Microsoft's RetroChimera retrosynthesis model published in Nature, open-sourced under MIT license
Microsoft Research · rss · 2026-09-21
Microsoft Research published RetroChimera, a retrosynthesis prediction model, in Nature, and open-sourced its implementation and weights (MIT license, also on Microsoft Foundry).
- Architecture: combines two complementary sub-models — Transformer-based R-SMILES 2 generates precursors de novo (flexible but prone to hallucination), while GNN-based NeuralLoc selects reaction templates (more grounded but limited by template coverage)
- Ensembling: a learned learning-to-rank strategy votes across both models' predictions, handling both common and rare reaction classes
- Results: in blind tests, PhD-level chemists preferred RetroChimera's single-step disconnections over prior models and even literature-recorded reactions; it succeeded on 9 of 10 challenging multistep targets versus 5/4/2 for baselines
- Impact: could accelerate design-make-test cycles in drug discovery and materials, enabling closed-loop self-improving synthesis planning with lab automation
More from Apps
- YC-backed Roma turns your to-do list into an agent command center across Gmail and Notion — testingcatalog · 2026-11-12
- Dev builds interactive quantum physics playground with Gemini: fire wavepackets at double slits — BenBlaiszik · 2026-11-12
- Gemini-built playground simulates double-slit experiment in near real time via WebGL2 — BenBlaiszik · 2026-11-12
- Ploid Deep Research finds your WeChat from a single video frame, in minutes — thetripathi58 · 2026-11-12
- Epic AI MegaJam finals in SF Nov 14-15: teams of 4 plus unlimited AI agents — chrisfirst · 2026-11-12
- Clinical pre/post-surgery video motion data trains one model across neurology, spine, senior care — zawthet · 2026-11-12