QSAR Classification Model for Antibacterial Compounds and Its Use in Virtual Screening

As novel and drug-resistant bacterial strains continue
to present
an emerging health threat, the development of new antibacterial agents
is critical. This includes making improvements to existing antibacterial
scaffolds as well as identifying novel ones. The aim of this study
is to apply a Bayesian classification QSAR approach to rapidly screen
chemical libraries for compounds predicted to have antibacterial activity.
Toward this end we assembled a data set of 317 known antibacterial
compounds as well as a second data set of diverse, well-validated,
non-antibacterial compounds from 215 PubChem Bioassays against various
bacterial species. We constructed a Bayesian classification model
using structural fingerprints and physicochemical property descriptors
and achieved an accuracy of 84% and precision of 86% on an independent
test set in identifying antibacterial compounds. To demonstrate the
practical applicability of the model in virtual screening, we screened
an independent data set of ∼200k compounds. The results show
that the model can screen top hits of PubChem Bioassay actives with
accuracy up to ∼76%, representing a 1.5–2-fold enrichment.
The top screened hits represented a mixture of both known antibacterial
scaffolds as well as novel scaffolds. Our study suggests that a well-validated
Bayesian classification QSAR approach could compliment other screening
approaches in identifying novel and promising hits. The data sets
used in constructing and validating this model have been made publicly
available.