Drug design is usually understood as “an inventive process of finding new medications based on the knowledge of the biological target” – according to the Wikipedia definition – where “drug” is usually defined as a relatively small organic ligand binding to a biomolecule to either inhibit or promote its activity. [1] Methods of thus-defined drug design are well established in the early stages of drug discovery and focus mainly on identifying new chemical entities (NCEs) – molecules with desired biological activity. Along with activity, however, drugs must possess acceptable ADMET (Absorption, Distribution, Metabolism, Elimination, and Toxicity) properties. Although a few recent trends push ADMET optimization into drug design, these properties are still largely neglected until the lead optimization phase occurring later in drug discovery when the available chemical space is already narrowed by affinity and selectivity concerns. Usually one group of drug discovery researchers hands out lead molecules to a separate group of drug development scientists: “Our job is done, here are the active molecules. Now you worry about their other properties.” This separation could, and often does, lead to frustrating failures where all the lead compounds may, for example, turn out to be insoluble in water, impermeable via the intended route of administration, unacceptably metabolized by enzymes common to co-medications, or produce unacceptable adverse effects. All the effort and cost of discovering them would then be wasted. Authors of this article strongly believe in “ADMET design” – a paradigm where ADMET properties are included in the earliest stages of drug discovery and treated on equal footing with biological efficacy. This multidimensional search and optimization should be a normal function of the team of discovery chemists and other pharmaceutical scientists. The difficulty is that humans are not well-adapted to thinking in more than a few dimensions at once. In silico approaches are the solution as they facilitate integration of multiple variables and the most efficient use of experimental data. In the following sections we focus on machine learning approaches, such as Artificial Neural Networks (ANN) [2] or Support Vector Machines (SVM) [3], to QSAR/QSPR modeling [4] of ADMET properties. In QSAR/QSPR, chemical structures are encoded by calculated values of molecular descriptors serving as inputs to mathematical predictive models. ANN models, in particular, have been determined to be superior predictors over other methods. [5] Training several ANN or SVM and taking an arithmetic average of their outputs as the final answer usually enhances accuracy of prediction and is known as ANN Ensemble (ANNE) or SVM Ensemble (SVME) approaches. [6] Some researchers regard these techniques as “black boxes” meaning that although ANNE or SVME models can produce numerical outputs, their internal complexity precludes any descriptive interpretation of these results. Contrary to these stereotypical beliefs, predictive mathematical models of this type are quite interpretable and thus amenable to a direct use in drug design. We demonstrate this claim on specific examples.