Half of people starting antidepressant medication experience no significant improvement with the first medication prescribed, leading either to discouragement and dropout or months of trial-and-error. More accurately matching individuals with specific medications could significantly increase treatment success. While selecting antidepressant medication is one of the most common clinical decisions in mental health and general medical care, actionable evidence to guide that decision is remarkably scarce. Matching patients with specific medications remains no better than chance. We identify three key gaps in previous research to personalize antidepressant selection. First, previous research has not clearly distinguished predicting overall prognosis or nonspecific response from predicting specific response to specific medications. Second, sample sizes have typically been inadequate to accurately detect differential treatment response. Third, previous research has focused on response to a single episode of treatment, and that is too “noisy” a clinical phenotype to support identification of useful endophenotypes. We propose to address these limitations using data resources, informatics tools, and analytic methods developed by the NIMH-funded Mental Health Research Network (MHRN). Records databases include comprehensive and harmonized data concerning sociodemographic characteristics, past mental health diagnoses and treatments, comorbid mental health or substance use disorders, concomitant treatments, dose and duration of antidepressant exposure, and structured assessment of depression symptoms at baseline and follow-up. We will assemble a database including over 500,000 antidepressant treatment episodes for over 200,000 patients treated since 2010 in order to: 1) Rigorously test core assumptions of previous research and treatment guidelines - Heterogeneity-of- treatment-effects analyses will evaluate two common beliefs regarding antidepressant response: a. Favorable or unfavorable prior response to a single medication predicts subsequent response b. Prior response to a medication in a specific class predicts subsequent response to that class 2) Identify and validate decision rules to inform selection of antidepressant treatment based on patterns of prior treatment response – Analyses will use machine learning methods (dynamic weighted ordinary least squares regression) to identify and validate practical decision rules for clinical practice. 3) Identify valid treatment response phenotypes to enable future research regarding genomic- or biomarker-guided personalization – Primary analyses will use k-means clustering to identify patterns of response across medications and medication classes. Additional analyses will explore clinical characteristics that distinguish response phenotypes as well as order effects. Findings will inform a next generation of research regarding genetic- and biomarker-guided antidepressant treatment.