Clearing Senescent Cells: When Could It Help or Harm?

Substantive correction, 20 September 2026: We correct the subgroup test's null hypothesis and a multiplicity conclusion caused by automatically switching rank-test methods. We add effect intervals, actual data coverage and the formal statistical plan. We withdraw claims that the radius signal disappeared under another p16 assay, that a negative SASP panel showed failure to engage the mechanism, and that clearing premalignant senescent cells removed an anticancer barrier. Both editions, figures and the reproducibility package have been revised. The biological question remains; certainty is now limited to what the evidence supports.

01Blocking senescence, clearing cells and suppressing immunity are different interventions

Cellular senescence comprises states that vary with cell type, damage and environment. p16, p21, SA-β-gal and the senescence-associated secretory phenotype (SASP) provide useful clues, but no single marker identifies every senescent cell. A p16-positive cell is not automatically a cell that should be removed; lower marker levels need not mean cell death.

Three experiments are particularly easy to confuse: causing damaged cells to stop proliferating, eliminating cells already in that state, and altering the immune cells that recognize and clear them. A protective role for the first step does not establish that the second is harmful.

Kang and colleagues' liver-cancer model illustrates this distinction. Oncogene-activated hepatocytes entered senescence and were then cleared through surveillance involving CD4 T cells and monocytes/macrophages. When surveillance was impaired, premalignant cells persisted and liver cancer developed. Protection here included clearance of premalignant senescent cells; those cells did not themselves perform the surveillance. The original article used this evidence in the wrong direction. Kang 2011

Krizhanovsky and colleagues similarly distinguished growth arrest from subsequent clearance in liver fibrosis. Disabling senescence regulators allowed activated hepatic stellate cells to keep proliferating and worsened fibrosis; NK-cell clearance of senescent stellate cells helped fibrosis resolve. This does not show that clearance disrupts a self-limiting antifibrotic process. Krizhanovsky 2008

Drugs and genetic tools are also not interchangeable. INK-ATTAC induces apoptosis according to particular promoter activity, without perfectly covering all senescent cells. Dasatinib plus quercetin (D+Q) affects multiple survival and signalling pathways, with cell selectivity that depends on experimental conditions. It is neither a perfectly selective eraser nor an intervention whose effects can be declared mostly unrelated to clearance without measurement. Zhu 2015

02What the Farr trial actually studied

Farr and colleagues' 2024 trial compared intermittent D+Q with no treatment at one centre, without blinding, over 20 weeks. Participants were postmenopausal women older than 60 whose T-cell p16 expression was already above the 95th percentile of a young-women reference group. The subsequent highest-tertile, or T3, subgroup is therefore a relative stratum within an enriched population, not a threshold applicable to all older adults. Original trial

Posted registry results cover 74 people: 30 assigned to D+Q, 14 to the original fisetin arm and 30 controls. The paper explains that the third arm was discontinued because of the pandemic and costs; its principal published comparison contains 60 participants. The registry retains fisetin results. These are parts of the same study, not independent trials. Study completion was 27 and 28 in the D+Q and control arms, while week-20 CTx had 28 values in each arm. Discontinuing treatment, completing all visits and contributing a particular measurement are different states. Registry and results

The primary endpoint was change in the bone-resorption marker CTx from baseline to week 20. P1NP reflects bone formation; DXA bone mineral density (BMD) provided other skeletal outcomes. None directly measures lifespan, fewer fractures or broad healthspan.

We obtained the publisher's workbook, supplementary material and the 8 December 2023 formal protocol posted on the registry. It specifies ANCOVA and extrapolation of missing week-20 measurements from time trends. The paper presents rank comparisons and reports ANCOVA and imputation sensitivity analyses. Our analysis uses public plotted values; it does not independently reconstruct the full intention-to-treat or covariate-adjusted analysis. Nor does a later uploaded protocol establish that every decision was locked before results were known. Source data, protocol and statistical plan

03Overall results: uncertainty at the primary endpoint and qualified early formation signals

We use two-sided asymptotic rank tests without continuity correction, a convention that matches the printed results. All seven selected P-value checks reproduce the paper's displayed precision. Week-20 CTx gives 0.611459 versus the printed 0.611; week-2 P1NP gives 0.020278 versus 0.020. These checks support the inspected data and method conventions; they do not validate the original measurements, every table or all missing-data procedures.

Median week-20 CTx changes were −7.7% in controls and −4.1% with D+Q, a treatment-minus-control difference of about +3.6 percentage points. Our within-arm bootstrap gives an approximate 95% interval of −15.5 to 12.4 points, compatible with changes in either direction. This is not proof of no effect or equivalence.

P1NP rank comparisons yielded nominal signals at weeks 2 and 4, without a clear difference at week 20. The estimands matter: a rank test compares distributions, not necessarily just medians. At week 4, the median difference was about +16.2 points, but its bootstrap interval crossed zero. That is not inconsistent with a rank-test P of about 0.024. The source arm includes a change of approximately +1,480%, also recorded as an extreme in registry results. We retained it rather than treating it as an error; it makes the distinction between mean and median results especially relevant.

The three overall BMD comparisons did not clearly detect differences. Whether a transient formation-marker change becomes sustained bone gain or fewer fractures requires those outcomes to be measured; an early P value cannot establish that translation.

04High p16: a signal, but the original “formal interaction test” needs correction

“Significant in the high stratum, nonsignificant in the low stratum” does not show that the effects differ. We retain the question—does the treatment-control difference change across p16 strata?—but revise its analysis.

The original code shuffled treatment labels within each stratum. That simulates the stronger null of no treatment effect in either stratum; it is not directly an exact test of the general hypothesis that two nonzero effects are equal. We now resample within each of the four cells to estimate the difference between median treatment effects, with a separate unequal-variance mean contrast as sensitivity analysis. These small-sample intervals and tests remain approximate.

For radius BMD, the difference between median effects is about +5.14 points. Its approximate pointwise 95% interval is 0.95–8.05; the Bonferroni interval allowing for nine comparisons is −0.70–9.23. The mean contrast is about 4.53 points, with unadjusted P≈0.0067 and Holm-adjusted P≈0.060 across nine tests. This supports further investigation of effect modification, without validating T3 as a treatment-selection rule. It does not prove the absence of heterogeneity either.

The week-2 P1NP median interaction is about 20.8 points, with a wide interval crossing zero. A specific coverage discrepancy also matters: the overall D+Q list has 30 values, but its two v5 strata total 29. Allowing for three-decimal rounding identifies one unlisted value, −7.71513%. If the overall and complete T3 lists are correct, adding it to the lower stratum changes the median interaction to about 23.0 points. This is an explicitly assumed source-reconciliation sensitivity analysis, not a recovered participant identity or a correction written into the source data.

The earlier claim that radius BMD remained significant after 18 comparisons arose from a software default. At the 11/8 sample-size cell, the method automatically switched to an exact rank test, producing P=0.00254, while other cells used an approximation. The consistent convention matching the paper gives P=0.003852; the minimum Holm or BH value across 18 comparisons is about 0.069. The 18 within-stratum comparisons and nine interaction comparisons are distinct test families. Neither adjustment is a complete multiplicity correction for the entire trial.

05Control trajectories and another assay do not independently refute the signal

Median T3 radius-BMD changes were +1.2% with D+Q and −1.5% in controls; lower-stratum controls changed by about +2.3%. The original article called the high-stratum decline anomalous and discounted the treatment effect on that basis. That attribution was too strong.

Different control trajectories may reflect prognosis or sampling variation. If treatment changes the course of a genuinely higher-risk stratum, that itself could constitute a larger benefit. Without an external reference or replication, these same control data do not split the randomized comparison into “real benefit” and “abnormal controls.”

The assay sensitivity analysis also has a specific scope. Public figures for variant 1+5 provide CTx and P1NP, without corresponding radius-BMD results. Some estimates and group counts change and nominal significance is lost, but a change in significance is not a formal test of an assay difference. We withdraw the claim that the radius signal disappeared under a different assay. Farr, Extended Data Fig. 5

The workbook actually contains age and continuous p16 values for 60 trial participants, plus absolute baseline clinical measurements. What is missing is a verified participant key linking these lists to subsequent outcomes. “Cannot reliably link these data” should not be rewritten as “no continuous or baseline values exist.” We did not construct a participant-level ANCOVA by assuming row-order linkage.

06Can a nonsignificant SASP panel rule out clearance?

The T3 data contain paired baseline and day-14 SASP measurements for 21 participants. Among 35 analytes, no between-arm change comparison reaches nominal P<0.05 under the consistent rank method. Log-change ratios and intervals also show no stable multiplicity-adjusted signal. For example, the ratio of geometric changes, D+Q relative to control, is about 1.12 for sclerostin (95% interval 0.95–1.33) and 0.72 for MMP9 (0.47–1.11). Precision differs across analytes.

This does not prove that all 35 factors were unchanged, nor is it a direct test of tissue clearance. Blood concentrations depend on secretion and clearance across tissues and cell types. Local cell numbers, secretion per cell and contributions from other cells may change together. A day-14 sample may also miss a transient early response. Four of 35 baseline comparisons are nominally higher in T3, with none retained after FDR correction; that does not by itself invalidate p16 as a construct.

Another mechanism remains plausible. Dasatinib can directly alter differentiation, signalling and resorption in bone-cell experiments. Farr's authors argue that its short plasma half-life and later bone responses better fit persistent effects after cell clearance. That is a meaningful argument, but downstream cell states can outlast drug exposure; it does not exclude persistent direct pharmacology. This trial cannot quantify how much each path contributes. Direct bone pharmacology

More discriminating evidence would measure multiple markers near the relevant tissue, at suitable times, alongside cell numbers and function, with controls that help distinguish clearance from other pharmacology. One blood panel cannot adjudicate the entire causal chain.

07Animal evidence includes benefits and a direct repair cost

Baker 2011 used a genetic clearance tool in a progeroid model to delay some phenotypes. Baker 2016 extended the work to normally ageing INK-ATTAC transgenic mice. Pooled-sex median lifespan increased by 27% and 24% in two genetic backgrounds, with differences in background, diet and sex-specific results. These experiments support particular genetic clearance strategies; they do not establish benefits of arbitrary drugs in humans. Baker 2011, Baker 2016

Xu 2018 separately studied functional impairment after senescent-cell transplantation and D+Q treatment of naturally old mice. The approximately 36% increase refers to remaining median survival after treatment began at 24–27 months, with 71 D+Q mice and 76 controls. It is not a one-third increase in whole lifespan, and the transplantation and survival cohorts should not be conflated. Xu 2018

Khan and colleagues' human temporal-lobe-epilepsy tissue associations and mouse intervention experiments identify another disease context worth testing. More markers in human tissue plus treatment effects in mice do not together establish efficacy in human epilepsy. Epilepsy study

An acute wound model provides more direct evidence of harm from clearance itself. Demaria and colleagues found that senescent fibroblasts and endothelial cells appearing early after injury helped wound closure. Clearing them in the relevant mouse model delayed healing; PDGF-AA supplementation rescued aspects of the process. This supports a repair cost at particular times and tissues, not a rule that every acute senescent cell must be retained or a direct human contraindication. Demaria 2014

Developmental evidence addresses a different intervention: p21 loss blocks some programmed senescence, with compensatory apoptosis and developmental abnormalities. This is not a trial of a senolytic in older adults. Separating development from the liver-cancer and fibrosis experiments makes the list of demonstrated harms more accurate—and more limited. Muñoz-Espín 2013

08Human evidence and safety depend on the drug, disease and route

Sources are Justice, Hickson, Nambiar, the final Gonzales paper, the UBX0101 announcement and BEHOLD. The sponsor's 2025 ASPIRE announcement additionally states that UBX1325 failed the prespecified noninferiority criterion against aflibercept for the primary average vision endpoint at weeks 20 and 24. Selecting a more favourable individual visit does not replace that result. ASPIRE announcement Local eye-disease findings do not substitute for evidence of systemic healthy ageing. This is not an exhaustive inventory of trials.

In Farr, 23/30 D+Q participants versus 5/30 controls reported nonserious adverse events: a 60-point risk difference, with an approximate Newcombe 95% interval of 36–75 points. Open-label expectations may affect reporting, but cannot be assumed to explain the whole difference. One D+Q participant withdrew because of QTc prolongation. Posted results report no serious events in either 30-person arm; this does not establish long-term safety. The discontinued 14-person fisetin arm separately reports one person with a serious event and must not be pooled into a D+Q drug-risk count. Registered safety results

The official dasatinib label includes myelosuppression, bleeding, fluid retention and QT-prolongation risks. Oncology regimens and patient circumstances cannot be transferred directly to intermittent experimental treatment, but known risks do not disappear because a small short trial observes no instance of a particular event. Official label

09Where the answer should stop

Clearance strategies can improve function or survival in some states of persistent pathological accumulation. They can also remove cells needed for short-term repair or incur the drug's own costs. Chronic versus acute settings, genetic tools versus drugs with multiple targets, and local disease versus healthy populations all require specific experiments. Intermittent clearance is not itself a guarantee of benefit.

Farr's high-p16 skeletal signal deserves replication but does not establish patient selection. Circulating SASP neither confirms nor refutes tissue clearance; recalculation cannot remove source-table and protocol limitations. These studies have not established human lifespan or broad healthspan benefits, and they do not support deriving a treatment regimen here.

This is a public correction made with the results already known. The original snapshot, methods and reasons for revision remain archived; no new preregistration is claimed. The reproducibility package contains pinned input hashes, analysis code, aggregate results and figures. Original individual plotted values are fetched from the publisher and are not redistributed in the package.