In 2004, researchers gave a name to the chromatin and granule proteins released by neutrophils into extracellular space: neutrophil extracellular traps, or NETs. These are physical structures. DNA forms the scaffold; proteins such as myeloperoxidase and neutrophil elastase decorate it; together they can ensnare microbes (Brinkmann et al.).
More than two decades later, however, we often use intracellular RNA to speak about this extracellular net. If inflammatory pathways, reactive oxygen species, degranulation, or PADI4-related genes rise, a gene set may acquire the label “NET signature.” The name comes easily. The evidence does not always follow.
The project I have been working on therefore begins by rewriting the question:
We are not asking whether a transcriptome contains something that looks NET-like. We are asking whether it can identify actual or impending NET release without mistaking general activation, degranulation, cell death, or disease severity for the endpoint itself.
Two spaces and a gap between them
RNA sequencing is an intracellular snapshot. NET release is an extracellular endpoint unfolding over time. The two can certainly be connected. Khan and Palaniyar showed that transcriptional initiation itself can contribute to chromatin decondensation, and that blocking transcription suppresses NETosis under several inducing conditions. Yet “transcription participates in NET formation” does not imply that one expression signature will remain valid across stimuli, diseases, time points, and sequencing platforms.
One upstream pathway can feed several neutrophil functions, while the same NET endpoint can be reached through different stimuli, kinetics, and molecular routes. A 2026 study offers a particularly useful counterexample: soluble uric acid altered phagocytosis, oxidative burst, cytoskeletal dynamics, and degranulation, but did not produce a corresponding change in NET formation in the reported experiments. Neighboring functions can move together, but they can also separate.
The most valuable evidence is therefore not that a dataset is “enriched for a NET pathway.” It is that RNA and a direct NET endpoint belong to the same biological unit—the same donor, the same sampling event, and preferably the same time point. Without that link, concordance across studies may show that two phenomena share an inflammatory background. It cannot prove that one measurement sees the other.
Arranging evidence by distance from the endpoint
I now place evidence into three layers:
1. Same-unit quantitative truth. RNA and an endpoint such as MPO–DNA complexes, imaged NET area, or another direct measurement come from the same donor and time point. This is what can test individual-level prediction. 2. Paired mechanistic direction. The same donors receive an intervention known to promote or suppress NET formation, and the transcriptomic score is tested for movement in the expected direction. This can support mechanistic consistency, but it is not yet clinical prediction. 3. Disease or stimulus association. A case group is more inflamed or a pathway is more active. This is useful for generating hypotheses, but it cannot certify a NET signature on its own.
This ranking changes the apparent value of datasets. A large cohort without sample-level functional mapping may be less informative than an experiment with only three donors but clean pairing and a mechanistic intervention. A small study can become the strongest positive control if its causal contrast is sharp enough.
What the data say so far
We froze the score, analysis rules, and unblinding conditions before examining validation outcomes, so that the ruler could not be redesigned after seeing the answer. Across two kinds of mechanistic positive control, the frozen score moved in the expected direction.
The transcriptome, then, is not blind to NET biology. A signal exists within at least some well-defined mechanistic windows.
The picture changed as soon as the task expanded across stimuli, diseases, and studies. Classification approached chance-level behavior. The model could more readily learn study identity, platform, and broad activation state than a stable NET endpoint.
The clinical dataset closest to quantitative truth taught a different lesson. It contained same-visit pairs of neutrophil RNA and serum MPO–DNA measurements, but it did not satisfy the measurement-support condition specified before the one-shot unblinding. We therefore declined to force the result into either a positive or a negative interpretation.
That is not an excuse invented after an analysis failed. It is the third answer that an identifiability framework must permit:
Not “associated,” and not “unassociated,” but “not decidable under the present measurement conditions.”
A provisional conclusion
For now, I trust a narrower but more testable statement:
A transcriptome can reflect NET propensity within particular stimuli, time windows, and biological contexts, but we do not yet have evidence that it can identify released NETs robustly across domains.
This has less rhetorical force than announcing a universal NET signature, but it has a boundary. A responsible model should return more than a score. It should know when interpretation is supported, when it is extrapolating, and when it must abstain.
In science, the most dangerous error is sometimes not an inaccurate measurement, but an overly confident name for whatever was measured. A transcriptome may indeed be able to see a net. Before claiming that it does, we must first show that it is not merely seeing the dust raised while the net is being woven.
Note: This essay records a provisional, non-peer-reviewed research judgment as of August 15, 2026. To avoid prematurely disclosing key implementation details, the composition of the frozen score, exact effect sizes, sample mappings, and one-shot validation details are intentionally omitted.