Life science companies rigorously scrutinize model architecture, training methods, assays, and data quality. Yet the physical condition of the sample that produced the data often receives far less scrutiny.
It is easy to forget that every dataset derived from patient blood begins with physical samples. The condition of those samples when they are measured determines the data that the assay generates. Scientists usually use “condition” to describe something set deliberately, a dose, a temperature point, a treatment. But samples are also exposed to conditions no one sets on purpose: temperature swings in transit, pressure changes during air shipment, mechanical jostling on a truck. If the samples undergo changes between collection and analysis due to certain unintended physical conditions, the assay captures those changes in addition to the biology of interest. By the time the data reaches the model, the effects from those unintended pre-analytical conditions are already embedded in the measurements (Moser et al.).
The model does not know the difference. If a collection or handling artifact tracks with the outcome being predicted, the model can learn that pattern just as readily as it learns a true biological signal. Nothing in the data inherently tells the model which is which.
How Sample Handling Becomes a Data Problem
Blood-based datasets provide a clear example of how sample handling can become a data problem. Blood does not become static once it enters a collection tube. Cells remain active after the draw and could eventually break down, altering the composition of the plasma.
This is different from degradation in the usual sense of a biological signal fading. Blood cells within a sample can break down, often driven by temperature extremes, rough handling in transit, or delayed processing, releasing material into the plasma that was not originally there at draw time. When such a sample is analyzed, the assay measures the contaminating material along with what was originally free-floating in the bloodstream.
Research using cell-free DNA (cfDNA), which is found in blood plasma, makes this problem especially clear. A cfDNA workflow begins with a whole-blood draw. At that point, the tube contains plasma along with red and white blood cells, platelets, and other cellular components. Before cfDNA is analyzed, the plasma must be separated from those cellular components. Until that separation occurs, changes within the blood sample can alter what ultimately ends up in the plasma.
In a controlled study of 27 healthy donors, delaying plasma separation for 24 hours at room temperature increased both the measured concentration of plasma DNA and the fraction of longer DNA fragments, consistent with the release of genomic DNA from white blood cells during storage (Chan et al.). No significant changes were observed at six hours. Those handling-induced differences ultimately become part of the training data and can be learned by the model.
Addressing this is challenging because these changes may not trigger an obvious quality-control failure. The sample could still look usable, the assay could run normally, and the resulting data could appear technically sound. The problem may only become visible later, when results do not reproduce across sites, batches, or studies.
When a model does not hold up on an outside cohort, that reveals a generalizability problem but not its source. Once the physical samples are gone, it is virtually impossible to reconstruct differences introduced during collection or handling.
Tracing Data Back to the Sample
If sample history can shape the data, then understanding a training dataset means tracing it back to the physical samples that generated it. How difficult that is depends largely on where the data came from.
Companies that collect their own samples have the most direct control over collection and handling conditions. The importance of that control is illustrated by a recent plasma proteomics study of amyotrophic lateral sclerosis (ALS). When researchers developed a machine learning classifier, blood collection tube type was retained as one of the final 20 features, alongside 17 proteins, age, and sex (Chia et al.). Collection context was useful enough that it survived feature selection. In this study, tube type was linked to where the samples came from, so the finding does not show that the tube itself altered the biology, but it does show that collection-related variables can carry information a model will use.
The connection becomes harder to see when companies acquire data from contract research organizations (CROs), biobanks, or other partners. Different sites may collect and process samples under different conditions, and those differences travel with the resulting data. The potential for variation grows as datasets span more collection sites. A 2025 cfRNA foundation model, for example, was developed using 13,014 serum and plasma samples drawn at more than 128 collection sites (Karimzadeh et al.).
Variation can also enter after specimen handling is complete. A multi-institution cfDNA study standardized the front end — one collection tube, double centrifugation, one extraction kit — and still found feature shifts from batch to batch. This drift was apparent in the control samples themselves, where no biological difference should exist. To correct it, the investigators carried five of those controls through every library preparation batch and rescaled each feature to a global reference before modeling (Lee et al.). This illustrates that sources of variation are layered: specimen handling and downstream analytical processing each require their own controls.
For companies training models on public or licensed datasets, the original sample is even further removed. The model developer may never handle a blood tube, yet the sequences, protein measurements, or other biological data used for training were still generated from physical samples somewhere upstream. If the available documentation does not capture how those samples were collected and handled, separating biological variation from pre-analytical variation becomes nearly impossible.
Pre-analytical conditions belong in the data-quality specification. For internally generated data, collection and handling requirements can be defined before the study begins. For data obtained from partners, those conditions can be documented and required as part of the dataset. And for public or licensed data, teams can trace the data back to the source studies and record what is, and is not, known about the samples that produced it.
The Missing Variables in Biological Data
Biological measurements are snapshots of a sample at a particular point in its history. Once that sample is gone, the only way to know what shaped the measurement is through the information that travels with the data.
For teams building models from biological data, that sample history should be treated as part of the dataset itself. Capturing it up front turns an invisible source of variation into something that can be evaluated rather than discovered later through failed replication.
Reading List
Moser T, Kühberger S, Lazzeri I, Vlachos G, Heitzer E. Bridging biological cfDNA features and machine learning approaches. Trends in Genetics. 2023;39(4):285-307. DOI 10.1016/j.tig.2023.01.004.
Chan KCA, Yeung SW, Lui WB, Rainer TH, Lo YMD. Effects of preanalytical factors on the molecular size of cell-free DNA in blood. Clinical Chemistry. 2005;51(4):781-784. DOI 10.1373/clinchem.2004.046219.
Chia R, Moaddel R, Kwan JY, et al. A plasma proteomics-based candidate biomarker panel predictive of amyotrophic lateral sclerosis. Nature Medicine. 2025;31(10):3440-3450. DOI 10.1038/s41591-025-03890-6.
Karimzadeh M, Sababi AM, Momen-Roknabadi A, et al. A multimodal cell-free RNA language model for liquid biopsy applications. Nature Machine Intelligence. 2025;7:1927-1938. DOI 10.1038/s42256-025-01148-x.
Lee TR, Ahn JM, Lee J, et al. Integrating plasma cell-free DNA fragment end motif and size with genomic features enables lung cancer detection. Cancer Research. 2025;85(9):1696-1707. DOI 10.1158/0008-5472.CAN-24-1517.
