. 2023 Jul 25;19(7):e1011323. doi: 10.1371/journal.pcbi.1011323

Table 1. Comparison of the results for the models trained on BF and FL images and CP features.

The subscript dmso is for the cases where DMSO plate-level normalization was applied to the data and the subscript site is for when the data was normalized at the imaging site level.

(a) Macro-F1 scores on the test sets for the five data splits.
	BF_dmso	FL_dmso	BF_site	FL_site	CP
Split 1	0.738	0.777	0.662	0.661	0.771
Split 2	0.821	0.799	0.770	0.762	0.801
Split 3	0.724	0.793	0.654	0.677	0.718
Split 4	0.710	0.738	0.676	0.645	0.739
Split 5	0.728	0.716	0.708	0.688	0.736
(b) F1 scores per MoA across all five test sets.
	BF_dmso	FL_dmso	BF_site	FL_site	CP
ATPase-i	0.605	0.701	0.650	0.683	0.779
AuroraK-i	0.683	0.675	0.713	0.671	0.746
HDAC-i	0.756	0.773	0.766	0.785	0.740
HSP-i	0.738	0.730	0.756	0.682	0.676
JAK-i	0.675	0.653	0.405	0.429	0.607
PARP-i	0.895	0.886	0.789	0.748	0.912
Prot.Synth.-i	0.711	0.793	0.520	0.646	0.711
Ret.Rec.Ag	0.740	0.769	0.767	0.796	0.786
Topo.-i	0.780	0.728	0.702	0.651	0.742
Tub.Pol.-i	0.887	0.854	0.850	0.865	0.845
DMSO	0.790	0.866	0.836	0.691	0.809
Macro average	0.751	0.766	0.705	0.695	0.759