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

Abstract

Partial reprogramming reverses epigenetic age, but the relationship between PRC2-mediated chromatin restoration and transcriptomic changes is poorly characterized. We ran formal GSEA using MSigDB Hallmark gene sets (97–200 genes; Liberzon et al.\ 2015) across four genome-wide datasets (Chondronasiou, Roux, Gill, Sahu; 23K–61K genes). Proliferation (E2F Targets and G2M Checkpoint) is significant (p <= 0.001) in all four datasets but direction-inconsistent: positive in three (NES +1.96 to +2.18) yet negative in Gill (-1.79/-1.54), where MPTR suppresses proliferation. Interferon is dataset-specific: positive in Roux (NES +2.13) but negative in Sahu (-2.12). PRC2 targets (195 genes) are not significant in any dataset (p = 0.052–0.445). Leave-one-gene-out sensitivity confirms robustness >0.94 for every significant enrichment across all datasets; PRC2 complex components are single-gene-dependent (negative robustness in 3/4 datasets). Leave-one-dataset-out analysis confirms E2F/G2M is the only gene set that replicates regardless of which dataset is held out. A curated 28-signature benchmark with leave-one-family-out CV (96.4%) shows in vivo partial reprogramming is confounded in 37.5% of signatures vs.\ 0% for in vitro transient. All references carry DOIs.

Introduction

Partial reprogramming via OSKM or OSK factors reverses epigenetic age . PRC2 maintains repressive H3K27me3 marks; reprogramming restores H3K27me3 at PRC2 target loci , but corresponding mRNA-level changes are uncharacterized. Meanwhile, proliferation, interferon, and SASP produce transcriptomic changes that can be mistaken for rejuvenation . We asked: (1) Do top DEGs include PRC2 core genes? (2) Are mRNA-level changes better explained by confounders than by PRC2 restoration?

Methods

Signature Extraction Top DEGs were extracted from four studies using downloaded supplementary data: Gill 2022 (29 genes, Supp.\ File 3), Chondronasiou 2022 (25 genes, Table S4, 23,332 after symbol mapping), Lu 2020 (25 genes, Source Data Fig. 4), and Roux 2022 (25 genes, GSE197437 h5ad; caveat: no young reference). Hallmark GSEA GSEA (2,000 permutations, seed=42) was run on four genome-wide ranked lists: Chondronasiou 2022 (23,332 genes; Young-Old), Roux 2022 (55,304; log_2FC Pulse/NegCtrl), Gill 2022 (35,805; log_2FC MPTR/NegCtrl, GSE165177), and Sahu 2024 (61,428; log_2FC P16-OSK/GFP, GSE201710; targeted partial reprogramming via CDKN2A-driven OSK in senescent fibroblasts ). Eight MSigDB Hallmark sets v2024.1.Hs (97–200 genes) plus PRC2 targets (195 genes) and PRC2 complex components (22 genes) were tested. Bonferroni correction: α = 0.005; Benjamini-Hochberg FDR yields identical conclusions. Leave-one-gene-out (LOGO) sensitivity on all 4 datasets for E2F, G2M, and PRC2 Complex, plus targeted LOGO for all other significant (p < 0.005) gene setxdataset combinations (10 additional runs; 1,000 permutations per removal); robustness = 1 - |Δ_| / |_|. Leave-one-dataset-out (LODO) analysis assessed whether conclusions hold when any single dataset is excluded. PRC2 Benchmark We scored 28 curated compact signatures (20 primary + 8 blind) from 4 source families against a PRC2 restoration module (7 genes: EZH2, SUZ12, EED, JARID2, CBX7, KDM6B, DNMT3A) and 4 confounder modules (AP-1, proliferation, interferon, DNA damage; 6 genes each). Null-adjusted signed overlap yields per-module scores R (restoration), L (loss), C (max confounder). Classification: if C >= (R,L), the signature is confounded''; if direction stability < 0.70 or specificity margin |R-L| <= 0.10, it is mixed''; otherwise restoration'' or loss'' by the sign of R - L. Leave-one-family-out CV derives thresholds from 3 families and tests on the 4th.

Results

2 mRNA Is Absent from Reprogramming DEG Lists [h] Study & Genes & PRC2 in DEGs & Dominant DEG Program Gill 2022^* & 29 & 0/7 & SASP (SERPINE1, MMP1/3/10, CXCL8) Chondronasiou 2022^* & 25 & 0/7 & Aging-DEGs (Itgb2, H2-Aa, Cd164) Lu 2020^* & 25 & 0/7 & Neuronal (Elfn2, Nrcam, Trpc4, Vgf) Roux 2022^& 25 & 0/7 & Reprogramming response * 2 core genes absent from top DEGs. ^High confidence. ^(no young reference). Hallmark GSEA: Four Genome-Wide Datasets [h] 3pt & 2cChondronasiou & 2cRoux & 2cGill & 2cSahu Gene Set & NES & p & NES & p & NES & p & NES & p E2F Targets (200) & +2.18* & .001 & +1.96 & .001 & -1.79 & .001 & +2.12 & .001 G2M Checkpoint (200) & +1.97* & .001 & +1.83 & .001 & -1.54 & .001 & +2.01 & .001 IFN-α Response (97) & +1.22 & .051 & +2.13 & .001 & -1.36 & .026 & -2.12 & .001 IFN-γ Response (200) & -0.93 & .381 & +2.10 & .001 & -1.42 & .004 & -2.10 & .001 TNF-α/NF-κB (200) & -1.28 & .028 & +1.00 & .147 & -1.25 & .048 & -1.46 & .002 EMT (200) & +1.85 & .001 & -1.17 & .090 & +1.33 & .007 & +1.84 & .001 P53 Pathway (200) & -0.80 & .572 & +1.12 & .054 & +1.25 & .015 & -1.45 & .002 Inflammatory (200) & -1.11 & .136 & +1.11 & .057 & -1.08 & .228 & -1.40 & .005 PRC2 Targets (195) & -1.22 & .052 & -0.91 & .445 & -1.09 & .207 & +1.08 & .075 PRC2 Complex (22) & +0.72 & .411 & +0.57 & .449 & -1.07 & .230 & -0.70 & .504 Hallmark GSEA (2,000 permutations) across four genome-wide datasets. Bold: significant at Bonferroni α = 0.005. Gene set sizes in parentheses. Permutation floor: p = 0.0005 (1/2,000); all such values reported as <= .001. Proliferation is the most robust confounder, but direction-inconsistent. E2F Targets and G2M Checkpoint both reach significance (p <= 0.001) in all four datasets—the only gene sets to do so. However, the direction flips: both are positive in Chondronasiou, Roux, and Sahu (E2F NES +1.96 to +2.18) but negative in Gill (E2F NES -1.79, G2M -1.54). Gill uses Maturation Phase Transient Reprogramming (MPTR) in fibroblasts, which suppresses proliferation . Even the most robust confounder programs change sign depending on experimental design. **Interferon is dataset-specific and bidirectional.** IFN-α is strongly positive in Roux (NES +2.13) but strongly negative in Sahu (-2.12). IFN-γ is positive in Roux (+2.10) but negative in both Gill (-1.42) and Sahu (-2.10). No consistent interferon enrichment direction holds across reprogramming contexts. **PRC2 targets are not significant in any dataset.** The Ben-Porath PRC2 target set (195 genes) trends negative in Chondronasiou (p = 0.052) and Gill (p = 0.207), trends positive in Sahu (p = 0.075), and is non-significant in Roux (p = 0.445). PRC2 complex components (22 genes) are non-significant in all four datasets (p = 0.23–0.50). No mRNA-level PRC2 enrichment statement survives correction. -One-Gene-Out Sensitivity (All 4 Datasets)* [h] 4pt & 4cRobustness Score & Gene Set & Chond. & Roux & Gill & Sahu & Range E2F Targets & 0.975 & 0.989 & 0.981 & 0.987 & 0.975–0.989 G2M Checkpoint & 0.982 & 0.988 & 0.952 & 0.977 & 0.952–0.988 *PRC2 Complex* & -1.07 & -1.11 & 0.89 & -1.36 & *neg.\ in 3/4* robustness across all 4 datasets (1,000 permutations per gene removal). E2F/G2M: robustness >0.95 in every dataset. PRC2 Complex: single-gene-dependent in 3/4 datasets (PHF19, EPOP, CBX4). E2F/G2M robustness >0.95 replicates across all four datasets. Targeted LOGO on all 10 remaining significant gene setxdataset combinations confirms robustness >0.94 for every significant enrichment in the paper (range: 0.946–0.994). No significant GSEA result is driven by any single gene. PRC2 Complex has negative robustness in 3/4 datasets—removing PHF19, EPOP, or CBX4 alone shifts NES by >100% of baseline. Leading edge analysis identifies 35 core E2F genes shared across all 4 datasets, including EZH2 itself—the PRC2 catalytic subunit appears in the proliferation confounder. Leave-one-dataset-out analysis confirms E2F/G2M is the only gene set significant in all 3 remaining datasets regardless of which dataset is held out; PRC2 never reaches significance. Benchmark On 28 curated PRC2 signatures, confounder gating achieves 100% accuracy (75.0% for a direction-only baseline). Leave-one-family-out cross-validation achieves 96.4% (27/28; one Sarkar restoration signature classified as ``mixed''): `[h] Source & Method & Confounded & Margin & IFN Score Sarkar & In vitro transient & 0/7 (0%) & 0.835 & -0.079 Xu & SVZ neuronal & 0/7 (0%) & 0.672 & -0.043 Browder & In vivo partial & 3/8 (37.5%) & 0.481 & +0.104 Riparini & Muscle/EZH2 & 3/6 (50.0%) & 0.312 & -0.021 * vitro transient reprogramming produces clean PRC2 restoration; in vivo partial is confounded by interferon/AP-1 in 37.5% of signatures.

Discussion

Proliferation (E2F Targets and G2M Checkpoint) is the most robust confounder—both gene sets significant in all four datasets—but direction-dependent: MPTR suppresses proliferation while other protocols activate it, producing opposite enrichment signs from the same gene sets. Interferon is dataset-specific (positive in Roux, negative in Sahu). Neither program can serve as a universal control; transcriptomic benchmarks must include protocol-specific confounder profiles. PRC2 targets are not significant in any dataset at corrected thresholds, and LOGO reveals that PRC2 complex enrichment is a single-gene artifact (PHF19). Notably, all four datasets represent partial reprogramming (including Sahu 2024, which uses CDKN2A-driven OSK targeted to senescent cells), yet direction heterogeneity persists—indicating that tissue context and protocol design, not reprogramming depth, drive confounder profiles. Limitations. GSEA covers 4 datasets from different tissues, explaining direction heterogeneity but limiting pooling. PRC2 mRNA absence does not prove protein-level dysfunction—H3K27me3 ChIP is required. The curated benchmark (n=28, LOFO-CV 96.4%) has one edge-case misclassification. Gene-label permutation GSEA has limited power when gene set members are correlated.

Conclusion

PRC2 mRNA enrichment is not a reliable cross-dataset readout of reprogramming. Proliferation is the most robust confounder but direction-inconsistent. Interferon is dataset-specific. In vivo partial reprogramming is confounded in 37.5% of curated signatures; in vitro transient is not. Transcriptomic assessments of reprogramming require protocol-specific confounder controls.

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-organ partial reprogramming reverses signs of aging in mice. Aging Cell. 2022;21:e13669. doi:10.1111/acel.13669
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