Partial Reprogramming: Younger Readings, Preserved Identity, Better Function?

01Our answer first

"Younger" holds at the molecular-readout layer, but the real magnitude is smaller, far more variable, and clock-version sensitive than the headline: 4–5 weeks after withdrawal, the median Horvath1 methylation-age offset of transiently reprogrammed cells versus same-day controls is about −9 years, with a per-sample range of −37 to +17 years and a donor×day range of −30.3 to +9.3; only fibroblast/skin-trained clocks move — blood-trained Hannum and PhenoAge and the stem-division clock EpiTOC are flat. A deeper confound: methylation clocks substantially track the dedifferentiation process itself — during exposure, clock readings co-move with identity-axis position (Spearman r = −0.33, p = 0.007), i.e. the clock is partly a "reprogramming-progress meter". At the identity layer the answer is "largely reacquired, not bit-perfect": fibroblast markers are fully restored, but a +0.25–0.30 SD residual elevation of the pluripotency score persists after withdrawal — and the same residual appears in the "exposed but failed to reprogram" arm, making it a shared OSKM-exposure scar rather than a rejuvenation signature. The function layer does not exist in this dataset: GEO carries no functional assays, and the paper-level scratch-migration result was variable across cells and — by the paper's own report — uncorrelated with the clock readings. The safety layer has no tumour, karyotype, or long-term endpoints; the harm boundary comes from the in-vivo literature: sustained OSKM produces teratomas, premature termination of reprogramming drives cancer, hepatic/intestinal expression causes premature death. Calling partial reprogramming "proven rejuvenation" exceeds the evidence; calling it "merely a measurement artefact" ignores the combination of a residual post-withdrawal clock offset and restored identity — the most defensible statement is that a population of cultured cells shifted several molecular readouts in a younger direction while largely retaining identity, with a central magnitude of a few to ~20 years rather than "30", and with the functional and human-meaning extrapolation still unestablished.

Scope is limited to what this public dataset can arbitrate at each evidence layer. This article describes evidence and provides no reprogramming protocol, dosing, or operational guidance; OSKM, doxycycline, and exposure durations appear only as facts of the study's registered design and GEO metadata.

02First, split "younger" into four layers

The headline conflates at least four separable meanings of "young", each of which can hold or fail independently in this kind of experiment:

1. Readouts: do molecular "age meters" — methylation clocks, transcriptome clocks — read lower? This is where the "~30 years" claim lives.
2. Identity: is the cell still a fibroblast? Reprogramming is identity rewriting by definition — "younger" and "became a different cell" must be judged separately.
3. Function: does the cell do its job better — migration, matrix secretion, stress response? That is what "young" means biologically.
4. Stability & safety: does the state persist, and at what cost (uncontrolled dedifferentiation, tumour risk, genomic damage)?

Four competing explanations of "younger" need arbitration:

• A. Local state restoration: cells reset molecular state to a younger position while retaining identity — the paper's reading, and the version the "rejuvenation" narrative needs.
• B. Dedifferentiation: cells simply travel toward pluripotency and clock reversal is a by-product — full reprogramming to iPSC zeroes the methylation clock, but that is "becoming embryo-like", not "rejuvenation".
• C. Selection / proliferative replacement: the measured "young cells" may be a subpopulation that retained a younger state or was enriched by sorting, not cells that each became younger. The failed and negative-control arms arbitrate this.
• D. Measurement-algorithm response: clocks are linear models over specific CpGs — if reprogramming rewrites those loci, readings move because "the algorithm is sensitive to reprogramming", not because "biological age" changed. Tests: do multiple clock versions agree, and are clock readings coupled to dedifferentiation position?

Gill's design happens to put all four explanations on data: it contains both "during-exposure intermediate" and "4–5-week post-withdrawal final" phases, a negative control plus an "exposed but failed" arm, and iPSC references.

03Data and calibration: what can and cannot be computed

GSE165180 is the SuperSeries for the Gill study, with four subseries: GSE165176/178 are the full Sendai reprogramming trajectory (RNA + methylation), and GSE165177/179 are the MPTR transient-reprogramming experiment at the core of this article (RNA + methylation). In MPTR, fibroblasts carrying a doxycycline-inducible polycistronic OSKM cassette are exposed for 10/13/15/17 days, sorted by surface markers into "successfully reprogrammed" (CD13−/SSEA4+) and "failed" (CD13+/SSEA4− — expressed the reprogramming cassette but never SSEA4) populations alongside a mock-treated negative control, then cultured without inducer for 4–5 weeks before measurement. Gill 2022 GSE165180

Three calibration facts first:

• The independent biological unit is 3 donors (O1/O2 = 53 y, O3 = 38 y), not 96 samples. Each condition×day cell holds only n = 2–6 samples, and donor composition is unbalanced across days: the 13-day reprogrammed group contains only O2+O3 (n = 4), the 15-day group only O3 (n = 2) — the paper's "optimal 13–15-day window" is partly made of donor composition, which we return to below.
• Cross-modal alignment is group-level only. Methylation and RNA were measured on different cell batches; they share donor×day×condition group labels, not cell identity — we analyze group medians jointly and make no per-cell pairing claims.
• Clock implementations are version-sensitive. The input matrix is the authors' NOOB-normalized EPIC betas; Horvath1 loses 19 CpGs on this array (noted by the paper itself; our coverage audit confirms 334/353), which we impute with row means; the paper's implementation also applies Horvath gold-standard normalization. All interpretation is therefore restricted to within-dataset contrasts and directions — no absolute cross-study age claims. Clock implementation

04Readout layer: the clock does rewind — conditionally

During exposure: the clock falls monotonically with dedifferentiation

Intermediate (pre-withdrawal) samples show a clean relation: the longer the OSKM exposure, the lower the methylation age and the closer the identity axis moves toward the iPSC endpoint. In the reprogrammed arm, Horvath1 falls from ~37 y at day 10 to ~16 y at day 17, and Horvath2 from ~27 to ~11; the identity axis's per-day medians sit at 0.78–0.98 over the same window, highest at the last day (per-sample range 0.76–1.08). Neither line moves in the control arm. The iPSC reference reads ~0 years — the clock direction itself is correct: complete reprogramming does zero it. [Figure 1]

But this decline does not require "rejuvenation" as its explanation: cells on the iPSC trajectory show falling clock readings as a matter of course — the paper itself uses the full-reprogramming trajectory to locate when the reversal accrues (Figure 1A: a cumulative ~20 y reversal by day 10 and ~40 y by day 17), not to claim the during-exposure readings as the post-withdrawal MPTR result.

After withdrawal: a real but highly variable residual offset

The discriminating observation is the final-phase samples, taken 4–5 weeks after withdrawal — when cells are off inducer and morphology and markers have largely returned to fibroblast. Whatever offset remains on the clock is the candidate evidence for persistent rejuvenation. There is one — but it is far more conservative than the headline:

• Per-day median offsets (reprogrammed − same-day control): day 10 −6.4 y, day 13 −22.9 y, day 15 −9.8 y, day 17 −0.1 y (Horvath1).
• Per-sample range: −36.6 to +16.8 y (Horvath1); donor×day range −30.3 to +9.3 y (Horvath1) and −33.7 to +19.0 y (Horvath2) — there are post-withdrawal cells that read older (Horvath1, donor O2, 17-day arm median +9.3 y vs control).
• Pooled medians: reprogrammed −8.6 y, failed −4.1 y, control 0 (Horvath1). The failed-arm offset is version-sensitive too: on Horvath2 its median is ≈0 with a per-sample spread of ±25 y. [Figure 5]

The paper's "~30 years" corresponds to the favourable end of this distribution (deepest per-sample offset −37 y; deepest donor×day is −33.7 y on Horvath2, in donor O1's 17-day cell) plus its bespoke transcriptome clock — not the central tendency. And since the day-13 group contains only O2+O3 and day-15 only O3, the "optimal window" is entangled with donor composition — the data cannot separate "day 13 is truly optimal" from "these two donors happened to respond strongly in that cell".

Version sensitivity: only fibroblast-trained clocks move

Of five clocks, only two — Horvath1 (multi-tissue) and Horvath2 (skin+blood) — show consistent post-withdrawal offsets; blood-trained Hannum and PhenoAge and the mitotic clock EpiTOC are flat or inconsistent. [Figure 2]

The paper itself notes this ("these other epigenetic clocks were not trained on fibroblast data"), but the caveat is lost in transmission: the widely quoted line is "methylation age reversed 30 years", not "only the two skin/fibroblast-relevant clocks move, with per-cell variability". Version sensitivity does not negate the change — it bounds where it holds: "methylation age" in this dataset is not one number but a set of model-specific numbers, of which only the skin/fibroblast-adjacent ones move.

05Identity layer: toward iPSC, then largely back — with a scar shared by both exposed arms

Averaging z-scores of 13 fibroblast markers (COL1A1, COL3A1, DCN, LUM, S100A4, VIM, …) and 12 pluripotency markers (POU5F1, NANOG, SOX2, LIN28A, DNMT3B, …), then normalizing the difference between the Sendai-trajectory fibroblast endpoint (=0) and the d54 SSEA4+ iPSC endpoint (=1), gives an identity axis: [Figure 3]

• During exposure: the reprogrammed arm's axis rises to a per-sample range of 0.76–1.08 (per-day medians 0.78–0.98) — cells do move toward pluripotency, about a quarter to half of the fibroblast→iPSC distance (the transient-experiment iPSC references read 2.29 on this axis, a purer state than the Sendai endpoint anchors). The failed arm reaches only 0.06–0.19; controls sit at −0.06 to 0.09.
• After withdrawal: group medians land in a narrow 0.26–0.39 band — reprogrammed 0.30, failed 0.32, control 0.26, donor baseline 0.39 (per-sample range 0.21–0.44) — with no between-arm difference in fibroblast marker score. At marker resolution, identity is largely reacquired — the paper's central claim holds at this layer.
• But not bit-perfect: the pluripotency score remains +0.29 SD above control after withdrawal — and the same elevation appears in the failed arm (+0.25 SD). Post-withdrawal cells keep a faint "touched by OSKM" transcriptional signature regardless of whether reprogramming succeeded. The paper itself observed a moderate transcription-age reduction in the failed arm and attributed it to reprogramming-factor expression alone — consistent with our reading of a non-specific exposure signature.

This cuts against both extremes: "identity completely intact" is wrong — a measurable residual signature exists; "cells became something else" is also wrong — on the marker axis they return to the control band. The honest statement: identity is largely reacquired with a non-specific OSKM-exposure signature retained; whether a residual subpopulation (e.g. a few still-intermediate cells lifting the group mean) hides inside the bulk measurement is indistinguishable here — which is exactly where explanation C (selection/subpopulation) cannot be fully excluded.

06The clock–identity coupling: the core confound

Putting group-level medians on one plot — identity axis on x, Horvath1 on y — exposes the confound's shape: during-exposure reprogrammed cells cluster along the upper-right→lower-right diagonal (identity 0.76–1.08, age 12–40 y); post-withdrawal points return to the upper-left region (identity ~0.3, age 33–78 y), overlapping controls and failures. Spearman r = −0.33 (p = 0.007, n = 68 group-level cells). [Figure 4]

Two conclusions. First, clock readings are significantly coupled to dedifferentiation position: explanation D (measurement-algorithm response) is partly true — in this dataset the methylation clock is substantially a "reprogramming-progress meter", and much of the during-exposure "youth" is the coordinate system moving with the cell. Second, coupling is not everything: post-withdrawal cells return to the control identity band while retaining a clock offset (median −8.6 y) — that residual is the candidate evidence for persistent rejuvenation, but it is modest, variable per cell, and visible only on fibroblast-trained clocks.

The paper's own explanation for the weaker effect at 15/17 days is that "extended reprogramming may make reversion more difficult and cellular stresses 're-age' the methylome during the process" — consistent with our readings, but indistinguishable from a simpler account: the deep day-13 offset is partly driven by donor composition and a small cell (n = 4). Both explanations remain live.

07Function layer: absent from the data, and the paper-level evidence discounts itself

The GEO matrices contain no functional assays — this layer cannot be verified from data. The paper-level functional evidence has three pieces, each weighed by the paper's own text:

• Migration (scratch wound): added at reviewer request. Middle-aged controls migrated more slowly than young controls; transient reprogramming "partially restored" migration speed — but the paper states that "the individual responses were quite variable and in some cases migration speed was improved and in other cases it was unaffected" and that it "did not appear to correlate with other aging measures such as transcription and methylation clocks". That is the paper's own report of readout–function dissociation: the deepest clock-reset cells are not necessarily the better-migrating cells.
• Collagen: collagen I/IV transcripts and immunofluorescence protein levels in reprogrammed cells moved toward "youthful levels" — directionally consistent, still a group-level measurement, and the paper itself notes the collagen I transcript restoration was not significant (small age difference, few samples).
• Morphology: roundness falls then recovers to the starting fibroblast state — consistent with the identity-axis account.

One in-paper negative result deserves keeping: the telomere-length clock shows MPTR does not extend telomeres and in some cases slightly shortens them — "rejuvenation" does not rewind all molecular attributes in step.

The honest functional statement: readouts changed and identity largely held, but "the cells function as younger" was never directly measured in this dataset; the paper-level functional evidence is small-n, variable, and decoupled from the clock readings.

08Stability & safety layer: molecular state persists 4–5 weeks; tumour endpoints absent

What this dataset directly observes about stability: the final-phase samples are themselves measured after 4–5 weeks of post-withdrawal culture — so "the molecular state persists at least 4–5 weeks after withdrawal" is an observation, not an inference. Longer-term stability, karyotypic integrity, and tumorigenic potential are not measured at all.

The harm boundary comes from the in-vivo literature (abstract-level reading):

• Sustained expression → teratomas: Abad 2013 produced teratomas and iPS cells with totipotency features under continuous in-vivo OSKM expression — the direct consequence of "not stopping in time". Abad 2013
• Premature termination → cancer: Ohnishi 2014 reported that incomplete in-vivo reprogramming leads to cancer development through altered epigenetic regulation. Ohnishi 2014
• Hepatic/intestinal expression → premature death: Parras 2023 linked in-vivo reprogramming to premature death via hepatic and intestinal failure. Parras 2023
• Damage/senescence signals are double-edged: Mosteiro 2016 showed tissue damage and senescence provide critical signals for in-vivo reprogramming — aged/damaged tissue environments amplify reprogramming, complicating the dose-context relation. Mosteiro 2016

Genuine favourable evidence must be placed symmetrically: Lu 2020 expressed OSK (minus c-Myc) via AAV in retinal ganglion cells, restored vision, and reported no tumour increase under long induction; Ocampo 2016 extended lifespan in progeroid mice with cyclic OSKM; Chondronasiou 2022 rejuvenated naturally aged tissues multi-omically with a single transient cycle; Browder/Chondronasiou 2024 extended lifespan in aged mice with gene-therapy partial reprogramming. Lu 2020 Ocampo 2016 Chondronasiou 2022 Browder 2024 Factor sets, dose windows and tissue specificities differ enormously — in-vivo results cannot extrapolate to "a clock offset in a dish = safe", nor can in-vivo efficacy be taken as proof that the in-vitro mechanism is understood.

09Verdict on the four competing explanations

The four-layer verdicts, summarized in Figure 6:

10What we know, and what we don't

Known:

1. In the Gill dataset, methylation-clock reversal is real and still measurable 4–5 weeks after withdrawal — but the central tendency is a ~9-year median offset, not the "~30 years" in circulation; per-sample range −37 to +17 y (Horvath1), including cells that read older. That is still markedly deeper than earlier transient reprogramming — Sarkar 2020's initiation-phase transient transfection of the Yamanaka factors managed only ~3 years of methylome reversal Sarkar 2020.
2. Clock movement is partly a dedifferentiation gauge — it falls with the identity axis during exposure; the discriminating residual only counts after withdrawal.
3. Fibroblast identity is largely reacquired at marker resolution; the residual pluripotency signature is a shared OSKM-exposure scar, present even in the failed arm.
4. The "optimal 13–15-day window" is entangled with donor composition and small cells and cannot be separated from noise.
5. Functional improvement was never directly measured; the paper-level migration evidence is variable and decoupled from clock readings; telomeres do not rewind.
6. The in-vivo literature gives a concrete harm list — teratomas, cancer after premature termination, hepatic/intestinal failure — alongside genuine benefit reports stratified by factors, dose and tissue.

Unknown:

• The per-cell joint clock–identity trajectory (cross-modal measurements are not same-cell);
• Whether the residual pluripotency signature carries functional consequences or marks a risk subpopulation;
• Long-term (>4–5 week) molecular state, karyotype, tumorigenicity;
• Any extrapolation from these in-vitro readings to human function or healthspan;
• Whether other transient-reprogramming protocols decompose the same way (this analysis covers MPTR only).

11Methods and update notes

• Data: GSE165180 (GEO, downloaded 2026-09-18), 96 methylation + 95 RNA samples; processed matrices (NOOB betas / Log2 RPM), not raw idat/FASTQ.
• Implementation: five clocks per biolearn coefficients (Horvath1 334/353 CpGs, row-mean imputed; others 65–514 of 71–514); identity axis is an analyst-defined 25-marker z-score axis anchored on the Sendai trajectory.
• Statistics: all contrasts are group-median differences with nominal Mann–Whitney p; the independent unit is 3 donors and per-cell n is 2–6 — p-values are exploratory flags only; cross-modal analysis joins group medians, not cells.
• Code and outputs: research/calculations/partial-reprogramming-identity/ (analysis plan, reanalysis script, per-sample outputs, joint table, log) and research/assets/partial-reprogramming-identity/ (figures and reproducibility bundle).
• Self-review: performed by the same agent that authored the article; score and revision record are in the reviews field of the research record; not an independent review.
• Competing interests note: Gill 2022 senior author Wolf Reik's listed affiliation is the Altos Labs Cambridge Institute — partial reprogramming is Altos's core program; the study is in-vitro methodology and does not constitute product evidence.

12Sources

• <a id="source-s1"></a>s1: Gill D, Parry A, Santos F, et al. Multi-omic rejuvenation of human cells by maturation phase transient reprogramming. eLife 2022;11:e71624 (PMC9023058, full text including reviewer exchange)
• <a id="source-s2"></a>s2: GEO SuperSeries GSE165180 (subseries GSE165176–179), "Multi-omic rejuvenation of human cells by maturation phase transient reprogramming", downloaded 2026-09-18
• <a id="source-s3"></a>s3: biolearn clock coefficients (Horvath1/Horvath2/Hannum/PhenoAge/EpiTOC) + Horvath 2013 Genome Biology 14:R115 method paper
• <a id="source-s4"></a>s4: Sarkar TJ et al. Transient non-integrative expression of nuclear reprogramming factors promotes multifaceted amelioration of aging in human cells. Nat Commun 2020 (PMC7093390, full text)
• <a id="source-s5"></a>s5: Abad M et al. Reprogramming in vivo produces teratomas and iPS cells with totipotency features. Nature 2013 (abstract-level)
• <a id="source-s6"></a>s6: Ohnishi K et al. Premature termination of reprogramming in vivo leads to cancer development through altered epigenetic regulation. Cell 2014 (abstract-level)
• <a id="source-s7"></a>s7: Parras A et al. In vivo reprogramming leads to premature death linked to hepatic and intestinal failure. EMBO Mol Med 2023 (abstract-level)
• <a id="source-s8"></a>s8: Mosteiro L et al. Tissue damage and senescence provide critical signals for cellular reprogramming in vivo. Science 2016 (abstract-level)
• <a id="source-s9"></a>s9: Lu Y et al. Reprogramming to recover youthful epigenetic information and restore vision. Nature 2020 (PMC7752134, full text)
• <a id="source-s10"></a>s10: Ocampo A et al. In vivo amelioration of age-associated hallmarks by partial reprogramming. Cell 2016 (PMC5679279, full text)
• <a id="source-s11"></a>s11: Chondronasiou D et al. Multi-omic rejuvenation of naturally aged tissues by a single cycle of transient reprogramming. Aging Cell 2022 (PMC8920440, full text)
• <a id="source-s12"></a>s12: Browder KC / Chondronasiou D et al. Gene therapy-mediated partial reprogramming extends lifespan and reverses age-related changes in aged mice. Cell Reprogram 2024 (PMC10909732, full text)
• <a id="source-s13"></a>s13: Legacy site article 122 "Cellular reprogramming: Yamanaka factors for organ rejuvenation" — the article this study revises (snapshot archived)