Machine learning (ML) methods realize direct mappings between features of molecules and properties of interest, thus allowing fast prediction of properties of new molecules. On the other hand, we can determine which of the features of the molecules the ML model was trained on are important for the prediction of the desired properties. This helps us decide, in combination with physics-based knowledge, which features, such as effective nuclear charge, spin state, and types of ligands, are to be optimized, yielding chemical insights of driving forces behind the properties of interest.