Discovering multi-omics predictors for lithium response in bipolar disorder by integrating proteomic and genomic data: A 6-month follow-up study
Hyunju Lee1, Dohyun Han2, Kyung Sue Hong3
1Department of Neuropsychiatry, Seoul National University Hospital, Seoul, Republic of Korea; Department of Psychiatry, Seoul National University College of Medicine, Seoul, Republic of Korea.
Background:
Reliable predictors of lithium response in bipolar disorder (BD) are unavailable, and reported predictors have rarely been temporally validated. We examined whether baseline proteomic, clinical, and polygenic information predicts lithium response using a temporal validation design.
Methods:
Patients with BD from three hospitals were assessed over six months (acute) or at baseline (stable maintenance); lithium response was rated with the Alda scale. Of 295 participants with complete multi-omic data, 228 in the acute phase entered prediction analyses: elastic-net models (responders, Alda A ≥ 7) developed in an earlier cohort (n = 110) were applied to a later, separately assayed validation cohort (n = 118).
Results:
In the validation cohort, the protein model showed modest discrimination (AUC = 0.58, 95% CI 0.47-0.70; permutation p = 0.060), whereas clinical (AUC = 0.51) and polygenic (AUC = 0.46) models did not transfer, and combining modalities did not improve prediction. In an exploratory comparison of clear responders and non-responders (Alda A ≥ 7 vs ≤3), protein models reached AUC 0.67-0.70. Protein discrimination was similar with validation-blind preprocessing but weaker in sensitivity analyses.
Limitations:
Validation was temporal rather than site-independent, and most patients were already taking lithium at baseline sampling.
Conclusions:
Under temporal validation, baseline proteomic profiles carried at most modest information about lithium response, and clinical and polygenic predictors did not transfer. Candidate markers need external validation in larger cohorts sampled before lithium initiation.