AI-Identified Biomarkers Need More Than Accuracy to Reach Patients

A promising signal emerges from the data. An artificial intelligence model has found a pattern in blood work, imaging scans, or wearable device readings that predicts disease with impressive accuracy. The numbers look good. The researchers publish. And then, often, the biomarker fails to work when tested in a different patient population, or it doesn't actually change how doctors treat anyone, or it simply cannot be reproduced. A new review in Signal Transduction and Targeted Therapy examines why this happens so often—and what needs to change for AI-discovered biomarkers to actually reach patients.

The problem is not that AI is bad at finding patterns. It is excellent at that. The problem is that finding a pattern and proving it matters in real clinical care are two entirely different things. Researchers can combine data from blood samples, imaging studies, genetic sequencing, protein measurements, cellular analysis, and wearable devices. AI can integrate all of it and identify signals that humans would miss. But a biomarker that performs well in the lab—one that correctly predicts disease in the dataset it was trained on—may collapse when applied to a different group of patients, or when the equipment changes, or when the patient population shifts in age, ancestry, or disease stage. The review's authors argue that predictive accuracy alone is not enough. A useful biomarker needs to be reproducible across independent groups, biologically plausible, and proven to actually improve clinical decisions.

The technical obstacles are substantial. Classical machine learning can identify candidate biomarkers from datasets with many variables but few subjects, yet the methods are vulnerable to overfitting—the model learns noise instead of signal. Deep learning works well for images and complex data structures, but it requires large amounts of training data and can encode hidden biases related to ancestry, site of measurement, or batch processing that masquerade as disease signals. Multi-modal approaches that combine imaging, lab work, and clinical history are promising, but they introduce new sources of error: scanner differences, variations in how images are processed, incomplete data from some patients, and socioeconomic factors that affect who has access to certain tests. Circulating biomarkers—cell-free DNA, proteins, metabolites in the blood—offer a minimally invasive window into disease, but their clinical value depends on how the sample is handled before testing, how sensitive the measurement is, how much tumor burden is present, and when the sample is collected relative to treatment. Extracellular vesicles, tiny particles that carry molecular signals between cells, are theoretically valuable but plagued by inconsistent isolation methods and lack of standardized measurement.

Most AI-derived biomarkers are correlational: they identify patterns associated with outcomes without explaining why those patterns occur. That may be sufficient for risk stratification—telling a patient their risk is high or low—but it is insufficient when a biomarker is meant to guide treatment decisions or identify a drug target. The review emphasizes mechanistic approaches: AI models that incorporate known biological pathways, that use data from genetic perturbation experiments to distinguish cause from effect, that assess how networks of interacting molecules behave rather than measuring single components in isolation. Neutrophil extracellular traps, for example, may reflect the interaction between inflammatory and clotting processes, but their value in guiding treatment has not yet been tested. Pathway-state biomarkers that capture functional activity of signaling networks may be more stable and more clinically relevant than measurements of individual molecular abundance, but this remains an area requiring experimental validation.

The gap between discovery and clinical use is wide and well-documented. A biomarker that performs well in the initial study often fails when tested in independent cohorts. Data heterogeneity—differences in how samples are collected, processed, and measured across sites—undermines reproducibility. Limited diversity in the patient populations studied means the biomarker may not work equally well across age groups, ancestries, or disease stages. Incomplete biological annotation means researchers cannot understand what the biomarker actually measures. Treatment-related confounding—the fact that patients receiving different therapies may have different biomarker values for reasons unrelated to disease—is often not controlled for. And even when a biomarker is validated, it may not integrate smoothly into clinical workflows, or it may require equipment or expertise that not all hospitals have.

The authors propose a structured development pathway: discovery, external validation across independent cohorts and subgroups, testing under different conditions and in diverse populations, demonstration of clinical utility through prospective studies, and finally deployment. Evaluation must go beyond standard discrimination measures like the area under the curve to include decision analysis—does using this biomarker actually change what doctors do and improve outcomes?—and workflow integration. Prospective studies are particularly important when biomarkers guide treatment selection or clinical trial enrollment, because retrospective studies can overestimate effectiveness.

The review identifies priorities for moving forward: prospective multicenter validation studies, harmonized data collection across sites, transparent reporting of methods and results, standardized assessment of measurement consistency and fairness, regulatory coordination, and interdisciplinary collaboration between AI researchers, biologists, clinicians, and statisticians. The authors conclude that AI can expand biomarker discovery by integrating diverse data types, but accurate predictions alone do not establish clinical value. A biomarker must be reproducibly measured, independently tested, biologically plausible, and proven to improve care. Without these elements, even the most impressive accuracy in the discovery phase means little when the biomarker reaches the clinic.