Combining machine learning and molecular simulations to predict the stability of amorphous drugs
Trent Barnard, Gabriele C. Sosso
Abstract
Amorphous drugs represent an intriguing option to bypass the low solubility of many crystalline formulations of pharmaceuticals. The physical stability of the amorphous phase with respect to the crystal is crucial to bring amorphous formulations into the market-however, predicting the timescale involved with the onset of crystallization a priori is a formidably challenging task. Machine learning can help in this context by crafting models capable of predicting the physical stability of any given amorphous drug. In this work, we leverage the outcomes of molecular dynamics simulations to further the state-of-the-art. In particular, we devise, compute, and use "solid state" descriptors that capture the dynamical properties of the amorphous phases, thus complementing the picture offered by the "traditional," "one-molecule" descriptors used in most quantitative structure-activity relationship models. The results in terms of accuracy are very encouraging and demonstrate the added value of using molecular simulations as a tool to enrich the traditional machine learning paradigm for drug design and discovery.