About This Demo

This is an interactive one-step retrosynthesis tool built as part of the AI for Chemistry (AI4Chem) course assignment (HW4). Given a target molecule in SMILES notation, the model predicts plausible sets of reactants that could be used to synthesize it.

What is retrosynthesis?

Retrosynthesis is the process of working backwards from a target molecule to identify simpler precursors. Traditionally performed by expert chemists, it can now be assisted by sequence-to-sequence deep learning models trained on large reaction databases.

How it works

  1. Input: Paste a SMILES string, or click one of the 10 curated example molecules.
  2. Preview: The backend generates a 2D depiction and an RDKit 3D conformer via ETKDG.
  3. Beam search: Chemformer (a BART-based transformer) decodes up to N reactant SMILES in parallel beams. A confidence estimate is derived from the beam score distribution.
  4. Output: Each predicted reactant set is rendered in 2D and 3D for inspection.

The backend runs as a Hugging Face Space (FastAPI) using the fine-tuned USPTO-50K checkpoint released by MolecularAI.

Features

  • 2D molecular rendering (SmilesDrawer 2.0)
  • 3D conformer rendering (3Dmol.js)
  • Data-driven success probability estimation
  • Beam search retrosynthesis with 1–10 parallel outputs
  • 10 curated example drug molecules

References

  1. Irwin, R.; Dimitriadis, S.; He, J.; Bjerrum, E. J. Chemformer: A Pre-Trained Transformer for Computational Chemistry. Mach. Learn.: Sci. Technol. 2022, 3 (1), 015022. https://doi.org/10.1088/2632-2153/ac3ffb

Open-source repositories:

Acknowledgements

This demo was developed as Homework 4 for the AI for Chemistry course. The full-stack implementation (FastAPI backend, Vue 3 frontend, Hugo integration) was built with the assistance of GitHub Copilot powered by Claude Sonnet 4.6. The fine-tuned Chemformer model and training data are credited to the MolecularAI team at AstraZeneca.