Current technologies used for separation of trivalent lanthanides and actinides are based on multistage extraction cascade with relatively low efficiency at each stage. Difficulties of separation process arises from similarity of trivalent lanthanides and actinides properties and liquid-liquid extraction is used either as a basic separation technique or as a model experiment for more complicated system. Significant improvement can be reached by development selective and effective extractants, that could form a stable complex with a single ion (or at least with a single group of ions) in presence of competing ones. Now we know several groups of such ligands with satisfactory selectivity and other properties (as solubility in target diluent, nitric acid resistance, etc) and a process of development is usually devoted to modifications of these basic frameworks. On the other hand, a process of selective ligand development is very similar to drug design process, where we try to find a molecule with the best affinity to the biological target. In the case of drug design, using computational models, especially based on machine learning is an essential part of development process. Here we would like to present an example of adopting drug design approaches to lanthanides and actinides complexation problem. We have collected databases containing complex stability constants for trivalent f-elements. For each element we built models based on neural network architectures predicting target constants (R2~0.8, RMSE~2[logK]). We also ‘unboxed’ these models to determine key fragments of molecule with the most significant influence on stability constant value. This approach can ‘offer’ molecular fragments increasing or decreasing constant value. Finally, we combined models with genetic algorithm available not only to predict values of a drawn molecule, but to construct new molecules with increasing target property.