Foundation Models and AI Agents Are Reshaping Cancer Research
Researchers at the American Association for Cancer Research 2026 annual meeting presented advances in AI tools that move beyond data classification to tackle complex problems in oncology and drug discovery. Foundation models-AI systems trained on massive datasets-can now predict protein structure, mutations, and gene regulation, opening new pathways for precision cancer treatment.
Single-Cell Models Predict Gene Expression Changes
Bo Wang, a computer scientist at the University of Toronto, built scGPT, a foundation model trained on over 33 million cells. The model performs cell-type annotation, genetic perturbation prediction, and gene network inference-tasks that previously required manual analysis.
Wang's team extended this work with X-Cell, a foundation model that predicts how gene expression changes when cells are perturbed. The model was trained on nearly 26 million perturbed single-cell transcriptomes across seven screens, representing what Wang called "probably the largest high-quality genome-wide perturbation datasets out there."
The researchers also developed BioReason, which combines a foundation model for DNA sequences with a reasoning language model to interpret biological sequences. When tested on disease pathway predictions, the model improved at understanding biology with each iteration.
For protein function annotation, Wang's team created BioReason-Pro. When evaluated against 27 protein experts and biologists, the model generated predictions faster than human annotators and produced higher-quality results 80 percent of the time.
AI Agents Streamline Scientific Workflows
Large language models like ChatGPT and Claude have limited usefulness in biomedicine because they rely on user prompts and lack autonomy. Jure Leskovec, a computer scientist at Stanford University, is developing AI agents that function as collaborative partners in research.
In summer 2025, Leskovec's team released Biomni, an open-source biomedical agent that integrates data upload, visualization, and scientific reasoning into a single interface. Scientists remain "at the steering wheel" while the AI handles routine tasks like scanning literature, identifying relevant databases, and designing experimental protocols.
Biomni includes safeguards against hallucination-a known limitation of large language models-by asking clarifying questions, supporting brainstorming, and providing transparent self-evaluations. The system adapts to researcher preferences over time and works alongside robotic wet labs for automated experimentation.
Clinical AI Tools Face Adoption Barriers
Faisal Mahmood at Harvard Medical School develops foundation models that integrate histopathology, genomics, radiology, and clinical data for cancer diagnosis. His team created TITAN, which represents entire pathology slide images as single data points, and THREADS, which links histological images to genomic data.
These models typically require 10,000 whole slide images to reach clinical performance. By condensing patient data into single vectors, Mahmood's approach enables downstream applications for diagnosis, prognosis, and treatment response prediction.
Translating these advances to clinical practice requires overcoming specific hurdles. Suchi Saria, a computer scientist at Johns Hopkins University, said current clinical AI tools are not built for real-time decision support. Existing sepsis alerting systems, for example, often miss early signs or flag the condition after treatment has already begun, resulting in low adoption among clinicians.
Saria implemented an AI system at Cleveland Clinic that routes each patient sample to a specialized model best suited to interpret the data. The system improved sepsis detection by nearly 60 percent compared to the existing tool, enabling faster treatment initiation, shorter hospital stays, and reduced mortality.
Successful clinical AI requires handling patient heterogeneity-ensuring data encompasses symptom variations-and delivering results in real time. Saria emphasized that deploying these tools to drive clinical action is what ultimately affects patient outcomes.
Researchers building AI for cancer biology are now working across the full pipeline from early hypothesis generation to precision oncology. AI for Science & Research training can help professionals understand how these tools integrate into existing workflows.
Your membership also unlocks:
