IBM RXN for Chemistry

What it does

IBM RXN for Chemistry is a web platform that applies AI to two core synthesis tasks: retrosynthesis (working backward from a target molecule to identify feasible synthetic routes) and forward reaction prediction (predicting the product of a reaction given starting materials and reagents). You enter a molecule in SMILES format or draw it using the built-in editor, and the tool proposes routes and reaction conditions ranked by predicted feasibility.

The platform was recently updated with integration of Thieme’s Science of Synthesis database — a curated expert-authored collection covering reaction classes underrepresented in the patent literature that most models train on. IBM Research reports approximately 9× improvement in retrosynthesis prediction accuracy and ~3× improvement in forward prediction following this integration.

Best for

Chemists planning synthetic routes to target molecules, particularly for drug discovery, natural product synthesis, or materials chemistry. Also useful as a first-pass sanity check when evaluating whether a proposed molecule is synthetically feasible before running a de novo design campaign.

Pricing

Freemium. A free tier provides basic retrosynthesis and forward prediction. The full multi-step synthesis planning feature (planning routes up to 40+ steps) and higher-throughput access require a paid account. Institutional licenses available.

Strengths

• End-to-end synthesis planning from a single target molecule — not just one retrosynthetic step but multi-step route proposals
• Forward prediction validates proposed routes: given a reaction you’re planning, it predicts the likely product and flags low-confidence predictions
• Web-based with a molecule drawing editor — no Python required; accessible to researchers without ML infrastructure
• Science of Synthesis integration substantially improved accuracy, especially for reaction classes outside the patent literature mainstream
• Routes can be exported for use in electronic lab notebooks

Limitations

• Training data is biased toward well-documented reaction classes in patents and literature — truly novel or exotic reaction types remain less reliable
• Stereochemistry handling (which face of a molecule a reagent attacks, enantioselectivity) is less reliable than connectivity prediction
• Route proposals are ranked by model confidence, not by practical ease, cost of starting materials, or lab availability — always filter suggested routes through your practical constraints
• A proposed route that the model rates highly still needs expert evaluation and experimental validation before trusting it

How it compares

Related content

• Field Guide: Chemistry
• Glossary: Retrosynthesis, SMILES Notation