Active Learning Enables Extrapolation in Molecular Generative Models

Although generaIve models hold promise for discovering molecules with opImized desired properIes, they oNen fail to suggest synthesizable molecules that improve upon the known molecules seen in training. We find that a key limitaIon is not in the molecule generaIon process itself, but in the poor generalizaIon capabiliIes of molecular property predictors. We tackle this challenge by creaIng an acIve-learning, closed-loop molecule generaIon pipeline, whereby molecular generaIve models are iteraIvely refined on feedback from quantum chemical simulaIons to improve generalizaIon to new chemical space. Compared against other generaIve model approaches, only our acIve learning approach generates molecules with properIes that extrapolate beyond the training data (reaching up to 0.44 standard deviaIons beyond the training data range) and out-of-distribuIon molecule classificaIon accuracy is improved by 79%. By condiIoning molecular generaIon on thermodynamic stability data from the acIve-learning loop, the proporIon of stable molecules generated is 3.5x higher than the next-best model. More recently, generaIve or inverse-design models have been proposed as a new paradigm for materials discovery due to their ability to efficiently navigate chemical space beyond what is present in exisIng databases. The goal of property-constrained molecular generaIon is to generate novel molecules that possess desirable properIes for the applicaIon of interest. Typically, a ground-truth oracle funcIon is defined for each molecule design task to quanItaIvely assess how well the generated molecules meet the desired molecular properIes. As a means to quickly approximate this oracle funcIon, property predicIon models are used as a surrogate model. These property predicIon models are first trained on a pre-exisIng dataset of molecular properIes to learn the mapping between the chemical structure of the molecules and their target molecular properIes.