SPEAR: Self-supervised post-training enhancer for molecule optimization

Abstract

The molecular optimization task is to generate molecules that are similar to a target molecule but with better chemical properties. Deep Generative Models (DGMs) have shown initial success in automatic molecule optimization. However, the training of DGMs often suffers from limited labeled molecule pairs due to the ad-hoc and restricted molecule pair construction. To solve this challenge and leverage the entire unpaired molecule database, we propose Self-Supervised Post-training EnhAnceR method (SPEAR) to enhance any graph-based DGMs for molecule optimization. SPEAR mines molecular structure knowledge and learns the molecule generation procedure in a purely self-supervised fashion. Unlike most self-supervised deep learning models that rely on pre-training for better molecule representation, the SPEAR method is applied as post-processing step to enhance molecule optimization during inference time for DGMs without additional training. Our SPEAR model can be efficiently incorporated into any DGM model as part of the inference procedure. We evaluated SPEAR against several state-of-the-art DGMs, SPEAR successfully improved the performance of all DGMs and obtained 5-21% relative improvement over its corresponding DGM models in terms of success rate.