Diagnostics and Biomarker Translation
Diagnostics and biomarker translation deserve a separate chapter because discovery biomarkers often fail at the measurement and context-of-use step. AI systems may find molecular, imaging, pathology, digital, or multi-omic patterns associated with disease or treatment response, but translation requires a defined assay, locked model, intended-use population, threshold, comparator, and decision. The existing real-world evidence chapter covers causal evidence and biomarker use in therapeutic development; this chapter focuses on the assay, diagnostic, and qualification pathway.
Use this chapter to:
- Move candidate signals from discovery toward diagnostic, prognostic, predictive, pharmacodynamic, safety, monitoring, or risk biomarker use.
- Context of use, analytical validity, clinical validity, clinical utility, locked model status, and regulatory pathway matter more than association strength alone.
What is this field trying to solve? Move candidate signals from discovery toward diagnostic, prognostic, predictive, pharmacodynamic, safety, monitoring, or risk biomarker use.
What is the core idea? Context of use, analytical validity, clinical validity, clinical utility, locked model status, and regulatory pathway matter more than association strength alone.
What is the current state of the field? AI can discover and rank biomarkers in defined datasets, but qualification and diagnostic use require stronger evidence and intended-use clarity.
What do we know, and what remains open? Known reference points include FDA Biomarker Qualification Program, BEST, STARD, REMARK, companion diagnostic guidance, AI-enabled medical device lists, liquid-biopsy datasets, and validation cohorts. What remains open is whether discovered signals become locked, measured, externally validated tools for a defined context of use.
Why does this matter? Biomarker translation connects omics, imaging, pathology, trials, real-world evidence, therapeutics, diagnostics, and regulatory science.
Introduction
The FDA Biomarker Qualification Program frames biomarker development around a well-defined context of use, reliable measurement, and analytical performance that supports that context (FDA, 2026). The FDA-NIH BEST resource defines a biomarker as a measured characteristic that indicates normal biological processes, pathogenic processes, or responses to exposure or intervention, and separates biomarker categories such as diagnostic, prognostic, predictive, pharmacodynamic or response, safety, monitoring, and susceptibility or risk (FDA-NIH Biomarker Working Group, 2025). That discipline is exactly what AI biomarker discovery often lacks. A model may find a molecular, histologic, imaging, digital, or multi-omic feature associated with outcome, but translation depends on whether the feature can be measured reliably and used for a defined decision.
The scope is narrower than the Real-World Evidence and Biomarker AI chapter, which covers evidence generation using real-world data and biomarkers in therapeutic development. This chapter covers diagnostic and assay translation.
The Translation Ladder
Biomarker translation has a ladder. The order matters.
| Step | Question | Typical evidence |
|---|---|---|
| Discovery | Is there a signal associated with disease, response, toxicity, or outcome? | Exploratory cohort, omics, imaging, pathology, or clinical data |
| Measurement definition | What exactly is measured? | Specimen, assay, platform, preprocessing, feature definition, quality controls |
| Analytical validity | Does the measurement work reliably? | Precision, accuracy, limit of detection, interference, reproducibility, site-to-site variation |
| Clinical validity | Is the measurement associated with the clinical state or outcome in the intended-use population? | External validation, appropriate reference standard, locked threshold, calibration, subgroup performance |
| Clinical utility | Does acting on the result improve a decision? | Impact study, decision curve, trial evidence, workflow evidence, patient outcome, or accepted drug-development use |
| Regulatory or qualification status | What claim is accepted and for which context of use? | Device authorization, companion diagnostic labeling, biomarker qualification, or drug-application-specific acceptance |
Most AI biomarker claims sit in the first two steps. That is not a defect. It becomes a defect when discovery evidence is described as diagnostic readiness.
Biomarker Categories and Context of Use
The same measured feature may have different meanings depending on context. A tumor mutation may be diagnostic in one setting, predictive for treatment response in another, prognostic in a third, and a resistance marker after therapy. A pathology image feature may distinguish tumor grade, predict recurrence, or enrich for a molecular subtype. A liquid-biopsy signature may be used for minimal residual disease monitoring, treatment selection, recurrence surveillance, or early detection. Each use needs a separate context-of-use statement.
The context of use should include:
- Population: disease state, treatment history, setting, eligibility criteria, and intended geography or health-system context
- Specimen: tissue, blood, plasma, urine, CSF, image, wearable signal, or another measurement source
- Assay: platform, laboratory process, pre-analytical handling, quality threshold, and failure rules
- Model: locked version, inputs, preprocessing, feature extraction, threshold, calibration, and report format
- Decision: diagnosis, risk stratification, treatment selection, trial enrichment, safety monitoring, response assessment, or follow-up
- Comparator: standard of care, reference diagnosis, clinical adjudication, molecular assay, pathology review, imaging read, or clinical endpoint
- Consequence: what changes if the result is positive, negative, indeterminate, or failed
Without those elements, performance metrics are hard to interpret. A high AUC in a case-control dataset does not reveal what happens in the clinic where prevalence, specimen quality, comorbidity, workflow, and downstream action differ.
Current Models, Datasets, and Benchmarks
Biomarker AI is multi-domain. The evidence standard changes with the measurement source.
| Domain | Data or benchmark layer | What it supports | What it does not support |
|---|---|---|---|
| Multi-omics | TCGA, CPTAC, HTAN, cohort-specific genomics, transcriptomics, proteomics, methylation, metabolomics | Discovery of molecular subtypes, prognostic or predictive hypotheses, pathway features, assay candidates (NCI TCGA, 2026; NCI OCCPR, 2026; HTAN, 2026) | Clinical validity without external cohorts and locked assays |
| Liquid biopsy | ctDNA mutation, methylation, fragmentomics, multi-analyte signatures | Minimal residual disease, treatment monitoring, early-detection hypotheses, resistance tracking (Cohen et al., 2018; Cristiano et al., 2019; Klein et al., 2021) | Population screening utility without prospective evidence |
| Digital pathology | Whole-slide images, H&E, immunohistochemistry, spatial markers | Tumor detection, grading, subtype prediction, microenvironment features, trial enrichment | Companion diagnostic use without locked workflow and validation |
| Imaging biomarkers | Radiology, MRI, CT, PET, ultrasound, quantitative imaging | Segmentation, lesion measurement, response assessment, radiomics hypotheses | Device authorization or clinical utility without intended-use validation |
| Clinical prediction | EHR, claims, registries, labs, notes | Risk stratification and enrichment hypotheses | Transportable diagnostic claims without site-level validation |
| AI medical devices | FDA AI-enabled medical device list | Authorized device landscape and device-specific intended uses (FDA, 2026) | Proof that a discovery biomarker is qualified or that a different model is safe |
Reporting standards are also part of the benchmark layer. STARD 2015 lists essential items for diagnostic accuracy studies (Bossuyt et al., 2015). REMARK addresses reporting for tumor marker prognostic studies (McShane et al., 2005). These frameworks do not validate a biomarker by themselves. They make the evidence interpretable for external assessment.
What is demonstrated?
Demonstrated capability includes AI-supported biomarker discovery in bounded datasets when the feature, assay, population, endpoint, and validation cohort are specified.
Demonstrated capability also includes AI-assisted diagnostic assay development when analytical performance is evaluated against an appropriate reference standard and the intended use is clear.
Demonstrated capability includes image-based and molecular feature extraction in defined settings. Deep-learning work in dermatology showed that image classifiers could reach expert-level performance on selected binary skin-lesion tasks in a curated validation design (Esteva et al., 2017). That kind of result demonstrates task performance under study conditions. It does not automatically create a general diagnostic service.
Demonstrated capability includes liquid-biopsy discovery and validation studies with defined cohorts. CancerSEEK combined circulating proteins and mutations to detect several cancer types in a study setting (Cohen et al., 2018). Fragmentomics showed that genome-wide cell-free DNA fragmentation patterns contain cancer signal (Cristiano et al., 2019). Targeted methylation assays have been clinically validated in independent validation sets for multi-cancer early detection research (Klein et al., 2021). Each example illustrates the same rule: the claim is defined by specimen, assay, cohort, endpoint, and intended use.
Demonstrated capability includes companion diagnostic regulation as a clear decision pathway. FDA describes a companion diagnostic as a medical device, often an IVD, that provides information essential for safe and effective use of a corresponding drug or biological product (FDA, 2023). That definition is intentionally narrower than “biomarker.” A predictive biomarker becomes a companion diagnostic only when the test is essential to a therapeutic decision in the approved or proposed use.
Demonstrated capability includes regulatory transparency for AI-enabled devices. FDA’s AI-enabled medical device list identifies authorized devices that include AI-related terms in authorization summaries or classifications and links to device-specific database entries (FDA, 2026). The list is useful for seeing how intended use and device-specific review are handled. It should not be used as generic proof that any biomarker model is clinically valid.
What is theoretical?
Theoretical capability includes models that discover biomarkers that generalize across assay platforms, cohorts, sites, and treatment contexts. This is plausible in some domains, but platform shift and cohort bias remain common failure modes.
Theoretical capability also includes companion-diagnostic discovery from multimodal data. The decision standard is high because the diagnostic affects treatment selection.
Theoretical capability includes foundation models that learn reusable pathology, radiology, single-cell, and multi-omics features for biomarker discovery. Reuse is plausible because many disease features recur across cohorts. Clinical translation remains context-specific because each intended use has its own population, specimen, device, threshold, action, and harms.
Theoretical capability includes adaptive biomarker programs where early multi-omics data identify a candidate signature, the assay is locked before the pivotal cohort, and the model is updated only under a prespecified change-control plan. That approach is reasonable in research programs. It is not equivalent to letting an exploratory model change thresholds during validation.
Theoretical capability includes AI-assisted biomarker qualification packages. A program may use models to define features, detect measurement artifacts, select thresholds, identify subgroups, and analyze external validation. Qualification still depends on the FDA context of use and evidence package, not the use of AI in analysis.
What is beyond current capability?
Beyond current capabilities includes treating opaque model features as qualified biomarkers without measurement validity, biological interpretation, or context-of-use evidence.
Beyond current capabilities also includes assuming that model discrimination in a discovery cohort establishes clinical utility. Clinical utility requires a decision and evidence that acting on the result improves that decision.
Beyond current capabilities includes a universal multi-omics diagnostic model that transfers across platforms, sites, specimens, populations, ancestries, disease stages, treatments, and clinical workflows without revalidation. Platform shift is not a technical nuisance. It changes the measurement.
Beyond current capabilities includes automatic companion-diagnostic status from a predictive model. A model may predict response. Companion diagnostic status requires a linked therapeutic product, an intended-use claim, assay validation, clinical evidence, and the relevant device or drug-review pathway.
Beyond current capabilities includes treating an unsupervised feature as clinically meaningful because it is biologically interesting. Interesting biology may justify experiments. It does not by itself define a diagnostic test.
What would make this more promising?
Diagnostics and biomarker work become more promising when it advances the result from discovery to a locked, decision-linked, externally tested measurement.
| Claim | Evidence that raises or lowers confidence |
|---|---|
| “The model found a biomarker” | Feature definition, measurement method, specimen source, biological plausibility, and replication in an independent cohort |
| “The biomarker is analytically valid” | Precision, accuracy, reproducibility, limit of detection, interference, batch effects, pre-analytical variables, and site-to-site testing |
| “The biomarker is clinically valid” | Locked threshold, intended-use population, reference standard, external validation, calibration, sensitivity, specificity, and subgroup performance |
| “The biomarker is clinically useful” | Evidence that acting on the result improves diagnosis, treatment selection, trial efficiency, safety monitoring, or patient outcome |
| “The model is a diagnostic” | Locked model, intended use, input specification, threshold, report, failure mode, clinical validation, and comparison with current workflow |
| “The model supports a companion diagnostic” | Therapeutic linkage, predictive evidence, co-development or bridging logic, assay validation, and FDA companion-diagnostic pathway alignment |
| “The biomarker is qualified” | FDA qualification decision for a stated context of use, or a clear statement that the biomarker is not formally qualified |
| “The assay generalizes” | Multi-site, multi-platform, multi-operator, and population-relevant validation with prespecified acceptance criteria |
The most important shift is the locked-model moment. Before a signature is locked, performance estimates are discovery estimates. After locking, the validation cohort tests a product-like system: specimen handling, lab process, preprocessing, model version, threshold, and reporting.
Failure Modes
Case-control inflation. Many discovery studies compare clean cases against clean controls. Clinical deployment sees indeterminate symptoms, comorbidities, preclinical disease, treatment effects, and low prevalence. Predictive value changes when prevalence changes.
Leakage across cohorts. Slides, sequencing batches, sites, scanner types, collection dates, and duplicate patients may leak across train and test splits. The model may learn institution, platform, or batch rather than biology.
Threshold drift. Exploratory thresholds selected after looking at validation results inflate performance. A diagnostic claim needs a threshold locked before confirmatory evaluation.
Specimen drift. Time to processing, fixation, storage, hemolysis, ischemia time, tumor purity, plasma tube type, extraction kit, and sequencing depth may change the signal. The assay is part of the biomarker.
Platform drift. RNA-seq, targeted panels, methylation arrays, mass spectrometry, scanners, staining protocols, and image compression are not interchangeable. Model performance may shift when the platform changes.
Population shift. Ancestry, age, sex, comorbidity, disease stage, treatment history, and access to care may affect both measurement and outcome. Subgroup performance is part of validity.
Endpoint mismatch. A biomarker associated with survival may not predict treatment response. A prognostic marker is not automatically predictive. A monitoring marker is not automatically diagnostic.
Clinical utility gap. A model may classify accurately and still fail to improve a decision. Utility depends on what action follows the result and whether that action changes outcomes, safety, cost, access, or trial efficiency.
Companion-diagnostic overclaim. A predictive marker in a retrospective cohort is not a companion diagnostic. FDA’s definition requires essential information for safe and effective use of the linked therapeutic product (FDA, 2023).
What should researchers, biotech teams, funders, and program leaders do with this?
- Start with a one-sentence context of use before reviewing any performance metric.
- Classify the claim: discovery biomarker, analytical assay, diagnostic test, prognostic marker, predictive marker, pharmacodynamic marker, safety marker, monitoring marker, companion diagnostic, or qualified biomarker.
- Lock the specimen, assay, preprocessing pipeline, model version, and threshold before confirmatory validation.
- Use external validation cohorts that represent the intended-use population, not only a convenient archive.
- Report sensitivity, specificity, calibration, indeterminate rate, failed-test rate, missingness, subgroup performance, and predictive values at realistic prevalence.
- Separate model evaluation from assay validation. Both are necessary.
- Predefine how positive, negative, indeterminate, and failed results change the decision.
- Treat multi-omics integration as a measurement problem first and a modeling problem second.
- In pathology and imaging, evaluate slide or scan quality, scanner or staining variation, site effects, reader workflow, and report integration.
- In liquid biopsy, evaluate tumor fraction, clonal hematopoiesis, fragment size, methylation platform, limit of detection, and pre-analytical handling.
- For companion-diagnostic claims, align the assay plan with the therapeutic-development plan early.
- State explicitly when a biomarker is exploratory, accepted in a specific drug application, or formally qualified for a context of use.
Practical Validation Checklist
- Write the context of use before reviewing model performance.
- Separate analytical validity, clinical validity, and clinical utility.
- Name the specimen, assay, platform, endpoint, and threshold.
- Do not call a model feature a biomarker until measurement is defined.
- Link companion-diagnostic claims to therapeutic decisions.
FAQ
Is an AI feature automatically a biomarker?
No. A model feature becomes a biomarker candidate only when the measured characteristic, specimen, assay, and biological or clinical interpretation are defined.
What is the difference between clinical validity and clinical utility?
Clinical validity asks whether the result is associated with the clinical condition or outcome in the intended-use population. Clinical utility asks whether acting on the result improves a decision or outcome.
What makes a diagnostic model locked?
A locked diagnostic model has fixed inputs, preprocessing, architecture or algorithm, weights or coefficients, threshold, report format, failure rules, and intended use before validation.
Is a predictive biomarker the same as a companion diagnostic?
No. A predictive biomarker may identify likely treatment response. A companion diagnostic is a medical device that provides information essential for safe and effective use of a corresponding therapy.
Why do liquid-biopsy biomarkers need special caution?
Liquid biopsy depends on tumor fraction, shedding biology, clonal hematopoiesis, assay sensitivity, specimen handling, and disease prevalence. Discovery discrimination does not establish screening utility.