Ancestry component as a major predictor of lithium response in the treatment of bipolar disorder

Bipolar Disorder (BD) represents the seventh major cause of disability life-years-adjusted. Lithium remains as a first-line treatment, but clinical improvement occurs only in 30 % of treated patients. Studies suggest that genetics plays a major role in shaping the individual response of BD patients to lithium. We used machine-learning techniques (Advance Recursive Partitioned Analysis, ARPA) to build a personalized prediction framework of BD lithium response using biological, clinical, and demographical data. Using the Alda scale, we classified 172 BD I-II patients as responders or non-responders to lithium treatment. ARPA methods were used to build individual prediction frameworks and to define variable importance. Two predictive models were evaluated: 1) demographic and clinical data, and 2) demographic, clinical and ancestry data. Model performance was assessed using Receiver Operating Characteristic (ROC) curves. The predictive model including ancestry yield the best performance (sensibility = 84.6 %, specificity = 93.8 % and AUC = 89.2 %) compared to the model without ancestry (sensibility = 50 %, Specificity = 94.5 %, and AUC = 72.2 %). This ancestry component best predicted lithium individual response. Clinical variables such as disease duration, the number of depressive episodes, the total number of affective episodes, and the number of manic episodes were also important predictors. Ancestry component is a major predictor and significantly improves the definition of individual Lithium response in BD patients. We provide classification trees with potential bench application in the clinical setting. While this prediction framework might be applied in specific populations, the used methodology might be of general use in precision and translational medicine.