Separating drugs from nondrugs: a statistical approach using atom pair distributions

A computational approach to quantify the druglike character of chemical compounds is presented. For this purpose, the distribution of atom types and their pair-wise combinations in known drugs and nondrugs was examined. Statistical analysis of the occurrence probabilities was used to derive a drug-likeliness score on a logarithmic scale. "Typical" pharmaceutical agents exhibit scores greater than 0.3, while for ordinary substances, values below 0 are expected. Although any kind of fitting or error minimization scheme is absent in this method, confirmed drugs are predicted with an accuracy of at least 71%. Many falsely predicted nondrugs were found to closely resemble actual drugs or to contain unsuitable substitution patterns that can easily be ruled out by applying medicinal knowledge. As the outlined method is computationally inexpensive, this drug-likeliness score can therefore be used as a filter for the in silico screening of large substance databases.