The ENCODE4 long-read RNA-seq collection reveals distinct classes of transcript structure diversity

Abstract

The majority of mammalian genes encode multiple transcript isoforms that result from differential promoter use, changes in exonic splicing, and alternative 3’ end choice. Detecting and quantifying transcript isoforms across tissues, cell types, and species has been extremely challenging because transcripts are much longer than the short reads normally used for RNA-seq. By contrast, long-read RNA-seq (LR-RNA-seq) gives the complete structure of most transcripts. We sequenced 264 LR-RNA-seq PacBio libraries totaling over 1 billion circular consensus reads (CCS) for 81 unique human and mouse samples. We detect at least one full-length transcript from 87.7% of annotated human protein coding genes and a total of 200,000 full-length transcripts, 40% of which have novel exon junction chains.

To capture and compute on the three sources of transcript structure diversity, we introduce a gene and transcript annotation framework that uses triplets representing the transcript start site, exon junction chain, and transcript end site of each transcript. Using triplets in a simplex representation demonstrates how promoter selection, splice pattern, and 3’ processing are deployed across human tissues, with nearly half of multi-transcript protein coding genes showing a clear bias toward one of the three diversity mechanisms. Evaluated across samples, the predominantly expressed transcript changes for 74% of protein coding genes. In evolution, the human and mouse transcriptomes are globally similar in types of transcript structure diversity, yet among individual orthologous gene pairs, more than half (57.8%) show substantial differences in mechanism of diversification in matching tissues. This initial large-scale survey of human and mouse long-read transcriptomes provides a foundation for further analyses of alternative transcript usage, and is complemented by short-read and microRNA data on the same samples and by epigenome data elsewhere in the ENCODE4 collection.

INTRODUCTION

Most mammalian genes produce multiple distinct transcript isoforms¹. This transcript structure diversity is governed by promoter selection, splicing, and polyA site selection, which respectively dictate the transcript start site (TSS), exon junction chain (the unique series of exon-exon junctions used in a transcript), and transcript end site (TES, and the resulting 3’ UTR) used in the final transcript. Each of these processes is highly regulated and is subject to a different set of evolutionary pressures^2–5. In protein coding genes, missplicing can lead to nonfunctional transcripts by disrupting canonical reading frames or introducing premature stop codons that predispose the transcript to nonsense mediated decay (NMD). Conversely, the cellular machinery involved in promoter or polyA site selection for protein coding genes is only constrained by the need to include start and stop codons for the correct open reading frame (ORF) in the final mRNA product.

Transcript structure diversity poses challenges for both basic and preclinical biology. As computational gene prediction and manual curation efforts have identified ever more transcripts for many genes^6,7, a common assumption in genomics and medical genetics is that we only need to consider one or at most a handful of representative transcripts per gene such as those from the MANE (Matched Annotation from NCBI and EMBL-EBI) project⁸. MANE transcripts are chosen with respect to their expression levels in biologically-relevant samples and sequence conservation of the coding regions, and are perfectly matched between NCBI and ENSEMBL with explicit attention to the 5’ and 3’ ends. This decision to focus on one transcript per gene was driven in part by the difficulties in transcript assembly using ESTs and short-read RNA-seq, which is the assay used for most bulk and single-cell RNA-seq experiments^9,10. The advent of long-read platforms heralded the promise of full-length transcript sequencing to identify expressed transcript isoforms, thus potentially bypassing the error-prone transcript assembly step^11,12. However, as long-read RNA-seq (LR-RNA-seq) produces more novel candidate transcripts, there is a need to find organizational principles that will allow us to cope with the diversity of transcripts observed at some gene loci in catalogs such as GENCODE⁷, while at the same time distinguishing the genes that do not seem to undergo any alternative splicing.

Short-read RNA-seq has been the core assay for measuring gene expression in the second and third phases of the ENCODE project for all RNA biotypes, regardless of their lengths, in both human and mouse samples^13–15. Short-read RNA-seq has also been used by many groups to comprehensively characterize TSS usage¹⁶, splicing¹⁷, and TES usage¹⁸, but the challenges of transcript assembly given the combinatorial nature of the problem have precluded a definitive assessment of the transcripts present. In addition to continuing Illumina-based short-read sequencing of mRNA and microRNA, the fourth phase of ENCODE (ENCODE4) adds matching LR-RNA-seq using the Pacific Biosciences Sequel 1 and 2 platforms in a set of human and mouse primary tissues and cell lines in order to identify and quantify known and novel transcript isoforms expressed across a diverse set of samples. We report the resulting ENCODE4 human and mouse transcriptome datasets. We implement a novel triplet scheme that captures essential differences in 5’ end choice, splicing, and 3’ usage, which allows us to categorize genes based on features driving their transcript structure diversity using a new software package called Cerberus. We introduce the gene structure simplex as an intuitive coordinate system for comparing transcript usage between genes and across samples. We then compare transcript usage between orthologous genes in human and mouse and identify substantial differences in transcript diversity for over half the genes.

Results

The ENCODE4 RNA dataset.

This LR-RNA-seq study profiled 81 tissues or cell lines by using the PacBio sequencing platform on 264 human and mouse libraries that include replicate samples and multiple human tissue donors (Tables S1–2). Without consideration for the seven postnatal timepoints in mouse, they represent 49 unique tissues or cell types across human and mouse (Fig. 1a, Fig. S1). In addition, we sequenced matching human short-read RNA-seq (Fig. S1c) and microRNA-seq (Fig. S2, Supplementary results) for most samples as well as for an additional 37 that were sequenced with short-read RNA-seq only. We detect the vast majority of polyA genes (those with biotype protein coding, pseudogene, or lncRNA) whether we restrict the analysis to short-read samples that have matching data in the LR-RNA-seq dataset (93.9% of GENCODE v40 polyA genes and 90.6% of protein coding genes) or if we use all of the short-read samples (Fig. 1b, Fig. S3a). 31.1% of all expressed genes are detected in most (>90%) of the samples, and 34.0% are detected more specifically (<10% of samples) (Fig. 1c, Fig. S3b).

Figure 1. Overview of the ENCODE4 RNA datasets.

a, Overview of the sampled tissues and number of libraries from each tissue in the ENCODE human LR-RNA-seq dataset. b, Percentage of GENCODE v40 polyA genes by gene biotype detected in at least one ENCODE short-read RNA-seq library from samples that match the LR-RNA-seq at > 0 TPM, >=1 TPM, and >= 100 TPM. c, Number of samples in which each GENCODE v40 gene is detected >= 1 TPM in the ENCODE short-read RNA-seq dataset from samples that match the LR-RNA-seq. d, Data processing pipeline for the LR-RNA-seq data. e, Percentage of GENCODE v40 polyA genes by gene biotype detected in at least one ENCODE human LR-RNA-seq library at > 0 TPM, >= 1 TPM, and >= 100 TPM. f, Number of samples in which each GENCODE v40 gene is detected >= 1 TPM in the ENCODE human LR-RNA-seq dataset. g, Boxplot of TPM of polyA genes at the indicated rank in each human LR-RNA-seq library. Not significant (no stars) P > 0.05; *P <= 0.05, **P <= 0.01, ***P <= 0.001, ****P <= 0.0001; Wilcoxon rank-sum test.

For each LR-RNA-seq dataset, we first mapped the reads using Minimap2¹⁹ and corrected non-canonical splice junctions and small indels using TranscriptClean²⁰, after which we ran TALON²¹ and LAPA²² to identify each transcript by its exon junction chain and assign each transcript a supported 5’ and 3’ end. Finally, to catalog transcript features and summarize transcript structure diversity in our datasets, we ran Cerberus, which is described below. It is important to emphasize that this pipeline (Fig. 1d) does not attempt to assemble reads, so that every reported known transcript is observed from 5’ to 3’ end in at least one read. We further required support from multiple reads for defining valid ends. Overall this is a conservative pipeline that was designed to detect and quantify robust novel and known transcripts (Materials and Methods).

Our LR-RNA-seq reads are oligo-dT primed and we therefore expect to see high detection of transcripts from polyA genes, which we define as belonging to the protein coding, lncRNA, or pseudogene GENCODE-annotated biotypes, across our datasets. Consistent with this expectation, we detect 75.9% of annotated GENCODE v40 polyA genes and 93.7% of protein coding genes at >= 1 TPM in at least one library in our human dataset (Fig. 1e). The overwhelming majority of undetected polyA genes are pseudogenes and lncRNAs, which are likely to be either lowly expressed or completely unexpressed in the tissues assayed. As expected, GO analysis of the undetected protein coding genes yielded biological processes such as smell and taste-related sensory processes that represent genes specifically expressed in tissues that we did not assay (Fig. S3c). We find that many genes are either expressed in a sample-specific manner (27.8% in <10% of samples) or are ubiquitously expressed across many samples (28.2% in >90% of samples), consistent with the short-read samples (Fig. 1f).

Transcriptionally active regions that are absent from GENCODE are candidates for novel genes. Applying conservative thresholds that included a requirement for one or more reproducible splice junctions (Supplementary methods), we found 214 novel candidate genes with at least one spliced transcript isoform expressed >= 1 TPM in human and 96 in mouse at our existing sequencing depth. Applying the same criteria to annotated polyA genes, we find 20,716 and 18,971 genes for human and mouse respectively, meaning that plausible novel genes constitute less than 1.0% in human and 0.5% in mouse. We subsequently focus analysis on transcripts from known polyA genes.

We then examined the distribution of gene expression values across our human LR-RNA-seq dataset to characterize the abundance of genes and to assess whether we would be able to measure differences in transcript abundance at our current sequencing depths. For each library, we ranked each gene by TPM and found that the most highly expressed genes have higher TPMs in primary tissue-derived libraries than cell line-derived libraries (Fig. 1g, Supplementary methods). In particular, the tissue-derived liver libraries have the most highly expressed genes at ranks 1, 5, and 10, which include ALB and FTL, as expected. We also observed that the top 1,000 genes expressed in all but one liver library are expressed >= 100 TPM and that the top 5,000 are expressed >= 10 TPM. We can therefore confidently measure major transcript expression usage with a conservative threshold of 10 TPM for at least a third of expressed genes in each sample.

From genes expressed >= 10 TPM, we are able to capture over half (54.0%) of MANE transcripts that are 9–12 kb long (Fig. S3d). Coupled with our read length profiles, we estimate that we can reliably sequence the 99.7% of annotated GENCODE v40 polyA transcripts that are <12 kb long from end-to-end if they are highly expressed (Fig. S3d–g). In mouse, we observe similar read length profiles, sample separation, and gene detection patterns (Fig. S3h–j), including detection of 84.9% of annotated GENCODE vM25 protein coding genes at >= 1 TPM (Fig. S3h). In summary, we are able to detect most of human and mouse protein coding genes in our ENCODE LR-RNA-seq datasets at similar rates to short-read RNA-seq, and our long reads are long enough to capture the vast majority of annotated polyA transcripts.

Different sources of transcript structure diversity.

We compared the transcript start sites (TSSs), exon junction chains (ECs), and transcript end sites (TESs) observed in the human LR-RNA-seq data with other prior assays of these features and with established catalogs of these features. Cerberus is designed to identify unique TSSs, ECs, and TESs from a wide variety of inputs that include LR-RNA-seq data, reference atlases, and external transcriptional assays such as CAGE, PAS-seq, and the GTEx LR-RNA-seq dataset^18,23 (Supplementary methods). Cerberus numbers each TSS, EC, and TES (triplet features) based on the annotation status of the transcript that it came from (e.g. the most confidently annotated will be numbered first) as well as the order in which each source was provided (Supplementary methods). Cerberus outputs genomic regions for each unique TSS and TES and a list of coordinates for each unique EC. In all cases, the gene of origin is also annotated (Fig. S4a). Using the integrated series of cataloged triplet features, Cerberus assigns a TSS, EC, and TES to each unique transcript model to create a transcript identifier of the form Gene[X,Y,Z], which we call the transcript triplet (Fig. S4b, Fig. 2a). This strategy distinguishes the structure of two different transcripts from the same gene solely on the basis of their transcript triplets. Additionally, it gives us the ability to sum up the expression of TSSs, ECs, and TESs across the transcripts they come from to enable quantification of promoter usage, EC usage, and polyA site usage respectively.

Figure 2. Triplet annotation of transcript structure maps diversity within and across samples.

a, Representation of structure and transcript triplet naming convention for 3 different transcripts from the same gene based on the transcript start site (TSS), exon junction chain (EC), and transcript end site (TES) used. b-d, Triplet features detected >= 1 TPM in human ENCODE LR-RNA-seq from GENCODE v40 polyA genes broken out by novelty and support. Known features are annotated in GENCODE v29 or v40. Novel supported features are supported by b, CAGE or RAMPAGE c, GTEx, d, PAS-seq or the PolyA Atlas. e-g, Triplet features detected >= 1 TPM in human ENCODE LR-RNA-seq per GENCODE v40 polyA gene split by gene biotype for e, TSSs, f, ECs, g, TESs. h, Number of transcripts from GENCODE v40 polyA genes detected >= 1 TPM from human ENCODE LR-RNA-seq that have a known EC split by gene biotype. i, Novelty characterization of triplet features in each transcript detected >= 1 TPM in the human ENCODE LR-RNA-seq. j, Number of transcripts detected >= 1 TPM in human ENCODE LR-RNA-seq per GENCODE v40 polyA gene split by gene biotype. k, COL1A1 (gene expressed at 548 TPM) transcripts expressed >= 1 TPM in the ovary sample from human ENCODE LR-RNA-seq. l, PKM (gene expressed at 506 TPM) transcripts expressed >= 1 TPM in the ovary sample from human ENCODE LR-RNA-seq colored by expression level (TPM). m, Expression level of gene (TPM) versus the percent isoform (pi) value of the predominant transcript for each gene expressed >= 1 TPM from human ENCODE LR-RNA-seq in the ovary sample. Points are colored by whether or not pi = 100. n, Number of unique predominant transcripts detected >= 1 TPM across samples per gene.

We applied Cerberus to the ENCODE human LR-RNA-seq data and to annotations from GENCODE v40 and v29 to obtain transcript triplets for the transcripts present in each transcriptome. Cerberus labels each triplet feature as known if it is detected in a reference set (here defined as transcripts derived from GENCODE) of transcripts or novel if not. Additionally, using the information from the GENCODE reference transcriptomes, we assign the triplet [1,1,1] to the MANE transcript isoform for the gene, if it has one.

Altogether, we detected 206,806 transcripts expressed >= 1 TPM from polyA genes, 76,469 of which have exon junction chains unannotated in GENCODE v29 or v40. From these transcripts, we first sought to characterize the observed triplet features (expressed >= 1 TPM in at least one library from polyA genes) in our dataset (Fig. 2b–d, Fig. S5a–f). We found that 18.0% of TSSs, 37.3% of ECs, and 22.1% of TESs are novel compared to both GENCODE v29 and v40 (Fig. 2b–d). We furthermore determined whether any novel triplet features were supported by sources outside of the GENCODE reference. We used CAGE and RAMPAGE data to support TSSs, GTEx transcripts to support ECs, and PAS-seq and the PolyA Atlas regions to support the TESs (Supplementary methods). Of the novel triplet features, 42.8% of TSSs, 17.9% of ECs, and 79.0% of TESs were supported by at least one external dataset (Fig. 2b–d, Fig. S5a–c). While the intermediate transcriptome (in general transfer format; GTF) for our LR-RNA-seq dataset from LAPA has single-base transcript ends, the majority of our Cerberus TSSs and TESs derived from the LR-RNA-seq data are 101 bp in length and 99.8% are shorter than 500 bp, which is consistent with how Cerberus extends TSSs and TESs derived from GTFs by n bp (here n=50) on either side (Fig. S4a, Fig. S5d–e, Supplementary methods).

We further annotated the novelty of ECs compared to GENCODE using the nomenclature from SQANTI²⁴. For detected ECs from polyA genes, we find that the majority (62.7%) of ECs were already annotated in either GENCODE v29 or v40. Novel ECs are primarily annotated as either NIC (16.1%; novel in catalog, defined as having a novel combination of known splice sites) or NNC (11.6%; novel not in catalog, defined as having at least one novel splice site) (Fig. S5f). Given the high external support for our triplet features, we were also able to predict CAGE and RAMPAGE support for our long-read derived TSSs both in and across cell types using logistic regression (Fig. S6, Supplementary results). In aggregate, a majority of our triplet features observed in our LR-RNA-seq data show were in prior annotations or have external support from additional assays.

We then examined the number of observed triplet features per gene. We find that most protein coding genes (89.8%) express more than one triplet feature across our dataset (Fig. 2e–g). By contrast, only 33.7% of lncRNAs and 14.4% of pseudogenes express more than 1 transcript and therefore triplet feature per gene. These biotypes exhibit far less transcript structure diversity as compared to protein coding genes. Overall, we find that our observed triplet features are individually well-supported by external annotations and assays. We also show that protein coding genes are far more likely to have more than one triplet feature than lncRNAs and pseudogenes.

The ENCODE4 LR-RNA-seq transcriptome.

Following our characterization of individual triplet features, we moved on to examining our full-length transcripts. We first note that most observed transcripts with known ECs belong to the protein coding biotype (Fig. 2h). In contrast to the gene-level analysis, transcripts with known ECs are expressed in a more sample-specific manner, with 49.0% expressed in <10% of samples and 4.4% expressed in >90% of samples (Fig. S7a, Fig. 1f). Of the remaining protein coding transcripts with novel ECs, 53.0% are predicted to have complete ORFs which are not subject to nonsense mediated decay (Supplementary methods). Examining detected transcripts in our dataset based on the novelty of each constituent triplet feature, we find that 52.0% of observed transcripts have each of their triplet features annotated, and that more transcripts contain novel TSSs than novel TESs (Fig. 2i). Consistent with our observation that protein coding genes generally have more than one triplet feature per gene compared to the other polyA biotypes, we find that most protein coding genes also have more than one transcript per gene (Fig. 2j, e–g). We investigated the extent that lower expression levels of lncRNAs contribute to their overall lower transcript diversity compared to protein coding genes. We found that the 137 lncRNAs expressed >100 TPM in one or more samples have the same median number of expressed transcripts per gene as the 8,436 protein coding genes at the same expression level (median = 7) (Fig. S7b, Supplementary methods). Therefore the lower reported overall diversity of lncRNAs is due to a combination of their lower expression levels and our sequencing depth.

We compared the number of TSSs and TESs that are detected per EC across the observed transcripts from GENCODE v40 versus in our observed transcripts. We found that in GENCODE v40, each multiexonic EC has a maximum of 3 TSSs or TESs across the polyA transcripts and that the overwhelming majority of ECs are only annotated with 1 TSS and TES (99.7% and 99.4% respectively) (Fig. S7c–d). In contrast, our strategy of transcriptome annotation yields a substantial increase in the number of distinct TSSs and TESs observed per EC, which more accurately reflects the biology of the coordination of promoter choice, polyA site selection, and splicing across the diverse samples in our dataset (Fig. S7e–f). The effect of this increase in annotated TSSs and TESs is also apparent when analyzing our transcripts using traditional alternative splicing event detection methods, which are not written to consider the more subtle differences in transcript structure at the 5’ and 3’ end (Fig. S8, Supplementary results).

Predominant transcript structure differs across tissues and cell types.

Different multiexonic genes with similar expression levels within the same sample can exhibit vastly different levels of transcript structure diversity. For instance, the genes COL1A1 and PKM have a high number of exons (60 and 47 exons, respectively across our entire human dataset) and are highly expressed in ovary (548 and 506 TPM respectively). Yet, we detect only one 6.9 kb long transcript for COL1A1 (Fig. 2k) whereas we detect 18 transcript isoforms that vary on the basis of their TSSs, ECs, and TESs for PKM (Fig. 2l).

We then asked what fraction of overall gene expression is accounted for by the predominant transcript, which is the most highly expressed transcript for a gene in a given sample. Comparing the TPM of genes expressed in ovary to the the percentage of reads from a gene that come from that transcript (pi - percent isoform)²⁵ of the predominant transcript, we find that 19.5% of protein coding genes expressed >100 TPM have a predominant transcript that accounts for less than 50% of the reads, and therefore are highly expressed with high transcript structure diversity. Conversely, 26.8% of protein coding genes are expressed >100 TPM and have a predominant transcript that accounts for more than 90% of the expression of the gene (Fig. 2m). Globally, we generated a catalog of predominant transcripts for each sample. The median number of predominant transcripts per protein coding gene across samples was 2, and that 73.0% of protein coding genes have more than one predominant transcript across the samples surveyed (Fig. 2n). Thus, the majority of human protein coding genes use a different predominant transcript in at least one condition represented in our sample collection.

Quantifying transcript structure diversity across samples using gene triplets and the gene structure simplex.

We developed a framework to systematically characterize and quantify the diversity between the detected transcripts from each gene by computing a summary gene triplet, which is related to but distinct from transcript triplets. For each set of transcripts from a given gene, we count the number of unique TSSs, ECs, and TESs (Fig. 3a, Fig. S9). As the number of exon junction chains is naturally linked to the number of alternative TSSs or TESs (for instance, a new TSS with a different splice donor will lead to a novel EC regardless of similarities in downstream splicing), we calculate the splicing ratio as $\frac{2\times N_{EC}}{N_{TSS}+N_{TES}}$ to more fairly assess the contribution of ECs to transcript diversity in each gene (Fig. 3a, Fig. S9). We then compute the proportion of transcript diversity that arises from each source of variation: alternative TSS usage, alternative TES usage, or internal splicing (Fig. 3a). Representing these numbers as proportions allows us to plot them as coordinates in a two-dimensional gene structure simplex (Fig. 3b, Fig. S9). This enables us to visualize how transcripts from a gene typically differ from one another and categorize genes based on their primary driver of transcript structure diversity. Genes with a high proportion of transcripts characterized by alternative TSS usage (>0.5) will fall into the TSS-high sector of the simplex, those with a high proportion of transcripts characterized by alternative TES usage (>0.5) will fall into the TES-high sector of the simplex, and those with a high proportion of transcripts characterized by internal splicing (>0.5) will fall into the splicing-high portion of the simplex. Genes with more than one transcript that do not display a strong preference for one mode over the other lie in the mixed sector, and genes with just one transcript are in the center of the simplex, henceforth the simple sector (Fig. 3a–b, Fig. S9, Supplementary methods).

Figure 3. The gene structure simplex represents distinct modes of transcript structure diversity across genes and samples.

a, Transcripts for 5 model genes; 1 of each sector (TSS-high, splicing-high, TES-high, mixed, and simple). Table shows the gene triplet, splicing ratio gene triplet, and simplex coordinates that correspond to each toy gene. b, Layout of the gene structure simplex with the genes from a, plotted based on their simplex coordinates. Proportion of TSS usage is the blue axis (left), proportion of TES is the orange axis (bottom), and proportion of splicing ratio is the pink axis (right). Regions of the simplex are colored and labeled based on their sector category (TSS-high, splicing-high, TES-high). Gene triplets that land in each sector are assigned the concordant sector category. c-e, Gene structure simplices for the transcripts from protein coding genes that are c, annotated in GENCODE v40 where the parent gene is also detected in our human LR-RNA-seq dataset, d, the observed set of transcripts, those detected >= 1 TPM in the human ENCODE LR-RNA-seq dataset, e, the observed major set of transcripts, the union of major transcripts from each sample detected >= 1 TPM in the human ENCODE LR-RNA-seq dataset. f-j, Proportion of genes from the GENCODE v40, observed, and observed major sets that fall into the f, TSS-high sector, g, splicing-high sector, h, TES-high sector, i, mixed sector, j, simple sector. k, Gene structure simplex for AKAP8L. Gene triplets with splicing ratio for H9 and H9-derived pancreatic progenitors labeled. Simplex coordinates for the GENCODE v40, observed set, and centroid of the samples also shown for AKAP8L. l-m, Transcripts of AKAP8L expressed >= 1 TPM in l, H9 m, H9-derived pancreatic progenitors colored by expression level in TPM. Alternative exons that differ between transcripts are colored pink.

We first used the gene structure simplex to compare different transcriptomes. We computed gene triplets for protein coding genes for the following transcriptomes: GENCODE v40 transcripts from genes we detect in our LR-RNA-seq dataset; observed transcripts in our LR-RNA-seq dataset (observed); and the union of detected major transcripts (observed major), which we define as the set of most highly expressed transcripts per gene in a sample that are cumulatively responsible for over 90% of that gene’s expression in any of our LR-RNA-seq samples (Fig. 3c–j, Supplementary methods). The observed and observed major gene triplets describe the diversity of transcription in each gene across all samples in the dataset. Unsurprisingly, GENCODE genes show less density in the TSS or TES sectors of the simplex, largely because the main focus of GENCODE is to annotate unique ECs rather than 5’ or 3’ ends. This causes a concomitant drop in diversity in these sectors in GENCODE compared to the observed and observed major transcripts (Fig. 3f, h). Interestingly, there is also a distinct enrichment of genes that occupy the splicing-high portion of the simplex in our observed set compared to GENCODE (Fig. 3g). When considering the observed major transcripts, we see an increase in the percentage of genes in the TSS and splicing-high sectors over the set of all transcripts detected in our entire LR-RNA-seq dataset, but a decrease in the TES-high sector (Fig. 3f–h). Overall, we compared gene triplets for transcripts as annotated by GENCODE and observed in our LR-RNA-seq dataset and found higher proportions of genes with high TSS and splicing diversity as compared to GENCODE.

Calculating sample-level gene triplets identifies genes that show distinct transcript structure diversity across samples.

The observed gene triplets represent the aggregate repertoire of triplet features for each gene globally across our entire LR-RNA-seq dataset. However, the overall transcript structure diversity of a gene does not necessarily reflect the transcript structure diversity of a gene within a given sample. Therefore, we computed gene triplets for each sample in our dataset using all detected transcripts in each sample (sample-level gene triplets) or just the major transcripts in each sample (sample-level major gene triplets). These gene triplets can also be visualized on the gene structure simplex where each point represents the gene triplet associated with a different sample (Fig. 3k).

In order to find genes that display heterogeneous transcript structure diversity across unique biological contexts, we computed the average coordinate (centroid) for each gene from the sample-level gene triplets and calculated the distance between it and each sample-level gene triplet (Supplementary methods). 2,892 unique genes had a distance z-score >3 in at least one sample and therefore demonstrate dissimilar transcript structure diversity from the average. One such example is AKAP8L in the H9-derived pancreatic progenitors (z-score: 5.23). AKAP8L can bind both DNA and RNA in the nucleus and has been shown to have functional differences on the protein level resulting from alternate transcript choice^26,27. In our data, transcripts of this gene generally differ in terms of the EC or TES choice, but this behavior differs from sample to sample (Fig. 3k). For example, transcripts of AKAP8L differ only in their ECs in H9 embryonic stem cells, whereas transcripts differ in their ECs and TESs in the H9-derived pancreatic progenitors (Fig. 3l–m).

We also compared our sample-level gene triplets to the observed gene triplets to understand how transcript structure diversity differs globally versus within samples (Fig. 3d, Fig. 2j, Fig. S10a–h). First, we simply counted the number of triplet features or transcripts per gene and found that while most genes have more than one triplet feature or transcript globally (Fig. 2e–g, j), on the sample level, most genes have far fewer triplet features and transcripts; with a particularly pronounced difference for the TSS (Fig. S10a–d). We found that the distributions of triplet features overall and in each sample were significantly different from one another (two-sided KS test) (Fig. S10e–l, Supplementary methods).

To determine how transcript structure diversity for each gene changes from the global to sample level using the gene structure framework, we computed distances between the global observed gene triplets for non-simple genes to sample-level gene triplet centroids (Supplementary methods). We find that 3.2% of tested genes have a distance z-score >2 between their observed and sample-level centroid gene triplets. In support of our analysis on the individual triplet feature level, we find that 94.8% of genes from the TSS-high sector in the observed set do not share this sector with their sample-level centroid, indicating that genes with a large number of promoters typically use them in a sample-specific manner. ACTA1, a gene that encodes for an actin protein²⁸, is the gene with the highest distance between observed and sample-level centroid. Its observed gene triplet is (1,18,1) and therefore splicing-high. However, in most samples where ACTA1 is expressed, it has only one transcript isoform (Fig. S10m). This drives the sample-level centroid behavior into the mixed sector (Fig. S10n). In contrast, in heart and muscle ACTA1 expresses 18 and 15 transcripts respectively, which all differ on the basis of their ECs (Fig. S10m–o). This illustrates how the gene structure framework can be used to highlight differences between sample-specific and global transcript structure diversity, and also shows that individual genes are substantially different.

Sample-specific and global changes in predominant and major transcript isoform usage.

Nevertheless, the transcript structure diversity pattern for the majority of genes is consistent across samples where they are expressed at substantial levels. Elastin (ELN), which is an important component of the extracellular matrix²⁹, is the gene with the greatest number of detected transcripts in our dataset (283 in total). We find that in most samples, distinct transcripts of ELN are characterized by different ECs (Fig. 4a). For example, in lung, ELN has 32 major transcripts with 21 different ECs, but in 31 of its major transcripts, uses only one TSS and two TESs (Fig. 4b). By contrast, the four transcripts from the transcription factor CTCF expressed in lung use three TSSs but only one TES (Fig. 4c–d).

Figure 4. Sample-specific and global changes in predominant and major transcript isoform usage.

a, Gene structure simplex for major transcripts of ELN. Gene triplets with splicing ratio for lung and H9-derived chondrocytes labeled. Simplex coordinates for the GENCODE v40 and observed major set are labeled. b, Major transcripts of ELN expressed >= 1 TPM in lung colored by expression level in TPM. Alternative exons that differ between transcripts are colored pink. c, Gene structure simplex for major transcripts of CTCF. Gene triplets with splicing ratio for lung labeled. Simplex coordinates for the GENCODE v40 and observed major set are labeled. d, From top to bottom: Major transcripts of CTCF expressed >= 1 TPM in lung, TSSs of CTCF major transcripts expressed >= 1 TPM in lung, ENCODE cCREs colored by type. e, Gene structure simplex for E4F1. Gene triplet with splicing ratio for observed E4F1 transcripts labeled. Simplex coordinates for the GENCODE v40 and observed set also shown for E4F1. f, Gene structure simplex for major transcripts of E4F1. Gene triplet with splicing ratio for observed major E4F1 transcripts labeled. Simplex coordinates for the GENCODE v40 and observed major set also shown for E4F1. g, Sector assignment change and conservation for protein coding genes in the human ENCODE LR-RNA-seq dataset between the observed set of gene triplets (left) and the observed major set of gene triplets (right). Percent of genes with the same sector between both sets labeled in the middle. h-k, Percentage of libraries where a gene with an annotated MANE transcript is expressed and the MANE h, transcript i, TSS j, EC k, TES is the predominant transcript or triplet feature.

While the observed gene triplets for a gene represent the overall transcript structure diversity, the observed major gene triplets capture diversity of the most highly expressed transcripts in each sample. We computed the distances between the observed and observed major simplex coordinates for protein coding genes. The transcription factor E4F1³⁰ has a high distance between the observed and observed major gene triplets, which corresponds to a change from the mixed to splicing-high sector (Fig. 4e–f). This sector change is driven by the use of fewer TSSs and TESs in major transcripts. Overall, 83.7% of protein coding genes retain their sectors between our observed and observed major triplets, while 4.8% genes in the mixed sector move to one of the three corners of the simplex (TSS, splicing, or TES-high) (Fig. 4g). Thus, the differences between the observed and observed major gene triplets in a subset of genes can be substantial.

One criterion for the identification of MANE transcripts is how highly expressed the transcript is compared to others⁸. Therefore, we assessed how frequently the MANE transcript was the predominant one in each of our LR-RNA-seq libraries. Limiting ourselves to only the genes that have annotated MANE transcripts in GENCODE v40, we found that 64.1% of genes have a non-MANE predominant transcript in at least 80% of the libraries where the gene is expressed (Fig. 4h). At the individual triplet feature level, 30.8% of TSSs, 40.9% of ECs, and 45.2% of TESs have a non-MANE predominant feature in at least 80% of libraries (Fig. 4i–k). Therefore, though the MANE transcript typically is the most highly expressed transcript in a library, most genes with MANE transcripts have some libraries where this is not the case. For non-MANE predominant transcripts, only 17.0% were predicted to have the same ORF as the MANE transcript. Furthermore, 62.1% of non-MANE predominant transcripts are predicted to encode for a full ORF that does not undergo NMD. These results indicate that in many cases, the alternative predominant transcript in a sample likely encodes for a distinct, functional protein. The genes where the MANE transcript and triplet features are frequently not the predominant one represent loci that would suffer more from restricting analyses to only a single transcript isoform.

For a subset of gene / library combinations where the MANE transcript or feature was not the predominant one, the MANE transcript or feature was still expressed, albeit at a lower level. For these gene / library combinations, we compared the expression of the predominant transcript to the MANE one (Fig. S11a–d). We found that for predominant transcripts or triplet features expressed <30 TPM, the MANE counterpart was expressed at a comparable level. By contrast, for the opposite situation, where the MANE transcript or triplet feature was the predominant one, we found that the secondary transcript was not expressed at a similar level (Fig. S11e–h). Overall, for most gene / library combinations, the MANE transcript or triplet feature is the predominant one (Fig. S11i–l).

Comparing transcript structure diversity between species.

We ran Cerberus on the ENCODE4 mouse LR-RNA-seq dataset to calculate transcript and gene triplets to enable comparison of transcript structure diversity between the two species. Compared to human GENCODE v40, GENCODE vM25 genes are less enriched in the TSS, splicing, and TES-high sectors (Fig. 3f–j, Fig. 5a–e, Fig. S12a–c). For the mouse observed and observed major gene triplets, we see relatively similar percentages of genes in each sector (Fig. 5a–e, Fig. S12a–c). We found fewer predominant transcripts across samples per protein coding gene in mouse than in human, which is expected due to the overall lower number of tissues in our mouse data, with a median of 2 predominant transcripts per gene and 57.7% of protein coding genes with more than one (Fig. S12d). Furthermore, we observe that 54.5% of protein coding transcripts with novel ECs are predicted to encode for full ORFs without NMD. Thus, the two transcriptomes have similar distributions of genes in our gene structure simplex.

Figure 5. Conservation of gene triplets from human and mouse.

a-e, Proportion of genes from the GENCODE vM25, observed, and observed major sets that fall into the a, TSS-high sector, b, splicing-high sector, c, TES-high sector, d, mixed sector, e, simple sector. f, Gene structure simplex for ARF4 in human. Gene triplet with splicing ratio for ARF4 transcripts in H1 labeled. Simplex coordinates for the GENCODE v40, sample-level centroid, and observed set also shown for ARF4. g, Gene structure simplex for Arf4 in mouse. Gene triplet with splicing ratio for Arf4 transcripts in F121–9 labeled. Simplex coordinates for the GENCODE v40, sample-level centroid, and observed set also shown for Arf4. h, Transcripts of ARF4 expressed >= 1 TPM in human H1 sample colored by expression level in TPM. i, Transcripts of Arf4 expressed >= 1 TPM in mouse F121–9 sample colored by expression level in TPM. j, Sector assignment change and conservation for orthologous protein coding genes between the observed major human set of gene triplets (left) and the observed major mouse set of gene triplets (right). Percent of genes with the same sector between both sets labeled in the middle. k, Sector assignment change and conservation for orthologous protein coding genes between the sample-level H1 major human set of gene triplets (left) and the sample-level F121–9 major mouse set of gene triplets (right). Percent of genes with the same sector between both sets labeled in the middle.

In order to make gene-level comparison for orthologous genes in both species, we subset the human samples on those that are the most similar to the mouse samples and computed “mouse matched” observed, observed major, and sample-level gene triplets (Supplementary methods, Table S1–2). We computed the sample-level centroids for each gene in both species and computed the distance between each pair of 1:1 orthologs. Of the 13,536 orthologous genes, 4.3% have a distance z-score >2 between the species and therefore exhibit substantial changes in transcript structure diversity between the species. One of these is ADP-Ribosylation Factor 4 (ARF4), which is the most divergent member of the ARF4 family³¹. Human ARF4 sample-level gene triplets are nearly always splicing-high whereas mouse Arf4 sample-level gene triplets are mainly TES-high (Fig. 5f–g). We examined the ARF4 / Arf4 transcripts expressed in matching embryonic stem cell samples (H1 in human and F121–9 in mouse) and found that, despite the homologous samples, all 3 of the expressed human transcripts use the same TSS and TES but differ in the ECs whereas all 3 expressed mouse transcripts use the same TSS and EC but differ at the TES (Fig. 5h–i). We find globally that when comparing the observed major gene triplets between human and mouse, only 42.2% of genes have the same sector in human and mouse (Fig. 5j). This result holds even when restricting ourselves to comparing human tissues with adult mouse samples or just a comparison between human and mouse embryonic stem cells (Fig. 5k). Thus, we find substantial differences in splicing diversity for orthologous genes between human and mouse.

Discussion

The ENCODE4 LR-RNA-seq dataset is the first large-scale, cross-species survey of transcript structure diversity using full-length cDNA sequencing on long-read platforms. We identify and quantify known and novel transcripts in a broad and diverse set of samples with uniformly processed data and annotations available at the ENCODE portal. A new framework was introduced for categorizing transcript structure diversity based on their exon junction chains and ends using gene and transcript triplets, which allowed us to use the gene structure simplex to visualize and compare gene triplets between samples and across species. The results showed a full range of transcript structure diversity across the transcriptome, based on promoter, internal splicing, and polyA site choice. As expected, the existing gene annotation catalogs such as GENCODE have successfully captured individual features such as TSSs and exons. However, GENCODE annotated full-length transcripts only represent a subset of the TSS, EC, and TES combinations that we observe using our conservative pipeline that requires full end-to-end support in a single read and support from multiple reads for defining ends. From the human LR-RNA-seq quantification, we found more than one predominant transcript across samples for 73.0% of genes, which is in contrast with prior reports^32,33. We also found that for a substantial number of genes, transcript structure diversity and major transcript usage for the same gene differs between tissues, samples, and developmental timepoints. The majority of genes had at least one library where the MANE transcript is not the predominant transcript. This could confound analyses such as variant effect prediction in which it is common practice to consider only one transcript per gene. Finally, we found that transcript structure diversity behavior differs quite strikingly between human and mouse on a gene-by-gene basis. In matching samples, the dominant source of transcript structure diversity differed for more than half of orthologous protein coding genes.

Our data and framework provide a foundation for further analyses such as the functional impacts of alternative 5’ and 3’ ends, RNA modifications, RNA binding protein function, allele-specific expression, and transcript half-life. Together with the accompanying tissue and cell type annotations, this constitutes a transcript-level reference atlas that is structured appropriately for integration of future single-cell long-read analysis. Exploration using the gene structure simplex analysis will yield additional genes showing sample specific variance compared to their average behavior when extended to new tissues, differentiation time courses, or disease samples. The triplet annotation scheme for transcripts, based on mechanistically distinct transcript features, organizes and simplifies high-level analysis of transcripts from the same gene. We find it to be a useful and commonsense improvement over arbitrary transcript IDs and we expect it to be widely applicable to transcriptomes of any organism that uses regulated alternative splicing, promoter choice, or 3’ end selection.

Our annotations are consistent and extensive, yet they have several limitations. With the current sample preparation protocol and depth of sequencing, we reliably detect transcripts that are expressed above a minimum expression level of 1 TPM and are less than 10 kb long. While 99.3% of GENCODE v40 polyA transcripts are less than 10 kb long, we are undoubtedly underrepresenting transcripts that are at the long end of the distribution, especially when they are expressed at low levels or in rare cell types within a tissue. RNA integrity differences between the human cell lines and mouse tissues, both of which produce very high quality RNA compared to human postmortem tissues, are expected to affect our results, because we have focused on full-length transcript sequencing rather than read assembly. Our imposition of minimum expression and inherent length limitations could also lead to lower sensitivity of splicing diversity in lncRNAs, which have thus far generated staggering transcript structure complexity when sequenced after enrichment capture³⁴. We also had lower detection of pseudogenes, and we hypothesize that the PacBio platform’s accuracy and read length reduce the multimapping errors typical of short reads, especially for pseudogenes of highly expressed genes.

Within these boundaries outlined above, we were able to assess the sources and specifics of transcript structure diversity for major transcripts of most protein coding genes. Nearly all studies that have examined alternative splicing have emphasized transcript isoform multiplicity per gene. Also as expected, studies that have applied more permissive processing pipelines, used transcript assembly, or focused on nuclear RNA typically find evidence for far more RNA transcripts, especially at lower expression ranges^24,35–38. Assigning biological functions to new transcripts from our collection or any other contemporary study is a major challenge for the field. Unlike DNA replication’s elaborate mechanisms to ensure fidelity, the three major processes of RNA biogenesis mapped here are understood to operate with less stringent fidelity, and though it has long been debated, we consider evidence for the existence of a new transcript isoform simply makes it a candidate of interest for a protein coding, precursor, or regulatory function.

The range of regulation used by different genes was illuminating. COL1A1, a complex gene in terms of number of exons, exhibited minimal transcript structure diversity in spite of high expression versus other genes of similar expression levels, such as PKM with its many transcripts resulting from all three mechanisms. This implies that transcript structure diversity is a property of the gene that has been optimized in evolution. This has major implications for evaluating the functions of regulatory factors such as PAX6, which has 81 transcripts in GENCODE v40, and 33 transcripts in our dataset. Conventional gene-level short-read RNA-seq profiling is likely obscuring important distinctions in transcript usage. While not every one of these transcripts leads to a difference in the protein product, changes at the 5’ and 3’ end are likely to alter the regulation of those transcripts. The incorporation of transcript usage as well as its regulation within the framework of gene regulatory networks, where appropriate, is a major challenge going forward.

Considering transcript structure diversity as a fundamental, tunable property of gene function, the mouse-human comparative results were the most surprising to us. In genomics and in the wider biology community we often use orthology of mouse and human genes to predict and interpret gene function in vivo, including many uses of mice as mammalian models for both basic and preclinical purposes. The differences in transcript structure diversity that surfaced when we compared matching tissues from human and mouse suggests that this diversity is rapidly evolving on a per gene basis, even between primates and rodents. This is, however, consistent with prior observations of a large population of rapidly evolving candidate cis-regulatory elements³⁹. The results presented here provide a roadmap for evaluating the evolution of transcript structure diversity across species and impetus to focus on it, especially for genes with substantial differences that would affect interpretation of existing animal models and expectations for humanized gene-locus mouse models. It is hard to underestimate the need for better methods to test the functional significance of different transcript isoforms.

DATA AVAILABILITY

• Human LR-RNA-seq processed data / processing pipeline

• Human LR-RNA-seq datasets

• Mouse LR-RNA-seq processed data / processing pipeline

• Mouse LR-RNA-seq datasets

• Human short-read RNA-seq datasets

• Human microRNA-seq datasets