Deep neural networks have become state-of-the-art for small molecule property prediction. While these networks have excellent accuracy, they are limited in their generalization to out of domain examples, and struggle with the relative lack of labeled data. Relatedly, it is also expensive to acquire new sample labels for various molecular properties. Thus, various active learning techniques have been developed in order to better decide which samples should be labeled to maximize model gain. However, the majority of these techniques require additional compute, potentially including training entire new models. In a resource limited setting, it is desireable for the model to output an uncertainty estimation directly without additional compuational burden. Here we present various architectures and strategies for regressing to different methods of uncertainty estimation in order to predict the pKa of small molecules.

Regression uncertainty estimation for small molecule pKa active learning

ACS Spring 2022

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.