The problem first walked onto the oncology construction site in 1891, when William Coley tried to jolt tumors with bacterial toxins; after more than 1,000 treated patients and a century-plus of vaccines, cytokines, checkpoints, cell therapies, and biomarker panels that often missed the mark, immunotherapy still looks like a brilliant building with several staircases to nowhere.
OpenIO, short for Open Immune Oncology, is a proposal to redraw the plans. In Wu et al., 2026, the authors argue that immunotherapy should move from empirical screening - try a thing, squint at survival curves, repeat - toward AI-native engineering. That means foundation models trained on omics, immune context, clinical outcomes, and experimental feedback, then used to propose and refine therapies. Not "ChatGPT, prescribe pembrolizumab," which would be malpractice wearing a novelty hat. More like a computational design studio where every wall has to survive biological inspection.
The Building Has Too Many Materials
Immunotherapy sounds simple: help the immune system recognize and attack cancer. The facade is elegant. The interior plumbing is chaos.
A tumor is not just a lump of bad cells. It sits inside a tumor microenvironment filled with immune cells, fibroblasts, blood vessels, signaling molecules, metabolic weirdness, and escape routes. Checkpoint inhibitors can release immune brakes such as PD-1, PD-L1, and CTLA-4, but many patients do not respond, and some tumors politely install a second door behind the first one.
That is where omics enters: genomics, transcriptomics, proteomics, metabolomics, spatial omics - the entire building survey, from foundation bolts to suspicious stains in the basement. OpenIO wants to integrate those layers instead of treating them as separate annexes designed by committees that never met.
Foundation Models As Flying Buttresses
The OpenIO idea leans on two load-bearing concepts: biological scaling laws and foundation models. Scaling laws are the tempting observation that, in AI, larger and better datasets can produce more capable models. Biology adds a catch: cells are not words. They do not autocomplete. They sulk, mutate, migrate, and occasionally make your model look like it was trained on soup.
Still, the neighboring architecture is real. scGPT trained a generative transformer across more than 33 million single-cell profiles. GET modeled transcription across human cell types. MUSK joined pathology images and text for precision oncology tasks, including prognosis and immunotherapy-response prediction. These are early structural systems, not finished cathedrals, but the sight lines are improving.
For humans trying to sketch all these loops - tumor, immune cell, antigen, treatment, response, relapse - a visual reasoning tool like mapb2.io is less decoration than scaffolding.
The Pretty Rendering Is Not The Building
OpenIO is most interesting because it treats AI as part of the experiment cycle, not just a dashboard bolted onto the clinic. A model might suggest a neoantigen target, predict which immune cells matter, design a combination therapy, or flag why a treatment failed. Then wet-lab and clinical data come back, and the system learns. In architectural terms: the blueprint argues with the building inspector, then updates the blueprint.
But the paper also leaves necessary negative space. Generalization remains ugly. A model that behaves nicely in one dataset may collapse in another because batch effects, population differences, tissue handling, and clinical noise all have the manners of a drunk contractor. Recent benchmarks have shown that single-cell foundation models can underperform simpler baselines in zero-shot settings, and perturbation-prediction work keeps reminding everyone that evaluation metrics can flatter the wrong thing.
That is not a reason to dismiss OpenIO. It is a reason to demand open benchmarks, shared data standards, transparent validation, and models that can explain their load distribution before anyone lets them carry clinical weight.
Why This Blueprint Matters
If OpenIO becomes reproducible and expands beyond a framework, it could shift immunotherapy research from artisanal guessing toward rational design. Fewer dead-end screens. Better patient stratification. More testable hypotheses. Maybe even treatments built for the actual immune architecture of a patient’s tumor, rather than the oncology equivalent of "standard two-car garage, hope it fits."
The appeal is not hype. It is form meeting function. Immunotherapy already changed cancer care for some patients. OpenIO asks whether AI can help make that building less exclusive, less mysterious, and less full of locked doors.
References
- Wu Y. et al. "OpenIO: An open framework for AI-native immunotherapy." Cancer Cell (2026). DOI: 10.1016/j.ccell.2026.06.002
- Cui H. et al. "scGPT: toward building a foundation model for single-cell multi-omics using generative AI." Nature Methods 21, 1470-1480 (2024). DOI: 10.1038/s41592-024-02201-0
- Fu X. et al. "A foundation model of transcription across human cell types." Nature 637, 965-973 (2025). DOI: 10.1038/s41586-024-08391-z
- Xiang J. et al. "A vision-language foundation model for precision oncology." Nature 638, 769-778 (2025). DOI: 10.1038/s41586-024-08378-w
- Kedzierska K.Z. et al. "Zero-shot evaluation reveals limitations of single-cell foundation models." Genome Biology 26, 101 (2025). DOI: 10.1186/s13059-025-03574-x
Disclaimer: This blog post is a simplified summary of published research for educational purposes. The accompanying illustration is artistic and does not depict actual model architectures, data, or experimental results. Always refer to the original paper for technical details.