Responsive to calls from lawmakers, the USPTO has recently announced a broad set of measures to increase the quality of drug patents ex ante, before they are granted, as a way of in the US. However, there is currently no way to tell which patent applications cover inventions that will lead to FDA-approved drugs, potentially compromising the efficiency and effectiveness of the agency’s efforts. Nor is it known how drug patent applicants differ from others in their use of examination tactics such as those that increase the number of patents that cover a drug. We address these informational deficits predictively and descriptively through an analysis of patents issued in 2005-2015 that cover drugs as identified through their listing in the FDA’s “Orange Book.” We find that even within the same technology areas, patent applications that mature into drug patents differ from other patent applications along several dimensions, showing intensified use of continuations, terminal disclaimers, Track One examination acceleration, and applicant- submitted prior art. Applying machine learning models, we find traits publicly observable at publication and grant to be reasonably predictive of a patent’s eventual designation as a drug patent. A random forest model trained on publication characteristics is associated with an area under the curve (AUC) statistic of 0.83, which improves to 0.91 when grant characteristics are used. The AUC statistic for predicting the first patent associated with a drug to be listed in the OB based on grant characteristics is ~0.9, and for subsequent patents, it is 0.97.
Note: Research papers posted on SSRN, including any findings, may differ from the final version chosen for publication in academic journals.