Moleculeace

After publication, it came to our attention that a software bug affected the split between training and test set, due to incorrect labeling of certain activity cliffs pairs. We fixed this bug and retrained all models on the corrected data. The same findings and conclusions as the original paper were observed. The data, figures, and results in this repository were corrected accordingly.

I changed the use of non-consecutive patience to consecutive patience. This is a more sensible way of doing early stopping and implementing it slightly affected individual results (models became slightly better on average). Changes in the code had no effect on any conclusions we draw in the paper after re-training all models. Along with the updated code, I updated the results file, so any changes in the results are traceable. For replicating our results from the paper, you should use release 2.0.1.

This release is the original code used to produce the results for the 'Exposing the Limitations of Molecular Machine Learning with Activity Cliffs' paper in 2022 (10.1021/acs.jcim.2c01073). If you want to exactly reproduce the results presented in the paper, you should use this release.