Dr. Lu will highlight recent work using specimens from the NRG B-20 trial, in which investigators applied multimodal AI approaches to integrate histopathology features with clinical and molecular data to refine recurrence risk prediction in early-stage breast cancer (Geyer CE, et al. Clin Cancer Res. 2026;32[suppl 4]:RF3-03). “These types of models illustrate how AI can leverage existing clinical trial data sets to generate more nuanced prognostic insights and may ultimately complement or enhance established genomic assays,” she says.
Dr. Lu will also examine the emerging intersection of AI and ctDNA. Applications for AI and ctDNA are still being validated and their use varies by setting, she notes. “But the direction is very clear: ctDNA becomes much more powerful when interpreted in context, and AI is a way to model that context.”
One emerging direction for integrated AI and ctDNA is in early risk and disease characterization, where ctDNA signals alone can be weak. “AI can help integrate ctDNA with imaging, pathology, and clinical variables to improve interpretability and risk modeling,” Dr. Lu says. Other areas in which AI and ctDNA can work together are in assessing neoadjuvant treatment response and measurable residual disease and recurrence risk. And as more longitudinal ctDNA and multimodal data sets mature, AI may help identify patterns that predict who benefits from treatment escalation, de-escalation, or specific targeted strategies.
Growing data show that ctDNA on its own has strong prognostic and predictive value in many settings and that AI can improve classification and risk stratification. “What is still developing is the highest level of evidence showing that AI-guided, ctDNA-informed decisions directly improve patient outcomes in prospective interventional settings,” she says.
Dr. Cooper, in his talk, will share his laboratory’s experience with the prognostic model the group developed for breast cancer, built to gain insight into the tissue microenvironment, including the contributions of noncancer tissue elements to patient prognosis. “We’ll use this application to discuss issues of model interpretability and explainability,” he says. One of his research laboratory’s findings: “There are latent but measurable patterns in breast cancer stroma that are very predictive of breast cancer-specific survival.”
What advice do they have for physicians who might want to wade into this world?
“Demand evidence at the level of intended use,” Dr. Lu says. “A model can have excellent technical accuracy and still fail clinically if it was not tested in your patient population, your workflow, or your setting.”
“You need to see strong validation in real-world data,” Dr. Cooper says. “The availability of public data and software has made it easier to generate results. Do they have access to high-quality data, often clinical trial data, to validate the tool?”
As for collaborating with partners in industry, the best partnerships go beyond a transactional relationship in which the laboratory supplies the vendor with data, he says. “Some partners just want data and are not interested in other input to make the tool useful.”
In Dr. Lu’s view, “A good AI pitch should sound less like a technology demo and more like a clinical implementation plan.” A strong commercial partner should understand clinic and pathology workflow, integration burden, and user experience, as well as the regulatory and compliance issues at hand. “They should be clear about what is research use, what is clinically deployable, and what regulatory pathway applies,” she says.
Is there a role for pathology-oncology collaboration in navigating this new terrain?
“In many ways, AI makes pathology-oncology collaboration even more important than before,” Dr. Lu says. “AI tools fail when they optimize for a narrow technical endpoint, ignoring clinical workflow.” A collaboration between the two disciplines can prevent that, she says.