A fair experimental comparison of neural network architectures for latent representations of multi-omics for drug response prediction

Table 9 Comparison of p values for all architectures computing the triplet loss or not

Data	Method	AUROC	AUPRC
Test	Early integration	0.047 (6/7)	0.016 (7/7)
	Omics stacking	0.156 (5/7)	0.156 (6/7)
	Super.FELT	0.219 (5/7)	0.813 (4/7)
	MOLI	0.813 (4/7)	0.813 (3/7)
	OmiEmbed	0.938 (3/7)	0.578 (4/7)
	MOMA	0.296 (2/7)	0.078 (1/7)
External	Omics stacking	0.219 (5/7)	0.078 (5/7)
	MOLI	0.463 (4/7)	0.297 (6/7)
	Super.FELT	0.578 (4/7)	0.938 (4/7)
	OmiEmbed	0.578 (4/7)	0.938 (4/7)
	MOMA	0.375 (2/7)	0.578 (2/7)
	Early integration	0.156 (1/7)	0.219 (1/7)

The p values are from a Wilcoxon signed-rank test, where an \(\alpha <0.05\) was regarded as significant. Additionally, we counted the number of times the method was better using the triplet loss and added it in parentheses. The methods are sorted in descending order with the criterion of the times the triplet loss improved the results separately for test and external data

ISSN: 1471-2105