Machine learning based prediction of major depressive disorder symptoms increases genomic discovery power

Abstract

Background: Large biobank samples such as the UK Biobank (UKB) offer great potential for discovering loci influencing complex disorders. However, data may not be collected uniformly with only a subset of participants being assessed using clinically based instruments. There is evidence that broadly defined Major Depressive Disorder (MDD) using non-clinical instruments is not sufficient to detect disorder specific loci and results in lower genetic correlation (rG) with clinically ascertained MDD. With the availability of thousands of additional variables, Machine Learning (ML) methods offer the potential to predict clinical risk in subjects not directly measured thereby increasing power and specificity. This research has been conducted using the UKB (application 30782) and aimed to predict lifetime clinically based MDD symptoms using an ML method adapted to the non-random missing patterns observed in biobanks. Method(s): This Missingness Adapted Group-wise Incremental Clustered (MAGIC)-LASSO was applied in UKB to predict the CIDI-SF MDD symptom count (MDDsx). Measured MDDsx is available for ~150K subjects and was used to construct a model predicting unobserved MDDsx in the remaining ~350K subjects. This method implements hierarchical clustering of variables by missingness structure followed by applying sequential Group LASSO by cluster. Variables are filtered for missingness balance between training and target prediction sets and variables for the final model are included stepwise by rank completeness and balance. Genome-wide association study (GWAS) analyses were performed using unrelated individuals of European ancestry for measured (n=117,572) and predicted unmeasured (n=242,381) MDDsx. Heritability (h2) and rG analyses were performed in LDSC to assess the relationship of the predicted scores and other definitions of MDD and risk factors. Result(s): After filtering for invariance, excessive missingness, and balance across missingness patterns, 1,246 phenotypes were included in the MAGIC-LASSO procedure. The algorithm selected 61 variables to enter the final model and retained 27 in the final prediction with a phenotypic correlation of 0.69 between MDDsx directly assessed (n=157,166) and predicted (n=345,315) participants. GWAS of MDDsx detected 2, 50, and 152 independent genome-wide significant loci (including 24 overlapping intervals with recent PGC MDD loci) using the observed (h2=0.025), unobserved predicted (h2=0.053), and predicted all (h2=0.24, n=339,953) sample sets, respectively. rG between observed and predicted MDDsx were 0.92 (se=0.02). We calculated phenotype correlations and rG for other MDD diagnostic definitions previously used in UKB, and variables retained in the final model with known relationships to MDD. Predicted MDDsx is only moderately correlated with MDD risk factors such as neuroticism (r2obs 0.15, r2predicted=0.38) but more highly correlated with probable diagnosis from cardinal symptoms (r2obs =0.34, r2predicted=0.73). Discussion(s): The analysis represents a first attempt to predict one dimension of more strictly defined MDD by applying a novel ML approach. These predictions represent an effective sample size increase of 49% for genetic discovery studies of MDD in the UKB. Results support the interpretation that the predicted MDDsx are specific and leverage the correlated variables to produce a novel score indexing CIDI based symptoms. Future studies are planned to leverage predicted MDDsx to define probability of lifetime case-control status. Disclosure: Nothing to disclose.Copyright © 2022