Table 2 Comparison of the linear probing evaluation of the learned representations against fully supervised methods²⁰

Type	Method	AUC	F1	AUC > 0.9	AUC > 0.8	AUC > 0.7
Linear probing	CLOOME	0.714 ± 0.20	0.395 ± 0.32	57	84	109
	CellProfiler	0.655 ± 0.20	0.273 ± 0.32	35	63	84
Supervised	ResNet	0.731 ± 0.19	0.508 ± 0.30	68	94	119
	DenseNet	0.730 ± 0.19	0.530 ± 0.30	61	98	121
	GapNet	0.725 ± 0.19	0.510 ± 0.29	63	94	117
	MIL-Net	0.711 ± 0.18	0.445 ± 0.32	61	81	105
	M-CNN	0.705 ± 0.19	0.482 ± 0.31	57	78	105
	SC-CNN	0.705 ± 0.20	0.362 ± 0.29	61	83	109
	FNN	0.675 ± 0.20	0.361 ± 0.31	55	71	90

For each method the performance metrics area under the receiver operating characteristic curve (AUC) and F1-score are shown, along with their standard deviation (n = 209 tasks), and the number of tasks with an AUC higher than 0.9, 0.8, and 0.7. Note that the CLOOME encoders do not have access to any activity data. The features produced by the CLOOME encoder are still predictive for activity data as shown by fitting a logistic regression model, considered as linear probing. CLOOME reaches the performance of the several supervised methods, which indicates transferability of the learned representations²³. The best method in each category is marked in bold.

Quick links

Search