Learning individualized treatment strategies for mental disorders is confronted with many challenges. The true underlying mental status is not observable, substantial between-patient heterogeneity is present, the high-dimensional combinations of measures of symptomatology need to be addressed, and treatment mechanisms can be complex and unknown. In addition, inferring optimal treatments of mental disorders in a target population needs to consider potential distribution disparities between the patient data in a study and the target population of interest. In order to address these challenges, we provide a new paradigm that connects measurement theory, efficient weighting procedure, and flexible neural network architecture through latent variables to discover individualized treatment strategies that integrate evidence from multi-domain data sources and multiple studies to increase generalizability. Simulation studies demonstrate consistent superiority of the proposed method and the weighting scheme when applied to the target population. Application of the proposed method to real-world studies is conducted to recommend treatments to patients with major depressive disorder.