PubChem BioAssays as a data source for predictive models

Predictive models are widely used in computer-aided drug discovery, particularly for identifying potentially biologically active molecules based on training sets of compounds with known activity or inactivity. The use of these models (amongst others) has enabled "virtual screens" to be used to identify compounds in large data sets that are predicted to be active, and which would thus be good candidates for experimental testing. The PubChem BioAssay database contains an increasing amount of experimental data from biological screens that has the potential to be used to train predictive models for a wide range of assays and targets, yet there has been little work carried out on using this data to build models. In this paper, we take an initial look at this by investigating the quality of naive Bayesian predictive models built using BioAssay data, and find that overall the predictive quality of such models is good, indicating that they could have utility in virtual screening.