Modeling of properties of amino acids with random, semirandom, and molecular connectivity descriptors

Sets of random indices, sets of molecular connectivity indices (MCIs), and mixed sets of random and MCIs have been used to model nine properties of amino acids. The number of amino acids throughout the concerned properties is not constant, it ranges from 21 for the isoelectric point to eight for the unfrozen water content. The quality of the descriptors has been submitted to the leave‐one‐out (loo) and to the training/evaluation statistical tests. Molecular connectivity descriptors made of MCIs only show exceptionally good model characteristics for the volume side‐chain throughout both tests, while the model of the solubility holds only at the loo level. Random descriptors made of random indices, and semirandom descriptors, made of random and MCIs show good descriptive characteristics of all other properties at the loo level, while they fail at the training/evaluation level. Molecular connectivity descriptors achieve a good model at the training/evaluation level for the hydration potential and for the specific rotation. Practically, at the loo level all properties, excluding the unfrozen water content, can satisfactorily be modeled, while at the training/evaluation level, only side‐chain volume, hydration potential, and specific rotation can be modeled in a satisfactory way. To notice is the importance, at both statistical levels, of the hydrogen content of the amino acids, which is encoded with a perturbation parameter at the level of the valence delta number of the valence MCIs. © 2011 Wiley Periodicals, Inc. Int J Quantum Chem, 2011