Exemplar scoring identifies genetically separable phenotypes of lithium responsive bipolar disorder

摘要

Note that this step of the analysis is performed using the clinical feature dataset alone, and therefore has no overlap in the feature space with genomic data used in Step 2. A Demonstration of heterogeneity in the relationship between lithium responsiveness (depicted as “Li(+)” for responders and “Li(−)” for non-responders) and clinical features across four hypothetical sites. Overall (A) demonstrates how a classifier trained on different sites’ data may yield different discriminative functions. B Points demonstrate the aggregated dataset (“+” and “−” are responders and non-responders, respectively). Contours demonstrate regions of clinical feature space in which site-level classifiers from (A) agree with high accuracy on the predicted class (i.e., where they overlap). A clinical exemplar score can be computed for each subject in the clinical dataset by (1) holding his data out of the training set, (2) predicting his lithium responsiveness using site-level classifiers trained on the remaining subjects, then (3) using the site-wise prediction results to compute the clinical exemplar score. C Stratification of the clinical dataset according to lithium responsiveness and clinical exemplar score quartile. The “LRBest” and “NRBest” exemplars are those responders and non-responders with clinical exemplar scores above the 75th percentile, respectively. The “LRPoor” and “NRPoor” exemplars are those responders and non-responders with clinical exemplar scores below the 25th percentile, respectively. This stratification can be used to evaluate the clinical features that differentiate good from poor clinical exemplars of lithium response and non-response, respectively.