The role of AI in oncology: present applications and future horizons

Abstract

Artificial intelligence (AI) offers a powerful means to accelerate precision oncology by individualizing care in an era of rapidly evolving treatment paradigms. However, there is limited regulatory oversights for safe clinical implementation of AI, and concerns surrounding data bias, ownership, and privacy have hindered broad integration into healthcare practice. In this perspective, we offer a forward-looking roadmap for the dissemination of AI in oncology. We discuss the role for AI in guiding biomarker-driven patient selection for clinical trials and in facilitating both personalized drug selection and de novo drug design by the potential to predict therapeutic response. We highlight the capabilities, and current limitations, of AI to inform realistic pathways for practical implementation.

Leveraging AI in clinical oncology: an introduction

Artificial intelligence (AI) holds tremendous promise for streamlined decision-making and data-derived treatment in healthcare. As it pertains to oncology, where patient selection, therapy response, and drug development are highly nuanced, AI offers a powerful means to accelerate precision medicine and narrow the gap between academic centers and community care.

Yet, for many oncologists, terms like “AI” and “precision medicine” remain largely conceptual if not aspirational. While there is current authoritative guidance on the clinical trial implementation of AI, concerns surrounding data have hindered broad integration^1–3. In this perspective, we offer a forward-looking roadmap for the dissemination of AI in oncology. This evolutionary timeline begins with AI’s present use in abstraction for clinical trial selection, progresses toward phenotype-driven therapy selection, and ultimately envisions AI-enabled drug design (Fig. 1). At each stage, we highlight both capabilities and current limitations to inform realistic pathways for integration. Though not a formal systematic review, this perspective draws on a thorough literature search and input from academic and commercial leaders.

Fig. 1. Visual orientation to the integration of artificial intelligence (AI) in application to oncology.

Created in Biorender without the use of AI assistance.

AI-assisted patient selection in oncology clinical trials

Randomized clinical trials are the cornerstone of oncologic advancement. Such studies rely on sufficient enrollment of patients who meet specific eligibility criteria. Traditionally, this is a process involving manual chart review and data abstraction, presenting a major barrier to trial accrual. AI, particularly through natural language processing (NLP) and large language models (LLMs), offers an unparalleled ability to synthesize large datasets and parse unstructured clinical data to streamline clinical trial matching. This improvement is achieved without a loss in key performance metrics and diagnostic yield, with systematic reviews reporting accuracy, sensitivity, and specificity consistently exceeding 80%^4,5. Initial AI efforts in oncology clinical trials have leveraged NLP to extract cancer phenotypes from rich clinical text, with notable efforts focused on breast and lung malignancies⁴. Tools such as Mendel.ai and Watson for Clinical Trial Matching have demonstrated automated eligibility screening by filtering patient chart information to standardized medical terminology matched to clinical trials⁶. This workflow significantly reduced manual labor time in the early stages of patient selection⁷. A 2025 review by Saady et al. identified 16 studies focused on AI in clinical trial enrollment and highlighted near-universal improvements in screening efficiency, with time reductions of 25% or more⁸. More recently, as utilized in 7 out of 33 clinical trial studies published in 2023, LLMs have advanced the interpretation of complex trial protocols, optimizing patient-trial matching⁹. TrialGPT is one such system that uses a three-stage pipeline: extracting relevant trials, evaluating eligibility criteria, and reversing the trial paradigm to generate a ranked list of candidate trials for the patient to consider. GPT-4-based TrialGPT achieved a combined-feature performance of 0.73 while reducing screening time, suggesting that newer model iterations (GPT-5.2) could further enhance trial ranking.¹⁰. Notably, this NCI/NIH-built system provides explanations for eligibility decisions to enhance interpretability and an open-source codebase¹¹. In parallel, LLM-based tools, Criteria2Query 3.0 and AutoCriteria, have refined the extraction of granular eligibility criteria from clinical trial documents^12,13. These systems generate structured queries from trials formatted for downstream software to search large patient databases, achieving a similar goal to TrialGPT without an embedded workflow.

Beyond protocol interpretation or reliance on synthetic patient datasets, emerging AI systems can be integrated directly within point-of-care pathways. This allows the use of real-world patient data to refine trial eligibility. More importantly, embedded workflow promises the democratization of clinical trial recruitment beyond high-volume centers and potentially enables smaller, peripheral sites to accrue patients. MatchMiner, an open-source platform developed at the Dana-Farber Cancer Institute, has streamlined the consent-to-trial pathway by incorporating genomic profiling into the matching process^14,15. This system has been used in innovative ways, such as automatically notifying oncologists via email when a solid tumor patient with panel DNA next-generation sequencing is matched to a trial. This process achieved a 17% discordance rate between AI-generated matches and expert physician assessments¹⁶. Similarly, OncoLLM, a domain-specific, EHR-embedded model trained on single-center data, has demonstrated physician-level accuracy in identifying appropriate clinical trials for oncology patients¹⁷. Together, these integrated models may bridge the gap between abstract model development and point-of-care deployment.

While trial matching solutions continue to mature, the next frontier lies in AI-powered adaptive trials that iteratively learn from participant data streams to refine eligibility criteria, modify treatment arms, and recalibrate endpoints^18,19. This shift from static to dynamic protocol is largely advanced by industry leaders in direct partnership with pharmaceutical sponsors and promises the ability to make trials more efficient, personalized, and effective.

Ideally, AI-enabled patient-trial matching would reduce recruitment costs and shorten the timeline for trial activation. Yet, current systems still require human oversight to validate AI-generated matches, and software vendors are acutely aware of the legal risk posed by using AI in patient care²⁰. In addition, AI-driven selection can reinforce existing biases, particularly when training datasets underrepresent diverse patient populations and are not transparent about which parameters are valued (closed-weight)^21,22. For example, tools such as TrialGPT rely on a synthetic, closed-weight dataset for development and have critical dependencies on cloud-based vendors. While platforms like MatchMiner mitigate this issue by using real patient data, they are still closed-weight and face barriers to scalability due to dependence on institution-specific infrastructure. Expert reviews consistently echo these limitations, citing small sample sizes, restricted access to patient records, and limited generalizability as persistent challenges across published models^6,8. Alternatively, open weight LLMs, such as LlaMA2 and Mistral, may offer more transparency and create reproducibility and reinforce provider trust in clinical trial recruitment²³.

Although LLMs may appropriately select patients meeting trial criteria¹⁷, a concern is that most LLM-based patient selections operate as “black boxes,” with limited insight into how variables may have been weighted in patient selection. This lack of transparency, coupled with the potential for AI hallucinations or confabulated inputs, can undermine clinician trust and clinical adoption, a challenge emphasized by expert review of clinical trial applications^24–26. If clinicians cannot understand how a match is generated or created, the tool, no matter how accurate the underlying logic is, becomes a source of concern. Many LLMs are also prompt-dependent; narrow queries can drive blind data analysis in clinicians without expertise, potentially propagating misinformation and mistrust. To fully realize the potential of AI in clinical trial recruitment, future systems must prioritize data diversity, interpretability, and transparency. As emphasized by Verlingue et al., AI must not disrupt the face-to-face communication between the patient and physician²⁷. There should be multi-institutional efforts to build data sources that capture the rich diversity of patients, ensuring that AI models are truly representative. While initiatives like Truveta have begun to establish shared data repositories, the lack of clear guidelines around IP ownership has hindered collaborative model training²⁸. If these challenges are addressed, AI holds significant promise to fundamentally reshape how oncology clinical trials are conducted.

AI-predicted drug selection and response in oncology

Beyond optimizing clinical trial recruitment, AI has significant potential to proactively guide treatment selection and reduce therapy inefficiency²⁹. Given that many patients receiving oncology drugs do not achieve the desired effect, multimodal AI platforms incorporating disparate data sources like genetic or laboratory assays, demographic variables, clinical information, and biowearable physiological signals may optimizedrug efficacy. When applied to large, multimodal datasets, AI techniques, especially deep learning (DL) approaches, can stratify patients into clinically similar cohorts to improve forecasting of treatment responses^30,31.

As we gain a deeper understanding of disease heterogeneity and the expanding role of immunotherapy, AI is increasingly being used for initial treatment selection in oncology. An outline of existing models and their contributions is provided in Table 1. In parallel, industry partners have advanced AI-driven platforms for drug selection (Table 2). Using machine learning (ML) with EHR data, Pan et al. derived predictive clinical subphenotypes for non-small cell lung cancer to determine response to immune checkpoint inhibitor (ICI) therapy, outperforming other models³². At an individual level, AI has been harnessed to predict ICI outcomes across cancer types using routine blood tests, improving on FDA-approved biomarkers³³. At the cellular level, neural networks can “learn” tumor archetypes, facilitating selection of targeted therapy³⁴. The large computational power of ML has even been harnessed to predict drug responses from single-cell RNA sequencing data^35,36.

Table 1.

Representative list of peer-reviewed, published models of AI in application to drug selection, identified through targeted PubMed and Scopus searches

Model name	Model type	AI approach	Oncology application	Stage	Description	Performance metrics	Reference	Year
PERCEPTION	Open Source	ML	Multiple myeloma, breast cancer, NSCLC	Selection	Predicts resistance and sensitivity to FDA-approved therapies from single-cell and bulk RNA-seq data.	AUC = 0.81 for predicting sensitive cell lines, AUC = 0.83 for stratifying responder status.	Sinha et al.35	2024
ATSDP-NET	Proprietary	DL	Oral squamous cell carcinoma, prostate cancer, and multiple myeloma	Selection	Predicts treatment response from single-cell and bulk RNA-seq data.	High correlation (R = 0.89) between predicted and actual gene sensitivity scores.	Zhou et al.34	2025
DL-Based MSI/dMMR	Proprietary	DL	Colorectal cancer	Selection	Detects MSI/dMMR from whole-slide H&E pathology images.	AUC = 0.89 for predicting MSI/dMMR status.	Echle et al.36	2022
DL-Based MSI-H	Proprietary	DL	Prostate cancer	Selection	Detects MSI-H from whole-slide H&E pathology images.	AUC = 0.78 for predicting MSI status.	Hu et al.37	2024
SCORPIO	Open Source	ML	Pan-Cancer (21 types)	Selection	Predict the clinical benefit of ICI therapy from CBC/CMP and clinical characteristics.	AUC = 0.76 for predicting 6–30 months overall survival (OS), outperforming tumor mutational burden.	Yoo et al.32	2025
GEMS	Proprietary	ML	NSCLC	Treatment Response	Stratifies ICI-treated patients into predictive subphenotypes based on EHR clinical data.	C-index = 0.67, outperforming unsupervised clustering methods to predict OS.	Pan et al.30	2025
AANet	Proprietary	ML	Breast cancer	Treatment Response	Identifies key cell archetypes using spatial transcriptomics.	Minimized mean square error (MSE) between the input expression and reconstructed profile.	Venkat et al.33	2025
H&E DL Angio	Proprietary	DL	Renal cancer	Treatment Response	Predicts response to anti-angiogenic therapy from tumor regions on whole-slide H&E pathology images.	Strong correlation (c-index = 0.67) with costly gold standard, Angioscore.	Jasti et al.35	2025
CHAI	Proprietary	DL	NMIBC	Treatment Response	Predicts BCG response from histologic assays and clinical variables.	Model stratified patients to predict 3.9 higher odds of progression, 2.3 higher odds of BCG-unresponsive disease, and 3.4 higher odds needing cystectomy.	Lotan et al.39	2024
GMLF	Proprietary	DL	MIBC	Treatment Response	Predicts neoadjuvant response from whole-slide images and tissue gene expression.	AUC = 0.74 for predicting response to neoadjuvant chemotherapy.	Bai et al.40	2025
PDRP	Open Source	ML	NSCLC	Treatment Response	Predicts treatment response from the geometry of the drug-target binding site.	97.5% accuracy, 93% recall, 96.5% precision, and 94% F1-score for a 4-class drug response prediction task.	Qureshi et al.41	2022
DRN-CDR	Open Source	DL	Pan-Cancer (24 types)	Treatment Response	Models dose-response using drug molecular structures, IC50, Cancer Cell Line Encyclopedia (CCLE).	AUC = 0.76 for classifying drugs as sensitive or resistant.	Saranya et al.49	2024
DrugCell	Open Source	DL	Pan-Cancer	Treatment Response, Synergy	Predicts single-agent response and synergy.	High correlation (R = 0.80) between actual and predictive drug response.	Kuenzi et al.48	2020
CURATE.AI	Proprietary	DL	Advanced, solid tumors on palliative care, metastatic prostate cancer	Dosing	Creates and modulates a dosing regimen based on individualized biomarkers.	Expert clinician users chose to follow dosing recommendations for 97.2% of decisions.	Blasiak et al.42	2025

Table 2.

Selected commercial AI platforms for clinical trial and therapy selection in oncology, including their computational approaches and clinical use cases

Company	AI approach	Oncology application	Stage	Description	Country of origin
CertisAI	ML	Pan-cancer (Solid and liquid tumors)	Drug Selection	Tumor molecular profiling is completed, and 5–20 genes of interest are identified to create a proprietary gene expression signature that is compared to FDA-approved medications.	USA
CureMatch	ML	Pan-cancer (solid tumor)	Drug Selection	Tumor molecular profiling is completed, and the efficacy of various treatments is ranked, including drug combinations.	USA
Foundation Medicine (Roche)		Pan-cancer (solid and liquid tumors)	Clinical Trial Selection, Drug Selection, Treatment Response	Genomic and tumor profiling is used to match patients to targeted therapy and clinical trials.	Denmark
Predictive Oncology Inc	DL	Breast, colon, ovary, AML, MM	Drug Selection, Drug Repurposing	Assesses FDA-approved drugs against ovarian tumor samples. Drug repurposing from the biobank of dissociated tumor cells. Develops organoid models for hematological (AML, MM) malignancies.	USA
Tempus	DL	Pan-cancer (solid and liquid tumors)	Clinical Trial Selection, Drug Selection	Partners with academic centers for site activation of oncology trials. Commercially available for multi-omic sequencing to predict patient treatment response to ICI, generation of tumor organoids.	USA
ArteraAI	DL	Prostate	Treatment Response	A prognostic model that combines digital pathology images with clinical data to predict response to hormonal therapy and combination treatments in high-risk prostate cancer.	USA
ConcertAI, Guardant Health	ML	Pan-cancer (>60 solid tumor types)	Drug Selection, Treatment Response	A combination of solid tumor liquid biopsies and clinical data is used to select treatment and monitor therapy response in advanced cancer.	USA
Lantern Pharma (RADR)	ML	Pan-cancer (solid and liquid tumors)	Treatment Response	Uses a proprietary multi-omic database to stratify patients into responder, partial responder, and non-responder categories.	USA

Given that microsatellite instability (MSI) is an indicator of response to immunotherapy, AI has been applied to characterize this molecular-level biomarker. In colorectal cancer, DL has predicted MSI status directly from pathology slides with clinical-grade performance, accelerating therapeutic decision-making³⁷. A similar model was developed for prostate cancer, where MSI status is not routinely tested; an AI screening tool may broaden access to ICI by identifying eligible patients³⁸. DL digital histopathology models have shown promise in multiple settings, including identifying patients with metastatic clear cell renal cell carcinoma responsive to anti-angiogenic therapy at a fraction of the cost and predicting BCG-responsiveness in non-muscle-invasive bladder cancer (NMIBC)^39,40. With AI, whole-slide image data can be supplemented with gene expression profiles for superior predictive accuracy. This has been implemented in predicting the response to neoadjuvant chemotherapy in muscle-invasive bladder cancer (MIBC)⁴¹. Macroscopically, AI models contextualizing protein-drug interaction features with patient data have accurately predicted EGFR-TKI responses in NSCLC⁴².

Multimodal assessment with AI can also anticipate dynamic changes in treatment response, enabling earlier intervention before toxicity or resistance is clinically evident. CURATE.AI is an agnostic platform that uses clinical biomarkers alongside patient medical history to dynamically adjust a palliative chemotherapy dosing regimen⁴³. This system has also been trialed in metastatic castration-resistant prostate cancer to modulate combination chemotherapy dosing on an individual-by-individual basis⁴⁴. Given that targeted radionuclide therapy has become a treatment option in select metastatic cancers, emerging digital twin platforms in nuclear oncology promise patient-specific simulations of radiopharmaceutical kinetics to optimize dosing and minimize adverse toxicity⁴⁵. Dose modulation based on patient-specific multi-omics has drawn considerable attention from industry, with companies like OncoKDM and DoseMeRx aiming to personalize dosing strategies. Similarly, when this dynamic dosing regimen is paired with outcomes-based feedback loops, it challenges the need for large-scale clinical trials, enabling continuous learning mechanisms for personalized treatment regimens. AI also holds promise in assisting with response-adaptive radiotherapy, increasing physician confidence to maximize tumor control while minimizing side effects^46,47. Building on these applications, AI has been employed to identify synergistic drug combinations from existing large-scale datasets^48,49.

In an ideal setting, trends in biomarkers, along with clinical and genetic data, could be continuously integrated using AI to proactively modify treatment plans before detrimental effects or suboptimal response^43,50. However, broad implementation remains limited by data fragmentation and the challenges of synthesizing high-volume information in real-time. High-quality and well-annotated data are necessary to train models, yet relevant data sources, such as oncology notes and imaging, are difficult to incorporate. Even when high-quality data is extracted, small datasets are prone to overfitting, resulting in poor generalization to new patient data, particularly when the data the model is applied to is significantly larger in magnitude than the training data itself. Beyond this, ground truth data for predicting pathologic or physiologic response is not standardized. In the models discussed above, some assign a complete response as ground truth, whereas others do not. As such, models may differ in their utility based on what the training set comprises and how training was achieved.

Achieving real-time AI-driven treatment adaptation would require centralized data repositories (data lakes) and seamless EHR integration that is difficult to coordinate within institutions, let alone between institutions. Even when large data synthesis is achieved, LLMs can become overwhelmed and inadvertently overweight irrelevant factors. Small language models may be a good initial step before federated learning with LLMs, accomplishing targeted computing tasks and rapid iteration locally within the healthcare systems. Their ability to function locally also bypasses privacy concerns and shows promise for N-of-One therapy⁵¹.

AI in de novo drug design: the final frontier

While the use of AI in optimizing existing treatments continues to evolve, transformative potential may lie even further upstream. The discovery and development of new oncologic therapies often spans decades and costs millions, only to have the majority fail to gain regulatory approval. By processing vast amounts of data and simulating human decision-making through neural networks, AI can accelerate the bench-to-bedside workflow of pharmaceutical development.

Just as AI may uncover overlooked patterns in existing treatment selections, it is now used to generate entirely new therapeutic compounds. The presence of AI-discovered molecules in clinical trials has grown exponentially from the early stages of PubChem bioassays and the clinical trial introduction of AI-discovered molecules in 2015^52,53. AI can contribute to every stage of drug design: identifying molecular targets, predicting protein-ligand interactions, and modeling in-vivo drug behavior with simulation of pharmacologic properties.

For example, druggability, the likelihood a gene can be targeted for therapeutic effect, can be predicted with DrugnomeAI and PINNED, ML frameworks that use genomic data to identify oncologic therapeutics^54,55. Other AI-learning systems also exploit the concept of druggability, outperforming existing methods^56–58. Similarly, AI has enhanced the identification of synthetic lethal interactions, a strategy that selectively kills cancer cells by targeting gene pairs, with neural network models advancing prior computational methods⁵⁹. A major leap came with the development of AlphaFold, a neural network platform to predict protein structure with high accuracy at a fraction of the cost and time of traditional techniques, like X-Ray crystallography⁶⁰. When coupled with generative adversarial networks, AI can further refine predictions of protein-ligand interactions to accelerate lead compound development^61,62.

In addition to early-stage discovery, AI is also reshaping the modeling of drug-host interactions. Simulations now allow for the prediction of pharmacokinetics/pharmacodynamics. As highlighted in a preliminary analysis by Jayatunga et al., these models have shown improved success rates in Phase I trials compared to industry standards^52,63.

Several of these AI-driven approaches have already led to in silico investigational compounds. Although, as of this publication, no AI-designed drugs have completed the full approval process for oncology indications, a growing number have received Investigational New Drug clearance. One notable example includes a CDK12/13 dual inhibitor, identified through a multi-omic AI pipeline to block upregulation of DNA-damage response pathways and the subsequent tumorigenesis⁶⁴. AI has also aided the design of proteolysis-targeting chimeras (PROTACs); bifunctional molecules that induce the degradation of disease-relevant proteins within a cell. AI assistance led to a PROTAC targeting p300, a chromatin regulator implicated in prostate cancer progression⁶⁵.

Beyond designing new compounds, AI promises the ability to repurpose existing therapies that may otherwise be orphaned, a faster and cheaper process than developing new agents. For example, the neural network-based GraphRepur model extracts drug gene expression signatures and structural information to identify new therapeutic indications, such as repurposing agents for breast cancer⁶⁶. Interest in AI for drug development and repurposing is ever-growing, and the field is already populated by several clinical-stage companies (Table 3).

Table 3.

Representative AI enterprises for drug discovery and repurposing in oncology, summarizing their applications and design features

Company	AI approach	Oncology application	Stage	Description	Country of origin
Numerion Labs	DL	Pan-cancer (solid and liquid tumors)	Drug Design	Developed a protocol to quickly screen chemical libraries and identify drug-like molecules that can bind disease-related targets (i.e., androgen receptor sites).	USA
Benevolent AI	ML	Pan-cancer (solid and liquid tumors)	Drug Design, Drug Repurposing	Uses “patient-level data” to group patients into endotypes for subsequent drug discovery. Partnered with Novartis to investigate new indications for drugs in clinical development.	Luxembourg
Exscientia	DL	Pan-cancer (solid and liquid tumors)	Drug Design	Trains on pharmacology data and patient multi-omics to iteratively design and chemically synthesize targeted compounds, including an A2a receptor antagonist.	UK
Insilico	DL	Pan-cancer (solid and liquid tumors)	Drug Design	Uses AI generative chemistry platforms (Chemistry42) to generate molecular scaffolds tailored to target binding affinity, with a specific focus on tumors that invade the immune system.	USA
Recursion Pharmaceuticals	DL	Pan-cancer (solid tumor, B-cell lymphoma)	Drug Design	Conducts molecular design using multi-omics data and iterative loops, discovering a CDK12-adjacent target for DNA-damage response.	USA
Schrodinger	DL	Pan-cancer (solid tumor)	Drug Design	Integrates ML and physics-based algorithms to predict protein-ligand binding affinity and identify chemotypes.	Germany
Xaira Pharmaceuticals	DL	Pan-cancer (solid and liquid tumors)	Drug Design	Created the largest Perturb-seq platform and uses diffusion-based generative AI for protein design. Has implemented early research in CTLA-4 antibodies.	USA

While the term “precision medicine” is often applied broadly, many current programs fall short of true patient-level specificity. The full actualization would involve drugs custom-tailored to an individual’s gene expression, somatic mutations, and clinical context. This vision may become possible with the maturation of General AI, systems capable of integrating germline sequencing, tumor-specific genomic alterations, EHRs, environmental exposures, and more to generate drug candidates. Achieving this goal requires not only high-volume, high-quality data but also systems that deliver interpretable recommendations. In addition, the concept of General AI requires a combination of multiple complex computational skills: a higher level of transfer reasoning, short-term “memory” to support self-directed and single-shot learning, and datasets that support reasoning beyond narrow AI’s capabilities.

While AI has shown strong performance across both well-defined and increasingly complex tasks, current systems remain dependent on human oversight and are constrained by several limitations. Datasets used in drug discovery often involve millions of candidate compounds and can overwhelm traditional ML algorithms⁶⁷. Although AI is promoted to reduce cost, the computational demands of high-complexity modeling impose a significant resource burden that should not be overlooked⁶⁸. Moreover, AI-generated molecules often resemble known ligands, suggesting current models are better suited for optimizing existing chemical scaffolds versus generating entirely novel hits⁶⁹. While the majority of AI-based drug design has leveraged supervised learning to follow existing chemical scaffolds, unsupervised learning may hypothetically generate truly novel compounds. Yet, many of these drug designs fail secondary to toxicity, synthesis feasibility, pharmacokinetics, or biocompatability⁷⁰. As with AI in clinical trial matching and drug selection, external validity is also a major challenge⁶. The diversity of models and the absence of a standardized protocol hinder regulatory acceptance^71,72.

Perhaps the most formidable barrier lies in this latter domain⁷³. Oncology drugs already undergo rigorous evaluation for safety and efficacy; the introduction of AI-generated variants raises several questions. Would every unique molecular permutation require separate approval? What role would AI systems play in submission and review?⁷⁴ What safeguards would be needed to protect patient data, particularly when therapies are derived from genomic profiles⁷⁵?

No matter how new therapies are generated, the bottleneck may be obtaining real-world clinical data⁶⁸. Until models can reliably simulate accurate, physiologic responses for novel therapies, clinical validation will continue to be the limiting step^76,77. The use of AI in drug design is one of the most ambitious frontiers in oncologyand the potential is shadowed by the ethical, technical, and regulatory hurdles. Realizing individualized cancer therapies will depend not only on algorithmic innovation, but on the development of infrastructure and governance standards to ensure safety and reproducibility in the new era of oncologic therapeutics.

Conclusion

AI now touches virtually every stage of oncologic care. LLMs substantially reduce trial-screening burden, while industry groups have launched AI-enabled adaptive trials to scale recruitment and prioritize the most promising studies. Advances in computation have enabled academic and commercial systems that integrate “multi-omic” data to forecast therapeutic benefit and resistance on an individual level. In parallel, AI-driven platforms for de novo discovery and repurposing have accelerated drug design. Yet, broad clinical impact will depend on establishing the infrastructure, standards, and oversight necessary for equitable implementation.

Author contributions

A.W. contributed to methodology, original manuscript drafting, and revisions. N.S. contributed to conceptualization, original manuscript drafting, revisions, and supervision. A.R. contributed to conceptualization, original manuscript drafting, revisions, and supervision.

Data availability

No datasets were generated or analyzed during the current study.

Competing interests

The authors declare no competing interests.