Table 2. Summary of the metrics to compare performance of the different search approaches over the entire search space. Any approaches with pretraining components were pretrained on 58 000 molecules. The starting dataset here refers to the dataset with 58 000 molecules.

Search algorithm	Unique new evaluations	Unique molecules with Fcomb > 0	Rate of discovery of molecules with Fcomb > 0	Molecules found better than starting dataset. (outstanding molecules)
BO-Learned	841	458	54%	4
BO-Mord	1273	406	31%	5
BO-Prop	1799	893	50%	2
SUEA	1037	1004	86%	1
EA	1637	544	32%	7
Rand	1236	0	0%	0