Compounds created from previously working drugs for other diseases may be able to aid in the treatment of COVID-19. We can determine whether these new drugs are good or bad based on machine learning/algorithms that can analyze their properties.
There are several viral proteins with identified structures against which inhibitors have been identified and for which clinical studies are running at the moment. An example thereof is the SARS-2 RNA polymerase.
Favipiravir as well as related structures (see video) have been identified as binders (as noted in the FDA Landscape of Therapeutics), however to our knowledge no wider related structure search has been conducted. We have sucessfully identified analogues to some of the starting structures using the SwissSimilarity System on SwissADME. Crude docking analyses were done to a viral RNA dependant RNA polymerase to proove that these still bind at the appropriate site. One of the structures was chosen to dock using Swissdock to the true COVID-19 enzyme. This revealed binding at the appropriate site.
A machine learning algorithm could be applied to candidate data to determine whether it was a "good" or "bad" candidate. This succeeded with a 65% accuracy. Furthermore, we attempted to generate a classifier to identify the similarity between candidates and the endogenous binder ATP. Unfortunately due to the lack of data, this did not produce consistent results.