Dubiety exists over whether clinical symptoms of schizophrenia can be distinguished from affective psychosis, the assumption being that absence of a "point of rarity" indicates lack of nosological distinction, based on prior group-level analyses. Advanced machine learning techniques, using unsupervised (hierarchical clustering) and supervised (regularized logistic regression algorithm and nested-cross-validation) were applied to a dataset of 202 patients with functional psychosis (schizophrenia n = 120, affective psychosis, n = 82). Patients were initially assessed with the Present State Examination (PSE), and followed up 2.5 years later, when DSM III diagnoses were applied (independent of initial PSE). Based on PSE syndromes, unsupervised learning discriminated depressive (approximately unbiased probability, AUP = 0.92) and mania/psychosis (AUP = 0.94) clusters. The mania/psychosis cluster further split into two groups - a mania (AUP = 0.84) and a psychosis cluster (AUP = 0.88). Supervised machine learning classified schizophrenia or affective psychosis with 83.66% (95% CI = 77.83% to 88.48%) accuracy. Area under the ROC curve (AUROC) was 89.14%. True positive rate for schizophrenia was 88.24% (95%CI = 81.05-93.42%) and affective psychosis 77.11% (95%CI = 66.58-85.62). Classification accuracy and AUROC remained high when PSE syndromes corresponding to affective symptoms (those that corresponded to the depressive and mania clusters) were removed. PSE syndromes, based on clinical symptoms, therefore discriminated between schizophrenia and affective psychosis, suggesting validity to these diagnostic constructs.