From foundation models to AI scientists: Inside oncology's AI shift at AACR 2026

“I think ultimately the whole conference should probably be about AI”, in the words of Pr. Jakob Kather, Technical University Dresden, at this year’s AACR Annual Meeting 2026. It’s a bold statement but one that captures a broader shift: there is no moving forward in cancer research without AI. We are well past the hype cycle; AI is now firmly embedded in the healthcare ecosystem.

Under the bright skies and palm trees of San Diego, more than 22,000 participants gathered for four days of scientific exchange and discovery. For the first time in its history, the conference featured a dedicated AI plenary session alongside multiple agentic AI-focused talks and poster sessions: a clear signal of how quickly the field is evolving. If last year marked the rise of AI at AACR, this year confirmed its central role.

Setting the tone for the week, AACR President Dr. Lillian Siu, University of Toronto, opened the conference on a tone of optimism. With cancer mortality at an all-time low after three decades of decline, and over 20 new anticancer therapeutics approved in the past year alone, progress is tangible. “Our purpose is clear, our mission is urgent”, she emphasized, adding that the discoveries shared here “will travel far beyond these halls”.

From models to foundation systems

As Dr. Kather outlined, AI in oncology has evolved through three distinct waves. The early 2000s were defined by classical machine learning, largely limited to structured, tabular data. The 2010s brought deep learning, enabling major advances in imaging and pathology, but at the cost of massive data requirements. Today, the field has entered the era of foundation models: large, reusable systems trained once and deployed across multiple downstream tasks.

These models are helping address one of the field’s core challenges: extracting meaningful biological insight from increasingly complex and multimodal data, enabling a shift from detection toward molecular prediction, biomarker discovery, and patient stratification directly from routine clinical data.

This evolution was further highlighted by Dr. Faisal Mahmood, Harvard Medical School, who introduced Apollo, a multimodal temporal foundation model for virtual patient representations. By integrating data across modalities and time, Apollo enables earlier risk prediction, treatment response modeling, clinical trial matching, and biomarker discovery, pointing towards a future where the medical record itself becomes a computable representation of patient biology.

The rise of AI agents

If foundation models are the engine, AI agents are the next frontier. We are moving from passive tools to active collaborators: systems capable of planning, executing, and iterating on complex scientific tasks.

Agentic systems exemplify this shift. Rather than simply answering questions, they operate within environments: accessing databases, running analyses, designing experiments, and refining hypotheses. As Dr. Jure Leskovec, Stanford University, emphasized, “for the first time, from question to results, all steps are captured fully”.

The implication is significant. AI is no longer just accelerating isolated tasks, it is beginning to scale the entire scientific process. This is particularly impactful in clinical development, where complex, multi-step workflows can be streamlined by AI agents, from evidence generation and trial design to execution and monitoring, while still requiring careful oversight, governance, and trust to ensure safe deployment.

Bridging the gap to real-world care

Despite this progress, the main bottleneck is no longer model performance. Across sessions, a consistent theme emerged: the challenge now lies in implementation, translating these advances into real-world clinical impact remains difficult.

Key barriers include data fragmentation, lack of integration into clinical workflows, and the difficulty of generating reliable, actionable insights in real time. Addressing these challenges requires not just better models, but robust systems that are designed for real-world deployment.

This gap, between where AI models are developed and where medicine actually happens, is critical to close. At Owkin, this is a central focus. Our Phase 1a INVOKE trial with OKN4395 integrates a continuous feedback loop between clinical data and AI models, enabling real-world data to refine predictions over time. This approach grounds AI in clinical reality and creates a cycle where models and patient outcomes continuously inform each other.

AI meets next-generation therapeutics

AI will almost certainly play a foundational role in advancing the therapeutic innovations showcased at AACR, from target selection in next-gen CAR-T to payload design and bispecific formats in ADCs. 

In the opening plenary, Dr. Carl June, University of Pennsylvania, reviewed the impact of immunotherapies, which have reshaped outcomes across multiple cancer types and now make a significant proportion of previously deadly cancers treatable. Efforts are now focused on extending these successes to more challenging indications, particularly solid tumors, with next-generation CAR T approaches exploring more advanced engineering strategies.

Next-generation antibody-drug conjugates (ADCs) were also widely discussed, with innovation focused on improved payloads, bispecific formats, and more advanced delivery strategies to increase efficacy while reducing toxicity and overcoming resistance. This is another area that Owkin’s R&D is actively pursuing, exploring how AI and spatial biology can guide smarter ADC design, using the tumor microenvironment to predict response, anticipate resistance, and inform payload strategy.

Together, these developments reflect a broader shift toward more sophisticated therapeutic design, grounded in deeper biological understanding. Agentic AI will play an increasing role in this evolution, supporting hypothesis generation, target selection, trial design, and the translation of biological insight into clinical applications.

Why domain-specific AI matters

As AI systems move deeper into biology, a clear distinction is emerging: general-purpose models can orchestrate workflows and summarize information, but they lack the depth required for complex biomedical reasoning. Across AACR sessions, two questions kept resurfacing: how do we move from foundation models to agents that actually function in clinical environments, and how do we ground them in real biology rather than retrieved text?

This is the gap Owkin’s AI scientist, K Pro, is built to close. Introduced at AACR last year as our first agentic research co-pilot, K Pro now integrates multimodal patient data from our network of 100+ partner hospitals with a library of expert-defined workflows, what we call “skills”.  Each skill is a reproducible, executable analysis or procedure built from a decade of biopharma experience, such as multimodal target characterization or spatial differential expression analysis. Crucially, this library isn't static. It grows through ongoing collaboration with the scientists shaping the field. At AACR this year, our team spent the week with world-leading oncologists and translational researchers, showing them what we've been building and, just as importantly, learning how they reason through hard cancer questions. Those conversations can become new skills, encoding expert thinking into workflows K Pro can run reproducibly, at scale.

The result is an agentic system that doesn’t just generate plausible answers, but it runs the analysis, shows its work, and produces results that hold up to scientific scrutiny. That combination of data, tools, and domain expertise is what turns AI from a productivity layer into a genuine collaborator in oncology research.

Looking ahead

AACR 2026 made one thing clear: the future of cancer research is inseparable from AI. From foundation models to agentic systems, the field is converging towards a new paradigm, one where human and machine intelligence work hand in hand.

As partnerships between AI and biopharma continue to deepen, and as systems like K Pro bring cutting-edge models into real-world settings, the potential to accelerate progress in oncology is significant.

If this year confirmed AI’s central role, the next will likely show how far it can go.