ML model predicts chemical reactions with much higher accuracy than chemists

> The algorithm developed by Lee and his group uses tools in pattern recognition to recognise how chemical groups in molecules react, by training the model on millions of reactions published in patents.

While I am excited for the potential machine learning to improve retrosynthesis, patent literature would seem like an especially unlikely source for reliable information on reaction-condition training data.

Actual paper appears to be accessible here: https://pubs.acs.org/doi/pdf/10.1021/acscentsci.9b00576

>The algorithm developed by Lee and his group uses tools in pattern recognition to recognise how chemical groups in molecules react, by training the model on millions of reactions published in patents.

https://en.wikipedia.org/wiki/Survivorship_bias