Adversarial validation tests whether a classifier can distinguish molecular fingerprints from the training and test partitions. Easy discrimination signals different chemical-space distributions and possible test-set extrapolation beyond the model’s applicability domain. section 5.3.2
Apply it as a practical distribution-shift check:
- Label training examples as 0 and test examples as 1, concatenate them, shuffle, and split the combined data. section 5.3.2
- Train a binary classifier on molecular fingerprints and evaluate its ability to recover the original partition labels. section 5.3.2
- Interpret low performance as structural similarity and high performance as distinguishable chemical distributions. section 5.3.2
Random-forest applicability-domain check
A Random Forest trained for adversarial validation produced PR AUC 0.504, indicating that the original training and test fingerprints were difficult to distinguish in this assessment.

