AI Finds New Applications in Oncologic Surgery, Drug and Biomarker Screening, and Clinical Trial Design

AI is rapidly moving beyond the realm of future promise and into everyday oncology practice. From enhancing surgical precision and diagnostic accuracy to refining biomarker-driven treatment selection and streamlining clinical trial enrollment, AI-powered tools are beginning to fundamentally reshape how cancer care is delivered.

Although these technologies are not replacing clinical judgment, they are providing oncologists with increasingly sophisticated decision support that has the potential to improve efficiency, personalize treatment, and ultimately enhance patient outcomes. With a growing body of evidence showing the benefit of utilizing these tools, experts across oncology are weighing in on where AI is already making an impact and where its greatest opportunities still lie.

“[AI] is not acting on its own. [We] are not acting [on a] piece of information that was given to us [via] AI and blindly following it to its end,” James T. McCormick, DO, FACS, FASCRS, said in an interview with OncLive®. “[AI] is just giving highly trained and highly functioning [oncologists] an additional piece of information that we can use to make critical decisions, but we're not making the decisions solely based upon [its output].”

McCormick is a professor of surgery at Drexel University College of Medicine as well as the chairperson of the Colon/Rectal Cancer Program at Allegheny Health Network Cancer Institute, in Pittsburgh, Pennsylvania.

How can AI-based tools assist surgical oncologists?

McCormick noted that AI-assisted systems such as the Eureka AI Platform have shown promise in aiding oncologists in more precisely administering surgery. Eureka uses a deep learning model to help surgeons recognize anatomic features and define a safe dissection plane.¹ When Eureka is connected to a laparoscopic system, it is able to highlight loose connective tissue fibers in the dissected layers, nerves, pancreas, and ureters in the surgical field and display them on an attached monitor in real time throughout the course of the procedure.

“The primary focus right now is training,” McCormick explained. “We’re training the AI to identify structures that we know to be the appropriate structures for cutting or for saving, and so when we go into the training field of trying to get general surgery residents or colorectal surgery residents through the process of becoming surgeons that can identify these structures independently, the AI can augment reality and help them to see things in the background that they may not appreciate at first glance.”

In a prospective study published in Surgical Endoscopy, investigators connected the Eureka system to the laparoscopic system side by side, and it highlighted nerves on the monitor during surgery. The study authors examined the rate of failure to recognize nerves by trainee surgeons over the course of 101 scenarios after it was recognized intraoperatively by a supervising surgeon. They also evaluated the system’s recognition assistance rate, which was defined as the frequency of nerve recognition by the trainee physicians viewing the Eureka monitor when recognition was not possible.

Findings from the study revealed that the nerve recognition failure rates with the Eureka system were 43.6% and 26.7% for right hypogastric nerves during sigmoid colon mobilization and left hypogastric nerves during dissection of the dorsal rectum, respectively; the recognition rates were 31.7% and 43.6% for the right and left lumbar splanchnic nerves, respectively. The respective recognition assistance rates were 43.2%, 48.1%, 90.6%, and 88.6%.

“At my level of experience, I don’t necessarily need AI to [assist me during surgery],” McCormick explained. “My hope is that at some point the AI [will be able to] give me a warning ahead of time, like a blind spot indicator on your car. It’s creating additional safety elements, both for experienced surgeons and for less experienced surgeons who really do need help identifying [certain] structures.”

How is AI being used in drug screening, biomarker selection, and clinical trial protocols?

Beyond enhancing the delivery of surgical procedures, AI could also aid in optimizing treatment selection for patients, novel biomarker identification, and creating adaptive clinical trial designs. For example, in a study published in Annals of Oncology, Gerhardt Attard MD, PhD, FRCP, and his coauthors examined a digital pathology multimodal AI model that had been previously validated as a prognostic biomarker for the prediction of benefit from abiraterone acetate (Zytiga) in patients with nonmetastatic prostate cancer. Investigators generated scores for patients enrolled in 2 sequential abiraterone acetate trials in the STAMPEDE platform protocol (NCT00268476) using digital pathology images, prostate-specific antigen (PSA) levels, tumor stage, and age.

“The opportunities are large, and this is one of the first studies to do this in prostate cancer,” Attard, a John Black Charitable Foundation Endowed Chair in Urological Cancer Research and a professor of medical oncology in the Research Department of Oncology at University College London in England, said in an interview with OncLive. “Clearly, it introduces so many opportunities across other cancers. Over the next five years, these tests will become routine, and [we] will have the score and recommendation available when [we] treat patients.”

The study authors applied the 75^th percentile threshold to classify patients as very high-risk or standard high-risk per the multimodal AI model. The primary end point was metastasis-free survival (MFS).

The study included 1137 patients who were randomly assigned to receive long-term androgen deprivation therapy (ADT; n = 583) or long-term ADT in combination with abiraterone (n = 554). Patients who were determined to have very high-risk disease per the model (n = 268) experienced a significant MFS improvement with the addition of abiraterone (HR, 0.47; 95% CI, 0.31-0.70); the 5-year MFS rates with long-term ADT and long-term ADT plus abiraterone were 62% (95% CI, 53%-70%) and 81% (95% CI, 74%-87%), respectively. Moreover, the benefit with abiraterone acetate was more limited in patients with standard high-risk disease per the model (n = 869; HR, 0.83; 95% CI 0.63-1.09). The 5-year MFS rates in the investigational and control arms in this subgroup were 82% (95% CI, 78%-85%) and 84% (95% CI, 80%-87%), respectively.

“This test can be easily implemented into clinical practice, because all the features and the slides that we use are the standard of care [SOC],” Attard said. “There's no additional cost or technology to obtain those. The algorithm itself can be run in the cloud or locally at institutions, so this can be very rapidly and widely implemented.”

AI is also emerging as a valuable tool for biomarker identification and tumor profiling. “Predicting responses to certain drugs and helping to select the right drug for the patient—there’s huge opportunity [there], and lots of us are working on these different [AI] applications. When we pair a patient’s molecular results with their clinical history, we start to get some really useful data sets to build models,” James Hamrick, MD, MPH, the chairman of the Caris Precision Oncology Alliance, said in an interview with OncLive.

Another AI model, Genomic Probability Score AI (GPSai), aims to predict tumor tissue of origin in cancers of unknown primary and identifies possible misdiagnoses for additional workup during routine molecular testing.³ The model was trained on whole exome and whole transcriptome data from 201,612 cases that were submitted for tumor profiling and also underwent retrospective and prospective validations. Investigators then examined the clinical utility of GPSai over 8 months of live testing and physician surveys.

Findings from the evaluation period showed that GPSai changed the diagnosis of 0.88% of all profiled cased (n = 704); these changes were supported by orthogonal evidence including imaging, immunohistochemistry data, mutational signatures, hallmark fusions, or viral reads. Changes in diagnosis led to changes in targeted therapy eligibility based on level 1 clinical evidence in 86.1% of cases. In terms of physician responses (n = 97), most respondents indicated that they accepted the findings from GPSai and 53.6% said that the findings led to a change in treatment plan.

“Sometimes patients have more than one cancer; I’ve had multiple patients in my clinic with 2 or even 3 cancers,” Vitor Pastorini, MD, an oncologist at Florida Oncology & Hematology in Cape Coral and Fort Myers, said in an interview with OncLive. “We try to figure out where the metastasis is coming from, and GPSai helps a lot in those cases. [This tool] gives us another layer of certainty and security that we have the right diagnosis.”

AI is also being rapidly integrated into clinical trial design and enrollment, showing the potential to help oncologists more quickly navigate the complex enrollment criteria that accompany man studies, according to Pastorini. “A computer is faster than any of us, no matter how intelligent we are. The AI can quickly scan through the patient's characteristics and the inclusion and exclusion criteria of the trial and then [determine] whether the patient is eligible or not. It's much better for me as a clinician when I can walk into a patient's room and say, ‘You have 2 or 3 clinical trials that are available, or we can [proceed with] SOC [therapy],’” he said.

Findings from a review article published in ESMO Real World Data and Digital Oncology revealed that the large language model AutoTrial was able to assist clinical trial design by creating eligibility criteria, incorporating scalable knowledge through in-context learning, and providing explicit reasoning chains for better interpretability.⁴ Other AI models have been used to predict the success of clinical trials prior to launch using protocol design, eligibility criteria, molecular targets, and historical context to predict outcomes from early termination to approval. The AI model PrOCTOR showed the ability to predict agent toxicity via chemical and target-based features and was able to outperform traditional drug-likeness rules (AUC, 0.83) and correlate with the severity of adverse effects.

Data from a meta analysis highlighted in the narrative review revealed that AI tools reached greater than 80% accuracy, sensitivity, and specificity based on 10 studies, using 19 datasets that included more than 50,000 patients. However, the authors noted that AI’s effect on patient recruitment, alignment with projected accrual time, and impact on the physician’s workload remain understudied. Further validation is required for broad generalizability across sexes, ages, ethnicities, and various electronic health record systems.

“[AI] is another tool in our armamentarium,” Hamrick said. “It's becoming more and more embedded in clinical infrastructure, [but ultimately] it's a tool that helps us to get through the day, make the right decisions, and do all the other work that needs to happen operationally to deliver world-class care.”

References

1. Ryu S, Imaizumi Y, Goto K, et al. Artificial intelligence-enhanced navigation for nerve recognition and surgical education in laparoscopic colorectal surgery. Surg Endosc. 2025;39(2):1388-1396. doi:10.1007/s00464-024-11489-0
2. Parker CTA, Huang H-C, Grist E, et al. Multimodal artificial intelligence prediction of abiraterone efficacy in two STAMPEDE phase 3 trials of non-metastatic very high-risk prostate cancer. Ann Oncol. Published online June 5, 2026. doi:10.1016/j.annonc.2026.05.708
3. Ghani H, Helmstetter A, Ribeiro JR, et al. GPSai: a clinically validated AI tool for tissue of origin prediction during routine tumor Profiling. Cancer Res Commun. 2025;5(9):1477-1489. doi:10.1158/2767-9764.CRC-25-0171
4. Knapen DG, van Kruchten M, de Groot DJA, Broekman KE, Fehrmann RSN. Artificial intelligence for clinical trial design, conduct, and analysis: a narrative review. ESMO Real World Data Digit Oncol. 2026;11:100682. doi:10.1016/j.esmorw.2026.100682