We were recently approached by a clinical-stage biotech client to investigate the potential repositioning of one of their lead compounds used to treat COPD. Specifically, we wanted to know whether this compound would have any activity against SARS-CoV-2.
As a basis, we have previously built machine learning models to predict antiviral activity using the RxRx-19 data set for training. https://covid.blog.biosymetrics.com/
A total of 1,237 compounds were used to train the data set (146 actives, 1,091 inactive) resulting in 8,367 chemical features. A robust feature selection algorithm was employed and tested using cross-validation to capture the 836 most important features. This algorithm reached an overall accuracy of 0.91 (0.93 precision at 10% recall). With this in mind, we used this model to predict the chemical activity of nearly 40K compounds not included in the assay. This model was also used to predict the antiviral activity of the test candidate compound, resulting in a predicted efficacy of 0.159, suggesting that this compound would likely be inactive against SARS-CoV-2 if tested in the RxRx-19 assay.
It is important to note that this finding does not necessarily preclude the compound from being repositioned as a succesful COVID-19 therapeutic as there are likely multiple mechanisms that can contribute to clinical improvements. However, this did indicate that the drug was unlikely to be succesful through this mechanism, and that other available treatments should be given priority. Recent works have demonstrated that depending on the specific assay, there can be large discrepancies in the activity profiles of identical compounds.