The application of machine learning within psychiatric research has proliferated in recent years, driven by the growing availability of large-scale neurobiological, behavioral, and clinical datasets. This development has enabled novel approaches to modeling the complex and multifactorial processes underlying mental disorders. Yet, a persistent tension remains in the field between predictive accuracy and mechanistic interpretability, which constrains the translational potential of ML-derived insights for both neuroscientific theory and clinical application. This systematic review critically synthesizes contemporary ML-based methodologies in psychiatric research, with particular attention to the distinction between predictive models, designed to maximize classification performance or prognostic precision, and mechanistic models, which seek to elucidate the latent cognitive and neural processes implicated in psychopathology. This review synthesizes peer-reviewed machine learning studies in psychiatric research spanning neurobiological, behavioral, and clinical domains, with emphasis on predictive and mechanistic models. The extant literature reveals a pronounced predominance of predictive modeling approaches, typically employing supervised learning algorithms such as support vector machines and deep learning architectures to address tasks ranging from diagnostic categorization and symptom severity estimation to treatment response prediction in conditions such as schizophrenia, mood disorders, and anxiety disorders. In contrast, mechanistic machine learning frameworks, often situated within computational psychiatry paradigms such as reinforcement learning and Bayesian modeling, are comparatively underutilized, though they offer critical explanatory power by linking model parameters to underlying cognitive and neural mechanisms. A subset of emerging hybrid approaches seeks to reconcile predictive utility with mechanistic clarity, though they remain methodologically heterogeneous and insufficiently validated. The findings highlight enduring methodological challenges in the field, including limited generalizability across cohorts, inconsistent validation strategies, and difficulty aligning learned representations with underlying neurocognitive processes. These limitations underscore the imperative for computational models that are not only predictive but also mechanistically interpretable, thereby advancing both theoretical insight and translational relevance.