5 key criteria for biomarker development

AI-driven biomarker development can accelerate the identification of biological patterns, but it only provides clinical value when it combines high-quality data, independent validation, reproducibility and professional oversight. A statistically promising result is not, on its own, a useful biomarker.

In summary:

• A clinical biomarker must be linked to a measurable biological characteristic and address a specific clinical question.
• Artificial intelligence can prioritise candidates, integrate data and automate analyses, but it does not replace clinical validation.
• Biomarker validation must assess analytical performance, reproducibility, clinical utility and the population studied.
• AI-assisted interpretation requires traceability, professional review and a clear distinction between statistical association and clinical decision-making.

Sophia’s clinical AI can help structure the interpretation of reports, always within the limits of the available evidence and under the responsibility of the healthcare professional.

What turns a biological pattern into a clinical biomarker?

A biological pattern becomes a clinical biomarker when it can be measured reproducibly, is related to a relevant process or outcome, and provides useful information for a specific healthcare decision, beyond an association observed in a dataset.

It is useful to distinguish three levels:

• Biological signal: a difference or variation observed in a sample.
• Candidate marker: a signal that raises a hypothesis and requires further evaluation.
• Clinical biomarker: a measure whose usefulness has been studied for a specific clinical question.

For example, detecting a difference in microbial composition is not enough to inform a decision. It is necessary to know what is being measured, using which method, in which population and for what purpose. The difference between an orientational measurement and an analysis intended for clinical use is addressed in this comparison between clinical and consumer microbiota testing. A result that is statistically associated with a condition does not, on its own, demonstrate diagnostic, prognostic or predictive ability.

What tasks can artificial intelligence automate in biomarker development?

Artificial intelligence can automate the preparation and integration of omics, clinical and functional data, identify candidate variables, classify samples, detect non-linear relationships and prioritise hypotheses. However, its results depend on sample quality, study design and an independent evaluation to assess whether the model generalises.

In practice, an algorithm can:

• Standardise data, identify missing values and reduce repetitive preparation tasks.
• Explore many variables to select candidates that warrant further evaluation.
• Classify samples according to combined patterns, without turning that classification into a clinical decision.
• Detect non-linear relationships that may go unnoticed with simpler methods.
• Rank hypotheses to guide further analyses, such as integrating clinical data with a blood metagenomic analysis.

The result is a list of prioritised patterns or candidates, not a validated biomarker. Professional oversight should review data quality, biological coherence and the possibility that the model is learning characteristics specific to the sample or centre, rather than a generalisable signal.

How is an AI-developed biomarker validated?

An AI-developed biomarker is validated when the measurement is reliable, the association holds in an independent population, and the result informs a specific clinical decision; strong performance within the training data alone does not demonstrate reproducibility or clinical utility.

The evaluation should separate the samples used to develop the model from those used to test it. This separation reduces the risk of overfitting, meaning that the system learns features specific to the initial dataset rather than a generalisable pattern.

It is useful to assess:

• Analytical validity: the measurement produces consistent and technically reliable results.
• Clinical validity: the observed relationship holds in independent and representative people.
• Calibration: the predictions correspond to the observed frequency of the outcome.
• Clinical utility: the biomarker provides relevant information for a clinical decision.

Reproducibility also requires documenting the samples, variables, processing and method. If it incorporates functional signals, functional metabolomics may form part of the evaluation, without replacing clinical validation of the complete set.

What risks does the automation of biomedical data science introduce?

Automation can amplify selection bias, confuse correlation with causation and produce less reliable models outside the population studied. It can also conceal errors caused by overfitting or information leakage, which is why it requires documentation, quality controls, subgroup evaluation and review by healthcare professionals.

The main risks are:

• Selection bias: the data may represent certain groups, centres or clinical situations less accurately.
• Confounding variables: a characteristic associated with the outcome may appear to be a biomarker without being the cause or providing independent clinical value.
• Overfitting and information leakage: the model may learn features specific to the development data or incorporate information that would not be available when making the decision.
• Loss of performance: changes in patients, measurement techniques or prevalence may reduce the model’s usefulness.
• Lack of explainability: a prediction that is difficult to review makes it harder to detect errors and justify its use.

Monitoring and training for healthcare professionals help maintain responsible oversight, but they do not replace independent validation or clinical judgement.

How can clinical AI be integrated into biomarker interpretation?

Clinical AI can help structure biomarker results, relate them to the available context and identify missing information, provided that it preserves source traceability, makes the analysis limitations explicit and keeps diagnostic or therapeutic decisions under the responsibility of the healthcare professional.

In practice, it can make a report easier to interpret by:

• organising results from different tests;
• distinguishing findings, associations and hypotheses;
• linking each interpretation to the available data;
• indicating when the evidence is insufficient or cannot be extrapolated.

This integration is particularly relevant when combining layers of information, such as the microbiome, functional data and clinical data, in integrated profiles. The system should allow users to review which data it has used and how it has reached each conclusion. AI supports interpretation; it does not replace clinical assessment, patient examination or professional decision-making.

Frequently asked questions

What is the difference between a biomarker and clinical data?

A biomarker is a biological measurement related to a process, state or outcome of interest, while clinical data describes information about a patient, such as a symptom, medical history or care outcome. Both can be combined in a model, but not every item of clinical data or biological signal constitutes a biomarker with demonstrated clinical utility.

Can an algorithm become a biomarker on its own?

No. An algorithm is a method for analysing data and generating a classification, prediction or score; a biomarker is the measurable characteristic related to a biological or clinical question. For an algorithmic result to be used as a biomarker, it must be defined, measured consistently and evaluated in independent settings.

What data does an AI model need to identify biomarkers?

A model needs relevant, well-characterised data linked to a specific clinical question. This may include biological measurements, clinical information and follow-up outcomes, collected according to consistent criteria. It is also necessary to understand the population studied and the quality of the labels, because errors or imbalances in the dataset can influence the conclusions.

How can bias in biomedical models be reduced?

Bias can be reduced through transparent participant selection, quality controls, review of relevant variables and separate evaluation of different subgroups. It is also advisable to test the model with data that were not used during its development and to document its limitations. These measures do not eliminate bias completely, but they help to identify it and assess its impact.

Can the usefulness of a biomarker vary depending on the population studied?

Yes. The usefulness of a biomarker may vary when the characteristics of the population, the prevalence of the process being studied, the measurement criteria or the care setting change. Therefore, a result observed in a specific group should not be automatically extrapolated to other patients. Applicability requires assessing whether the available evidence represents the intended use.

Learn how Sophia can help interpret clinical reports with AI support