mRNA is a single-stranded RNA molecule that carries genetic information transcribed from DNA. It plays a crucial regulatory role in protein synthesis within the cell (Qin et al., 2022). mRNA expression data provide insights into gene function and activity. Investigating fluctuations in gene expression levels can elucidate disease development mechanisms. In cancer research, mRNA expression profiling has emerged as an essential element in elucidating cancer progression mechanisms. Studies show that dysregulation of specific genes can result in uncontrolled cell proliferation, a major factor in cancer development (Leibovitch and Topisirovic, 2018). For example, Li et al. (2017) used GA with a KNN classifier to classify mRNA data from 9,096 tumor samples of 31 types with 90% precision. Similarly, Kim et al. (2020) identified key genes that accurately distinguish 21 types of tumors by using ANOVA tests on mRNA data from cancer and normal samples. Therefore, studying mRNA expression data to find oncogenes helps in early cancer diagnosis and more accurate classification, improving treatment.

miRNAs are small noncoding RNAs present in plants and animals, typically 20 to 24 nucleotides long. They play a critical role in the regulation of cellular processes (Cui et al., 2025). miRNA controls oncogenes and tumor suppressor gene expression by degrading mRNAs or inhibiting their translation (Tang et al., 2021; Galagali, 2020). For example, in non-small cell lung cancer, high let-7 expression reduced lung cancer cell growth and inhibited differentiation (Pop-Bica et al., 2020). In gastric cancer, certain miRNAs inhibit the expression of the phosphatase and tensin homolog (PTEN) gene and promote cancer cell growth and invasion (Ashrafizadeh et al., 2020). Wang et al. (2019) combined GA with random forest (RF) for pan-cancer classification of miRNA data from 32 tumor types, achieving 92% sensitivity. To more robust and reliable set of miRNA features capable of distinguishing different types of tumor, Lopez-Rincon et al. (2019). developed an integrated feature selection algorithm for an accfor ante classification of 28 types otypes of tumorsth reliable miRNA features. Therefore, studying miRNA functions is vital for accurate cancer classification and early diagnosis, significantly impacting treatment and prognosis.

lncRNAs are RNA molecules with transcript sequences of more than 200 nucleotides. Although they do not encode proteins, they regulate biological processes such as gene expression, development, and differentiation (Chen et al., 2021). Initially considered as genomic noise, lncRNAs are now recognized as important in cancer development. Changes in their expression can serve as diagnostic markers (Nandwani et al., 2021; Fang and Fullwood, 2016). Analyzing lncRNA data has identified potential biomarkers and distinguished between tumor types (Al Mamun and Mondal, 2019a; Al Mamun and Mondal, 2019b; Al Mamun et al., 2020). Therefore, understanding the roles of lncRNAs is crucial for early cancer diagnosis and treatment.

CNV refers to the variation in the number of copies of a particular gene present in an individual’s genome (Pös et al., 2021). Genes such as BRCA1, CHEK2, ATM, and BRCA2 have strong associations with cancers like breast cancer (Hu et al., 2018). Zhang et al. (2016) proposed using a Dagging classifier to categorize CNV data from six cancer types, highlighting key features for accurate classification. Therefore, studying CNV helps explore cancer pathogenesis, aiding early diagnosis and treatment selection.

DNA methylation, an epigenetic modification, involves adding a methyl group to DNA, usually suppressing gene expression (Liu et al., 2016). It is crucial for normal cellular functions and implicated in cell differentiation and tumorigenesis. Dysregulated methylation, such as hypermethylation of CpG islands in promoter regions, can silence tumor suppressor genes or reduce oncogenic miRNA transcription, increasing cancer risk (Formosa et al., 2013). Liu et al. (Liu et al., 2019) used methylation data from 27 cancers types and proposed machine learning and deep learning strategies for accurate cancer differentiation. Therefore, DNA methylation is closely related to the occurrence and development of cancer, and the analysis and study of methylation is very important in the field of cancer diagnosis.

The development of cancer is a very complex process that is not simply caused by the occurrence of abnormalities in one type of data, but often involves multiple histological pathological processes. Therefore, data mining analysis based on single omic data has certain one-sidedness and limitations. In recent years, with the rapid development of next-generation genomic technologies, a large amount of genomic data of different types of cancers has been accumulated, and more and more researchers have started to integrate multiple omic data to conduct systematic and complete analysis of the mechanisms of cancer occurrence, and cancer research is developing from single omic to multi-omics. Integrated multi-omics analysis can make up for the lack of information in single-omics data and provide a comprehensive view of patients, and enable researchers to explore the relationship between cancer and genes from multiple perspectives, so as to perform early cancer diagnosis more accurately.

Table 1 summarizes the characteristics of common pan-cancer data types, including mRNA, miRNA, and DNA methylation.

TCGA is the largest human tumor genome sequencing database globally (Weinstein et al., 2013). Jointly sponsored by the National Human Genome Research Institute (NHGRI) and the National Cancer Institute (NCI), this major research project was officially launched in 2005. TCGA has sequenced 33 common cancers and over 11,000 tumor samples, using genomic analysis technology to enhance understanding of tumor mechanisms and improve cancer diagnosis and treatment capabilities (Tomczak et al., 2015). TCGA currently provides mRNA expression data, miRNA expression data, DNA methylation data, CNV data, and other high-throughput sequencing data. Researchers can access these datasets through the Genomic Data Commons (GDC) Data Portal, the primary data source for many cancer researchers.

GEO is a subdatabase of the National Center for Biotechnology Information (NCBI). This free and publicly accessible repository houses biological data from gene chips, second-generation sequencing, and other high-throughput functional genomics experiments. It includes submissions from over 16,000 laboratories and research teams worldwide, featuring 175,825 datasets with 5,069,606 data samples. GEO supports data download capabilities, enabling users to obtain samples or datasets of interest. Additionally, it offers tools to discover genes of interest and their expression profiles, as well as to identify genes with similar expression patterns.

UCSC Xena is a cancer genomics data analysis platform developed by the UCSC Cancer Genome Browser (Navarro Gonzalez et al., 2021). This platform collects and standardizes data from several major cancer research projects such as TCGA, ICGC, and TARGET, facilitating subsequent analysis (Consortium et al., 2010). UCSC Xena encompasses multiple levels of data, including copy number, methylation, gene expression, protein expression, and mutation data. It provides user-friendly data analysis and visualization tools. Researchers can easily analyze or download organized data with link clicks and can also upload their data for analysis. This flexibility considerably aids in the advancement of genomic research.

The Clinical Proteomic Tumor Analysis Consortium (CPTAC) is a comprehensive proteomic and genomic research program initiated by the National Cancer Institute (NCI) that aims to accelerate the understanding of cancer biology through the integration of large-scale proteomic and genomic analysis (Mesri et al., 2024). The consortium systematically identifies, quantifies, and analyzes proteins from cancer biospecimens characterized by genomic data to improve cancer prevention, early diagnosis, treatment, and prognosis. CPTAC provides a rich source of public data, serving as a critical resource for researchers studying pan-cancer proteomics. Its data, which includes protein abundance, post-translational modifications, and mass spectrometry data, is often used in combination with genomic data to provide a multi-layered view of tumors, enabling the discovery of new biomarkers and therapeutic targets.

The Cancer Genomics Cloud (CGC), an NCI-funded resource powered by Seven Bridges, is a secure and scalable cloud-based platform designed to overcome the challenges associated with accessing, sharing, and analyzing massive, diverse multi-omics datasets (Subramanian et al., 2021). The platform achieves this by co-localizing three essential components within the cloud: major cancer datasets like The Cancer Genome Atlas (TCGA) and Clinical Proteomic Tumor Analysis Consortium (CPTAC); over 400 bioinformatics tools and best-practice workflows; and the high-performance computational capabilities for large-scale analysis. The CGC simplifies the user experience by enabling researchers to browse, query, and filter datasets, run their entire analysis workflow on the platform, and even integrate their own private tools and data.

Building on the data sources described above, the following section reviews computational methods for pan-cancer classification.

Recent advancements in supervised deep learning have demonstrated remarkable efficacy in pan-cancer classification through tailored architectural innovations. Sun et al., 2018) introduced GeneCT, an artificial neural network (ANN) framework designed to classify 11 tumor types using raw mRNA expression data without feature engineering, achieving 98.2% accuracy and underscoring the potential of end-to-end learning in omics analysis. Complementing this approach (Cava et al., 2023), applied principal component analysis (PCA) to reduce data dimensionality before deploying the model. The neural network achieved a mean accuracy of 84%, the random forest reached 86%, and XGBoost achieved the highest performance with a mean accuracy of 90%. To address the challenges of limited sample sizes in specific cancer types (Cho et al., 2023), proposed a meta-learning method that integrates multi-omics data (transcriptomics, proteomics, and clinical data from TCGA) to create predictive models using survival information from 17 cancer types. Their approach requires fewer samples than conventional deep learning models, effectively mitigating data scarcity issues. Expanding this paradigm (Divate et al., 2022), employed deep neural networks (DNNs) to classify 33 cancer types. Their methodology integrated expression-based gene screening with SHAP (Shapley Additive exPlanations) interpretability, identifying critical biomarkers and achieving superior performance in distinguishing cancers from healthy controls.

To address high-dimensional data challenges (Wu et al., 2024) developed DeepMoIC, a method combining deep graph convolutional networks (GCNs) with autoencoders for cancer subtype classification. By constructing a patient similarity network (PSN) and leveraging GCNs, DeepMoIC outperformed existing models on multi-omics datasets, highlighting its potential for precision oncology. (Li et al., 2025) introduced DGHNN, a deep graph and hypergraph neural network for pan-cancer related gene prediction that takes biological pathways into consideration. This method applies a deep graph and hypergraph neural network to encode higher-order information in protein interaction networks and biological pathways. This approach, along with the introduction of skip residual connections and a feature tokenizer with a transformer for classification, demonstrates how advanced network architectures can capture the multi-level complexity of biological systems, setting a new standard for performance. (Li et al., 2020) tackled CNV sparsity by coupling Monte Carlo feature selection (MCFS), which evaluates feature stability via randomized sampling, with self-normalizing neural networks (SNNs) to enhance training robustness. Their framework achieved 79.8% accuracy in classifying four cancer types. These studies collectively highlight how supervised architectures can be customized to diverse omics modalities while balancing performance and biological interpretability.

In recent years, due to the excellent performance of convolutional neural networks (CNNs) on image classification tasks, more and more researchers have started to apply these networks to the classification problem of pan-cancer. For instance (Ameen et al., 2025) proposed a stacked deep learning ensemble model for multi-omics cancer type classification, demonstrating that deep learning can be effectively applied to high-dimensional biological data. Similarly (Lyu and Haque, 2018), firstly proposed the use of a convolutional neural network to classify mRNA expression data by embedding high-dimensional gene expression data into a two-dimensional image as the input of the convolutional neural network to classify 33 different types of tumors. Building on this, Mostavi et al. (Mostavi et al., 2020) systematically compared CNN architectures (e.g., Inception modules, residual connections), revealing that deeper networks achieved 95. 82% precision on 33-class tasks that highlight the impact of structural optimization. Addressing computational inefficiency Khalifa et al., 2020), applied binary particle swarm optimization (BPSO) to reduce the dimensionality of mRNA from 20,531 to 512 features before CNN training, achieving accuracy of 96. 9% on five types of tumors. Hybrid models also emerged as a promising frontier: (Huynh et al., 2019) combined deep CNNs (DCNN) with SVM classifiers, where DCNNs extracted high-order features and SVMs performed classification, reaching 76. 33% precision for 25 cancers. (Abdullahi et al., 2020) further demonstrated the efficiency of fine-tuning pre-trained AlexNet models on mRNA data, reaching 98.1% accuracy for five cancers with minimal computational overhead. Beyond expression data (Ye et al., 2021) encoded somatic mutation profiles into heatmap-like “mutation maps,” enabling ResNet-50 and Inception-v3 models to outperform traditional methods (89.7% vs. SVM’s 72.3%). Finally (AlShibli and Mathkour, 2019) validated CNNs’ versatility in CNV analysis, showing that a six-layer residual network (ResCNN6) surpassed standard CNNs and VGG-16 (86% accuracy for six cancers), underscoring the efficacy of residual connections in combating gradient vanishing. These innovations exemplify CNNs’ adaptability to multi-omics integration through data transformation, architectural refinement, and cross-domain transfer learning.

Unsupervised deep learning techniques have emerged as powerful tools for pan-cancer classification, particularly in scenarios with limited labeled data. Rong et al. (Rong et al., 2022) proposed a computational approach, multi-omics clustering variational autoencoders (Mcluster-VAEs), based on a new probabilistic model of a deep learning method consisting of clustering algorithm for multi-omics data to estimate posterior cancer subtypes. Building on this (Al Mamun et al., 2020) introduced the Concrete Autoencoder (CAE), an unsupervised framework for identifying discriminative lncRNAs. The CAE outperformed supervised methods (Lasso, RF, SVM-RFE) in classifying 33 tumors, achieving 93% accuracy. To address feature instability across CAE iterations (Al Mamun et al., 2021) later proposed the multi-run CAE (mrCAE), which aggregated high-frequency lncRNAs from 100 CAE runs to derive a stable subset of 69 markers. This refined set enabled accurate classification of 12 cancers, resolving reproducibility challenges inherent to stochastic deep learning models. Expanding to multi-omics integration (Zhang et al., 2019) developed OmiVAE, an end-to-end model combining VAEs with a classification network. OmiVAE first compressed the mRNA and DNA methylation data into low-dimensional embeddings, then predicted 33 tumor types using a three-layer neural network, achieving precision of 97. 49%. Finally (Albaradei et al., 2021) designed MetaCancer, which used convolutional VAE to extract features from mRNA, miRNA and methylation data. When fed into a deep neural network (DNN), this multi-omics integration classified 11 cancers with 88.85% accuracy-surpassing mRNA-only approaches by 14.2%. (Li et al., 2024) proposed AVBAE-MODFR, a two-phase framework that combines adversarial variational Bayes autoencoder for multi-omics embedding with a dual-net feature ranking module. Tested on TCGA pan-cancer data, AVBAE-MODFR outperformed four state-of-the-art methods, highlighting its robustness in representation learning and biomarker discovery. Compared with earlier VAE-based models such as OmiVAE and MetaCancer, AVBAE-MODFR not only integrates heterogeneous omics but also incorporates an explicit feature ranking mechanism, thereby enhancing interpretability and facilitating the identification of biologically meaningful markers. These innovations underscore unsupervised learning’s potential to uncover robust biomarkers and integrate heterogeneous omics data without reliance on labeled datasets.

Figure 3 illustrates the growing prominence of deep learning in pan-cancer research. It shows the percentage of all pan-cancer-related articles that used deep learning methods for classification over the past few years. A systematic review of papers published on the PubMed and Web of Science platforms using search terms “pan-cancer classification”, “deep learning” and “machine learning” from 2018-2024 revealed a steady increase in this ratio from 2018 to 2024. To summarize the current landscape of pan-cancer classification, we present an overview of relevant studies in recent years in Table 3. This table highlights the variety of machine learning and deep learning approaches, as well as the multi-omics data they employ.