Artificial intelligence is rapidly changing the way scientists investigate cancer. AI systems can analyze vast biological datasets, detect previously unrecognized patterns, generate hypotheses, and propose new therapeutic candidates. Some can even help design experiments to test their own predictions.
These capabilities are already translating into clinical development. One example is IAM1363, an AI-designed HER2 inhibitor currently being evaluated in patients with advanced solid tumors. Early clinical findings have demonstrated antitumor activity, providing evidence that AI-driven drug discovery can generate promising therapeutic candidates for oncology.[1, 2]
Beyond the development of new therapeutic candidates, AI is also opening opportunities for drug repurposing. In a recent study, researchers used an AI Co-Scientist system to identify potential treatments for acute myeloid leukemia. Several of the proposed candidates subsequently demonstrated activity in laboratory experiments, illustrating how AI-generated hypotheses can lead to experimentally validated findings and uncover new therapeutic opportunities.[3]
Together, these advances raise an intriguing question: Could artificial intelligence be behind the next major breakthrough in cancer research?
Increasingly, the answer appears to be yes. AI is already contributing to the discovery or new therapeutic targets and the development of promising drug candidates. But identifying a potential breakthrough is only the beginning; establishing biological significance, validating its therapeutic potential, and ultimately demonstrating meaningful benefit for patients remain key challenges.
Understanding both the possibilities and the limitations of AI is essential to assessing how it could reshape the future of oncology research.
AI Is Already Transforming Cancer Research
The impact of AI extends well beyond the design of individual drug candidates. Across oncology research, computational approaches are changing how scientists investigate disease biology, identify therapeutic targets, and develop potential treatments.
One of the most significant advances has been in protein structure prediction. For decades, determining the three-dimensional structure of proteins through experimental methods was technically demanding and often time-consuming, requiring researchers to isolate and purify proteins, prepare suitable samples, and use specialised techniques (such as X-ray crystallography, nuclear magnetic resonance spectroscopy, or cryo-electron microscopy) that could take months or even years.
AlphaFold, an AI system developed by Google DeepMind to predict the three-dimensional structures of proteins from their amino acid sequences, demonstrated that deep learning could predict many protein structures with remarkable accuracy, providing researchers with valuable insights into biological mechanisms and potential drug-binding sites.[4] The work contributed to the 2024 Nobel Prize in Chemistry.
AI is also being used to identify potential therapeutic targets within complex genomic and molecular datasets, to design and optimize drug candidates, and to discover biomarkers that help predict how a patient’s cancer will behave. Deep learning models have shown, for example, that certain genetic features of gastrointestinal tumors, such as microsatellite instability, can be predicted directly from routine pathology slides.[5] In clinical research, AI is supporting trial design and helping match patients to the studies most likely to benefit them.
Early results from drug development are encouraging, although important questions remain. A 2024 analysis of AI-discovered drug candidates reported Phase I success rates of approximately 80–90%, substantially higher than historical industry averages. However, Phase II success rates were closer to conventional benchmarks, at approximately 40%, based on a limited sample of programs.[6] These findings suggest that AI may improve the identification and development of candidates with properties suitable for early clinical testing. Whether those advantages consistently translate into better therapeutic outcomes remains to be established.
This distinction is crucial: accelerating drug discovery does not necessarily mean overcoming the biological and clinical challenges that cause promising cancer therapies to fail.
The Challenge of Understanding Cancer Biology
AI is exceptionally good at recognizing complex patterns, generating hypotheses, and exploring relationships across enormous volumes of information. But every AI system is fundamentally shaped by the data it learns from. While it can identify previously unrecognized connections, propose novel biological mechanisms, and challenge existing assumptions, its predictions are not equivalent to experimental evidence. Establishing whether these hypotheses reflect biological reality still required rigorous scientific investigation and validation.
Consider cancer genomics. A tumor can accumulate hundreds or even thousands of genetic alterations, yet only a small subset of these, known as “driver” mutations, directly contribute to disease development and progression. Distinguishing these driver mutations from passenger mutations, which accumulate along the way without contributing to cancer progression, is essential for understanding cancer biology and identifying therapeutically relevant targets.[7]
AI can help identify candidate drivers, uncover relationships between molecular alterations, and predict which mechanisms may be important. But these predictions must ultimately be tested against experimental evidence. Researchers must establish whether a proposed target plays a causal role in disease, whether modifying it produces the intended biological effect, and whether that effect can be reproduced in relevant experimental models.
Even a validated target does not guarantee a successful treatment; tumor heterogeneity, compensatory signaling pathways, acquired resistance, and effects on healthy tissues can all undermine an otherwise promising therapeutic approach.
This helps explain why translating scientific discoveries into effective cancer therapies remains so challenging, even as our understanding of cancer biology continues to advance. Historically, only around 3–4% of cancer drug programs entering clinical trials have reached approval.[8]
Many of Oncology’s Greatest Advances Began With the Unexpected
Some of oncology’s most transformative breakthroughs have emerged from unexpected observations, unconventional lines of investigation, or discoveries that challenged established scientific assumptions. History has shown that scientific progress does not always follow a predictable path, and that experimentation can reveal possibilities that researchers had neither anticipated nor initially set out to investigate. This raises an important question about the role AI might play in establishing such discoveries.
For instance, in the 1960s, physicist Barnett Rosenberg was studying the effects of electric currents on bacteria when he noticed that bacterial cells had stopped dividing. Further investigation revealed that the effect was caused not by the current itself, but by platinum compounds released from the electrodes.[9] This unexpected observation ultimately led to cisplatin, a chemotherapy drug that helped transform testicular cancer into one of the most curable solid tumors.
In the early 1980s, Barry Marshall and Robin Warren challenged the widely held belief that bacteria could not survive in the acidic environment of the stomach. Their identification of Helicobacter pylori [8] reshaped understanding of stomach disease and ultimately established a bacterial infection as a major cause of gastric cancer.
A different kind of breakthrough emerged from research into the relationship between the immune system and cancer. For much of the twentieth century, efforts to stimulate antitumor immunity produced inconsistent results until James Allison and colleagues demonstrated that blocking CTLA-4, a protein that helps regulate T-cell activity, could enhance antitumor immune responses and induce tumor rejection in experimental models.[10] This work laid important foundations for modern immune checkpoint inhibition therapy.
Similarly, KRAS, one of the most frequently mutated oncogenes in human cancer, was long considered “undruggable.” Its molecular structure appeared to offer few suitable binding sites for small molecules. This perspective changed when researchers identified a previously unrecognized pocket in the KRAS G12C mutant protein [11], opening the way for the development of targeted inhibitors, including sotorasib, which became the first FDA-approved KRAS-targeted therapy in 2021, initially for patients with advanced non-small cell lung cancer harboring the KRAS G12C mutation.[12]
These discoveries followed very different paths, but each transformed our understanding of cancer or opened new possibilities for treatment. They also illustrate that scientific breakthroughs are not always predictable: sometimes they emerge from questions researchers set out to answer, and sometimes from findings they never expected to encounter.
AI could help accelerate both kinds of discovery, identifying promising directions for investigation while bringing previously overlooked relationships to light. But determining which findings represent genuine advances in our understanding of biology, and which can ultimately translate into effective therapies, remains a question for experimental science.
From Scientific Discovery to Clinical Success
Identifying a promising therapeutic target is only the beginning of cancer drug development. Even when a biological mechanism has been established, a target may be important in one tumor type but not another. Its role may vary between patients or change as the disease progresses. And a treatment that demonstrates promising activity in preclinical models may ultimately fail to deliver sufficient benefit in the clinic.
Scientists must still determine whether a newly identified mechanism genuinely drives disease or simply appears alongside it. They must judge whether a promising target is causally important in patients, not only in cell lines or datasets. They must decide which biological observations are meaningful enough to justify years of development and substantial investment, and how new findings fit within the broader understanding of how tumors grow, adapt and resist treatment.
Above all, they must work out how promising biology can be translated into meaningful benefit for patients. That translation involves choosing the right patient population, the right dose, the right combination and the right moment in treatment. These are questions that data can inform, but rarely resolve on their own.
A Partnership, Not a Replacement
As AI becomes more sophisticated, its role in scientific discovery will continue to expand. Systems are already moving beyond data analysis to generating hypotheses, proposing experiments, and contributing to the interpretation of scientific findings.
What is likely to change is the role of the scientist. As AI takes on more of the work of analyzing data and generating hypotheses, researchers will increasingly focus on the tasks that remain distinctly human: interpreting unexpected findings, challenging existing models, designing experiments that expand what is known, and making sound development decisions under uncertainty. Meaningful innovation will come from pairing the speed of machines with the curiosity and judgment of scientists.
The next major breakthrough in cancer research may well come from AI. The greater achievement will be turning that discovery into a therapy that makes a meaningful difference to patients’ lives.
References
1. [Author(s)]. IAM1363-01: A phase 1/1b study of a selective and brain-penetrant HER2 inhibitor for HER2-driven solid tumors. J Clin Oncol. 2024;42(16_suppl):TPS3186. doi:10.1200/JCO.2024.42.16_suppl.TPS3186
2. Iambic Therapeutics. Iambic shares clinical data from Ph1/1b study of IAM1363, a highly selective type-2 HER2 inhibitor, demonstrating monotherapy activity in heavily pretreated patients with HER2 alterations, at the 2025 ESMO Congress [press release]. October 20, 2025. Available from: https://www.iambic.ai/post/esmo-iam1363
3. Gottweis J, Weng WH, Daryin A, et al. Towards an AI co-scientist. arXiv. 2025. arXiv:2502.18864.
4. Jumper J, Evans R, Pritzel A, et al. Highly accurate protein structure prediction with AlphaFold. Nature. 2021;596(7873):583-589. doi:10.1038/s41586-021-03819-2
5. Kather JN, Pearson AT, Halama N, et al. Deep learning can predict microsatellite instability directly from histology in gastrointestinal cancer. Nat Med. 2019;25(7):1054-1056. doi:10.1038/s41591-019-0462-y
6. Jayatunga MKP, Ayers M, Bruens L, Jayanth D, Meier C. How successful are AI-discovered drugs in clinical trials? A first analysis and emerging lessons. Drug Discov Today. 2024;29(6):104009. doi:10.1016/j.drudis.2024.104009
7. Vogelstein B, Papadopoulos N, Velculescu VE, Zhou S, Diaz LA Jr, Kinzler KW. Cancer genome landscapes. Science. 2013;339(6127):1546-1558. doi:10.1126/science.1235122
8. Wong CH, Siah KW, Lo AW. Estimation of clinical trial success rates and related parameters. Biostatistics. 2019;20(2):273-286. doi:10.1093/biostatistics/kxx069
9. Rosenberg B, Van Camp L, Krigas T. Inhibition of cell division in Escherichia coli by electrolysis products from a platinum electrode. Nature. 1965;205:698-699. doi:10.1038/205698a0
10. Leach DR, Krummel MF, Allison JP. Enhancement of antitumor immunity by CTLA-4 blockade. Science. 1996;271(5256):1734-1736. doi:10.1126/science.271.5256.1734
11. Ostrem JM, Peters U, Sos ML, Wells JA, Shokat KM. K-Ras(G12C) inhibitors allosterically control GTP affinity and effector interactions. Nature. 2013;503(7477):548-551. doi:10.1038/nature12796
12. Skoulidis F, Li BT, Dy GK, et al. Sotorasib for lung cancers with KRAS p.G12C mutation. N Engl J Med. 2021;384(25):2371-2381. doi:10.1056/NEJMoa2103695