Profiling lariat intermediates reveals genetic determinants of early and late co-transcriptional splicing

SUMMARY

Long introns with short exons in vertebrate genes are thought to require spliceosome assembly across exons (exon definition), rather than introns, thereby requiring transcription of an exon to splice an upstream intron. Here, we developed CoLa-seq (co-transcriptional lariat sequencing) to investigate the timing and determinants of co-transcriptional splicing genome wide. Unexpectedly, 90% of all introns, including long introns, can splice before transcription of a downstream exon, indicating that exon definition is not obligatory for most human introns. Still, splicing timing varies dramatically across introns, and various genetic elements determine this variation. Strong U2AF2 binding to the polypyrimidine tract predicts early splicing, explaining exon definition-independent splicing. Together, our findings question the essentiality of exon definition and reveal features beyond intron and exon length that are determinative for splicing timing.

In brief

Zeng et al. report the timing of co-transcriptional splicing for tens of thousands of human introns, revealing that splicing timing varies dramatically across introns and early splicing occurs before transcription of a downstream 5′ splice site. Such variation in splicing timing is dependent on multiple genetic elements, including U2AF2 binding.

Graphical abstract

INTRODUCTION

Nuclear pre-mRNA splicing is catalyzed by the spliceosome, a large and dynamic ribonucleoprotein complex that removes an intron to ligate flanking exons through two sequential reactions.¹ In the first step, lariat formation, the 2′ hydroxyl of a conserved branch point (BP) adenosine attacks the 5′ splice site (SS) phosphate, forming a free 5′ exon and a branched lariat intermediate, characterized by a 2′−5′ phosphodiester linkage between the BP and the 5′ SS. In the second step, exon ligation, the 3′ hydroxyl of the free 5′ exon attacks the 3′ SS phosphate at the 3′ end of the intron, ligating the exons and excising the lariat intron. Pre-mRNA splicing in vivo is intimately coupled with transcription both physically and functionally.² Defining co-transcriptional splicing mechanisms remains a challenge, however, in part, due to limited insight into the timing of splicing relative to transcription.

During initial spliceosome assembly, the 5′ SS is recognized by the U1 small ribonucleoprotein particle (snRNP); the BP, by SF1; and the polypyrimidine tract (PPT) and 3′ SS AG dinucleotide, by U2AF2 and U2AF1 of the heterodimeric U2AF, respectively. Because both intron ends encode limited information, efficient and accurate SS recognition is thought to require cooperative binding of U1 snRNP and U2AF, in one of two ways.^3–5 First, for short introns, most common in lower eukaryotes, U1 snRNP and U2AF cooperate across an intron (intron definition).^6–8 Second, for longer introns separated by short exons, most common in vertebrates, U1 snRNP and U2AF instead cooperate across an exon (exon definition)^5,9; in exon definition, U1 snRNP binds to the 5′ SS downstream of the intron to be spliced. Because most human genes contain short exons separated by long introns, exon definition is thought to dominate in humans.^3,5,10 Indeed, perturbing splicing factor expression or SS recognition often induces exon skipping rather than intron retention, although to varying degrees.^3,5 Recent studies have suggested that exon ligation can happen before, rather than after, transcription of a downstream 5′ SS, which would be inconsistent with exon definition-mediated splicing^11–13; however, low sequencing depth and complications due to mRNA contamination question the generality and significance of these findings. Therefore, whether exon definition represents an obligatory mechanism of SS recognition in humans remains unknown. Furthermore, the relative importance of specific genetic features, including intron and exon length, to spliceosome assembly remains to be defined.

Studies indicate thatmostsplicing(>75%) occurs while RNA polymerase II (RNAPII) synthesizes RNA.^2,14–16 Co-transcriptional splicing is more likely than post-transcriptional splicing to be influenced directly by RNAPII and chromatin. Furthermore, at any given moment during co-transcriptional splicing, only a subset of SSs and regulatory elements of a gene are transcribed and available to the spliceosome and trans-acting splicing factors.^2,17 Thus, SS selection critically depends on splicing timing, which we operationally define here as RNAPII position at the time of splicing and which reflects the interplay between transcription speed and splicing speed. For example, studies have implied that slower transcription favors earlier splicing and the upstream of two competing 3′ SSs, whereas faster transcription favors later splicing and a downstream, stronger 3′ SS.^3,18 Furthermore, faster splicing favors in-order splicing in which introns are spliced in the order transcribed, whereas slower splicing permits frequently observed out-of-order splicing,^15,19 which can impact alternative splicing outcomes by changing the proximity of regulatory RNA elements in downstream regions.^20,21 Therefore, defining co-transcriptional splicing timing and the cis-regulatory sequences and trans-acting factors that govern such timing is paramount to define how SSs are selected and regulated during co-transcriptional splicing.

Recent genome-wide studies implementing distinct methodologies have begun to reveal the timing of co-transcriptional exon ligation.^{11–13,22–26} For example, some studies in lower eukaryotes and mammals have implicated fast splicing, occurring immediately after transcription of a 3′ SS.^{11–13,23,24} Other studies in budding yeast and mammals, although, imply slower splicing timings.^11,22,25,26 These differences could reflect splicing timing stochasticity or the impact of genetic factors on splicing timing. Alternatively, these differences could reflect methodological shortcomings, which include mRNA contamination during chromatin isolation, limited long-read sequencing depth, limited sampling during live imaging, as well as the preferential enrichment of U-rich transcripts or the perturbation of splicing kinetics when metabolically labeling nascent RNA^27,28; although metabolical labeling modestly impacts steady-state mRNA levels,²⁷ co-transcriptional splicing is much more sensitive to perturbation,¹¹ and the impact of metabolic labeling on co-transcriptional splicing remains unknown. Furthermore, due to limited sequencing depth, many of these methods only provided an analysis of global splicing timing and none have yet yielded insights into features that predict splicing timing for individual introns on the scale of the human transcriptome.

Although BP selection can impact SS selection, and mutations altering BP selection induce human disease,^29,30 BP usage, in comparison to 5′ and 3′ SS usage, is poorly defined in humans and many other organisms because of its degenerate sequence motif and the low abundance of lariat RNAs. To annotate BPs and study BP usage, studies have examined RNA-seq datasets for the incidentally captured, inverted reads that reveal juxtaposition of the 5′ SS with the BP. However, these reads are extremely rare, found in only ~1 in a million reads.³¹ Although the low abundance of these reads can be partially mitigated by mining thousands of RNA-seq libraries across different tissues and cell types,^32,33 such an approach precludes studies of differential BP usage. Alternatively, lariat RNAs have been enriched by degrading linear RNAs, isolating spliceosomes, capturing introns, or knocking down Dbr1, which debranches excised lariat introns (ELIs).^22,34–37 However, these methods are biased toward lariat RNA subpopulations, difficult to implement generally, or restrictive in that they uncouple the BP from the 3′ SS.

In this study, we present a novel transcriptome-scale approach, called CoLa-seq (co-transcriptional lariat sequencing), to map BP usage and to study human co-transcriptional splicing. CoLa-seq enabled BP mapping with a 1,000-fold increase in efficiency and revealed coupling between BP choice and 3′ SS selection. CoLa-seq also revealed that splicing timing varies at least five orders of magnitude and that a majority of introns can splice independent of exon definition. Multiple features predict splicing timing, including strong U2AF2 binding to the PPT, which emerges as a key predictor of early, exon definition-independent splicing. Overall, our findings question an obligatory status for exon definition, reveal features beyond intron and exon length that are determinative for splicing timing, and establish CoLa-seq as an approach to investigate the regulation of BP usage and co-transcriptional splicing.

RESULTS

A strategy to capture co-transcriptional lariat intermediates and excised lariat introns

To investigate the timing of co-transcriptional lariat formation and to define BP usage genome wide in humans, we developed CoLa-seq to capture and sequence the tails of nascent lariat intermediates (NLIs; Figures 1A and 1B). In reporting splicing timing, NLIs provide several advantages over nascent linear transcripts (Figure 1A). First, NLI tails indicate inherently that splicing is in progress; whereas the 3′ end reports on elongating RNAPII position, the 5′ end reveals that the RNA is branched and undergoing splicing chemistry. Second, NLIs, along with ELIs, encode BP usage and in a manner coupled to 3′ SS usage. Third, the unique chemical nature of the BP allows for NLI enrichment and the potential to capture splicing timing broadly in humans, where the genome encodes >375,000 introns.³⁸

Figure 1. CoLa-seq workflow and efficient, genome-wide BP mapping.

(A) Illustration of co-transcriptional splicing.

(B) CoLa-seq workflow. BP+1 marks RT stop. Gray boxes represent sequencing adaptors.

(C) Plots show meta-analysis of all 5′ or 3′ read ends, relative to annotated BPs or constitutive 3′ SSs.

(D) BP annotation workflow. (Left panel) Conversion of CoLa-seq 5′ ends (bars) to BP likelihood scores (green line). (Middle panel) Conversion of 5′ ends and sequence features to composite likelihood scores, with peaks greater than a threshold (dotted line) classified as BPs (vertical lines). (Right panel) Receiver operating characteristic curve tests likelihood functions against gold-standard BPs. (Inset) Zoomed view shows BP calling threshold (dotted line).

(E) Number of BPs identified by various studies.

(F) Stacked bar graph shows the usage of CoLa-seq-annotated BPs per 3′ SS.

To define NLI tails, we extracted chromatin-associated RNAs,^39–41 excluding MgCl₂ and including EDTA to inhibit splicing. We enriched for NLIs by degrading linear RNAs; captured RNAPII positions by ligating a UMI (unique molecular identifier)-containing adaptor to RNA 3′ ends; captured BP positions by reverse transcription, which stops at the 2′−5′ linkage; and amplified the cDNA for short-read sequencing (Figure 1B). Because 90% of BPs reside within the last 50 nts of an intron, and splicing can occur as soon as an intron is transcribed,^12,13 inserts derived from early-forming NLIs are short enough for short-read sequencing, which allows for greater depth than longer-read sequencing. Furthermore, despite using short-read sequencing, CoLa-seq reported on later-forming NLIs (see below). Thus, a CoLa-seq read, corresponding to an NLI tail, reveals both BP location at its 5′ end and RNAPII position at its 3′ end (Figures 1A and 1B).

CoLa-seq also captures ELI tails because a co-transcriptionally spliced ELI remains associated with chromatin until spliceosome disassembly^42–44 (Figure 1A). Like the 5′ end of an NLI read, the 5′ end of an ELI read reveals the BP location, but its 3′ end reveals a 3′ SS, coupling BP usage to 3′ SS usage (Figures 1A and 1B).

In a first application of CoLa-seq, we prepared two libraries from human K562 cells (Figure S1A). Confirming lariat RNA enrichment by CoLa-seq, the 5′ ends of CoLa-seq reads peaked one nucleotide downstream of annotated BPs,³² and their 3′ ends peaked at annotated 3′ SSs (Figure 1C), indicating ELI enrichment. This enrichment decreased drastically in cells treated with the splicing inhibitor pladienolide B or the transcription elongation inhibitor flavopiridol (Figures S1B and S1C).

CoLa-seq efficiently defines BP usage genome wide

For CoLa-seq reads with 5′ ends located within 100 nts upstream of a 3′ SS, only 24% of their 5′ ends map to an annotated BP (Figure S2A), but given that BP annotations are far from complete,^32–34,36 some 5′ ends may map to unannotated BPs. Thus, to more comprehensively annotate BPs and to more sensitively identify lariat RNA-derived reads, we developed a computational pipeline to systematically identify BPs using CoLa-seq data.

First, we identified reads with a 5′ end within 100 nts upstream of a 3′ SS and classified them as putative NLI reads if they traversed a 3′ SS or as putative ELI reads if they ended at a 3′ SS, a more stringent definition. After this simple classification, 5′ ends of both putative NLI and ELI reads already resembled annotated BPs in terms of their sequence context and distance to the 3′ SS (Figure S2B),^32,36,45 especially for putative ELI reads, suggesting together that a substantial fraction of putative NLI reads derived from authentic NLIs and most putative ELI reads derived from authentic ELIs.

To refine BP mapping, we built a classifier that integrates 5′ end CoLa-seq coverage with BP motif and position information³² to compute a composite BP-likelihood score for each nucleotide (Figure 1D). We evaluated the BP calling accuracy of the com- posite BP-likelihood score using gold-standard BPs observed in both a previous large-scale study³² and our own analysis of K562 cells by lariat-seq (Table S1). Our classifier identified BPs with high accuracy (accuracy under curve = 0.98; Figure 1D) and recalled 80% of the gold-standard BPs at a 1% false positive rate.

At this false positive rate, our classifier identified 165,282 high-quality BPs (Figures 1E, S2C, and S2D;Table S2), including 750 U12-type BPs (Figure S2E) and 151 BPs associated with 106 previously identified recursive SSs⁴⁶ (e.g., Figure S2F). These high-quality BPs showed expected characteristics (Figure S2D); indeed, 14,750 BPs were also observed by lariat-seq (Figure S2G). Notably, our classifier identified more BPs than the most recent large-scale study (Figure 1E), which analyzed 1,000 times as many reads pooled across 17,164 RNA-seq libraries from diverse tissues and cell types.³² Importantly, CoLa-seq BPs overlap with previously, experimentally identified BPs to a similar extent as this large-scale study (Figure S2G). Of CoLa-seq BPs, 58% (88,881) were novel (Figure S2D), underscoring that BP annotations remain unsaturated, and we confirmed 5,733 by lariat-seq (Table S1). Previously annotated BPs not identified by CoLa-seq are enriched in genes lowly or not expressed in K562 cells (Figure S2H). Together, these observations highlight the utility of Cola-seq in mapping BPs in a single cell type.

Consistent with previous findings,³² many constitutive 3′ SSs (32%) were associated with more than one BP (Figure S2I). However, although a subset of these 3′ SSs used two or more BPs, most 3′ SSs used a single dominant BP (Figure 1F). Even when considering the more inclusive BP set from putative ELI reads (Figure S2B), more than 50% of the 3′ SSs used predominantly a single BP over 50% of the time (Figure S2J). Thus, in a single cell type, a single BP dominates in the majority of introns.

Excised lariat introns reveal coupling between BP and 3′ SS usage

Using CoLa-seq BP annotations, we classified 75% of putative ELI reads as ELI reads. Compared with the largest published BP annotation,³² CoLa-seq BP annotations identified twice as many ELI reads (Figure S2A). ELI reads derived primarily from protein-coding genes, showed expected BP characteristics, and displayed strong reproducibility between biological replicates (Pearson’s r = 0.98; Figures 2A and 2B). In total, we detected 2,574,753 ELI reads associated with 57,309 3′ SSs (≥3 ELI reads; Figure 2C). To our knowledge, these data represent the first genome-wide dataset of ELIs in humans.

Figure 2. CoLa-seq captures excised lariat introns, revealing BP-3′ SS coupling.

(A) Top panel: seqlogo of ELI-derived BP motif matches YUNAY consensus.⁴⁵ Bottom panel: histogram of BP-3′ SS distances of ELI reads.

(B) Scatterplot of ELI reads (≥3) associated with a 3′ SS, showing biological replicates. “R,” Pearson’s correlation coefficient.

(C) Bar plot shows cumulative intron counts as a function of ELI read coverage.

(D) Example of ELI reads.

(E) Bar plot shows the percentage of alternative 3′ SS pairs using the same set of BPs, as a function of 3′ SS separation.

(F) Example of ELI reads.

ELI reads revealed cases of strict coupling between BP and 3′ SS choice. For example, for two alternative 3′ SSs in RNF111, the upstream 3′ SS always coupled to an upstream BP, whereas the downstream 3′ SS always coupled to a downstream BP (Figure 2D); in this case, the downstream 3′ SS is likely too far (59 nts) from the upstream BP, and the upstream 3′ SS is too close (3 nts) to the downstream BP. Although 3′ SS recognition by U2AF can impact BP selection prior to lariat formation, in this example, BP selection and lariat formation appear to restrict 3′ SS selection. Conversely, ELI reads revealed that alternative 3′ SSs that are near to one another are often chosen independently of the BP; when competing 3′ SS pairs were within 9 nts, 87% used the same set of BPs (Figures 2E and 2F; cf. Dou et al.⁴⁷). Thus, by capturing coupled BP and 3′ SS information, ELI reads directly reveal whether BP selection restricts 3′ SS selection.

Nascent lariat intermediates reveal in-order, out-of-order, and concurrent splicing

Using our BP annotations, we classified 39% of putative NLI reads as NLI reads. Compared with the largest published BP annotation,³² CoLa-seq BP annotations identified nearly twice as many NLI reads (Figure S2A). NLI reads derived primarily from protein-coding genes, showed expected BP characteristics, and displayed strong reproducibility between biological replicates (Pearson’s r = 0.97; Figures 3A and 3B). In total, we detected 2,201,036 NLI reads associated with 73,562 3′ SSs (R3 NLI reads; Figure 3C).

Figure 3. Nascent lariat intermediates capture in-order, out-of-order, and concurrent splicing, revealing dramatic splicing timing variation.

(A) Top panel: seqlogo of NLI-derived BP motif matches YUNAY consensus.⁴⁵ Bottom panel: histogram of BP-3′ SS distances of NLI reads.

(B) Scatterplot of NLI reads (≥3) showing biological replicates.

(C) Bar plot shows cumulative intron counts as a function of NLI read coverage.

(D) Venn diagram shows the overlap between NLI- and ELI-associated BPs.

(E) Histogram of BP-3′ SS distance for indicated reads (≥10).

(F) Schematic of co-transcriptional splicing orders. The dotted line indicates completed splicing.

(G) Example of NLI reads.

(H) Cumulative plot of 3′ SS-RNAP II distance distributions for all in-order and out-of-order NLI reads (black line), 50 randomly sampled introns (colored lines; >30 NLI reads), or a theoretical exponential distribution with a splicing half-life set to a major tick mark (dashed lines).

(I) Histogram of splicing timing variation, per intron, defined from in-order and out-of-order NLI reads.

(J) Histogram of the ratio of IQR_obs to IQR_exp across introns. Shaded area marks introns (52.4% of total) with a ratio of IQR_obs/IQR_exp between 0.5 and 2. Inset shows IQR_obs (black line) and IQR_exp (gray line) for an example intron.

(K) Boxplots show % in-order splicing as a function of US-DS 3′ SS distance, binned by deciles. p-values were calculated by Mann-Whitney U test; ns (not significant), p > 0.05; *, p ≤0.05; **, p ≤ 0.01; ***, p ≤ 0.001; ****, p ≤ 0.0001.

(L) Line plot shows % in-order splicing as a function of US-DS 3′ SS distance for observed (black line) and simulated data (colored lines), in which t_1/2, the median, genome-wide splicing half-life, was varied.

(M) Top panel: cassette exon architecture. Bottom panel: cumulative plot of % out-of-order splicing for indicated introns. The vast majority of cassette exons were mostly included.

(N) Example of NLI reads showing slower splicing of introns flanking a cassette exon (gray).

Although NLIs might reflect stalled intermediates, our results strongly imply that NLIs captured by CoLa-seq represent functional splicing intermediates. Specifically, despite size bias in the capture of ELI cDNAs (Figure S2K), 75% of NLI-associated BPs overlapped with ELI-associated BPs (Figure 3D), and NLI and ELI read numbers positively correlated (Pearson’s r = 0.32), consistent with functional conversion of NLIs to ELIs. Introns that yielded NLI reads but not ELI reads did so for technical reasons; these introns have short BP-3′ SS distances, precluding the efficient capture of the corresponding ELI tails (Figure 3E). Additionally, we found no major differences between introns that yielded NLI reads and all annotated introns, including for features that limit the conversion of NLIs to ELIs or favor the conversion of pre-mRNA to NLIs (Figure S2L). To test whether CoLa-seq NLIs reflected splicing events that occurred co-transcriptionally in vivo rather than during chromatin fractionation, we directly captured intron-containing RNAs from total RNA extracted under denaturing conditions. Indeed, we observed reads corresponding to lariat intermediates, including short-tailed intermediates (Figure S2M). To our knowledge, the NLI reads identified by CoLa-seq represent the first major dataset of NLIs in any organism. The depth and breadth of these datasets allowed us to study co-transcriptional splicing at the level of individual introns, which has been difficult with lower-depth, long-read sequencing.

To characterize NLI reads, we first interrogated the splicing order of adjacent intron pairs by defining the splicing status of a downstream intron relative to an upstream intron in which an NLI was detected. We observed three types of NLI reads, reflecting: (1) in-order splicing, (2) out-of-order splicing, and, unexpectedly, (3) concurrent splicing (Figures 3F, 3G, and S3A) reads, which were highly reproducible between replicates (Figure S3B). In-order splicing represents the earliest splicing timing for an intron and is indicated by NLI reads reflecting an incompletely transcribed or unspliced downstream intron. By contrast, out-of-order splicing represents the latest splicing timing for an intron and is indicated by NLI reads reflecting the transcription and splicing of one or more downstream introns (Figures 3F and 3G), including very long introns indicating very late splicing (Figure S3K). We observed in-order splicing in 93% of introns and out-of-order splicing in 46% (Figure S3C), although CoLa-seq is biased toward shorter reads and earlier splicing events (Figure S3E).

Strikingly, NLI reads also indicated concurrent splicing, in which two adjacent introns splice at the same time. To our knowledge, concurrent splicing has only been observed in vitro and for a synthetic substrate.⁴⁸ Concurrent splicing represents an intermediate splicing timing and is indicated by NLI reads that end at the last nucleotide of the downstream exon (Figure S3G), reflecting cleavage of the downstream 5′ SS. Thus, these concurrent NLI reads indicate that the downstream intron is transcribed and at the intermediate stage of splicing (Figures 3F and 3G); indeed, 64% of these downstream introns also yielded NLIs (R5 reads). In our dataset, 90% of introns reported concurrent splicing (Figure S3C). Although concurrent NLI tails are relatively short and consequently preferentially captured, the prevalence of concurrent splicing suggests coordinated splicing of adjacent introns. Because concurrent splicing reports on splicing order independent of RNAPII position, concurrent splicing can report on late splicing timings, dependent on downstream intron transcription and length, even when assayed by short-read sequencing.

Co-transcriptional splicing timing varies dramatically across introns

We investigated splicing timing by measuring the genomic distance from the 3′ SS to the RNAPII position (3′ SS-RNAPII distance) for each NLI. Aggregating in-order and out-of-order NLI reads across introns revealed that the timing of all these splicing events in our dataset varied by at least five orders of magnitude (Figures 3H and S3I). By concurrent splicing, which reports on the minimum splicing timing (Figure 3F), splicing timing was similarly variable (Figure S3J).

The depth of CoLa-seq allowed us to assess splicing timing variation at the level of individual introns. Considering 17,448 introns with R10 in-order or out-of-order NLI reads, splicing timing varied substantially (Figures 3I, S3H, and S3K), with interquartile (IQR) variation larger than 10,000 nts for 6% of introns. Such large variation would be expected for NLIs with long median 3′ SS-RNAPII distances, assuming first-order kinetics (i.e., exponential decay). To test this possibility, we compared the observed variation of each intron with the variation expected for an exponential model with the observed median 3′ SS-RNAPII distance. Consistent with first-order kinetics, the majority of introns showed variation (IQR_obs) within 2-fold of the expected variation (IQR_exp; Figure 3J). Interestingly, some introns (11%) varied 10 times greater than expected, whereas others (4.2%) varied 10 times narrower than expected, suggesting heterogeneity in splicing or transcription. These splicing timing variations have important implications for splicing regulation.

To estimate splicing timing, we calculated the read distribution between in-order and out-of-order splicing as a function of the distance between the upstream and downstream 3′ SSs (US-DS 3′ SS distance). To eliminate size bias favoring short in-order NLI reads, we calculated % in-order splicing using only in-order and out-of-order NLI reads that traverse the downstream 5′ SSs (5,840 intron pairs) and thus have similar length distributions (Figures S3D–S3F). If splicing were stochastic and slow such that the transcriptional lag between an upstream and downstream intron were negligible, then the probability of splicing in-order versus out-of-order would be equally likely (50% in-order splicing), as observed for shorter distances (<428 nts; Figure 3K). However, as the US-DS 3′ SS distance increases (>428 nts), the median % in-order splicing increases, and when >1.5 kb, it exceeds 75% (Figure 3K). This observation implies that during co-transcriptional splicing, the observed introns splice mostly before RNAPII has transcribed 1.5 kb downstream of the 3′ SS.

To assess how this correlation informs splicing timing, we simulated splicing timings and examined how our observed data fit various simulated scenarios. Briefly, we assumed that splicing half-lives of introns follow a lognormal distribution genome-wide with an unknown genome-wide median splicing half-life (Figures S3L–S3N). To estimate this median splicing half-life, we simulated % in-order splicing as a function of the US-DS 3′ SS distance for different median splicing half-lives (Figure S3M). For a US-DS 3′ SS distance of 1.5 kb and a simulated median splicing half-life of ~5 kb, the simulated % in-order splicing approaches 50% (Figures 3L and S3N), rather than the observed 75% (Figure 3K), implying the median splicing half-life is <5 kb. Indeed, in all simulated scenarios, the median splicing half-life must be ≤2.5 kb for the simulated % in-order splicing to match the high observed % in-order splicing (Figures 3L and S3N). Although splicing timing can vary dramatically, the analyses of these ~6,000 introns indicate that at least one third of co-transcriptional splicing events occur before a median-sized downstream intron is fully transcribed.

Consistent with previous evidence that U12 introns splice slower than U2 introns,^25,49,50 U12 introns yielded lower % in-order splicing than U2 introns (Figure S3O). Previous findings also provide evidence that alternative introns splice slower than constitutive ones.^39,51,52 To test for slower splicing, we examined splicing order for intron pairs that are upstream and downstream of cassette exons (Figure 3M). Indeed, intron pairs upstream of the cassette exons (−2 and −1 introns) show lower % out-of-order splicing than all intron pairs (mostly constitutive), whereas intron pairs downstream of the cassette exons (+1 and +2 introns) show higher % out-of-order splicing (Figures 3M and 3N). This observation indicates that introns flanking cassette exons (−1 and +1 introns) splice later than their neighboring constitutive introns (−2 and +2 introns), consistent with recent studies on the or- der of exon ligation.^11,19 The 3′ SS-RNAPII distance distribution of NLIs revealed similar results (Figure S3P). The later splicing of introns flanking cassette exons supports a role for splicing timing variation in regulating alternative splicing.

Intronic elements, gene architecture, and genomic context predict co-transcriptional splicing order

The substantial variation in splicing timing across introns suggested that specific genetic features influence splicing timing. Indeed, many features correlate with a given splicing order (Figure S4A). For example, introns with high % in-order splicing tend to have strong splicing signals, longer downstream introns, and lower regional GC content. However, unintuitively, higher % in-order splicing also correlated with a longer upstream intron. Furthermore, many of the correlated features correlate among themselves (e.g., Figure S5G), complicating the identification of key features that impact splicing timing.

To disentangle the impact of sequence features on co-transcriptional splicing, we trained computational models that use sequence features to predict a % splicing order for an intron pair, using a decision tree-based ensemble machine learning algorithm, XGBoost.⁵³ We selected 22 features of 235 screened for modeling (Table S3). We trained models on 23,324 introns for % in-order splicing, % concurrent splicing, or % out-of-order splicing. These XGBoost-based models accounted for 40% ± 1% (standard deviation), 34% ± 1%, and 15% ± 1% of the overall variance for in-order splicing, concurrent splicing, and out-of-order splicing, respectively, outperforming linear regression-based models by 1.3- to 2.2-fold (Figure S4B).

For each trained model, we calculated the impact of each feature on each prediction using SHAP (Shapley additive explanation).⁵⁴ Briefly, for each prediction of a % splicing order for an intron pair, SHAP determines the impact of each feature by comparing the prediction of a % splicing order based on the feature value of the intron pair to the prediction based on the average feature value of all intron pairs. Additionally, SHAP breaks down the impact of each feature into its “main impact” (independent of other features) and “interaction impact” (dependent on other features), which was minimal in our models. Importantly, the prediction for each intron pair depends on multiple features (e.g., Figures 4A, 4B, S4C, and S4D). Three general feature classes emerged: (1) intronic splicing elements, (2) intron and exon lengths, and (3) GC content (Figures 4A and S4C). Most features are consistent with a role in accelerating or delaying splicing of either the upstream or downstream intron.

Figure 4. Intronic elements, gene architecture, and genomic context predict the co-transcriptional splicing order.

(A) The impact, averaged across introns, of the top features is shown for models for the orders of splicing. The average impact from 5-fold cross validations is plotted as points, and averages of these points, as bars. Average impact is calculated as the mean of absolute SHAP values, and the sign of the impact was determined by Spearman’s correlation. US, upstream intron harboring NLI; exon, downstream exon; DS, downstream intron.

(B) Force plot of features, for randomly sampled introns, that increase or decrease the predictions from the average % in-order splicing of all introns (44%).

(C) Averaged eCLIP signals⁵⁵ of U2AF in the indicated regions of upstream and downstream introns for NLIs with high or low % in-order splicing. Confidence interval, 95%.

(D) Scatterplot of US intron length main impact on the % in-order splicing prediction.

(E) Scatterplot of DS intron length main impact on the % order of splicing predictions.

(F) Line plots of GC content of intron pairs with different in-order splicing levels. Confidence interval (shaded area), 95%.

Supporting the significance of % splicing order values and the validity of the models, the models identified predictive features expected to compromise the conversion of concurrent splicing NLIs to out-of-order splicing NLIs, which reflects exon ligation for the downstream intron.⁵⁶ Specifically, a longer downstream BP-3′ SS distance, a weaker downstream 3′ SS, or a down- stream 3′ SS with high GC content predicts a higher % concurrent splicing and a lower % out-of-order splicing (Figure S5A). At the same time, a longer BP-3′ SS distance of the upstream intron, a distance that correlates with PPT strength (Pearson’s r = 0.3), predicts earlier lariat formation of the upstream intron relative to the downstream intron (Figures 4A, S4C, and S5B). BP-3′ SS distances contribute significantly to the models, accounting, for example, for 11% of the predictive power of the in-order splicing model, a contribution determined as the mean absolute SHAP value of a feature across all introns divided by the sum of the mean absolute SHAP values for all features.

As anticipated, the strength of intronic splicing elements also emerged as predictive in our models and accounted, for example, for 13% of the predictive power of the in-order splicing model. U content in the upstream BP-3′ SS region emerged as the most important feature, with higher U content predicting higher % in-order splicing (Figure 4A), which is notable because this U content correlates with PPT strength (Pearson’s r = 0.34). Further, by k-mer analysis and enhanced cross-linking immuno- precipitation sequencing (eCLIP-seq),⁵⁵ introns with higher % in-order splicing are enriched for U-rich hexamers (Figure S5C) and show increased interactions with U2AF (Figure 4C), implying that strong U2AF binding to the PPT promotes early splicing. Interestingly, the BP-3′ SS region had a greater impact than the 5′ SS region on the models (Figure S4C), suggesting that the later transcription of the intron’s 3′ end imposes a greater constraint on 3′ SS recognition for early splicing of an intron.

Importantly, the upstream and downstream intron lengths contribute to the model predictions, accounting for 3% and 21%, respectively, of the predictive power of the in-order splicing model, for example (Figures 4A, 4D, and 4E). (Although exon length also contributes to model predictions, correcting for size bias eliminates the correlation between exon length and % in-order splicing (Figures S3D–S3F, S4A, and S5D). Notably, contrary to a simple correlation analysis¹³ (also see Figure S4A), the model revealed that shorter upstream introns predict higher % in-order splicing, especially for introns shorter than 200 nts (Figure 4D), consistent with expectations (e.g., Fox-Walsh et al.⁹ and Toan et al.⁵⁷).

The model revealed that longer downstream introns (>1 kb) predict increasingly higher % in-order splicing (Figure 4E). Indeed, for downstream intron lengths > 5 kb, this feature dominates in-order splicing predictions (Figure S5E). The importance of downstream intron length is not due to its impact on splicing of the downstream intron because the importance of upstream intron length in predicting upstream intron splicing is marginal beyond ~1 kb (Figure 4D). Instead, these findings imply that for downstream introns longer than 1 kb, % in-order splicing increases with increasing downstream intron length because the upstream intron increasingly splices before the downstream intron is transcribed, consistent with an estimated median splicing timing of %2.5 kb (Figures 3L and S3L–S3N). These findings highlight the importance of transcription time on co-transcriptional splicing outcomes.

The GC content of the upstream intron, the exon, and the downstream intron feature prominently in the models, accounting for 12% of the predictive power of the in-order splicing model, for example. High GC content in all three regions predict low % in-order splicing (Figures 4A and S5F). Because the GC content of all three regions strongly correlate (Figure S5G), their importance in the models suggests that regional GC content bias influences splicing timing. High GC content correlates with shorter introns having flanking exons with similar GC content, whereas low GC content correlates with longer introns having flanking exons with higher GC content.⁵⁸ Notably, the high GC class associates with low % in-order splicing, whereas the low GC class associates with high % in-order splicing (Figure 4F). Regional GC content bias often extends beyond an intron or gene.^58–60 Indeed, introns from the same gene have similar GC content (p value < 4.3 × 10⁻¹⁶⁷; F test) and similar splicing timings, as measured by % in-order splicing (p value < 1.2 × 10⁻²⁴; chi-square test). High GC content weakens SS strength⁵⁸ and increases local secondary structure potential (also see Figure S5H), but GC content could impact splicing timing through multiple mechanisms (see discussion). Taken together, our findings imply that the broad genomic context of an intron is a critical determinant of splicing timing.

3′ splice site features and GC content predict the timing of co-transcriptional in-order splicing

We examined the timing of early lariat formation in terms of the 3′ SS-RNAPII distance distribution of in-order NLIs reads, the medians of which are reproducible between replicates (Pearson’s r = 0.68). The 3′ SS-RNAPII distances vary by over hundreds of nucleotides for individual introns and, importantly, between introns (Figures 5A–5C, S6A, and S6B), implying a genetic basis for in-order splicing timing.

Figure 5. Specific features predict early in-order splicing.

(A) Density plot of the 3′ SS-RNAP II distance distribution of all in-order NLI reads.

(B) Density plots of the 3′ SS-RNAP II distance distribution for each intron.

(C) The first 20 in-order NLI reads are visualized for OPLAH intron 20, EIF4G2 intron 16, and HNRNPK intron 16 (of 56, 85, and 440 in-order reads, respectively).

(D) Force plot of features that increase or decrease the predictions from the average of the median 3′ SS-RNAPII distance across introns (131 nts).

(E) The average impact of the top features for the median 3′ SS-RNAPII distance model. Details, as in Figure 4A.

(F) Hexamer analysis of the BP-3′ SS region. Details, as in Figure S5C.

(G) Seqlogos illustrate motifs enriched in the BP-3′ SS region of early or late introns.

(H) Averaged eCLIP signals⁵⁵ of splicing factors for early or late introns. The eCLIP signal for Aquarius (AQR) is shown as a negative control. Details, as in Figure 4C.

To investigate the impact of sequence features on splicing timing, we trained two computational models on 14,437 introns using XGBoost.⁵³ In the “earliest” model, the target variable was the shortest 3′ SS-RNAPII distance for in-order NLIs of a given intron (Figure 5B; median, 18 nts), reflecting the requirements for the earliest events; in the “median” model, the target variable was the median 3′ SS-RNAPII distance for in-order NLIs of a given intron (Figure 5B; median, 107 nts), reflecting the requirements for the population of events. We selected 37 features of 235 screened (Table S3) for modeling, including the number of NLI reads per intron to account for sampling bias. The resulting earliest and median models accounted for 31% ± 2% and 28% ± 1% of the variance in 3′ SS-RNAPII distance, respectively. Both models reveal that multiple features contribute to splicing timing prediction for each intron (Figure 5D) and highlight the importance of two general features—the 3′ SS region and GC content (Figure 5E and S6C–S6E; note: upstream and downstream intron lengths emerged as contributing features but for technical reasons—see STAR Methods). Importantly, in the median model, relative to the earliest model, the 3′ SS features decrease in importance, whereas downstream, exonic features generally increase in importance (Figure S6C), validating that the median 3′ SS-RNAPII distances represent later splicing events.

Consistent with an early timing for in-order NLI events, in both models features flanking the 3′ SS dominated in importance (Figures S6D and S6E), excluding BP-3′ SS distance for technical reasons (see STAR Methods). In the median model, the 3′ SS-flanking features accounted for 45% of the predictive power (Figure 5E and S6E). In particular, high U and low C content in the BP-3′ SS region predict shorter 3′ SS-RNAPII distances and earlier in-order splicing (Figure 5E and S6D–S6G). Notably, the BP-3′ SS region of “early in-order” introns is enriched for U-rich hexamers and a strong U2AF2 consensus site, whereas the same region of “late in-order” introns is depleted of U-rich hexamers and instead enriched for C/U-rich hexamers (see STAR Methods; Figures 5F, 5G, and S6H; cf. Figures S6I and S6J), indicative of weak U2AF2 binding.⁶¹ Indeed, by eCLIP-seq,⁵⁵ early in-order introns showed increased interactions between an intron 3′ end and U2AF (Figure 5H). These observations indicate that optimal U2AF binding promotes early lariat formation.

Early co-transcriptional lariat formation implicates widespread splicing independent of exon definition

Because exon definition is thought to predominate in humans,^3,5 we expected in the modeling of 3′ SS-RNAPII distance that downstream 5′ SS strength and exon length would contribute to predictions. Indeed, a stronger downstream 5′ SS and a shorter exon predict earlier in-order splicing in the median model (Figure S6E); however, compared with the 3′ SS-flanking features, these features accounted only modestly (5% versus 45%) for the predictive power of the median model (Figure S6E). Furthermore, exon length does not correlate with % in-order splicing (Figure S5D). These results suggest that exon definition, dependent on a 5′ SS, may only modestly contribute to in-order splicing timing, at least when RNAPII is within hundreds of nucleotides of the 3′ SS.

To test further for a role for exon definition in splicing timing, we asked whether in-order splicing timing depends on transcription of the downstream 5′ SS by examining the 3′ end locations of in-order NLI reads relative to the downstream 5′ SSs. The 3′ ends distributed broadly and extended both downstream and upstream of 5′ SSs (Figures 6A, S7A, and S7B). Similar distributions can be observed for individual introns (Figure 6B). When considering all spicing orders, only 10% of assayed introns splice exclusively downstream of the downstream 5′ SS (Figure S7C), although we estimate that for an average-sized intron a large majority of splicing events occur downstream of a 5′ SS. Compared with introns upstream of constitutive exons, introns upstream of cassette exons tend to splice more frequently when RNAPII is downstream of the downstream 5′ SS (Figure S7D), consistent with evidence that cassette exons splice later^11,19 (Figures 3M, 3N, and S3P) and the expectation that cassette exons are more dependent on exon definition.⁴ Although the size bias of CoLa-seq precludes rigorously quantitating the absolute fraction of splicing events that occur upstream of a downstream 5′ SS, these observations imply that the vast majority of introns can splice, at least to some extent, independent of a downstream 5′ SS and consequently independent of a classic mode of exon definition.

Figure 6. Early co-transcriptional lariat formation indicates widespread exon definition-independent splicing.

(A) Density plot of in-order NLI 3′ ends around the downstream 5′ SS. Confidence interval, 95%; a.u., arbitrary units.

(B) Example of NLI 3′ ends upstream (purple) or downstream (blue) of the downstream 5′ SS.

(C) Intron length density plot for early introns (≥25% in-order NLIs ending in exons).

Because intron definition is thought to be favored for short introns and exon definition for longer introns,⁴ we assessed splicing timing relative to the downstream 5′ SS as a function of intron length. Remarkably, introns spanning a wide range of lengths yielded NLIs with 3′ ends upstream of the downstream 5′ SS (Figures 6C and S7E), implying that longer introns do not absolutely necessitate U1-dependent exon definition, contrary to a fundamental premise of the exon definition model.^3,5 Early in-order, exon definition-independent splicing of long introns likely reflects many additional features that contribute to the timing of co-transcriptional lariat formation—most notably, strong binding of U2AF2 to the PPT (Figures 5D–5H and S6D–S6J).

Evidence that early lariat formation driven by U2AF2 recognition of the PPT is AG/U2AF1-independent

Biochemical studies have found that some introns splice in an AG and U2AF1-independent manner and that such introns generally have strong PPTs for strong U2AF2 binding.^62,63 Because high U content and strong U2AF2 binding in the BP-3′ SS region predict earlier splicing (Figures 5E–5H, 7A, and S6D–S6H), we investigated whether strong PPT recognition by U2AF2 in early in-order introns renders lariat formation AG- and U2AF1-independent. Indeed, for many of the early in-order introns, the 3′ end of the shortest NLI is located less than 15 nts downstream of the 3′ SS (Figure 7B), and thus, the AG would still be inside of RNAPII when splicing occurred.^64,65 Although only a portion of the BP-3′ SS region of these “ultra-early” NLIs would be available for U2AF2 binding, these emerged, exposed regions are long enough to accommodate U2AF2 binding (Figure 7C) and have high U content, U-rich hexamers, and a strong U2AF2 binding motif (Figures 7D–7F); UUUUAG hexamers were de-enriched (Figure 7E), arguing against the possibility that upstream, cryptic AGs account for ultra-early splicing. Together, these observations support the hypothesis that early lariat formation is dependent on strong U2AF2 binding and independent of the 3′SS AG and U2AF1 (Figure 7G).

Figure 7. Early lariat formation is AG/U2AF1-independent.

(A) Boxplots of U content in the BP-3′ SS region as a function of the shortest 3′ SS-RNAPII distance of NLIs.

(B) Density plot of the shortest 3′ SS-RNAPII distance of early in-order introns.

(C) Density plot of the exposed region length downstream of the BP for the shortest NLI of each intron.

(D) Boxplots of U content in the exposed BP-3′ SS region. P values are calculated as in Figure 3K.

(E) Hexamer analysis of the exposed BP-3′ SS regions. Details, as in Figure S5C.

(F) Motif similarity search after de novo analysis for motifs in the exposed BP-3′ SS region.

(G) Model illustrating the interplay between splicing timing and AG/U2AF1-dependence. “1,” U2AF1.

(H) Co-transcriptional splicing timing is highly variable across introns; examples of early and late splicing are highlighted. This variation was used to train models to identify features contributing to model predictions. Many genetic features, including several key features, contribute to splicing timing.

DISCUSSION

The timing of splicing relative to transcription can impact splicing outcomes^2,66; however, our understanding of co-transcriptional splicing timing has been limited. In particular, the genetic determinants of splicing timing remain to be defined. In this work, we have developed CoLa-seq to assay lariat formation timing relative to transcription on a genome-wide scale (Figure 1). Lariat formation timing varies dramatically both across introns and within an intron (Figures 3G–3I and S3H–S3K), and the splicing timing of regulated introns is later than constitutive introns (Figures 3M, 3N, and S3P). Furthermore, machine learning-based modeling revealed that multiple features contribute combinatorially to predictions of lariat formation timing, with intron length, GC content, and the PPT playing major roles (Figures 4, 5, and S4–S6); strong binding of the PPT by U2AF2 correlates with early, in-order splicing (Figures 4C, 5F–5H, 7A, 7D–7F, S5C, and S6F–S6H). Unexpectedly, many lariats form before transcription of a downstream 5′ SS (Figures 6 and S7) and some form even before the 3′ SS has emerged from RNAPII (Figures 7A–7G), both implying that exon definition is not obligatory, even for long introns; although intron definition has been perceived historically as the alternative to exon definition, our data suggest a third pathway to spliceosome assembly (see below). By enriching for lariat RNA species, CoLa-seq also enabled BP mapping with unprecedented efficiency and yield (Figure 1), yielding the largest BP dataset to-date (165,282 BPs) and revealed coupling of BP and 3′ SS choice (Figures 2 and S2A–S2G), paving the way for future studies on the regulation and perturbation of BP usage, such as the impact of somatic cancer mutations on BP and 3′ SS usage³⁰ (cf. Figures 2D–2F). Overall, our analysis of splicing by CoLa-seq has revealed that lariat formation timing varies widely, can be quite early, and depends particularly on several genomic features (Figure 7H), all of which have important implications for splicing regulation.

Due to limited read depth of long-read sequencing, recent genome-wide studies of co-transcriptional splicing resorted to analyzing global splicing timing by aggregating reads across different introns. In this study, leveraging short-read sequencing and its depth, we were able to examine and model lariat formation timing of individual introns across the transcriptome. The wide range of splicing timing observed (Figures 3H, 3I, and S3H–S3K) provides opportunities for splicing regulation. For instance, changes in splicing timing could alter the % out-of-order splicing, changing the availability of RNA regulatory elements (e.g., Nasim et al.²⁰ and Takahara et al.²¹). Additionally, during co-transcriptional splicing, an average intron splices on a time-scale that is frequently shorter than the timescale for transcribing a downstream intron that is average in length (Figures 3K, 3L, and S3L–S3N), and longer downstream introns predict higher % in-order splicing (Figure 4E). Consequently, the transcription time of the downstream intron could be regulated to influence exon skipping, consistent with implications of transcription as a key target for splicing regulation.⁶⁶

High GC content in the exon and flanking introns predicts low % in-order splicing (Figures 4A, S4C, and S5F). Because GC content bias extends well beyond an intron and even a gene^58–60 (also see Figure S5G), our detected correlation of splicing timing of nearby introns, also observed by others,^11,12 may not reflect splicing coordination between adjacent introns but rather reflect the impact of regional GC content bias on splicing timing across multiple introns within a gene. Indeed, strong U2AF binding, favored in GC-poor introns, not only predicts early splicing but also correlates between adjacent introns (Figure 4C). High GC content likely impacts splicing timing in additional ways, such as weakening SS strength,^58,67 stabilizing local RNA secondary structure (Figure S5H), accelerating^68–70 or decelerating^71,72 transcription, influencing nucleosome occupancy,^58,73,74 localizing introns to different nuclear regions,^75,76 or impacting genome conformation.^67,77 Although future studies will be required to discern the primary impacts of GC content, our findings imply that the larger genomic context of an intron modulates splicing timing.

Our work raises important questions concerning the mechanisms of SS recognition in humans. Consistent with evidence for exon definition in humans,^3,5,11,13 our analysis indicates that an average intron undergoes lariat formation after transcription of a downstream 5′ SS because with a median global splicing half-life of ≤2.5 kb (Figures 3K, 3L, and 4E), a median exon length of ~150 bps, and a majority of introns splicing with kinetics that approximate exponential decay (Figure 3J), we estimate that for an average intron as much as 96% of splicing events can occur after the downstream 5′ SS is transcribed. Furthermore, in the tail length models (Figures S6D and S6E), the predictions are modestly influenced by downstream 5′ SS sequence and position. However, many introns also undergo lariat formation before transcription of a downstream 5′ SS (Figures 6A, 6B, and S7A–S7C), which is inconsistent with downstream 5′ SS-dependent exon definition and is consistent with recent reports of exon ligation before transcription of a downstream 5′ SS in mammalian cells.^12,13 Strikingly, lariats formed when RNAPII was in the exon for both short (<300 nts) and long (>10,000 nts) introns (Figures 6C and S7E), which is unexpected because inefficient SS recognition across long introns is thought to necessitate exon definition.^3,5 However, although an important feature, intron length is just one of multiple features predictive of splicing timing (Figures 4A, 4B, 5D, 5E, S4C, S4D, S6D, and S6E). We envision several mechanisms by which long introns could splice early. First, in a variation of exon definition, splicing enhancer-binding factors could aid 3′ SS recognition in lieu of U1 snRNP binding to the downstream 5′ SS.^78,79 Second, upstream exon tethering mediated by U1 snRNP bound to transcribing RNAPII could mitigate any penalty for long introns and obviate the need for U1 snRNP bound to the downstream 5′ SS.^80–83 Notably, however, the contribution of intron length to predicting splicing timing implies that exon tethering is insufficient to overcome a penalty for longer introns (Figures 4A and 4D). Finally, because early splicing introns are characterized by a strong PPT and strong U2AF binding (Figures 5F–5H), we suggest that at some introns, U2AF2 binds to a PPT in a manner in dependent of either a downstream or upstream 5′ SS—that is, by neither canonical exon definition nor canonical intron definition. Indeed, U2AF2 along with cofactors can bind in extract to a substrate that includes a BP, PPT, and 3′ SS AG but lacks any 5′ SS (e.g., Mackereth et al.⁸⁴ and Maul-Newby et al.⁸⁵). Furthermore, in introns with strong PPTs, ultra-early lariat formation occurred before the 3′ SS has emerged from RNAPII (Figure 7B), consistent with previous biochemical studies that introns with strong PPTs can undergo lariat formation independently of the 3′ SS AG and U2AF1,^62,63,86 suggesting independence of both a 5′ SS and 3′ SS. Additionally, the association of U2AF2 with RNAPII⁸⁷ could allow rapid and efficient PPT recognition. Together, these findings imply that U2AF2 can drive early splicing. Reconciling evidence for early splicing and exon definition will be an important goal for the future.

Limitations of the study

Several limitations of this study warrant further investigation. First, our estimate of median splicing timing does not account for splicing not captured by CoLa-seq: post-transcriptional splicing, splicing events that did not pass stringent thresholds for BP identification (Figures 1D and S2C), or co-transcriptional splicing of poorly expressed genes, although features do not differ between introns captured or not by CoLa-seq (Figure S2L). Second, although splicing might have the potential to occur during chromatin isolation, from cell lysis onward, we excluded MgCl₂ and included EDTA, an effective inhibitor of splicing catalysis^88–92 and ATP-dependent spliceosome rearrangements. In support of a quenched reaction, our isolation of branched introns supports inhibition of the metal ion-dependent Dbr1.^93,94 Furthermore, we identified NLIs independent of chromatin isolation (Figure S2M). Third, because CoLa-seq reads are biased toward shorter lengths due to sequencing constraints (Figure S3D), we have mitigated such bias in our analyses (e.g., Figures 3K–3L and S3F), focused our analyses on relative comparisons (e.g., Figures 4A, 4B, and 5C–5E), and highlighted features influenced by size biases (e.g., exon length, Figure 4A), but such bias still precludes important quantitative measures, including a quantification of splicing events that occur independent of a classic exon definition model. Such bias also all but precludes a kinetic analysis of the second chemical step of splicing, potentially reflected in the disappearance of NLIs. Additionally, we cannot rule out cases of complex splicing kinetics, such as lag phases, although most introns approximate simple first-order kinetics (Figure 3J). Fourth, although our models identify features that account for a significant fraction of splicing timing variation, much of the variation remains unaccounted for, requiring the inclusion of additional features. Fifth, in principle, the variation of NLI steady-state levels and tail lengths could be sensitive to processes beyond the first chemical step of splicing. Although features expected to impact the second step of splicing help predict the partitioning between concurrent and out-of-order splicing (Figure S5A), the predictions for % in-order splicing and in-order NLI tail length do not appear sensitive to such effects. Similarly, spliceosome disassembly and RNA decay could also impact NLIs, but these processes do not easily account for the predictive features, and an impact of decay is not supported by the coherent 3′ ends of CoLa-seq reads at 3′ SSs (Figure 1C) and at 5′ SSs (Figure S3G). We envision that future experiments will address these limitations.

STAR★METHODS

RESOURCE AVAILABILITY

Lead contact

Further information and requests for resources and reagents should be directed to and will be fulfilled by the lead contact, Jonathan P. Staley (jstaley@uchicago.edu).

Materials availability

This study did not generate new unique reagents.

Data and code availability

• The sequencing data generated during this study are available at the NCBI Sequence Read Archive database under GSE204834. Gel images for western and PCR have been deposited at Mendeley data.

• This paper does not report original code.

• Any additional information required to reanalyze the data reported in this paper is available from the lead contact upon request.

EXPERIMENTAL MODEL AND SUBJECT DETAILS

Cell culture conditions

Human K562 cells were a generous gift from Alex Ruthenburg lab at The University of Chicago. K562 cells were maintained with RPMI + GlutaMAX (Life Technologies) supplemented with 10% (v/v) fetal bovine serum (Life Technologies). Cells were cultured at 37°C with 5% CO₂ and were tested to be negative for mycoplasma using VENORGEM mycoplasma detection kit (Sigma).

METHOD DETAILS

Chromatin-associated RNA isolation

Chromatin RNA was isolated from K562 cells similar to previously described methods,^39–41 with modifications noted below. All the steps were performed with pre-chilled, ice-cold, freshly prepared RNase-free buffers using RNase-free equipment on ice or at 4°C. Roughly 5% of subcellular fractions were saved for quality control (western and RNA fragment size analysis). Fifty million K562 cells were spun down at 100 x g for 3 minutes and washed twice with 15 mL of ice-cold 1X PBS (phosphate-buffered saline) and once with 1 mL of ice-cold 1X PBS. Washed cells were then gently resuspended in 1mL of 0.1% (v/v) Triton X-100-containing buffer A (10 mM HEPES (pH 7.5), 10 mM KCl, 10% glycerol, 0.116 g/mL sucrose, 1 mM DTT, 1x protease inhibitor, 2 mM EDTA) and lysed on ice for 9 minutes. After incubation, the lysed cells were centrifuged at 800 x g for 5 minutes at 4°C. The supernatant corresponding to the cytoplasmic fraction was carefully removed. The remaining nuclei pellet was gently resuspended in 250 μL of NRB buffer (20 mM HEPES (pH 7.5), 75 mM NaCl, 50% glycerol, 0.5 mM EDTA, 1 mM DTT, 1x protease inhibitors). Then, 250 μL of NUN buffer (50 mM HEPES (pH 7.5), 300 mM NaCl, 0.2 mM EDTA, 1 M urea, 1 mM DTT, 1% NP-40) was slowly added and mixed by inversion. The resulting nuclei suspension was incubated on ice for 5 minutes and centrifuged at 800 x g for 5 minutes at 4°C. The supernatant corresponding to the nucleoplasmic fraction was carefully removed. The chromatin pellet was washed once with 500 μL of premixed NRB/NUN buffer (1:1) and centrifuged at 800 x g for 5 minutes at 4°C. The supernatant was carefully removed. The washed chromatin pellet was gently resuspended in 200 mL of NRB buffer. RNA from whole cells, the cytoplasmic fraction, the nucleoplasmic fraction, and the chromatin pellet were extracted by Trizol (Thermo Fisher Scientific) followed by a second extraction with phenol:chloroform:isoamyl alcohol (25:24:1) and isopropanol precipitation. The isolated RNA was furthered treated with Turbo DNase (Thermo Fisher Scientific) to remove residual genomic DNA, followed by extraction with phenol:chloroform:isoamyl alcohol (25:24:1).

Confirmation of fractionation

The efficiency of fractionation was tested by western blotting of the cytosolic fraction, the nucleoplasmic fraction, and the chromatin fraction with appropriate antibodies. The antibodies used were GAPDH (Santa Cruz Biotechnology), RNA Pol II (Ser-5 phosphorylated, MBL International), and Histone H3 (Abcam).

Enrichment for lariat RNA species

For each CoLa-seq library, chromatin-associated RNA from 200 million cells was used. To enrich for lariat RNA species (nascent lariat intermediates and excised lariat introns) from the chromatin-associated RNA, RNAs less than roughly 200 nts were removed using SPRI paramagnetic beads (Beckman Coulter) with a beads-to-sample ratio of 1:1.2. Note that this step also removes excised lariat introns derived from short introns and nascent lariat intermediates derived from short introns with short tails (i.e., short introns that form early during in-order splicing), which likely leads to underrepresentation of ELIs with short BP-3′ SS distances and impact the estimate of BP-3′ SS distance distribution based solely on ELI reads (see Figure 3E). Next, ribosomal RNA was depleted using oligo based Ribo-Zero Magnetic Gold Kit (Illumina) according to the manufacturer’s manual. During the optimization of CoLa-seq, we noticed abundant ncRNA reads in our libraries, similar to previous observations in mNET-seq.¹¹¹ To overcome this issue, biotinylated anti-sense oligos were designed to specifically target and deplete ncRNAs that were found to be abundant in our initial libraries (Table S4). Briefly, for every μg of rRNA-depleted RNA, a depletion reaction was assembled, containing 82.5 pmol of total depletion oligo mix (Table S4), 2X SSC, and 5 mM EDTA. The depletion reaction was put through an oligo annealing cycle (5 minutes at 75°C, cool down to 22°C at 0.1 °C/second), after which ncRNA bound biotinylated depletion oligos were pulled down using Dynabeads^™ MyOne^™ Streptavidin C1 beads (Thermo Fisher Scientific), according to the manufacturer’s manual. The RNA in the supernatant was purified using the RNA Clean and Concentrator kit (Zymo Research) and excess anti-sense oligos were removed by treatment with Turbo DNase (Thermo Fisher Scientific). Finally, according to the manufacturer’s instructions, purified RNA was treated with RppH (NEB) to decap linear RNA and then was degraded using Terminator 5′−3′ exonuclease (Lucigen). The resulting RNA was put through a size selection step to remove ≤200 nts RNA using SPRI paramagnetic beads (Beckman Coulter) with a beads to sample ratio of 1:1.2. The final RNA quantity is about 0.5–1% of the starting material.

CoLa-seq library prep

Once the chromatin RNA was enriched for lariat species, ~500 ng of RNA was used for CoLa-seq library prep. The 3′ ends of RNA were ligated to a pre-adenylated 3′ end adapter (Table S5), which includes a 6 nt UMI (unique molecule identifier) at the 5′ end to allow correction for PCR duplication and RT mis-priming. The ligation reaction contained 17.5% PEG-8000 (w/v), 1X T4 RNA ligase buffer (NEB), 1 μM of the pre-adenylated 3′ end adaptor, and 10 units/μL T4 RNA Ligase 2 (truncated K227Q, NEB). The ligation reaction was incubated for 3 hours at 22°C. After the reaction excess adapters were removed by size selection using SPRI paramagnetic beads (Beckman Coulter) with a beads to sample ratio of 1.6:1. Then, the adapter-ligated RNA was used for cDNA synthesis using 0.25 μM of RT primer (Table S5) and 10 units/μL of Superscript III (Thermo Fisher Scientific) at 50°C for 1 hour, according to the manufacturer’s manual. Afterwards, to degrade excessive RT primer, 3.5 μL of ExoSAP-IT was added to the cDNA synthesis reaction and incubated for another 30 minutes at 37°C. The reaction was stopped by adding 10 μL of 0.5M EDTA and RNA was degraded by adding 10 μL of 1 N NaOH. The resulting cDNA was cleaned up using Oligo Clean and Concentrator kit (Zymo Research) and was circularized with Circligase (Lucigen), according to the manufacturer’s manual. Final library amplification was performed on the circularized cDNA using Phusion polymerase (Thermo Fisher Scientific) with primers that add sequencing adaptor sequences and specific sequencing indices, which are located in the reverse primers (Table S5). The number of PCR cycles was between 15 and 18 cycles. The amplified product was run on a 6% TBE gel and DNA fragments between 175 and 750 bps were gel purified and 3 nM of the library was sequenced in a 100 bp paired-end mode on an Illumina HiSeq 4000 (Genomics Facility at University of Chicago). Reads were pre-processed and analyzed as described below. Note that while the sequence reads have well balanced GC content, we could not completely exclude the possibility that GC content of the library impacts the PCR or Illumina sequencing.

Sequencing of chromatin-associated RNA

For each library, 10 μg of chromatin-associated RNAs (extracted as described above) were depleted of ribosomal RNA depletion using Ribo-Zero Magnetic Gold Kit (human/mouse/rat; Illumina), according to the manufacturer’s instruction. Abundant ncRNAs were depleted as described in the section of enrichment for lariat RNA species (see above). Then, 100 ng rRNA/ncRNA-depleted RNA was used to prepare an RNA-seq library using NEBNext Ultra II™ Directional RNA Library Prep Kit for Illumina, according to the manufacturer’s instruction. The final library was sequenced in a 100 bp paired-end mode on an Illumina HiSeq 4000 (Genomics Facility at University of Chicago). Nascent RNA-seq data were pre-processed as follows: (1) adaptors were trimmed using Cutadapt,⁹⁶ with default parameters except discarding reads shorter than 15 nts, (2) remaining reads were aligned to the human reference genome (hg38) with GENCODE transcriptome annotation (v26) using the STAR software package⁹⁸ with the following settings: –out- SAMattributes NH HI AS nM NM MD jM jI XS, –out-FilterMultimapNmax 50, –outFilterMismatchNmax 15, –outSJ-filterDistToOtherSJ- min 10 0 5 10 –alignIntronMin 15, –outSAMtype BAM SortedByCoordinate, –twopassMode Basic, (3) aligned reads were further filtered for unique mapping and proper alignment using samtools.⁹⁹ The resulting reads were used for downstream analyses. Separately, after the adapter removal step, reads were used for gene expression quantitation using Kallisto with default settings.¹⁰¹

Lariat sequencing

Lariat-seq was used to identify BPs via inverted reads that traverse through the branch in the lariat RNAs and performed similarly to previously described methods^36,112 with modifications. Briefly, chromatin-associated RNA was first isolated as above to capture both lariat intermediates and excised lariat introns. Then, 10 μg chromatin-associated RNAs were depleted of ribosomal RNA depletion using Ribo-Zero Magnetic Gold Kit (human/mouse/rat; Illumina), according to the manufacturer’s instruction. To further enrich lariat RNA species, RNase R (Lucigen) digestion was performed to degrade linear RNA, according to the manufacturer’s instruction; treatment also trims tails of lariat RNA species. The resulting RNA was used to prepare an RNA-seq library using NEBNext Ultra II™ Directional RNA Library Prep Kit from Illumina, according to the manufacturer’s instruction except for the RT step, during which 1 mM MnCl₂ was included to facilitate reverse transcriptase to go through the BP. The final library was sequenced in a 100 bp paired-end mode on an Illumina Hiseq 4000 (Genomics Facility at University of Chicago). The resulting reads from three biological replicates were combined and used to identify BPs according to previously described methods.^32,36

Splicing inhibition and transcription inhibition

K562 cells at 90% confluency were treated for 2 hours with one of the following reagents (1) pladienolide B (Santa Cruz Biotechnology) in DMSO to a final concentration of 10 μM or (2) flavopiridol (Sigma) in DMSO to a final concentration of 1 μM. A control treatment was also performed with an equal volume of DMSO. Splicing inhibition was confirmed by RT-PCR to detect previously observed splicing changes¹¹³; primers used are listed in Table S6. Chromatin fractionation and CoLa-Seq libraries were prepared from the DMSO or drug treated cells as described above.

Direct RNA pull-down and sequencing

Total RNA were extracted from K562 cells using Trizol (Thermo Fisher Scientific) followed by a second extraction with phenol:chloroform:isoamyl alcohol (25:24:1) and isopropanol precipitation. The isolated total RNA was furthered treated with Turbo DNase (Thermo Fisher Scientific) to remove residual genomic DNA, followed by extraction with phenol:chloroform:isoamyl alcohol (25:24:1). Then, 1 mg of total RNA was incubated with three sets of biotinylated anti-sense oligos (ASOs), which target intron 16 of HNRNPK, intron 1 of EIF, and intron 6 of PPP2CA (1 μM per oligo; Table S7), in a solution containing 2x SSC, 5 mM EDTA, and 10% Formamide. The resulting hybridization reaction was put through an oligo annealing cycle (5 minutes at 80°C, cool down to 22°C at 0.1 °C/second), after which RNA bound biotinylated ASOs were pulled down using Dynabeads^™ MyOne^™ Streptavidin C1 beads (Thermo Fisher Scientific), according to the manufacturer’s manual. The ASOs were digested with Turbo DNase and the pulled-down RNA was extracted using Trizol and precipitated by ethanol. The resulting RNA was used for library prep as described in the CoLa-seq library prep section.

Data pre-processing of CoLa-seq data

CoLa-seq data were processed as follows: (1) 6 nt UMIs were trimmed from reads and attached to read names to maintain their association with reads using fastp,⁹⁷ (2) adaptors were trimmed using Cutadapt. Reads shorter than 15 nts were discarded, (3) the remaining reads were aligned to the human reference genome (hg38) with GENCODE transcriptome annotation (v26) using the STAR software⁹⁸ with the following settings: –outSAMattributes NH HI AS nM NM MD jM jI XS –outFilterMultimapN-max 50 –outFilterMismatchNmax 15 –outSJfilterDistToOtherSJmin 10 0 5 10 –alignIntronMin 15 –outFilterScoreMinOverLread 0 –outFilterMatchNminOverLread 0 –outFilterMatchNmin 15 –outSAMtype BAM SortedByCoordinate –twopassMode Basic, (4) aligned read pairs were further filtered for unique mapping using samtools, (5) to remove RT mis-primed reads, raw reads were separately mapped without UMI trimming. If the UMI portion of a read was mapped to the genome without any mismatch, it is considered as a mis-primed read and removed from the reads that were mapped with UMI trimming. These RT mis-primed reads were removed using samtools and pysam, (6) PCR duplicates were removed using UMI-tools,¹⁰² (7) reads overlapping with ncRNAs (snRNAs, snoRNAs, miRNAs, rRNAs) were removed using samtools based on the GENCODE reference (v26). The remaining reads were used for downstream analyses.

Analysis of CoLa-seq read ends around branch points, 5′ splice sites, or 3′ splice sites

The 5′’ ends of reads were intersected with 21-nt regions centered on previously annotated BPs, whereas read 3′ ends were intersected with 21-nt regions centered on either constitutive 5′ SSs or constitutive 3′ SSs (see below). To minimize expression differences, reads numbers across each region for each intron were normalized such that the total read count ranged from 0 to 1. Then, regions were aligned with their respective feature (BP, 5′ SS, or 3′ SS) in the center, and the mean, normalized signal was calculated at each position.

Sequence logo of branch point motif

Using Bioconductor tools and ggseqlogo package in R,^114,115 the sequence logo was generated as follows: (1) for each BP, an 11 nt long sequence (with the BP in the middle) was extracted, (2) the resulting sequence set was used to calculate the position weight matrix and then to plot the sequence logo for the BP motif.

Branch point calling algorithm

Putative ELI and NLI reads were first identified as sequenced fragments for which their 5′ ends mapped to within 100 nts of an annotated 3′ SS. Next, putative ELI reads were further defined as sequenced fragments for which their 3′ end maps to an annotated 3′ SS (Gencode v29), and putative NLI reads were further defined as extending without gaps past an annotated 3′ SS. Note that the filter for putative ELI reads is more stringent than putative NLI reads in that it requires a putative ELI read to perfectly align with a 3′ SS, whereas a putative NLI read only needs to cross a 3′ SS. To finally filter for authentic ELI and NLI fragments, we aimed to require that their 5′ ends align with a BP. Because published databases of known BPs³² are relatively incomplete, we built a probabilistic classifier to expand the set of BP positions, from which NLI reads are anchored. Specifically, each position in the 5 to 100 nt region upstream of annotated 3′ SSs is assigned a probability that it is a BP based on multiple features and a naive Bayes assumption of independence. The probability distribution of features over the non-BPs is assumed uniform and therefore a constant that can be ignored at all positions. Therefore, at each position, $X$, we can assign an unscaled posterior probability that the position is a BP:

$$Pr\left(C_{bp}\right)\propto \prod _{i=1}^{n}Pr\left(x_i\mid C_{bp}\right)$$

where $Pr\left(C_{bp}\right)$ is the posterior probability of classification as a BP $C_{bp}$ (as opposed to non-BP $C_n$) at position $X$, and $x_i$ is the value of feature $i$ at position $X$. The features considered are the pentamer motif (−3 to +1 relative to position $X$), the distance to the annotated 3′ SS, and the relative coverage of CoLa-seq 5′ read ends. The probability distribution of pentamers in $C_{bp}$ was trained on BPs empirically identified previously.³² Specifically, all U2-recognized, high-quality BPs were used to generate a position weight matrix, which is weighted by the read count supporting each BP in Pineda and Bradley,³² to assign BP probabilities for each pentamer. Similarly, the empirical distance from a U2-recognized, high-quality BP to the annotated 3′ SS (across all BPs) was used to create a smoothed probability distribution using the density function in R with bw=3 bandwidth adjustment. Feature training and classification were repeated separately for introns annotated as U12 introns.¹¹⁶ To determine a BP probability from relative coverage of CoLa-seq 5′ ends ($Pr\left(x_i\mid C_{bp}\right)$) where $x_i$ is the relative coverage of CoLa-seq 5′ ends at known BPs, $C_{bp}$, we did not consider the empirical probability of relative coverage from known BPs, but, rather, we devised an alternate probability function as follows: CoLa-seq reads from all experiments were combined and filtered for fragments that end at (putative ELI) or cross (putative NLI) an annotated 3′ SS. For each 3′ SS region with at least 10 ELI and/or putative NLI fragments, read coverage of the 5′ position of fragments was normalized to the total coverage in the region after adding a 0.1 pseudo-count to all positions. Then, a smoothed probability distribution was created with the density function in R with a bandwidth setting of bw=0.5. Smoothing was introduced as a way to emulate the natural imprecision of reverse transcription termination, as we have observed that though reverse transcriptase usually stops one nucleotide downstream of the BP, reverse transcriptase sometimes terminates one nucleotide further downstream or one nucleotide upstream (i.e., at the BP; Figure S2A). Following the evidence that putative ELI and putative NLI CoLa-seq reads most often terminate at the base downstream of the BP (Figure S2A), the probability distribution was shifted upstream by one base such that the center of smoothed peaks corresponds to more likely BPs. Finally, to classify bases as belonging to a BP region, we defined a threshold

$$\widehat{y}=C_{bp}\text{if}\prod _{i=1}^n\mathrm{Pr}\left(x_i\mid C_{bp}\right)>\text{threshold;}{C}_n\text{otherwise}$$

Or equivalently:

$$\widehat{y}=C_{bp}\text{if}\sum _{i=1}^n\log \left(\mathrm{Pr}\left(x_i\mid C_{bp}\right)\right)>\text{threshold;}{C}_n\text{otherwise}$$

The threshold was chosen based on ROC analysis using a validation set of BP positions identified in both Pineda and Bradley,³² as well as lariat-seq from K562 (this study; see above; Table S1), as the true response. We chose a threshold that yields a false positive rate of 0.01 in the ROC analysis. Finally, after applying this classifier to all 114,540 3’ SS regions with >10 ELI/LI reads, we merged BPs with one or zero bases in between each other into a single most likely BP. In total, we identified 164,532 BPs amongst 109,256 U2- type 3’ SSs and 750 BPs amongst 583 U12-type 3’ SSs. The novelty of identified BPs was determined by comparing to previously annotated BPs.^32–34,36

Identification of nascent lariat intermediates and excised lariat introns

Exons from the GENCODE transcriptome annotation (v26) that we also identified in chromatin-associated RNA-seq and total RNA-seq were used for identifying NLIs and ELIs. Exons whose ends can be used as alternative 3′ SSs were removed to avoid ambiguity. To identify reads derived from NLIs, the following filtering steps were performed using a custom snakemake¹⁰³ pipeline wrapped around bedtools,¹⁰⁰ samtools, and Picard tools (https://broadinstitute.github.io/picard/): (1) reads were defined as putative NLI reads, as noted above, if they overlapped with an annotated 3′ SS and their 5′ ends were within 100 nts upstream of 3′ SSs; and (2) using BPs annotated above, putative NLI reads were designated as genuine NLI reads if their 5’ ends corresponded to the BP or BP+1 position. In total, we detected 2,201,036 NLI reads associated with 73,562 3′ SSs (≥ 3 NLI reads) (Figure 3C), more than 20 unique NLI reads in 25,476 introns (Figure 3C), and NLI reads in every intron of the major isoform for 2,941 genes. To identify reads derived from ELIs, the following filtering steps were performed using a custom snakemake pipeline wrapped around bedtools, samtools, and Picard tools: (1) reads were defined as putative ELI reads, as noted above, if their 3’ ends corresponded to the last nt of an intron, and their 5’ ends were within 100 nts upstream of 3’ SSs; and (2) using BPs annotated above, putative ELI reads were designated as genuine ELI reads if their 5’ ends corresponded to the BP or BP+1 position. In total, we detected 2,574,753 ELI reads associated with 57,309 3′ SSs (≥ 3 ELI reads) (Figure 2C), more than 20 unique ELI reads in 23,272 introns (Figure 2C), and ELI reads in every intron of the major isoform for 1,432 genes.

Identification of constitutive and alternative splice sites

Using Regtools¹⁰⁴ and LeafCutter,¹⁰⁵ 5′ SSs and 3′ SSs were extracted from chromatin-associated RNA-seq data (this study), nuclear RNA-seq data from ENCODE, as well as total RNA-seq data from ENCODE. Splice sites were considered as constitutive sites when their usage percentages were higher than 95%.

Quantification of branch point usage

Branch point usage was quantified as described in Mercer et al.³⁶ Briefly, ELI read (or NLI read) counts of a BP were divided by ELI read (or NLI read) counts of all BPs associated with a 3′ SS.

Quantification of usage of coupled branch point and 3′ splice sites

Alternative 3′ SSs using the same 5′ SSs were paired together, and then the nucleotide distances between the alternative 3′ SSs were calculated and used to sort these pairs into different groups. If both 3’ SSs within a pair use the same set of BPs, the coupling index of this pair was set to 1, whereas if they used different ones, the coupling index was set to 0. The coupling fraction of each group was then calculated by taking the average of the coupling index for all 3′ SS pairs in the group. NAGNAG sites were extracted from GENCODE transcriptome annotation (v26) and supplemented with NAGNAG sites from Bradley et al.¹¹⁷

Classification of in-order, concurrent, and out-of-order splicing

To further classify NLI reads into classes of in-order, concurrent, and out-of-order splicing, NLI reads that ended at the last nt of an annotated exon were classified as concurrent splicing. NLI reads that were split reads (N in the CIGAR string) were considered as out-of-order splicing. NLI reads that did not belong to the other two classes and were shorter than 650 nts (upper insert size limit of the library) were considered as in-order splicing; note that for intron pairs in which the downstream intron is short, <600 nts, the likelihood of concurrent splicing and out-of-order splicing increases, and consequently the likelihood of capturing in-order NLIs with longer tails decreases.

Splicing timing simulation

The level of in-order splicing of an intron relative to its downstream intron strongly correlates with the distance from its 3′ SS to the 3′ SS of the downstream intron (US-DS 3′ SS distance; Figure 3K). Importantly, the median level of in-order splicing reaches over 75% when the US-DS 3′ SS distance exceeds 1500 nts. We thus reasoned that the timing of splicing must be generally faster than the transcription time of a 1500-nt intron. To derive an estimate of the median timing of splicing for all introns, we set out to simulate the median proportion of in-order splicing for adjacent introns as a function of the US-DS 3′ SS distance ($ds$, in nts). To this end, we sampled splicing timing for an intron (1) and its downstream intron (2) from distributions with a range of possible median splicing timingshalf-lives, ranging from 0 to 5000 nt. More specifically, for a specific overall median splicing half-life $\mu \left(t_{1/2}\right)$ (i.e., the timing of lariat formation), we draw a median splicing timing half-life $t_{1/2\text{\_}1}$ and $t_{1/2\text{\_}2}$ for each intron 1 and 2 from a lognormal distribution $\log \text{N}\left(\mu \left(t_{1/2}\right),{\sigma }^2\right)$ with several possible log-scaled variance parameters of ${\sigma }^2$ that yield a reasonable distribution of splicing timing (Figure S3M). Next, for each median splicing timing half-lives ($t_{1/2\text{\_}1}$ and $t_{1/2\text{\_}2}$) and distance ds, we drew 1,000 samples each from ${\text{T}}_1\sim expon\left(\ln (2)/t_{1/2}1\right)$ and ${\text{T}}_2\sim expon\left(\ln (2)/t_{1/22}\right)$, where ${\text{T}}_1$ and ${\text{T}}_2$ represent individual splicing events, resulting in 1,000 pairs of simulated splicing timings $\left({\text{t}}_{1,1},\dots ,{\text{t}}_{1,1000}\right)$ and $\left({\text{t}}_{2,1},\dots ,{\text{t}}_{2,1000}\right)$. To calculate the number of in-order splicing events, we counted the number of intron pairs such that ${\text{t}}_{1,\text{i}}<{\text{t}}_{2,\text{i}}+ds+{\text{t}}_{\text{s}}$, as this formula indicates that the timing of splicing of intron $1\left({\text{t}}_{1,\text{i}}\right)$ is faster than the timing of splicing b of intron $2\left({\text{t}}_{2,\text{i}}\right)$ plus $ds$, which corresponds the delay in splicing of intron 2 imposed by the time to transcribe the exon, and the downstream intron, and plus${\text{t}}_{\text{s}}$, which is fixed to 50 nts to roughly estimate the time of the second step of the downstream intron as it must have occurred to observe an out-of-order NLI. To calculate the number of out-of-order splicing events, we counted the number of intron pairs such that ${\text{t}}_{1,I}>{\text{t}}_{2,\text{i}}+ds+{\text{t}}_{\text{s}}$. To calculate the proportion of in-order splicing, we divided the number of in-order splicing events by the sum of the in-order and out-of-order splicing events. Of note, the conclusion from our simulations did not change qualitatively when using reasonable variance parameters (${\sigma }^2$) or different values (0 to 200 nts) for the timing of the second step of splicing (${\text{t}}_{\text{s}}$).

To compare the empirical proportion of in-order splicing, we binned intron pairs according to the US-DS 3′ SS distance in 750 nt bins, from 500 nts to 10k nts. We then plotted the median proportion of in-order splicing of all intron pairs in each bin with at least 3 informative NLI reads. We then performed simulations as outlined above by using $ds$ as the midpoint of each bin and by varying the overall median splicing half-life. When we plotted the simulated data along with the empirical data for different variance parameters or different choices of splicing timing distributions; we found that the median splicing half-life that best matched the empirical data was less than 2.5 kb.

To reduce the impact of read size bias on the empirical and simulated data, we calculated the proportion of in-order splicing using only in-order and out-of-order NLI reads that overlap with the downstream 5′ SSs, which resulted in reads with similar length distributions in the empirical data (Figures S3E and S3F). Moreover, in the simulations, we restricted the calculation of the proportion of in-order splicing to intron pairs (${\text{x}}_{a,\text{i}}$,${\text{x}}_{\text{b},\text{i}}$) such that ${\text{x}}_{a,\text{i}}$ or ${\text{x}}_{\text{b},\text{i}}$ are both less than 500 nts, given similar length constraints on the empirical CoLa-seq data.

Sequence feature generation

For exons and their neighboring introns, sequence features such as length and nucleotide compositions (%A, %G, %C, %U, and %GC) were extracted and calculated using Bioconductor tools within R. For 5′ and 3’ SSs, their scores were calculated using MaxEntscan algorithm,¹⁰⁷ and a 50-nt window surrounding each site was used to calculate nucleotide compositions and RNA secondary structure potentials (using viennaRNA¹⁰⁶). The polypyrimidine tract (PPT) score is calculated using the script from Taggart et al.³³ Binding energy was calculated using viennaRNA for U1 snRNA and the 5′ SS, U6 snRNA and the 5′ SS, and U2 snRNA and the BP. For BPs annotated by CoLa-seq, the following features were generated: BP position, the percent usage of a BP relative to alternative BPs, BP-3’ SS distance, and nucleotide composition and RNA secondary structure around BPs and in regions between the BP and 3’ SS. NLI read counts were also included in the 3′ SS-RNAP II distance models, because introns with greater numbers of NLI reads are more likely to sample earlier splicing events.

XGBoost models to identify features that predict the order of splicing and the timing of in-order splicing

For the splicing order models, 23,324 introns pairs containing at least 10 CoLa-seq reads were used, and % in-order splicing, % con- current splicing, and % out-of-order splicing were used as target variables. For the NLI 3′ SS-RNAP II distance models, 14,437 introns containing at least 10 in-order NLI reads were used and the minimum 3′ SS-RNAP II distance and the median 3′ SS-RNAP II distance were used as target variables. For each model, introns were randomly split into a training set and a test set in a 7:3 ratio. In the training set, when two features were more than 80% correlated (Pearson correlation), one of them was randomly removed. Afterwards, 22 fea- tures were used for the splicing-order models (Table S3) and 37 features were selected for the 3′ SS-RNAP II distance models. XGBoost⁵³ was used to train each model. The basic parameters for all models were as follows: objective=reg:squarederror, boos- ter=gbtree, predictor=gpu_predictor, tree_method=gpu_hist. The hyperparameters for models were tuned using Hyperopt.¹¹⁸ The resulting hyper-parameters for the splicing-order models were as follows: max_depth=5, learning_rate=0.01, subsample=0.6, col- sample_bynode=0.6, reg_alpha=0.5, reg_lambda=0.8, random_state=123, min_child_weight=60, n_estimators=600. The resulting hyper-parameters for the 3′ SS-RNAP II distance models were based on the default settings with the following changes: max_depth=5, learning_rate=0.01, subsample=0.6, reg_alpha=0.5, reg_lambda=0.8, random_state=123, min_child_weight = 200, n_estima- tors=550. The test set was used for model performance evaluation (metric: R-squared). Furthermore, 5-fold cross-validation was per- formed on the training set to derive the final averaged model performance as well as the variance of the performance (metric: R-squared). To evaluate the contribution of each feature within each model, SHAP values were computed and visualized using the SHAP package in python.^54,119 In the 3′ SS-RNAP II models, upstream intron length is modeled but not considered further because the CoLa-seq protocol discriminates against in-order nascent lariat intermediates that are derived from short introns and have short tails, thereby distorting the distribution of in-order splicing tail lengths (see enrichment for lariat RNA species). Despite the biased depletion of short introns with short tail lengths, the remaining short introns with longer tail lengths do reveal a significant impact of short intron length in favoring early, in-order splicing over concurrent splicing or out-of-order splicing (Figure 4).Further, downstream intron length is modeled in the 3′ SS-RNAP II models but not considered further, because shorter downstream intron length correlates with shorter in-order NLI tails; this correlation emerges simply because when the downstream intron is short, in-order NLI tails would position RNAP II downstream of the downstream 5′ SS and thus directly compete with concurrent splicing and out-of-order splicing and deplete longer NLI tails. Moreover, because a lower length cut off (175 nts) used in the gel purification of the final PCR step likely leads to under capture of NLIs with very short tails, the fact that longer upstream BS-3′ SS distances predicts shorter NLI tails (Figure 5E) likely reflects that short NLI tails derived from introns with short BP-3′ SS distances are largely excluded from the library prep.

GC content calculation across splice sites for introns with different levels of in-order splicing

Intron pairs with different levels of in-order splicing were divided to three groups: low (≤20% in-order splicing), mid (40–60% in-order splicing), and high (≥80% in-order splicing). To calculate GC content around splice sites for each intron pair, 50 nts flanking each splice site were extracted. Sequences corresponding to splice site signals were not included in GC content calculation; specifically, for 5’ SSs, the last 3 nts in the exons and first 6 nts in the introns were removed, and for 3′ SSs, the last 3 nts in the introns and first 2 nts in the exons were removed. GC content was calculated at each nucleotide position with a 10 nt sliding window. Within each group, the average GC content and 95% confidence intervals were calculated at each position and plotted across splice sites of both upstream and downstream introns.

Analysis of eCLIP-seq data in K562 cells from ENCODE

Bam files of input and IP for each RBP were downloaded from the ENCODE project website (https://www.encodeproject.org/).¹²⁰ For each tested RBP (Figures 4C and 5H), eCLIP-signals were processed using RBP-Maps¹¹⁰ with default settings for introns with low in-order splicing (≤20%), introns with high in-order splicing (≥80%), early in-order introns (≥70% of 3′ SS-RNAP II distances ≤50 nts, n=1,712), and late in-order introns (≥70% of 3′ SS-RNAP II distances ≥200 nts, n=2,735). The resulting signals were further processed in R for visualization.

De novo motif identification and motif enrichment analysis

To identify de novo motifs in specific regions of early splicing and late splicing introns, meme from MEME SUITE^108,121 was used with a Markov model generated from the same regions of all introns containing at least 10 NLI reads. The identified motifs were further processed in R using universalmotif¹²² and ggseqlogo for making sequence logos. The identified motifs were also fed into tomtom from MEME SUITE to search for enriched known motifs of RNA binding protein against the human CISBP-RNA database.⁹⁵

Hexamer analysis

To identify enriched and depleted hexamers in specific regions of early splicing and late splicing introns, k-mer analysis was performed using transite package in R.¹⁰⁹

QUANTIFICATION AND STATISTICAL ANALYSIS

All quantification and statistical analyses were done in R and python. Analysis details can be found in figure legends and the result sections. All plots were prepared using data.table,¹²³ ggpubr,¹²⁴ cowplot,¹²⁵ patchwork,¹²⁶ ggrepel,¹²⁷ or tidyverse tools¹²⁸ in R, except for SHAP summary plots, which were made using SHAP^54,119 and matplotlib¹²⁹ in python.

KEY RESOURCES TABLE

Reagent or resource	Source	Identifier
Antibodies
GAPDH Antibody (0411)	Santa Cruz Biotechnology	sc-47724
anti-histone H3	Abcam	ab1791
anti-RNAP II (ser5-p)	MBL International	MABI0603
Chemicals
Fetal Bovine Serum	Life Technologies	26140079
RPMI 1640 Medium, GlutaMAX™	Life Technologies	61870127
10X PBS	Sigma	P7059–1L
flavopiridol	Sigma	F3055–5MG
Pladienolide B	Santa Cruz Biotechnology	sc-391691
cOmplete, EDTA-free Protease Inhibitor Cocktail Tablets	Sigma	11873580001
TRIzol Reagent	Thermo Fisher Scientific	15596018
SSC (20X), RNase-free	Thermo Fisher Scientific	AM9763
Formamide	Thermo Fisher Scientific	AM9344
Critical Commercial Assays
Turbo DNase	Thermo Fisher Scientific	AM2238
RppH	NEB	M0356S
Terminator™ 5´-Phosphate-Dependent Exonuclease	Lucigen	TER51020
T4 RNA Ligase 2 truncated K227Q	NEB	M0351L
Superscript III	Thermo Fisher Scientific	18080044
ExoSAP-IT	Thermo Fisher Scientific	78250.40.UL
Circligase	Lucigen	CL4115K
Phusion polymerase	Thermo Fisher Scientific	F530L
SPRI paramagnetic beads	Beckman Coulter	B23317
Ribo-ZeroTM Gold Kit H/M/R Kit	Illumina	MRZG126
Dynabeads™ MyOne™ Streptavidin C1 beads	Thermo Fisher Scientific	65002
NEBNext Ultra II™ Directional RNA Library Prep Kit	NEB	E7760S
Power SYBR™ Green PCR Master Mix	Thermo Fisher Scientific	4367659
NEBNEXT MULTIPLEX OLIGOS	NEB	E7335S
Venorgem mycoplasma detection kit	Sigma	MP0025–1KT
RNA Clean and Concentrator kit	Zymo Research	11–325
Oligo Clean and Concentrator kit	Zymo Research	11–380
6% TBE GEL 1.0MM 10W	Thermo Fisher Scientific	EC6265BOX
Deposited data
Raw and analyzed data	This study	GEO: GSE204834
Gel images	This study	10.17632/p5mj87pj8w.1
RNA-seq data in K562	ENCODE	www.encodeproject.org
eCLIP-seq data in K562	Van Nostrand et al56	www.encodeproject.org
Branch points from Mercer et al., 2015	Mercer et al36	Supplementary Data 2
Branch points from Pineda & Bradley, 2018	Pineda and Bradley32	Supplemental File S2
Branch points from Taggart et al., 2017	Taggart et al33	Supplemental Table S2
Branch points from Briese et al., 2019	Briese et al34	Supplementary Dataset 2
CISBP-RNA database	Ray et al121	cisbp-rna.ccbr.utoronto.ca
Experimental models: cell lines
Human K562 cells	ATCC	CCL-243
Oligonucleotides
See table S4–7	This study	N/A
Software and algorithms
cutadapt v2.3	Martin96	N/A
fastp v0.20.0	Chen et al102	N/A
STAR v2.6.1b	Dobin et al97	N/A
samtools v1.9	Li et al98	N/A
bedtools v2.27.1	Quinlan and Hall108	N/A
kallisto v0.43.0	Bray et al99	N/A
UMI-tools v0.5.5	Smith et al103	N/A
pysam v0.15.3	github.com/pysam-developers/pysam	N/A
snakemake v5.5.0	Köster and Rahmann107	N/A
Picard toolkit v2.21.8	broadinstitute.github.io/picard	N/A
Regtools v0.5.1	Feng et al109	N/A
LeafCutter	Li et al110	N/A
viennaRNA v2.0	Lorenz et al113	N/A
MaxEntscan algorithm	Yeo and Burge112	github.com/kepbod/maxentpy
PPT score algorithm	Taggart et al33	github.com/allisontaggart/PPT_scoring
XGBoost v1.5.1	Chen and Guestrin53	N/A
SHAP v0.40.0	Lundberg and Lee54	N/A
MEME SUITE v5.0.2	Bailey et al118	N/A
transite v1.0	Krismer et al122	N/A
rbp-maps v0.1.4	Yee et al117	N/A

Highlights.

• CoLa-seq maps cell-type-specific branch points with unprecedented efficiency

• CoLa-seq reveals splicing timing varies dramatically across human introns

• Early splicing occurs independent of exon definition

• Modeling uncovers key genetic determinants of splicing timing, such as U2AF2 binding