Combining machine learning and molecular simulations to predict the stability of amorphous drugs
This work leverages the outcomes of molecular dynamics simulations to build bespoke descriptors that can inform machine learning models aimed at predicting the solubility of amorphous drugs. The latter are attracting an ever-growing interest from the pharmaceutical industry, as they are more soluble (and thus more bioavailable) than their crystalline counterparts. However, they also tend to crystallise – whether in a day or a year, it is very hard to tell. Machine learning can help in making predictions with respect to the tendency of amorphous drugs to re-crystallise, but up to now we have been using the traditional “one-molecule-in-vacuum” approach, where molecular descriptors re: the molecular species in isolations are used. Here, we generated models of amorphous glasses for each datapoint available (more than 150 glasses!), which allowed us to access dynamical and structural properties that really improved the accuracy and the reliability of our models. This paper has been published in JCP as part of the Special Collection: “Machine Learning Hits Molecular Simulations”, and it has also been selected as a Featured Article by the editors. Have a look at the paper online or grab a pdf here.