Can senolytic CAR T cells selectively remove harmful cells? Target choice and the cost of persistence

We examined the intestinal study, its sorted-cell raw counts and per-mouse source measurements, then checked RNA co-detection in an independent lung dataset. A separate uPAR-targeting study published in October 2026 supplies actual liver-toxicity counterevidence. Positive and negative experiments used different constructs, doses, hosts and timelines. Neither cancels the other, and they cannot jointly establish a safe human dose. 6

01Recognizing uPAR does not establish harmful-state specificity

CAR binding can activate T cells and killing. The infused cells can expand, migrate and form memory, so the initial cell dose is not the effective exposure on every later day. 10

Selective removal requires four linked observations: which cells have sufficient surface antigen; whether the CAR binds and kills them; whether those cells belong to the harmful state intended for removal; and how removal versus preservation changes tissue function. RNA, a protein gate, killing and net functional benefit each cover only part of that chain.

The 2025 intestinal study found increased surface uPAR in old mice and added flow cytometry, EdU, p21, transcriptional programs and organoid function. This is stronger evidence than a single age-associated transcript. However, a negative four-hour EdU pulse does not track irreversible arrest, and p21 is not a ground-truth harmful-state label. Normally differentiated cells may also be nonproliferative. 1

02The published percentages have different denominators

Figure 1c measured the surface-uPAR gate among SPiDER-beta-gal-positive cells. The source-table uPAR-positive column averages 73.43%; that cannot be reversed into a claim that 73% of all uPAR-positive cells are harmful senescent cells. Separate staining used uPAR-positive cells as the denominator: 97.25% were EdU-negative and 70.71% p21-positive, with three mice for each measurement. These are not a joint accuracy estimate in the same cells. 1 2

The transcriptional “approximately 60%” is another conditional fraction: among signature-labelled cells, approximately 60% belonged to the surface-positive sorted group. The captured positive/negative proportions do not reproduce their original tissue abundance. The author's SenMayo gene list contains Plaur and uses a module-score cutoff of 0.02, so this label is not an independent certification of target precision. 1 3

Together these observations support co-occurrence of uPAR and some senescence-associated features. They do not provide complete lists of persistent harmful cells covered and useful cells inadvertently removed.

Human evidence is narrower: the small-intestinal RNA analysis included one donor per age group. The three “normal colon” protein-staining samples came from surgical resections in patients with colon adenocarcinoma. They support examining co-expression in those samples, rather than a healthy-population age pattern or a human CAR-treatment result. 1

03RNA detection and a surface gate are not interchangeable

We obtained four original 10x matrices from GSE233431: male/female surface-uPAR-positive/negative pools from old mice. There were 28,882 raw barcodes. Applying the author's mitochondrial, count and feature filters retained 16,809 cells: 9,430 positive and 7,379 negative, matching the paper. Four mice per gate were pooled into two libraries. Thousands of cells are not thousands of animals. 3 4

The ≥2-count threshold and all unfiltered barcodes are also retained. These data directly show that RNA detection cannot substitute for the author's surface label. Transcript sampling, protein turnover, gate placement, antibodies and processing could contribute. We did not independently certify the original protein threshold or antigen density, so we do not call these RNA false-negative rates, and certainly not CAR off-target killing rates.

Cdkn1a-or-Cdkn2a RNA detection was approximately 42.8%–44.0% in filtered positive pools and 38.6%–44.6% in negative pools. This also provides no sharp harmful-state boundary. It does not invalidate protein staining: the assays observe different quantities under different conditions. Target selection needs paired surface abundance, state tracking and actual killing.

The full control and treatment h5ad archives are approximately 7.1 and 6.9 GB compressed. We did not obtain them or redo the complete integration, differential-expression or post-treatment cell-profile analysis. Metadata, sorted raw matrices and per-mouse workbooks have separate, limited uses. 4

04Independent tissue: changing the denominator changes the answer

GSE124872 contains lung data from 15 mice aged 3 or 24 months: 14,813 barcodes, with 14,316 retained by the original annotations. We freshly extracted Plaur, Cdkn1a and Cdkn2a from the original RData and checked the corresponding official MTX rows and total counts. There was no CAR treatment or paired surface/killing truth. 5

At a ≥1-count threshold, with equal weight per mouse, mean Plaur detection was 3.98% at 3 months and 3.99% at 24 months. Among cells detecting the p21/p16 RNA union, mean Plaur co-detection was 6.45% and 8.25%. Reversing the denominator, union detection among Plaur-detecting cells was 21.9% and 24.8%. The intersection is the same; the questions differ.

This does not contradict the intestinal surface-antigen increase: tissue, composition and measurement differ. It shows why a co-marker percentage should not be transferred directly between organs. We retain the ≥2 threshold, all barcodes and cell-type results; zero denominators remain missing. Cells do not enlarge animal replication, and RNA states are not clinical risk classes.

05Preserve the functional benefits and their limits

The intestinal study used an anti-mouse-uPAR antibody fragment with mouse CD28/CD3-zeta signalling, 200 mg/kg cyclophosphamide 18 hours before transfer, and 0.5×10⁶ CAR-positive cells. Treatment in old mice and prevention from youth were separate experiments. We specified two per-mouse comparisons before reading the source values: six-week FITC-Dextran after old-age treatment, and FITC-Dextran 15 months after treatment at 3 months. 1 2

Units are the workbook's normalized arbitrary units, not human permeability rates. Both differences favor CAR. Raw-scale Welch P values are approximately 0.052 and 0.046; both become approximately 0.092 after Holm adjustment across the two comparisons. Log-scale sensitivity retains a positive long-term prevention signal (adjusted P approximately 0.049), with less certainty for treatment. Scales are not independent replications, and the paper's Student test assumes something different from our unequal-variance comparison. 2

Plasma FITC also depends on delivery, transit, distribution and clearance. Batch/sex links to individual normalized values were not provided; column order was not used to invent them. The paper's Mice section says assignment was random, whereas Statistics says no randomization method was used. This internal conflict remains unresolved. There was an additional blinded observer, but the main investigator, data collection and analysis were not blinded. Flow samples with viability below 30% were excluded. 1

Other positive proliferation, organoid, inflammation and injury-regeneration results remain relevant. A more conservative P value for one measurement does not erase them. Organoid source fields/wells lacked donor links; we did not turn more than 20 fields into more than 20 mice for a new test. 1 2

Mechanistic controls also help: marrow transplantation from treated mice was insufficient to reproduce intestinal improvement; CAR-treated old organoids had better subsequent formation capacity; uPAR-null organoids were protected. This supports epithelial contribution and antigen dependence. Marrow transplantation did not replace every resident immune cell, and organoids do not reproduce the intact intestine. Exclusive mediation by deletion of one epithelial senescent state is therefore unproved. 1

06Persistence creates a possibility, not a proven optimum

The 2024 metabolic and later intestinal studies support detectable transferred T cells at late time points following early treatment, alongside functional benefits. The intestinal prevention cohort retained lower markers and barrier readouts after 15 months. CD45.1/CD3 identifies transferred cells; it does not directly count daily target killing or certify active CAR expression in every remaining cell. 1 7

Persistence also depends on context. In the 2020 liver-fibrosis experiments, related CAR cells expanded and then contracted relatively quickly. In natural-aging models they could remain long term. Permanent residence or rapid disappearance is not a fixed property of every construct. 7 8

To isolate the value of persistence, compare persistent and terminable interventions with the same target and similar initial killing and tissue exposure, then track new targets, normal repair and function. Cross-paper comparisons change constructs, doses and disease processes together. A long effect after one dose does not establish that longer is always better.

The intestinal authors reported no intestinal malignancies. We did not find a complete at-risk denominator, monitoring coverage and per-mouse censoring/death timeline, so no cancer-risk upper bound is calculated. Positive repair after 15 Gy irradiation is informative, but one injury and dose cannot cover every infection or wound challenge. 1

07Actual counterevidence: antigen recognition can damage other tissue

The 2020 CCl₄-fibrosis study reported hypothermia, weight loss and cytokine elevations at higher 2–3×10⁶-cell doses, with better tolerability at lower effective doses. Its 2024 correction replaced an incorrectly reused control micrograph; it should be disclosed without describing the entire study as retracted. 8 9

An independent October 2026 study used the uPA ligand GFD rather than the antibody recognition domain, connecting it to CD3 and also testing GFD-CAR. High-dose BiTE caused liver injury in old mice and could cause death. Hepatic colocalization of engager, uPAR and CD31 accompanied endothelial loss. Young mice also developed liver toxicity, and global uPAR-null mice were protected. This supports antigen-dependent injury beyond a purely hypothetical concern. 6

Global receptor deletion is not endothelial-specific rescue. Colocalization and protection do not isolate endothelial necessity or exclude other uPAR cells, distribution and inflammatory intermediates. GFD-CAR used human T cells, 4-1BB signalling and young tumor-bearing NSG mice, with 5×10⁶ cells in the methods. The intestinal study used mouse antibody/CD28 cells, aging syngeneic mice and 0.5×10⁶ CAR-positive cells. The new toxicity therefore does not prove inevitable harm from that low-dose intestinal regimen; construct differences also do not make it irrelevant to target appraisal.

Lower BiTE doses retained mouse marker and functional benefits. The primate component included only three old macaques, with single-animal dose escalation before repeat dosing. One animal had appetite loss and aminotransferase abnormalities at the selected lower dose, followed by recovery. Broad abstract-level safety wording cannot replace these observations. There are no matched human results or direct safety comparisons with persistent CAR at equivalent effective exposure. 6

ALT/AST can monitor the liver injury observed here; they do not certify absence of unintended removal in every organ. Joint tissue/state gates and controlled residence or exit are testable improvements, requiring demonstrated preservation of benefit and reduction of actual injury.

08What this study establishes

uPAR targeting has experimental support for antigen dependence and functional benefits in mice. The evidence does not establish exclusive removal of persistent harmful senescent cells: marker denominators differ, RNA and surface gates are not interchangeable, and other target states have actual toxicity evidence. Long-term effects have not been matched against transient exposure.

Our contribution is two per-mouse functional recalculations, RNA co-occurrence in four surface-sorted pools and 15 independent lung mice, and an appraisal of construct, timing and safety boundaries. We did not redo complete Seurat or treatment transcriptomics, certify the original protein gates, or recalculate the newest toxicity supplements. The latter DOCX returned a verification page/403 and is not registered as read. No human anti-aging CAR precision, safe dose, healthspan or lifespan conclusion follows.

Download the reproduction package. The same Codex agent researched, calculated, self-reviewed and translated this article; there was no independent human or professional review. Third-party originals and cell-level working tables stay private; the package contains our code, protocol, source locations and aggregate outputs.