Incomplete DNA methylation underlies a transcriptional memory of somatic cells in human iPS cells

Abstract

Human induced pluripotent stem (iPS) cells are remarkably similar to embryonic stem (ES) cells, but recent reports indicate that there may be important differences between them. We carried out a systematic comparison of human iPS cells generated from hepatocytes (representative of endoderm), skin fibroblasts (mesoderm) and melanocytes (ectoderm). All low-passage iPS cells analysed retain a transcriptional memory of the original cells. The persistent expression of somatic genes can be partially explained by incomplete promoter DNA methylation. This epigenetic mechanism underlies a robust form of memory that can be found in iPS cells generated by multiple laboratories using different methods, including RNA transfection. Incompletely silenced genes tend to be isolated from other genes that are repressed during reprogramming, indicating that recruitment of the silencing machinery may be inefficient at isolated genes. Knockdown of the incompletely reprogrammed gene C9orf64 (chromosome 9 open reading frame 64) reduces the efficiency of human iPS cell generation, indicating that somatic memory genes may be functionally relevant during reprogramming.

Figure 1. Pluripotency validation for the Hepatocyte-iPS cells derived and used for the microarray studies

(a) The 3 Hep-iPS clones used in this analysis showed strong, positive staining for all human ES cell specific-markers analyzed by immunostaining. Scale bar represents 300 μm. (b) All Hep-iPS clones showed high expression of endogenous pluripotency markers and negligible levels of transgene expression by qRT-PCR. Values were standardized to GAPDH and Ubb, then normalized to H9 ES cells (endogenous) or 5-factor infected hepatocytes + dox for 4 days (viral). Data are from triplicate reactions. Error bars represent standard deviations. (c) All Hep-iPS clones formed embryoid bodies in vitro when grown under non-attachment conditions. Shown here are d8 EBs alongside with control ES cell-derived EBs. Scale bar represents 200 μm. (d) Pluripotency of the Hep-iPS cell clones was further confirmed by their ability to form teratomas in vivo, comprised of tissues derived from all 3 germ layers. (i) Neural tissue (ectoderm), (ii) Striated muscle and adipocytes (mesoderm), (iii) gut-like epithelium (endoderm). Also see Supplementary Fig. S2 for pluripotency validation of Fib-iPS cells used for the microarray analysis. Mel-iPS cells have previously been described.

Figure 2. Multiple cell types undergo extensive transcriptional reprogramming to the human iPS cell state

(a) Average-linkage hierarchical clustering of the RMA-normalized expression profiles shows that the replicate data cluster together tightly, confirming the reproducibility of the experiments, and that the somatic cells have been successfully reprogrammed. (b) The box-plot of log expression fold changes for all RefSeq genes further shows that the iPS cells have been reprogrammed to closely resemble the transcriptional profiles of ES cells. The black center line represents the median. The upper and lower edges of the box represent the first and third quartiles, and they define the inter-quartile range. Outliers farther than 1.5 times the inter-quartile range from the box are shown as circles.

Figure 3. iPS cells retain a transcriptional memory of the original somatic cell

(a) LOESS curves fitted to the scatter plots of t-test log p-values: −log(p) and log(p) are plotted for fold changes greater than 1 and less than 1, respectively. The black line is a curve fitted to our data, and other curves are fitted to the 1000 bootstrap simulation datasets obtained by assuming identically distributed iPS and ES cell expression levels. The black line shows clear deviation from the null hypothesis iPS=ES and thus reflects the trend that the transcriptional memory of the originating cell type is retained in low passage iPS cells: genes that were higher (or lower) in the somatic cell than in ES cells tend to be significantly repressed (or induced) during reprogramming, but nevertheless remain higher (or lower) in iPS cells than in ES cells. (b) Box-plots of expression levels for 191 genes that are higher in both iPS cells and somatic cells relative to ES cells (upper right corner genes in (a)) and 391 genes that are lower in both iPS cells and somatic cells relative to ES cells (lower left corner genes) at a t-test p-value cutoff of 0.01. The plots illustrate progressive convergence of somatic gene expression towards the ES cell state. (c) The top panel shows progressively reprogrammed genes (somatic>iPS>ES or somatic<iPS<ES), and the bottom panel shows over-reprogrammed genes (somatic<ES<iPS or somatic>ES>iPS). The p-values for the overlaps are from Fisher’s exact test, and show significant overlaps only for progressively reprogrammed genes. The standard deviations indicate variation among the 3 cell types.

Figure 4. DNA methylation can partially explain somatic gene expression in iPS cells

(a) The genes which maintain higher expression levels in Fib-iPS cells compared to ES cells tend to be also methylated at higher levels in H1 ES cells compared to the fibroblast cell line IMR90. The Pearson correlation coefficient between the log expression fold change and single-nucleotide resolution differences in CpG island methylation was 0.80 (R²=0.64, p-value=0.002). (b) The correlation was 0.88 (R²=0.78, p-value=0.02) for six genes which remain higher in all three iPS cell types compared to ES cells. CpG_ES>IMR90 is the number of cytosines in CpG islands with higher methylation in H1 than IMR90. (c) The overall DNA methylation of 4 of the top somatic genes whose expression persists in low passage iPS cells. DNA methylation was examined with bisulfite sequencing analysis in 3 types of somatic cells (hepatocytes, fibroblasts and melanocytes), 2 clones for each iPS cell type and H1 and H9 ES cells. The detailed bisulfite sequencing data for all samples can be found in Supplementary Data 2. (d) Higher passage iPS cells retain incomplete DNA methylation at somatic cell memory genes. CpG island methylation levels were examined for our validated somatic memory genes (Fig. 4c) in 5 ES cell lines and 6 iPS cell lines with passage number >30 (passage range 30–58, data from a recent study). The boxplot shows the difference in methylation levels between the higher passage ES and iPS cells. One-sided Wilcoxon test p-values confirm that C9orf64, TRIM4 and COMT are still resistant to promoter DNA methylation (i.e., they are hypomethylated) in high passage iPS cells relative to high passage ES cells. No significant difference in DNA methylation was found for the more variable of the 4 genes, CSRP1.

Figure 5. Meta-analysis of DNA methylation-associated transcriptional memory in independent datasets

(a) 37 genes expressed at higher levels in our Fib-iPS cells relative to ES cells tend to show higher expression in the iPS cells generated by Guenther et al. and Warren et al. but there is high variability when expression data alone are used (cyan box-plots). However, when we use 10 differentially expressed genes from our data that were also differentially DNA methylated in ES cells, a greater proportion show persistent higher expression in the iPS cells of the two datasets (yellow box-plots). (b) The heat-map shows the iPS/ES fold-change ranking of the 10 genes that are higher in our Fib-iPS cells and also methylated in ES cells. (The higher the rank, the greater the fold-change). Shown are the Spearman rank correlation coefficients of fold-changes between our data and those of Guenther et al. and Warren et al. (c) 29 genes were significantly higher in iPS cells relative to ES cells in a pooled analysis of the Guenther and Warren datasets and also differentially methylated in ES cells. Differential expression was determined by applying meta-DEDS analysis to the pooled dataset at a stringent cutoff of 0% FDR. The figure shows that the fold-change levels of those genes correlate significantly with DNA methylation levels (p-value = 9.9x10⁻⁴): the higher the fold change in iPS cells relative to ES cells, the higher the promoter DNA methylation in H1 ES cells relative to IMR90 fibroblasts.

Figure 6. The somatic cell memory gene C9orf64 is required for efficient generation of iPS cells

(a) The number of Tra1-81+ iPS cell colonies was counted on d20 after infection of BJ foreskin fibroblasts with 4f alone (4f), 4f+non-targeting shRNA (4f+NTi), 4f+C9orf64 shRNA (3 different short hairpins targeting C9orf64 were independently tested, 4f+Ci1, 4f+Ci2 and 4f+Ci3) or 4f+p53 shRNA (4f+p53i). Infections were performed in duplicate. Knockdown of C9orf64 resulted in a significant reduction in the number of Tra1-81+ iPS cell colonies compared to 4f alone, 4f+NTi or 4f+p53i. (b) Reduction in the levels of C9orf64 expression achieved by each of the 3 shRNA constructs was confirmed by qRT-PCR. The 4f, 4f+NTi and 4f+p53i conditions showed no significant reduction in C9orf64 expression. Fib-iPS and H9 hES cells served as positive and negative controls for C9orf64 expression, respectively. p53 expression analysis further validated the specificity of the shRNAs. Values were standardized to GAPDH and Ubb, then normalized to uninfected BJ fibroblasts. Note log2 scale in Y-axis: e.g., −2 equals down 4x, −3 equals down 8x, etc. Data are from triplicate reactions. (c) Growth curves of fibroblasts infected with 4f, 4f+NTi, 4f+Ci1 and 4f+Ci2, counted on d0, d1, d4, d7 and d10 post-infection. Infections were performed in triplicate. C9orf64 RNAi did not substantially alter total cell numbers during the first 10 days of reprogramming. In all panels data shown are representative of two independent experiments, and error bars represent standard deviations.

Figure 7. Proximity in the genome affects efficiency of gene silencing in iPS cells

(a) The heatmap shows the expression levels of 4 different DNA methlytransferases (DNMTs) in all of the cell types analyzed in our microarray study. The expression level of these DNMTs is relatively equivalent between iPS cells and ES cell controls, suggesting that any differential DNA methylation in iPS cells is not due to insufficient DNMT expression. (b) We considered 62 silenced genes (“Reprogrammed”) and 5 genes whose expression persists in iPS cells (“Fib-iPS>ES”), with at least 10 cytosines methylated only in ES and also showing higher expression in Fib than ES cells. The “Reprogrammed” genes tend to have nearby genes that also require silencing, while the “Fib-iPS>ES” genes are more isolated. The one-sided Wilcoxon p-values for the difference in the number of nearby genes between the two groups are 0.054, 0.022 and 0.028 for the 20kb, 50kb and 100kb distance restrictions, respectively. The local density of genes, irrespective of expression status, was also slightly lower near genes whose expression persists in iPS cells.

Figure 8. Model for the role of DNA methylation in reprogramming to the human iPS cell state

It has previously been shown that DNA de-methylation and re-activation of pluripotency genes is an essential component of reprogramming (top). In addition, incomplete de-methylation of genes silenced in the somatic cell, including developmental regulators of other lineages, has been shown to persist in mouse iPS cells and may affect their differentiation^, (middle). We report here that differential methylation of somatic cell genes underlies their differential expression in human iPS cells (bottom “somatic” panel), and that somatic genes whose expression persists in low passage iPS cells tend to be isolated from other genes that undergo silencing. Clustering of genes requiring simultaneous repression may facilitate recruitment of the silencing machinery, including DNMTs, and regional DNA methylation (top “somatic” panel). Extensive passaging (pink arrows), may lead to further epigenetic silencing of somatic genes in human iPS cells. Arrows indicate active transcription, while hooks indicate repression.