Individuals with depression, psychosis and clinically high risk for psychosis face cognitive decline and show altered functional connectivity in resting-state networks. An unsupervised machine-learning method, sparse partial least squares analysis (SPLS) captures latent variables from clinical features and brain mechanisms, allowing the examination of relationships between functional connectivity and cognitive functions across disorders. This study aimed to explore relationships between functional connectivity, cognitive domains and study groups using SPLS. 990 participants with recent-onset psychosis (227), clinical high-risk for psychosis (218), recent-onset depression (192), and healthy controls (353) underwent resting-state fMRI scans and neuropsychological assessments covering working memory, processing speed, verbal learning, social cognition, attention, and reasoning. To yield 30 independent component maps, the independent component analysis was performed. Then, SPLS was utilized using a 10-fold cross-validated framework that selects significant weight vectors by using a 60% density grid search, 5000 permutation tests for significant vector pairs, a mass-univariate statistical test, and Fisher multiple comparisons. The results revealed increased functional connectivity between the default mode network (DMN) and control network, DMN and dorsal attentional network (DAN), and within DMN. Furthermore, decreased functional connectivity within the salience network (SN), between SN and DMN, as well as SN and DAN was detected. In addition, increased and decreased connectivity between and within networks were related to low performance in cognitive domains in the recent-onset psychosis group, rs = .41, p <.001. Our study identified prominent features in recent-onset psychosis by associating functional connectivity, cognitive domains and study groups.