Multi-target spectral moments for QSAR and Complex Networks study of antibacterial drugs.

There are many of pathogen bacteria species which very different susceptibility profile to different antibacterial drugs. There are many drugs described with very different affinity to a large number of receptors. In this work, we selected Drug-Bacteria Pairs (DBPs) of affinity/non-affinity drugs with similar/dissimilar bacteria and represented it as a large network, which may be used to identify drugs that can act on bacteria. Computational chemistry prediction of the biological activity based on one-target Quantitative Structure-Activity Relationship (ot-QSAR) studies substantially increases the potentialities of this kind of networks avoiding time and resource consuming experiments. Unfortunately almost all ot-QSAR models predict the biological activity of drugs against only one bacterial species. Consequently, multi-tasking learning to predict drug's activity against different species with a single model (mt-QSAR) is a goal of major importance. These mt-QSARs offer a good opportunity to construct drug-drug similarity Complex Networks. Unfortunately, almost QSAR models are unspecific or predict activity against only one receptor. To solve this problem, we developed here a multi-bacteria QSAR classification model. The model correctly classifies 202 out of 241 active compounds (83.8%) and 169 out of 200 non-active cases (84.5%). Overall training predictability was 84.13% (371 out of 441 cases). The validation of the model was carried out by means of external predicting series, classifying the model 197 out of 221 (89.4%) cases. In order to show how the model functions in practice a virtual screening was carried out recognizing the model as active 86.7%, 520 out of 600 cases not used in training or predicting series. Outputs of this QSAR model were used as inputs to construct a network. The observed network has 1242 nodes (DBPs), 772,736 edges or DBPs with similar activity (sDBPs). The network predicted has 1031 nodes, 641,377 sDBPs. After edge-to-edge comparison, we have demonstrated that the predicted network is significantly similar to the observed one and both have distribution closer to exponential than to normal.