To customize the optimal treatment strategy for patients and achieve the purpose of precision medicine, it is of great significance to improve the accuracy of a cancer diagnosis. Specifically, cancer is generally further divided into multiple molecular subtypes, and different molecular subtypes adopt different treatment strategies to achieve the best therapeutic effect (8). For example, breast cancer is subdivided into four molecular subtypes: HER2-enriched, Luminal A, Luminal B, and Basal-like (9). Each subtype is associated with a unique panel of mutated genes. Therefore, the task of cancer subtype classification is to automatically identify defined subtypes based on the multi-omics measurement results of patients (10–13). In addition, with the growth of multi-omics data, there are still some potential cancer subtypes that need to be mined (14–16). Therefore, the identification of cancer subtypes is usually treated as a supervised classification problem, and the discovery of new subtypes is generally treated as a clustering problem.

To improve the survival rate of cancer patients, a large number of researchers studied and analyzed the factors affecting their survival times by collecting the survival times of cancer patients and using machine learning methods to discover possible survival rules (17, 18). Cancer survival analysis can be defined as a binary classification or a risk regression problem (19, 20). In a binary classification task, patients are divided into short-term and long-term survival groups, or low-risk and high-risk survival groups, according to a predefined survival time threshold (e.g., 5 years) (21). In risk regression tasks, the Cox proportional hazards model and its extensions are often used to calculate the risk score for each patient.

IC50 is widely used to assess the sensitivity of drug response, and it is the concentration of drug required to reduce the number of viable cells by half after drug administration (22). Drug response prediction tasks are the same as survival analysis tasks. The binary classification tasks and regression tasks are often used in drug response research (19, 23–26). In regression problems, drug response directly predicts IC50 values, while in binary classification tasks, a predefined threshold based on the distribution of IC50 values is used to predict drug response as sensitive or resistant.

In this review, the goal of biomarker discovery is to find genes associated with cancer prognosis by combining multi-omics data, which can advance the understanding of molecular mechanisms of cancer and offer new ideas for clinical diagnosis and treatment (27, 28). For example, in clinical practice in gastrointestinal cancer, CEA is the most commonly used marker. Furthermore, some biomarkers have been used for cervical cancer including miR-215-5p, miR-192-5p, KAT2B, PCNA, and CD86 (29). Biomarkers are widely identified by using the methods of feature selection and feature importance ranking in traditional ML (10, 30, 31). When analyzing the contribution of each feature in multiple omics sources, the feature will be set to 0 in turn, and then the performance of the classification or regression model will be calculated, and it will be compared to performance using all features (10).

In this section, we introduce the multi-omics cancer datasets that are widely used in the literature. The multi-omics datasets are shown in Table 1.

The Cancer Genome Atlas (TCGA) is a project jointly launched by the National Cancer Institute (NCI) and the National Human Genome Research Institute (NHGRI) in 2006 (35). It includes clinical information, histopathological images, and multiple omics data like genomics, transcriptomics, proteomics, and epigenomics. Especially, genomics and transcriptomics, are the most commonly used types of omics. For example, the data types of DNA methylation and copy number variation in genomics, as well as the data types of mRNA expression and miRNA expression in transcriptomics, appear most frequently in the literature on multi-omics-based ML integration methods. TCGA dataset currently includes a total of 33 types of cancer. In particular, all gastrointestinal cancer, including gastric cancer, colorectal cancer, liver cancer, etc., can be obtained in TCGA. TCGA is free and open, which greatly helps cancer researchers to improve the prevention, diagnosis, and treatment of cancer.

The Genomics of Drug Sensitivity in Cancer (GDSC) omics database was jointly developed by the Wellcome Trust Sanger Institute in the United Kingdom and the Massachusetts General Hospital Cancer Center in the United States (36). GDSC collects drug response data (IC50) of about 200 anticancer drugs in more than 1,000 human cancer cell lines. Variations in the cancer genome can affect the effectiveness of clinical treatment, and different targets have very different responses to drugs. Therefore, the GDSC data are really important for the discovery of potential tumor therapeutic targets, which have been widely used in anticancer drug screening.

In addition to the TCGA and GDSC datasets, other widely used databases also appear in the relevant literature, such as Molecular Taxonomy of Breast Cancer International Consortium (METABRIC), COSMIC Cell Lines, CPTAC, LinkedOmics, and the Cancer Imaging Archive (TCIA) (37–41).

Although multi-omics data can be used for cancer diagnosis using ML integration methods, there are still some problems with multi-omics data. We list some of the challenges that are quite general in the relevant literature (6, 42) as follows.

Here we will briefly introduce three subgroups of models applied in data integration: traditional ML models, classical deep learning models, and auto-encoder models.

Logistic regression (LR), support vector machine (SVM), random forest (RF) and Xgboost are widely used traditional ML models (30, 31, 47), before feeding the data to them, it generally needs to reduce the dimensionality of high-dimensional features of multiple omics data based on feature extraction methods such as nearest component analysis (NCA) (19, 23) and principal component analysis (PCA) (21), and then concatenate the dimensionality-reduced features and finally feed the concatenated features to the model.

In contrast, it is not necessary for classical deep learning models like fully connected neural networks (FCNNs) and convolution neural networks (CNNs) to reduce the dimensionality of omics features to very low dimensions, due the models can automatically learn useful information from high-dimensional space (11, 48–51).

Auto-encoder is an unsupervised neural network model where the network can be replaced by FCNN, CNN, or other DNNs (16). Auto-encoder compresses the data to a lower dimension, which is called encoding, and then reconstructs the original input data back, which is called decoding (11). Intuitively, auto-encoder can be used for dimensionality reduction which is similar to PCA, but its performance is stronger than PCA due to the neural network model can extract more effective new features (52). In addition to dimensionality reduction, new features learned by the auto-encoder can be fed into the supervised learning model for the tasks of classification or regression.

In recent years, Graph Neural Networks (GNNs) have shown strong capabilities in handling non-Euclidean graph-structured data by naturally combining network topology structure and the information of node and link, GNNs have been employed to integrate multi-omics data since the last 2 years. Wang et al. (10) proposed MOGONET, for each omics data type, a weighted sample similarity network was constructed according to omics characteristics as the input of GNN and used for the identification of biomarkers. In Xing et al. (53), MLE-GAT is presented to explore the correlation information between genes contained in omics data. It assumes that genes usually interact rather than acting alone, so the weighted correlation network analysis (WGCNA) is firstly used to convert each patient's omics data into a co-expression map as the input of GAT, and then GAT outputs each node feature as a weighted combination of its neighbors and the current node.

The transformer model is widely used in different fields such as natural language processing and computer vision, it is becoming one of the most frequently used deep learning models (54–56). The success of the transformer architecture depends on the multi-head attention mechanism that calculates the attention between different positions in the input sequence multiple times. The Transformer is applied to multi-omics-based integration techniques since 2021, which is a relatively new data integration method. For example, in these two papers (57, 58), the transformers are used to calculate the cross-attention between multi-omics features and histopathological image features.