysis and eQTM data. Random forest classifiers achieved the highest performance (AUC 0.73 and 0.76) in CV and hold-out tests respectively on the batch-level processed data. In contrast, the methylation showed low predictive power (all AUCs < 0.57) for all classifiers in CV and no predictive power in hold-out tests when used with harmonized data. All models achieved significantly better performance (>14% gain in AUCs) with pre-selected features (selection bias), with some of the models (joint autoencoder-classifier) reaching AUCs of up to 0.91 in the final testing regardless of data preparation. Different algorithmic feature selection approaches may outperform limma, however, random forest models perform well regardless of the strategy. The results provide an overview over potential future biomarkers for depression and highlight many important methodological aspects for DNA methylation-based depression profiling including the use of machine learning strategies.