Bidirectional Generation of Structure and Properties Through a Single Molecular Foundation Model

Abstract Despite the impressive successes of deep learning approaches for various chemical problems such as property prediction, virtual screening, and de novo molecule design, separately designed models for specific tasks are usually required, and it is often difficult to synergistically combine these models for novel tasks. To address this, here we present a bidirectional molecular foundation model that can be used for both molecular structure and property inferences through a single model, inspired by recent multimodal learning methods such as VLP. Furthermore, thanks to the outstanding structure/property alignment in a common embedding space, experimental results confirm that our method leads to state-of-the-art performance and interpretable attention maps in both multimodal and unimodal tasks, including conditional molecule generation, property prediction, molecule classification, and reaction prediction.