Toward Predictive Chemical Deformulation Enabled by Deep Generative Neural Networks

The design of chemical formulations is a challenging, high-dimensional problem. In typical formulations, tens thousands ingredients are available for use, yet only tiny fraction end up in given formulation. Deformulation, the problem reverse engineering precise amounts each ingredient starting from just list ingredients, similarly challenging but key capability staying up-to-date with industry competitors. Here, we take advantage large, curated dataset CAS, division American Chemical Society, which offers consistent and highly structured representation identities their components to show that variational autoencoder neural network learns meaningful representations various product classes such as antiperspirants oral care. Furthermore, it can be used conjunction two-step sampling algorithm generate accurate amount suggestions deformulation. Deformulation using produces estimates significantly more than nearest neighbor methods, extrapolates better different previously seen provides way leverage large datasets industrially relevant capabilities.