Neuroscientific research in mental disorders is plagued by unclear nosological boundaries, phenotypic diversity, and high intra- and interindividual variability of identified neurobiological correlates. Likewise, genotypes associated with an increased risk for, e.g., schizophrenia are regularly found to also increase the risk for uni- and/or bipolar affective disorders. Therefore, one major research strategy of the last decade was to avoid correlation of genetic variation with complex clinical disorders and instead to focus on so-called intermediate or endophenotypes, i.e., neurobiological variables such as in vivo receptor expression or neuronal activation patterns, which are hypothetically more closely related to direct gene effects. We describe one such attempt and show that intermediate phenotypes such as brain activation patterns elicited by more or less complex cognitive tasks underlie complex regulations and influences and may thus not be the best target for neurobiological research. We suggest that instead of reifying brain activation as correlates of mental disorders, such disorders may best be conceptualized as results of alterations/biases in basic learning mechanisms (e.g., Pavlovian and operant conditioning) interacting with individual and social environments and that neuroscientific research can rely on animal models and computationalized modeling to reveal their neurobiological correlates.