- Biomarker measurements provide valuable information about biological processes, disease and treatment response, but they represent selected features of a much more complex biological system.
- Many biomarker assays provide a time-fixed snapshot. DNA, RNA and protein measurements each capture different layers of biology, and no single molecular readout provides a complete picture of what is happening inside a living cell.
- As datasets grow and AI-based biological models become more sophisticated, understanding what the underlying data represent, and what they leave out, becomes increasingly important.
What are Biomarkers?
Biomarkers are used across many areas of medicine and drug development, including diagnosis and disease monitoring, prediction of prognosis and treatment response, safety and toxicity assessment and patient selection for clinical trials. According to the FDA/NIH Biomarker Working Group, a biomarker is a ‘defined characteristic that is measured as an indicator of normal biological processes, pathogenic processes, or responses to an exposure or intervention, including therapeutic interventions’ (1). Biomarkers can include DNA variants, RNA transcripts, proteins, metabolites or other measurable components of cells and tissues.
One well-known biomarker is prostate-specific antigen (PSA), a protein measured in blood. Elevated PSA levels can be associated with prostate cancer, although they may also occur in benign conditions, so PSA is used as part of prostate cancer assessment rather than as a standalone diagnostic test. In Alzheimer’s disease, phosphorylated tau biomarkers are increasingly being used to support disease assessment. For example, the FDA-cleared Lumipulse G pTau217/β-Amyloid 1-42 Plasma Ratio uses a blood sample to help identify amyloid pathology associated with Alzheimer’s disease (2).
Within oncology, HER2 status, which is based on HER2 protein expression and/or amplification of the ERBB2 gene, can help determine whether a cancer patient may benefit from HER2-targeted therapies. Circulating tumor DNA (ctDNA) can be analyzed to detect tumor-specific genomic alterations, assess molecular residual disease and, through serial sampling, track changes in tumor burden or the emergence of treatment-resistant clones over time.
A Biomarker Measurement Is a Snapshot
Biomarkers are widely used across research and medicine, but their interpretation begins with a simple question: what does the measurement actually represent?
Cells are dynamic living systems that respond continuously to nutrients, signaling molecules, neighboring cells, drugs, pathogens, stress and other changes that may occur in their environment. Gene expression varies with time and cellular conditions, proteins are produced and degraded, signaling pathways are activated or suppressed, metabolites fluctuate, and cells transition between distinct activation, differentiation or metabolic states.
Most biomarker assays rely on samples collected at defined time points. In this sense, a biomarker measurement is a snapshot: a static observation of biological state at a specific, well-defined time point and under a particular set of conditions. RNA expression and protein phosphorylation, for example, may change rapidly in response to cellular conditions, but the resulting assay measurement captures only the state detected at that time.
Most biomarker assays rely on samples collected at defined time points. In this sense, a biomarker measurement is a snapshot: a static observation of biological state at a specific, well-defined time point and under a particular set of conditions. RNA expression and protein phosphorylation, for example, may change rapidly in response to cellular conditions, but the resulting assay measurement captures only the state detected at that time.
Not all biomarker measurements reflect an instantaneous biological state. Some biomarkers provide information about biological exposure or activity over a longer period. For instance, glycated hemoglobin (HbA1c), which is used in the diagnosis of diabetes and monitoring of long-term glucose control, reflects glycemic exposure over a period of months rather than a single point in time. When feasible, serial biomarker measurements can also provide information about how a biological parameter changes over time. Even so, each measurement represents an observation made within a defined temporal and biological context.
What Different Biomarker Measurements Reveal, and What They Do Not
Different biomarker types provide information at different levels of biological organization, from relatively stable genomic features, i.e. by analyzing genomic DNA, to more dynamic changes in gene expression and protein expression, including expression of alternative transcripts and protein post-translational modifications.
These different molecular layers provide distinct but incomplete views of cellular state and function. DNA sequencing can identify inherited variants, somatic mutations, copy-number changes and other sequence-level alterations, while targeted genotyping can be used to assess specific single nucleotide polymorphisms (SNPs). Because genomic DNA is comparatively stable, these analyses can provide valuable information about disease susceptibility or mechanisms, but sequence alone does not tell us which genes are currently active in a particular cell.
Transcriptomic approaches, including RNA sequencing, can shed light on cellular state by revealing which genes are transcribed and at what abundance, albeit at a defined point in time and under a particular set of conditions. Transcript abundance therefore represents only one layer of cellular regulation; high RNA abundance does not necessarily translate into high protein abundance, and transcript levels alone cannot establish a protein’s function.
Protein assays, including ELISA, immunoassays and mass spectrometry-based approaches, interrogate another molecular layer. But even knowing how much of a protein is present does not necessarily reveal whether it is active, where it is located within the cell, which proteins it is interacting with, or whether it has undergone phosphorylation or other post-translational modifications.
Beyond the molecular layer being measured, biological context can also influence how a biomarker result is interpreted. The same molecular feature may have different implications in different cell types or cellular states. For example, STAT3 phosphorylation can indicate active JAK-STAT signaling, but persistent activation in a tumor cell may support proliferation and survival, whereas transient activation in an immune cell may reflect a normal response to cytokine stimulation (3). Spatial context can also be important. In cancer, for example, identifying a particular immune-cell population may be informative, but whether those cells are located adjacent to tumor cells, excluded from the tumor, or interacting with other components of the microenvironment can provide substantially different biological information.
Biomarker interpretation therefore requires consideration not only of what was measured, but also in which cells or tissues, under what biological conditions, and what conclusion that measurement can reasonably support.
More Data Does Not Automatically Mean More Understanding, So What’s the Solution?
The need to understand what a biomarker measurement actually represents becomes even more important as datasets increase in size. Modern sequencing and single-cell technologies can generate measurements from millions of cells and thousands of molecular features. Such datasets can be extremely valuable, but their scale does not remove the limitations of the underlying measurements. Increasing the number of observations does not, by itself, change the type of biological information being captured.
This distinction is increasingly relevant in AI. Many biological foundation models have been trained on large observational datasets such as gene-expression atlases. These models can learn patterns and relationships within those data, but predicting what will happen after a biological intervention is a different problem. Recent work on Xaira Therapeutics’ X-Cell virtual-cell model illustrates this distinction. X-Cell was developed using large-scale CRISPRi Perturb-seq data in which genes were deliberately perturbed and the resulting transcriptomic responses measured across millions of cells (4). The aim is to model not simply what cellular states look like, but how those states change following perturbation.
From Snapshots to Dynamic Biology
Biomarker measurements remain valuable, but their interpretation can be strengthened by approaches that capture biological change across time, context and molecular layers. Perturbational studies can help to reveal how cells respond when a gene or pathway is altered, while longitudinal sampling can reveal how a biomarker or cellular state changes over time. Spatial approaches can add information about where molecular changes occur, while multimodal approaches can show how different types of biological measurements, including genomic, transcriptomic and proteomic data, relate to one another.
Virtual-cell models, a concept that dates back to the 1990s but is now being revisited with modern AI and large-scale biological datasets, represent one attempt to integrate such information and predict how cellular states may change following genetic, chemical or other perturbations (5). Their performance will be shaped by the quality and biological relevance of the data used for training. More broadly, the value of these models will depend not simply on the volume of data available, but on whether those data capture the biological processes the models are being asked to predict.
Working on a virtual cell project? → fill out the form and let us know!
References
- Cagney DN, Sul J, Huang RY, Ligon KL, Wen PY, Alexander BM. The FDA NIH Biomarkers, EndpointS, and other Tools (BEST) resource in neuro-oncology. Neuro Oncol. 2018 Aug 2;20(9):1162-1172.
- U.S. Food and Drug Administration. FDA Clears First Blood Test Used in Diagnosing Alzheimer’s Disease. May 16, 2025. Available at: https://www.fda.gov/news-events/press-announcements/fda-clears-first-blood-test-used-diagnosing-alzheimers-disease.
- Johnson DE, O’Keefe RA, Grandis JR. Targeting the IL-6/JAK/STAT3 signalling axis in cancer. Nat Rev Clin Oncol. 2018 Apr;15(4):234-248.
- Wang C, Karimzadeh M, Ravindra NG, et al. X-Cell: Scaling Causal Perturbation Prediction Across Diverse Cellular Contexts via Diffusion Language Models. bioRxiv. 2026.
- Casini T. In vivo, in vitro, in silicio: towards a virtual cell. Trends Genet. 1993 Apr;9(4):105.
—
Karen O’Hanlon Cohrt is an independent Science Writer with a PhD in biotechnology from Maynooth University, Ireland (2011). After her PhD, Karen relocated to Denmark where she held postdoctoral positions in mycology and later in human cell cycle regulation, before moving to the world of drug discovery. Karen has been a full-time science writer since 2017, and has since then held numerous contract roles in science communication and editing spanning diverse topics including diagnostics, molecular biology, and gene therapy. Her broad research background provides the technical know-how to support scientists in diverse areas, and this in combination with her passion for learning helps her to keep abreast of exciting research developments as they unfold. Karen is currently based in Ireland, and you can follow her on Linkedin here.
