AI Moves From Lab Curiosity to Cancer Care’s New Backbone

Artificial intelligence has quietly become one of the most consequential forces in modern oncology, and a new comprehensive review published in Clinical Cancer Bulletin lays out just how far the technology has traveled. Written by Abuhurera Subhan and Geetha Manoharan, the analysis traces AI’s fingerprints across the entire cancer care pipeline, from the earliest days of molecular target identification to the regulatory offices of the U.S. Food and Drug Administration. The stakes could hardly be higher. Cancer claimed roughly 10 million lives in 2020 alone, and the traditional drug development machine has struggled to keep pace with the disease’s notorious complexity and heterogeneity. For every 5,000 to 10,000 compounds screened in conventional pipelines, only one typically makes it to market, a brutal attrition rate that has made the search for faster, smarter tools an existential priority for the pharmaceutical industry.

The core advantage AI brings to oncology is its ability to digest high-dimensional, multi-omic data at a scale no human team could match. By integrating genomics, proteomics, and clinical trial data, machine learning algorithms can detect subtle, non-linear relationships that traditional statistical tools simply miss. The review highlights how AI-driven platforms have compressed the journey from target identification to lead compound optimization, potentially shrinking a process that once took years into a matter of months. That acceleration does more than save money; it enables the rapid identification of patient subgroups most likely to respond to a given therapy, which is the practical heart of precision medicine. Techniques spanning machine learning, deep learning, natural language processing, and reinforcement learning now touch every stage of the oncology drug lifecycle, from discovery benches to post-marketing surveillance.

Among the technical standouts is DeepDTA, a model built on convolutional neural networks that predicts binding affinities between drug compounds and target proteins using nothing more than sequence information. Benchmarked on the widely used KIBA and Davis datasets, it has outperformed traditional molecular docking approaches on metrics like mean squared error and concordance index. Equally intriguing is the Cascade Deep Forest model, an ensemble framework of layered decision trees that beats deep neural networks on multiple datasets while resisting overfitting, even when training data is scarce. That robustness, the authors argue, makes it a strong candidate for real-world drug discovery pipelines where pristine datasets are a luxury rather than a given. These architectures matter because drug-target interaction prediction is the bottleneck that determines which molecules ever see the inside of a laboratory.

Generative AI is pushing the frontier even further. Generative adversarial networks and variational autoencoders, trained on massive chemical libraries like ZINC and ChEMBL, can design entirely novel molecules optimized for high binding affinity, low toxicity, and favorable ADMET profiles. Transformer-based models such as ChemBERTa apply natural language processing techniques to SMILES strings, the text-based representations of chemical structures, to predict properties like solubility and drug-likeness with competitive accuracy. Graph-based neural networks including GraphDTA and Mol2Vec encode molecular structures as graph embeddings, uncovering hidden bioactivity patterns in vast chemical spaces. And then there is AlphaFold 3, which delivers highly accurate predictions of protein, DNA, RNA, and ligand structures, effectively revolutionizing structure-based drug design and druggability assessments. Large language models are now layering on top of these tools, proposing new molecular entities with a speed and novelty that would have seemed like science fiction a decade ago.

The translational promise is no longer theoretical. Biopharmaceutical companies such as Exscientia and BenevolentAI have reported AI-discovered candidates advancing into Phase I and Phase II clinical trials, offering concrete proof that computational pipelines can produce molecules fit for human testing. AI is also proving adept at predicting synergistic drug combinations, with machine learning models trained on transcriptomic and proteomic data identifying novel drug pairs that show synergistic effects in preclinical cancer models. This is particularly valuable for stubborn malignancies like glioblastoma and pancreatic adenocarcinoma, where monotherapy routinely fails. Meanwhile, natural language processing tools such as PubTator mine the exploding biomedical literature to surface hidden associations between drugs, genes, and diseases, and their integration with large language models is supercharging hypothesis generation across the field.

Beyond the lab, AI is reshaping the full cancer care continuum. Deep learning algorithms now analyze mammograms, CT scans, and MRIs with radiologist-level accuracy, flagging suspicious images for urgent review and cutting false-negative rates in breast, lung, and prostate cancer screening. In pathology, whole-slide image classifiers have exceeded 90 percent accuracy in distinguishing benign from malignant lesions and in grading prostate and breast carcinomas, while also identifying predictive biomarkers like HER2 expression and PD-L1 status. One convolutional neural network trained on more than 40,000 CT scans matched or exceeded radiologists in detecting lung nodules. AI-driven auto-segmentation tools in radiotherapy delineate tumors and organs-at-risk with less inter-observer variability, and clinical decision support systems integrate genomic data with real-world evidence to recommend optimal treatment regimens. Wearable sensors paired with AI monitor recovering patients for complications like lymphedema, and models trained on electronic health records can flag patients at risk of sepsis or cardiotoxicity before symptoms become critical.

But raw accuracy, the review stresses, is not enough. The authors champion a concept they call model actionability, the degree to which an AI system can genuinely influence clinical or research decisions by reducing diagnostic and therapeutic uncertainty. One promising evaluation approach is decision impact analysis, which estimates how model recommendations change outcomes. In a prospective study of a machine learning tool predicting immunotherapy response in melanoma, the model’s guidance altered treatment strategies in 32 percent of patients and was associated with improved progression-free survival at six months. Interpretability techniques such as SHAP and LIME help reveal which features drive predictions, fostering clinician trust and easing regulatory review. Validation strategies including external cohorts and synthetic minority oversampling guard against overfitting, and the FDA’s proposed Total Product Life Cycle framework pushes sponsors to define predetermined change control plans so adaptive models can be updated post-approval without sacrificing safety oversight.

Regulators worldwide are scrambling to keep up. The FDA’s 2019 discussion paper on AI-based Software as a Medical Device outlined a lifecycle approach to oversight, while its Digital Health Software Precertification program evaluates companies rather than individual products, potentially streamlining approvals for AI-enabled oncology platforms. The European Medicines Agency has issued reflection papers emphasizing transparency, reproducibility, and explainability, warning that black-box models may face heightened scrutiny unless their outputs can be clearly justified. International bodies like the International Coalition of Medicines Regulatory Authorities are working toward harmonized principles, and sponsors now increasingly must submit validation datasets, explainability assessments, and performance benchmarks during investigational and new drug applications. Regulators are also demanding evidence that AI models generalize across populations, a critical concern in oncology where treatment responses vary by ethnicity, tumor subtype, and molecular profile.

Formidable challenges remain on the road to widespread adoption. Oncology data is often siloed, incomplete, and inconsistently formatted, and models trained on retrospective single-center datasets frequently fail to generalize. Algorithmic bias is a serious threat, with studies showing that some radiology and genomics models underperform in minority populations due to underrepresentation in training data. Federated learning offers a partial fix by training models on decentralized datasets without sharing raw patient information, preserving privacy while improving diversity. Interpretability, legal accountability for AI-driven errors, intellectual property questions around AI-generated discoveries, and workforce readiness all demand attention, and low-resource settings risk being left behind entirely. The authors’ prescription is a collaborative one: machine learning experts, oncologists, ethicists, regulators, and patient advocates working together toward transparent, equitable, and rigorously validated AI. If that vision holds, the future of cancer care may be more predictive, more personalized, and more patient-centered than anything the field has yet seen.

Subject of Research: Applications of artificial intelligence and machine learning in oncology drug development, clinical decision-making, and regulatory integration

Article Title: Advancing cancer care through artificial intelligence: from innovative models to clinical decision-making and regulatory integration

Article References: Subhan, A., & Manoharan, G. (2025). Advancing cancer care through artificial intelligence: from innovative models to clinical decision-making and regulatory integration. Clinical Cancer Bulletin, 4(1), Article 23. https://doi.org/10.1007/s44272-025-00052-0

Image Credits: AI Generated

DOI: 10.1007/s44272-025-00052-0

Keywords: artificial intelligence, oncology, drug discovery, machine learning, deep learning, AlphaFold 3, clinical decision support, FDA regulation, precision medicine, drug-target interaction, model actionability, federated learning