PREVAIL: Pre-trained Variational Adversarial Active Learning for Molecular Property Prediction

Abstract

Molecular property prediction is a fundamental task in drug discovery. The majority of the high-performing molecular property prediction methods currently were developed using deep learning techniques, which rely on massive labeled data. However, accurate molecular property annotation is time-consuming and expensive. Due to the fact that different samples usually have unequal importance in model training, we propose a pre-trained variational adversarial active learning, PREVAIL for short, to query the most informative samples to be annotated to reduce the annotation cost. Specifically, different from previous active learning whose initial set is sampled randomly, PREVAIL selects the most informative initial dataset by an autoencoder and K-Center greedy algorithm, which can avoid biases that affect the accuracy of the early decision-making process. Furthermore, PREVAIL simultaneously adapts the distribution of molecules and the information of the prediction task by incorporating the loss information of the molecular property prediction task into the latent space using task-aware variational adversarial active learning. Our benchmark experiments demonstrate that PREVAIL outperforms state-of-the-art active learning methods on molecular property prediction tasks.