April 2026
The Year the
Agents Arrive
An analysis of artificial intelligence trends across ~9,600 accepted abstracts at the American Association for Cancer Research Annual Meeting 2026. What the data reveals about where AI in oncology is heading — and where it's starting to gain real traction.
AI has moved from method
to infrastructure
The accepted abstract program for AACR 2026 suggests that oncology's relationship with AI is undergoing a structural shift — from isolated proof-of-concept models to embedded operational systems.
across the full program
system references
abstracts
validation or deployment
Three years ago, the typical AI abstract at AACR was a proof-of-concept: a single-center study training a convolutional neural network on a few hundred H&E slides to classify tumors. That archetype hasn't disappeared. But at AACR 2026, it will be crowded out by something considerably more ambitious.
The data reveals five structural shifts worth tracking: the emergence of agentic AI systems that execute multi-step workflows, the maturation of LLMs as clinical infrastructure, the rapid buildout of pathology and genomic foundation models, the rise of multimodal integration as a dominant design pattern, and early but real signals of clinical and commercial traction.
This brief is drawn entirely from the ~9,600 abstracts accepted for AACR 2026. All counts are approximate and reflect keyword-based classification, which may over- or under-capture abstracts depending on terminology. Interpretive conclusions represent inferences from abstract-level data and should be treated as directional, not definitive.
Numbers are intentionally rounded to avoid false precision — the goal is to identify the shape of the landscape, not to audit it. Where we draw conclusions beyond what the data explicitly states, we flag it.
The Agentic Moment
AACR 2026 will feature a dedicated session called "Agentic AI in Cancer" — a phrase that would have drawn blank stares at AACR 2023. The premise: AI systems that don't just classify, but act.
The session will host roughly 20 poster presentations and a half-dozen invited technology talks, with speakers from Tempus, Memorial Sloan Kettering, MD Anderson, Siemens Healthineers, Brigham and Women's, and the University of Mainz.
Tempus will present an agentic workflow for automated cancer diagnosis and staging abstraction from unstructured clinical records. MSK will show an MCP-enabled AI agent embedded in its Isabl genomics platform, automating multimodal analysis across petabyte-scale datasets.
MD Andersonwill debut "Charles" — a self-critical, self-correcting agentic AI drug discovery analyst — notable for its emphasis on traceability and hallucination mitigation, framed as a prerequisite for agents to function as credible members of research teams.
AstraZeneca — the most-represented biopharma in the AI corpus with ~10 affiliated abstracts — will present an agentic system for automating RNA-seq analysis pipelines. Other notable entries include Vizgen's AI analysis agent for spatial biology, Keiji AI's adaptive LLM system for clinical trial cohort extraction, and NYU's ImmunoVerse-Chat for immunotherapeutic target discovery.
The field appears to be moving from models to systems — from tools that answer a question to tools that manage a process.
LLMs Move from Novelty
to Infrastructure
A dedicated session — "Large Language Models in the Clinic" — will draw ~30 abstracts and may represent the clearest evidence that LLMs are transitioning from curiosity to operational tooling.
Several abstracts address what might be called the plumbing problems of oncology data: clinical trial matching, pathology report abstraction, EHR data extraction, and patient message triage.
Trial Matching at Scale
Hartford HealthCare will validate a clinical trial knowledge platform across a community cancer network. Tempus presents reasoning-guided retrieval for eligibility matching from clinical notes.
Structured Data from Chaos
Semmelweis University will present CIDER for structured extraction from unstructured medical records. Mayo Clinic demonstrates LLM applications for CAR-T clinical data using Google Cloud.
Patients Prefer the Machine
A University of New Mexico study reports that cancer patients preferred ChatGPT-generated responses to physician-authored ones in a blinded comparison — a culturally revealing finding.
Around ~160 abstracts across the full conference reference LLMs, GPT, ChatGPT, generative AI, or agentic systems — approximately 15% of all AI-related abstracts, forming a substantial and distinct sub-community.
Foundation Models
Stake Their Claim
Approximately 80 abstracts will reference foundation models, vision transformers, self-supervised learning, or pretraining — the architectural paradigm now being applied aggressively to pathology, genomics, and clinical data.
Bioptimus, the Paris-based startup, will deliver three abstracts including two late-breakers: H-optimus-1, a foundation model for computational histopathology, and M-Optimus-1, a multimodal model unifying H&E imaging, bulk RNA-seq, single-cell RNA-seq, and Visium spatial transcriptomics.
Data4Cure will contribute two late-breaking abstracts — one for PDX model selection, another for an RNA foundation model. A UCSD–Lunit collaboration will present a foundation model of cancer genotype that predicts therapeutic response.
In the Digital Pathology minisymposium, MSK will present multimodal modeling of detailed cancer subtypes across >60,000 patientsusing co-registered H&E images and clinical data — a scale that would have been exceptional even two years ago.
A third Bioptimus abstract offers practical benchmarking guidelines for foundation models on spatial transcriptomics data — a meta-contribution that implies the field may be maturing past "build it and publish" into something resembling evaluation rigor.
- H-optimus-1: A foundation model for computational histopathologyBioptimus · Paris, France
- A foundation model of cancer genotype enables precise predictions of therapeutic responseUCSD / Lunit
- AI-guided engineering of pH responsive antibodies enables tumor selective targetingChangping Laboratory · Beijing, China
- Clinical feasibility study of an AI-driven liquid biopsy platform using Raman spectroscopyEXoPERT Corporation · Seoul, Korea
- RNA1-DA: A domain-adaptive RNA foundation model for forward and reverse translationData4Cure · Cambridge, MA
Digital Pathology:
The Beachhead Consolidates
Roughly 100 abstracts will address digital and computational pathology — the most commercially mature AI application domain in oncology. The competitive landscape appears to be intensifying.
Natera will appear in a minisymposium with a deep learning model for virtual genomic profiling in colorectal cancer — presented by a Natera scientist — which may suggest the company is expanding beyond its MRD franchise into computational pathology territory. An Ohio State team will describe deploying AI-driven digital pathology for real-world clinical decision-making in pancreatic cancer, notably using the language of deployment, not validation.
The Multimodal
Integration Play
Around 140 abstracts will reference multimodal approaches — potentially the most important structural trend in the data, implying a shift from single-modality feature extraction to AI as an integrative layer.
A Duke team will present a multimodal framework integrating spatial omics and radiomics for glioblastoma recurrence prediction — bridging the historically separate domains of radiology AI and pathology AI.
A Harvard/Brigham late-breaker describes a healthcare-system-scale multimodal temporal foundation model built on longitudinal EHR data across clinical reports, imaging, and lab tests — one of several abstracts suggesting that the temporal dimension of patient data may be the next frontier for multimodal integration.
Bristol Myers Squibb and Tempuswill co-present an exhibitor spotlight on "scaling multimodal AI and lab-in-the-loop for breakthrough R&D" — phrasing that implies these companies are attempting to operationalize multimodal AI in drug development, not just publish on it.
The shift is from AI as a feature-extraction tool within a single modality to AI as an integrative layer across data types.
Note: Categories overlap — a single abstract may appear in multiple sub-themes. Counts are approximate and not mutually exclusive.
Where Clinical Traction
May Be Emerging
Roughly 225 abstracts in the AI corpus use language suggesting real-world validation, prospective study, regulatory relevance, or clinical deployment.
PanClaudinAI
A multicenter, prospective validation of a deep learning system predicting claudin 18.2 expression from contrast-enhanced CT in pancreatic cancer. Maps directly to the zolbetuximab commercial opportunity.
Guardant's AI Stack
Blood-based integration of epigenomic profiles, TMB, and MSI to predict ICI response — layering AI on top of existing liquid biopsy infrastructure.
Artera Prostate AI
Two abstracts on AI-digital pathology algorithms for prostate cancer prognostication, including external validation in the randomized phase III CHHiP trial.
EXoPERT Raman LBx
A clinical feasibility study of an AI-driven liquid biopsy platform using Raman spectroscopy of plasma extracellular vesicles. Late-breaker status suggests early but compelling data.
Early Detection Wave
~170 AI-related abstracts will touch on screening or early cancer detection — the commercial stakes are enormous and the regulatory pathway remains challenging.
Tempus Deploys Agents
Agentic workflow for automated staging abstraction from unstructured clinical records, plus real-world deployment language in digital pathology — suggesting operational maturity.
The Cancer Map
AI at AACR 2026 won't be evenly distributed. Lung, breast, and colorectal will dominate, but the relative density in liver/HCC and pancreatic cancer may be worth watching.
Liver/HCC and pancreatic highlighted — disproportionate AI interest relative to incidence may reflect diagnostic difficulty and imaging/pathology AI applicability.
Drug Discovery AI:
Pre-Clinical, But Louder
Approximately 50 abstracts will address AI in drug discovery, target identification, or molecular design. The gap between a proposed target and a clinical candidate remains vast — but investment appears to be accelerating.
MD Anderson's "Charles" — a self-critical agentic AI drug discovery analyst — and Bio LIMS's multi-agent system for autonomous CAR-T development will be the headline acts. Owkinwill pitch an "AI agent for pharma R&D" across two exhibitor spotlight sessions.
The agentic framing may accelerate the integration of these tools into actual discovery workflows. A Changping Laboratory late-breaker on AI-guided engineering of pH-responsive antibodies for tumor-selective targeting represents one of the more tangible applications.
The cluster remains largely pre-clinical and methodological. But the involvement of well-funded companies and the growing sophistication of the abstracts — multi-agent orchestration, self-correcting systems, integrated toxicity prediction — suggests the field is building toward something more operationally serious than the typical academic poster.
Who's Playing
The institutional and commercial leaderboard reveals which organizations are investing most visibly in AI-driven oncology research at AACR 2026.
The emerging startup layer — Bioptimus, Keiji AI, Data4Cure, EXoPERT, Certis Oncology, Portrai — is increasingly present, often in late-breaking slots, which may indicate that review committees are recognizing novelty in these approaches.
The Quiet Signal
At least 31 abstracts across the full conference include explicit disclosures that AI was used in the writing process.
The disclosures typically note that AI assisted with language editing, clarity, or summarization, with authors retaining responsibility for content. This is a modest number relative to ~9,600 abstracts, but the fact that it's visible at all — and that AACR appears to be normalizing such disclosure — may signal a broader methodological transparency trend worth monitoring.
The true number is almost certainly higher. These 31 represent the authors who chose to disclose — a self-selected group. As disclosure norms evolve, this figure will likely grow, and with it, new questions about reproducibility, authorship, and the boundary between AI as tool and AI as collaborator in scientific communication.
AI writing disclosure
conference abstracts
abstracts
remains unknown
What It May Mean
AACR 2026 isn't going to produce a single "AI cures cancer" headline. What it will produce may be more significant.
The agentic paradigm, the foundation model buildout, the multimodal integration play, and the quiet spread of LLMs into clinical workflows all point in the same direction — toward AI systems that are embedded in the daily practice of oncology research and care, rather than demonstrated in isolation.
Whether this translates to improved patient outcomes remains an open and genuinely difficult question. The gap between a validated model and a changed clinical decision is real, and the regulatory and reimbursement frameworks for AI-derived diagnostics are still being written.
But the trajectory suggested by this data is clear: the field is no longer asking whether AI belongs in oncology. It's arguing about how fast to let it drive.
The field is no longer asking whether AI belongs in oncology. It's arguing about how fast to let it drive.