Decoding cancer with artificial intelligence: Transforming research, diagnosis, and therapy with future insights Authors: Muneera Anwer, Krupa Bhaliya, Ming Q Wei Published: 2026 ABSTRACT Highlights•Artificial intelligence (AI) is transforming the cancer research from drug discovery and biomarker identification to diagnosis and therapy.•Deep learning frameworks accelerate oncology drug discovery and repurposing, reducing timelines and costs through predictive modelling and molecular simulation.•Multi-omics integration powered by AI enables the discovery of robust biomarkers for diagnosis, prognosis, and therapy response prediction.•Federated learning, digital twins, and explainable AI guide future clinical use.•Emerging AI paradigms offer future-ready frameworks for personalized, privacy-preserving oncology care.Keywords: Artificial intelligence, Precision oncology, Biomarker discovery, Cancer diagnosis, Cancer research, Digital twins Introduction Cancer remains one of the leading causes of morbidity and mortality worldwide, accounting for nearly 10 million deaths annually (1 in 6 deaths) and imposing an immense burden on patients, healthcare systems, and economies [1]. Additionally, there is no equal distribution of the burden; it is estimated that 70% of the cancer-related deaths occur in low- and middle-income countries, where early detection and effective treatment options are not available [2]. Despite remarkable advances in early detection, targeted therapies, and immunotherapy, cancer continues to present significant clinical challenges due to its heterogeneity, adaptability, and complex biological nature [3]. The era of high-throughput technologies has generated an exceptional volume of cancer-related data, spanning genomics, transcriptomics, proteomics, imaging, and clinical records [4]. While these datasets provide invaluable insights into tumour biology and therapeutic response, their sheer scale, diversity, and complexity surpass the capacity of traditional analytical approaches [5]. Extracting clinically actionable knowledge from such multidimensional data requires innovative computational strategies [6]. The concept of artificial intelligence (AI) was first introduced in 1956 at the Dartmouth Summer Workshop, where it was broadly described as the idea of creating “thinking machines” [7]. AI refers to the capability of machines to learn from representative examples, identify patterns and relationships, and apply this knowledge to make informed decisions on new, unseen data. The term AI is often used broadly and is closely associated with, or sometimes used interchangeably with, machine learning (ML) and deep learning (DL) [8]. AI is expected to transform all aspects of society and shape the real-world impact, including disease diagnosis, clinical support, drug and vaccine development, and disease prevention, management, control, and patient outcome [9,10]. AI, particularly machine learning and deep learning, has emerged as a transformative tool in addressing the challenges in cancer research. By enabling pattern recognition, predictive modelling, and integration of multimodal data, AI offers unparalleled opportunities to improve cancer research, diagnosis, and therapy. The integration of AI in medical imaging, genomic analysis, and pathology enhances diagnostic precision and introduces a novel, less invasive approach to cancer screening. This not only boosts the effectiveness of medical facilities but also reduces operational costs. AI systems have already demonstrated the potential to enhance imaging interpretation, uncover novel biomarkers, personalize treatment decisions, and accelerate drug discovery [8]. While several recent reviews have examined the application of AI in isolated oncology domains such as radiology, pathology, genomics, or drug discovery, most remain modality-specific and do not comprehensively address the convergence of multimodal AI ecosystems in translational oncology. In contrast, this review integrates advances across imaging, computational pathology, genomics, liquid biopsy, drug development, and precision therapeutics within a unified translational framework. Importantly, we further highlight emerging paradigms, including multimodal foundation models, federated learning, explainable AI, privacy-preserving architectures, and patient-specific digital twins, which are increasingly shaping next-generation oncology systems. By emphasizing the integration of these technologies rather than evaluating them in isolation, this review provides a broader translational perspective on how AI may evolve from domain-specific analytical tools toward clinically deployable oncology platforms. Accordingly, this review adopts a narrative approach to synthesise current evidence and emerging trends across multiple domains of artificial intelligence in oncology. AI in cancer research The integration of artificial intelligence into cancer research has opened new avenues for understanding tumour biology and developing more effective interventions. AI techniques have shown strong potential in revolutionizing the field of cancer [8,11]. By harnessing the power of machine learning (ML) and deep learning (DL) in healthcare, researchers are now able to interrogate complex datasets and generate insights that were previously unattainable with conventional approaches [11]. Fig. 1 highlights the application of AI across different domains in the field of oncology. From medical images and genomic profiles to clinical records, these technologies have demonstrated an unparalleled capacity to process and interpret data. The following sections will demonstrate the integration of AI with oncology in various aspects of cancer screening, prevention, and patient care. To clarify the conceptual framework that underlies these applications, Fig. 2 presents a schematic of the general AI workflow in cancer care, illustrating how multi-scale data (e.g., genomic, proteomic, and radiological inputs) are collected and pre-processed, how features are extracted and the data partitioned into training and validation cohorts, how models are constructed and validated, and how the resulting outputs are applied across cancer research, diagnosis, and therapy. Application of artificial intelligence in different domains of oncology. Schematic of the general workflow of artificial intelligence in cancer care. Multi-scale data, including genomic, transcriptomic, proteomic, radiological (MRI/CT/histopathology), and clinical records, are collected and integrated, then pre-processed through quality control, normalization, harmonization, and feature extraction and selection. The processed data are partitioned into training and validation (internal and external) cohorts for model construction using machine-learning and deep-learning architectures (e.g., convolutional neural networks, transformers, and graph neural networks), followed by performance evaluation, interpretability, and bias assessment. The validated model outputs are then applied across cancer research (drug discovery and repurposing, genomic analysis, biomarker discovery), diagnosis (radiology, computational pathology, liquid biopsy), and therapy (precision oncology, radiotherapy planning, treatment-response prediction), with iterative refinement informed by real-world feedback. Drug discovery and repurposing Drug discovery and development remain a complex process and require a significant financial investment. The pipeline from the identification of a lead compound to obtaining market authorization takes approximately 10 to 15 years [12]. Traditional drug development methods often rely on manual review and selection, which are time-consuming and resource-intensive, with high attrition rates [13]. AI-driven models, particularly deep learning architectures, are accelerating this process by predicting drug-target interactions, optimizing lead compounds, and identifying potential candidates for drug repurposing [14]. The multifaceted role of AI across the field of cancer drug development and repurposing is demonstrated in Fig. 3. Starting with drug design, AI facilitates rational discovery of novel compounds and optimization of leads through structural modelling and docking simulations. In the screening phase, machine learning models predict pharmacokinetics, absorption, distribution, metabolism, excretion, and toxicity (ADME/T) properties, helping to eliminate unsuitable compounds early. AI enhances the drug synthesis process through retrosynthesis and reaction prediction algorithms that enable automated, efficient synthesis of novel molecules. Finally, in drug repurposing, AI analyses complex drug-target and disease networks to reveal new applications for approved compounds, offering a cost-effective and faster route to clinical translation. This integrative approach demonstrates how AI can shorten timelines, improve precision, and reduce attrition in cancer drug development [15]. Role of artificial intelligence in cancer drug discovery and repurposing. Process includes AI-driven drug design, screening of the drug by predicting physicochemical and ADME/T properties to optimize candidate selection. Drug synthesis by retrosynthesis prediction, reaction prediction, and automated chemical synthesis of drug candidates. Finally, drug repurposing leverages AI to uncover new therapeutic uses of existing drugs by analysing interactions. Together, these applications accelerate cancer drug discovery while reducing costs and development timelines. In drug discovery, AI-driven deep learning models have been applied to analyse vast chemical libraries and predict anticancer activity with high precision. For example, convolutional neural networks and generative models can screen large chemical libraries to detect novel anticancer molecules, while natural language processing enables mining of biomedical literature to uncover new therapeutic uses for existing drugs. Such approaches not only shorten discovery timelines but also reduce costs and increase the likelihood of success [22]. AtomNet, a structure-based deep learning platform, has been used to screen millions of compounds for potential oncology targets, identifying candidates that would have been overlooked by conventional methods [16]. Similarly, reinforcement of learning and generative adversarial networks (GANs) have been employed to design novel molecules with optimized pharmacological profiles, such as kinase inhibitors with improved specificity [17,18]. AI also shows promising applications in drug repurposing, where algorithms integrate genomic, transcriptomic, and clinical data to reveal hidden therapeutic potential in approved drugs [19]. For instance, drug repurposing has been explored across different disease indications as a strategy to identify new therapeutic applications for existing drugs [20]. Baricitinib, an anti-rheumatoid arthritis drug, has been investigated as a potential treatment for certain cancers due to its inhibition of JAK–STAT signalling, a pathway implicated in tumour progression [21]. Studies reported the repurposing of the antipsychotic drugs for cancer treatment, for example, thioridazine (an antipsychotic drug) is used for inducing immunogenic cell death in colorectal cancer [22,23]. The same drug is investigated for targeting cancer stem-like cells in lung cancer [24]. AI-driven drug discovery platforms, such as DeepCure and BenevolentAI, have demonstrated applications in drug design and target identification [25,26]. For example, BenevolentAI’s knowledge graph approach has been applied to target identification and drug repurposing [27,28]. Moreover, AI-guided drug combination studies and clinical trials have investigated metformin (an antidiabetic agent) as a sensitizer to chemotherapy in breast and colorectal cancers. For instance, a randomized trial of metformin plus neoadjuvant chemotherapy in non-diabetic breast cancer patients showed improved pathological complete response rates compared to chemotherapy alone [29]. Additionally, epidemiological and preclinical data in colorectal cancer report reduced incidence and improved outcomes among patients treated with metformin, though with variability in response across patient subgroups [30,31]. By bridging computational prediction with laboratory validation, AI enables both the discovery of novel agents and the rapid repositioning of known drugs, reducing costs and development timelines. As these technologies mature, AI-driven discovery and repurposing are poised to play a central role in delivering personalized and more effective cancer therapies [32]. To explicitly link specific AI methodologies to their corresponding drugs or compounds and cancer indications, a structured summary of representative AI applications in cancer drug discovery and repurposing is provided in Table 3, organised by AI function (drug–target interaction prediction, virtual screening, ADMET prediction, and de novo molecule generation). Table 3Representative AI applications in cancer drug discovery and repurposing, organised by AI function.Table 3 dummy alt textAI functionAI method/ platformDrug or compoundCancer indicationReferenceDrug-target interaction prediction and virtual screeningStructure-based deep learning (AtomNet)Novel oncology lead compoundsMultiple solid tumours[16]De novo molecule generationGenerative adversarial networks; reinforcement learningKinase inhibitors with improved specificityMultiple kinase-driven cancers[17,18]Drug repurposingJAK-STAT inhibitionJAK inhibitorsMultiple cancers[21]Drug repurposingRepurposing evidence from preclinical studiesThioridazineColorectal cancer, leukemia, lung cancer[[22], [23], [24]]Knowledge-graph-based repurposingKnowledge-graph mapping of molecular networks (BenevolentAI)Established oncology drugs (repositioned)Multiple tumour types[27,28]AI-guided drug combinationAI-guided combination analysis with clinical-trial validationMetformin plus neoadjuvant chemotherapyBreast and colorectal cancer[[29], [30], [31]] Representative AI applications in cancer drug discovery and repurposing, organised by AI function. Genomic data analysis Cancer is fundamentally a disease of the genome, and the ability to decode large-scale sequencing datasets is crucial for precision oncology. AI has shown significant improvement in genomic analysis by providing efficient and accurate genetic mutation calling, identifying variant classification, and understanding disease pathway analysis, which enables precision medicine for personalized treatments and diagnosis. AI models use DL and ML algorithms, are trained on large-scale datasets, and can handle complex genomic regions, including highly variable or repetitive areas, on which conventional methods struggle [33]. AI tools like DeepVariant and DeepFilter, are DL based variant caller that use deep convolutional neural networks have been shown to distinguish true variants from sequencing artifacts, achieving higher accuracy in identifying single-nucleotide variants (SNVs) and insertions/deletions (indels), especially in challenging genomic regions [34]. DeepTrio, an advanced DL based variant caller developed by Google Health, is an extension of DeepVariant, designed specifically for analysing genomic data from family trios, typically consisting of a child and their two parents [35]. Clairvoyante [36], Clair [37] and Clair3 [38] are DL-based variant callers that specialize in genomic variant detection from both short-read and long-read sequencing data and utilize CNN for improved accuracy in calling SNPs and InDels. Other tools include Medaka, precisely designed for analysing long-read sequencing data, and Hybrid and stand-alone Estimation of smaLL genOmic variants (HELLO), an open-source variant caller designed to address the challenges of detecting genomic variants of both SNPs and InDels [39]. Machine learning (ML) and ensemble approaches similarly enhance mutation detection. For example, NeoMutate, an ensemble-ML framework trained on thousands of cancer variants, has improved sensitivity and specificity for somatic mutation calling compared to individual variant callers [40]. Beyond coding regions, deep learning models are increasingly being employed to interpret the non-coding genome. Illuminating the Noncoding Genome in Cancer Using Artificial Intelligence (2025) reviews models that identify functional noncoding variants, predict their effects on gene expression and enhancer/promoter activity, and map risk-associated regulatory mutations [41]. AI-based tools are also advancing variant effect prediction in protein-coding regions. AlphaMissense, which leverages AlphaFold2 models, has demonstrated high sensitivity and specificity in predicting the effects of missense variants, although reduced sensitivity has been reported for variants in intrinsically disordered protein regions [42]. Together, these applications highlight how AI is not only enhancing the accuracy of genomic interpretation but also translating into clinically meaningful insights, from mutation discovery to precision treatment planning and supporting personalized treatment strategies [43]. Biomarker discovery through multi-omics integration The advent of high-throughput omics platforms has generated datasets across genomics, transcriptomics, proteomics, metabolomics, and epigenomics, which AI can integrate to discover robust biomarkers for cancer detection, prognosis, and treatment response [44]. Recent studies demonstrate these applications: for example, a multi-omics model combining whole-exome sequencing and RNA-seq data in melanoma outperformed nonsynonymous tumour mutational burden in predicting anti-PD-1 immunotherapy response, achieving an AUC of ∼0.70 in training and ∼0.64 in validation [45]. A similar melanoma study integrated genomic, transcriptomic, epigenomic, and proteomic data to compare responders versus non-responders, confirming known biomarkers (e.g., TMB, PD-L1, ITH) and identifying new features, including enrichment of JAK-STAT, RAS, PI3K-Akt pathways, and specific microRNAs and proteins [46]. In breast cancer, integrative approaches using transcriptomics, epigenomics, and microbiome, and comparing statistical methods with deep learning integration methods, succeeded in defining distinct molecular subtypes with significant prognostic value [47]. Another study combines multi-omics to uncover the clustering of expression, methylation, and mutation data to identify novel prognostic subtypes and markers in immunotherapy for melanoma [48]. For early detection of breast cancer and therapeutic response prediction, a panel of four metabolites, including inosine and uridine, showed strong discrimination in breast cancer using metabolomics coupled with machine learning [49]. Collectively, these emerging studies underscore that AI-enabled multi-omics integration not only enhances predictive performance over single omics layers but also enables the discovery of new, clinically actionable biomarkers. AI in cancer diagnosis Accurate and timely diagnosis is the cornerstone of effective cancer management. However, diagnostic modalities such as imaging, pathology, and molecular assays generate massive and complex datasets that can be challenging to interpret consistently. Artificial intelligence has emerged as a powerful tool to augment diagnostic accuracy, improve reproducibility, and uncover subtle features that may be imperceptible to human experts. AI is quickly changing cancer diagnosis across various methods, from large-scale imaging to detailed pathology and molecular liquid biopsies, as shown in Table 1. This section summarizes recent advances, practical strengths, and challenges of AI in (1) radiology (image segmentation, detection, classification), (2) computational pathology (analysing histopathology images for grading and classification), and (3) liquid biopsy analysis (using machine learning on circulating tumour DNA and extracellular vesicles). Table 1Applications of AI in cancer care across different modalities.Table 1 dummy alt textDomainAI applicationsExample / outcomeReferenceDrug discoveryPredict drug–target interactions; virtual screening; ADMET prediction; de-novo molecule generationFaster identification of candidate inhibitors and improved hit rates in lead optimisation pipelines; reduced cost/time for early discovery workflows.[89]Genomics / mutation analysisMutation calling, driver-mutation prioritisation, somatic variant classification; ML for predicting functional impactImproved prioritisation of candidate driver mutations from tumour genomes; integration with clinical annotation to nominate targets.[90]Biomarker discovery / multi-omicsMulti-omics integration (genomics, transcriptomics, epigenomics, proteomics); radiogenomics/ pathomics fusion; predictive model building for response/survivalIdentification of composite biomarkers predicting ICI response and survival; meta-biomarkers from combined radiomic + molecular features.[91]Radiology (image analysis)Image segmentation, lesion detection, classification, risk scoring, temporal comparisonAutomated mammogram/DBT assistance that improves detection rates and reduces reader workload; improved tumour delineation consistency used for radiotherapy planning.[52]Pathology (computational histopathology)Whole-slide image analysis, weakly supervised learning / MIL for detection and grading; prediction of molecular phenotypes from H&EClinical-grade detection/triage systems; ML models predicting microsatellite instability and tumor subtypes directly from H&E.[55]Liquid biopsy (ctDNA / fragmentomics / exosomes)Fragmentome analysis (e.g., DELFI), methylation signatures, mutation calling, multi-analyte classifiersEarly detection / tissue-of-origin signals from cfDNA fragmentation & methylation; improved sensitivity when signals are combined.[60] Applications of AI in cancer care across different modalities. Radiology Medical imaging is central to cancer screening, diagnosis, staging, and treatment monitoring. AI, particularly deep learning methods, has shown strong technical performance in image segmentation, lesion detection, and disease classification [50]. Deep convolutional neural networks (CNNs) have become the main method for radiology tasks like segmentation, lesion detection, and image classification. Reviews and large surveys indicate that since around 2015, deep learning has surpassed traditional feature-engineering methods in effectiveness for tasks such as tumour delineation on CT and MRI and lesion detection on mammography and CT. This shift has been driven by access to larger annotated datasets and improved architectures [51]. CNN-based segmentation networks, like U-Net families and multi-scale variants, provide consistent delineations used in radiotherapy planning and quantitative imaging biomarkers. Meanwhile, object-detection networks (Faster R-CNN, RetinaNet, YOLO variants) enable automated lesion detection and triage workflows. Multi-reader studies have validated these methods. For example, stand-alone AI systems have been directly compared with hundreds of radiologists in mammography screening, showing similar sensitivity. In some cases, AI improved specificity when used as a second reader or for pre-screening triage [52]. Recent reviews highlight the importance of multi-scale context and transfer learning in radiology image analysis for breast tumour detection and more [53]. Nonetheless, robust generalisation across imaging settings remains a key challenge. Models can perform poorly when faced with different scanners, patient groups, or acquisition methods. This requires domain adaptation, training at multiple sites, and clinical evaluation before deploying. Pathology Pathology remains the diagnostic gold standard in oncology, providing essential insights into tumour type, grade, and molecular features. The digital transformation of histopathology through whole-slide imaging has generated vast visual datasets that require advanced computational tools for efficient analysis. Artificial intelligence, particularly deep learning, now plays a pivotal role in automating tissue classification, grading, and molecular phenotype prediction, thereby reducing inter-observer variability and increasing diagnostic speed and accuracy [54]. The digitization of whole-slide images (WSI) alongside deep learning has made automated tumour detection, grading, and even molecular phenotype inference possible from routine H&E histology. Weakly supervised learning and multiple instance learning (MIL) approaches have significantly impacted model training. These methods use slide-level labels, like diagnoses and outcomes, without needing detailed pixel-level annotation. Large studies have shown that MIL-based systems can achieve clinical-grade sensitivity for specific tasks and help reduce a pathologist's workload by sorting negative slides while maintaining nearly 100% sensitivity for malignant cases [55]. Beyond prostate pathology, AI systems have achieved high diagnostic accuracy across multiple cancers. For instance, convolutional neural networks (CNNs) trained on colorectal and gastric histopathology slides can predict microsatellite instability (MSI) and tumour mutational burden directly from hematoxylin-eosin images with area-under-curve (AUC) values between 0.90 and 0.96, showing promising concordance with molecular testing in retrospective studies [56]. In breast cancer, weakly supervised learning models successfully classified histologic grades and hormone-receptor status from digitized biopsies, demonstrating clinical-grade performance while reducing manual review time [57,58]. Likewise, lung cancer classifiers based on multiple-instance learning reached ROC values (Receiver Operating Characteristic, a measure of diagnostic accuracy) around 0.94 for distinguishing adenocarcinoma from squamous cell carcinoma [54]. A survey of AI in computational histopathology was provided by a study, covering methods from supervised to transfer learning approaches [59]. Nonetheless, computational pathology faces unique challenges. WSI files are massive and diverse, requiring scalable patch-level training methods and careful calibration of magnification levels. Label noise and variability in grading among observers complicate supervised learning, while biases in training data can lead to unforeseen confounding. Additional challenges include scanner variability, staining heterogeneity, and batch effects, all of which may reduce model generalisability across institutions. Tackling these problems calls for multi-centre datasets, stain normalization methods, explainability analyses, and clinical trials to confirm diagnostic benefits. Liquid biopsy The development of liquid biopsy represents a paradigm shift in oncology diagnostics, offering a minimally invasive means to capture the molecular landscape of tumours from body fluids. However, the data generated from circulating tumour DNA, exosomes, and other biomarkers are highly complex and noisy. Machine-learning algorithms can be used for identifying clinically relevant patterns within these data, enabling early detection and treatment monitoring, [53,60]. One key method examined genome-wide cfDNA fragmentation patterns and employed gradient-boosting machines to identify cancer patients versus healthy individuals and suggest tissue-of-origin candidates. These fragmentome-based classifiers, like DELFI, achieved high area-under-curve values across various tumour types and show promise for early detection and monitoring [60]. Other approaches combine mutation detection, methylation signatures, and fragmentomics into multi-modal classifiers that boost sensitivity, especially for early-stage diseases where mutant allele fractions are low [61]. Recent efforts are focused on multi-cancer early detection (MCED) using blood-based liquid biopsy panels, leveraging machine learning to detect cancer signals across multiple tissue types. Foser et al. discuss how imaging and liquid biopsy can work together using AI to improve sensitivity while reducing false positives. Practical aspects of machine learning in liquid biopsy involve extreme class imbalance and low signal-to-noise ratios in early stages [53]. Importantly, the population-level clinical utility and cost-effectiveness of MCED assays remain under active investigation. This requires careful selection and regulation of features, as well as harmonization across different sequencing depths or library protocols. The clinical utility question remains: a technically accurate classifier must be tested on prospective cohorts to determine if earlier detection improves outcomes and manages acceptable harms. Integrated insights and future diagnostic outlook In radiology, pathology, and liquid biopsy, several common themes arise: (1) modern AI methods have demonstrated performance comparable to expert clinicians in selected, well-defined diagnostic tasks when trained and evaluated on large representative datasets; (2) weak supervision, multi-scale modelling, and multi-modal fusion are effective strategies for using routinely generated clinical data; and (3) translational success relies on thorough external validation, prospective evaluation, clear reporting (for example, TRIPOD, STARD-AI), and attention to bias and regulatory standards. Future advancements may be improved by federated and privacy-preserving training models that allow multi-institutional development without sharing raw data and by integrating AI into clinical workflows as supportive tools that enhance, not replace, expert judgment. AI in cancer therapy Artificial intelligence is increasingly revolutionizing cancer therapy by optimizing treatment planning, individualizing therapeutic strategies, and predicting patient outcomes. AI models analyse multimodal clinical, imaging, and genomic data to tailor interventions for each patient, improve dose delivery in radiotherapy, and forecast treatment response and toxicity. By integrating predictive modelling and real-time decision support, AI enhances the precision, safety, and efficacy of modern oncology care. Precision oncology: AI for therapy selection based on patient profiles Artificial Intelligence is increasingly important in precision oncology. It helps customize therapy selection based on individual patient profiles, which include clinical, imaging, genomic, and other multi-omics data. In recent years, researchers have developed AI models to predict how patients will respond to immunotherapy, identify patients for targeted treatments, and find biomarkers that could guide treatment decisions. AI can analyse radiological, genomic, and pathological data to anticipate the effectiveness of immune checkpoint inhibitors (ICIs) and the related adverse events (irAEs) [62]. Researchers have also created AI-driven scores to predict immunotherapy responses. These scores integrate various data types, including genomics, imaging, and pathology, to pinpoint patients who are likely to benefit. In hepatocellular carcinoma (HCC), a study using CT radiomics from both liver tissue and viable tumours found that combining features from both improved predictions of immunotherapy responses. This approach achieved an accuracy of about 86%, with a sensitivity of around 70% and a specificity of about 94% [63]. Similarly, in lung cancer, researchers combined advanced radiomics with deep learning to predict progression-free survival (PFS) in patients receiving immunotherapy, utilizing imaging and clinical features [64]. Multimodal longitudinal models also show potential. Yeghaian et al. created a transformer-based multimodal model that integrated blood biomarkers from before and during treatment, CT-derived organ volumes, and medication data in a pan-cancer cohort (n = 694). This model predicted survival at 3, 6, 9, and 12 months, achieving area under the curve (AUC) scores between 0.81 and 0.84 [65]. These models demonstrate that combining different patient data over time can strengthen therapy selection. However, precision oncology faces several challenges; hence, there is a need for large, diverse datasets and a variety of data types. Ensuring models are understandable is also important. Many studies rely on retrospective data, which can lead to overfitting and limited generalizability. Additionally, many biomarkers have not yet been validated in prospective trials. It should be emphasised that the performance metrics cited above (e.g., accuracy and AUC) are difficult to compare directly across studies, as they are frequently reported with limited methodological context, including heterogeneous cohort sizes and composition, differing validation strategies (internal, external, or cross-validation), and a predominance of single-centre, retrospective designs. To aid interpretation, representative precision-oncology studies are summarised in Table 4, organised by prediction task (response prediction, survival prediction, and adverse-event prediction) together with cohort characteristics, validation strategy, and reported performance. At present, most of the AI approaches described in this section remain at the proof-of-concept or retrospective-validation stage; relatively few have undergone prospective, multi-centre evaluation, and this distinction is essential when assessing their translational readiness and clinical applicability. Table 4Representative studies of AI in cancer therapy and precision oncology, organised by prediction task, with cohort characteristics, validation strategy, and reported performance.Table 4 dummy alt textPrediction taskCancer typeData modalityAI methodCohort/ validationReported performanceResponse predictionHepatocellular carcinomaPretreatment CT radiomics (liver and viable tumour)CT radiomics modelSingle-centre, retrospectiveAccuracy ∼86%, sensitivity ∼70%, specificity ∼94% [63]Response prediction (progression-free survival)Lung cancerRadiomics plus clinical imaging featuresAdvanced radiomics with deep learningRetrospectiveProgression-free survival prediction under immunotherapy [64]Survival predictionPan-cancer (n = 694)Longitudinal blood biomarkers, CT-derived organ volumes, medication dataTransformer-based multimodal modelRetrospective longitudinal cohortAUC 0.81–0.84 for 3–12 month survival [65]Response and adverse-event predictionMelanomaHeterogeneous data across multiple studiesVarious ML algorithms (systematic review and meta-analysis)Multiple retrospective studiesVariable predictive accuracy; selection bias and small samples noted [70] Representative studies of AI in cancer therapy and precision oncology, organised by prediction task, with cohort characteristics, validation strategy, and reported performance. Radiotherapy planning: dose optimization, tumour contouring with AI tools AI is changing radiotherapy (RT) planning, especially in two key areas: automated tumour and organ at risk (OAR) contouring and predicting or optimizing doses. Auto segmentation tools for tumour volumes, such as gross tumour volume (GTV) and clinical target volume (CTV), and OARs are becoming more common. However, it emphasized that human supervision is essential to ensure correct dosing. This study compared two commercial auto contouring systems. It found that these tools significantly cut down contouring time. While some differences compared to manual contours were noted, many dose volume metrics stayed within acceptable limits [66]. For dose prediction, deep learning models are used to estimate spatial dose distributions and dose volume histogram (DVH) metrics. The goal is to speed up planning and improve consistency. For example, the study deep learning‐powered radiotherapy dose prediction [67] looked at models in 622 patients with different tumour types. It found a good match between predicted and actual dose distributions, indicating potential for clinical use. Another example is the multi-task learning (MTL) framework developed by Kim et al. This framework learns both contouring and voxel-based dose prediction together. It improved DVH metric errors by around 19.8% for prostate and 16.3% for head and neck compared to doing tasks separately, while keeping or enhancing contouring accuracy [68]. AI is also used to guide the setup of parameters for inverse treatment planning to minimize manual work. A review by Shan et al. discussed AI-based optimization of treatment plans and automated dose map prediction, including both DVH and voxel-based methods [69]. These are key tools for cutting planning time and improving consistency across treatment centres. However, limitations exist, such as differences in image quality, variability in contouring between observers, and inaccuracies in dosing, especially at organ boundaries when using auto contours. Studies show that even with high geometric similarity, dose-dependent errors can affect target coverage if the contours are not checked manually [66]. Predicting treatment response and toxicity It's crucial to predict which patients will respond to cancer treatments, whether immunotherapy, targeted therapy, or radiotherapy, and who might face side effects. For immunotherapy response prediction, AI models have been created to forecast progression-free survival, overall survival, and response rates across different cancers. The lung cancer radiomics study mentioned earlier [64] is one example. The work on HCC CT radiomics [63] shows that imaging features from the liver and tumour enhance response prediction. Multimodal longitudinal methods, like those by Yeghaian et al. [65], also demonstrate strong predictive ability. Systematic reviews of machine learning models for immunotherapy response in melanoma reveal that while many algorithms are promising, their performance varies. Additionally, issues like selection bias and small sample sizes are common. The meta-analysis on Machine Learning in Prediction of Immunotherapy Response and Toxicity evaluated melanoma cases and found unreliable predictive accuracy across the studies [70]. Challenges and limitations of AI in cancer care Despite its remarkable progress, the integration of artificial intelligence into oncology is not without significant obstacles. Current limitations arise from data heterogeneity, algorithmic bias, interpretability issues, and regulatory uncertainty. Understanding these challenges is essential to ensure that AI systems are not only accurate and efficient but also transparent, equitable, and clinically reliable across diverse patient populations. Data heterogeneity, quality, and bias Artificial Intelligence (AI) models in oncology often face challenges because of the differences in the quality of data. Cancer datasets usually come from various sources, including hospitals, imaging centres, and research institutions. This leads to inconsistencies in data formats, imaging protocols, and patient demographics. Such variability can make it difficult for AI models to work well across different populations and settings. For example, a study by Esteva et al. highlighted the fine-grained variability in the appearance of skin lesions as a challenge for automated image classification [71]. This happens due to the differences in imaging protocols and patient populations. In addition, issues with data quality, like missing values, noise, and inconsistencies, can reduce the effectiveness of AI models. Techniques such as data augmentation, normalization, and imputation are often used to tackle these problems, but they may not eliminate the effects of poor-quality data. Bias in AI models is another major issue. If the training data over-represents certain demographics or under-represents others, the resulting AI models may produce biased predictions. A notable example is the study by Obermeyer et al., which found that an algorithm used in U.S. hospitals was less likely to refer Black patients for high-risk care compared to white patients, despite having similar medical needs [72]. This was due to biases in the training data. Interpretability and clinician trust in AI outputs Understanding AI models is essential in healthcare. Clinicians need to comprehend and trust the decisions made by AI systems. Black-box models, like deep neural networks, often lack transparency, making it hard for clinicians to grasp how decisions are made. LIME (Local Interpretable Model-agnostic Explanations) is one method used to enhance interpretability. It approximates black-box models with simpler, interpretable models around specific predictions. Ribeiro et al. introduced LIME to explain predictions from any classifier in a more understandable way [73]. However, even though LIME and similar techniques can give some insights into model behaviour, they might not always match clinical reasoning. This can lead to mistrust among healthcare professionals. A study by Caruana et al. showed that clinicians were more likely to trust and act on predictions from interpretable models than from black-box models, even when the latter had greater accuracy [74]. A central limitation across current AI applications in oncology is the heterogeneity of the supporting evidence base. While some radiology and computational pathology tools have progressed to regulatory clearance and early clinical deployment, many applications in drug discovery, liquid biopsy, and immunotherapy-response prediction remain at the retrospective or proof-of-concept stage. Consequently, reported performance metrics should be interpreted cautiously, as they are strongly influenced by cohort composition, validation strategy, and study design. Table 5 summarises the current translational readiness of major AI applications in oncology. Table 5AI applications in oncology are summarised by translational readiness, with representative tasks, example studies or tools, current level of validation, and the principal barrier to clinical adoption.Table 5 dummy alt textApplication areaRepresentative AI tasksExample studies/toolsCurrent level of validationMain barrier to clinical adoptionDrug discovery and repurposingDrug-target prediction, virtual screening, de novo design, repurposingAtomNet, BenevolentAI;Mostly preclinical/experimental; selected candidates in clinical evaluationExperimental validation and clinical-trial translationGenomic and multi-omics analysisVariant calling, driver-mutation prioritisation, multi-omics biomarker discoveryDeepVariant, NeoMutate; multi-omics ICI-response modelsTechnically validated; emerging clinical useStandardisation and prospective clinical validationRadiologyDetection, segmentation, classification, triageTranspara, ProFound; mammography AI systemsClinically deployed (FDA/CE-cleared tools)Generalisability across scanners and patient populationsComputational pathologyWSI detection, grading, molecular phenotype predictionPaige Prostate; MSI prediction from H&EClinically deployed (de novo FDA authorisation)Scalability, stain variability, multi-centre validationLiquid biopsyFragmentomics, methylation signatures, multi-cancer early detectionDELFI; multi-analyte MCED assaysAdvanced clinical validation; not all FDA-clearedProspective demonstration of clinical utilityPrecision oncology and therapyResponse, survival, and adverse-event prediction; therapy selectionHCC and lung radiomics; transformer-based multimodal modelsMostly proof-of-concept/retrospectiveProspective, multi-centre validationRadiotherapy planningAuto-contouring, dose prediction and optimisationCommercial auto-contouring systems; DVH-prediction modelsEarly clinical adoption with mandatory human oversightDose accuracy at organ boundaries; quality-assurance requirementsDigital twins and real-time AIVirtual patient modelling, dynamic dose adjustmentCURATE.AI platformEarly-stage feasibility trialsReal-time data integration, calibration, regulatory approval AI applications in oncology are summarised by translational readiness, with representative tasks, example studies or tools, current level of validation, and the principal barrier to clinical adoption. Future perspective Multimodal AI models integrating imaging, genomics, and clinical data The next frontier in AI for oncology is combining multiple data types, including imaging (like radiology, pathology, and radiomics), genomics/multi-omics (such as transcriptomics, proteomics, and mutational profiles), and clinical data (including electronic health records, lab values, and demographics). Multimodal AI models can offer more accurate, robust, and general predictions of prognosis, treatment response, and toxicity than unimodal models. Recent reviews indicate that multimodal modelling approaches using graph neural networks (GNNs), transformer architectures, or fusion strategies perform better than those using only one type of data, such as imaging alone or genomics alone [75]. For instance, Jennings et al. systematically reviewed 48 studies across 19 cancer types. They combined histopathology images (whole slide images) with omics data and found that multimodal models often surpassed unimodal ones (c-indices ranged from 0.55 to 0.86), even though many lacked external validation and assessment of clinical use [76]. Another study, Multimodal deep learning approaches for precision oncology [77] highlights that the entire cancer care process, from early detection to treatment planning to survivorship, can improve with multimodal fusion. However, it also notes technical challenges, such as missing data, aligning different data types, and understanding combined models. Data integration still faces challenges. These include matching image data to omics samples on time, harmonizing platforms (like different sequencing methods or imaging scanners), managing absent data types, and addressing computational demands. However, ongoing advancements in architectures, like transformer-based fusion and attention mechanisms, along with transfer learning and foundation models tailored for healthcare, are likely to speed up progress. For example, recent studies show promise in transitioning from traditional machine learning to foundation models, which could enhance the use of large pretrained models for multimodal tasks in cancer and improve feature extraction across different data types [77]. Real-time AI in clinical decision making: digital twins and personalized oncology Real-time AI systems, especially digital twins, are set to change oncology by providing more dynamic, continuously adapting care. A digital twin is a virtual model of a patient or their disease that updates in real time with incoming data. This model can simulate different treatment scenarios, monitor disease progression, predict negative events, and optimize interventions. For example, the study Digital twins: a new paradigm in oncology in the era of big data explains how virtual models can include continuous updates from clinical records, multi-omics data, imaging, and treatment feedback [78]. This helps provide personalized predictions and simulations of treatment outcomes with various strategies. One example is the CURATE.AI platform (National University of Singapore), which has created patient-specific digital twins using real-time biomarker data to adjust chemotherapy doses dynamically [79]. Early reports suggest that this approach can reduce drug use by about 20%. Digital twins also allow for “what if” scenarios, simulating the effects of dose adjustments, the addition of new therapies, or the timing of interventions before they are applied to real patients, thus lowering risk. Moreover, these models can support ongoing monitoring and modifications, such as adjusting therapy if a biomarker or imaging pattern changes, enabling a more flexible approach to oncology. However, several challenges still need to be addressed. These include ensuring real-time data accuracy and updates, managing latency and computational demands, calibrating models so simulations closely match real patient outcomes, integrating into clinical workflows, obtaining regulatory approval, and keeping clinicians involved. Clinical trials examining digital twin-driven decision-making are still rare. Role of federated learning and privacy-preserving AI in cancer care A major challenge in cancer AI is gathering large, diverse, and high-quality datasets. However, sharing or centralizing data between institutions, regions, or countries raises concerns about privacy, legality, and ethics. Federated Learning (FL) and other privacy-preserving technologies (PPT) are increasingly seen as potential solutions. Federated Learning lets institutions train AI models together without sharing raw patient data. Instead, they share only model weights or gradients, which are then aggregated. A recent study on brain tumour segmentation collected data from 71 sites across six continents using an FL framework. It trained models that identified glioblastoma boundaries in radiology scans, showing strong performance even in a distributed system [80]. Another study used FL with differential privacy for breast cancer detection across different institutions. This approach ensured patient privacy while achieving good model accuracy. Privacy-preserving frameworks go beyond FL. Techniques such as differential privacy, encrypting model weights, and secure aggregation protocols enhance data security and model integrity [81]. Regulatory and ethical issues Integrating AI into cancer care brings various regulatory and ethical challenges. Regulatory and governance frameworks are increasingly being developed to address the evaluation and oversight of AI applications in healthcare [82,83]. However, the rapid pace of AI development can outpace existing regulatory and governance mechanisms. For instance, the FDA has approved several AI-based imaging devices, Paige Prostate Detect, the first AI-based pathology product to receive FDA de-novo marketing authorisation for the detection of prostate cancer foci on digitised whole-slide images [84]. Yet, the approval process remains complex, balancing safety, effectiveness, and transparency. Examples of regulatory status and clinically deployed AI tools in oncology are summarised in Table 2. The evolving nature of AI models, which can learn continuously, further complicates regulatory oversight. From an ethical standpoint, AI in cancer care raises concerns about patient privacy, informed consent, and data ownership. AI models rely on extensive datasets often containing sensitive clinical information; therefore, it is crucial to ensure proper anonymization and obtain explicit consent for secondary use [85]. Additionally, questions regarding who owns the data and the trained models must be addressed to prevent misuse and ensure accountability. Prior studies emphasize the importance of transparent algorithmic auditing, clinician oversight, and equitable data governance in healthcare AI [86]. Table 2Regulatory-authorised or clinically deployed AI tools in oncology.Table 2 dummy alt textTool / vendorModality & intended useRegulatory status / deployment (example)Key evidence / notesPaige Prostate (Paige.ai)Digital pathology — assist detection of prostate cancer foci on needle biopsiesFDA de-novo marketing authorization (prescription device; IVD use via a cleared digital pathology viewer)First AI pathology product to receive de-novo FDA authorisation; peer and regulatory documentation demonstrate improved detection/triage in validation cohorts [84].ProFound / ProFound Detection (iCAD)Breast imaging (2D mammography and Digital Breast Tomosynthesis — DBT): lesion detection and decision supportMultiple FDA clearances for versions of ProFound (including DBT updates); company / clinical press summarise improved detection in dense breasts.Shown to increase detection in dense breasts and serve as a second-reader/triage tool in clinical studies and regulatory submissions [92]Transpara (ScreenPoint Medical)Breast AI: abnormality detection and density assessment in 2D & 3D mammography (concurrent reading aid)FDA cleared (2D & 3D) and CE-marked — widely deployed internationally for screening support.Multiple peer-reviewed and real-world evaluations report improved workflow and sensitivity; used by screening programmes [93]Multi-analyte / MCED (commercial research stage)Liquid biopsy multi-cancer early detection (fragmentomics, methylation panels + ML)Several commercial MCED assays are in advanced clinical validation or have regulatory submissions pending; not all are FDA-cleared as of latest public reports. See DELFI-based research and multi-analyte studies.Evidence base growing (fragmentomics, methylation + ML) — clinical utility / population screening impact remains under active investigation [94]. Regulatory-authorised or clinically deployed AI tools in oncology. Conclusion Artificial intelligence is rapidly transforming cancer research, diagnosis, and therapy, offering powerful tools to manage the growing complexity of genomic, imaging, and clinical data. From accelerating drug discovery and biomarker identification to enabling more accurate diagnostics and personalized treatment strategies, AI has demonstrated considerable potential to augment multiple aspects of oncology research and clinical practice [87]. However, challenges such as data heterogeneity, algorithmic bias, lack of interpretability, and regulatory constraints remain significant barriers to widespread clinical adoption. Addressing these issues through rigorous validation, explainable models, and collaborative frameworks will be critical for ensuring safe and equitable integration of AI into healthcare systems. Emerging paradigms such as multimodal learning, federated approaches, and digital twins promise to further enhance precision oncology and bridge the gap between research innovations and real-world patient care [88]. By fostering multidisciplinary collaboration and innovation, AI has the capacity to augment human expertise and open a new era of predictive, preventive, and personalized cancer care. Given the broad scope of this review, Table 5 provides an integrative synthesis that summarises the major application areas by their current level of validation and translational readiness, together with representative AI tasks, example studies or tools, and the principal barrier to clinical adoption, allowing readers to distinguish early-stage research applications from tools that are already clinically deployed or near deployment. Nevertheless, prospective multi-centre validation, regulatory oversight, and integration into real-world clinical workflows remain essential prerequisites for widespread clinical adoption. Ethical statement Not applicable. Funding sources Not applicable. CRediT authorship contribution statement Muneera Anwer: Writing – review & editing, Writing – original draft, Validation, Software, Investigation, Formal analysis, Data curation, Conceptualization. Krupa Bhaliya: Writing – review & editing, Writing – original draft. Ming Q. Wei: Resources. Declaration of competing interest There are no conflicts to declare. Acknowledgements The authors would like to acknowledge the library resources at Griffith University. Data availability This article contains no datasets generated or analysed during the current study. References Associated Data Data Availability Statement This article contains no datasets generated or analysed during the current study.