AI technology involves the development of systems capable of executing tasks typically requiring human intelligence, such as reasoning, learning, and problem-solving. It is designed to replicate cognitive processes like perception, language processing, and decision-making, these systems draw from a diverse range of disciplines, including computer science, mathematics, psychology, and linguistics. AI technology has penetrated all aspects of human activities (12–15). In the cancer research field, AI is characterized by the use of machine learning and deep learning algorithms (16), which are important in processing and analyzing large-scale datasets (17–19).

Machine learning (ML), an integral part of artificial intelligence, encapsulates the autonomous identification of patterns and formulations within vast datasets (20). By discerning and extracting significant features from the data, ML can make accurate predictions and decisions .It fundamentally extracts patterns and rules from the data and apply them to new data. The workflow of ML comprises the acquisition, pre-processing, feature extraction of the data, model training, and evaluation optimization application of the obtained model. During model training, parameters are adjusted to minimize the discrepancy between predicted and actual outcomes, known as ‘error’ or ‘loss,’ which is quantified to direct the optimization process towards enhanced accuracy. Based on the model training approaches, there are four different types of ML: supervised (21), unsupervised (22, 23), semi-supervised (24), and reinforcement learning (25) ( Figure 1 ).

Deep learning (DL), a subfield of machine learning, employs artificial neural networks (26) to represent important information from massive amounts of data. DL comprises an input layer, multiple hidden layers, and an output layer, each of which receives the output of the previous layer as input and performs nonlinear transformations that progressively distill raw data into meaningful feature abstractions. There are several popular DL architectures: multilayer perceptron (MLP), convolutional neural networks (CNNs), recurrent neural networks (RNNs), auto-encoders (AEs), generative adversarial networks (GANs) and transformer ( Figure 2 ). These architectures can be used according to the specificity of the data.

Early cancer detection: By facilitating cancer detection at the precancer stage, AI allows for early interventions that significantly prolong the overall survival time of the patients. For example, Klein et al. used a blood-based multi-cancer early detection (MCED) test and applied cell-free DNA sequencing, combined with machine learning, which predicted the origin of cancer signals with high specificity and accuracy in a variety of cancers (27). Similarly, Stark et al. constructed machine learning models using Gail model inputs and personal health data. These models exhibit strong performance in predicting breast cancer risk and can be used as non-invasive tools to increase early detection and prevention of breast cancer (28). Additionally, to develop a machine learning model to predict the risk of lymph node metastasis in renal carcinoma, Feng et al. filtered clinical features through LASSO and univariate and multivariate logistic regression analyses and then used statistically significant risk factors to build the XGB model. It could distinguish about 89% of LNM patients when the threshold probability was set to 54.6%, suggesting a promising application prospect in the clinic (29).

Machine learning and deep learning emerge as potent tools to identify biomarkers from intricate datasets (30). For example, Halner et al. established a random forest-based machine learning pipeline, “Decancer,” to analyze liquid biopsies. Decanter enhanced the sensitivity for detecting stage I cancer from 48% to 90% regardless of cancer type. Promisingly, DEcancer’s performance using a 14-43 protein panel is comparable to 1,000 original proteins (31). To identify metabolomic biomarkers for the diagnosis and prognosis of gastric cancer, Chen et al. used the LASSO regression algorithm to build a 10-metabolite GC diagnostic model, which is validated in an external test set with a sensitivity of 0.905. This model exhibited superior performance to traditional models that utilized clinical parameters and identified two distinct biomarker panels, enabling early diagnosis and prognosis of cancer (32). Tayob et al. developed the parametric empirical Bayes algorithm and the Bayesian screening algorithm to improve the early detection of cancer, which improved sensitivity to cancer biomarkers (33). Furthermore, Konstantinos et al. tested the miRNA expression profiles of Gastrointestinal stromal tumors (GISTs) and applied machine learning to identify the miRNAs associated with the risk of GIST development. They found that several miRNAs, with hsa-miR218-5p as the best, may strongly affect the prognosis of GISTs and can serve as predictors for their development (34). In short, the application of AI and machine learning in oncology clinics has improved diagnostic time and clinical outcomes in various cancers.

Medical imaging: Imaging is at the forefront of clinical care. The integration of AI into image interpretation helps radiologists streamline workflow and improve patient care (35). Within imaging, convolutional neural networks (CNNs) and Deep Learning (DL) are exceptionally useful in computer vision and enable machines to see and interpret visual data (36, 37). Al-Masni et al. developed the ROI-based Convolutional Neural Network “You Only Look Once (YOLO)” to accurately detect and classify the masses in mammograms. It achieves an overall accuracy of 96.33% in detecting the mass location and 85.52% in distinguishing between benign and malignant lesions (38). Zhao et al. built the deep-learning-based, fully automated lymph node detection and segmentation (auto-LNDS) model based on multiparametric magnetic resonance imaging (mpMRI). The auto-LNDS achieved a sensitivity, PPV, and FP/vol of 80.0%, 73.5%, and 8.6 in internal testing and 62.6%, 64.5%, and 8.2 in external testing, respectively, significantly better than the performance of junior radiologists, therefore holding great potential for facilitating N-staging in clinical practice (39). Jin et al. developed a CNN-based algorithm to Improve the accuracy in Optical Diagnosis of Colorectal Polyps. It increased the accuracy of novice endoscopists to 85.6% and significantly reduced the skill-level dependence of endoscopists and costs (40).

Pathological identification: As the gold standard for confirming cancer, pathological identification holds paramount significance in diagnosis, prognosis, and therapeutic strategies. However, the heterogeneity of tumors poses a big challenge to precise diagnosis (41, 42). AI has transformed the landscape of cancer pathology by empowering it with enhanced diagnostic accuracy and streamlined decision-making frameworks, leveraging sophisticated histology image analysis (43). For example, to achieve an AI-based pathological prediction of the origins of unknown cancers. Lu et al. build the Tumor Origin Assessment via Deep Learning (TOAD), a deep-learning-based algorithm that provides a differential diagnosis for the origin of the primary tumor based on routinely acquired histology slides (44). In addition, Lee et al. presented a graph deep neural-network model to analyze the whole-slide images (45). This model considers histopathological features from the tumor microenvironment. in gigapixel-sized WSIs in a semi-supervised manner and was trained to provide interpretable prognostic biomarkers in patients with kidney, breast, lung, and uterine cancers.

Treatment: The outcomes of cancer treatment are affected by several key factors, such as the patient`s health status, cancer subtype, and stage. Additionally, molecular cancer research has recently revealed the contribution of genetic mutations to patients` responses to a specific treatment. The complex interplay of the above factors in the real world poses a significant challenge for oncologists in selecting the appropriate treatment regimen for a specific patient.

In this scenario, AI is inherently a powerful approach to the integration and aggregation of intricate and multi-dimensional datasets and providing comprehensive data support for decision-making (46, 47). For example, Luo et al. proposed a collaborative filtering method with machine learning. It can identify the most suitable compounds for patients without genetic data, making it feasible to predict drug sensitivity and achieve personalized drug selection in a cost-effective way (48). Abajian et al. used Supervised Machine Learning with both Logistic Regression (LR) and Random Forest (RF) algorithms to explore the treatment response to transarterial chemoembolization for hepatocellular carcinoma. Both LR and RF models achieved an overall accuracy of 78% and identified cirrhosis status and relative tumor signal intensity (>27.0) as the two strongest predictors of treatment response (49). Kong et al. introduced a NetBio-based machine learning, which accurately predicted the treatment responses to Immune checkpoint inhibitors (ICIs) in three different cancer types-melanoma, gastric cancer, and bladder cancer. This model demonstrated superior performance in comparison with conventional ICI treatment biomarkers, such as the expression profiles of ICI targets (50).

Prognosis and management: Understanding the progression and survival time of patients is essential for cancer management. Oncologists used to predict patients’ prognoses based on their experience of understanding patients’ clinical profiles (age, health status) and tumor characteristics (subtype, stage, and grade). Nevertheless, this strategy is inherently limited and has a low predictive capability due to individual variation. AI has exhibited great promise to deal with these constraints and achieves accurate prognosis prediction for individual patients (51). Qiu et al. developed an XGBoost model to help physicians make clinical decisions. It employed clinicopathological information and predicted the risk of distant metastasis in patients with rectal cancer (52). Based on LASSO regression and Pearson correlation coefficients, Cai et al. identified metastasis-associated genes from different cancer tissues and then used them to build a CNN-based model, “Multi-Dimensional Convolutional Neural Network (MDCNN).” It achieved satisfactory prediction accuracy in bone metastasis, lung metastasis, and liver metastasis (53). The combination of AI and the Internet of Things (IoT) technology enables telemedicine and intelligent monitoring functions, allowing patients to receive scenario-based remote management (54).

Drug Discovery: Drug discovery and development of anti-cancer drugs is the goal of translational medicine. However, this work is quite a costly and time-consuming operation. Additionally, although the advances in muti-omics and clinical trials provide quite meaningful information, their complexity also imposes a huge obstacle. AI and computer-aided drug design, along with modern experimental technical knowledge, has energized data mining for faster drug design and development in the pharmaceutical industry (55). For example, AlphaFold2, a deep neural network algorithm, demonstrates high accuracy in predicting the three-dimensional structures of proteins, particularly when sequences of multiple homologs are available. It helps us understand protein function changes underlying carcinogenesis and improve our approaches to counter them (56). Meanwhile, low-cost cancer drug repurposing can be achieved by deep learning approaches, which aid the modeling of existing drugs for discovering novel drug targets. For example, Zhou et al. designed a prediction approach called an ensemble of multiple drug repositioning approaches (EMUDRA). Using EMUDRA, they predicted and experimentally validated the antibiotic rifabutin as an anti-cancer drug for triple-negative breast cancer (57).

scRNA-seq analysis: scRNA-seq allows researchers to conduct in-depth analysis of molecular characteristics, such as gene expression and epigenetic modifications within individual cells, generating vast amounts of genetic information data (104). The analysis of these data is crucial for revealing cellular heterogeneity and functional characteristics (105). With the intervention of AI, rapid processing and interpretation of massive scRNA-seq data can be achieved with enhanced data accuracy. AI algorithms can automatically identify and filter out noise, retaining the true biological differences between cells and thereby improving data reliability (106). Additionally, batch effects, a common issue in scRNA-seq analysis, can be caused by various factors such as experimental samples, platforms, and library construction methods. AI technologies can effectively eliminate these batch effects while preserving biological differences by projecting high-dimensional data into a low-dimensional cellular embedding space through an asymmetric autoencoder structure (107). In fact, these methods not only improve the accuracy of data integration but also enable online data integration and comparative analysis of new data with existing data.

By use of different machine learning approaches, cell type identification models are developed to recognize cell types and subtypes (108). These models can extract key biological insights to predict the changes in gene expression levels or even dynamic changes in gene interaction networks. For example, scRobust, a self-supervised learning strategy built on the transformer architecture, has demonstrated effectiveness in cell-type annotation and drug tolerance detection (109). A deep learning model “Enformer Celltyping” predicts epigenetic signals across cell types. It overcomes the limitations of existing machine learning approaches, which are confined to the cell types they were trained on (110). Here, we summarized cell-type identification models in Table 2 .

Multi-omics analysis: Complex and dynamic networks of molecules in LC-IME make a single layer of “omics” unable to provide deep insights into the underlying mechanisms. Recent technological advancement in high-throughput measurement of genome (111, 112), epigenome (113, 114), metabolome (115), transcriptome (116), and proteome (117) allows comprehensive multi-omic studies. Multi-omics approaches are pivotal in identifying new therapeutic targets (118) and predicting patients’ responses to treatments (119). The data from different omics can be cross-fused and mutually verified, providing a more reliable, comprehensive, and systematic perspective (120, 121). Through the integrated analysis of omics data, in-depth biological data that cannot be obtained by a single omics technology can be uncovered (122, 123). However, the advantage of multi-omics data integration comes with the extra complexity deriving from inherently diverse types of omics datasets, which may pose a challenge to integrateing the omics data in a biologically meaningful manner (124). The experimental data generated across diverse laboratories often cannot be seamlessly amalgamated due to inherent constraints. Additionally, the inherent heterogeneity of multi-omic datasets, stemming from technical, biological, chemical, and physical sources, poses significant challenges for interpretation (125).

With the continuous development of AI technologies, the integration of AI and multi-omics has emerged as one powerful solution to these challenges (126–128). AI has remarkable capabilities in deciphering complex patterns and extracting meaningful insights from large and intricate datasets (129–131). This enables researchers to more systematically analyze the complexity of biological systems (132), reveal the interactions and regulatory mechanisms between different molecular layers, and more accurately identify disease-related molecular markers and potential drug targets (133, 134). This subsequently contributes to the development of personalized medicine and precise treatment plans, improving therapeutic effects and reducing side effects (16, 135, 136).

Data-based Integration. This methodology has proven effective in several studies. Zhang Team merged information from single-nucleotide polymorphisms (SNPs) and transcriptomic profiles into a single matrix, which uses a Bayesian integrative model to facilitate the investigation of their interplay and enable the prediction of quantitative phenotypes (137). To predict remission rates and survival outcomes in ovarian cancer, Mankoo and colleagues integrated the data of copy number alteration, DNA methylation, microRNA, and gene expression and performed a multivariate Cox-LASSO analysis (138). Shen et al. proposed the iCluster framework for glioblastoma subtyping. This framework harmoniously and integrated, with a common set of latent variables, three distinct omics data of copy number alteration, gene expression, and DNA methylation (139).

Model-based Integration. In a model-centric integration framework, distinct models tailored to individual data perspectives are initially formulated, subsequently converging through a fusion process of their respective outputs. For example, the ATHENA tool (140–142), which is designed for investigating heritable and environmental network associations, integrates different omics data of copy number alterations, DNA methylation, miRNA, and gene expression to uncover correlations with clinical endpoints. This integration involves constructing foundational models and neural networks per omics type, ultimately leading to the construction of an integrated model (137). Wang’s team used Similarity Network Fusion (SNF) for cancer subtyping. It begins by creating patient similarity matrices based on DNA methylation and the expression of mRNA expression or miRNA and moves to an iterative nonlinear integration, where the three foundational similarity matrices converge into a unified matrix (143). To predict drug resistance in HIV protease mutants, Dr. Ghici and Potter devised an ensemble-based strategy. It sets up the basic predictive models with structural characteristics of the HIV protease-drug inhibitor complex and DNA sequence variations, respectively, and then orchestrates a majority voting system to enhance the accuracy of drug resistance prediction (144).