Box 1 | Application of a Bayesian model for clinical forecasting

We have developed a Bayesian network model that calculates the probability of Phase III success and New Drug Application (NDA) approval for a drug, given pre-Phase III safety and efficacy data¹². The model operates on summary statistics (for example, mean, variance or frequency distributions) and does not require access to raw, elemental data, which can be difficult to obtain. Publicly available data — including therapeutic indices, in vitro and in vivo efficacy data, early clinical efficacy data, true versus surrogate biomarkers or endpoints and drug-invention source (either in-licensed or developed in-house) — for 500 new chemical entities (NCEs) stratified by therapeutic class were used to develop the model.

The model was validated on an independent data-set consisting of successful and failed drugs for one class (antineoplastics), and found to perform with 78% accuracy (80% sensitivity and 76% specificity)¹². Entering hypothetical models with a range of sensitivity and specificity values into the same Monte Carlo simulation framework used to assess the Bayesian network model showed that model sensitivity and specificity values of only 60% or better demonstrate a potential financial benefit over the recent reported performance of the pharmaceutical industry (see figure).