Challenges in AI-driven multi-omics data analysis for Oncology: Addressing dimensionality, sparsity, transparency and ethical considerations

Artificial intelligence, particularly deep learning, is becoming increasingly prominent in multi-omics research, especially since traditional statistical models struggle to handle the complexity and high dimensionality of such data. By effectively combining different types of omics data, AI techniques can unveil hidden connections, detect biomarkers, and improve disease prediction through the integration of multi-omics layers and modalities, which can lead to significant advancements in precision medicine. In this review, we gathered published methods of deep learning-based multi-omics integration specialized in oncology since 2020. We concentrated exclusively on studies utilizing cancer omics data mainly sourced from The Cancer Genome Atlas (TCGA) database. As a result, we identified 32 articles that generally fulfilled the criteria. We studied their techniques and their ability to handle challenges in analyzing multi-omics data, particularly regarding missing data, dimensionality, and processing workflows. We also discuss how well these methods consider explainability, interpretability, and ethical aspects in developing solutions that treat private medical and sensitive information. From the 32 studies, we can divide deep learning-based multi-omics integration methods into two types: non-generative and generative models. Non-generative approaches, such as feedforward neural networks (FFNs), graph convolutional networks (GCNs), and autoencoders, are designed to extract features and perform classification directly. On the other hand, generative methods such as variational autoencoders (VAEs), generative adversarial networks (GANs), and generative pretrained transformers (GPTs) focus on creating adaptable representations that can be shared across multiple modalities. These methods have advanced the handling of missing data and dimensionality, outperforming traditional approaches. However, most reviewed models remain at the proof-of-concept stage, with limited clinical validation or real-world deployment. • Incorporate interaction networks (miRNA–mRNA …) for more biologically meaningful integration over simple concatenation. • Use generative models (GANs, VAEs …) to synthesize minority-class samples and address imbalance. • Prioritize fairness evaluation with subgroup-specific metrics and adversarial debiasing to ensure equitable clinical performance. • Adopt hybrid, privacy-preserving frameworks (federated transformers …) to improve robustness and enable real-world deployment across institutions.