Programmed Delayed Splicing: A Mechanism for Timed Inflammatory Gene Expression

Reviewer #1 (Public review):

Summary:

In this work, the authors revisit a well-defined experimental system for studying temporal gene expression mechanisms in TNF-alpha-stimulated macrophages, bringing new tools to the process. Using a hybrid-capture approach, they are able to obtain deeper RNA sequencing of target genes, which allows them to identify potential differences in splicing kinetics of individual introns. Further implementing transcriptional blocks to measure intron half-lives, and predictive machine learning models to identify potential contributing cis-acting RNA elements, they define a group of 'bottleneck' introns whose delayed splicing is a rate-limiting step in mRNA maturation.

Strengths:

(1) The hybrid-capture approach enables deeper RNA sequencing of target transcripts.

(2) The neural network application to identify motifs outside of splice sites could be related to intron removal kinetics.

(3) The paper uses splicing reporters with modulation of 5' splice sites to test the effect on reporter gene expression in the context of 'bottleneck' introns.

Weaknesses:

(1) While evidence is provided that these introns are distinct from previously published splicing kinetics studies, 'bottleneck' introns are not adequately placed in context for assessment of how they are similar or different.

(2) Splicing reporters are a good approach, but the complexities of post-transcriptional gene expression regulation are not adequately addressed

(3) Deep learning models are a potentially powerful tool for identifying novel regulatory sequences; however, their use here is underdeveloped.

Reviewer #2 (Public review):

Summary:

The authors analyzed the temporal dynamics of gene expression patterns within the inflammatory response transcriptome following TNF stimulation, and proposed that the splicing rate of certain introns is a key mechanism of regulating mature mRNA expression rate.

Strengths:

The measurement strategy is generally well-designed to understand the core question of splicing rate and gene expression. The following computation analysis, as well as the mutation or repair studies, further supported the claims. The writing and presentation of the results are also generally clear and easy to follow. I think this manuscript will be of interest to a wide audience.

Weaknesses: 

I do have some questions regarding some of the results and conclusions, and I think either more analysis or more explanation and discussion can make the claims more solid. Please see below for details:

(1) On the hybrid capture method and the RNA coverage results: The strategy of enriching for the last exon before sequencing does have significance in linking pre-mRNA and mature mRNA. If I understand correctly, this enriches for pre-mRNA molecules that are about to finish the full-length elongation of RNA polymerase. However, is this strategy biased towards measuring the splicing rate variation on introns closer to the 3-prime end? For example, if a gene takes 5 minutes for the RNA polymerase to elongate through the full length of the gene, for intron #1 that's very close to the 5' end, you can't tell if it takes 20s to be spliced out or 4 minutes, as both will show as fully spliced out in the sequencing library. In other words, for introns near the 5' end, a consistent "CoSI=1" pattern in the data doesn't necessarily suggest a true consistent fast splicing of that intron. Do you observe any general pattern of the measured "slowliness" in relation to the 5'-3' location of the introns? If so, should the 5' introns be specially considered or even excluded from certain analyses that use all introns?

(2) Following on my last point, it may benefit the readers if the author can provide a more detailed comparison of possible sequencing library construction choices. For example, is it feasible to also enrich for other exons for the sequencing library, etc?

(3) Figure 1C: Are there biological replicates, and should there be error bars and statistics on the plot? Similarly, in places like Figure 2, Supplemental Figure 4C, Supplemental Figure 6, etc., is there any statistical analysis that can be done to show if the claimed differences are statistically significant?

(4) The logic behind measuring the half-lives of introns seems a little unclear to me.  From the time-dependent RNA coverage plots in Figure 2, it seems that, if we assume a constant transcription elongation rate, then the splicing rate of a specific intron can vary across time after TNF stimulation, as represented by the temporal change of CoSI values, or the heights of the coverage plot relative to neighboring exons. This means the splicing rate or half-life of an intron is not necessarily constant but may be time-dependent, at least in the case of TNF stimulation. Shouldn't the half-life measurements be designed in a way to measure the half-life at multiple time points after TNF stimulation? And maybe the measured half-lives of some introns will show as time-dependent?

(5) In Supplemental Figure 6, the interpretation is a little confusing to me: If delayed splicing is causing delayed expression of the corresponding gene, shouldn't the non-immediate gene groups (early/intermediate/Late) have low CoSI beginning from the early time points (e.g. 4 minutes)? Why does the slowdown of splicing seem to peak at a later time point? Does it mean immediately after TNF stimulation, there's a different mechanism in delaying the expression of the non-immediate gene groups? Maybe it's better to have more explanation or use a different visualization to show what non-immediate gene groups are experiencing at very early time points.

(6) On the fine-tuning of the deep sequence model: it's a little unclear whether the input and output are time-dependent. It's stated that expression at multiple time points is used for training, but it's unclear whether the model outputs time-dependent expression patterns and whether the time information is used as input.

Reviewer #3 (Public review):

Summary:

The manuscript by Dearborn et al investigates the kinetics of intron splicing in inflammation-associated transcripts after TNF-stimulation of macrophages, using targeted sequencing of chromatin-associated RNA to obtain high coverage across a focused set of induced genes. The authors' main conclusion is that splicing kinetics are heterogeneous across these transcripts, and that delayed introns (which they term "bottleneck introns") are associated with weak donor sequences. Using a deep learning approach, they have also identified additional sequence features that might contribute to intron splicing kinetics.

Overall, I think the findings in the manuscript are very intriguing and will be of interest to readers working on RNA biology. The changes the authors have made to the manuscript in response to some very valid comments from reviewers have strengthened the manuscript. While the existing data might not be sufficient to directly address some of the broader mechanistic claims made by the authors, I think the findings are nonetheless very interesting and should contribute towards a better understanding of the post-transcriptional regulation of gene expression.

Strengths:

A strength of the manuscript is the experimental design. The targeted capture approach is innovative and well-suited to the goal of measuring intron-specific splicing behaviour across time. The inclusion of experimental validation in minigene assays of some of the computational predictions also strengthens the claims made by the authors.

The authors have made a constructive effort to address some of the concerns raised in a previous round of review. The revised manuscript reads as a balanced text.

Weaknesses:

The study still does not fully resolve the downstream consequences of delayed splicing. In particular, it remains unclear whether the bottleneck introns lead primarily to delayed production of mature transcripts, reduced productive transcript output, or some combination of the two.

On a related point, the minigene reporter assays measure a steady-state level of the transcript and don't provide insights into the kinetics directly.

Lastly, given that the detailed analyses were performed on a selected subset of (inflammation-induced) transcripts, a broader evolutionary interpretation needs to be restrained given the current data.

Author response:

We thank the Reviewing Editor and reviewers for their thoughtful and constructive evaluation of our manuscript, Programmed Delayed Splicing: A Mechanism for Timed Inflammatory Gene Expression. We are encouraged that the reviewers found the study valuable, the experimental design strong for the core findings. We appreciate the reviewers’ careful attention to the limits of inference in several parts of the manuscript, and will address these points in a revised version. We especially want to acknowledge that this paper has benefited from the abiding interest in splicing regulation by the editors and reviewers who have meticulously improved nearly every aspect of this multifaceted work in its present state.

Our planned revisions will focus on five areas. First, we will more carefully evaluate and discuss the extent to which the hybrid-capture strategy may impose position-dependent constraints on apparent splicing behavior, particularly across 5′ and 3′ introns. Second, we will clarify the use of the term “bottleneck introns,” distinguishing descriptive use in the main text from the ranked subsets used in downstream analyses. Third, we will revise the framing of the reporter assays to make explicit that these measure steady-state reporter output and do not, on their own, resolve all downstream kinetic consequences of delayed splicing. Fourth, we will clarify the interpretation of the actinomycin D experiments as providing estimates of intron excision behavior under transcriptional arrest rather than a complete time-resolved model of splicing during TNF induction. Fifth, we will substantially revise the scope and stated limitations of the deep learning-aided interpretations of data in this work.

Reviewer #1

We thank Reviewer #1 for the positive assessment of the hybrid-capture strategy, the splice-site reporter experiments, and the potential value of the neural-network-based analysis. We appreciate the reviewer’s view that these approaches help extend a well-established system for studying temporal gene expression in TNF-stimulated macrophages. We address the main concerns raised in the public review below.

(1) While evidence is provided that these introns are distinct from previously published splicing kinetics studies, “bottleneck” introns are not adequately placed in context for assessment of how they are similar or different.

We appreciate this point and agree that the current manuscript does not yet place these introns in sufficiently clear context relative to prior literature. Our study builds on foundational work describing regulated changes in splicing kinetics, widespread intron retention, and detained introns as biologically meaningful modes of gene regulation, including transcript-specific regulation of splicing in response to stress (Pleiss, Mol Cell., 2007), widespread functional intron retention in mammals (Braunschweig, Genome Res., 2014), and the definition of detained introns as a distinct class of post-transcriptionally spliced introns (Boutz, Genes Dev., 2015). In revision, we will expand the comparison to previously described classes of delayed or retained introns and clarify more explicitly how the introns studied here are defined in the setting of inducible inflammatory transcripts and their temporal resolution over the course of stimulation. We will also revise the relevant Results and Discussion text so that the distinction is made directly in the manuscript rather than relying on inference from the broader presentation.

(2) Splicing reporters are a good approach, but the complexities of post-transcriptional gene expression regulation are not adequately addressed.

We agree that the interpretive limits of the reporter assays should be stated more clearly and consistently. In revision, we will revise the presentation of the minigene experiments to make explicit that these are steady-state reporter assays and therefore do not, on their own, resolve all downstream kinetic consequences of delayed splicing in the endogenous context. At the same time, we believe the assay remains informative because it provides a controlled system in which the contribution of splice donor sequence can be tested directly in matched reporter constructs. In that sense, the reporter experiments are valuable as a reductionist test of whether weak donor sequences are sufficient to alter reporter output, even if they do not fully recapitulate the broader endogenous post-transcriptional environment. We will emphasize that these data support an association between weak donor sites and altered reporter output, while moderating any broader mechanistic claims that extend beyond what the assay directly measures.

(3) Deep learning models are a potentially powerful tool for identifying novel regulatory sequences; however, their use here is underdeveloped.

We appreciate this concern and agree that the deep-learning section should be revised substantially. In a revised manuscript, we will clarify the training setup, the definition of the slow-intron subsets used in downstream analyses, and the interpretation of the attribution and motif analyses. Alongside, we believe the assay remains informative because it provides a controlled system in which the contribution of splice donor sequence can be tested directly in matched reporter constructs. In that respect, the reporter experiments are valuable as a reductionist test of whether weak donor sequences are sufficient to alter reporter output, even if they do not fully recapitulate the broader endogenous post-transcriptional environment. We will revise the framing of these results so that they are presented more explicitly as identifying candidate sequence features associated with delayed splicing, rather than as direct evidence of specific causal regulatory mechanisms.

Reviewer #2

We thank Reviewer #2 for the thoughtful and detailed comments, and for recognizing the strengths of the measurement strategy and the clarity of the manuscript. We appreciate the reviewer’s view that the study will be of interest to a broad audience, and we agree that several conclusions will be strengthened by additional analysis and clearer explanation. We address the main concerns raised in the public review below.

(1) Concern regarding possible bias of the hybrid-capture strategy toward introns closer to the 3′ end, and whether 5′ introns should be treated separately in some analyses.

We thank the reviewer for this careful and important point. We agree that this is a potential limitation of the approach and that it should be addressed more explicitly in the manuscript. Our assay begins with poly(A)-selected RNA and then enriches transcripts of interest through terminal-exon capture, so the molecules analyzed are completed, polyadenylated transcripts rather than nascent partial transcripts. This feature is important for reducing ambiguity arising from incomplete transcription, particularly in the chromatin-associated fraction. At the same time, we agree that for introns near the 5′ end, the assay may have limited power to distinguish very rapid splicing from moderately rapid splicing if excision is largely complete by the time the transcript is fully synthesized and polyadenylated.

In revision, we will address this concern directly in two ways. First, we will revise the Results and Discussion to clarify that the assay provides a population-level measure of splice completion in completed transcripts and that interpretation is strongest for introns whose excision is not already fully resolved before transcript completion. Second, we will more systematically evaluate whether apparent slow splicing covaries with transcript position, distance from the 3′ end, and intron length, and we will perform sensitivity analyses with and without the most 5′ introns to determine which conclusions are robust to these positional constraints. We will also examine transcript coverage patterns in greater detail to better assess the extent to which library construction and  cDNA generation may contribute to apparent positional bias. Our preliminary inspection suggests that transcript position is not the sole determinant of the observed heterogeneity, but we agree that a more explicit treatment of this issue is warranted in the revised manuscript.

(2) Request for more detailed discussion of alternative library-construction choices.

We appreciate this suggestion and agree that the revised manuscript would benefit from a fuller discussion of the strengths and limitations of the current enrichment strategy. We chose poly(A) selection followed by terminal-exon capture because this design enriches completed transcripts of interest and reduces ambiguity from nascent partial transcripts, which is particularly important in the chromatin-associated fraction. This approach also provides greater read depth over the selected inflammatory transcripts, enabling more informative intron-level comparisons within the targeted dataset. In revision, we will clarify this rationale more explicitly in the manuscript. We will also discuss the tradeoffs of this design relative to alternative exon-targeting strategies and how those alternatives might provide different, but complementary, views of splicing kinetics.

(3) Questions regarding biological replicates, error bars, and statistical analysis in Figure 1C and other plots.

We agree that the replicate structure and intended interpretation of these plots should be clarified more explicitly. In revision, we will revise the figure legends and Methods to distinguish panels that display a single bulk RNA-seq time course (for example, Figure 1C) from panels that summarize distributions across many introns (for example, Figure 2 and Supplementary Figure 6). We will also add statistical comparisons where they are most appropriate and informative, such as in sequence-feature comparisons like Supplementary Figure 4C, while making clear that some CoSI panels are intended as descriptive summaries of intron-level heterogeneity rather than replicate-based inferential plots.

(4) Concern that intron half-lives may be time-dependent during TNF induction, and that the logic of the actinomycin D measurements is therefore unclear.

We appreciate this point and agree that the manuscript should distinguish more clearly between two related but non-identical quantities: the CoSI trajectories observed during ongoing TNF induction, and the interruption-based half-life estimates derived from actinomycin D treatment. The actinomycin D experiments were performed using multiple post-treatment timepoints, but they were designed to estimate intron excision behavior after transcriptional arrest under a defined set of conditions, rather than to measure whether an individual intron’s effective splicing rate changes across all phases of the TNF response. We agree that these estimates should therefore be interpreted as constrained measurements under the assay conditions used, rather than as a complete time-resolved model of splicing kinetics during induction. In revision, we will clarify this point in the Results, Methods, and Discussion, and we will more explicitly acknowledge that effective splicing behavior could vary across the induction time course.

(5) Concern that the interpretation of Supplementary Figure 6 is unclear, particularly why delayed splicing in non-immediate groups appears to peak later rather than at the earliest time points.

We appreciate this point and agree that the current presentation of Supplementary Figure 6 does not explain this behavior clearly enough. Our interpretation is not that delayed splicing is the sole determinant of early versus later induction classes. Rather, the earliest time points reflect a combination of transcriptional induction timing and RNA processing state. In this framework, the dip in CoSI shortly after stimulation reflects the appearance of newly induced, incompletely spliced transcripts, and the later kinetic groups appear to recover from this dip more slowly than the immediate-early group. Thus, the strongest signal of delayed splicing may become most apparent only after sufficient transcript accumulation, rather than necessarily at the very earliest time point. In revision, we will revise the text to make this logic clearer and will consider a more intuitive visualization of these group-specific CoSI trajectories.

(6) Concern that the deep-learning setup does not make clear whether the model input and output are time-dependent.

We appreciate this concern and agree that the current manuscript does not explain the model setup clearly enough. Briefly, we will clarify the role of the three TNF timepoints in model training, including the fact that these outputs were modeled jointly and that time itself was not provided as an explicit input to the model. We will also revise the Results and Methods so that the scope and interpretation of the resulting analyses are more explicit.

Reviewer #3

We thank Reviewer #3 for the positive assessment of the targeted capture design, the evaluation of overall interest of the findings, and the improvements in the current version. We appreciate the reviewer’s view that the study is intriguing and that the manuscript has been strengthened in revision. We agree, however, that the manuscript should more clearly distinguish what is directly demonstrated from what remains mechanistically unresolved. We address the main concerns raised in the public review below.

(1) The study still does not fully resolve the downstream consequences of delayed splicing, including whether bottleneck introns lead primarily to delayed production of mature transcripts, reduced productive transcript output, or some combination of the two.

We agree with this assessment. The current data do not fully resolve whether delayed splicing primarily delays mature transcript production, reduces productive transcript output, or reflects some combination of the two. In revision, we will further moderate the framing of the downstream consequences of delayed splicing and will revise the Abstract, Results, and Discussion to make clear that the present data do not fully distinguish among delayed mature transcript production, reduced productive transcript output, or a combination of both. We will ensure that the manuscript consistently presents these possibilities as alternatives not fully resolved by the current data.

(2) The minigene reporter assays measure a steady-state level of the transcript and do not provide direct insight into kinetics.

We agree and will revise the manuscript to make this limitation explicit throughout. In particular, we will ensure that the reporter assays are described consistently as steady-state reporter assays that support an association between splice donor strength and altered reporter output, while avoiding stronger claims that they directly resolve endogenous splicing kinetics or downstream transcript fate.

(3) Given that the detailed analyses were performed on a selected subset of inflammation-induced transcripts, a broader evolutionary interpretation should be restrained.

We agree that the broader evolutionary and mechanistic framing should be more carefully defined. In revision, we will restrain these interpretations so that they remain closely aligned with the inflammation-focused and targeted-transcript scope of the current study, and we will moderate language that extends beyond what is directly supported by the present dataset.

Closing Remarks

We again thank the reviewers for their constructive comments. We believe that the planned revisions will strengthen the manuscript by clarifying the scope of the mechanistic conclusions, sharpening the interpretation of the experimental approaches, and more carefully defining the role of the computational analyses. We appreciate the opportunity to revise the work and to provide this provisional response to accompany the Reviewed Preprint.