A new class of antimicrobial agents, 3-nitrocoumarins and related compounds, has been chosen as a subject of the present study. In order to explore their activity and molecular properties that determine their antimicrobial effects, QSAR models have been proposed. Most of the 64 descriptors used for the development were extracted from semiempirical and density functional theory (DFT) founded calculations. For this study literature data containing results of microbiological activity screening of 33 coumarin derivatives against selected clinical isolates of C. albicans (CA) and S. aureus (SA) have been selected. Multivariate predictive models based on random forests (RF) and two hybrid classification approaches, genetic algorithms (GA) associated with either support vector machines (SVM) or k nearest neighbor (kNN), have been used for establishment of QSARs. An applied feature selection approach enabled two-dimensional linear separation of active and inactive compounds, which was a necessary tool for rational candidate design and descriptor relevance interpretation. Candidate molecules were checked by cross-validated models, and selected derivatives have been synthesized. Their antimicrobial activities were compared to antimicrobial activities of the representative derivatives from the original set in terms of minimal inhibitory concentration (MIC) against chosen SA and CA ATCC strains. High ranking of descriptors consistent with the degree of hydrolytic instability of selected compounds is common to models of antimicrobial activity against both microorganisms. However, descriptor ranking indicates different antimicrobial mechanisms of action of chosen coumarin derivatives against selected microbial species.