A single stained tissue slide, the kind pathologists have examined under microscopes for more than a century, can now be fed into an AI model that simultaneously predicts a tumor’s cancer subtype, its underlying genetic mutations and its likely survival trajectory — across 32 different solid cancers at once. The model, described in a study published this month in The American Journal of Pathology, marks one of the broadest attempts yet to connect routine diagnostic imaging directly to molecular oncology without requiring a separate, expensive genomic sequencing test for every patient.
Most AI pathology tools to date have been narrow by design: a model trained to spot lung adenocarcinoma, another trained on breast cancer receptor status, a third for colorectal polyps. The new model breaks from that pattern, training on whole-slide histopathology images spanning dozens of tumor types simultaneously, then testing whether a single shared architecture could pull out cancer subtype and mutation signals that are normally invisible to the human eye.
Why Genotype Usually Waits for a Separate Test
In current clinical practice, a pathologist reviews a biopsy slide to confirm cancer is present and identify its general type. If more precise treatment decisions are needed — for instance, whether a lung tumor carries an EGFR mutation that would make it responsive to a targeted therapy — the sample is typically sent for separate molecular sequencing, a process that can take one to three weeks and cost hundreds to thousands of dollars per panel, depending on the assay and institution. That delay matters enormously for aggressive cancers where treatment decisions are time-sensitive.
The researchers behind the new model argue that many mutations leave subtle, indirect fingerprints on tissue architecture and cell morphology that a sufficiently large AI system, trained across enough cases, can learn to detect directly from the image — effectively inferring genotype from phenotype at a fraction of the turnaround time.
How the Model Was Built and Tested
The team trained the algorithm on a large multi-institutional dataset of histopathology slides paired with confirmed genomic and outcome data spanning the 32 cancer types, then validated its predictions against independent held-out cases where the true mutations and survival outcomes were already known. The published results show the model performing with meaningful accuracy in flagging key driver mutations and biomarkers across most of the cancer types tested, though the researchers are careful to note that performance varies by tumor type and that findings are not yet intended to replace confirmatory genomic testing before treatment decisions.
Part of a Broader Wave in Computational Pathology
The study lands amid a string of similar advances in 2026. Researchers publishing in Nature Cancer earlier this year described PRET, a plug-and-play AI pathology system that recognized 18 cancer types from tissue slides without needing institution-specific retraining, reporting notably strong accuracy in colorectal cancer screening. Separately, other 2026 research has shown AI systems outperforming physicians at summarizing complex pathology reports for patients. Together, these efforts reflect a shift in computational pathology away from single-purpose diagnostic tools and toward broad, general-purpose models that behave more like foundation models than narrow classifiers.
The Skeptics’ Case
Not everyone in oncology is convinced multi-cancer prediction models are ready to influence care. Pathologists and oncologists who have reviewed similar tools caution that predicting a mutation’s presence from morphology is fundamentally probabilistic — a model might flag a likely mutation with reasonable confidence, but a false negative could mean a patient misses out on a targeted therapy, while a false positive could trigger unnecessary treatment. Regulatory bodies, including the FDA, have generally required that AI-inferred genomic predictions be confirmed by orthogonal molecular testing before they inform prescribing decisions, and there is no indication that will change soon. Critics also point out that training data drawn heavily from certain institutions can embed biases that hurt performance in underrepresented patient populations, a recurring concern across AI pathology research broadly.
Where This Could Fit Into Care
Proponents see the more realistic near-term use case as triage rather than replacement: a same-day AI read that flags a high probability of a specific mutation could help oncologists prioritize which patients need expedited sequencing, rather than waiting in a generic queue. In resource-limited settings where genomic sequencing is expensive or slow to access, an image-based pre-screen could meaningfully shorten the path to a treatment decision even if a confirmatory test is still required.
What’s Next
The researchers say their next steps involve expanding validation to more diverse patient cohorts and additional cancer types, along with prospective studies that track whether AI-flagged mutation predictions actually change clinical decision-making and outcomes in real time — not just in retrospective analysis. If those trials pan out, the tool could eventually be positioned for FDA review as a triage aid rather than a diagnostic replacement, following the same cautious, human-in-the-loop path that most AI pathology and radiology tools have taken through the agency’s clearance process so far in 2026.
