Table 5:

Type and number of data modalities tested by each inter-modality data harmonization approach (free modality scope)

Method name	No. of modalities trained	3D chromosome structure	DNA methylation	Epigenetic peak data	DNA-Protein binding	DNA-RNA interactions	RNA-Protein interactions	Protein-Protein interactions	Genomics	Transcriptomics	Citation
DeepMF	1	X	X	X	X	X	X	X	X	O	[70]
JIVE	1	X	X	X	X	X	X	X	X	O	[71]
GCCA	2	X	X	X	X	X	X	X	O	O	[72]
NetICS	3	X	O	X	X	X	X	X	O	O	[73]
DIABLO	2	X	O	X	X	X	X	X	X	O	[20]
iCluster	3	X	O	X	X	X	X	X	O	O	[62]
GFA	2	X	O	X	X	X	X	X	X	O	[63]
MOFA	2	X	O	X	X	X	X	X	X	O	[74]
seurat*	2	X	X	O	X	X	X	X	X	O	[69]
SNF	2	X	O	X	X	X	X	X	X	O	[75]
NMF	2	X	O	X	X	X	X	X	X	O	[65]
iNMF	2	X	O	X	X	X	X	X	X	O	[66]
LIGER*	2	X	O	X	X	X	X	X	X	O	[67]
sMBPLS	3	X	O	X	X	X	X	X	O	O	[76]

Note that GCCA [72], seurat [69], and LIGER [67] are specific to single-cell data and the others are intended for bulk data. “DNA methylation” in this context refers specifically to the ratio of signal between methylated and unmethylated alleles. In contrast to Table 4, the quantity of modalities represents the quantity of modalities on which the algorithm was tested and does not reflect the modalities with which the algorithm is compatible. For simplicity, some modalities have been aggregated, e.g., transcriptomics data include both gene expression and small RNA data, which gives the illusion that DeepMF [21] and JIVE [71] were trained on unimodal data. Some methods are capable of handling proteomics, metabolomics, or medical images, but these are excluded because they are not a focus of this review. A link to each method is provided for easy reference.