Improving synthesis prediction of small molecules at scale with RetroChimera
Microsoft says RetroChimera beats its component models by learning when to trust each one.
The retrosynthesis system combines R-SMILES 2 and NeuralLoc, then re-ranks their proposed reaction routes. Microsoft says blind tests showed PhD-level chemists preferred RetroChimera’s predictions over prior models and even recorded literature reactions. In a ten-target route assessment, it succeeded on nine targets, ahead of the compared systems. The implementation and weights are open-sourced under MIT license. Microsoft Research AI's note
The retrosynthesis system combines R-SMILES 2 and NeuralLoc, then re-ranks their proposed reaction routes. Microsoft says blind tests showed PhD-level chemists preferred RetroChimera’s predictions over prior models and even recorded literature reactions. In a ten-target route assessment, it succeeded on nine targets, ahead of the compared systems. The implementation and weights are open-sourced under MIT license. Microsoft Research AI's note
score 6