AI Biomarker Discovery: What Evidence a Model Needs Before It Reaches the Clinic

AI Biomarker Discovery: What Evidence a Model Needs Before It Reaches the Clinic

AI biomarker discovery is no longer a side project for data science teams. Deep learning models can read routine H&E slides and predict molecular status, quantify IHC staining more consistently than the human eye, and combine pathology, genomics and clinical data into prognostic scores. Some of these models now carry FDA marketing authorisations, and at least one is selecting patients for a phase 3 oncology trial.

Most published models, though, never reach a patient. The gap is rarely the algorithm. It is the evidence: whether the training data were representative, whether performance held up in an external cohort, whether the locked model was analytically validated like any other assay, and whether anyone showed that using it changes decisions or outcomes. This article sets out the evidence ladder for an AI-derived biomarker in oncology, the reporting standards and regulatory routes that apply, and the questions a pharma partner will ask before building a programme around one.

Key takeaways

• An AI biomarker is judged on the same three pillars as any biomarker (analytical validity, clinical validity, clinical utility), plus extra scrutiny of data provenance, model locking and change control.
• Internal cross-validation is not validation. Regulators and pharma partners expect performance on independent external cohorts, ideally from different sites, scanners and populations, and eventually prospective data.
• TRIPOD+AI (2024) is the reporting standard for prediction model studies; CONSORT-AI and SPIRIT-AI (2020) cover trials that test an AI intervention.
• In the US, FDA finalised its predetermined change control plan (PCCP) guidance for AI-enabled devices in December 2024; its broader lifecycle guidance for AI-enabled device software remains a draft from January 2025.
• In the EU, AI that is part of an IVDR device requiring a notified body is high-risk under the AI Act; the Digital Omnibus moved those obligations to 2 August 2028.
• Real examples span detection aids (Paige Prostate), prognostic tools (ArteraAI Prostate) and an AI-scored companion diagnostic in development (Roche’s TROP2 device using AstraZeneca’s QCS).

What counts as an AI-derived biomarker?

In oncology, “AI biomarker” covers several different things, and the evidence required depends on which one you are building:

• Quantification of a known analyte. The algorithm scores an existing stain more precisely, for example continuous membrane and cytoplasmic TROP2 scoring rather than a manual H-score. The biology is established; the novelty is the measurement.
• Prediction of a molecular feature from morphology. Models infer MSI, mutations or expression signatures directly from H&E, usually as a pre-screen ahead of confirmatory molecular testing.
• Novel prognostic or predictive signatures. Multimodal models combine images with clinical or omics data to produce a risk score that has no single-analyte equivalent.
• Detection and triage aids. Software that flags suspicious regions for a pathologist. These are diagnostic devices rather than biomarkers, but they set the regulatory precedent for the rest.

The further you move from a known analyte towards a novel signature, the more of the evidence burden shifts to clinical validation, because there is no reference method to compare against. Our article on spatial proteomics vs spatial transcriptomics covers the tissue platforms that often supply the features these models learn from.

The evidence ladder for an AI biomarker

1. Data: the foundation most programmes underinvest in

A model can only be as general as the slides and labels it was trained on. The questions that matter early are:

• Cohort size and event counts. For prognostic models, the number of outcome events matters more than the number of patients.
• Diversity. Ancestry, age, sex, tumour subtype, treatment era, and specimen type (biopsy vs resection, primary vs metastasis).
• Batch effects. Staining protocols, fixation, scanner make and model, and file format all shift image statistics. A model trained at one centre can learn the centre rather than the biology.
• Label quality. Which assay defined ground truth, how discordant cases were handled, whether pathologist annotations were adjudicated, and whether outcome labels came from trial-grade follow-up or real-world records.
• Leakage. Splits must be at patient level (and ideally site level), with no slides from the same patient on both sides.

2. Internal vs external validation

Cross-validation within a development set estimates how well the model fits that data. It does not show transportability. External validation means an independent cohort the developers did not touch during training or tuning, preferably from other institutions, scanners and populations. The MSI literature is a useful example: Kather and colleagues showed in Nature Medicine in 2019 that deep learning could predict MSI from H&E, and Echle and colleagues followed in Gastroenterology in 2020 with a model trained on more than 6,000 specimens and tested on a separate external cohort of 771. That second step, not the first, is what made clinical use plausible.

3. Prospective validation

Retrospective external validation still uses archived, often curated material. Prospective validation tests the locked model on consecutive samples as they arrive in a real workflow, with the cut-off fixed in advance. For a predictive biomarker intended to select patients, the strongest evidence comes from a trial that enrols on the AI-defined status, or from a prospectively specified analysis of a randomised trial. The general principles are the same as for any IO biomarker; see our guide on how to validate an IO biomarker in a clinical trial.

4. Analytical validation of an algorithm

Once locked, the algorithm is part of an assay and needs the same analytical characterisation: repeatability and reproducibility across runs, operators, scanners, staining batches and sites; robustness to tissue artefacts, section thickness and image quality; defined acceptance criteria and failure modes (what happens when the slide is out of specification); and agreement with a reference method where one exists. The wet-lab components (antibody clone, staining platform, scanner) are validated together with the software as a system.

5. Clinical validation and clinical utility

Clinical validation shows the output is associated with the clinical state or outcome it claims to measure, in the intended-use population, at the chosen cut-off. Clinical utility shows that acting on the result improves outcomes or decisions compared with not using it. Many AI tools stop at the first; payers, guideline committees and pharma partners care about the second.

Reporting standards: TRIPOD+AI, CONSORT-AI and SPIRIT-AI

Reporting guidelines are the fastest way for a reviewer to judge whether a model study is credible. TRIPOD+AI, published in the BMJ in April 2024, is a 27-item checklist for studies that develop or evaluate clinical prediction models using regression or machine learning, and it replaces TRIPOD 2015. It asks for, among other things, clear descriptions of data sources, handling of missing data, sample size rationale, performance including calibration, and fairness across subgroups.

For trials in which the AI system is the intervention, CONSORT-AI (14 new items) and its protocol counterpart SPIRIT-AI (15 new items) were published in 2020. They require details such as the algorithm version, how input data were acquired and pre-processed, how outputs were used in decisions, and how errors were analysed. A partner reading a data package will expect it to map onto these checklists.

Regulatory routes for AI in pathology and companion diagnostics

United States

Most standalone AI pathology tools are regulated as software as a medical device (SaMD). Novel low-to-moderate-risk tools have mostly gone through De Novo classification, which creates a product code later devices can follow through 510(k). An AI-scored companion diagnostic would normally need premarket approval (PMA) alongside the drug, following the co-development process described in our FDA companion diagnostic approval guide.

Two FDA guidance documents shape the AI-specific expectations:

• Marketing Submission Recommendations for a Predetermined Change Control Plan for Artificial Intelligence-Enabled Device Software Functions, finalised in December 2024. A PCCP describes modifications planned in advance (for example retraining, or adding compatible scanners), the protocol for validating them and an impact assessment, so that changes within the plan do not each need a new submission.
• Artificial Intelligence-Enabled Device Software Functions: Lifecycle Management and Marketing Submission Recommendations, issued as a draft in January 2025 and still listed as draft on FDA’s site at the time of writing. It covers device description, data management, model validation, performance monitoring, transparency and bias.

Laboratory developed tests are a parallel route. A federal court vacated FDA’s LDT rule in March 2025 and FDA rescinded it in September 2025, so AI-based tests run within a single CLIA laboratory remain outside premarket review for now. Pharma partners, however, generally want an FDA-authorised device when a test will be named in a drug label.

European Union

Under the IVDR, companion diagnostics are class C and need notified body review, and software that drives a diagnostic decision is itself a device. The AI Act adds a second layer: an AI system that is a device, or a safety component of one, and needs third-party conformity assessment under the MDR or IVDR is classed as high-risk. The Commission’s joint FAQ (MDCG 2025-6, June 2025) explains how the two regimes fit together, including integrating AI Act requirements into existing quality and technical documentation. The Digital Omnibus on AI, formally adopted in 2026, moved the application date of high-risk obligations for these product-embedded systems from 2 August 2027 to 2 August 2028. For the IVDR side, see our IVDR companion diagnostics guide.

Real AI biomarkers and tools in oncology

The TROP2 programme is the clearest example of an AI biomarker moving into drug development. Daiichi Sankyo and AstraZeneca linked TROP2 NMR positivity to outcomes with datopotamab deruxtecan in retrospective analyses of TROPION-Lung01, then in January 2026 dosed the first patient in TROPION-Lung17, described as the first phase 3 trial to prospectively enrol patients on TROP2 NMR status. That sequence (retrospective signal, locked algorithm, device co-development, prospective enrolment) is the evidence ladder in practice. We discuss the wider ADC selection problem in ADC and immunotherapy patient selection.

Note what the Paige authorisation rested on: a reader study in which 16 pathologists reviewed 527 slide images with and without the software. FDA noted the study did not evaluate the effect on final patient diagnosis. Even an authorised tool, in other words, carries explicit limits on what its evidence shows.

What a pharma partner will ask

If you are a diagnostics, tools or AI company pitching an AI biomarker to a drug developer, expect the questions below. Having the answers in a single data package shortens diligence considerably.

The last row is often the hardest. Our article on structuring pharma–diagnostics partnerships covers the contractual side.

Where these conversations continue in 2027

AI in pathology is one of the threads running through Stream 2 at ImmunoMark Summit San Diego on 10–11 February 2027. Stream 2: Biomarkers & Companion Diagnostics includes the panel “Panel Discussion : How AI-Driven Digital Pathology is Transforming Biomarker Discovery and Precision Medicine”, with Sripad Ram, Director, Digital Pathology and Image Analysis Group, Pfizer, among the panellists. The stream also features “Decoding Response and Resistance to Immunotherapy in NSCLC Through Integrative Multi-Omic and AI Approaches” from Ariel (Yung-Chia) Chen, Director, Biomarker and Translational Sciences, Interventional Oncology, Johnson & Johnson, and a session on “Leveraging Multi-Omics Technologies to Identify Predictive Biomarkers for Precision Oncology”.

Across both streams, 300+ senior pharma, biotech and diagnostics leaders meet for keynotes, panels, case studies, pre-arranged 1:1 meetings and a Day 1 drinks reception. The series continues in London on 3–4 March 2027 and in Boston in June 2027.

Register for ImmunoMark San Diego 2027

If your company wants to meet this audience directly, whether you build digital pathology platforms, AI biomarker algorithms, scanners or companion diagnostic services, see the San Diego 2027 sponsorship packages, or read why ImmunoMark works as a biomarker technology conference for sponsors.

FAQs

What is AI biomarker discovery?

AI biomarker discovery uses machine learning, often deep learning on digitised pathology slides combined with clinical or omics data, to find or measure features that predict diagnosis, prognosis or response to treatment. Examples include predicting MSI from H&E slides, quantifying TROP2 staining continuously, and multimodal prognostic scores in prostate cancer.

Is external validation enough for an AI biomarker to be used clinically?

No. External validation shows that a model transports to new sites and populations, but clinical use also needs analytical validation of the locked algorithm with its staining and scanning system, clinical validation in the intended-use population at a fixed cut-off, and, for most decisions, prospective evidence of clinical utility.

Which reporting standards apply to AI biomarker studies?

TRIPOD+AI, published in the BMJ in 2024, covers studies that develop or evaluate prediction models using regression or machine learning. CONSORT-AI and SPIRIT-AI, published in 2020, cover reports and protocols of clinical trials in which an AI system is the intervention.

Have any AI pathology tools been authorised by FDA?

Yes. Paige Prostate received De Novo authorisation in September 2021 as an aid for detecting cancer on prostate biopsy images, and ArteraAI Prostate received De Novo authorisation in August 2025 as a prognostic tool for non-metastatic prostate cancer. Roche’s AI-scored TROP2 companion diagnostic received Breakthrough Device Designation in April 2025 but has not yet been approved.

How does the EU AI Act affect AI companion diagnostics?

AI software that is an IVDR device, or a safety component of one, and needs notified body assessment is classed as high-risk under the AI Act. Manufacturers can integrate AI Act requirements into their IVDR documentation, and the Digital Omnibus moved the application date for these obligations to 2 August 2028.