In rare tumors, oncologists know the usual story. PD-L1 is a weak compass. Tumor mutational burden is often unavailable. The evidence base was never built. So when MD Anderson investigators revisited biopsies from 84 patients in a phase 2 basket trial of pembrolizumab (NCT02721732), they added a second reader.
A deep learning system, Lunit SCOPE IO, scored ordinary hematoxylin and eosin slides for lymphocytes inside the tumor and for remaining tumor content, at baseline and during treatment. Derbala, Stephen, Naing and colleagues reported the results in the Journal for ImmunoTherapy of Cancer.1
Patients whose lymphocytes rose and tumor content fell during treatment had a median overall survival of 42 months, against 10 months for everyone else. The combined signal carried a hazard ratio of 0.36 for both progression free survival and overall survival. In 27 paired samples, the rise in lymphocytes tracked with CD8 T cells sitting closer to tumor cells. That is checkpoint blockade working, seen by an algorithm.1,2
The number that matters most for policy got less attention. Before the first dose, lymphocyte density was a much weaker guide. Across the whole cohort, the baseline association with progression free survival did not hold (hazard ratio 0.62, 95 percent confidence interval 0.37 to 1.06). The strong signal is dynamic. It appears only after the drug has been given.1
For an immunologist this is the right kind of AI. It measures named biology, the hot versus cold tumor distinction, at slide scale in diseases too uncommon for a companion diagnostic. For anyone who works on AI strategy, the paper is a different object. It describes a decision system that could be used, quietly, to decide who counts as a responder. Institutions are not ready for that.
Prediction is not the same act as exclusion
A model that ranks patients by likely benefit can be used in at least four ways, and they are not morally equivalent. It can enrich a trial, enrolling the patients it favors. It can intensify monitoring: treat everyone, watch predicted nonresponders harder. It can sequence therapy, sending predicted nonresponders first to a bispecific, a vaccine, or a cell therapy. Or it can withhold: do not offer the drug, or do not fight the payer, because the slide said no.
The MD Anderson analysis is written for the first two uses, which suits a retrospective study. But once the same scores appear on Monday morning before tumor board, clinics will drift toward sequencing and withholding. That is how scarce, toxic, expensive immune drugs get rationed when the evidence is thin.
Rare tumors are the stress test. There will never be a registrational study of pembrolizumab in 3,000 patients with adrenocortical cancer. An AI score will be the most confident looking number in the room, and confidence is not calibration. A model can perform well in lung adenocarcinoma and badly in a sarcoma subtype barely present in its training data, and in a rare tumor nobody may have enough cases to tell.
In January 2026 the FDA and the European Medicines Agency jointly issued ten Guiding Principles of Good AI Practice in Drug Development.3 They are nonbinding and aimed at sponsors, but their emphasis on a defined context of use and validation proportionate to risk carries a corollary we think is unavoidable: an AI derived subgroup must not silently become an exclusion rule.
The error is asymmetric
False reassurance means the model says responder, the tumor is cold, and the patient loses months to a drug that may cause colitis or pneumonitis. False gatekeeping means the model says nonresponder and a patient who would have done well never gets the infusion. In common cancers with many later lines of therapy, the first error is sometimes recoverable. In rare tumors the second can be the whole story. Governance that collapses both into an area under the curve is statistics with no account of stakes.
The biology sharpens the point. Because the useful signal was change during treatment, a responsible deployment looks like treat, rebiopsy, update the bet. That is not free. A biopsy during treatment is invasive, is not routine care outside trials, and many rare tumors are hard to reach. Centers will be tempted to score the diagnostic block alone. A static veto before the first dose is cheaper for the health system, worse for the biology, and the use the paper supports least.1
Who overrules the model?
Most hospital AI committees were built for radiology worklists and sepsis alerts. They ask about contracts, data security, and device clearance: the wrong questions for a score that can reclassify a person as ineligible for what may be their best remaining option. An override rule has to exist before the first score is posted to the record. A workable minimum:
A named owner. A physician of record, not the model, signs any decision to withhold or delay a checkpoint inhibitor. The score is an input, like a PD-L1 percentage, not a protocol.
A required dissent channel. When the oncologist and the model disagree, the case goes to a standing immunotherapy review, not an IT ticket. Minutes record whether the override was for biology (mismatch repair deficiency, viral etiology, lymphocytes the algorithm undercounted) or for values (the patient accepts toxicity for a small chance).
Separate thresholds for research and care. A cutoff that makes a trial cleaner is not the cutoff that makes treatment in the final line ethical.
Performance by slice. Report accuracy in the histology in front of you, not the pooled statistic. If the slice is too small to know, say so alongside the score.
A drift watch. Stains, scanners, and case mix change, and an unmonitored model rots into a confident antique. FDA’s draft lifecycle guidance for AI-enabled devices expects monitoring after deployment,4 and a recent Nature Medicine Comment on scaling one screening model to over a million patients in three countries shows it must be made to work again in each new setting.5 Most academic deployments do not budget for that.
None of this requires a new statute. It requires a cancer center to admit that decision support is a political category, not a software one.
What this is not
This is not an argument against the MD Anderson result. Measuring lymphocytes and tumor content on routine slides, in diseases that never got a companion diagnostic, is exactly the machine vision immunologists should want. The investigators themselves say it needs validation in larger populations before it guides treatment.2
It is an argument against a familiar failure in clinical AI. A paper demonstrates association with survival. A vendor demonstrates a dashboard. Nobody demonstrates who is allowed to say the dashboard is wrong. Rare tumor immunotherapy is where that failure will hurt first, because these patients have the fewest alternative numbers to put on the table.
The model can say no. A person still has to be able to say yes, out loud, on the record, and take the consequences. That is the governance specification the science now requires.
Author Bio
Matt Hasan, PhD, is CEO of aiRESULTS and founder of the AI Humanist Movement. He advises Fortune 50 organizations on AI strategy and governance and previously held leadership roles at Deloitte, IBM Global Business Services, and Capgemini.
Farah Hasan, PhD, is a scientist in immunotherapy development with experience at BioNTech, NYU Langone Health, and the Icahn School of Medicine at Mount Sinai.
