The objective of the proposed research is developing data-driven, multi-view, and multimodal representation learning approaches to (1) characterize the neuropsychiatric continuum and heterogeneity from neuroimaging data, (2) mitigate pipeline-related variation in neuroimaging preprocessing, and (3) identify phenotypic and psychiatric biomarkers from multimodal neuroimaging data. We demonstrate that (1) our data-driven interpolation framework effectively characterizes individual differences within a group and continuous patterns between groups, (2) our multi-view representation learning methodologies greatly improve prediction robustness, and (3) our multimodal latent variable models successfully identify linked sources associated with phenotypes and psychosis. Our proposed methods and toolboxes contribute to reproducible neuroimaging analysis and reliable brain-behavior relationship discovery.