search

kaist-amsg / LocalRetro

Retrosynthesis prediction for organic molecules with LocalRetro
81 stars 24 forks source link

Information leak? #2

Closed chaoyan1037 closed 2 years ago

chaoyan1037 commented 2 years ago

Hi, some people report that your method can not reproduce reported results with canonical test input as preprocessed in RetroXpert. Have you run your model with data processed by RetroXpert?

shuan4638 commented 2 years ago

@chaoyan1037 Yes, I also noticed the issue of data leakage as mentioned in RetroXpert. We did run our model with data processed by RetroXpert using canonicalize_products.py, the results are identical with the reported results. Since our model is graph-based approach and the bond representation is directed, the data leakage should not be an issue in our approach.

chaoyan1037 commented 2 years ago

That is great! Thanks for the confirmation!