Partial reprogramming: what becomes younger—readouts, identity, or function?

Molecular states changed, and the selected marker profiles moved back toward fibroblast-like states after withdrawal. The magnitude depends on donor, exposure duration, clock and marker panel. These analyses cannot partition the changes into “true age restoration” and dedifferentiation, or convert a culture’s clock change into years of human rejuvenation. Functional assays provide useful but limited signals; a dependable relationship between clock change and functional improvement has not been established here.

Correction, 19 September 2026: This review corrected missing-CpG handling, the PhenoAge intercept, cross-matrix RNA standardization and the unit of statistical inference. We withdraw the claims that “about 9 years refutes 30 years,” that a significant correlation establishes dedifferentiation as the dominant explanation, and that residual markers establish a shared OSKM scar. Three PMID links and one author attribution were also wrong. All figures, values and the reproduction package have been updated. The original author’s review is retained as history; Codex is responsible for this revision.

01What was measured, and what is an independent observation?

The MPTR cells came from donors aged 38, 53 and 53 years. The study compared mock-infected negative controls subjected to corresponding culture and sorting procedures with factor-exposed populations sorted by surface markers. The population called “failed” did not enter the same selected state as the reprogramming population. This is a classification after treatment, not a randomized failure intervention. Design and sorting

The transient-reprogramming data contain 96 methylation and 95 RNA samples. Another 124 Sendai-reprogramming RNA reference samples define a common scoring coordinate system. Matrix columns represent cultures or sorted populations, not independent people. GEO SOFT metadata identify donor, experiment and duration. For example, “iPSC 13” and “iPSC 21” are sample names; their metadata specify 51 days, rather than exposure durations of 13 and 21 days. GEO

Each primary contrast subtracts the negative control with the same donor, experiment, exposure duration and collection phase. Repeated experiments are then averaged within donor and duration; donors receive equal weight within a duration. Experiment 1 collected cells four weeks after withdrawal and experiment 2 after five weeks. These are endpoints in separate experiments, not a continuous four-to-five-week trajectory in the same cells.

The final reprogrammed groups at exposure durations of 10, 13, 15 and 17 days contain three, two, one and three donors, respectively. A difference between duration groups therefore cannot directly identify an optimal window. We report descriptive donor contrasts, without population significance tests based on repeated cultures or selecting the most favorable condition as a general effect.

02Clock changes depend on the model and comparison

All five clocks use the authors’ processed NOOB methylation values. The corrected calculation retains model intercepts and transformations. PhenoAge has 513 CpGs plus an intercept of 60.664: the intercept is not a missing 514th probe. Horvath 2013, Hannum and EpiTOC lack 19, six and 31 required probes in this matrix. These are now explicitly filled using the reference/mean hybrid rule from a pinned Biolearn version; an entirely absent probe cannot be imputed from its own nonexistent sample mean. Implementation and coefficients

The table shows the mean donor difference between reprogrammed cells and matched negative controls after withdrawal. Negative values mean lower model readings. Parentheses show the minimum and maximum donor contrasts, not confidence intervals.

Some conditions show lower clock readings, with substantial donor and model variation. All 391 Skin & blood probes are observed, so that clock does not require filling whole absent probes. Complete inputs still do not validate it as a measure of functional age after reprogramming.

Blood-trained clocks cannot be described as universally flat. The four post-withdrawal PhenoAge mean contrasts are −17.0, +12.3, +7.3 and −9.5 years, with inconsistent directions. EpiTOC has score units, not years. Its small numerical changes cannot be compared directly with year-valued clocks. Divergent responses do not identify which model represents a cell’s “true age.”

Implementation checks compared five clocks × 96 samples × two probe-processing choices: all 960 values agreed with the pinned upstream formulas to floating-point precision. A sensitivity analysis masked probe values with detection P > 0.01 and imputed again. The maximum individual-sample change was about 1.27 years for Horvath 2013 and 0.015 years for Skin & blood. This checks the selected probes; it neither repeats raw-array quality control nor validates the biological meaning of a clock change.

Why withdraw “about 9 rather than 30 years”? The paper’s approximately 30-year interpretation involves its transcriptomic and methylation analyses. Our earlier single number combined different donors, durations and an incorrectly implemented calculation. Different preprocessing, models, summaries and contrasts cannot directly refute one another. We did not rerun the paper’s complete transcriptomic age pipeline or establish identical clock processing. Gill’s methods specify minfi/NOOB; they do not justify our earlier assertion that the authors applied an additional “gold-standard normalization.” The defensible contribution is the explicit matched comparison above, with its limits. Methods and Figures 3–4

03Identity requires one common reference scale

We selected 13 fibroblast and 12 pluripotency markers. Each gene is transformed using its mean and standard deviation in the Sendai reference matrix, and the same frozen transformation is applied to the MPTR matrix. The difference between marker panels is anchored at the mean starting fibroblast reference, zero, and the mean iPSC endpoint, one. This exploratory expression axis is not a cell-purity estimate: 0.3 does not mean that 30% of cells have reprogrammed.

After correction, the median axis position of the MPTR reprogramming population changes from approximately 0.26 during expression to −0.23 after withdrawal; the MPTR iPSC reference is approximately 1.02. Negative controls also have negative values. Culture and batch differences between series remain, and not every fibroblast population must equal zero. The direction is consistent with reacquisition of fibroblast-like features, but it cannot establish complete restoration in every cell. Original identity analyses

Cross-modal joins connect population measurements with matching culture-condition labels, not trajectories of individual cells. Within the three donors, descriptive correlations between the Skin & blood clock and identity axis are approximately −0.90, −0.83 and −0.55 during expression, and 0.06, 0.04 and −0.18 after withdrawal. These points retain condition and time structure, and combine controls with different sorted states. They describe concurrent movement during expression; they cannot establish mechanistic dominance or the fraction of age change explained by dedifferentiation. The earlier P value based on 68 supposedly independent group entries has been withdrawn.

04Residual pluripotency scores depend on the panel

The original article called elevated post-withdrawal pluripotency scores a shared OSKM “scar.” However, the panel includes POU5F1 and SOX2, two delivered reprogramming factors. These gene-level measurements cannot reliably distinguish endogenous expression from vector-derived expression.

For reprogrammed cultures, the four post-withdrawal donor-mean contrasts range from approximately +0.17 to +0.42 using all 12 markers. Excluding those two factors reduces the range to approximately +0.001 to +0.096 for the remaining ten. The failed population is also sensitive to panel choice. These units are means of gene scores standardized to the reference, not a calibrated biological effect-size SD.

This does not prove that vectors cause the entire difference, or that the smaller residual is meaningless. It shows that a sustained pluripotency interpretation depends on marker definition. Because successful and failed populations were sorted by post-treatment state, the failed group cannot provide a causal upper bound on selection effects.

These explanations can coexist. The design does not assign a percentage contribution to each.

05Functional evidence exists, but clocks do not substitute for it

Gill’s paper reports collagen I/IV expression and immunofluorescence, together with scratch-wound migration assays. Collagen I transcript restoration did not meet the reported statistical threshold; protein measurements shifted in a younger-associated direction. Migration improved in some paired culture populations but not clearly in others. The authors also noted that migration changes did not show consistent correspondence with the other age-related measurements. Figure 3, supplements, methods and review correspondence

These are functional and phenotypic signals. Their absence from the GEO molecular matrices does not mean that the study had no functional evidence. Conversely, a small sample without an apparent correlation does not establish independence between clocks and function. We did not obtain functional raw data that could be reliably joined to these molecular conditions, did not re-estimate a clock–migration correlation, and cannot extrapolate a culture assay to human tissue repair or healthspan.

Other approaches, including Sarkar and colleagues’ transient non-integrative expression study, also report improvements in age-associated cellular phenotypes. Different cell types, factor combinations and measurement pipelines cannot be ranked by “years rejuvenated” without appropriate common comparisons. Sarkar 2020

06In-vivo benefit and harm depend on the experimental setting

Observing a state four or five weeks after withdrawal does not demonstrate uninterrupted stability during that interval, or resolve long-term tumour and organ risks. Mouse studies provide boundaries, rather than direct replications of the Gill culture experiment.

Tissue injury and senescence can themselves affect the reprogramming environment, further complicating whole-body extrapolation. Mosteiro 2016 Species, factors, delivery and observation periods differ across these studies. We did not systematically pool them, and this article provides no reprogramming or drug regimen.

07Scope and reproducibility

We used processed methylation and expression matrices, without rerunning raw IDAT or FASTQ processing or the full transcriptomic age model. Three MPTR donors, sorted bulk populations, missing conditions, reference-series differences and exploratory markers limit generalization. Correct software does not by itself turn model outputs into verified changes in biological age.

The reproduction package contains corrected code, input sources and hashes, matched outputs, marker references and figures. Obtain the large original matrices from GEO using the input manifest. Earlier code and conclusions are retained in internal revision snapshots; obsolete significance tables are excluded from current attachments. Source checking was updated on 19 September 2026. The revising agent reviewed both language editions; this is not independent clinical or human professional review.