Chemist-Computer Interaction: Representation Learning for Chemical Design via Refinement of SELFIES VAE

Representation learning for molecular structure is essential in helping chemists with novel drug discovery and other scientific tasks. Here we refine neural networks for chemical representation learning, especially for SELFIES VAE, and thereby improve upon generative models for chemical design and discovery. For model evaluation we propose five metrics (syntactic/semantic validities; degeneracy; emptiness proportion; and diversity of generation) and experimentally demonstrate that our refined model outperforms the standard models that exist in the scientific literature today, which is achieved by integrating the symbolic grammatical structure of compounds with statistical representation learning.