Perspective: Where a Biomarker Is Located May Matter as Much as How Much Is Present
The finding was not simply “more CD8 is better”
Ampullary adenocarcinoma is uncommon, and much of its treatment evidence is necessarily drawn from related gastrointestinal cancers. Studies in rare tumors therefore face a familiar tension. Available cohorts may be small, retrospective, and assembled across long treatment eras, but they can still contain biologically and clinically useful information if the analysis preserves the dimensions of the data that matter.
In our recently accepted study in Cancer Research Communications, we examined mismatch repair status, PD-L1 expression, and compartment-specific CD3 and CD8 tumor-infiltrating lymphocytes in 99 resected ampullary adenocarcinomas.
Mismatch repair deficiency was relatively common. PD-L1 expression was also frequent. Neither, however, was associated with disease-specific survival in this cohort.
The more interesting observation concerned where CD8-positive T cells were located.
CD8-positive T cells within the tumor epithelium showed the clearest association with disease-specific survival. Stromal CD8 did not show the same pattern. After adjustment for age, histologic subtype, pathologic T stage, nodal status, and adjuvant chemotherapy, intraepithelial CD8 presence was associated with lower disease-specific hazard. Continuous intraepithelial CD8 counts pointed in the same direction.
The result was not definitive. The cohort was small, tissue microarrays sample only part of a tumor, and likelihood-ratio and Wald inference did not agree completely. We therefore described the association as borderline, method-dependent, and requiring external validation.
But the spatial finding itself raises a broader question.
Averages can be right and still hide something important
Evidence generation necessarily involves compression.
Patients become cohorts. Repeated observations become averages. Biological measurements become positive or negative categories. Heterogeneous populations are summarized by a treatment effect. Complex tumors are represented by biomarker scores.
None of that is inherently wrong.
The problem arises when a summary measure is assumed to preserve every dimension relevant to the question being asked.
Kent, Altman and colleagues made a related point in the clinical-trial setting. A valid average treatment effect can still be poorly representative of the effects experienced by many individual patients when baseline risk and absolute treatment benefit vary substantially across the trial population. Their argument was not that averages are misleading by definition. It was that heterogeneity can contain decision-relevant information that disappears when attention stops at the population mean.
That distinction is important:
The problem is not averaging. The problem is assuming that the average preserves the structure relevant to the decision.
The same principle applies far beyond treatment effects.
Biology also has structure
This is not the first time we have encountered that problem in tumor immunology.
In work published in Modern Pathology in 2009, Blaise Clarke and colleagues, including me, evaluated intraepithelial T-cell infiltration in ovarian carcinoma. In a retrospective cohort of 500 tumors, intraepithelial CD8-positive T cells were associated with improved disease-specific survival overall. But the relationship was not uniform across ovarian cancer types. It was particularly strong in serous carcinoma and was not associated with improved survival in endometrioid or clear-cell carcinoma. In a second prospectively characterized cohort, intraepithelial CD8 infiltration was also associated with BRCA1 mutation or epigenetic loss of BRCA1 expression.
The important point was not simply that CD8 was prognostic.
Its meaning depended on where the cells were located and the biological context in which they occurred.
That creates several layers of potentially relevant structure:
cell type → spatial location → tumor type → molecular context
Collapsing across any one of those dimensions may be reasonable for some questions and damaging for others.
The new ampullary study raises a related issue. Stromal and intraepithelial CD8 cannot automatically be treated as interchangeable manifestations of the same immune signal. The compartment itself may carry biological information.
Preserving structure does not mean making every analysis more complicated
There is an important counterpoint here.
The answer is not to stratify every dataset endlessly or build increasingly elaborate models until every observation becomes unique. Small datasets are particularly vulnerable to overfitting, unstable subgroup findings, and post hoc explanations.
The objective is not maximum complexity.
It is decision-relevant resolution.
Before combining categories, averaging measurements, or discarding context, the analyst should ask whether the dimension being removed has a plausible relationship to the biological, clinical, or decision process under study.
In the ampullary analysis, spatial compartment had such a rationale. Immune cells situated within tumor epithelium are not necessarily biologically equivalent to immune cells remaining in surrounding stroma. Preserving that distinction allowed us to ask whether epithelial engagement carried information that total or stromal measurements did not.
The same reasoning applies in other forms of evidence generation.
In real-world evidence, treatment history and timing can matter as much as the presence of an exposure.
In patient-preference research, the context in which a choice is made may influence what the observed preference actually represents.
In health economics, an average outcome can conceal meaningful variation in who gains, by how much, and under what circumstances.
In longitudinal research, a single baseline measurement may fail to represent a process that changes as experience accumulates.
Different fields use different language for these problems, but the methodological question is similar:
Which dimensions can safely be collapsed, and which are part of the signal?
Rare datasets require restraint as well as creativity
Small or unusual datasets can encourage two opposite mistakes.
The first is to ask too little of them. If a cohort cannot provide definitive validation, it can be dismissed as incapable of producing useful evidence.
The second is to ask too much. An interesting association can be pushed into claims of independent prognosis, treatment prediction, or clinical utility that the study was never capable of establishing.
Neither is necessary.
In the ampullary study, presence-versus-absence, continuous maximum-core counts, and duplicate-core mean analyses produced directionally concordant evidence for intraepithelial CD8. At the same time, differences between likelihood-ratio and Wald inference meant that the signal could not reasonably be presented as established.
That led to a narrower conclusion:
spatial CD8 engagement deserves prospective evaluation, and future studies should preserve the epithelial versus stromal distinction rather than assume that broader immune measures are sufficient.
That is a smaller claim than declaring a validated biomarker.
It is also a more useful one.
Prognostic evidence is not predictive evidence
The same discipline is required when interpreting treatment implications.
No patient in the ampullary cohort received immune checkpoint blockade. The study therefore cannot establish that intraepithelial CD8 predicts immunotherapy benefit.
Similarly, the relatively high frequency of mismatch repair deficiency is clinically relevant because dMMR already has an established relationship with checkpoint inhibition across tumor types. The present cohort, however, contributes descriptive and prognostic evidence, not a new treatment-effect estimate.
These distinctions matter because biomarker studies are particularly susceptible to conceptual slippage:
association with outcome is not prediction of treatment benefit;
biological plausibility is not clinical utility;
and an interesting subgroup is not yet a treatment-selection population.
Preserving structure in the data should therefore be accompanied by equal restraint in the claims drawn from it.
The larger evidence-generation lesson
The ampullary study began as an investigation of immune biomarkers in a rare cancer. Its broader methodological lesson is much less disease-specific.
Useful evidence does not always emerge from collecting more variables, using larger models, or reducing uncertainty to a single number.
Sometimes the important step is recognizing that a dimension we are tempted to average away may itself carry information.
Kent and Altman approached that problem through heterogeneity in clinical-trial populations. Our earlier ovarian work demonstrated how the meaning of immune infiltration could depend on histologic and molecular context as well as cellular localization. The ampullary study provides another example in which preserving spatial information exposed a signal that broader measures could have obscured.
The general principle extends across evidence generation:
Before compressing complex evidence into a summary measure, ask whether the feature being removed could be part of the signal.
The goal is not complexity for its own sake.
It is to retain enough structure that the evidence still answers the question that matters.
Publication
Kalloger SE, Chow C, Gao D, et al.Compartment-specific CD8 infiltration and disease-specific survival in resected ampullary adenocarcinoma.Cancer Research Communications. [DOI: 10.1158/2767-9764.CRC-26-0302 https://aacrjournals.org/cancerrescommun/article/doi/10.1158/2767-9764.CRC-26-0302/788402/Compartment-specific-CD8-infiltration-and-disease
Related work
Clarke B, Tinker AV, Lee C-H, et al. Intraepithelial T cells and prognosis in ovarian carcinoma: novel associations with stage, tumor type, and BRCA1 loss. Modern Pathology. 2009;22:393–402.
Kent DM, Rothwell PM, Ioannidis JPA, Altman DG, Hayward RA. Assessing and reporting heterogeneity in treatment effects in clinical trials: a proposal. Trials. 2010;11:85.