Improving the Prediction of Clinical Success Using Machine Learning

Abstract

AbstractIn pharmaceutical research, assessing drug candidates’ odds of success as they move through clinical research often relies on crude methods based on historical data. However, the rapid progress of machine learning offers a new tool to identify the more promising projects. To evaluate its usefulness, we trained and validated several machine learning algorithms on a large database of projects. Using various project descriptors as input data we were able to predict the clinical success and failure rates of projects with an average balanced accuracy of 83% to 89%, which compares favorably with the 56% to 70% balanced accuracy of the method based on historical data. We also identified the variables that contributed most to trial success and used the algorithm to predict the success (or failure) of assets currently in the industry pipeline. We conclude by discussing how pharmaceutical companies can use such model to improve the quantity and quality of their new drugs, and how the broad adoption of this technology could reduce the industry’s risk profile with important consequences for industry structure, R&D investment, and the cost of innovation