Table 1.

AUPR score for the two transfer learning modifications of the chemogenomic neural network CN, based on a single train/validation/test split of the DBEColi dataset

	Raw ($S_1$)	Orphan proteins ($S_2$)	Orphan molecules ($S_3$)	Double orphan ($S_4$)
Chemogenomic neural network (CN)	$39.57\pm 4.17$	$26.74\pm 2.49$	$43.74\pm 2.35$	$\mathit{24.63}\pm \mathit{1.89}$
Curriculum learning	$45.06\pm 2.64$	$21.43\pm 3.62$	$\mathit{51.45}\pm \mathit{3.03}$	$20.97\pm 2.70$
$CN-feaMLP$	$\mathit{50.616}\pm \mathit{2.71}$	$\mathit{30.89}\pm \mathit{4.93}$	$42.04\pm 2.95$	$\mathit{25.77}\pm \mathit{3.60}$

The curriculum learning line corresponds to pre-training the molecule encoder of CN on the DBHuman dataset. The standard deviations are obtained by repeating 5 times the evaluation procedure