Enhancing Molecular Expressiveness through Multi-View Representations

In the field of foundation models for materials science and chemistry, the quality of molecular representations plays a critical role in the success of machine learning models in downstream tasks. Transformer-based molecular representation models have shown great potential in these areas by generating high-quality latent representations of molecules. However these representations often fail to capture the full complexity of molecular structures, leading to suboptimal performance in predictive tasks. In this work we propose a simple yet novel multi-view representation method to improve the expressiveness of the molecular latent representations. We provide preliminary analysis of the proposed method which shows promising improvements compared to the conventional method, suggesting that the multi-view approach improves the quality of the latent representations.