Mental disorders affect more than one billion people worldwide, and timely intervention is critical; however, current diagnostic practice remains largely subjective. Magnetic resonance imaging (MRI) provides a noninvasive way to measure brain structure and function and identify objective biomarkers of mental disorders. Despite these advances, several challenges remain in mental disorder diagnosis and neuroimaging analysis, including objective characterization of the neuropsychiatric continuum and heterogeneity, preprocessing-related variability, and heterogeneous information integration from high-dimensional, multimodal data. This dissertation develops data-driven, multi-view, and multimodal representation learning approaches, which learn low-dimensional representations from high-dimensional MRI data, to characterize the neuropsychiatric continuum and heterogeneity, mitigate preprocessing-related variability, and identify phenotypic and psychiatric biomarkers from structural and functional MRI. For unsupervised representation learning, our data-driven interpolation framework effectively characterizes individual differences within a group and continuous patterns between groups. For multi-view representation learning, our methods substantially improve both neural network representational similarity across preprocessing pipelines and prediction robustness for brain-phenotype relationships. For multimodal representation learning, our multimodal latent variable models, developed in the MISA PyTorch framework, successfully identify linked sources associated with phenotypic and psychiatric measures. Together, our methods and toolboxes contribute to reproducible neuroimaging analysis and reliable brain-behavior relationship discovery.