Table 2.

Deep learning-based multi-omics integration approaches including case studies on breast cancer. Subtype-GAN [39], Denoising autoencoder for accurate CAncer Prognosis prediction (DCAP) [41], DeepProg [42], BRCA Multiomics [38], Multi-Omics Late Integration (MOLI) [43], Survival Analysis Learning with Multi-Omics neural Networks (SALMON) [37], DeepType [36], Concatenation AutoEncoder (ConcatAE) and Cross-modality AutoEncoder (CrossAE) [64], IntegrativeVAEs [65], Drug Response analysis Integrating Multi-omics (DRIM) [66].

Software	Arch.a	Purpose	Highlights
Subtype-GAN	GAN	To extract low-dimension features for predicting novel biomarkers and patient stratifications.	The first algorithm to explore the potential of generative adversarial network (GAN) architecture to improve the feature selection process by autoencoder (AE) methods.
DCAP	AE	To predict differentially expressed genes (DEGs) and to discriminate high- and low-risk groups of patients based on predicted DEGs.	Pan-cancer risk prediction system. It ranks the importance of omics data types by mRNA expression > miRNA expression > DNA methylation > copy number variations (CNVs).
DeepProg	AE	To predict patient survival subtypes using supervised machine learning algorithms from reduced dimensions by AE.	Trains on pan-cancer datasets to allow learning from well-established survival of cancer types to predict that for other less-studied cancer types. Flexible using input data types (e.g., mRNA expression).
BRCA Multiomics	MLP	To predict survival and drug responses at the same time by combining two multilayer perceptron (MLP) inferences using survival datasets from TCGA and drug response datasets from GDSC, respectively.	The tool focuses on breast cancer omics data and clinical outcomes and tries to build a connection between patient survival and treatment outcomes to predict if the treatment indeed improves the patient's condition.
MOLI	AE	To predict drug responses from selected features by training on each omics data type separately and then concatenating them into one representation.	MOLI employs a ‘late integration’ strategy and trains on drug response datasets targeting biological pathways rather than specific cancer types to hypothesise other non-traditional drugs for treating BC.
SALMON	MLP	To predict patient survival and characterise which data types are most pivotal predictors by incorporating omics data and clinical annotations (e.g., age).	SALMON groups patients by their ages at diagnosis (young: 26–50, middle: 51–70, elderly: 71–90). It identified that PR status is most predictive for the young group, ER status for the middle group and mRNA co-expression modules for the elderly group.
DeepType	MLP	To extend gene markers (218 DEGs) for breast cancer patient stratification by integrating omics data types and previous PAM50 subtypes.	The first deep learning-based method for patient stratification using mRNA expression only. The involvement of prior knowledge (PAM50 subtypes) addresses de novo clustering problems.
ConcatAE and CrossAE	AE	To question the essence of multi-omics integration, the expression similarity or the difference between omics data types, which is more informative for patient survival prediction.	By comparing learning from the similarity and from the difference between the expression in omics data types, it reports that the expression difference is a stronger predictor.
IntegrativeVAEs	AE	To investigate the inner architectures of AE for feature selection for classifying patient data by clinical annotations (e.g., PAM50 labels and metastasis status).	Patient samples are labelled by distance relapse and the co-effects of gene expression, CNA and clinical annotations are learned by different inner designs of AE for predicting relapse possibilities.
DRIM	AE	To model drug sensitivity from cancer cell lines and drug perturbation by selecting DEGs and analysing them according to pathway enrichment analysis.	DRIM provides a user-friendly website to select drug/cell line of interest for non-experts and allows users to customise the feature selection methods.

^aArch.: The deep learning architecture mainly used in these studies.