Nature Biotechnology
(2026) Cite this article
Although human pluripotent stem cells (hPSCs) can generate all tissues of the body, hPSCs in vitro frequently exhibit differentiation biases or failure that pose substantial challenges for disease modeling and regenerative medicine. The origins of these biases remain incompletely understood and extend beyond reprogramming artifacts. Here we show that loss of default neural differentiation capacity and failure to form brain organoids are linked to erosion of bivalent chromatin marks at developmental gene loci, independent of DNA methylation, driving acquisition of a posterior epiblast-like state and premature developmental gene expression. We develop a chemical chromatin restoration (CHR) approach that rescues this differentiation bias by reinstating transcriptional programs and chromatin landscapes characteristic of the competent anterior epiblast-like state, restoring broad differentiation potential. These findings establish locus-specific patterns of repressive and activating histone post-translational modifications as a tractable and experimentally targetable determinant of hPSC fate competency, and offer an effective route to rescue differentiation-compromised hPSC lines for applications in disease modeling and regenerative medicine.
Human pluripotent stem cells (hPSCs), both embryonic stem cells (hESCs) and induced pluripotent stem cells (hiPSCs), cultured in vitro can vary markedly in their differentiation potential. This variation is partially explained by a pluripotency spectrum characterized by distinct epigenetic and metabolic states1,2,3. The current consensus distinguishes naïve, formative and primed pluripotency4, reflecting different states of the peri-implantation embryo in time and space. These states have, at least in rodents, different differentiation propensities toward germ line and extraembryonic identities. The default lineage of the post-implantation epiblast is neural5,6,7 and, similarly, formative and primed cells in vitro exhibit a tendency to differentiate towards neural identities spontaneously1. Induction of non-neural identities requires WNT and BMP morphogens to first pattern the antero-posterior axis of the epiblast and next to specify germ layers8,9,10,11,12.
Heterogeneity in primed hPSC developmental capacity affects in vitro differentiation to various identities, including neural differentiation and regional central nervous system identity. Inter-individual differences in hPSC differentiation potential are believed to be driven by natural and acquired genetic variation and by alterations in DNA methylation13,14,15,16,17,18,19,20,21,22,23, but recent focus has shifted toward more complex epigenetic differences, such as X chromosome dosage or deposition of Polycomb repressive complex 2 (PRC2) marks24,25.
Here, we explore variability in hPSC differentiation potential and the underlying epigenetic processes through the lens of epiblast anterior-posterior (A-P) regionalization in vivo. We harness the natural diversity among human PSC lines to examine how regionalization influences the competency to differentiate, using anterior neural differentiation as a readout. Leveraging the prolonged embryonic fate transitions in human, we capture subtle and gradual states free of compensatory mechanisms of natural embryos26,27,28,29,30,31.
We show that the failure to undergo default neural differentiation and form brain organoids is linked to a loss of the uncommitted anterior epiblast-like state and the acquisition of a posterior epiblast-like state, with gene locus-specific erosion of bivalent chromatin marks independent of DNA methylation, leading to premature expression of developmental genes. Furthermore, we establish a defined chromatin restoration (CHR) protocol that reinstates the full differentiation capacity. Notably, after CHR, previously differentiation-compromised hiPSC lines display chromatin and transcriptomic signatures that closely resemble naturally anterior-competent lines and are able to generate derivatives of all germ cell layers.
To investigate the mechanisms of unrestricted differentiation competency, we utilized an unguided cerebral organoid (CO) model, which reflects the default pathway of neural differentiation upon withdrawal of TGF-β and FGF232,33. To ensure isolation of cell-intrinsic differences, we reanalyzed single-cell RNA sequencing (scRNA-seq) data from COs generated from a pool of hiPSCs19 grown under identical conditions before and during organoid formation. The individual cell lines produced varying proportions of neural and non-neural cell identities (Fig. 1a and Extended Data Fig. 1a–c).
a, Uniform manifold approximation and projections (UMAPs) of the pooled hiPSC-derived COs; left UMAP summarizes the organoid cell types and right UMAP summarizes the cell line genotype distribution. UL, upper layer; DL, deep layer; INP, intermediate progenitor, RG, radial glia; oRG, outer radial glia; ChP, choroid plexus. b, Heatmaps of z-scores of lineage marker expression in day 20 organoids generated from single lines. FMB, fore- and midbrain; NC, neural crest; SP, sensory placode; R, retinal pigment epithelium; CHP, choroid plexus; ECE, ectoderm epithelium; M, mesenchyme; CV, cardiovascular; EE, endoderm epithelium; AM, amnion. c, Scatter plot of the first two principal components (PCs) of the principal-component analysis (PCA) of bulk transcriptomes of the hPSC lines used in this study (differential genes defined by VST threshold, competent encircled in green). d, Volcano plot of differentially expressed genes between competent (n = 5 lines) and noncompetent (n = 6 lines). Differential expression was tested using the Wald test, and the resulting P values were adjusted for multiple testing using the Benjamini-Hochberg procedure to control the false discovery rate (FDR); significance cutoff is P < 0.05 and log2 fold change >1.5. e, Heatmap of normalized expression (transcripts per million (TPM)) of Wnt ligands in competent and noncompetent hiPSC lines. f, Heatmap of normalized expression (TPM) of canonical Wnt targets in human hPSCs in competent and noncompetent hiPSC lines. g, Representative images of day 20 organoids from either parental lines or after at least 8 passages of culture with IWP2; scale bars 200 μm. h, Quantification of marker expression in day 20 organoids from lines treated with IWP2 or controls; line represents mean, and error bars represent standard deviation (s.d.). SOX2: mean burb1 IWP2 = 55.91 ± 7.51 (n = 6 organoids) versus burb1 = 2.752 ± 3.73 (n = 6), P = 0.0022; mean sojd3 IWP 0.31 ± 0.23 (n = 4) versus sojd3 4.44 ± 9.09 (n = 6), P = 0.7619, EOMES: mean burb1 IWP2 = 2.67 ± 0.94 (n = 6) versus burb1 = 0.38 ± 0.35(n = 6), P = 0.0022, mean sojd3 IWP2 0.14 ± 0.06 (n = 4) versus sojd3 0.46 ± 0.49 (n = 6), P > 0.99, all two-sided Mann-Whitney test).
From these data, we identified anterior-competent hPSC lines (deyz2, hehd1 used in this study and eesb1, fawm2 not used in this study) and anterior-noncompetent lines that predominantly generated non-brain identities (burb1, cuhk2, feec2, fiaj1 and sojd3) (Extended Data Fig. 1b–d). We then generated organoids from individual lines to control for the influence of the pool on differentiation, adding additional lines to achieve representation of competent male lines and ESCs with known competency (kolf2, competent hiPSC; H1, H9, competent hESC; HUES8, noncompetent hESC)34. RNA-seq analysis of tissue markers in day 20 organoids confirmed the competent or noncompetent status of those cell lines (Fig. 1b), whereas through a two-dimensional neural differentiation, we demonstrated that competency is assay independent (Extended Data Fig. 1e). Despite these differences in competency, all hiPSC lines displayed expected patterns of pluripotency-associated gene expression (Extended Data Fig. 1f).
To assess intrinsic differences in pluripotency state that could predict competency, we assayed the transcriptome of the starting cells. PCA revealed a clear separation between competent and noncompetent lines along PC1 (Fig. 1c). To identify the genes responsible, we performed differential expression (DESeq2) on the five competent lines (deyz2, H1, H9, hehd1 and kolf2) and six noncompetent lines (burb1, cuhk2, feec2, fiaj1, HUES8 and sojd3) (Fig. 1d, Supplementary Table 1). Gene Ontology (GO) term analysis of genes upregulated in noncompetent lines revealed several biological process (BPs) terms linked to embryonic patterning and differentiation, particularly those related to posterior structures (Extended Data Fig. 1g,h, Supplementary Table 2). Molecular function analysis pointed toward transcriptional regulators and in particular developmental transcription factors (Extended Data Fig. 1i,j). This suggested a more caudal identity in noncompetent lines.
GO BP analysis suggested elevated Wnt signaling in noncompetent lines (Extended Data Fig. 1g). In mouse, the A-P axis is established by a Wnt signaling gradient that emerges about 1 day before the appearance of the primitive streak35,36. Wnt signaling is also sufficient to drive A-P regionalization in vitro, and it does so before germ layer specification11,12,37. Based on these findings, we hypothesized that the differences between competent and noncompetent hPSC lines are driven by variations in Wnt activity, resulting in a caudalized phenotype analogous to epiblast patterning.
We analyzed the expression of WNT ligands (Fig. 1e) and found that WNT transcripts were generally expressed at very low levels38,39, but canonical WNT3 and non-canonical (WNT5A) and WNT5B) tended to be more highly expressed in noncompetent lines. We also detected expression of previously reported ESC canonical Wnt targets35,40 (Fig. 1f), with higher expression of MIXL1 and HHEX in five out of six noncompetent lines.
To assess whether WNT inhibition could eliminate the posterior bias41, we treated two noncompetent lines, burb1 and sojd3, with the pan-Wnt inhibitor IWP2 for eight passages34. We chose sojd3 as a line with strong Wnt activity and burb1 as a line with weak Wnt activity. We then tested their competency by generating COs analyzed at day 20. The burb1 line showed an improvement in organoid identity and morphology, whereas the sojd3 line did not improve (Fig. 1g,h). These results demonstrate that although spontaneous Wnt activity may contribute to noncompetence in some cell lines, inhibiting Wnt is insufficient to reverse caudalization in all lines, and that the impact of Wnt inhibition on differentiation outcomes cannot be predicted by WNT ligand and target expression.
We next considered the possibility that Wnt signaling memory could alter differentiation outcomes42,43,44,45. We hypothesized that such a phenomenon could underlie the regionalization observed in noncompetent cell lines. In the case of the IWP2 nonresponsive line (sojd3), the memory of prior WNT exposure might not be easily erased by simply blocking de novo Wnt signaling.
Because the establishment and propagation of cellular memory often involves epigenetic changes at the chromatin level46, we explored global chromatin accessibility in competent and noncompetent lines by ATAC-seq. In general, chromatin accessibility was similar between all samples and almost identical in regions within ±5 kb of transcription start sites (Extended Data Fig. 2a), indicating no differences in global chromatin accessibility. However, differential accessibility analysis revealed common trends among the noncompetent compared to the competent lines. PCA of differentially accessible peaks showed separation along PC1, with all five competent lines clustering together and four of six noncompetent lines separately (Fig. 2a).