Teaching free energy calculations to learn

Alchemical free energy calculations have been a useful tool for prioritizing compounds for synthesis in structure-based drug discovery programs, but suffer from numerous limitations---the foremost of which is the inability to learn from newly collected data to improve future predictions. Conversely, machine learning approaches are typically data-hungry, requiring large quantities of experimental data before useful predictions can be made. In this talk, we consider the future of computer aided drug discovery where hybrid physical / machine learning models can exploit the physical framework of statistical mechanics to achieve data efficiency while taking advantage of the learnability that deep learning models provide.