Which senescent cells are we actually identifying? Do different markers describe the same state?

01One marker answers a limited question

Both p16 and p21 constrain the cell cycle, but their regulation, timing and cellular contexts differ. p16 primarily restricts progression toward replication through CDK4/6–RB; p21 participates in damage responses, cell-cycle control and differentiation. Secreted factors, lysosomes, DNA damage, morphology and function provide other kinds of evidence. SenNet recommendations

Single-cell RNA adds another distinction. Our calculations measure the fraction of cells with detectable Cdkn1a, Cdkn2a or Plaur gene counts, not protein-positive fractions. The Cdkn2a locus has different products, so gene-level counts cannot identify the p16 transcript alone. A zero can reflect low expression or failed detection. When a gene is absent from a file's feature list, even an observed zero cannot simply be assumed. SenePy methods

SenNet recommends combining hallmarks within the same cell and probing at least three hallmarks in tissue studies. Three RNA markers reflecting similar cell-cycle or stress programs do not automatically constitute three independent kinds of evidence. Several positive results are also not an already validated diagnostic rule. s5

02Multimodal progress and its current validation scope

The 2026 RamanOmics study combines single-nucleus RNA, spatial transcript measurements and Raman imaging in young and old mouse lung and skin. Raman spectroscopy detects chemical-bond vibrations and can add biochemical information that RNA does not directly describe. Raman imaging precedes spatial measurements on the same section, with registration connecting measurements to cells. Single-nucleus RNA comes from adjacent material and is not another measurement of those exact cells. s1

Target cells are defined by Cdkn1a expression greater than zero. The authors acknowledge coverage of a p21-positive subset. Combined-model AUCs are approximately 0.765 in lung and 0.722 in skin; testing randomly splits downsampled cells. The code also selects some features using labels from all candidate cells before the training/test split. This evaluation cannot replace validation in new animals with feature selection restricted to training animals. Paper, pinned code

The directly measured spatial panel contains 890 genes; additional spatial expression is imputed from reference RNA. The authors separately check selected genes by in situ hybridization, so experimental validation is present. However, the entire imputed transcriptome cannot count as a second independent measurement. Two modalities predicting the same p21 label support biochemical and transcriptional relationships associated with that label. They have not validated all senescent cells, persistence or clearing value. s1, supplements

Balanced testing samples roughly equal numbers of positive and negative cells. Its positive predictive value cannot be transferred directly to tissues where positives are rare. The proportion of predicted positives that are truly positive also depends on prevalence and false-positive rate; AUC alone cannot supply either.

03What changes when we redefine the same cells?

We use original counts and metadata from another lung study: seven 24-month-old and eight 3-month-old mice, comprising 14,813 cell records. Excluding the authors' low-quality and unclassified labels leaves 14,316 cells in the main comparison. The statistical units remain the 15 animals. Every entry agrees between the R count object and companion matrix; each cell's UMI total and detected-gene count also match metadata. Original study, public data

This study is absent from SenePy's listed training/validation projects, permitting an external-study check of fixed definitions. RamanOmics did use it as a cell-type reference, so it cannot be presented as independent validation of the RamanOmics classifier.

With counts greater than zero, the main sample includes 1,743 Cdkn1a-positive, 19 Cdkn2a-positive and 570 Plaur-positive cells. Only three detect both Cdkn1a and Cdkn2a, giving an intersection/union of approximately 0.17%. This is RNA detection overlap within the same dataset, not sensitivity or specificity for true senescence.

Changing the definition changes the estimand in the age comparison. We calculate proportions within each mouse and compare animal means. Small-sample Welch intervals depend on independent animals and approximate distribution assumptions. Exact permutations also require exchangeable age labels, which observational aging data do not guarantee. Neither removes confounding or measurement error.

The Cdkn2a age association has a Holm-adjusted P of approximately 0.049 across five primary comparisons. Yet only 19 cells are positive. Requiring at least two UMIs reduces the estimated difference to 0.03 percentage points, with an interval crossing zero. This signal deserves retention and depends visibly on definition. Reproducibility package

“Cdkn2a significant, Cdkn1a not significant” does not establish different age effects. A paired, within-animal comparison added after the primary results estimates their age-effect difference at about 0.38 percentage points, with an interval from −3.97 to 4.72. We do not use this to declare one marker correct and another wrong.

Low-expression genes are more likely to escape detection. Rare Cdkn2a and more common Cdkn1a detections will consequently often fail to overlap. Omitting each animal in turn retains a positive Cdkn2a age difference of about 0.17–0.24 percentage points. Ten fixed-seed, 50% UMI-thinning runs yield about 0.03–0.16. Direction is somewhat stable while magnitude and threshold conclusions remain sensitive to detection. Thinning is a technical sensitivity check, not an estimate of biological false-negative rate. No cell type meets the predefined threshold of at least 100 cells in every one of the 15 mice. We neither fill missing strata nor claim full common-composition standardization.

A 2025 study also finds different p16 and p21 distributions using protein measurements and imaging across tissues, suggesting RNA dropout need not explain every discrepancy. These findings do not establish the same irreversible state in every positive cell; computational trajectories do not substitute for lineage tracking. s4

04A fixed signature retains an age signal—but what does it measure?

SenePy develops tissue- and cell-type-specific transcriptional signatures and examines them in p16 reporter and induced models. The authors also report weaker performance in some cultured-cell settings. Its contribution is richer, more specific measurement, rather than biological ground truth obtained merely by adding genes. s3

We apply the authors' fixed 635-gene universal mouse list exploratorily to the external lung data, which contain 594 of these genes. Background genes and a high-score threshold are fitted only in young training animals, holding out each animal and weighting training mice equally. No old-animal result selects the threshold. This frozen-background application differs from the original program's default of recomputing backgrounds in the input dataset. Code and signature

With raw-UMI background bins, the mean fraction exceeding the training-reference threshold is approximately 4.25% in old mice and 0.74% in young mice: a difference of 3.51 percentage points, with an approximate interval of 0.75–6.27. Normalizing before background binning gives a similar direction; continuous scores are also higher overall in old mice.

In the main analysis, held young mice have seven other young training animals while old mice have eight. Matching reference training size to seven gives approximately 4.03% high-scoring cells in old mice versus 0.74% in young mice, retaining the direction. Training sets overlap: this is an animal-level descriptive check, not new independent replicates or another P value.

Stratified findings are not uniform. Among strata meeting the fixed 100-cell gate, the raw-UMI-background continuous mean in alveolar macrophages is slightly lower in old mice; high-scoring type II pneumocytes are extremely rare in both groups. These strata cover only some animals. Within the same animal and cell type, score–UMI-depth correlations can be positive or negative. The overall association does not establish an increase in every cell type, and simple depth checks do not exclude technical effects.

This positive finding supports some external-study age association of a fixed signature. It does not validate the 4.25% as truly senescent cells. Cell identity, inflammation, detection depth and age-related programs can still contribute. This branch was added after reading the primary marker results and is exploratory. Without clearing or persistence references, it cannot estimate therapeutic target coverage.

05Constraining biological states requires evidence beyond markers

In 2017 macrophage experiments, p16 and SA-β-gal change following immune stimulation, and some changes reverse. This does not deny that macrophages can become senescent. It shows that two familiar positive markers can still reflect another physiological response. The study has relevant company funding and author interests; its particular models do not represent all tissues. s6

In a 2026 mouse brain-barrier study, several p21-positive cell classes share senescence-associated transcriptional features. Cajal–Retzius cells lack several additional cellular hallmarks and are not classified by the authors as the same senescent state. Choroid plexus epithelium exhibits a persistent, relatively non-inflammatory state and barrier-related function. Positivity, persistence and harm are three distinct questions. s14

Broad p21-cell ablation in that study produces severe developmental phenotypes but removes multiple cell classes and other functions. All damage cannot be attributed uniquely to losing a senescence program, nor can its magnitude be extrapolated to adult clearing therapy.

The frequently cited Demaria 2014 acute wound-healing study has a journal expression of concern in 2026. We verified the formal bibliographic record but did not obtain the notice's full text. We do not speculate about its reason or use this study as a settled benchmark of repair benefit. Original study, expression of concern

06From cell identification to intervention evaluation

Raman, RNA, protein and functional measurements address different questions. More complete studies need to establish whether the same cells have mutually supporting hallmarks, how those hallmarks persist or recede with time and environmental change, and whether changing state or removing cells improves actual function. Surface targets such as uPAR additionally require protein-density measurements, overlap with normal states and killing controls. Plaur RNA positivity cannot be directly translated into CAR-T selectivity. s16

The public RamanOmics RNA derivatives we obtained cover 12 tissue samples and 19,096 barcodes from the same six mice. Salmon processing, spliced/unspliced definitions and retained sample scope differ from the authors' pipeline. Missing gene columns remain unknown. Current public indexes list no original Raman/STARmap files; we have not obtained reproducible same-cell paired matrices or completed animal-held-out multimodal classifier reanalysis. s9, pipeline audit

The next step is to obtain paired data, restrict feature selection to training animals, test multimodal gains in new animals, and use orthogonal state and functional references to determine which program is identified. Current evidence supports boundaries between detection definitions and applications. It does not support selecting human cells for clearance from a single marker, spectrum or high score.

Search cutoff: 2026-10-08. This is a focused question-driven investigation, not an exhaustive systematic review. Research, calculation, self-review and translation are performed by the same Codex agent. The package supplies code, minimal aggregate materials and acquisition instructions, without redistributing third-party paper text or complete raw matrices.