Global Bayesian Models for the Prioritization of Antitubercular Agents

To aid the creation of novel antituberculosis (antiTB) compounds, Bayesian models were derived and validated on a data set of 3779 compounds which have been measured for minimum inhibitory concentration (MIC) in the Mycobacterium tuberculosis H37Rv strain. The model development and validation involved exploring six different training sets and 15 fingerprint types which resulted in a total of 90 models, with active compounds defined as those with MIC < 5 mu M The best model was derived using Extended Class Fingerprints of maximum diameter 12 (ECFP_12) and a few global descriptors on a training set derived using Functional Class Fingerprints of maximum diameter 4 (FCFP_4). This model demonstrated very good discriminant ability in general, with excellent discriminant statistics for the training set (total accuracy: 0.968; positive recall: 0.967) and a good predictive ability for the test set (total accuracy: 0.869; positive recall: 0.789). The good predictive ability was maintained when the model was applied to a well-separated test set of 2880 compounds derived front a commercial database (total accuracy: 0.73-, positive recall: 0.72). The model revealed several conserved substructures present in the active and inactive compounds which are believed to have incremental and detrimental effects on the MIC, respectively. Strategies for enhancing the repertoire of antiTB compounds with the model, including virtual screening of large databases and combinatorial library design, are proposed.