Samsung Advanced Institute of Technology

Retrosynthesis is an essential task in organic chemistry to identify a synthesis pathway of newly discovered materials. With the recent advance of deep learning, there is growing attempts to solve the retrosynthesis problem through transformer models, the state-of-the-art in neural machine translation, by converting the problem into a machine translation problem. However, the vanilla transformer provides dissatisfactory results lacking grammatical validity, chemical plausibility, and diversity in reactant candidates. In this study, we propose tied two-way transformers with latent modeling to solve those problems by cycle consistency check, parameter sharing, and multinomial latent variables. Experimental results with public and in-house datasets demonstrated that the proposed model improves retrosynthesis accuracy, grammatical error, and diversity. Qualitative evaluation results further confirm its ability to suggest valid and plausible results.