Application of Metric Transformation in One-Step Retrosynthesis

Ngoc Trinh Hung Nguyen, Milo Roucairol, Tristan CazenaveOpenReview 2025 · Submission to ICLR 2025 · Rejected · OpenReview

Abstract

In this article, we investigate the impact of Deep Metric Learning and Transformer architecture on predicting the retrosynthesis of Simplified Molecular Input Line Entry System (SMILES) chemical compounds. We demonstrate that combining the Attention mechanism with Proxy Anchor Loss is effective for classification tasks due to its strengths in capturing both local and global contexts and differentiating between various classes. Our approach, which requires no prior chemical knowledge, achieves promising results on the USPTO-FULL dataset, with accuracies of 53.4\%, 83.8\%, 90.6\%, and 97.5\% for top-1, top-5, top-10, and top-50 predictions, respectively. We further validate the practical application of our approach by correctly predicting the retrosynthesis pathways for 63 out of 100 randomly selected compounds from the ChEMBL database and for 39 out of 60 compounds selected by Bayer's chemists and from PubChem.
BibTeX
@misc{nguyen2025application,
  author = {Ngoc Trinh Hung NGUYEN and Milo Roucairol and Tristan Cazenave},
  title = {{Application of Metric Transformation in One-Step Retrosynthesis}},
  year = {2025},
  url = {https://openreview.net/forum?id=o1efpbvR6v},
  howpublished = {OpenReview},
  note = {Submitted to ICLR 2025},
}

Loading related content...

Figure