The psychopathological features of affective disorders have been continuously debated ever since these disorders were first conceptualized in the 19th century by psychiatric pioneers such as Falret, Baillarger, Kahlbaum, and Kraepelin. Even today, classification remains problematic in clinical psychiatry: some patients’ manifestations are hard to fit into one of the affective disorders as defined by DSM-IV or ICD-10. In modern scientific psychiatry, too, such issues still engender lively debate—about, for instance, the classification of schizoaffective disorder (1), or the various attempts to subclassify the affective disorders. A major step was taken in the late 1960s by, among others, Perris and Angst when they demonstrated the importance of distinguishing uni- and bipolar types of what had, till then, been the all-encompassing manic-depressive disorder described by Kraepelin. The effort to characterize affective disorders more precisely by their psychopathological features is inspired by the hope that disease subtypes thus identified will respond better to targeted treatment than the more broadly defined disorders of the current classification. A case in point is the hypothesis that affective disorders are actually much more commonly bipolar than previously thought (2). Another illustration of the search for clinically meaningful psychopathological characteristics is the attempt to subdivide bipolar disorders by the predominance of depressive or manic recurrences, as described in a highly interesting study by Baldessarini et al. (3) in the current issue of Acta Psychiatrica Scandinavica. The authors, working at mood disorder centers in five different cities on four continents—Boston, Buenos Aires, Barcelona, Cagliari, and Bundang in South Korea—studied the clinical course of 928 patients with bipolar I disorder. All patients who had at least twice as many depressive episodes as manic and mixed episodes taken together were classified as mainly depressed; all with at least twice as many manic and mixed episodes as depressive episodes were classified as mainly manic. About one-fifth of patients were found to be mainly depressed, some 30% mainly manic, and half of the sample neither. Remarkably, the predisposition to depression was already clearly recognizable in the first 5 years of illness. The mainly depressed and mainly manic subtypes both showed statistically significant associations with certain features of the patients and their disorders: thus, for example, mainly manic patients were more likely to be well educated and to abuse drugs, while mainly depressed patients more often attempted suicide. Suicidal behavior was almost three times as common in mainly depressed as in mainly manic bipolar patients. In sum, the two subtypes were significantly associated with a total of 12 different factors (of 30 for which potential associations were sought). When these factors were evaluated in patients who did not fit into either subtype, they were found to take on intermediate values between those of the two subtypes. This may suggest a dose–effect relationship and lends credibility to the model. It would be valuable if the subtyping described in this paper was to ‘contribute to improve planning of clinical care and to biological studies of BPD’, as the authors write. Clearly, bipolar disorder subgroups such as those defined here may well attract the attention of geneticists. At present, for example, the subgroup of bipolar patients who respond well to lithium is the subject of a genome-wide association study (4). Nor is there any denying the importance of the association, stressed by Baldessarini and co-workers, between depressive episodes and suicidal behavior. On the other hand, subtyping by the polarity of recurrences needs to be further refined before it can be used for clinical purposes. With regard to suicidal behavior, for example, the presence of the depressive subtype of bipolar disorder predicts suicidal acts in less than 30% of cases, and Cohen’s kappa is 0.2, indicating no more than modest agreement. More factors may need to be added to make the model more predictive. Such major problems are not unique to psychopathological predictor research; ultimately, they simply reflect the complexity of illness. The biomarkers of biological predictor research are often associated with only slightly differing probabilities that the patient will develop the illness in question: for example, Ioannidis et al. (5) recently showed that, among the 35 most often cited biomarkers in medicine, only 15 were associated with a statistically significant way with illnesses or illness outcomes—and, of these 15, only 7 were associated with relative risks higher than 1.37. Predictor research, as this meticulous study by Baldessarini and co-workers (3) reminds us, has already come a long way, but it still has a long road ahead.