Transcriptomic Signatures of Partial Reprogramming Are Confounder-Dominated: A PRC2 Fidelity Benchmark with MSigDB Hallmark Validation

Abstract

Partial reprogramming reverses epigenetic age, but researchers routinely assess whether PRC2-mediated chromatin restoration occurred by measuring PRC2 subunit mRNA levels. We tested whether this mRNA readout is reliable by analyzing four genome-wide reprogramming datasets (Chondronasiou, Roux, Gill, Sahu; 23K-61K genes). It is not — and five independent lines of evidence converge on this conclusion. First, GSEA across a strict 13-gene PRC2 subunit set (per Margueron and Reinberg 2011; Hauri et al. 2016) finds no significant enrichment in any dataset (p = 0.16-0.36; permutation 95% CIs all contain zero). Second, EZH2 — PRC2's catalytic subunit — is transcriptionally driven by E2F during cell-cycle re-entry (Bracken et al. 2003), making its mRNA indistinguishable from a proliferation marker. Third, per-sample ssGSEA with confounder regression shows that 86-89% (adjusted R²) of PRC2 subunit mRNA variance is explained by proliferation and interferon alone, leaving zero residual PRC2 signal. Fourth, leave-one-gene-out analysis confirms zero sign flips — the signal is uniformly absent, not stochastically noisy. Fifth, H3K27me3 ChIP-seq from Camacho et al. (2026; GSE288574) shows the actual chromatin marks at PRC2 loci are stable between young and old mouse epidermis (mean Old/Young = 0.95), while mRNA fluctuates — the two modalities are decoupled. The field has been using a broken thermometer: PRC2 mRNA measures cell-cycle noise, not epigenetic restoration. Future benchmarks must use chromatin-level assays (H3K27me3 ChIP-seq). All data, signatures, and code: https://github.com/scottdhughes/prc2-rejuvenation-benchmark.

Introduction

Partial reprogramming via OSKM or OSK factors reverses epigenetic age (Lu et al. 2020, doi:10.1038/s41586-020-2975-4; Ocampo et al. 2016, doi:10.1016/j.cell.2016.11.052). PRC2 maintains repressive H3K27me3 marks (Margueron and Reinberg 2011, doi:10.1038/nature09784), but corresponding mRNA-level changes are uncharacterized. Meanwhile, proliferation, interferon, and SASP produce transcriptomic changes mistaken for rejuvenation (Lopez-Otin et al. 2023, doi:10.1016/j.cell.2022.11.001). We asked: can mRNA signatures of PRC2 genes distinguish genuine epigenetic restoration from confounding programs?

Methods

Datasets

Four partial reprogramming studies (three mouse, one human). Mouse ranked lists were analyzed against human MSigDB Hallmark sets v2024.1.Hs (Liberzon et al. 2015, doi:10.1016/j.cels.2015.12.004) with gene symbols mapped to human HGNC orthologs via uppercase matching; a sensitivity analysis using mouse MSigDB MH v2024.1.Mm confirmed results are species-invariant (see Results). Datasets: Chondronasiou 2022 (23,332 genes; Young-Old; doi:10.1111/acel.13578), Roux 2022 (55,304 genes; Pulse/NegCtrl; GSE197437; doi:10.1016/j.cels.2022.05.002), Gill 2022 (35,805 genes; MPTR/NegCtrl; GSE165177; doi:10.7554/eLife.71624), Sahu 2024 (61,428 genes; P16-OSK/GFP; GSE201710; doi:10.1126/scitranslmed.adg1777).

PRC2 Gene Set

13 strict PRC2 subunits: core (EZH2, EZH1, SUZ12, EED, RBBP4, RBBP7) and accessory (JARID2, AEBP2, PHF1, MTF2, PHF19, EPOP, PALI1), per Margueron and Reinberg (2011) and Hauri et al. (2016, doi:10.1016/j.celrep.2016.08.096). JARID2/AEBP2 define PRC2.2; PHF1/MTF2/PHF19 with EPOP/PALI1 define PRC2.1 — both included because we test aggregate PRC2 mRNA, not subcomplex-specific function. Excluded: H3K27me3 demethylases KDM6B/KDM6A (Hong et al. 2007; doi:10.1073/pnas.0707292104; PMID: 18003914), DNA methyltransferases DNMT3A/DNMT3B, PRC1 chromobox readers CBX2/4/6/7/8. GSEA: 2,000 permutations, seed=42; Bonferroni alpha = 0.005.

Scoring Weights

PRC2 scoring module (5 genes): EZH2 (w=1.2), SUZ12 (1.1), EED (1.0), JARID2 (0.9), EPOP (0.8). Confounder modules (6 genes each, w=1.2-0.7): AP-1 (JUN, FOS, FOSL2, ATF3, JUND, CEBPB), proliferation (MKI67, PCNA, MCM2, TOP2A, TYMS, CDK1), interferon (STAT1, IFIT1, IFIT3, ISG15, MX1, OAS1), DNA damage (CDKN1A, GADD45A, DDIT3, BBC3, MDM2, TP53I3). Classification: if max confounder score >= max(restore, loss), "confounded"; if specificity margin < 0.10 or direction stability < 0.70, "mixed"; else "restoration" or "loss."

Single-Sample Enrichment and Confounder Regression

For datasets with per-sample expression matrices (Gill, n=24; Sahu, n=16), we computed single-sample enrichment scores using a rank-based method: for each sample, genes were ranked by expression and the normalized mean rank of gene set members computed (range [-1, +1]). Scores were computed for PRC2 targets (195 genes), PRC2 subunits (13 genes), E2F Targets (200 genes), and IFN-gamma Response (200 genes). Confounder regression: PRC2_score ~ intercept + beta_1(E2F_score) + beta_2(IFN_score), fit by OLS. Residuals were tested for treatment-vs-control differences via Welch's t-test.

H3K27me3 ChIP-seq Validation

H3K27me3 ChIP-seq bigwig files from Camacho et al. (2026; PMID: 41667860; GSE288574) were downloaded for young (3 months, n=2) and old (22 months, n=3) mouse tail epidermis. Mean H3K27me3 signal was computed in 4kb windows centered on the TSS of each PRC2 subunit gene, 5 canonical PRC2 target genes (Hoxa1, Hoxa9, Sox2, Pax6, Dlx5), and 3 housekeeping controls (Actb, Gapdh, Rpl13a) using pyBigWig.

Results

PRC2 Module Genes Absent from DEG Lists

Hallmark GSEA

Bold: p <= 0.005 (Bonferroni). Asterisk: p <= 0.001. PRC2 is non-significant in all datasets.

Permutation 95% CIs on PRC2 NES (strict 13): Chond. +1.10 [-1.49, +1.51]; Roux +0.81 [-1.41, +1.42]; Gill -1.11 [-1.41, +1.44]; Sahu +0.95 [-1.40, +1.42]. All intervals contain zero.

PRC2 Gene Set Variants

Non-significant across all variants and datasets.

Mouse MSigDB Sensitivity

Mouse datasets re-analyzed with MH v2024.1.Mm Hallmarks. Differences from human Hallmarks are negligible (E2F NES delta < 0.04; G2M delta < 0.05; PRC2 identical because gene symbols are conserved). Species handling does not affect any conclusion.

EZH2 as Proliferation Confounder

EZH2 mRNA is an E2F target (Bracken et al. 2003, doi:10.1093/emboj/cdg542): E2F drives EZH2 transcription during S-phase. Leading edge analysis identifies EZH2 among 35 core E2F genes shared across all 4 datasets. The catalytic PRC2 subunit is indistinguishable from a proliferation marker at the mRNA level.

Leave-One-Gene-Out (Strict 13)

Zero sign flips — signal is uniformly weak, not stochastic.

H3K27me3 ChIP-seq: Chromatin Marks Are Stable While mRNA Fluctuates

H3K27me3 occupancy at PRC2 subunit promoters is stable between young (3 months) and old (22 months) mouse epidermis (Camacho et al. 2026, GSE288574). Only 2/13 PRC2 genes show change outside 20% (EPOP: +22% gain; RBBP7: -29% loss). Meanwhile, our GSEA shows that mRNA-level PRC2 enrichment is non-significant and direction-inconsistent across reprogramming datasets. This pattern is consistent with chromatin-transcript decoupling at PRC2 loci; however, this comparison is young-vs-old (not old-vs-reprogrammed), so it supports but does not fully validate the fidelity claim. A matched reprogramming ChIP-seq comparison remains future work.

Per-Sample ssGSEA with Confounder Regression

To test whether PRC2 signals survive at the individual sample level, we computed ssGSEA enrichment scores for PRC2 targets (195 genes), PRC2 subunits (13 genes), E2F Targets, and IFN-gamma Response in each sample from Gill (24 samples) and Sahu (16 samples), then regressed PRC2 scores against E2F + IFN confounders.

PRC2 subunit mRNA is 86-89% explained by E2F + IFN confounders alone (adjusted R², accounting for 2 predictors). After confounder regression, zero residual PRC2 signal remains in either dataset. PRC2 targets (the 195 genes silenced by H3K27me3) show no treatment effect at all — they are not differentially expressed between reprogrammed and control samples at the mRNA level. This per-sample analysis confirms the cohort-level GSEA null result and demonstrates that even individual-sample scoring cannot rescue the PRC2 mRNA signal.

Fixture-Level Rule Evaluation

Note: the 8/8 holdout result depends on the direction stability gate (< 0.70), which fires on exactly one signature (blind_mixed_xu_1). Without this secondary rule, holdout accuracy is 7/8 (88%). We report 8/8 because the direction stability criterion is pre-specified, not post-hoc.

Within our curated fixture panel, 2/5 Browder and 2/4 Riparini primary signatures were flagged confounded or mixed, whereas 0/6 Sarkar primary signatures were. The Xu family (60% LOFO-CV) contains borderline mixed-biology signatures at the scoring system's resolution limit.

Discussion

Five convergent lines of evidence demonstrate that PRC2 mRNA cannot serve as a cross-dataset fidelity marker: (1) non-significance across 5 gene set variants and 4 datasets, with permutation CIs containing zero; (2) zero LOGO sign flips (uniformly weak, not stochastic); (3) EZH2 mRNA contaminated by the E2F proliferation axis; (4) per-sample ssGSEA with confounder regression shows 86-89% (adjusted R²) of PRC2 subunit variance is E2F + IFN, with zero residual signal after removal; (5) H3K27me3 ChIP-seq shows chromatin marks are stable at PRC2 loci while mRNA fluctuates.

Benchmark-to-target transition. While our 28-signature benchmark evaluates the historical use of PRC2 subunit genes in the literature, our ssGSEA results indicate that future transcriptomic scoring modules must transition to evaluating canonical PRC2 target repression networks (the 195-gene Ben-Porath set), not the writers themselves.

Limitations. The fixture-level benchmark shares gene vocabulary with the scoring modules; 18/20 and 8/8 assess rule consistency, not generalization. Wilson CIs are wide. External validation against genome-wide DEG lists is future work. Species handling used uppercase ortholog matching; mouse MSigDB MH sensitivity confirmed invariance for all reported gene sets. The 5-gene scoring module mixes PRC2.1 (EPOP) and PRC2.2 (JARID2) accessories; subcomplex-specific scoring requires larger gene panels than the 5-7 gene signatures available. PALI1 (also known as C10orf12/LCOR isoform) may require alias resolution in some annotation databases.

Reproducibility. All 28 signatures with weights, exact .rnk files, scoring module definitions, and GSEA results are available at https://github.com/scottdhughes/prc2-rejuvenation-benchmark.

Appendix: 28 Benchmark Signatures

Signatures contain genes as reported in original studies. The scoring module (EZH2, SUZ12, EED, JARID2, EPOP) uses strict PRC2-only definitions; other epigenetic genes (CBX7, DNMT3A) in signatures are not scored by the PRC2 module. Note: signatures may contain fewer confounder genes (typically 4) than the full 6-gene confounder modules because individual study signatures only include the subset of confounder genes that were differentially expressed in the original context.

Primary (20):

Within-Family Holdouts (8):

*blind_mixed_xu_1 is classified "mixed" due to direction stability < 0.70 (opposing direction signals within the PRC2 genes), not due to confounder presence.

Full gene lists with weights: https://github.com/scottdhughes/prc2-rejuvenation-benchmark

References

1. Margueron R, Reinberg D. The Polycomb complex PRC2 and its mark in life. Nature. 2011;469:343-349. doi:10.1038/nature09784
2. Lu Y, et al. Reprogramming to recover youthful epigenetic information and restore vision. Nature. 2020;588:124-129. doi:10.1038/s41586-020-2975-4
3. Ocampo A, et al. In vivo amelioration of age-associated hallmarks by partial reprogramming. Cell. 2016;167:1719-1733. doi:10.1016/j.cell.2016.11.052
4. Lopez-Otin C, et al. Hallmarks of aging: an expanding universe. Cell. 2023;186:243-278. doi:10.1016/j.cell.2022.11.001
5. Gill D, et al. Multi-omic rejuvenation of human cells by maturation phase transient reprogramming. eLife. 2022;11:e71624. doi:10.7554/eLife.71624
6. Chondronasiou D, et al. Multi-omic rejuvenation of naturally aged tissues by a single cycle of transient reprogramming. Aging Cell. 2022;21:e13578. doi:10.1111/acel.13578
7. Roux A, et al. Diverse partial reprogramming strategies restore youthful gene expression and transiently suppress cell identity. Cell Systems. 2022;13:574-587. doi:10.1016/j.cels.2022.05.002
8. Liberzon A, et al. The Molecular Signatures Database (MSigDB) hallmark gene set collection. Cell Systems. 2015;1:417-425. doi:10.1016/j.cels.2015.12.004
9. Ben-Porath I, et al. An embryonic stem cell-like gene expression signature in poorly differentiated aggressive human tumors. Nature Genetics. 2008;40:499-507. doi:10.1038/ng.127
10. Sahu SK, et al. Targeted partial reprogramming of age-associated cell states improves markers of health in mouse models of aging. Science Translational Medicine. 2024;16:eadg1777. doi:10.1126/scitranslmed.adg1777
11. Bracken AP, et al. EZH2 is downstream of the pRB-E2F pathway, essential for proliferation and amplified in cancer. EMBO J. 2003;22:5323-5335. doi:10.1093/emboj/cdg542
12. Hong S, et al. Identification of JmjC domain-containing UTX and JMJD3 as histone H3 lysine 27 demethylases. Proc Natl Acad Sci USA. 2007;104:18439-18444. doi:10.1073/pnas.0707292104 | PMID: 18003914
13. Camacho O, Koldobskiy MA, Reddy P, Oak A, Sun Y, Sherman K, Izpisua Belmonte JC, Feinberg AP. Convergence of aging- and rejuvenation-related epigenetic alterations on PRC2 targets. Molecular Systems Biology. 2026 Feb 10 (epub). doi:10.1038/s44320-026-00195-9 | PMID: 41667860 | GEO: GSE289856
14. Hauri S, et al. A high-density map for navigating the human Polycomb complexome. Cell Reports. 2016;17:583-595. doi:10.1016/j.celrep.2016.08.096