An Imaging Risk Score for Bipolar Disorder: Multimodal Comparison and Integration with Polygenic Risk
Listed in
This article is not in any list yet, why not save it to one of your lists.Abstract
Background
Biological risk scores provide objective markers of bipolar disorder (BD) liability. Large-scale genome-wide association studies (GWASs) have led to polygenic risk scores (PRSs) that capture genetic liability to BD at the individual level. We lack a comparable large-scale effort to leverage a brain-wide association study (BWAS) to create an imaging risk score (IRS) that captures brain structural liability to BD at the individual-level. Here, we conduct the largest multivariate BWAS of BD and develop a BD IRS. We compare the BD IRS to the PRS and assess their combined predictive performance in multimodal risk score (MRS) models.
Methods
Our multivariate BD BWAS used 153 structural neuroimaging features from 3902 cases and 6855 controls (mean age (sd)=35.47 years (13.42), 56.70% female) in a cross-validated machine learning approach. Using the resulting model coefficients, we constructed an IRS. We then evaluated the out-of-sample performance of the IRS, PRS derived from the largest GWAS to date, and combined MRS model using logistic regression in a subset of individuals with imaging and genetic data (n BD =486, n controls =1701, age (sd)=33.10 (12.02), 56.47% female). We also performed stratified analyses for demographic and BD clinical groups. Prediction performance was quantified using the area under the receiver operating characteristic curve (AUC). Subsequent analyses examined associations between each risk score and BD relevant symptoms in developing youth.
Outcomes
In predicting case status, the BD BWAS model achieved a mean AUC BD of 0.62 (sd BD =0.08) and in stratified analyses, had higher prediction than subgroup-specific BWAS models. The BD type II BWAS was the only exception (AUC BDII =0.60, sd BDII =0.07 vs AUC BD =0.58, sd BD =0.09). The BD PRS had an AUC BD of 0.65 (CI 95% =0.63-0.68) and a higher AUC than subtype-specific PRS in stratified analyses. The PRS typically attained a higher AUC than the IRS. Notably, the IRS and PRS were uncorrelated (r=0.03, p=0.11). MRS models outperformed the prediction of individual PRS and IRS with the strongest prediction for BD type I (AUC BD =0.74, CI 95% =0.70-0.78). The IRS had the strongest associations with BD-related symptoms in developing youth including increased energy (OR=1.21, p fdr =0.03) and distress from psychotic-like experiences (OR=1.23, p fdr =0.04).
Interpretation
Our BD IRS is a novel biological indicator of BD risk. The lack of correlation between the IRS and PRS along with improved MRS prediction suggests the two risk scores capture some orthogonal liability to BD. We show that integrating imaging and genetics is useful for biological characterization of BD and should be considered for the development of clinically relevant prediction models.
