Fig. 5
From: Applying atomistic neural networks to bias conformer ensembles towards bioactive-like conformations
Median BEDROC of bioactive-like (A, B) and non-bioactive conformations (C, D) on harder test sets containing only molecules having 250 generated conformers and less than 5% bioactive-like conformations, for the random (A, C) and scaffold splits (B, D). Each point represents a split. AtNNs and TFD2SimRefMCS show higher median BEDROC of bioactive-like conformations than bioactivity-unaware baselines indicating better early enrichment of desirable conformations. They also show lower median BEDROC of non-bioactive conformations, indicating higher impoverishment (lower early enrichment) of undesirable conformations