Sometimes no better than flipping a coin
At the intersection of artificial intelligence and oncology, researchers at MD Anderson Cancer Center have offered a new answer to one of cancer medicine's most persistent questions: not merely whether immunotherapy exists, but whether it will work for this particular patient. Path-IO, a model trained on the tissue slides already collected in routine care, reads patterns invisible to the human eye and sorts patients into risk groups with a clarity that the current standard test—PD-L1 expression—has rarely achieved. Validated across more than a thousand patients in multiple countries, it arrives as a reminder that precision in medicine is not just a technical ambition but a deeply human one, sparing some patients from treatments that would burden without healing.
- Immunotherapy has transformed lung cancer care, yet for a troubling share of patients it offers nothing—and doctors currently lack a reliable way to know in advance who will benefit.
- The PD-L1 protein test, medicine's standard predictor, performs so poorly in some patient groups that it is statistically indistinguishable from a coin flip, leaving life-altering treatment decisions on shaky ground.
- Path-IO scans routine pathology slides for subtle tumor structures and spatial patterns that no single pathologist could consistently identify, then separates patients into risk groups where the high-risk cohort faces double the mortality of the low-risk cohort.
- Unlike opaque AI systems, Path-IO explains its reasoning in terms of tissue features scientists already recognize, a design choice made deliberately to earn the trust of clinicians who must act on its findings.
- The model has now been validated across 1,000-plus patients at multiple institutions internationally, and a prospective clinical trial is the immediate next step toward real-world deployment.
- The team is already expanding Path-IO's inputs—CT scan data, genomics, clinical history—with a long-term vision of a digital twin that could recommend not just whether to use immunotherapy, but precisely which strategy to pursue.
When Rukhmini Bandyopadhyay presented Path-IO at the 2026 American Association for Cancer Research Annual Meeting, she was addressing a problem that has shadowed immunotherapy's success story: the treatment works brilliantly for some lung cancer patients and not at all for others, and medicine has had no reliable way to tell them apart. The current standard, a test measuring PD-L1 protein expression on tumor cells, sounds precise but frequently performs no better than chance. In some patient cohorts the researchers examined, that was literally true.
Path-IO was built to do better. Rather than hunting for a single molecular marker, it analyzes the pathology slides already collected as part of routine care, searching for complex spatial structures within tumors—niches and tissue arrangements too subtle and numerous for any pathologist to reliably assess by hand. From those patterns, the model assigns patients to risk groups. Those in the high-risk category faced double the likelihood of disease progression or death compared to those in the low-risk group, a gap that dwarfs anything PD-L1 testing has consistently produced.
Crucially, Path-IO is not a black box. It grounds its predictions in tissue features that scientists already understand to be clinically meaningful, making its reasoning legible to the physicians who would need to trust and act on it. That transparency was a deliberate design choice, not an afterthought.
The validation was extensive—more than 1,000 patients across multiple institutions and countries—and the model outperformed the standard test in every setting. The team built it with clinical translation as the explicit goal, and the next step reflects that: a prospective trial testing the model on new patients in real time rather than on historical records.
Looking further ahead, Bandyopadhyay envisions Path-IO absorbing additional data streams—CT imaging, genomics, clinical variables—to sharpen its predictions and eventually recommend not just whether immunotherapy is likely to help, but which specific regimen offers the best odds. The furthest horizon is a digital twin of each patient's disease, a living model updated as treatment unfolds. For now, the immediate promise is more modest and more urgent: giving a lung cancer patient a genuinely reliable answer to the question that shapes everything else—will this treatment actually work for me?
Rukhmini Bandyopadhyay stood at the American Association for Cancer Research Annual Meeting in 2026 with news that could reshape how doctors choose immunotherapy for lung cancer patients. The model she and her team at The University of Texas MD Anderson Cancer Center had built, called Path-IO, could predict which patients would actually benefit from these expensive, demanding treatments—and which ones would not.
Immunotherapy has been one of oncology's great victories. It works by unleashing the immune system to attack cancer cells. But here is the stubborn problem: it does not work for everyone. Some patients see their tumors shrink. Others see no benefit at all. Right now, doctors rely on a test called PD-L1 expression to guess who will respond. The test measures a specific protein on cancer cells. It sounds scientific and precise. In practice, it is barely better than a coin flip. In some of the patient groups the researchers studied, PD-L1 was exactly that unreliable.
Path-IO takes a different approach. Instead of looking for a single protein marker, it examines pathology slides—the tissue samples doctors already collect routinely—and searches for specific structures within tumors called niches, along with other complex patterns that are difficult for human eyes to spot consistently. The model then uses what it finds to sort patients into high-risk and low-risk groups based on their likelihood of disease progression after immunotherapy. A patient in the high-risk group faced double the risk of death or worsening disease compared to someone in the low-risk group.
What makes Path-IO different from other artificial intelligence tools in medicine is that it does not operate as a black box. It does not identify mysterious patterns that no one can explain. Instead, it focuses on tissue features that scientists already know matter for treatment response—features that are simply too subtle and numerous for any single pathologist to reliably spot and measure by hand. This explainability matters enormously for getting doctors to trust and use the tool in real clinical settings.
The researchers tested Path-IO on more than 1,000 patients across multiple institutions and multiple countries. Every time, it outperformed the standard PD-L1 test. The validation was rigorous and the results held up. Bandyopadhyay emphasized that the team designed Path-IO with clinical translation in mind from the start—not as a research curiosity, but as something that could actually change how doctors treat patients.
The next phase is a prospective clinical trial, where the model will be tested on new patients going forward rather than on historical data. The team is already working to include more diverse patient populations in their testing. They are also beginning to feed the model additional information: radiomics data from CT scans, genomic information, and other clinical variables. Each new data stream makes the predictions more accurate.
Bandyopadhyay sees the path ahead extending further still. Eventually, she believes Path-IO could predict not just whether a patient will respond to immunotherapy, but which specific immunotherapy strategy would work best—whether a single checkpoint inhibitor or a combination of agents. Years down the line, the model could become part of a comprehensive digital twin, a virtual replica of each patient's disease that incorporates imaging, genetics, and clinical history all at once.
For now, the significance is this: a lung cancer patient facing the decision of whether to undergo immunotherapy could soon have a much more reliable answer to the question that matters most—will this treatment work for me? That is not a small thing. It means avoiding unnecessary toxicity for those unlikely to benefit, and directing resources toward those most likely to respond.
Citas Notables
Path-IO was designed from the outset for clinical translation—it makes explainable decisions based on known factors and holds up across different datasets.— Rukhmini Bandyopadhyay, postdoctoral fellow, UT MD Anderson Cancer Center
This is the most rigorously validated deep-learning pathomics framework to date, but we're really just getting started.— Rukhmini Bandyopadhyay
La Conversación del Hearth Otra perspectiva de la historia
Why does PD-L1 testing fail so badly if it's the standard of care?
Because cancer is more complicated than any single protein. PD-L1 tells you one thing about the tumor's surface, but it doesn't capture the whole ecosystem—the architecture of the tissue, the way immune cells are positioned, the niches where resistance hides. A single marker was never going to be enough.
So Path-IO is looking at the same slides pathologists already examine, just seeing more in them?
Exactly. The data is already there. Pathologists have been looking at these slides for decades. But there are patterns—spatial relationships, subtle structural features—that are too intricate for any human to reliably quantify across thousands of cases. The AI can do that consistently.
What does it mean that it's "explainable"?
It means when the model says a patient is high-risk, you can point to the tissue features it identified and understand why. It's not magic. It's not a black box spitting out numbers. A clinician can look at the reasoning and decide whether they trust it.
If it's validated in over 1,000 patients already, why isn't it in use now?
Because validation on historical data is not the same as proving it works in real clinical practice. You need a prospective trial—new patients, real decisions, real outcomes. That's the next step. And you need to make sure it works across different populations, different hospitals, different healthcare systems.
What happens if the model says high-risk but the patient wants to try immunotherapy anyway?
That's a conversation between doctor and patient. The model is a tool for better decision-making, not a mandate. But having that information—knowing the odds are against you—changes what you might choose.
Could this eventually replace pathologists?
No. It augments them. A pathologist still needs to prepare the slide, still needs to make the initial diagnosis. Path-IO helps them see patterns they might miss and quantify things that are hard to quantify by eye. It's partnership, not replacement.