Transposable element expression and sub-cellular dynamics during hPSC differentiation to endoderm, mesoderm, and ectoderm lineages

Babarinde, Isaac A.; Fu, Xiuling; Ma, Gang; Li, Yuhao; Liang, Zhangting; Xu, Jianfei; Xiao, Zhen; Qiao, Yu; Lin, Zheng; Oleynikova, Katerina; Akinwole, Mobolaji T.; Zhou, Xuemeng; Ruzov, Alexey; Hutchins, Andrew P.

doi:10.1038/s41467-025-63080-3

Download PDF

Article
Open access
Published: 18 August 2025

Transposable element expression and sub-cellular dynamics during hPSC differentiation to endoderm, mesoderm, and ectoderm lineages

Nature Communications volume 16, Article number: 7670 (2025) Cite this article

5430 Accesses
1 Citations
2 Altmetric
Metrics details

Subjects

Abstract

Transposable elements (TEs) are genomic elements present in multiple copies in mammalian genomes. TEs were thought to have little functional relevance but recent studies report roles in biological processes, including embryonic development. To investigate the expression dynamics of TEs during human early development, we generated long-read sequence data from human pluripotent stem cells (hPSCs) in vitro differentiated to endoderm, mesoderm, and ectoderm lineages to construct lineage-specific transcriptome assemblies and accurately place TE sequences. Our analysis reveals that specific TE superfamilies exhibit distinct expression patterns. Notably, we observed TE switching, where the same family of TE is expressed in multiple cell types, but originates from different transcripts. Interestingly, TE-containing transcripts exhibit distinct levels of transcript stability and subcellular localization. Moreover, TE-containing transcripts increasingly associate with chromatin in germ layer cells compared to hPSCs. This study suggests that TEs contribute to human embryonic development through dynamic chromatin interactions.

Identifying transposable element expression dynamics and heterogeneity during development at the single-cell level with a processing pipeline scTE

Article Open access 05 March 2021

ETVs dictate hPSC differentiation by tuning biophysical properties

Article Open access 26 February 2025

A single-cell chromatin accessibility dataset of human primed and naïve pluripotent stem cell-derived teratoma

Article Open access 02 July 2024

Introduction

TEs are genomic elements with multiple copies resulting from autonomous and non-autonomous duplication in genomes^1,2. About half of the human genome consists of TEs^1,3,4 that come in families with multiple evolutionary histories and different genomic distributions across species⁵. Because of their repetitive nature and their underrepresentation in protein-coding genes, TEs were considered to have little functional importance^6,7. In comparative genomics studies, TE-containing genomic regions were often actively removed from analysis^7,8. However, recent studies have shown that TEs are functional in diverse biological processes^2,4, including transcription^9,10,11, post-transcription expression regulation^12,13, transcript processing and stability^12,14,15, chromatin regulation^16,17, development^18,19 and disease progression^20,21. While TEs have been co-opted in normal developmental processes^22,23, TEs pose a risk to genomic integrity^4,24, and can cause mutations that interfere with regulatory networks and from chimeric spliced coding and noncoding transcripts that alter function^2,4. Therefore, TEs have both positive and negative aspects⁶. Although TEs have been studied in pluripotent stem cells^12,23 and some terminally differentiated somatic tissues^20,25, their expression patterns and their roles in post-implantation human development and gastrulation have not been well explored.

The presence of multiple copies of TEs in the genome makes the investigation of their functions difficult^26,27. This is especially the case when using short-read RNA-seq as TE fragments often cannot be uniquely mapped to a genomic location²⁸, and replicate TE sequences make assembling TE-containing transcripts from short-reads fraught with difficulty even in well-annotated species^28,29,30. Previous efforts have considered global TE expression without consideration for the specific TE loci^20,31,32. Efforts are now being made to investigate TE expression at specific loci^12,26, with an emphasis on understanding how TE sequences are spliced to form chimeric transcripts. Locus-level expression quantification can improve transcript assembly³³. Although both short reads and long reads can be used for transcript assembly^12,27, the quality of the transcript assembly based on long-reads is superior^12,34, and benefits the study of TE transcriptomics.

We have previously shown that TE sequences are richly expressed in hPSCs, including inside coding and noncoding transcripts, and their presence is associated with changes in transcript biophysical properties¹². It is generally assumed that TEs are most active in the pluripotent stage of development and decline in somatic tissues. However, this has not been explored, particularly in the context of TE sequences inside transcripts. In this study, we investigated human early development transcripts (hEDTs) assembled exclusively from long-read data to investigate TE expression dynamics. We differentiated hPSCs in vitro to the three germ layers; endoderm, mesoderm, and ectoderm, to mimic human gastrulation and conducted long-read RNA sequencing for each cell type. Using the assemblies, we showed that TE-containing transcripts were dynamically regulated. Our results show that TE-containing transcripts alter RNA stability and chromatin interaction in a cell type-specific pattern. Overall, this study presents the expression dynamics of TEs in early development and implicates TEs in somatic differentiation.

Results

hPSC transcriptome assembled from long-read RNA-Seq data

We have previously described an hPSC-specific transcriptome based on a combined long-read and short-read pipeline¹². Here we relied exclusively on deeply sequenced long-read data to assemble transcripts. This strategy has the advantage that TEs will be accurately placed into their transcript context. To this end, we generated PacBio long-read RNA-sequence data for hPSCs. Consensus reads were generated from the PacBio subreads and then processed to produce noiseless reads using a standard PacBio read pipeline (see Supplementary Note and Supplementary Data 1). The reads were mapped to the human genome using Minimap2³⁵ and assembled into transcripts using StringTie2^36,37. After quality assessment, 33,375 hPSC transcripts, corresponding to 16,814 genes, were assembled.

One advantage of transcript assembly using long-reads only is the ability to sequence full-length mostly intact transcripts^27,28. However, transcript assembly from long-reads is not error-free^29,38,39, due to region-bias errors⁴⁰, reference-induced errors⁴¹, semi-random RNA shearing, and even differences in analysis pipelines^41,42,43. We, therefore, checked if our assembled transcripts were full-length and supported by long-reads, as opposed to annotation-aided assembly.

The assembled transcripts were covered entirely by multiple consensus reads (Fig. 1a), and most assembled transcripts were completely covered by at least one consensus read (Fig. 1b), suggesting that full-length transcripts were assembled. Interestingly, almost all the consensus reads were used in the assembly (Supplementary Fig. 1a). As a negative control, we generated a ‘random GTF set’ with the same number and structure as the assembled transcripts (see Supplementary Methods). This random GTF was not supported by our long-read data (Supplementary Fig. 1b). The pileup of consensus reads shows that the reads map from the transcription start sites (TSSs) to the transcription end sites (TESs) of the hPSC assembly (Fig. 1c). The pileup of short-read data indicates end-to-end transcript coverage for the hPSC assembly, but not for random GTF coordinates (Supplementary Fig. 1c).

**Fig. 1: Quality assessment of the assembled hPSC long-read transcriptome.**

To assess the completeness of the assembled transcripts, we utilized published hPSC deepCAGE⁴⁴ and polyadenylated (polyA) data that mark the TSS and TES, respectively. The long-read assembled TSSs were enriched for deepCAGE tags, while the TESs were enriched for polyA (Fig. 1d). These enrichments were not found in the random GTF set (Supplementary Fig. 1d). Similarly, ChIP-Seq data for the TSS marks POL2RA, H3K27ac and H3K4me3 were enriched at the promoters of the hPSC assembly, but not in the random GTF (Fig. 1e, Supplementary Figs. 1e, f). Additionally, H3k36me3 was enriched on the transcript bodies (Supplementary Fig. 1g). As examples, POU5F1 and SLC2A3 were marked by deepCAGE, POLR2A, H3K4me3 and H3K27ac at the TSS and polyA at the TES (Supplementary Fig. 1h, i). This pattern was also seen on novel transcripts on chromosome 1 (Fig. 1f), and a variant isoform of the LAS1L gene with a skipped exon (Fig. 1g).

We next checked the splice junctions of the assembled transcripts. The evolutionary conservation scores across the junctions were higher relative to introns, and the highest conservation was in the dinucleotide at the splice junctions for both the 5’ and 3’ ends (Fig. 1h). Indeed, the nucleotide frequencies showed the typical GT dinucleotide at the donor site, and AG at the acceptor site. Analysis of short-read RNA-seq data, showed that more than 99% of the splice junctions have at least one short-read, and 85% have 10 or more reads (Fig. 1i, Supplementary Fig. 1j). Taken together, these results, from multiple data, demonstrate the fidelity of the terminals and the splice junctions of the assembled transcripts.

In vitro differentiation transcriptome trajectory in germ layers

To investigate transcript expression dynamics during human embryonic development, we in vitro differentiated hPSCs to three germ layers using previously published protocols⁴⁵ (Fig. 2a). Lineage-specific marker genes were up-regulated in each lineage while pluripotent markers decreased along the differentiation time course (Supplementary Fig. 2a). We examined the expression patterns of a set of lineage-specific markers⁴⁶ and found a close match (Supplementary Fig. 2b). We also compared the expression profiles of our cells to published data and found that our cells agreed well with the RNA-Seq data from in vitro differentiation experiments⁴⁷, gastrulating human embryos⁴⁸ and human embryoid bodies⁴⁹ (Supplementary Fig. 2c–e). To confirm the fidelity of our differentiation experiments, we performed Fluorescence-Activated Cell Sorting (FACS) using PAX6, the marker for ectoderm cells. Ectoderm differentiation was confirmed by the expression of PAX6, and 90% of cells were positive (Supplementary Fig. 2f). Ectoderm has four main lineages, each with a unique gene expression pattern⁵⁰. To investigate the specific lineage of our ectoderm cells, we retrieved short-read RNA-Seq data. Consistent with PAX6 expression (Supplementary Fig. 2f), the correlation analysis showed that our ectoderm cells were closer to neuroectoderm (Supplementary Fig. 2g). These results suggest that our in vitro cells captured a normal developmental trajectory.

**Fig. 2: Dynamics of the assembled human early development transcripts.**

After confirming the cell types, long-read sequencing for two biological replicates per cell type was performed to a depth of 20-45 million PacBio raw subreads (Supplementary Fig. 3a). The consensus reads generated from each replicate ranged from 0.5 to 1.2 million, and resulted in transcript assemblies containing 31,245 endoderm, 37,161 mesoderm, and 56,063 ectoderm transcripts (Fig. 2b). These corresponded to 14,989 endoderm, 17,520 mesoderm, and 24,766 ectoderm genes (Supplementary Fig. 3b). To assess the reliability of the assembled transcripts, we checked the splice sites of the assembled transcripts. Across cell states, base-level conservation levels and nucleotide frequencies of the splice junctions showed that the assembled transcripts were reliable (Supplementary Fig. 3c, d). The distributions of alternative splicing events were similar across the four cell states (Supplementary Fig. 3e). The consistency of the splicing signals and the similarity across cell states suggest that the quality of our assembled transcripts was reliable.

We investigated how exhaustive our transcript assembly was across different cell states. Saturation analyses by down-sampling the reads suggests that increasing the sequencing depth would discover more transcripts, especially novel transcripts (Supplementary Fig. 3f), consistent with a previous report⁵¹. As a further measure of the exhaustiveness of the hEDT transcripts, we investigated the proportion of short-read RNA-seq data that aligned to the assembled transcripts and found that more than 90% of the short-reads mapped to the long-read hEDT transcripts (Supplementary Fig. 3g), suggesting that our assembly was exhaustive and that the expression level of the transcripts remaining to be discovered is low.

Since the number of assembled transcripts can be affected by the sequencing depth²⁸, we asked if the higher transcript count in ectoderm cells is a genuine phenomenon or a consequence of greater sequencing depth (Supplementary Figs. 3a, b). To explore this, we subsampled 300,000 consensus reads from the replicates of each cell state. Interestingly, more transcripts were still assembled in ectoderm using only the subsampled reads (Supplementary Figs. 3h). As the sequencing depth preferentially impacts lowly expressed transcripts²⁸, we checked if the higher transcript count in ectoderm would persist when using a higher expression threshold. Indeed, that was the case (Supplementary Figs. 3i), supporting the increased complexity of the ectoderm transcriptome.

Principal component analysis (PCA) from transcript quantification using the merged transcript reference showed that the same cell states clustered together and distinct from other cell types (Fig. 2c and Supplementary Data 2), indicating that different cell states had a specific transcriptomic landscape. We then proceeded to investigate the features of hEDTs across different cell types. Transcript lengths and exon count per transcript were heterogenous, with ectoderm transcripts having significantly longer spliced transcripts with fewer exons per transcript (Fig. 2d). However, other parameters, including un-spliced transcript length, exon length and number of transcripts per gene were not substantially different (Supplementary Fig. 3j). We next predicted coding potential using FEELnc⁵², and transcript distributions based on protein coding potential showed that ectoderm tended to have more noncoding transcripts (Fig. 2e). Comparison of the assemblies of each cell state to the GENCODE (v43) transcripts⁵³ revealed that ectoderm indeed had a lower proportion of GENCODE-known (matching) transcripts (Supplementary Fig. 3k). Novel transcripts tended to be noncoding in all cell types, with a higher proportion of novel transcripts in ectoderm (Fig. 2f). Even at the gene level, the majority of the novel genes were predicted as noncoding (Fig. 2g). These results highlight the ability of assembly to identify new transcripts.

Functional regulation of cell type-specific transcripts

To investigate the expression dynamics of the assembled transcripts, we extracted uniform and lineage-specific marker genes across the four cell states (Fig. 2h). The uniformly expressed genes included housekeeping genes such as ACTB (Supplementary Fig. 4a). As examples of lineage-specific genes, the ectoderm-specific FEZF1 and PAX6 had both long-read and short-read support (Fig. 2i). EOMES, an endoderm marker was expressed only in the endoderm, and the mesoderm marker, IGFBP3, was restricted to the mesoderm (Supplementary Fig. 4b). To investigate the heterogeneity of transcript expression at the single cell level we reanalyzed previously published scRNA-seq data⁴⁷ of hPSCs and differentiated cells. UMAP clustering showed cell types clustered together (Supplementary Fig. 4c). The expression patterns of uniformly expressed and cell-specific transcripts (Fig. 2h) were consistent (Fig. 2j, Supplementary Figs. 4d–h). Both bulk (long and short-read) and scRNA-seq data consistently showed that the in vitro differentiation and bioinformatics procedures captured the expressed hEDT set.

Finally, we checked for potential clinical or functional relevance of the transcripts. Compared to random sequences, both novel and matching transcripts were enriched in SNPs with clinical and functional relevance (Fig. 2k, and Supplementary Fig. 4i). Although ectoderm SNP enrichment was lower, novel transcripts from all cell states had significantly higher SNP enrichment compared to matched random transcripts (Fig. 2k, and Supplementary Fig. 4i). These data suggest that the novel transcripts have potentially relevant biological functions in development and disease.

TE superfamilies in spliced transcripts during differentiation

Having established the reliability of the in vitro differentiation experiments and the assembled transcripts, we investigated TE splicing patterns across cell types. We used nhmmer to annotate TEs inside transcript sequences as previously described^2,54. The proportions of TE-containing transcripts revealed that ectoderm expressed more coding and noncoding TE-containing transcripts (Supplementary Fig. 5a). As previously reported¹², noncoding transcripts tended to contain more TEs than coding transcripts (Fig. 3a). This observation was true when we considered the number of TE-containing transcripts, or the number of reads assigned to each transcript class (Fig. 3a). Intriguingly, the TE-derived nucleotide composition was less than 50% for both coding and noncoding (Fig. 3b). However, TE-containing noncoding transcripts contained more TE-derived nucleotides, although most noncoding transcripts still contained less than 50% TE-derived sequences. This pattern was common to all cell types (Fig. 3b), with only small differences, as hPSCs and endoderm had fewer overall TE-containing transcripts (Fig. 3a), they tended to have higher TE sequence content in their noncoding transcripts (Fig. 3b).

**Fig. 3: Divergent TE expression dynamics across coding and noncoding transcripts.**

To investigate the TE splicing pattern, we checked which TE superfamilies were contained in hEDTs. We found that the most frequently spliced TE superfamilies included SINE, retroposon, LINE, LTR, and DNA (Supplementary Fig. 5b). It is important to note that these TE frequencies did not reflect the genomic TE distribution (Supplementary Fig. 5b) as different TE superfamilies were enriched/depleted versus the genomic background (Fig. 3c). This suggests TEs are not randomly spliced from the background. We found that the majority of SINE-containing transcripts also contained retroposons, suggesting that these two TE superfamilies co-exist in hEDTs (Supplementary Fig. 5c). Across the major TE superfamilies, frequent TE sequences were observed in ectoderm, with almost half of the transcripts containing at least one SINE, compared to about a third in hPSC transcripts (Supplementary Figs. 5b). Indeed, the higher frequency of SINE and retroposon splicing is observed across cell states in both coding and noncoding transcripts. Interestingly, TE-containing transcripts were expressed in a cell state-specific manner. For example, a locus containing a cluster of LTRs on human chromosome 3 is exclusively expressed in hPSCs (Fig. 3d) while an LTR-containing locus on human chromosome 17 was exclusive to endoderm (Fig. 3e). Indeed, multiple cell state-specific TE-containing transcripts such as endoderm-specific, ectoderm-specific and mesoderm-specific transcripts were detected, highlighting cell state-specific TE expression (Supplementary Fig. 5d). TEs were found in the isoforms of genes with roles in differentiation. For example, MIXL1, GATA6 and TGFB2^55,56,57,58, all have TE-containing isoforms (Supplementary Fig. 5e), that were cell type-specifically expressed (Supplementary Fig. 5f). Similarly, hPSC-specific ESRG^59,60 has multiple annotated isoforms that derive predominantly from an LTR (Supplementary Fig. 5e). These results revealed tissue-specific activation of TEs and suggest their potential roles in cell state specification.

Since only a small proportion of transcripts are entirely covered by TEs (Fig. 3b), we asked which part of the transcript TEs are located. In most transcripts, TEs are found in an internal exon; a sizable number of transcripts were TE-terminating (i.e. in the TES) while a few transcripts had TEs in the TSS (Fig. 3f). Across different TE superfamilies, the prevalence of TE transcripts was consistent, but the proportions of TE-initiating and TE-terminating were mixed (Fig. 3f). For example, more transcripts tended to terminate with SINEs than other TE types. Also, LTRs were more likely to be in the TSS in hPSCs and endoderm than in mesoderm and ectoderm. These results highlight the bias in TE presence in the hEDT assemblies.

To investigate the consequences of TE insertion into transcripts, we focused on TE-free and TE-containing transcripts. Ectoderm cells tended to contain more TE-containing transcripts from otherwise TE-free genes, and the pattern was particularly prevalent for SINE-containing transcripts (Fig. 3g). Across TE superfamilies and cell states, the major consequence of TE insertion was modifying, in which a TE chimera affected the CDS length but did not lead to the loss of coding potential (Fig. 3h). An example is the TE-containing isoform of WDR4, a gene that promotes cerebellar development⁶¹. Although TE presence does not lead to the loss of coding potential, the transcript was dysregulated across cell types (Supplementary Fig. 5h). Interestingly, we also found instances in which a TE-containing isoform led to the complete loss or gain of coding potential (Supplementary Fig. 5h). An example of a loss of coding potential is a TE-containing isoform is TXNL4A, a gene implicated in Burn-McKeown syndrome and isolated choanal atresia⁶² (Supplementary Fig. 5h). Conversely, TE insertion is an isoform of a pseudogene, DOC2G creates coding potential (Supplementary Fig. 5h). The functional homolog of the gene has been implicated in neurotransmission in mice^63,64,65. Overall, TE-containing isoforms altered coding potential in otherwise TE-free transcripts.

Because TE insertion has different consequences on coding potential, we investigated the expression dynamics of TE-containing coding and noncoding transcripts. For both coding (Supplementary Fig. 6a) and noncoding (Supplementary Fig. 6b) transcripts, SINEs and retroposons were prominent. Whereas about 80% of noncoding transcripts in ectoderm contained a SINE element, less than 10% of coding transcripts contained an LTR. The bias towards coding transcripts being TE-free was evident in the proportion of coding transcripts for various categories, 91% of all TE-free transcripts were coding, whilst for all TE superfamilies, it ranged from 33% for LTR in ectoderm to 63% for SINE in endoderm (Supplementary Fig. 6c and d). As TEs could potentially create a minor transcript as an alternative isoform of a major functional transcript, we investigated the frequency of TE sequence in major and minor transcripts. For coding transcripts, major isoforms tended to be TE-free, compared to the TE-containing transcripts (Supplementary Fig. 6e). Conversely, for non-coding transcripts the major isoforms tended to contain TE sequences. This suggests TE-containing coding transcripts are minor isoforms, whilst noncoding TE-transcripts are the major isoforms.

TEs might contribute to cell states as functional modules for cell state transitions^66,67. As previously reported⁶⁸, noncoding transcripts had higher expression variability across cell types (Supplementary Fig. 6f). Further, TE-containing coding and noncoding transcripts tended to have higher expression variability than TE-free transcripts. Interestingly, the Tau coefficient, which measures expression specificity, showed that both TE-containing coding and noncoding transcripts had higher cell type-specific expression than TE-free transcripts (Supplementary Fig. 6f), suggesting that TE expression is regulated in a cell type-specific manner.

We next checked the expression variability within cell types using scRNA-seq data and found that hPSC transcripts had lower expression variability than differentiated cells (Supplementary Fig. 6g). Except for mesoderm LTR-containing transcripts, noncoding transcripts tended to have higher expression variability than the coding transcripts. Notably, TE presence led to higher expression variability in coding transcripts but lower variability in noncoding transcripts. The differences in the expression variability might be influenced by the expression detectability, especially for lowly expressed transcripts, however TE-containing transcripts had higher detectability, especially for noncoding transcripts (Fig. 3i). Taken together, these results suggest that TE expression is coordinated during differentiation.

Biased TE-transcript patterns in coding and noncoding transcripts

Having established the dynamic expression of TE-containing transcript in different lineages, we investigated the TE frequencies of coding and noncoding transcripts in the four cell states. Analysis of TE patterns in coding transcripts showed that TE sequences are rare in coding sequences (CDS) (Fig. 4a, Supplementary Fig. 7a, b), presumably reflecting the evolutionary cost of a TE inserting into and disrupting a CDS. Across TE superfamilies and cell states, TEs were overrepresented in the 5’ and 3’ untranslated regions (UTRs) compared to the CDS, with higher frequencies in the 3’ UTRs (Supplementary Fig. 7a). However, TE subfamilies showed differences. For example, while HERVH and MSTB are both LTRs, their splicing patterns were different, with HERVH sequences being rare in the 3’UTR, whilst MSTBs were enriched (Fig. 4a). In addition, there was a higher frequency of HERVH TE-derived sequences in endoderm cells. Interestingly, the TE pattern of X24_LINE was very similar to that of MSTB and also showed higher presence in ectoderm UTRs, revealing the heterogeneity of the splicing patterns of specific TEs. The comparison of the MER3 and LTR6A frequencies further highlighted differences between these TEs and cell states (Supplementary Fig. 7a).

**Fig. 4: TE splicing patterns and frequencies of TE-containing coding and noncoding transcripts.**

We next looked at the frequency of all TEs in the TSSs and TESs, and interestingly there were no substantial differences across cell states (Fig. 4b). Across the TE superfamilies, TE proportion was low at the TSS. TEs were also reduced at the TES and this reduction extended 5’ into the transcript body (Supplementary Fig. 7c). Similarly, TEs at both the donor and acceptor splice junctions were low (Fig. 4c, Supplementary Fig. 7d). Although there were subtle differences across cell lines and TE superfamilies, the general pattern was similar (Supplementary Fig. 7d). Surprisingly, TE frequencies at splice junctions suggest that suppression of TE insertion at the exon starts is stronger than suppression at the exon ends. This suggests TEs are inhibited from splicing into TSSs, TESs, or splice junctions to prevent the disruption of transcript structure.

We next studied TE splicing patterns in noncoding transcripts and found that they were dynamic between TE family, cell states, and even positions within the transcript (Fig. 4d–f, Supplementary Fig. 7e–g). Specifically, while SINE and retroposon TEs are enriched towards the 5’ and 3’ ends of the transcripts, LINE, LTR, and DNA superfamily TEs are uniformly distributed. Additionally, the overall frequencies were different across cell states (Fig. 4d, Supplementary Fig. 7e). Ectoderm noncoding transcripts were enriched for LINE and DNA TEs, whilst LTRs were enriched in hPSCs, ectoderm, and endoderm (Fig. 4d). The dynamics of TE expression was more obvious for transcripts containing LTRs. For example, while HERVFH21 was concentrated in the center of noncoding transcripts, endoderm LTR6A had higher enrichment around 5’ and 3’ ends (Supplementary Fig. 7e), suggesting these transcripts are noncoding remnants of semi-intact ERVs. HERVH, conversely, was uniformly distributed across noncoding transcripts in hPSC and endoderm cells (Supplementary Fig. 7e), highlighting the differences among TEs.

TE frequencies in noncoding transcripts were also dynamic. SINE, LTRs, and retroposons were rare at the TSS (Fig. 4e, and Supplementary Fig. 7f). Conversely at the transcript ends, SINEs and LTRs were enriched, and then their enrichment dropped substantially after the TES, suggesting that they marked transcript ends. TE sequences were absent at the TSSs of noncoding transcripts, suggesting TE sequences rarely act as noncoding promoters. TE frequencies at noncoding transcript splice junctions were also low (Fig. 4e, and Supplementary Fig. 7g), but not as low as in coding transcripts (Supplementary Fig. 7g). Overall, noncoding transcripts showed distinct patterns compared to coding transcripts at the TSS, TES and at splice junctions.

To explore cell type-specific TE activity, we quantified their frequencies in coding and noncoding transcripts. Comparison of TE-containing coding transcripts identified 218 TEs that were significantly different upon differentiation into endoderm, mesoderm, or ectoderm (Fig. 4g). The majority of the TE subfamilies that were significantly different were enriched in ectoderm cells and were LINEs, LTRs and a few DNA TEs. In contrast, some LTRs like LTR7, HERVH, HERVH48, and MTSB2 were enriched in endoderm coding transcripts. On the other hand, 331 TE subfamilies were differentially enriched in noncoding transcripts in at least one of the four cell states (Fig. 4h), and the majority of the differentially spliced TEs in noncoding transcripts were LINEs and LTRs that were more frequent in the ectoderm. Some LTRs such as LTR7B and HERVH were more frequent in hPSCs, as previously reported^69,70 and many of them were in endoderm cells, suggesting these are not specific features of hPSCs. Indeed, LTR7B/Y/H and HERVH were enriched in endoderm (Fig. 4h), supporting the idea that HERVH expression extends to cell types beyond hPSCs. HERVH modulates 3-dimensional genome structure in hPSCs⁷¹ and may also perform this function in somatic cells.

Interestingly, there was a substantial overlap in the significantly enriched TE superfamilies in any cell type for both coding and noncoding transcripts (Fig. 4i). Breaking this down by TE superfamily, the overlap is high for LINEs, with just 21 TEs specific for noncoding transcripts and 134 LINE types enriched in both coding and noncoding. For example, although the frequencies were different, the overall abundance patterns for LINE L1MD_orf2 were similar for coding and noncoding transcripts (Fig. 4j). In contrast to LINEs, less LTR overlaps between coding and noncoding transcripts, and only 49 LTR-types were common. This was exemplified by the LTRs MSTB2 and LTR7Y which were ectoderm and hPSC/endoderm enriched, respectively (Fig. 4j). Overall, these data indicate that the TE content in different lineages is dynamic, particularly in the ectoderm in general, and for LTRs in all lineages which are cell type and transcript-type specific.

TE subfamily transcript switching

To explore this, we measured the relationship between TE superfamily and expression level at the individual transcript level. We used bulk short-read RNA-seq data for each lineage with the long-read based hEDT reference, as the dynamic range of expression is higher for short-read data than long-read data²⁸. In all four cell types, the expression levels of TE-free coding transcripts were significantly higher than those of TE-containing coding transcripts (Supplementary Fig. 8a). TEs in the 5’ UTRs resulted in transcripts with the lowest expression levels while 3’ UTR transcript chimeras led to higher expression levels, but not as high as TE-free transcripts. We next investigated the aggregate expression patterns and found that the overall expression patterns for SINE, LTR, LINE, retroposon, and DNA coding transcripts were not substantially different across the four states (Fig. 5a, Supplementary Fig. 8b).

**Fig. 5: Dynamic expression patterns of TE-containing coding and noncoding transcripts across different cell states.**

Above, we mainly consider TEs at the superfamily or family level, but with long reads we can place TEs into their specific transcript context. Although there is a substantial overlap in the subfamilies of TEs that are found in coding and noncoding transcripts, the exact genomic loci of the TE-containing transcripts are likely to be different. Potentially there are two scenarios. The first scenario is expression of the same TE subfamily from the same genomic locus, while the second is the expression of the same TE subfamily from different loci. If analyzed at the bulk level, the TE family/subfamily expression would remain unchanged in both scenarios, but in the second the underlying transcript providing the TE would be different. Differential transcript expression between hPSC and the three differentiated cell states revealed that hPSC-to-endoderm differentiation induced the least number of transcript changes while mesoderm differentiation drove the largest coding transcript expression changes (Supplementary Fig. 8c). Differentiation in all three lineages led to upregulation and downregulation of TE-containing coding transcripts, suggesting that transcript regulation is locus-specific and that TE expression is not restricted to hPSCs. We checked the overlaps of the differentiation-induced differential transcripts and found that many transcripts were regulated in a lineage-specific manner (Supplementary Fig. 8d). The aggregate expression of specific TE coding transcripts identified seven LTRs (LTR7Y, LTR7C, LTR7B, LTR7, HERVH, HERVH48 and HERVFH21) that were significantly downregulated in both ectoderm and mesoderm cells, and two LINEs (L1P4b_5end and L1HS_5end) that were substantially upregulated in ectoderm cells (Fig. 5a, and Supplementary Figs. 8e). This suggests that although the TEs are similar in all three cell types, they originate from different transcripts.

We then examined the expression levels of the individual TE-containing coding transcripts. Surprisingly, the majority of the HERVH-containing differentially expressed coding transcripts were highly expressed in endoderm cells but not in hPSCs (Fig. 5b). Different coding transcripts containing L2 and L1HS_5end were activated in different cell states, with activation of many L1HS_5end-containing coding transcripts in ectoderm cells. These data support the second of our scenarios that whilst similar TE types are expressed, they are derived from different loci that are expressed in cell type-specific transcripts.

As noncoding transcripts are more likely to be expressed in a cell type-specific manner^28,72,73, we expected that noncoding TE-containing transcripts would also be expressed from different transcript loci. Fewer differentially expressed noncoding transcripts were induced by the differentiation process (Supplementary Fig. 8f), reflecting the reduced number of noncoding transcripts in each cell type (Fig. 2e). Also, the differentially expressed transcripts were regulated in a cell type-specific manner, and only a few differentially expressed transcripts were shared among cell states (Supplementary Fig. 8g), and TE-containing transcripts tended to be lower expressed (Supplementary Fig. 8h). A comparison of the aggregate expression for noncoding transcripts containing specific TEs showed that 126 TEs were differentially expressed between the hPSCs and at least one of the differentiated cell states (Fig. 5c). These TEs included LTRs and SINEs that are activated in hPSCs and endoderm cells. LTR6A and other LTR TEs were also enriched in endoderm cells. Analysis of the differentially expressed LTR7-containing transcripts showed that many were downregulated in ectoderm and mesoderm cells, whilst LTR6A-containing noncoding transcripts were up in endoderm (Fig. 5d). These data support a model whereby TE superfamilies/families may appear stably expressed during differentiation, but TE types are switching transcripts and the same TE superfamilies/families are expressed from different transcripts.

Cell differentiation experiments involved the exposure of hPSC to different chemical signals that were able to induce cell state changes. To exclude the possibility that the observed TE regulation across lineages was simply due to exposure to differentiation signals, we quantified the expression of TE-containing differentiation-induced differentially expressed transcripts in hESFs (human embryonic skin fibroblasts). PCA suggested that hESFs were unresponsive to differentiation signals (Supplementary Fig. 8i), and the transcript switching of HERVH and LINE TE-transcripts from hPSCs to endoderm was not observed in hESFs treated with endoderm differentiation medium (Supplementary Fig. 8j). These results showed that the TE expression changes reflected differentiation, and not exposure to differentiation signals.

In vitro differentiation of hPSCs can produce four different types of ectodermal lineages, namely neuroectoderm, neural crest, cranial placode, and non-neural ectoderm⁵⁰. PAX6 expression and overall correlation suggest that our ectoderm cells were biased towards neuroectoderm (Fig. 2i, and Supplementary Figs. 2a, f). We asked if the TE activation in our ectoderm cells was specific to neuroectoderm or common to all the four ectodermal lineages. By reanalyzing published bulk RNA-Seq data⁵⁰, LINEs showed consistent transcript-switching in the neurectoderm, neural crest and cranial placode, but not in non-neural ectoderm (Supplementary Fig. 8k). Interestingly, downregulated hPSC-expressed LINE-containing transcripts were down in all four ectodermal lineages, demonstrating dynamic regulation of LINE-containing transcripts in ectodermal lineages.

TE expression dynamics are recapitulated in vitro and in vivo

We next investigated the expression dynamics of TE-containing coding transcripts using scRNA-seq data from in vitro differentiated germ layer cells⁴⁷. The scRNA-seq data recaptured the higher expression of multiple TEs in ectoderm cells and downregulation of some LTRs such as LTR6A, LTR7, LTR7C, and HERVH in ectoderm and mesoderm cells (Supplementary Fig. 9a). The data also revealed that the expression differences were not obvious at the level of TE superfamilies but became distinct at individual transcripts (Fig. 5e, and Supplementary Fig. 9b). Also, single-cell analyses of the noncoding transcripts were consistent with the observations from bulk data and revealed upregulation of multiple TEs in ectoderm cells (Supplementary Fig. 9c, d). Several LTR TE subfamilies were downregulated in ectoderm and mesoderm, while LTR6A was enriched in endoderm cells (Fig. 5e, and Supplementary Figs. 9c, d). As with bulk data, lineage-specific expression became was clear when TEs were considered at the transcript level. These results demonstrated that the observations in bulk RNA-seq were recaptured in single cells.

To elaborate on the expression dynamics of TE-containing transcripts, we extended our analysis to scRNA-seq from in vivo gastrulation⁴⁸ (Fig. 5f, and Supplementary Fig. 10a). As in the in vitro system, ectoderm cells were marked by TE-containing coding and noncoding transcripts (Supplementary Figs. 10a, b). Indeed, compared to other cell types, more ectoderm-expressed transcripts were found in almost all TE superfamilies in both coding and noncoding transcripts (Supplementary Figs. 10c). Also, many transcripts that were identified in the in vitro data were recaptured in the in vivo data (Fig. 5f, and Supplementary Fig. 10d), indicating that in vitro observations reflected in vivo phenomena. Taken together, these data revealed that whilst similar TE types are expressed in all cell types, they originate from multiple lineage-specific transcripts, suggesting that TEs of the same type from multiple loci are independently regulated.

Cell type-specific regulation of TE-transcripts

We next sought to explore the regulation of cell type-specific transcripts. There are two potential (not exclusive) modes: firstly, that the normal cell type-specific regulation apparatus regulates TE-transcripts, or secondly, that a separate TE-specific regulation system is in action. We measured the enrichment of transcription factor binding motifs (TFBMs) in the core promoters of TE-containing and TE-free transcripts that were differentially expressed. TFBMs were enriched in the TE-free differentially expressed transcripts in multiple cell types (Supplementary Fig. 11a). An example was the Maz motif which was found in the top 5 most enriched TFBMs in all the lineages. Comparison of the enriched TFBMs of TE-containing and TE-free differentially expressed transcripts showed surprisingly little overlap (Supplementary Fig. 11b). The only exception was in hPSC transcripts in which the overlap was significantly more than expected by chance. These results suggest that these TE-free and TE-containing transcripts are upregulated by different transcription factors during development.

Next, we checked the enrichment of RNA-binding protein (RBP) motifs in the differentially expressed transcripts using transcripts without differentiation-induced expression changes as the background. Motifs for RBPs such as LIN28A and PRPRC1 were enriched in differentially expressed TE-free transcripts of multiple cell types (Supplementary Fig. 11c). Comparison of enriched RBP motifs showed that the number of overlaps was less than expected by chance in all lineages (Supplementary Fig. 11d), reflecting that TE-free and TE-containing transcripts have different RBP motifs. Overall, the analysis of TFBMs and RBP motifs suggests that TE-transcripts tend to be regulated independently. It is illustrative that the RBP motif for m⁶A (N6-methyladenosine) binding protein IGF2BP3 was found only in the TE-free transcripts, suggesting m⁶A-mediated suppression of TEs is cell type-specific^74,75,76.

TE-subfamily switching in terminally differentiated somatic cells

As the same TE superfamilies/families are expressed from different transcripts, we explored differentiation time course data to investigate the endoderm-specific expression of TE-containing transcripts. As LTR6A and HERVH-containing transcripts are specifically expressed in the endoderm or hPSCs (Fig. 5b, d), we focused on transcripts containing both TE subfamilies. The analysis of time-course bulk RNA-seq showed that LTR6A noncoding transcripts were upregulated at 72 h (Fig. 6a, b). HERVH-containing endoderm-specific noncoding transcripts were upregulated as early as 24 h and reached maximum expression at 72 h. Conversely, HERVH-containing hPSC-specific noncoding transcripts were downregulated in endoderm as early as 12 h and were undetectable at 72 h. The situation was similar for HERVH-containing coding transcripts (Fig. 6a, b, and Supplementary Fig. 12a). Next, we analyzed scRNA-seq data of a 0–96-h differentiation time-course of hPSCs-to-endoderm (Fig. 6c). In agreement with the bulk RNA-seq analysis, endoderm-specific LTR6A, and HERVH-containing noncoding transcripts had high expression only at 72- and 96-h of differentiation (Fig. 6c). In contrast, hPSC-specific noncoding transcripts were shut down at 12 h of differentiation (Fig. 6d). These results demonstrate how HERVH and LTR6A can appear expressed in hPSCs and endoderm, but the transcripts providing these TE sequences are switching during differentiation.

**Fig. 6: Expression changes of TE-containing transcripts in endoderm lineage.**

If TE subfamily transcripts switch in hPSCs and endoderm, then they should be regulated differently. We found multiple RBPs with differentially enriched motifs between HERVH transcripts in endoderm and hPSC (Fig. 6e, f). The most represented RBP motifs in HERVH transcripts in endoderm were serine/arginine-rich splicing factor (SRSF) motifs (SRSF9, 1 and 2) (Fig. 6e). These splicing factors have been implicated in differentiation⁷⁷. Interestingly, the RBPs with biased binding in hPSC-biased HERVH transcripts were also represented in endoderm-biased HERVH transcripts (Fig. 6f). For example, the motif of PABPC4, which has roles in mRNA stability and differentiation⁷⁸, was found in 92% of hPSC-specific and 78% of endoderm-specific HERVH transcripts (Fig. 6f). Overall, though, hPSCs and endoderm had specific RBP motif enrichments. In addition to the RBP motif enrichment, TFBMs in the core promoters of HERVH transcripts showed lineage-specific enrichment. The most represented biased motif was PITX1, a TF that has been implicated in differentiation⁷⁹, with motif found in 84% endoderm-biased, compared to 72% hPSC-biased HERVH transcripts, while HOXA13 motif was found in 89% hPSC-biased and 78% endoderm-biased HERVH transcripts (Fig. 6g and h). However, like the RBPs, the endoderm and hPSCs had specific TFBMs, for example, the TEAD motif, associated with genes important for differentiation^80,81, was endoderm-specific (Fig. 6g), and LRF motif was hPSC-specific (Fig. 6h). These results suggest that HERVH transcripts were independently regulated in endoderm and hPSCs, supporting a transcript-switching model.

To investigate if the expression of the endoderm-induced upregulated transcripts persists in terminally differentiated somatic cells. We checked their expression in a panel of hepatocyte-related RNA-seq. Surprisingly, the expression of noncoding transcripts containing HERVH or LTR6A that were endoderm-specific was not sustained in terminally differentiated cells (Fig. 7a–c). Importantly, endoderm-specific HERVH-containing noncoding transcripts did not persist in endoderm-derived terminally differentiated somatic cells (Fig. 7b, c). In contrast to HERVH, the expression of ectoderm-specific noncoding transcripts containing L1P1_orf2 and LTRs were maintained in multiple ectoderm-derived somatic cells, such as neurons (Fig. 7d, e). Similarly, the expression of many ectoderm-activated coding transcripts containing L2, L1P1_orf2 and LTR remained active in ectoderm-derived terminally differentiated tissues. In summary, these results demonstrate how TE families can remain active across cell types, and how HERVH-related transcripts are transient in the endoderm, but L1 and L2 LINEs persist in the ectoderm.

**Fig. 7: Expression patterns of differentially expressed transcripts containing selected TEs in somatic cells.**

TE sequences influence transcript subcellular localization

Previous studies have shown that TE-containing transcripts tend to localize in the nucleus^82,83. To this end, we generated RNA-seq data for nuclear and cytoplasmic subcellular fractions of hPSCs, mesoderm, endoderm, and ectoderm cells. Western blots for GAPDH (cytoplasm) and Histone H2B (nucleus) confirmed the relative purity of the subcellular fractions (Supplementary Fig. 13a). We computed the relative concentration index (RCI)⁸⁴ to quantify the subcellular localization. This confirmed our previous observation that both coding and noncoding TE-containing transcripts preferentially localize to the nucleus in hPSCs¹² (Fig. 8a). This pattern was the same for mesoderm and endoderm (Fig. 8a). Surprisingly, this was not the case for ectoderm, where the opposite pattern was seen: TE-containing transcripts were more likely to localize in the cytoplasm. Indeed, analyses of TE superfamilies revealed that they all tended to preferentially localize to the cytoplasm in ectoderm cells, and to the nucleus in all other cell types (Supplementary Fig. 13b). Greater details were revealed at the transcript level. For example, LTR-containing noncoding transcripts tended to have different cytoplasm/nucleus enrichment when compared to other TE superfamilies in hPSC and endoderm (Supplementary Fig. 13c).

**Fig. 8: Subcellular localization and stability of TE-containing transcripts.**

Noncoding RNAs that localize to the nucleus are often recruited to chromatin^85,86. Hence, we performed RNA-seq on the nucleoplasm and chromatin fractions to understand the sub-nuclear localization of TE-containing transcripts (Supplementary Fig. 13a). TE-containing hPSC transcripts were reduced in the chromatin compartment, versus the nucleoplasm (Fig. 8b). On the contrary, the TE containing transcripts of endoderm, mesoderm and ectoderm tended to be more enriched in the chromatin fraction (Fig. 8b, and Supplementary Fig. 13d). Transcript-level analyses of nucleoplasm/chromatin enrichment revealed heterogeneity across coding potentials, cell states and TE superfamilies (Supplementary Fig. 13e). To investigate the subcellular transcript localization, we compared the expression levels of individual transcripts enriched in specific sub-cellular fractions (Supplementary Fig. 14a). The number of transcripts localized to the sub-cellular compartments varied across cell types (Fig. 8c). Importantly, compared to the differentiated cells, hPSCs had fewer transcripts localized to the nucleoplasm and chromatin, and transcript distribution revealed both cell state, coding/noncoding and location-specific difference. Specifically, while nucleoplasm-localized transcripts had the highest noncoding proportion in hPSC, chromatin-localized transcripts had the highest noncoding proportion in differentiated cells (Fig. 8c). This effect was also transcript-specific, as few transcripts overlapped in the distinct sub-cellular compartments between the different cell types (Supplementary Fig. 14b). Indeed, using Jaccard similarity, transcripts that localized to the nucleus or cytoplasm were more consistent in hPSC, endoderm and mesoderm but more divergent in ectoderm (Fig. 8d). Interestingly, we found that the distribution of transcripts based on the number of cells in which they were localized varied across subcellular structures (Fig. 8e). The most unique chromatin localization was found in the mesoderm (Fig. 8d). Overall, nucleoplasm localization was the least consistent across the cell states.

To further explore the relationship between transcript subcellular localization and cell state conversion, we checked the proportion of differentiation-induced DETs among the transcripts localized to various subcellular structures across cell states and found that subcellular localization significantly influenced the proportion of DETs (Fig. 8f), as the distribution of DETs were significantly different in localized transcripts compared to overall transcripts. For hPSCs, nucleoplasm-enriched transcripts had the lowest proportion of DETs. For differentiated cells, however, the lowest proportion of DETs was found in chromatin-localized transcripts. Overall, compared to the TE-free transcripts, TE-containing transcripts tended to be cell type-specific (Fig. 8e), with hPSCs having fewer chromatin-enriched TE transcripts.

An important question was how the localization of the transcript to different subcellular fractions was established, particularly in the ectoderm. We hypothesized that RBPs might contribute to the transcript localization. Consequently, we checked the RBP enrichment of sub-cellular localized transcripts. Interestingly, multiple RBP motifs were enriched in transcripts localized to subcellular compartments, and the patterns were cell type-specific (Supplementary Fig. 15a). For example, PABPC4, SART3 and PAPPC1 were associated with strong cytoplasm localization in the mesoderm but there was no significant bias in hPSCs. On the other hand, RBM3 was associated with cytoplasm-localized transcripts in all cell types. Next, we checked if subcellular localization of TE-containing and TE-free transcripts were associated with the same RBPs. Interestingly, several RBPs with significantly higher enrichments in TE-containing transcripts localized to different subcellular structures (Supplementary Fig. 15b). Indeed, multiple RBPs were consistently associated with TE-containing transcripts localized to subcellular compartments in different cell states. For example, RBMS3, DAZAP1, PABPC4, SART3, MSI1 were all found to be enriched in TE-containing transcripts localized to different subcellular compartments in the four cell types. These results suggest that RBPs may play significant role in transcript subcellular localization and that TEs bias RBP interactions.

TE sequences influence transcript stability

As subcellular localization can influence degradation, we wondered if this would impact transcript half-life in a TE, sub-cellular localization, and cell type-dependent manner. We performed RNA-seq in differentiated cells treated with Actinomycin D to block transcription for 1 and 8 h and measured the transcript levels relative to the 0-hour time-point. Surprisingly, there were divergent patterns of transcript half-life across TE superfamilies, coding potentials, and cell states (Fig. 8g). The lowest RNA stability was found in hPSCs and ectoderm, whilst endoderm and mesoderm had more stable RNAs. Interestingly, the impact of TE presence was also dynamic: TE-free transcripts were more stable in hPSC, ectoderm, and endoderm, but not in mesoderm (Fig. 8g, Supplementary Fig. 16a, b). TE-containing ectoderm transcripts also contrasted with hPSCs and endoderm, as whilst TE-containing coding transcripts were less stable, TE-containing noncoding transcripts were more stable than TE-free transcripts (Fig. 8g, and Supplementary Fig. 16a, b). These revealed that transcript stability varied across cell states and transcript coding ability.

We checked if TE presence influenced transcript stability in combination with sub-cellular localization on the basis that cytoplasmic RNAs tend to have a shorter half-life. As expected, nuclear-localized transcripts were more stable than cytoplasm-localized transcripts in the ectoderm, however, the opposite was true in the other cell types (Fig. 8h and Supplementary Fig. 16c). Conversely, chromatin-localized transcripts were more stable than nucleoplasm-localized transcripts in hPSCs, but the nucleoplasm-localized transcripts were more stable than chromatin-localized transcripts in differentiated cells.

A comparison TE-free to TE-containing transcripts identified multiple TEs associated with significant differences in stability after 8 h (Supplementary Fig. 16d). In hPSCs, TE presence mostly led to lower stability in all subcellular compartments (Supplementary Fig. 16d). However, in differentiated cells, mosaic patterns were found. For example, TE-containing transcripts localized to cytoplasm and nucleoplasm in mesoderm were more stable, along with cytoplasm-localized transcripts in ectoderm (Supplementary Fig. 16d). Overall, the data suggest that stability and subcellular localization of TE-containing transcripts are dynamic across different cell state conversions.

Discussion

TEs are difficult to place in their genomic context due to their repetitive nature. However, understanding expression dynamics would substantially benefit from robust TE-containing transcript assemblies^12,33. Here, we used long-read RNA-seq technology to generate an accurate TE-containing transcript assembly^27,28 in in vitro cells that mimic early human embryonic development. Using this assembly, we found widespread and dynamic TE expression in human early development. Multiple studies have shown the overexpression of TEs in hPSCs^9,10,87,88. However, we find that the expression of TE sequences is not restricted to hPSCs, and multiple TEs are expressed in differentiated germ layers. Interestingly, although the same TE subfamilies are expressed, they are found in different transcripts. For example, hPSCs and endoderm both expressed HERVH, but we show the transcripts that contribute those TEs are different. There was a similar case for LINE elements in ectoderm cells.

Our long-read RNA-seq data indicates that the ectoderm has a particularly complex transcriptome with higher levels of TE-transcripts. The expression of LINEs was particularly prevalent, and, unlike the endoderm-specific TEs, many ectoderm-specific TEs were expressed in terminal somatic cells and later stages of development. This is consistent with reports of TE activity in terminally differentiated ectoderm-derived cells and tissues^31,89,90,91. The exceptions to higher TE activities in ectoderm involved HERVH, LTR7, LTR6A, and several other LTRs that have previously been reported to be enriched in hPSCs^10,70. In addition to expression in hPSCs, HERVH is also expressed in endoderm, and some HERVH-containing transcripts were higher in endoderm while others were higher in hPSCs. Overall, all three lineages had specific patterns of TE expression, broken down by superfamily, family and subfamily level down to individual transcripts. This reveals the complex pattern of TE expression in chimeric transcripts and shows how TE expression is not just limited to the pre-implantation stages of embryonic development but is also prominent at gastrulation.

Consistent with previous studies^12,92,93, we found that TE presence leads to lower expression levels, probably due to the activities of RBPs^15,93. Indeed, SINE elements are overrepresented in 3’ UTRs, and STAU1 promotes mRNA decay by targeting Alu elements¹⁵. This indicates that meta-analysis of TEs has limitations, and TE sequences should be considered in their specific transcript context. From a biological perspective, it also suggests that cell-state specific regulation of TE expression occurs at the transcriptional and post-transcriptional stages. Interestingly, we found that TEs tend to be enriched in the chromatin fraction of the differentiated cells, suggesting that TE enrichment on chromatin might be associated with chromatin changes which might induce or regulate cell fate transitions. Indeed, sub-cellular enrichment of TE-transcripts was dynamic between the different lineages, and there was no simple relationship between TE, TE-transcript, cell type and subcellular localization. One prominent feature of mesoderm is the depletion of TE-transcripts in the nuclear fraction. Similarly, TE-transcript stability was also cell type-specific: unstable in hPSCs and endoderm, but more stable in mesoderm and ectoderm. The functional impact of these changes remains to be elucidated, and whether all TE expression changes are drivers of cell differentiation or bystanders requires detailed studies.

Several studies have demonstrated roles for TE-containing transcripts in pluripotency^70,94,95, cardiomyocyte development⁹⁶ and neurogenesis^91,97,98, and we predict that TE-transcripts, including coding TE sequence fragments have an unappreciated role in somatic development. Taken together, this study reveals how TE-containing transcripts are a normal part of germ layer development and are highly dynamic along multiple axes.

Methods

In vitro cell differentiation into the three germ layers and RT-qPCR validation

H1 human pluripotent stem cell lines were maintained on a Matrigel treatment plate, cultured in mTeSR1 (STEMCELL TECHNOLOGIES, 85850) medium and digested with Accutase (Sigma, A6964) every 5–7 days. In vitro germ layers were generated by changing the medium and adding inhibitors or cytokines when human PSC reached 60–70% confluency as previously reported⁴⁵.

For differentiation, cells were treated as previously described⁴⁵.

Briefly, for endoderm differentiation, hPSCs were cultured for 120 h in RPMI medium (Life Technologies) supplemented with 100 ng/ml Activin A (PEPROTECH, 120-14E-100), 50 nM/ml WNT3A (MedChemExpress, HY-P70453A), 0.5% FBS (BIOWEST, S1580), 2X GlutaMax (Thermo Fisher, 35050061), 0.2x MEM Non-Essential Amino Acids Solution (Thermo Fisher,11140050), and 55 µM β-mercaptoethanol (BBI, A600194).

For mesoderm differentiation, hPSCs were cultured in DMEM/F12 supplemented with 2X GlutaMax, 0.2X MEM Non-Essential Amino Acids Solution, 55 µM b-mercaptoethanol, 0.5% FBS, 100 ng/ml VEGF (PEPROTECH,100-20-50), 100 ng/ml BMP4 (PEPROTECH, 120-05ET), 10 ng/ml bFGF (PEPROTECH, 100-18B-100), and 100 ng/ml Activin A. From 24 to 120 h of mesoderm differentiation, Activin A was removed from the culture medium.

For ectoderm differentiation, hPSCs were cultured in DMEM/F12 supplemented with 0.2X MEM Non-Essential Amino Acids Solution, and 55 µM b-mercaptoethanol, 15% KOSR (Life Technologies), 2 µM A83-01 (Tocris, 2939), 2 µM PNU-74654 (Tocris, 113906), and 2uM Dorsomorphin (Tocris, 3093).

Medium was changed daily, and the cells were collected after 5 days. The success of each differentiation experiment was confirmed by the expression of lineage specific marker genes using RT-qPCR quantification. Oligonucleotides are listed in Supplementary Data 3.

Long-read and short-read sequencing

RNA was isolated using RNAzol RT (MRC, RN190) according to the manufacturer’s protocol. Samples for short-read RNA-Seq was prepared for sequencing with RNA-seq NEB Next Ultra RNA Library Prep Kit (NEB, #7530). Short-read RNA sequencing was performed on Illumina Novaseq 6000 platform. Samples for long-read RNA sequencing were purified by the RNAeasy Mini kit. The RNA library generated with PacBio binding kit (Pacbio, 101-849-000) was sequenced on Sequel II (Pacbio).

Data generation and retrieval

PacBio long-read bulk RNA-Seq data for the four cell states (PSC, endoderm, mesoderm and ectoderm) were generated in duplicate for the purpose of transcript assembly and long-read quantification. For each cell state, short-read bulk RNA-Seq data were also generated in duplicate for the purpose of expression quantification. We also generated replicated sub-cellular RNA-Seq data for nucleus, cytoplasm, chromatin and nucleoplasm for the four cell states. Further, we generated replicated stability RNA-Seq data for the four cell states 0, 1 and 8 h after actinomycin B treatment.

We retrieved a collection of short-read RNA-seq data generated from multiple studies. These data were previously analyzed for a different purpose²⁸. Short-read bulk RNA-Seq hepatocyte data to investigate transcript expressions in hepatocytes were retried from two studies^99,100. The scRNA-seq data and endoderm differentiation bulk RNA-Seq data were retrieved from Chu et al.⁴⁷. We also retrieved the mesoderm, endoderm and ectoderm scRNA-seq data from a human in vivo gastrulating embryo⁴⁸. POLR2A ChIP-Seq and a portion of deepCAGE data were retrieved from Encode database¹⁰¹ (https://www.encodeproject.org). Another set of deepCAGE data were retrieved from Poletti et al.¹⁰². Polyadenylated data from previous studies^103,104 were retrieved. For the analyses of histone modification in hPSC, published data from two studies^105,106 were retrieved. Nanopore long-read data of human early development⁴⁶ were retrieved to confirm the identify of our cell states. Short-read RNA-Seq data were retrieved from Tchieu et al.⁵⁰ to further confirm the identities of the ectoderm cells and investigate TE dysregulation in different ectodermal lineages.

Transcript assembly and assembly quality evaluation

Independent transcript assembly based on long-read RNA-Seq data was done for each cell state. Transcript assembly was done using IsoSeq V3 pipeline of the Pacific Biosciences (PacBio) (https://github.com/PacificBiosciences/IsoSeq/blob/master/isoseq-clustering.md). For each sample, the high-fidelity (HiFi) consensus reads were first generated from the raw PacBio subreads using ccs (https://github.com/PacificBiosciences/ccs). The following parameters were used: --min-rq 0.9 --min-passes 1 --min-snr 1. The full-length sequences were then generated by primer removal and demultiplexing using lima (https://github.com/lima-vm/lima). The full-length sequences were refined by isoseq3 (https://github.com/PacificBiosciences/IsoSeq) to remove the noise. The noiseless full-length sequences were then converted to the fastq format using PacBio’s bam2fastq (https://github.com/PacificBiosciences/pbtk#bam2fastx). After the processing of the PacBio raw reads, replicates of the same cell state were merged.

For the mapping of the noiseless full-length sequences to human genome, Minimap2³⁵ reference was prepared using version GRCh38 primary assembly of human genome and version 43 (GENCODEv43) of human reference GTF obtained from the GENCODE database⁵³. GENCODE reference GTF junctions were made using paftools in Minimap2 package. The sequence mapping was done using Minimap2. The output alignment was then sorted using SAMtools¹⁰⁷. The transcript assembly was sone using StringTie^36,37. For StringTie assembly, the following parameters were used: -s 2 -c 1 -L. The assembled transcripts were then filtered to remove transcripts with undefined strand or those shorter than 200 bp. The initial quality assessment of our assembly was based on the strand information and the transcript lengths. Further quality assessments were done using different genomic data and read coverage and conservation analyses. For expression comparison across different cell states, the assemblies from each cell state were merged into a human early development transcript (hEDT) reference using gffcompare¹⁰⁸.

Splice junction coverage

For splice coverage computation, Bowtie2¹⁰⁹ reference was first built from the fasta sequences of the assembled transcripts. Paired-end short read transcriptome sequences were then mapped to the built reference using Bowtie2. Using a custom python script (junction_coverage.py) with the assembly GTF file and bowtie2 alignment as inputs, the numbers of reads covering each splice junction was estimated. Only the coverage with at least 10-bp anchor length was considered.

Splice site conservation

For splice conservation, the flanking regions, including 10 bp in the exons and 10 bp in the introns were extracted for both the 5’ and 3’ sides of the splice junction. The 100-way phyloP conservation scores¹¹⁰ for version GRCh38 of the human genome was downloaded from the UCSC database¹¹¹. Using a custom Python script (wig_overlap_range_full_use.py), the PhyloP conservation score for each position, relative to the splice junction of each multi-exon transcript, was computed.

Splice junction nucleotide frequencies

The flanking regions, including 10 bp in the exons and 10 bp in the introns were extracted separately for both acceptor and donor splice junctions. The nucleotide sequences for each junction were then extracted using BEDtools. The nucleotide frequencies were then computed using WebLogo 3^112,113.

Alternative splicing event reference count

SUPPA2¹¹⁴ was used to investigate the numbers of alternative splicing event. The GTF files of hEDT assemblies were first converted to SUPPA-recognized GTF file. Then, generateEvents in SUPPA package was used to generate individual splicing events. The number of each event in all the assemblies were then counted.

Generation of random GTF sequences

For the generation of random GTF, the gene coordinates of the transcripts were first extracted. Then random coordinates of input gene coordinates were then shuffled with BEDtools¹¹⁵ shuffle. Using the marked shuffled coordinates, random transcripts were then extracted using the original isoform and exon structures. This procedure ensured that random sequences are found on the same chromosome, have the same lengths and isoform/exon structure as the original input GTF. A custom random GTF generator Python script was written to accomplish this purpose.

Transcript assembly pileup and heatmap for POLR2A and histone modification ChIP sequencing, deepCAGE and polyA tail sequencing

For short read heatmaps, Bowtie2¹⁰⁹ genome reference was built using the GRCh38.p13 version of human genome. The reads were then mapped to the Bowtie2 genome reference using Bowtie2. For long reads, the alignment made with Minimap2³⁵ was used. The alignment was sorted and the alignment index was prepared with SAMtools¹⁰⁷. Alignment was done individually for each sample. Individual alignment was then converted to bigwig file using bamCoverage in deepTools2¹¹⁶. For samples and regions to be plotted together, computeMatrix of deepTools2 package was used to prepare the matrix. The matrix was then plotted using plotHeatmap of deepTools package.

Estimation of the assembly exhaustiveness using short-read data

To investigate the proportion of potential transcripts that might have been missed by our assembly procedure, we took advantage of the short-read RNA-seq data. For each cell states, the short-read RNA-seq data were first cleaned with fastp. The reads were then mapped to the assembled transcripts using bowtie2 with --very-sensitive-local option. The overall alignment rates and the alignment rates of concordant read pairs were then retrieved. The percentages of the short-read data mappable to the assemblies from the long-read data were used as the indicators of the exhaustiveness of the exhaustiveness of our transcript assembly.

Coding potential and coding sequence prediction

FEELnc⁵² was used for the prediction of coding potential of the assembled transcripts for each cell state and the merged assembly. Version 43 of the human protein coding (pc) and long noncoding RNA (lncRNA) transcript fasta sequences obtained from the GENCODE database were used for the random forest training. Using the version GRCh38 primary assembly genome, FEELnc_codpot.pl was used to estimate the coding potential of the transcripts.

For the CDS annotation of the assembled transcripts, version 42 of the amino acid sequences of human protein-coding transcripts were downloaded from the GENCODE database⁵³. The sequences were used to make a BLAST¹¹⁷ database. Pfam-A.hmm database¹¹⁸ was also downloaded. CDS prediction started with the extraction of the longest open reading frame (orf) using TransDecoder.LongOrfs of the TransDecoder package (https://github.com/TransDecoder/TransDecoder) from the Trinity transcript assembler^119,120. The homology searches for the closest amino acid sequence of the longest orf were then conducted using BLASTP¹¹⁷ and hmmscan¹²¹. The outputs of the homology searches were then used for the prediction of CDS in the transcripts using TransDecoder.Predict in the TransDecoder package.

TE search

TE searches were performed individually for the assembly of each cell state, and for the merged assembly. The assembled transcripts were converted to fasta format using BEDTools¹¹⁵. For the search of TEs in the assembled transcripts, HMM TE sequence alignment was obtained from the version Dfam_3.6 of Dfam database¹²². TE search was conducted using nhmmer⁵⁴. TE splicing frequency was computed using a custom python script. For noncoding transcripts, body of the transcripts were divided into 20 equal bins. Then proportion of transcripts with the specified TE was then computed for each bin position. For coding transcripts, the transcripts were first divided into 5’ UTR, CDS and 3’ UTR. Each region was divided into 20 bins. For each region, the proportion of transcripts with the specified TE was then computed for each bin position.

For the analyses of TE presence at the TSS and TES, the TSS and TES for each transcript were first extracted. For each transcript, the regions of TSS and TES including 2 kbp upstream and downstream were extracted to make a 4 kbp region for each transcript end. The sequences of the TSS and TES regions were then searched in the TE database. The regions were then divided into 20 bins after which the proportion of the transcript with the specified TE was estimated for each bin.

Analysis of the consequences of TE chimeric transcripts

We focused our analyses of the consequence of TE chimeric transcripts on genes with both TE-containing and TE-free transcripts. We compared the exon structure of TE-containing transcript to the closest TE-free transcript. The closest TE-free transcript was defined as the transcript with the highest overlapping exon coverage. FEELnc⁵² prediction and CDS length from TransDecoder in Trinity assembler^119,120 were used to infer the consequences of TE insertion. The consequence was defined with respect to the coding potential and the CDS length. If both TE-containing and the closest TE-free transcripts were both noncoding, the consequence was defined noncoding. When the TE-free was predicted noncoding and the TE-containing was predicted coding, the consequence was termed creating. For TE-free coding transcript associated with TE-containing noncoding transcript, the consequence was termed disrupting. When both TE-containing and TE-free were both coding and their CDS lengths were within 90% of the other, the consequence was termed preserving. When both coding and non-coding were both predicted coding but the CDS lengths were not within 90% of the other, the consequence was termed modifying.

Differentially spliced TEs

To identify TEs that tend to be more frequently retrieved in the assembly of a specific cell states, the numbers of transcripts with or without the specified TE was compared. The comparisons were done separately for coding and noncoding transcripts. Because the differentiation process started with PSCs, the assemblies of the three germ layers were compared to the PSC assembly. Fisher exact tests followed by Bonferroni multiple test corrections were used to test statistical significance. Odd ratio threshold of 2 was set for statistical significance. Matrix of the proportion of transcripts containing the significant TEs were then made.

Long-read alignment and transcript quantification

Long read alignments were made with minimap2³⁵. Noiseless full-length sequences were aligned to the assembled transcripts. Transcript quantification was done using a custom Python script that quantifies the number of sequences that map to each transcript. When a sequence maps to more than one transcript, the sequence is assigned to the transcript with the highest percent identity and aligned transcript coverage.

Cell state-specific transcript with long-read quantification

For the evaluation of the cell state specific expression, previously reported MGFR¹²³ steps were adapted. GC-content expression normalization was done using EDASEQ¹²⁴. Then the samples were ranked by expression samples. Transcripts for which all replicates of a cell state were ranked highest were retained. Using DESeq2¹²⁵ to extract differentially expressed transcripts, transcripts for which the highest-ranking cell state are not significantly different from others were discarded. To get the marker transcripts, transcripts expressed at less than 10 in the highest-ranking cell state or more than 10 in other cell states were discarded. Additionally, for a marker transcript, every replicate of a marked cell state must have at least 10-fold higher expression compared to any of the replicate of the non-marked cell states. For the extraction of the top uniformly expressed transcripts, the transcripts were ranked by the expression variation (coefficient of variation). The top 20 uniformly expressed transcripts with the lowest expression variation were presented.

Short-read alignment and transcript quantification

Short read transcript assembly was done using STAR^126,127. STAR references were made from the assembled transcripts. Short read sequences were then mapped to the STAR references. For short read quantification, RSEM STAR references were made using the assembled transcripts. The short read sequences were first filtered with fastp¹²⁸. The filtered SR sequences were then used for quantification using STAR and RSEM¹²⁹. DESeq2 was used for the identification of differentially expressed transcripts. Adjusted P value of 0.05 was used as the threshold for statistical significance.

Human embryonic skin fibroblast (hESF) culturing and analyses

We used human embryonic skin fibroblasts, hESF (CCC-ESF-1; BNCC100546; RRID:CVCL_6884) to investigate the influence of differentiation-inducing factors on the transcriptional landscape to establish that the expression patterns were due to the actual change in expression patterns, and not just reflection of different factor treatments. The initial hESF cells were grown in 90% DMEM-H basal medium supplemented with 10% fetal bovine serum (FBS, NATOCOR). The medium was changed every day. We cultured the cell lines in endoderm and mesoderm media for five days with the normal differentiation-inducing factors, mimicking the differentiation process from the hPSCs. As a control, hESFs were also cultured in the normal hESF medium for five days. At the end of the fifth day, RNA was extracted and sequenced on short-read platforms. Transcript quantification was done as for hPSC-derived cell lines. Cell clustering analyses were then done using DESeq2.

Functional single nucleotide polymorphism (SNP) analyses for novel and matching transcripts

Clinically-associated and phenotype-associated SNPs were retrieved from version 113 of Ensembl database¹³⁰. The number of the functionally relevant SNPs found per 1000 bp (kbp) or 1,000,000 bp (Mbp) of the transcripts were then computed using a custom Python script. Similar analyses were performed for the matched random transcripts (with same length and exon structure). Functional SNP analyses were focused on the exons of the novel and matching transcripts for each cell lineage. Fisher Exact test was used to test statistical significance of the differences.

Short read differential TE expression

To investigate TEs that were differentially expressed during differentiation process, short read transcript expression quantification was performed using the merged reference for all the replicates and cell states. The average TPM expressions for all transcripts were computed. The expressions were log2-transfornmed. The expression levels of all TE-containing transcripts in each of the three germ layers were compared to PSC levels. T-test was used to test the statistical significance of average expression of transcripts containing specified TE. Additional, 2-fold expression change was required for a TE to be considered significantly differentially expressed.

Genomic data visualization using Integrative Genomics Viewer (IGV)

Visualization of genome regions of interest was done using version 2.17.4 of the Integrative Genomics Viewer¹³¹. For the high-throughput data to be visualized, the sorted alignment files were converted to bigwig format using bamCoverage in deepTools2¹¹⁶. Count per million (CPM) normalization with bin size of 10 bp was used. For TE visualization, repeat mask file of Hg38 human genome was retrieved from the UCSC database¹¹¹. BED file was then made using a custom Python script. Only LINE, SINE, LTR, Retroposon and DNA TEs were included in the TE track. Each of the five TEs were assigned different color tracks.

Single cell RNA (scRNA) transcript quantification

For scRNA-Seq data, SR sequences were first filtered with fastp¹²⁸. Transcript quantification was done with RSEM¹²⁹ using the merged hEDT assembly. scRNA-Seq was merged into a matrix. scRNA-Seq processing was done using Seurat^132,133. Transcripts with at least 2 expression count in less than 200 cells were discarded. Cells with at least 25% mitochondria-derived or 40% ribosome-derived transcripts were discarded. The filtered matrix was normalized using logNormalize. Using vst method, the top 10,000 variable transcripts were retained. The matrix was then used to run the PCA. The optimum number of dimensions was decided from 1 to 50 based on the Elbow analysis. Cell clustering was done using UMAP. For the expression of multiple transcripts containing the same TE, MetaFeature of Surat was used to merged the TE-containing transcripts. FeaturePlot was used to plot the expression levels of the individual and merged transcripts.

RNA-seq from subcellular structures

RNA extraction from subcellular structures were performed according to the reported pipelines¹³⁴. The plated cells were digested and washed once with ice-cold PBS. Cytoplasmic fractions were isolated by incubating the cells in ice-cold cytoplasmic lysis buffer containing 10 mM Tris-HCl pH 7.5, 0.05% NP-40, 150 mM NaCl, 25 μM α-amanitin (Sigma, A2263), 40 U/ml SUPERase.IN (Thermo Fisher, AM2694) and 1× complete protease inhibitor mix (Sigma, 11697498001). Cell lysate was carefully layered on top of a new 1.5 ml tube containing 500 μl of sucrose buffer (10 mM Tris-HCl pH 7.0,150 mM NaCl 25% (wt/vol) sucrose, 25 μM α-amanitin, 20 U SUPERase.IN, and 1x Protease inhibitor mix). The sample was centrifuged for 10 min at ~2000 x g in 4 °C. The cytoplasmic fraction was collected in the supernatant, transferred to a new 1.5 ml tube and put on ice. The nuclear pellet was washed with 500 μl of ice-cold PBS. Nuclear RNA fraction was extracted from a portion of the RNA pellet.

The nuclear pellet was resuspended in 100 μl of nuclear resuspension buffer (20 mM Tris-HCl (pH 8.0), 75 mM NaCl, 0.5 mM EDTA, 50% (vol/vol) glycerol, 0.85 mM DTT, 25 μM α-amanitin, 10 U SUPERase.IN, and 1x Protease inhibitor mix). Next, 100 μl of nuclear lysis buffer (1% (vol/vol) NP-40, 20 mM HEPES (pH 7.5), 300 mM NaCl, 1 M urea 0.2 mM EDTA, 1 mM DTT 25 μM α-amanitin,10 U SUPERase.IN, and 1x Protease inhibitor mix) was introduced, and vortexed for 5 s. After 3 min of ice incubation, the sample was centrifuged at approx. 26,000 × g at 4 °C for 2 min. Nucleoplasmic fraction was collected in a new 1.5-ml tube as the supernatant. The chromatin-associated fraction was collected as the pellet on ice-cold PBS.

RNA extraction for nucleus- and chromatin-associated fraction was carried out by adding 100 ul PBS to resuspend the pellet. RNA was isolated by adding 500 ul RNAzol RT (MRC, RN190), according to manufacturer’s protocol. RNA suspended in supernatant fraction (including cytoplasmic and nucleoplasmic fraction) was extracted by the TRIZOL-PLS region. In short, 3.5x sample volumes RLT buffer introduced and mixed. The initial sample was added to 2.5x of the ice-cold absolute ethanol. The sample was mixed before centrifuging at 16,000 g at 4 °C for 15 min. The ethanol was removed, resuspended and washed twice with 75% ethanol. The dry pellet was dissolved with RNase/DNase-free water. The success of the sub-cellular RNA extraction was confirmed by using western blotting for the genes known to be preferentially localized to specific sub-cellular structure: HNRNPU (1:2000, Santa Cruz; sc-32315), H2B (1:2000, Abcam; ab1790), GAPDH (1:2000, Novus; NB300-221SS) and LMNB1 (1:2000, Abcam; ab133741). Raw unprocessed western blots are available in Supplementary Data 4.

Computation of transcript relative concentration index (RCI)

For the quantification of subcellular RCI, short read transcript quantification was first done using RSEM¹²⁹ and STAR¹²⁶ as highlighted above. RCI were then computed using the average expression of each transcript as previously reported^12,84. For transcript x with the average TPM expressions xA in subcellular location A and xB in subcellular location B, the RCI for subcellular location B relative to location B is given as (1):

$${RCIA}/B=\frac{\log 2({xB}+0.01)}{\log 2({xA}+0.01)}$$

(1)

Retrieval of transcripts localized to each subcellular structure

The retrieval of differentially expressed transcripts was done at two levels for each cell state. For the extraction of transcripts localized to either cytoplasm and nucleus, differential transcript expression was conducted between cytoplasm and nucleus using DESeq2¹²⁵. Differentially localized transcripts were defined as transcripts with ≥1 absolute log-fold change and the adjusted p values ≤ 0.05 between the two subcellular structures. Transcripts were then classified into cytoplasm or nucleus transcripts. Next, we conducted the subcellular structure localization for chromatin and cytoplasm using the same procedure. Consequently, transcripts localized to different subcellular structures in the four states were retrieved.

Subcellular structure consistency

To investigate the consistency of subcellular transcript localization across cell states, we computed Jaccard similarity scores between the transcript sets localized to the same structures across each pair of cell states. Given Ax, the set of transcripts localized to subcellular location x in cell A, and Bx, the set of transcripts localized to subcellular location x in cell B, the Jaccard similarity score is given as (2);

$${{{\rm{J}}}}({{{\rm{Ax}}}},{{{\rm{Bx}}}})=|{{{\rm{Ax}}}} \cap {{{\rm{Bx}}}}|/|{{{\rm{Ax}}}} \cup {{{\rm{Bx}}}}|$$

(2)

Transcription factor binding motif enrichment

Transcription factor binding motif (TFBM) analyses were done using homer2¹³⁵. Known vertebrate motifs in the homer database (472 in number) were used for the analyses. We focused our search on core promoters (200 bp) upstream of the transcription start site. For each transcript categories, we extracted the 200 bp promoter regions. Overlapping promoter regions were concatenated such that no genomic region was used more than once. Significant enrichment was defined as the homer adjusted p value of 0.05. P value adjustment was done using Bonferroni correction. To investigate the TFBM enrichment during lineage differentiation, differentially expressed transcripts between PSCs and each of the three germ layers were extracted. These produced transcripts that were upregulated (up) or downregulated (down) in each germ layer cell. Transcripts classified as hPSC-enriched included transcripts with significantly higher expression in hPSC than in at least one of the differentiated cells. TFBM analyses were done independently for TE-containing and TE-free transcripts. As background, we retrieved transcripts with at least 10 DESeq2-normalized average expression but were not classified as differentially expressed by our definition. The procedure would identify TFBMs that might be relevant in each differentiation process. The significantly enriched TFBMs were shown in the scatter plots while the names of the five most significant TFs were shown. To compare the TFBM enrichment for HERVH transcripts in hPSC and endoderm, we extracted the sequences of HERV transcripts with significantly different expression in hPSC and endoderm. To get TFBM overrepresentation in endoderm-biased HERVH transcripts, hPSC-biased HERVH transcripts were used as background. Conversely, endoderm-biased HERVH transcripts were used as background to get hPSC-overrepresented TFBM.

RNA-binding protein (RBP) enrichment analyses for differentially expressed and localized transcripts

To investigate if transcript differential expression was driven by known RNA-binding proteins, we checked the enrichment of localized transcripts in each cell lineage using AME¹³⁶ in version 5.5.7 of The MEME Suite¹³⁷. Previously published human RBP motifs¹³⁸ available in The MEME Suite was used for the analyses. Differentially expressed TE-containing and TE-free transcripts and control transcripts were retrieved as described under TFBM analyses. The nucleotide sequences of the spliced transcripts were then extracted. To investigate RBPs that differentially bind TE-containing transcripts in each subcellular structure, we classified all localized transcripts into TE-containing and TE-free sets. We then used TE-free transcripts as control transcripts to identify RBP motifs with overrepresentation in TE transcripts across different subcellular structures in each cell states. Enrichment significance was set at the adjusted AME P value of 0.05.

Computation of transcript relative stability

Human PSC and differentiated three germ cells were cultured in 6-well plates. Then Actinomycin B was added to inhibit the synthesis of new RNA. RNA was extracted at 0, 1, and 8 h after the addition of Actinomycin B. Short-read sequencing was then done for RNA extracted at the time points after transcription stoppage. For each time point, short-read transcript quantification was first done using RSEM and STAR as highlighted above. Relative stability was then computed using the average expression of each transcript as the specified time, relative to the expression at time 0 h after actinomycin treatment. For transcript x with the average TPM expressions x0 at time 0 hr and xT at time T h, relative stability is given as (3):

$${Relative\; stability}=\frac{\log 2({xT}+0.01)}{\log 2(x0+0.01)}$$

(3)

Bin analyses of TE enrichment in ranked transcripts

TE enrichments of binned ranked transcripts were computed for transcript stability and sub-cellular localization. For each transcript, relative stability or relative concentration index (RCI) were first calculated as described earlier. The transcripts were then ranked in ascending orders. The ranked transcripts were divided into 20 bins. Given a total of N transcripts with detectable expression in the assembly, with NTE transcripts containing the specific TE of interest, for each bin containing n transcripts, including nTE with the specific TE, TE enrichment for the bin is computed as (4);

$${Bin\; TE\; enrichment}=\log 2\left(\frac{\frac{{nTE}}{n}}{\frac{{NTE}}{N}}\right)=\log 2\left(\frac{{nTE}}{n}\right)-\log 2\left(\frac{{NTE}}{N}\right)$$

(4)

Statistics & reproducibility

All statistical tests were conducted using Scipy in Python3. Statistical comparison of average log2-transformed values was done with T-test. For the comparison of median values, Mann-Whitney U test was employed. Kruskal-Wallis test was performed to check the homogeneity of median values. Fisher exact test was used to compare the distribution of 2 × 2 contingency table. To compare the distributions with more than two values, Chi square tests were done. In general, three biological replicates were generated, although the long-read was performed in biological duplicate for endoderm, mesoderm and ectoderm cell types. Some samples failed quality control and were omitted from the analysis. In this case, two biological replicates may only be available. Biological replicates were chosen based on cost versus benefit and is typical for the current state of the field. No statistical method was used to predetermine sample size.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Data availability

The sequencing data generated in this study have been deposited in the Gene Expression Omnibus under accession codes GSE269270, GSE269272, GSE269273, and GSE269274. Source Data are provided with this paper Source data are provided with this paper.

Code availability

Python scripts used in the analyses and the Seurat R commands used for scRNA analyses are deposited in (https://github.com/iababarinde/Transposable-Elements-Human-Gastrulation).

References

Hutchins, A. P. & Pei, D. Transposable elements at the center of the crossroads between embryogenesis, embryonic stem cells, reprogramming, and long non-coding RNAs. Sci. Bull. (Beijing) 60, 1722–1733 (2015).
Article ADS PubMed CAS Google Scholar
Ma, G., Babarinde, I. A., Zhou, X. & Hutchins, A. P. Transposable elements in pluripotent stem cells and human disease. Front Genet 13, 902541 (2022).
Biémont, C. A brief history of the status of transposable elements: from junk DNA to major players in evolution. Genetics 186, 1085–1093 (2010).
Article PubMed PubMed Central Google Scholar
Bourque, G. et al. Ten things you should know about transposable elements. Genome Biol. 19, 199 (2018).
Article PubMed PubMed Central CAS Google Scholar
Osmanski, A. B. et al. Insights into mammalian TE diversity through the curation of 248 genome assemblies. Science (1979) 380, eabn1430 (2023).
Naville, M. et al. Not so bad after all: retroviruses and long terminal repeat retrotransposons as a source of new genes in vertebrates. Clin. Microbiol. Infect. 22, 312–323 (2016).
Article PubMed CAS Google Scholar
Kim, Y.-J., Lee, J. & Han, K. Transposable elements: no more ‘Junk DNA’. Genomics Inf. 10, 226–233 (2012).
Article Google Scholar
Babarinde, I. A. & Saitou, N. Tempo and mode of conserved noncoding sequence evolution among four mammalian orders. Genome Biol. Evol. 5, 2330–2343 (2013).
Article PubMed PubMed Central Google Scholar
Wang, J., Huang, J., & Shi, G. Retrotransposons in pluripotent stem cells. Cell Regen. 9, 4 (2020).
Wang, J. et al. Primate-specific endogenous retrovirus-driven transcription defines naive-like stem cells. Nature 516, 405–409 (2014).
Elbarbary, R. A., Lucas, B. A. & Maquat, L. E. Retrotransposons as regulators of gene expression. Science (1979) 351, aac7247 (2016).
Babarinde, I. A. et al. Transposable element sequence fragments incorporated into coding and noncoding transcripts modulate the transcriptome of human pluripotent stem cells. Nucleic Acids Res 49, 9132–9153 (2021).
Article PubMed PubMed Central CAS Google Scholar
Fueyo, R., Judd, J., Feschotte, C. & Wysocka, J. Roles of transposable elements in the regulation of mammalian transcription. Nat. Rev. Mol. Cell Biol. 23, 481–497 (2022).
Article PubMed PubMed Central CAS Google Scholar
Farré, D., Engel, P. & Angulo, A. Novel role of 3’UTR-Embedded Alu Elements as facilitators of processed pseudogene genesis and host gene capture by viral genomes. PLoS One 11, e0169196 (2016).
Article PubMed PubMed Central Google Scholar
Gong, C. & Maquat, L. E. E. lncRNAs transactivate STAU1-mediated mRNA decay by duplexing with 3’ UTRs via Alu elements. Nature 470, 284–288 (2011).
Article ADS PubMed PubMed Central CAS Google Scholar
Raviram, R. et al. Analysis of 3D genomic interactions identifies candidate host genes that transposable elements potentially regulate. Genome Biol. 19, 216 (2018).
Article PubMed PubMed Central CAS Google Scholar
Sun, L., Fu, X., Ma, G. & Hutchins, A. P. Chromatin and epigenetic rearrangements in embryonic stem cell fate transitions. Front Cell Dev. Biol. 9, 174 (2021).
Google Scholar
Garcia-Perez, J. L., Widmann, T. J. & Adams, I. R. The impact of transposable elements on mammalian development. Development 143, 4101–4114 (2016).
Article PubMed CAS Google Scholar
Hackett, J. A., Kobayashi, T., Dietmann, S. & Surani, M. A. Activation of lineage regulators and transposable elements across a pluripotent spectrum. Stem Cell Rep. 8, 1645–1658 (2017).
Article CAS Google Scholar
Criscione, S. W., Zhang, Y., Thompson, W., Sedivy, J. M. & Neretti, N. Transcriptional landscape of repetitive elements in normal and cancer human cells. BMC Genomics 15, 583 (2014).
Article PubMed PubMed Central Google Scholar
LaRocca, T. J., Cavalier, A. N. & Wahl, D. Repetitive elements as a transcriptomic marker of aging: evidence in multiple datasets and models. Aging Cell 19, e13167 (2020).
Article PubMed PubMed Central Google Scholar
Kruse, K. et al. Transposable elements drive reorganisation of 3D chromatin during early embryogenesis. Preprint at bioRxiv https://doi.org/10.1101/523712 (2019).
Nellåker, C. et al. The genomic landscape shaped by selection on transposable elements across 18 mouse strains. Genome Biol. 13, R45 (2012).
Article PubMed PubMed Central Google Scholar
Gray, Y. H. M. It takes two transposons to tango:transposable-element-mediated chromosomal rearrangements. Trends Genet. 16, 461–468 (2000).
Article PubMed CAS Google Scholar
Chishima, T. et al. Identification of transposable elements contributing to tissue-specific expression of long non-coding RNAs. Genes (Basel) 9, 23 (2018).
Article PubMed Google Scholar
Berrens, R. V. et al. Locus-specific expression of transposable elements in single cells with CELLO-seq. Nat. Biotechnol. 40, 546–554 (2022).
Article PubMed CAS Google Scholar
Babarinde, I. A., Li, Y. & Hutchins, A. P. Computational methods for mapping, assembly and quantification for coding and non-coding transcripts. Comput Struct. Biotechnol. J. 17, 628–637 (2019).
Article PubMed PubMed Central CAS Google Scholar
Babarinde, I. A. & Hutchins, A. P. The effects of sequencing depth on the assembly of coding and noncoding transcripts in the human genome. BMC Genomics 23, 487 (2022).
Article PubMed PubMed Central CAS Google Scholar
Shumate, A., Wong, B., Pertea, G. & Pertea, M. Improved transcriptome assembly using a hybrid of long and short reads with StringTie. PLoS Comput Biol. 18, e1009730 (2022).
Article ADS PubMed PubMed Central CAS Google Scholar
Lagarde, J. et al. High-throughput annotation of full-length long noncoding RNAs with capture long-read sequencing. Nat. Genet 49, 1731–1740 (2017).
Article PubMed PubMed Central CAS Google Scholar
He, J. et al. Identifying transposable element expression dynamics and heterogeneity during development at the single-cell level with a processing pipeline scTE. Nat. Commun. 12, 1456 (2021).
Article ADS PubMed PubMed Central CAS Google Scholar
Rodríguez-Quiroz, R. & Valdebenito-Maturana, B. SoloTE for improved analysis of transposable elements in single-cell RNA-Seq data using locus-specific expression. Commun. Biol. 5, 1063 (2022).
Article PubMed PubMed Central Google Scholar
Shao, W. & Wang, T. Transcript assembly improves expression quantification of transposable elements in single-cell RNA-seq data. Genome Res 31, 88–100 (2021).
Article MathSciNet PubMed PubMed Central CAS Google Scholar
Bayega, A., Fahiminiya, S., Oikonomopoulos, S. & Ragoussis, J. Current and future methods for mRNA analysis: a drive toward single molecule sequencing. In Methods in Mol. Biol. 1783, 209–241 (Humana Press, New York, NY, 2018).
Li, H. Minimap2: pairwise alignment for nucleotide sequences. Bioinformatics 34, 3094–3100 (2018).
Article PubMed PubMed Central CAS Google Scholar
Pertea, M. et al. StringTie enables improved reconstruction of a transcriptome from RNA-seq reads. Nat. Biotechnol. 33, 290–295 (2015).
Article PubMed PubMed Central CAS Google Scholar
Pertea, M., Kim, D., Pertea, G. M., Leek, J. T. & Salzberg, S. L. Transcript-level expression analysis of RNA-seq experiments with HISAT, StringTie and Ballgown. Nat. Protoc. 11, 1650–1667 (2016).
Article PubMed PubMed Central CAS Google Scholar
Chen, Y., Zhang, Y., Wang, A. Y., Gao, M. & Chong, Z. Accurate long-read de novo assembly evaluation with inspector. Genome Biol. 22, 12 (2021).
Watson, M. & Warr, A. Errors in long-read assemblies can critically affect protein prediction. Nat. Biotechnol. 37, 124–126 (2019).
Ono, Y., Hamada, M. & Asai, K. PBSIM3: a simulator for all types of PacBio and ONT long reads. NAR Genom. Bioinform. 4, lqac092 (2022).
Nip, K. M. et al. Reference-free assembly of long-read transcriptome sequencing data with RNA-Bloom2. Nat. Commun. 14, 2940 (2023).
Article ADS PubMed PubMed Central CAS Google Scholar
Pardo-Palacios, F. J. et al. Systematic assessment of long-read RNA-seq methods for transcript identification and quantification. Nat. Methods 21, 1349–1363 (2024).
Article PubMed PubMed Central CAS Google Scholar
Tung, L. H., Shao, M. & Kingsford, C. Quantifying the benefit offered by transcript assembly with Scallop-LR on single-molecule long reads. Genome Biol. 20, 287 (2019).
Article PubMed PubMed Central CAS Google Scholar
Abugessaisa, I. et al. FANTOM5 CAGE profiles of human and mouse reprocessed for GRCh38 and GRCm38 genome assemblies. Sci. Data 4, 170107 (2017).
Article PubMed PubMed Central CAS Google Scholar
Tsankov, A. M. et al. Transcription factor binding dynamics during human ES cell differentiation. Nature 518, 344–349 (2015).
Article ADS PubMed PubMed Central CAS Google Scholar
Dobner, J., Diecke, S., Krutmann, J., Prigione, A. & Rossi, A. Reassessment of marker genes in human induced pluripotent stem cells for enhanced quality control. Nat. Commun. 15, 8547 (2024).
Article PubMed PubMed Central Google Scholar
Chu, L. F. et al. Single-cell RNA-seq reveals novel regulators of human embryonic stem cell differentiation to definitive endoderm. Genome Biol. 17, 173 (2016).
Article PubMed PubMed Central Google Scholar
Tyser, R. C. V. et al. Single-cell transcriptomic characterization of a gastrulating human embryo. Nature 600, 285–289 (2021).
Article ADS PubMed PubMed Central CAS Google Scholar
Rhodes, K. et al. Human embryoid bodies as a novel system for genomic studies of functionally diverse cell types. Elife 11, e71361 (2022).
Article PubMed PubMed Central CAS Google Scholar
Tchieu, J. et al. A modular platform for differentiation of human PSCs into all major ectodermal lineages. Cell Stem Cell 21, 399–410.e7 (2017).
Article PubMed PubMed Central CAS Google Scholar
Wang, L., Wang, S. & Li, W. RSeQC: quality control of RNA-seq experiments. Bioinformatics 28, 2184–2185 (2012).
Article PubMed CAS Google Scholar
Wucher, V. et al. FEELnc: a tool for long non-coding RNA annotation and its application to the dog transcriptome. Nucleic Acids Res 45, gkw1306 (2017).
Article Google Scholar
Frankish, A. et al. GENCODE: reference annotation for the human and mouse genomes in 2023. Nucleic Acids Res 51, D942–D949 (2023).
Article PubMed CAS Google Scholar
Wheeler, T. J. & Eddy, S. R. Nhmmer: DNA homology search with profile HMMs. Bioinformatics 29, 2487–2489 (2013).
Article PubMed PubMed Central CAS Google Scholar
Osteil, P. et al. MIXL1 activation in endoderm differentiation of human induced pluripotent stem cells. Stem Cell Rep. 20, 102482 (2025).
Li, J. et al. TGF-β2 and TGF-β1 differentially regulate the odontogenic and osteogenic differentiation of mesenchymal stem cells. Arch. Oral. Biol. 135, 105357 (2022).
Article PubMed CAS Google Scholar
Tang, Y. et al. TGFBR3 dependent mechanism of TGFB2 in smooth muscle cell differentiation and implications for TGFB2-related aortic aneurysm. Stem Cells Transl Med 14, (2025).
Yang, J. et al. GATA6-AS1 Regulates GATA6 Expression to Modulate Human Endoderm Differentiation. Stem Cell Rep. 15, 694–705 (2020).
Article CAS Google Scholar
Yasui, R., Matsui, A., Sekine, K., Okamoto, S. & Taniguchi, H. Highly Sensitive Detection of Human Pluripotent Stem Cells by Loop-Mediated Isothermal Amplification. Stem Cell Rev. Rep. 18, 2995–3007 (2022).
Article PubMed PubMed Central CAS Google Scholar
Sekine, K. et al. Robust detection of undifferentiated iPSC among differentiated cells. Sci. Rep. 10, 10293 (2020).
Article ADS PubMed PubMed Central CAS Google Scholar
Wu, P. R. et al. Wdr4 promotes cerebellar development and locomotion through Arhgap17-mediated Rac1 activation. Cell Death Dis. 14, 1–16 (2023).
Google Scholar
Goos, J. A. C. et al. Identification of causative variants in TXNL4A in Burn-McKeown syndrome and isolated choanal atresia. Eur. J. Hum. Genet. 25, 1126–1133 (2017).
Article PubMed PubMed Central CAS Google Scholar
Yao, J., Gaffaney, J. D., Kwon, S. E. & Chapman, E. R. Doc2 Is a Ca2+ Sensor Required for Asynchronous Neurotransmitter Release. Cell 147, 666–677 (2011).
Article PubMed PubMed Central CAS Google Scholar
Xue, M., Giagtzoglou, N. & Bellen, H. J. Dueling Ca2+ Sensors in Neurotransmitter Release. Cell 147, 491–493 (2011).
Article PubMed PubMed Central CAS Google Scholar
Pang, Z. P. et al. Doc2 Supports Spontaneous Synaptic Transmission by a Ca2 + -Independent Mechanism. Neuron 70, 244–251 (2011).
Article PubMed PubMed Central CAS Google Scholar
Sakashita, A. et al. Transcription of MERVL retrotransposons is required for preimplantation embryo development. Nat. Genet 55, 484–495 (2023).
Article PubMed PubMed Central CAS Google Scholar
Modzelewski, A. J. et al. A mouse-specific retrotransposon drives a conserved Cdk2ap1 isoform essential for development. Cell 184, 5541–5558.e22 (2021).
Article PubMed PubMed Central CAS Google Scholar
Cabili, M. N. et al. Integrative annotation of human large intergenic noncoding RNAs reveals global properties and specific subclasses. Genes Dev. 25, 1915–1927 (2011).
Article PubMed PubMed Central CAS Google Scholar
Santoni, F. A., Guerra, J. & Luban, J. HERV-H RNA is abundant in human embryonic stem cells and a precise marker for pluripotency. Retrovirology 9, 111 (2012).
Article PubMed PubMed Central CAS Google Scholar
Lu, X. et al. The retrovirus HERVH is a long noncoding RNA required for human embryonic stem cell identity. Nat. Struct. Mol. Biol. 21, 423–425 (2014).
Article ADS PubMed CAS Google Scholar
Zhang, Y. et al. Transcriptionally active HERV-H retrotransposons demarcate topologically associating domains in human pluripotent stem cells. Nat. Genet 51, 1380–1388 (2019).
Article PubMed PubMed Central CAS Google Scholar
Mattick, J. S. et al. Long non-coding RNAs: definitions, functions, challenges and recommendations. Nat. Rev. Mol. Cell Biol. 24, 430–447 (2023).
Article PubMed PubMed Central CAS Google Scholar
Harrow, J. et al. GENCODE: The reference human genome annotation for the ENCODE project. Genome Res 22, 1760–1774 (2012).
Article PubMed PubMed Central CAS Google Scholar
Billon, V. & Cristofari, G. Nascent RNA m6A modification at the heart of the gene–retrotransposon conflict. Cell Res 31, 829–831 (2021).
Article PubMed PubMed Central CAS Google Scholar
Liu, J. et al. The RNA m6A reader YTHDC1 silences retrotransposons and guards ES cell identity. Nature 591, 322–326 (2021).
Article ADS PubMed CAS Google Scholar
Chen, C. et al. Nuclear m6A reader YTHDC1 regulates the scaffold function of LINE1 RNA in mouse ESCs and early embryos. Protein Cell 12, 455–474 (2021).
Article PubMed PubMed Central CAS Google Scholar
Leclair, N. K. et al. Poison exon splicing regulates a coordinated network of SR protein expression during differentiation and tumorigenesis. Mol. Cell 80, 648–665.e9 (2020).
Article PubMed PubMed Central CAS Google Scholar
Kini, H. K., Kong, J. & Liebhaber, S. A. Cytoplasmic Poly(A) binding protein C4 serves a critical role in erythroid differentiation. Mol. Cell Biol. 34, 1300–1309 (2014).
Article PubMed PubMed Central Google Scholar
Byun, J. S. et al. The transcription factor PITX1 drives astrocyte differentiation by regulating the SOX9 gene. J. Biol. Chem. 295, 13677–13690 (2020).
Article PubMed PubMed Central CAS Google Scholar
Joshi, S. et al. TEAD transcription factors are required for normal primary myoblast differentiation in vitro and muscle regeneration in vivo. PLoS Genet 13, e1006600 (2017).
Article PubMed PubMed Central Google Scholar
Yuan, Y. et al. YAP1/TAZ-TEAD transcriptional networks maintain skin homeostasis by regulating cell proliferation and limiting KLF4 activity. Nat. Commun. 11, 1472 (2020).
Article ADS PubMed PubMed Central CAS Google Scholar
Carlevaro-Fita, J. et al. Ancient exapted transposable elements promote nuclear enrichment of human long noncoding RNAs. Genome Res 29, 208–222 (2019).
Article PubMed PubMed Central CAS Google Scholar
Lubelsky, Y. & Ulitsky, I. Sequences enriched in Alu repeats drive nuclear localization of long RNAs in human cells. Nature 555, 107–111 (2018).
Article ADS PubMed PubMed Central CAS Google Scholar
Mas-Ponte, D. et al. LncATLAS database for subcellular localization of long noncoding RNAs. RNA 23, 1080–1087 (2017).
Article PubMed PubMed Central CAS Google Scholar
Yin, Y. et al. U1 snRNP regulates chromatin retention of noncoding RNAs. Nature 580, 147–150 (2020).
Article ADS PubMed PubMed Central CAS Google Scholar
Mishra, K. & Kanduri, C. Understanding Long Noncoding RNA and Chromatin Interactions: What We Know So Far. Non-Coding RNA 5, 54 (2019).
Article PubMed PubMed Central CAS Google Scholar
Mustafin, R. N. The role of transposable elements in the differentiation of stem cells. Mol. Genet. Microbiol. Virol. (Russian version) 34, 67–74 (2019).
Article Google Scholar
Castro-Diaz, N. et al. Evolutionally dynamic L1 regulation in embryonic stem cells. Genes Dev. 28, 1397–1409 (2014).
Article PubMed PubMed Central CAS Google Scholar
Pereira Fernandes, D., Bitar, M., Jacobs, F. M. J. & Barry, G. Long Non-Coding RNAs in Neuronal Aging. Noncoding RNA 4, 12 (2018).
PubMed PubMed Central Google Scholar
Reilly, M. T., Faulkner, G. J., Dubnau, J., Ponomarev, I. & Gage, F. H. The role of transposable elements in health and diseases of the central nervous system. J. Neurosci. 33, 17577–17586 (2013).
Article PubMed PubMed Central CAS Google Scholar
Lapp, H. E. & Hunter, R. G. Early life exposures, neurodevelopmental disorders, and transposable elements. Neurobiol. Stress 11, 100174 (2019).
Article PubMed PubMed Central Google Scholar
Kelley, D. & Rinn, J. Transposable elements reveal a stem cell-specific class of long noncoding RNAs. Genome Biol. 13, R107 (2012).
Article PubMed PubMed Central Google Scholar
Kelley, D. R., Hendrickson, D. G., Tenen, D. & Rinn, J. L. Transposable elements modulate human RNA abundance and splicing via specific RNA-protein interactions. Genome Biol. 15, 537 (2014).
Article PubMed PubMed Central Google Scholar
Sexton, C. E., Tillett, R. L. & Han, M. V. The essential but enigmatic regulatory role of HERVH in pluripotency. Trends in Genetics. 38, https://doi.org/10.1016/j.tig.2021.07.007 (2022).
Göke, J. et al. Dynamic transcription of distinct classes of endogenous retroviral elements marks specific populations of early human embryonic cells. Cell Stem Cell 16, 135–141 (2015).
Article PubMed Google Scholar
Zhang, R. et al. A primate-specific endogenous retroviral envelope protein sequesters SFRP2 to regulate human cardiomyocyte development. Cell Stem Cell https://doi.org/10.1016/j.stem.2024.07.006 (2024).
Sekine, K., Onoguchi, M. & Hamada, M. Transposons contribute to the acquisition of cell type-specific cis-elements in the brain. Commun. Biol. 6, 631 (2023).
Article PubMed PubMed Central CAS Google Scholar
Mustafin, R. N. & Khusnutdinova, E. K. Involvement of transposable elements in neurogenesis. Vavilovskii Zhurnal Genetiki i Selektsii 24, https://doi.org/10.18699/VJ20.613 (2020).
Xie, B. et al. A two-step lineage reprogramming strategy to generate functionally competent human hepatocytes from fibroblasts. Cell Res 29, 696–710 (2019).
Article PubMed PubMed Central CAS Google Scholar
Wang, Q. et al. Generation of human hepatocytes from extended pluripotent stem cells. Cell Research 30, https://doi.org/10.1038/s41422-020-0293-x (2020).
Gertz, J. et al. Distinct Properties of Cell-Type-Specific and Shared Transcription Factor Binding Sites. Mol. Cell 52, 25–36 (2013).
Article PubMed CAS Google Scholar
Poletti, V. et al. Genome-wide definition of promoter and enhancer usage during neural induction of human embryonic stem cells. PLoS One 10, e0126590 (2015).
Article PubMed PubMed Central Google Scholar
Cheng, L. C. et al. Widespread transcript shortening through alternative polyadenylation in secretory cell differentiation. Nat. Commun. 11, 3182 (2020).
Article ADS PubMed PubMed Central CAS Google Scholar
Wang, R., Zheng, D., Yehia, G. & Tian, B. A compendium of conserved cleavage and polyadenylation events in mammalian genes. Genome Res 28, 1427–1441 (2018).
Article PubMed PubMed Central CAS Google Scholar
Dixon, G.et al. QSER1 protects DNA methylation valleys from de novo methylation. Science (1979) 372, (2021).
Ai, Z. et al. Krüppel-like factor 5 rewires NANOG regulatory network to activate human naive pluripotency specific LTR7Ys and promote naive pluripotency. Cell Rep. 40, 111240 (2022).
Article PubMed CAS Google Scholar
Li, H. et al. The Sequence Alignment/Map format and SAMtools. Bioinformatics 25, 2078–2079 (2009).
Article PubMed PubMed Central Google Scholar
Pertea, G. & Pertea, M. GFF Utilities: GffRead and GffCompare. F1000Res 9, 304 (2020).
Article Google Scholar
Langmead, B. & Salzberg, S. L. Fast gapped-read alignment with Bowtie 2. Nat. Methods 9, 357–359 (2012).
Article PubMed PubMed Central CAS Google Scholar
Pollard, K. S., Hubisz, M. J., Rosenbloom, K. R. & Siepel, A. Detection of nonneutral substitution rates on mammalian phylogenies. Genome Res 20, 110–121 (2010).
Article PubMed PubMed Central CAS Google Scholar
Nassar, L. R. et al. The UCSC genome browser database: 2023 update. Nucleic Acids Res 51, D1188–D1195-D1195 (2023).
Article PubMed CAS Google Scholar
Crooks, G. E., Hon, G., Chandonia, J. M. & Brenner, S. E. WebLogo: a sequence logo generator. Genome Res 14, 1188–1190 (2004).
Article PubMed PubMed Central CAS Google Scholar
Schneider, T. D. & Stephens, R. M. Sequence logos: a new way to display consensus sequences. Nucleic Acids Res 18, 6097–6100 (1990).
Article PubMed PubMed Central CAS Google Scholar
Trincado, J. L. et al. SUPPA2: fast, accurate, and uncertainty-aware differential splicing analysis across multiple conditions. Genome Biol. 19, 40 (2018).
Article PubMed PubMed Central Google Scholar
Quinlan, A. R. & Hall, I. M. BEDTools: a flexible suite of utilities for comparing genomic features. Bioinformatics 26, 841–842 (2010).
Article PubMed PubMed Central CAS Google Scholar
Ramírez, F. et al. deepTools2: a next generation web server for deep-sequencing data analysis. Nucleic Acids Res 44, W160–W165 (2016).
Article PubMed PubMed Central Google Scholar
Altschul, S. et al. Gapped BLAST and PSI-BLAST: a new generation of protein database search programs. Nucleic Acids Res 25, 3389–3402 (1997).
Article PubMed PubMed Central CAS Google Scholar
Mistry, J. et al. Pfam: the protein families database in 2021. Nucleic Acids Res 49, D412–D419 (2021).
Article PubMed CAS Google Scholar
Grabherr, M. G. et al. Trinity: reconstructing a full-length transcriptome without a genome from RNA-Seq data HHS public access. Nat. Biotechnol. 29, 644–652 (2011).
Haas, B. J. et al. De novo transcript sequence reconstruction from RNA-seq using the Trinity platform for reference generation and analysis. Nat. Protoc. 8, 1494–1512 (2013).
Article PubMed CAS Google Scholar
Eddy, S. R. Accelerated profile HMM searches. PLoS Comput Biol 7, e1002195 (2011).
Storer, J., Hubley, R., Rosen, J., Wheeler, T. J. & Smit, A. F. The Dfam community resource of transposable element families, sequence models, and genome annotations. Mob DNA 12, 2 (2021).
Amrani, K. El, Alanis-Lobato, G., Mah, N., Kurtz, A. & Andrade-navarro, M. A. Detection of condition-specific marker genes from RNA-seq data with MGFR. PeerJ 7, e6970 (2019).
Risso, D., Schwartz, K., Sherlock, G. & Dudoit, S. GC-Content Normalization for RNA-Seq Data. BMC Bioinforma. 12, 480 (2011).
Article CAS Google Scholar
Love, M. I., Huber, W. & Anders, S. Moderated estimation of fold change and dispersion for RNA-seq data with DESeq2. Genome Biol. 15, 550 (2014).
Article PubMed PubMed Central Google Scholar
Dobin, A. et al. STAR: Ultrafast universal RNA-seq aligner. Bioinformatics 29, 15–21 (2013).
Article PubMed CAS Google Scholar
Dobin, A. & Gingeras, T. R. Mapping RNA-seq Reads with STAR. Curr. Protoc. Bioinforma. 51, 11.14.1–19 (2015).
Article Google Scholar
Chen, S., Zhou, Y., Chen, Y. & Gu, J. fastp: an ultra-fast all-in-one FASTQ preprocessor. Bioinformatics 34, i884–i890 (2018).
Article PubMed PubMed Central Google Scholar
Li, B. & Dewey, C. N. RSEM: accurate transcript quantification from RNA-Seq data with or without a reference genome. BMC Bioinforma. 12, 323 (2011).
Article CAS Google Scholar
Harrison, P. W. et al. Ensembl 2024. Nucleic Acids Res 52, 495–502 (2024).
Robinson, J. T. et al. Integrative genomics viewer. Nat. Biotechnol. https://doi.org/10.1038/nbt.1754 (2011).
Hao, Y. et al. Integrated analysis of multimodal single-cell data. Cell 184, (2021).
Satija, R., Farrell, J. A., Gennert, D., Schier, A. F. & Regev, A. Spatial reconstruction of single-cell gene expression data. Nat. Biotechnol. 33, (2015).
Mayer, A. & Churchman, L. S. A detailed protocol for subcellular RNA sequencing (subRNA-seq). Curr. Protoc. Mol. Biol. 120, 4.29.1–4.29.18 (2017).
Heinz, S. et al. Simple combinations of lineage-determining transcription factors prime cis-regulatory elements required for macrophage and B cell identities. Mol. Cell 38, 576–589 (2010).
Article PubMed PubMed Central CAS Google Scholar
McLeay, R. C. & Bailey, T. L. Motif enrichment analysis: a unified framework and an evaluation on ChIP data. BMC Bioinforma. 11, 165 (2010).
Article Google Scholar
Bailey, T. L., Johnson, J., Grant, C. E. & Noble, W. S. The MEME suite. Nucleic Acids Res. 43, W39–W49 (2015).
Ray, D. et al. A compendium of RNA-binding motifs for decoding gene regulation. Nature 499, 172–177 (2013).
Djebali, S. et al. Landscape of transcription in human cells. Nature 489, 101–108 (2012).
Article ADS PubMed PubMed Central CAS Google Scholar
Dyer, S. C. et al. Ensembl 2025. Nucleic Acids Res 53, D948–D957 (2025).
Article PubMed Google Scholar

Download references

Acknowledgements

We acknowledge the assistance of SUSTech Core Research Facilities. Funding was from the National Natural Science Foundation of China (32270597) the Science Technology and Innovation Commission of Shenzhen (RCBS20221008093109033), and the Guangdong Basic and Applied Basic Research Foundation (2023A1515111170). A.R.’s laboratory is funded by the Russian Science Foundation (22-65-00022).

Author information

These authors contributed equally: Isaac A. Babarinde, Xiuling Fu, Gang Ma.

Authors and Affiliations

Department of Systems Biology, School of Life Sciences, Southern University of Science and Technology, Shenzhen, China
Isaac A. Babarinde, Xiuling Fu, Gang Ma, Yuhao Li, Zhangting Liang, Jianfei Xu, Zhen Xiao, Yu Qiao, Zheng Lin, Xuemeng Zhou & Andrew P. Hutchins
Laboratory of Inflammation and Vaccines, Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences, Shenzhen, China
Isaac A. Babarinde
Columbia Center for Human Development, Columbia University Irving Medical Center, New York, NY, USA
Xiuling Fu
Department of Medical Genetics, School of Basic Medical Sciences, Southern Medical University, Guangzhou, China
Gang Ma
Queensland Alliance for Agriculture and Food Innovation, The University of Queensland, St Lucia Qld, Australia
Yuhao Li
Institute of Bioengineering, Research Center of Biotechnology, Russian Academy of Sciences, Moscow, Russia
Katerina Oleynikova & Alexey Ruzov
Infochemistry Scientific Center, ITMO University, St. Petersburg, Russia
Katerina Oleynikova
Laboratories for Biomembrane Research and Biotechnology, Department of Biochemistry, University of Ibadan, Ibadan, Nigeria
Mobolaji T. Akinwole

Authors

Isaac A. Babarinde
View author publications
Search author on:PubMed Google Scholar
Xiuling Fu
View author publications
Search author on:PubMed Google Scholar
Gang Ma
View author publications
Search author on:PubMed Google Scholar
Yuhao Li
View author publications
Search author on:PubMed Google Scholar
Zhangting Liang
View author publications
Search author on:PubMed Google Scholar
Jianfei Xu
View author publications
Search author on:PubMed Google Scholar
Zhen Xiao
View author publications
Search author on:PubMed Google Scholar
Yu Qiao
View author publications
Search author on:PubMed Google Scholar
Zheng Lin
View author publications
Search author on:PubMed Google Scholar
Katerina Oleynikova
View author publications
Search author on:PubMed Google Scholar
Mobolaji T. Akinwole
View author publications
Search author on:PubMed Google Scholar
Xuemeng Zhou
View author publications
Search author on:PubMed Google Scholar
Alexey Ruzov
View author publications
Search author on:PubMed Google Scholar
Andrew P. Hutchins
View author publications
Search author on:PubMed Google Scholar

Contributions

I.A.B. planned and designed the study, performed most of the bioinformatic analyses, drafted the paper, and coordinated the experimental work. X.F. and G.M. performed the experiments, assisted by J.X., Y.Q., and Z.X., Y.L., M.T.A., Z.Liang, Z.Lin. and X.Z. assisted with analysis. K.O. and A.R. interpreted data and revised the manuscript. A.P.H. designed the study, revised the manuscript, supervised and funded the study.

Corresponding author

Correspondence to Andrew P. Hutchins.

Ethics declarations

Competing interests

The authors declare no competing interests.

Peer review

Peer review information

Nature Communications thanks Andrew Modzelewski who co-reviewed with Youjia Guo and the other anonymous reviewer(s) for their contribution to the peer review of this work. A peer review file is available.

Additional information

Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Supplementary information

Supplementary Information

Description of Additional Supplementary Files

Supplementary Data 1

Supplementary Data 2

Supplementary Data 3

Supplementary Data 4

Reporting Summary

Transparent Peer Review file

Source data

Source Data

Rights and permissions

Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/.

Reprints and permissions

About this article

Cite this article

Babarinde, I.A., Fu, X., Ma, G. et al. Transposable element expression and sub-cellular dynamics during hPSC differentiation to endoderm, mesoderm, and ectoderm lineages. Nat Commun 16, 7670 (2025). https://doi.org/10.1038/s41467-025-63080-3

Download citation

Received: 23 July 2024
Accepted: 09 August 2025
Published: 18 August 2025
Version of record: 18 August 2025
DOI: https://doi.org/10.1038/s41467-025-63080-3

This article is cited by

Dissecting the contribution of transposable elements to interphase chromosome structure
- Liyang Shi
- Zhen Xiao
- Andrew P. Hutchins
Genome Biology (2026)