Quantitative modeling of mRNA degradation reveals tempo-dependent mRNA clearance in early embryos

Abstract

As embryos transition from maternal to zygotic control, precise clearance of pre-loaded maternal mRNAs is essential for initiating new zygotic gene expression programs. Yet the kinetics of this process and how it adapts across different developmental speeds remain unclear. Here, we introduce QUANTA, a computational approach that uses time-series RNA-seq data to quantify mRNA turnover and polyadenylation dynamics of transcriptionally silent genes and find related regulatory motifs. Applying QUANTA to zebrafish, frog, mouse, and human embryos, we uncover a conserved regulatory logic: maternal mRNA degradation onset and rates align with species’ developmental tempo. However, a subset of transcripts deviates from this pattern, suggesting species-specific kinetic tuning, which is further supported by the distinct use of destabilizing 3′UTR motifs in fast-developing species. Using temperature-based manipulation of zebrafish developmental speed and a high-throughput reporter assay, we reveal a regulatory logic of mRNA degradation scaling. Unstable mRNAs are not well-adapted to altered tempos, but scaling improves when enhancing stability through poly(A) tails or 3′UTR motifs. We demonstrate the tempo-sensitive function of 3′UTR motifs, linking regulatory sequences with developmental scaling. Our work establishes a quantitative framework for investigating mRNA turnover and reveals how clearance dynamics is tuned to match developmental pace.

Graphical Abstract

Graphical Abstract.

Introduction

Messenger RNA (mRNA) turnover is a key determinant of gene expression. Turnover rates can vary by several orders of magnitude between transcripts [1, 2], and these differences significantly shape cellular transcriptomes. For example, high-turnover genes, such as those involved in stress responses or cell identity, enable rapid but energetically costly regulation, while housekeeping genes maintain low turnover to conserve resources [3–6]. mRNA degradation also shifts dynamically in response to stimuli, driving changes in gene expression [2, 7–9].

A striking example of transcriptome reprogramming happens during the maternal-to-zygotic transition [10]. Early embryos are transcriptionally silent and utilize pre-loaded maternal mRNAs and proteins for all functionality. As development proceeds, maternal mRNAs are cleared and replaced by newly synthesized zygotic mRNAs that establish distinct developmental programs [11]. While the principles of this regulation are conserved in all metazoans [11, 12], the kinetics and regulation of maternal mRNA clearance vary widely across species. For example, a massive maternal mRNA degradation initiates within 3–5 h in fast developing anamniotic embryos (e.g. fish and amphibians) but only after 1–2 days in amniotes embryos (e.g. mammalian) [10, 11]. Timing differs also relative to other developmental events, such as the cell-cycle [10], occurring after just 1–2 cell-cycles in mammalians versus 8–10 cell-cycles in anamniotes. It is still not clear, however, how mRNA clearance is regulated and tuned across species with vastly different developmental tempos.

mRNA stability is regulated by interactions between cis-acting elements within transcript sequences and trans-acting factors that bind them. Regulators such as RNA-binding proteins (RBPs) [13–17] and microRNAs (miRNAs) [18, 19] selectively bind specific mRNA sequences and affect their fate. For example, zebrafish miRNA miR-430 [18] or Drosophila RBP SMAUG [20] mediate targeted degradation of maternal mRNAs. Regulated changes in mRNA stability are tightly linked to changes in poly(A) tail length [21]. For example, miR-430 destabilization is associated with tail shortening [18], while regulation by CPEB promotes cytoplasmic tail extension and transcript stabilization [22]. In transcriptionally silent oocytes and early embryos, poly(A) tail length also influences translation efficiency. In those systems, long tails enhance translation, while short-tailed mRNAs can remain stable [23–26]. Regulators in oocytes and embryos dynamically adjust poly(A) tail lengths to modulate gene expression as needed [23–30]. For example, poly(A) tail extension of mos and ccnb1 mRNAs is required for frog oocyte maturation [31–33] and is conserved in mammals [34–36]. Coordinated waves of deadenylation and polyadenylation regulate translation prior to genome activation, driven by conserved cis-elements and trans factors [14, 22, 37]. After genome activation, changes in poly(A) tail length mainly affect mRNA stability in embryos [26], as in other systems.

Recent advances have enabled transcriptome-wide analysis of mRNA stability and polyadenylation, and their sequence-based regulation. Combining RNA-seq with strategies for transcriptional arrest [9, 38] or RNA metabolic labeling [5, 39–41] provide information on mRNA stability. Improved analyses of short-read [23, 26, 42] and long-read [27, 28] sequencing technologies allow measurement of poly(A) tail lengths and modifications. Despite their power, such dedicated tools often require system-specific adaptations and technical expertise, which could limit their use. Some may also involve manipulations that can confound biological interpretation. In contrast, standard RNA-seq is easy to use and vast amounts of data already exist and can be utilized to infer valuable, indirect insights into mRNA regulation. For example, quantification of intron levels [43, 44] and single nucleotide polymorphisms (SNPs) [45] distinguish pre-mRNA production and deduces mRNA degradation. Additionally, comparing polyA+ and total-RNA fractions can indirectly estimate poly(A) tail lengths [46, 47], but this approach has not yet been combined with quantitative modeling. Unraveling the sequence-based rules governing mRNA regulation further requires tools that can robustly link sequence features to genome-scale regulatory outcomes [48–51]. Motif finding tools [52, 53] can identify enriched elements, but are often limited by the complexity of genomic data, and particularly (i) the limited complexity of native transcriptomes, (ii) non-linear and combinatorial interactions between signals, and (iii) difficulty in extracting discrete effects from continuous regulatory outputs (e.g. mRNA stabilities). Massively parallel reporter assays (MPRAs) help overcome some of these challenges by testing synthetic sequences in controlled contexts and have uncovered elements affecting mRNA stability [54, 55] and polyadenylation [56]. Yet MPRAs remain limited in capturing the complexity of native sequences and their combinatorial logic [51]. Thus, scalable, quantitative tools to model mRNA turnover and decode its regulatory logic from standard RNA-seq data, without perturbing the system, remain lacking.

Here, we introduce QUANTA (QUantitative ANalysis of Total and A+ RNA), a kinetic modeling framework to infer mRNA degradation and polyadenylation kinetics of transcriptionally silent genes from standard RNA-seq time-series. We apply QUANTA to investigate maternal mRNA clearance across species and developmental tempos, and use a complementary MPRA to also explore the underlying sequence logic. This allowed us to decipher a regulatory code governing mRNA kinetics, and how it adapts across developmental tempos. By enabling kinetic modeling from RNA-seq data, QUANTA offers a generalizable approach to study mRNA regulation across diverse biological systems, from early development to disease.

Materials and methods

RNA-seq data processing and expression analysis

Data

RNA-seq datasets used in this study (Supplementary Table S8) were taken from published works involving zebrafish, clawed frog, mouse, and human samples.

Mapping

RNA-seq reads were aligned to the relevant reference genome using STAR [57] with default parameters. The following reference genomes and annotations were used for alignment: zebrafish genome GRCz11 (GenBank:GCA_000002035.4) and Ensembl [58] annotations release 102; clawed frog genome UCB_Xtro_10.0 (xt10) and Xenbase [59] annotations XENTR_10.0; mouse genome GRCm39 and Ensembl annotations release 108; human genome GRCh38 and Ensembl annotations release 110.

Expression

For all organisms except frog, gene annotations were filtered to include only genes with a “Protein stable ID” value or with “Gene type” equals to “lncRNA.” Genes in frog dataset were not filtered. Organism specific annotation files were supplemented to include a single precursor mRNA for each gene. Precursors span the entire gene locus, from the earliest annotated isoform start point to the last annotated isoform end point. Isoform specific FPKM values were calculated using Cufflinks [60]. Mature mRNA expression was calculated by summing the FPKM values of all mature mRNA isoforms. FPKM values were log2 transformed, and values below −3 were floored to this value.

Temporal normalization

Temporal FPKM values (log2 transformed) in each dataset were normalized by iteratively selecting a set of control genes, and normalizing all temporal samples by their mean expression. To select control genes, we fitted a linear regression model to log2-transformed means and standard deviations of genes, and select genes with a standard deviation that is at least 2-fold lower than expected by the linear regression model. We repeated the normalization multiple times, as long as the number of selected controls increased.

Selecting expressed genes

For each organism, genes were selected for further analysis by the following criteria: log2(normalized FPKM) >2 in at least one temporal sample, in >50% of ribosomal depleted datasets. For human analysis, all datasets (ribosomal depleted and polyA selected) were used for gene selection.

Identification of shut-off genes

Classification

A gene is defined as “contributed” (maternally deposited in the context of embryos) if its maximal mature-RNA expression before activation-time is larger than a threshold value, defined by 2 standard deviations below the mean (excluding the minimum value −3 that skews the distribution). A gene is defined as “produced” (zygotically expressed in the context of embryos) if it fits at least one of the following two criteria: (i) ratio of maximal precursor–RNA expression after activation-time to maximal precursor–RNA expression before activation-time is >log2(1.25), and maximal precursor-RNA expression after activation-time is > −2.5. (ii) Ratio of mean mature-RNA expression after activation-time to mean mature-RNA expression before activation-time is >log2(0.75), and maximal mature-RNA expression after activation-time is > –2.5. “Combined” genes are those that match both “contributed” and “produced” criteria, “shut-off” genes (maternal) match only “contributed” criteria, and “turn-on” genes (zygotic) match only “active” criteria. Activation-time for embryonic data was defined per organism: zebrafish 3.5 hpf, frog 4.5 hpf, mouse 22 hpf, and human 32 hpf. Activation-time for non-embryonic datasets was defined as 2 h in human and 0.6 h in mouse DCs. In all developmental data except human, each of the ribosomal depleted dataset was classified separately, and majority vote classification was assigned for each gene. To improve robustness in inherently noisier human data (owing to a variable genetic background), all datasets (ribosomal depleted and polyA selected) were used for classification.

Permutation tests

Classification of genes into “shut-off”, “turn-on,” or “combined” expression was compared by permutation tests. One classification was randomly permuted 1000 times relative to the other, and the number of inconsistent classifications was counted. Counts were normally distributed, and P-value was assigned using a normal distribution with mean and standard deviation as measured in permutation tests.

Degradation models

Degradation model

Model is defined using the following parameters: initial expression level (), degradation onset time (), degradation rate (), and degradation shut-off time (). Equations are:

polyA+ model

Model is defined using the following parameters: initial expression level (), degradation onset time (), polyadenylation rate (), degradation rate (), and degradation shut-off time (). Equations are:

For each gene, each model was fitted to temporal RNA-seq data using non-linear least squares regression with multiple initial values (Matlab implementation). Samples with expression <1/16*(mean of later samples) are considered dropouts and removed from the fit. The range of possible values for each parameter was restricted as follows. Initial expression level was bounded in the range (−9, 15). Other parameters were bounded by sample times: onset time (min: earliest sample, max: 60% of latest sample), switch time (min: 20% of latest sample, max: 60% of latest sample), offset time (min: 60% of latest sample, max: latest sample + 1), degradation rate (half-life max: 60% of latest sample, min: 3% of latest sample), and deadenylation rate (half-life plus/minus: 3% of latest sample).

Accuracy analysis

We compared the fits to the two nested models (“degradation” model is nested in “polyA+” model) by a likelihood ratio test for each gene, and included an additional null model with a single parameter (constant expression level). The likelihood was calculated by assuming an additive Gaussian noise model (μ=0,  = 0.2).

Sequence k-mer enrichment analysis

3′UTR sequences

For each organism, the “canonical” 3′UTR sequences (as annotated in ensembl) were downloaded from ensembl biomart [61]. 3′UTR sequences <10 bp were filtered out.

k-mer P-values

We associated a short sequence (k-mer) with a regulatory effect when genes that contain this k-mer in their 3′UTR had a significantly different distribution (one-sided Kolmogorov–Smirnov test, 1% false discovery rate (FDR)) of a specific property (e.g. degradation rate) than genes without this sequence. We assigned an effect size to each k-mer by calculating the standardized mean difference defined as , where is the mean of the first population, is the mean of the second population and is the standard deviation (based on both populations). We tested all short sequences (k-mers) between 4 and 7 nucleotides long. Since the number of “shut-off” genes in human and mouse was lower, reducing the statistical power of the test, we used 5% FDR in these organisms and tested only k-mers between 4 and 6 nucleotides long.

3′UTR length normalization

The length of 3′UTR sequences varies between genes. Length can affect occurrence of k-mers since these become more likely to occur as the sequence gets longer. Length can also affect mRNA regulation, as previously shown [62]. These create interactions that might bias the distributions and affect statistical tests (see below). To control for these effects, we (i) remove the top and bottom 10% of lengths from the analysis and (ii) normalize values by their length dependence. For that, we use a linear model to connect each property with 3′UTR length, and subtract the resulting model from measured values. Notably, this normalization is not required for MPRA reporters that all have the same UTR length.

Motif prediction

To perform motif prediction, we first identify positions within 3′UTRs with presence of significant k-mers. In this way, we define the native context of k-mers and to identify and merge co-occurring signals. Next, we use an Expectation–Maximization (EM) based approach [53] to iteratively optimize the assignment of the selected positions into motifs, represented by position specific weight matrices (PWMs). To initialize a set of motifs, we score k-mers of size 4–8 nt by summing the significance score [−log₁₀(P-value)] of all 3′UTR positions that contain them, and select between 1 and 15 such k-mers with maximal score. For each selected k-mer, we align all matching 3′UTR positions and use the alignment to initialize a PWM for the motif. Once an initial set of motifs was defined, we iteratively optimize assignment of 3′UTR positions to motifs. (i) We assign each 3′UTR position to the PWM that maximizes its score (based on sequence similarity). (ii) Once all sequences have been assigned, we use their alignment to re-calculate a PWM for the motif. We repeat this procedure by initializing an increasing number of initial motifs (between 1 and 15) and calculate a Bayesian Information Criteria (BIC) score for the final motif assignment, which weighs both the fit of motifs to sequences and the total number of motifs. Final result is the optimal result by the BIC score. We perform the entire procedure separately for k-mers with either a positive or a negative effect.

k-mer filtering

We provide an option to filter the significantly enriched k-mers for each property (e.g. degradation rate) based on P-value, effect size, and sequence similarity, using the following rules: (i) We filtered out a k-mer X with a less significant P-value whose sequence is fully contained within another k-mer Y with a more significant P-value. In this case, the group of genes defined by X fully contains the group of genes defined by Y and extends it with additional genes (Y is a subgroup of X). The subgroup defined by Y has a more significant effect, and therefore X is filtered out. (ii) We filtered out k-mer X with a less significant P-value whose sequence fully contains another k-mer Y with a more significant P-value. In this case, the group of genes defined by Y fully contains the group of genes defined by X and extends it with additional genes (X is a subgroup of Y). The subgroup defined by X has a less significant effect, and therefore X is filtered out. In cases of an equal P-value, the effect size is similarly used as a criterion for filtering k-mers. The filtering procedure is used for presentation of the results by QUANTA and not for any subsequent analysis.

Gene set enrichment analysis

Functional enrichment analysis was performed using g:Profiler R client (version e106_eg53_p16_65fcd97) with a 1% FDR multiple hypothesis testing correction method.

Comparison between species

Orthologies

Orthologies were calculated relative to human and relative to zebrafish. All orthology information was downloaded from ensembl biomart [61]. In addition, all genes with identical gene names were considered as orthologs. Orthology information relative to zebrafish was also downloaded from the following resources. Frog orthologies were downloaded from Xenbase [59]. Mouse orthologies were downloaded from zfin [63] and from [64]. Human orthologies were downloaded from zfin [63].

Developmental time scaling

Temporal samples were scaled to match zebrafish standard developmental times (at 28°C) by dividing all times (in hpf) by the time of the latest developmental stage used for the specie in our analysis (frog: 14 hpf, mouse: 72 hpf, human: 96 hpf) and multiplying by 10.

Unified scaled degradation model

Model

The unified scaled degradation model is based on the “degradation model”. We fit the model to data from two separate experiments after times in both experiments were scaled to match zebrafish standard developmental times (at 28°C). The unified model uses two initial expression level parameters (), one for each of the two separate experiments. All other parameters are shared between experiments: degradation onset time (), degradation rate (), and degradation shut-off time (). Therefore, this model uses a total of five parameters. The model was fitted using non-linear least squares regression with multiple initial values (Matlab implementation), as described above for the “degradation model.”

Accuracy analysis

We compared the fit of the unified scaled with the fit of a separate “degradation model” to each of the two datasets, generating a hierarchy of two nested models (the unified model is nested in the fit of two “degradation” models). We used a likelihood ratio test for each gene, by assuming an additive Gaussian noise model (μ= 0). For comparing between species, variance was estimated by pairwise-comparison of samples along the timecourse and averaging for all genes, plus a constant sigma factor ( = 0.5). In MPRA experiments, variance was estimated by pairwise-comparison between two biological replicates for each timepoint.

Web portal database for comparative maternal mRNA degradation

Web portal was prepared using shiny65 (1.7.4 version) in R (version 4.2.2). All the datasets of zebrafish, xenopus, mouse, and human embryos were included. Modeling results for maternal genes were also included.

Zebrafish

All protocols and procedures involving zebrafish were approved by the Harvard University/Faculty of Arts and Sciences Standing Committee on the Use of Animals in Research and Teaching (IACUC; Protocol #25-08) and the Hebrew University Ethics Committee (IACUC; Protocol #NS-15859). Embryos were grown and staged according to standard procedures [65]. Zebrafish embryos from wild-type AB/TL strains were used for all experiments.

Reporter library preparation, sample collection, and data processing

The MPRA reporters and spike-in controls plasmids were prepared as previously described [54, 66]. Five mRNAs were specifically synthesized to be used as spike-in controls and had identical structure as reporters, including a 40 nt poly(A) tail, and processed together with the sample. Controls were mixed at 2-fold decreasing amounts (highest spike-in is 16-fold higher than lowest spike-in).

Experiment comparing total-RNA and polyA+ reporters

MPRA library preparation, injection into embryos, sequencing, data processing, and normalization were all done as previously described [54, 66], with the following modifications. For polyA+ sequencing of reporters, total-RNA was polyA selected using the NEBNext Poly(A) mRNA Magnetic Isolation Module (NEB #E7490S). Libraries were sequenced at low-scale (∼2 million reads per library) on an Illumina miSeq platform with 168 nt single-end reads. We assayed two types of reporters, which were synthesized either with an initial 40 nt poly(A) tail (A₄₀) or without an initial tail (A₀). A 40-nt polyA sequence was encoded on the plasmids. Plasmid was linearized for invitro synthesis either before the polyA sequence, resulting in reporters without a tail (A₀) or after the polyA sequence, resulting in reporters with a 40 nt tail (A₄₀). After 1 h, polyA+ levels of A₀ reporters are on average 2-fold lower compared to total-RNA, while polyA+ levels of A₄₀ reporters are similar (Supplementary Fig. S6A), demonstrating a dependence on poly(A) tail length in recovery.

Temperature experiments

UTR-Seq plasmid library was linearized by enzymatic restriction at the end of its 3′UTR (A₀ library) or after the poly(A) sequence (A₄₀ library). HiScribe SP6 RNA Synthesis Kit (NEB) was used to in vitro transcribe a library of mRNA reporters from linearized plasmids. One-cell staged wild-type zebrafish embryos were collected at 26°C and injected with 100 pg of mRNA reporters. Following injection, embryos were grown at the specified incubation temperature (22°C or 34°C). Two samples of 20 embryos each were collected at each time-point. In the 22°C incubation experiment, samples were collected every 2 h between 1 and 15 hpf. In the 34°C incubation experiment, samples were collected every 2 h between 1 and 5 hpf and every hour between 5 and 8 hpf. Total RNA was isolated using TRI-reagent (Sigma–Aldrich), after adding 4 pg of spike-in control mix during the initial TRIzol lysis step. Total RNA was reverse-transcribed with Maxima H minus RT (Thermo Fisher). Resulting cDNA was amplified by 19 PCR cycles using LongAmp Hot Start Taq 2X Master Mix (NEB). RT and PCR primers were used as previously described [54, 66]. PCR product was digested with EcoRV to eliminate empty vectors/reporters and cleaned with 1.0 volumes of AMPure beads (Agencourt). Libraries were quantified by Qubit fluorometric quantification (ABP Biosciences) and sequenced at low-scale (∼2 million reads per library) on Illumina NovaSeq platform with 122 nt single-end reads. Sequencing data processing and normalization were all done as previously described [54, 66]. Times in reporter data were scaled between temperatures only after the first sample (1 hpf), since during the first hour injections were all performed at 26°C. Scaling was based on the observed developmental stages at each sample: the 100% epiboly stage was reached after 15 h at 22°C, and after 8.5 h at 34°C. Variance in reporter experiments was estimated by pairwise-comparison between replicates for each timepoint.

Validation reporters preparation, sample collection, and processing

We used six UTR-Seq reporter sequences selected for our previously designed validation set, and all their previously designed loss-of-function and gain-of-function modification (total of 36 sequences) [54]. These include two “neutral” sequences that did not contain any peaks (BG1 and BG2), three sequences with a late-onset high-weight peak (M430, ARE, and PUM), and a sequence with an early-onset high-weight peak (POLYU). To that, we added two sequences that did not scale between temperatures based on our analysis and contained a C-rich motif (POLYC and POLYC2). We designed additional loss-of-function and gain-of-function modification for these sequences (total of 16 sequences). Finally, we added to each of the previously selected six UTR-Seq reporters a combination with a C-rich motif (total of six sequences). This resulted in a set of 60 validation reporter sequences (Supplementary Table S9). Each of the validation sequences was synthesized as a separate oligo at nmole scale. Oligos were separately amplified via the universal adaptor sequences, mixed at equal concentrations, and cloned and transcribed as previously described. By using equal concentrations from each sequence, we minimized representation biases in the resulting plasmid and mRNA validation libraries. Wild-type zebrafish embryos were injected with 200 pg of validation library (either A₀ or A₄₀) as described above. Following injection, embryos were grown at the specified incubation temperature (22°C or 34°C). Two samples of 15 embryos each were collected at each time-point. In the 22°C incubation experiment, samples were collected every 2 h between 1 and 15 hpf. In the 34°C incubation experiment, samples were collected every hour between 1 and 8 hpf. Samples were processed and analyzed as described above.

qRT-PCR of maternal genes

Total RNA from MPRA experiments (22 and 34°C) was reverse transcribed using the iScript^™ cDNA Synthesis Kit (Bio-Rad, #1708891). For the 22°C time course, samples collected from 1 to 13 hpf were analyzed, and for the 34°C time course, samples from 1 to 8 hpf were used. For each time point, two biological replicates were included. Real-time quantitative PCR was then performed using iTaq^™ Universal SYBR^® Green Supermix (Bio-Rad, #1725125), with two technical replicates per biological sample and primer pair. Primers were designed for two control genes (actb2 and etf1b) and three maternal genes (buc, btg4, andbmp15). Cq measurements were averaged between two technical replicates and normalized relative to the control gene etf1b per sample. Samples within each timecourse were further normalized relative to the average expression between two biological replicates at the first sample (1 hpf).

Results

QUANTA enables genome-wide kinetic modeling of mRNA decay in shut-off genes

During dynamic responses, changes to the cellular transcriptome combine the shutdown of a subset of genes, while others are turned on [67]. To investigate this transcriptional shut-off component using standard RNA-seq data, we developed QUANTA: a computational framework for modeling mRNA degradation kinetics and uncovering sequence-based regulatory features.

QUANTA operates in three stages (Fig. 1; “Materials and methods” section). First, it identifies transcriptionally silent (shut-off) genes by quantifying intron-containing transcripts in total-RNA-seq data [43, 44] (Fig. 1A). Genes are classified as shut-off if both pre-mRNA and mature mRNA levels show no increase over time, indicating a lack of new transcription.

Figure 1.

QUANTA dissects the transcriptional shut-off component of gene expression programs. (A) Left: Analysis of intron and exon RNA-Seq reads identifies pre-existing genes that are not transcribed during a dynamic response (shut-off, red) and distinguish them from actively transcribed genes (blue). Pre-existing genes are present as fully spliced and processed transcripts, with total-RNA-Seq reads that match only exon sequences. Actively transcribed genes are processed and spliced, so their intron sequences are also detected in total-RNA-Seq. Following stimulation, changes to cellular transcriptome arise by degradation of a subset of pre-existing transcripts that are replaced with newly synthesized transcripts. Right: selection of shut-off genes. New transcription is defined by an increase in expression levels of either mRNA or pre-mRNA isoforms in comparison to their initial levels, and only genes which do not exhibit new transcription are classified as shut-off genes for subsequent analysis. (B) Two alternative models for the dynamics of maternal mRNA degradation. Left: a “degradation” model for analyzing total-RNA levels (M₀, initial level). A gene-specific rate function is parametrized by a constant decay rate (β, 1/h), an onset-time (t₀) and an offset-time (t₁). Right: a “polyA+” model models for analyzing polyA+ RNA levels (A₀, initial level). A gene-specific rate function is parametrized by two constant rates: an early deadenylation rate (α, 1/h) and a late decay rate (β, 1/h) with a switch-time (t₀) and an offset-time (t₁). (C) Associating cis-elements with mRNA regulatory parameters. Shut-off genes are divided into two groups based on the presence of a short (4–7 nt) sequence (k-mer): either containing the sequence (orange) or not containing it (black). The distribution of parameter values is compared between the two groups by Kolmogorov–Smirnoff test. k-mers that divide the genes into two groups with distinct distributions of parameter values are associated with this parameter. Finally, an EM-based algorithm merges k-mers into motifs by their sequence and activity similarity.

Second, QUANTA models the decay dynamics of these shut-off genes (Fig. 1B). Two alternative kinetic models are used, each reflecting one of two commonly applied RNA-seq strategies. The simpler “degradation model” describes changes in total RNA-seq (measured after ribosomal RNA depletion). It assumes exponential decay of pre-existing mRNA copies and estimates a constant gene-specific degradation rate (β), as well as onset (t₀) and offset (t₁) times. The alternative “polyA+ model,” applied to polyA+ selected RNA-seq, includes an additional parameter (α) to account for changes in polyA+ levels that may precede degradation [68]. Differences reflect changes in the average poly(A) tail length of mRNAs, which were shown to influence their recovery in polyA+ data [44, 47, 51, 69, 70]. This parameter (deA) can be either positive to represent a decrease in polyA+ levels, or negative and represent an increase in polyA+ levels. The ratio of polyA+ to total RNA levels (“polyA+ fraction”) serves as a proxy for average poly(A) tail length dynamics [44, 47, 51, 69, 70]. For each shut-off gene, QUANTA fits these models to temporal RNA-seq data to infer optimal degradation parameters.

At the final step, QUANTA incorporates a statistical approach [1, 71] to identify cis-regulatory signals associated with decay dynamics (Fig. 1C). It systematically tests all short (typically 3–7 nucleotide long) sequences (k-mers) within 3′UTRs for association with kinetic parameters estimated by QUANTA. Analysis is restricted both by significance (Kolmogorov–Smirnov, FDR < 1%) and effect size (standardized mean-difference). Subsequently, it groups relevant k-mers within 3′UTRs into discrete motifs based on their sequence and function similarity using an EM-based approach.

Overall, QUANTA provides a tool for systematically analyzing shut-off genes. It generates a direct, in-depth quantitative view of mRNA decay kinetics and its sequence-based regulation. By leveraging standard RNA-seq data and models specifically designed to analyze degradation patterns, it provides an accessible and scalable solution for dissecting transcriptional shutdown, a critical process in development, stress responses, and disease [7, 9, 72].

A quantitative analysis of in vivo decay dynamics of zebrafish maternal mRNAs

Using QUANTA, we analyzed temporal expression patterns in zebrafish embryos throughout gastrulation (0–10 h post fertilization, hpf). For this analysis, we used three total-RNA-seq datasets (measured after ribosomal RNA depletion) [51, 73, 74] and four polyA+ RNA-seq datasets [45, 51, 75, 76] of zebrafish embryos.

Classification successfully distinguished maternally provided from zygotically expressed or combined (maternal and zygotic) embryonic transcripts. As expected, intron levels increased after genome activation, both globally (Supplementary Fig. S1A) and specifically for zygotic genes (Supplementary Fig. S1B and C). Gene classification (Supplementary Table S1) was consistent between three total-RNA datasets (>70% genes classified identically, Supplementary Fig. S2A), and also agreed with metabolic labeling based classification [1, 77] (>56% genes classified identically, Supplementary Fig. S2B). Functional annotations within classes identified expected pathways (Supplementary Table S2), including reproductive process (5% FDR hypergeometric P< 3 × 10^–6) for maternal genes, developmental process (P< 2 × 10^–23) for zygotic genes, and housekeeping processes such as translation (P< 7 × 10^–49) for combined genes. A final majority classification across three total-RNA datasets selected 4381 zebrafish maternal genes (Fig. 2A, 35%) for subsequent analysis. These transcripts are inherited from the egg and not transcribed in the embryo; thus, their regulation allows a direct view into mRNA decay.

Figure 2.

QUANTA elicits the regulatory kinetics of zebrafish maternal mRNAs. (A) Classification (y-axis: class) of zebrafish embryonic genes (x-axis, fraction of genes). Number of genes and their factions are indicated on bars. (B) Fitting of models to temporal (x-axis, hpf) total-RNA (orange) and polyA+ (blue) expression levels (y-axis, normalized FPKM, log₂) of specific maternal genes in all analyzed datasets. Gene names and average parameters of the fitted models are indicated. (C–E) Distribution (y-axis, % of maternal genes) of per-gene averaged parameter values (x-axis) estimated by QUANTA in zebrafish embryonic RNA-Seq datasets (gray: degradation model, black: “polyA+” model). (C) initial RNA level (normalized FPKM, log2), (D) deadnylation rate (1/h), (E) onset/switch time (h), and (F) decay rate (1/h). (G) Left: correlations between deadnylation rate (x-axis, 1/h) and initial polyA+ fraction (y-axis). Right: correlations between poly(A) length change at 0–2 hpf (x-axis, log2) and initial polyA length (bp, y-axis) by Chang et al. Colors represent density (blue = low density; yellow = high density).

Kinetic models successfully captured the expression patterns of maternal genes across datasets (Fig. 2B). While total-RNA levels of most genes remained constant prior to zygotic genome activation (before 3 hpf), polyA+ levels of genes changed in corresponding measurement (Supplementary Fig. S2C). For example, polyA+ levels of genes such as mcm3l and buc (Supplementary Fig. S2D) increased before 3 hpf, and those of genes such as wee2 and dazl (Supplementary Fig. S2D) decreased before 3 hpf, while their total-RNA levels remained unchanged. Only a few genes (e.g. gdf3, Supplementary Fig. S2D) behaved similarly in both populations. As a result of these differences, the simpler “degradation model” successfully captured expression patterns in total-RNA datasets, while the “polyA+ model” offered significant improvement in polyA+ datasets (Supplementary Fig. S3). Model parameters were consistent across datasets (Supplementary Fig. S4A and B), demonstrating reproducibility. Deadenylation parameters reflected poly(A) tail lengths that were measured in zebrafish embryos [23]. Early deA rates (α) correlated with measured differences in poly(A) tail lengths between 1 and 4 hpf (Supplementary Fig. S4C). The “polyA+ fraction” also correlated to poly(A) tail lengths [23], both at early and later times (Supplementary Fig. S4D), supporting their interpretation as indirect estimates of poly(A) tail lengths.

Together, these findings demonstrate that QUANTA robustly classifies shut-off maternal transcripts and accurately quantifies their decay kinetics from standard RNA-seq data.

Two phases of cytoplasmic mRNA metabolism in zebrafish embryos

QUANTA reveals a maternal mRNA clearance dynamic in zebrafish embryos that aligns with much of current knowledge.

Before genome activation, maternal transcript degradation was minimal, yet changes in poly(A) tail lengths were prominent. Kinetic parameters show that maternal transcripts were inherited with short poly(A) tails that are globally elongated after fertilization, as previously reported [23]. Consistent with this, initial polyA+ were on average 3.2-fold lower than total-RNA levels (Fig. 2C). Before zygotic genome activation, polyA+ levels increased in 72% of maternal transcripts (α < 0, Fig. 2D; e.g. pdlim2 and mctp2b, Fig. 2B) and decreased in 28% (α > 0, Fig. 2D; e.g. org and dazl, Fig. 2B).

Changes in poly(A) tail length were not associated with specific functional annotations, suggesting they may not immediately relate to functional transitions from oocyte-related to embryo-related processes. However, poly(A) tail dynamics was tightly linked to initial tail length: shorter tails were extended and longer ones shortened (Pearson r = 0.86, Fig. 2G), a trend also confirmed by direct poly(A) measurements (Fig. 2G). Thus, poly(A) tail length changes in the embryo could reduce tail length differences inherited from oocytes.

The onset of maternal mRNA degradation occurred in the expected time window between 3 and 5 hpf, coinciding with zygotic genome activation. Degradation of over 90% of maternal genes initiated within this period (Fig. 2E), though onset times varied slightly between genes (e.g. org, dazl early onset < 3.5 hpf, mctp2b, lrmp late onset > 4.5 hpf; Fig. 2B).

Following genome activation, a coordinated degradation program affected both total- and polyA+- RNA. Degradation rates estimated from both RNA populations were highly correlated (Supplementary Fig. S3E). Maternal half-lives ranged from <30 min to >4 h (median = 45 min), consistent with previous in vivo estimates [2, 54] (Fig. 2F). For example, btg4 was rapidly degraded (0.4 h), while pdlim2 remained stable (4.3 h, Fig. 2B). Furthermore, degradation rates correlated with both the “polyA+ fraction” at 6 hpf (r = 0.35, Supplementary Fig. S4E) and its change between 2.3 and 5.2 hpf (r = 0.38, Supplementary Fig. S4E), linking poly(A) tail changes to the kinetics of mRNA clearance.

Together, these findings demonstrate how QUANTA quantitatively dissects maternal mRNA regulation and reveal a two-phase model: an early phase of poly(A) tail remodeling, followed by synchronized degradation after zygotic genome activation.

Decoding maternal regulatory sequence elements in zebrafish 3′UTRs

Using QUANTA, we analyzed 3′UTR sequences of zebrafish maternal mRNAs with respect to their decay dynamics, uncovering both known regulatory elements and novel sequence motifs with potential roles in mRNA regulation (Supplementary Table S3 and Fig. 3).

Figure 3.

QUANTA reveals regulatory sequence elements in zebrafish 3′UTRs. (A) Left: volcano plots showing the effect size (standardized mean difference, x-axis) and P-value (1% FDR, one-sided Kolmogorov–Smirnoff test, y-axis) of different k-mers for different kinetic parameters. Test compares mRNAs with or without a specific k-mer. Colored dots represent k-mers that pass the P-value and effect size thresholds. In those cases, colors represent density (blue = low density; yellow = high density). k-mers that do not pass the thresholds are colored in grayscale by density. Representative top k-mers are indicated. Right: boxplots showing the normalized distribution of kinetic parameters (scaled values, y-axis) in native maternal mRNAs, with (red) or without (gray) a specific sequence element. Central line represents the median, box edges are 25th and 75th percentiles, whiskers extend to largest/smallest value except outliers; outlier points are plotted individually. Non-significant differences are shaded in light colors. (B) Motif logos representing k-mers associated with positive (right) or negative (left) effect on kinetic parameters. (C) Motif logos representing k-mers associated with positive (right) or negative (left) effect on the ratio between expression values predicted by the polyA+ and total-RNA models, as a proxy for poly(A) tail lengths. On both panels, only motifs that are associated with 10% or more of k-mer positions are shown.

Our analysis (Fig. 3A and B) confirmed the known destabilizing activity of miR-430 seeds (GCACUU). Additionally, it identified several AC-rich motifs (e.g. CACA) associated with delayed onset of total-RNA degradation, suggesting a potential role in mRNA decay timing.

We also found that polyadenylation-related elements [78] modulate maternal mRNA dynamics (Fig. 3A and B). Canonical polyadenylation signals (PASs; e.g. AAUAAA) and cytoplasmic polyadenylation elements (CPEs; e.g. UUUUUA) were associated with higher initial polyA+ levels (and therefore also with higher deA rates). These findings align with their known roles in promoting poly(A) tail extension, leading to longer initial tails for those transcripts. Notably, PAS motifs were also linked to slower polyA+ degradation, suggesting a stabilizing effect. Conversely, A-rich motifs (e.g. AAAAAA) were linked with lower initial polyA+ levels (and thus also deA rates), associating them with possible poly(A) shortening, an activity that was not previously recognized.

By leveraging the dynamics of “polyA+ fraction” (ratio of polyA+ to total-RNA levels) as a proxy for poly(A) tail length (Fig. 3C), we could resolve the temporal activity patterns of these elements. CPEs promoted longer poly(A) tails specifically before genome activation (∼4 hpf), whereas PAS motifs extended tail length more consistently throughout gastrulation (up to 10 hpf). A-rich motifs correlated with shorter tails early on (0–1 hpf), while the miR-430 seed was linked to tail shortening after genome activation at 4 hpf.

Together, these results validate the in vivo regulatory impact of established 3′UTR signals and highlight additional sequence elements potentially involved in maternal mRNA clearance.

MPRA reporters confirm regulatory kinetics and sequence element activities invivo

To validate the genomic observations, we implemented an MPRA that is compatible with QUANTA. In this assay, we injected into 1-cell stage embryos non-polyadenylated MPRA reporters [54] with different 3′UTR sequences, and measured both their total-RNA and polyA+ levels over time (“Materials and methods” section, Fig. 4A). Our sequencing depth allowed a reliable analysis of 6117 different reporters.

Figure 4.

A massively parallel reporter assay recapitulates maternal mRNA regulatory kinetics invivo. (A) A large set of mRNA reporters with a constant backbone and different 3′UTR sequences were injected into 1-cell stage embryos and followed over time. Reporters were sequenced from temporal total-RNA and polyA+ fractions. (B) Left: temporal (columns) mRNA expression patterns relative to initial measurement (orange: increase; purple: decrease; white: unchanged) of 6117 mRNA reporters (rows) that were injected into zebrafish embryos at the 1-cell stage. Reporters were captured in total-RNA (left) or polyA+ (right) fractions. Right: examples of temporal (x-axis, hpf) expression levels (y-axis, normalized FPKM, log₂) of individual reporters (left: total-RNA, right: polyA+). (C) Distribution of ratios between measured polyA+ and total-RNA level (y-axis, log₂) of reporters in temporal samples (x-axis). The central dot is median; gray box bounds are 25th and 75th percentiles, upper and lower limits of whiskers are 1.5× interquartile ranges. Values outside of the upper and lower limits are defined as outliers. (D) Distribution (y-axis, % of reporters) of parameter values (x-axis) estimated by QUANTA in MPRA datasets (solid: “degradation” model, dashed: “polyA+” model). Left to right: initial RNA level (normalized FPKM, log₂), deadnylation rate (1/h), onset/switch time (h), and decay rate (1/h). (E) Motif logos representing k-mers associated with positive (right) or negative (left) effect on kinetic parameters. Only motifs that are associated with 10% or more of k-mer positions are shown. (F) Enrichments associated with 3′UTR sequences of MPRA reporters. Boxplots show the normalized distribution of parameter values (scaled values, y-axis) in reporters with (red) or without (gray) a specific k-mer (as noted) across different parameters estimated on total-RNA or polyA+ data (parameter name on top, data on bottom). Central line represents the median, box edges are 25th and 75th percentiles, whiskers extend to largest/smallest value except outliers; outlier points are plotted individually.

Reporters confirmed mRNA polyadenylation and degradation dynamics. Reporters’ total-RNA levels show few changes prior to genome activation, while their polyA+ levels increased in corresponding measurement (Fig. 4B). Thus, reporters’ “polyA+ fraction” increased over time (Fig. 4C), suggesting they are polyadenylated in vivo. QUANTA analysis fitted kinetic parameters to MPRA data, with the “polyA model” providing an improved fit to polyA+ data, and total-RNA data retaining the “degradation model” (Supplementary Fig. S5). Kinetic parameters (Fig. 4D) showed lower initial polyA+ compared to total-RNA levels, and negative deA rates, further support invivo polyadenylation of the initially non-adenylated reporters. Most reporters (60%, onset >3 h) remained stable until genome activation. However, a subset of reporters (40%, onset <3 h) initiated degradation before genome activation. Despite those differences, degradation rates of reporters span a similar range as native maternal transcripts (Fig. 2F).

Analysis of MPRA sequences with respect to their decay dynamics (Supplementary Table S4 and Fig. 4E), confirmed and expanded the genomic observations. It confirmed the known [54, 55] destabilizing activity of miR-430 seeds (GCACUU) and AU-rich element (ARE, e.g. UAUUUAU), and stabilizing effect of UUAG signals. MPRA also confirmed several CPEs (e.g. UUUUUU) to promote polyadenylation (evident by higher initial levels and “polyA+ fraction”, Supplementary Fig. S5D), and linked them to delayed degradation onset. A-rich signals (e.g. AAAAA) were linked to early deadenylation (evident by lower initial levels, deA rates and “polyA+ fraction”), further supporting their suggested role. Finally, C-rich signals were linked to shorter tails prior to genome activation (Supplementary Fig. S5D), as previously identified [54, 55].

These results confirm our genomic observations also on a set of synthetic mRNA reporters. Reporters recapitulated the regulatory kinetics of maternal mRNAs in-vivo, with some changes that may reflect regulatory differences between in vitro synthesized and native mRNAs. MPRA analysis confirmed the activity of 3′UTR sequence elements, with enhanced sensitivity for effects that may have been obscured in the genomic analysis.

Maternal mRNA degradation dynamics is scaled by species’ developmental pace

Next, we used QUANTA to compare maternal mRNA clearance across species, aiming to test whether the principles of maternal mRNA regulation observed in zebrafish are conserved in other vertebrates. Because species vary widely in the timing of biological processes such as development and cell-cycle [10], it remains unclear whether maternal mRNA degradation follows a conserved timescale or is tuned to each organism’s developmental pace.

We analyzed temporal total-RNA and polyA+ RNA-seq datasets of frog (Xenopus tropicalis) embryos [79, 80] during the first 14 h of their development (end of gastrulation, as in zebrafish), and of mouse [81–85] and human [85–89] pre-implantation embryos through the morula stage (“Materials and methods” section). Both zebrafish and frog are well-established egg-laying models with both conserved and divergent developmental features [90, 91], while parallels between mouse and human embryos are also well documented [92–97]. For each organism, we applied QUANTA to identify exclusively maternal genes (Supplementary Tables S1 and S5, Supplementary Fig. S6A and B) and analyze their regulatory dynamics (Fig. 5A).

Figure 5.

Maternal mRNA expression patterns in embryos of different species. (A) Distribution (y-axis, % of maternal genes) of per-gene average of parameter values (x-axis) estimated by models in embryonic RNA-Seq datasets of four different species (red: frog, yellow: mouse, purple: human, black: zebrafish; solid: “degradation” model, dashed: “polyA+” model). Left to right: initial RNA level (normalized FPKM, log₂), deadnylation rate (1/h), onset/switch time (h), and decay rate (1/h). Bar graph: fraction of maternal genes (y-axis) with either a negative deadenylation rate (increase in polyA+ fraction, light gray) or a positive deadenylation rate (decrease in polyA+ fraction, dark gray) for each organism (x-axis). A subset of 20% of genes with rates closest to zero (in absolute value) was defined as non-changing (white). (B) Mean values (y-axis) of predicted initial expression levels for maternal genes of four species (red: frog, yellow: mouse, purple: human, black: zebrafish), in total-RNA (top) or polyA+ (bottom) datasets. (C) Mean values (y-axis) of regulatory parameters predicted for maternal genes of four species (red: frog, yellow: mouse, purple: human, black: zebrafish). Mean value (hours) and scale relative to zebrafish are indicated on bars. (D) Examples of temporal (x-axis, scaled time, arbitrary units) expression levels (y-axis, normalized FPKM, log₂) measured in total-RNA datasets of four different species (red: frog, yellow: mouse, purple: human, black: zebrafish) for specific maternal genes (names are indicated on graphs). Names of species that reject a unified model with zebrafish are noted on bottom. (E and F) Motif logos representing k-mers associated with positive (right) or negative (left) effect on kinetic parameters. Only motifs that are associated with 10% or more of k-mer positions are shown: (E) frog, (F) mouse, and human.

Across all four species, we found that several core features of maternal mRNA decay programs were conserved. In polyA+ datasets, the “polyA+ model” consistently outperformed the simpler “degradation” model, while total-RNA levels retained the “degradation model” (Supplementary Figs S6C and S7). Maternal mRNAs were typically deposited with short poly(A) tails, as inferred by a 3-fold lower average initial polyA+ levels compared to total-RNA levels (Fig. 5B). Transcripts were predominantly polyadenylated after fertilization in all species (Fig. 5A), as reported in zebrafish [23], mice [27], and humans [28]. However, mammalian embryos showed weaker polyadenylation activity: while 68%–69% of maternal transcripts increased in polyA+ levels (α < 0) in zebrafish and frog, only 38%–48% showed similar trends in mouse and human. In all species, the onset of degradation aligned closely with genome activation, and maternal mRNA decay was minimal before this point but accelerated following genome activation (Fig. 5A).

While these regulatory principles were conserved, their timing scaled to match species’ developmental pace (Fig. 5C). The onset of maternal degradation aligned with each species’ zygotic genome activation time: 3.5 hpf zebrafish, 4.5 hpf frog, 17 hpf mouse, and 34 hpf human. Importantly, both deA and degradation rates were also tuned to match species-specific timelines. For example, maternal degradation rates in frog were on average 1.7-fold slower than zebrafish (frog mean half-life 1.4 h compared to 0.8 h in zebrafish, Fig. 5C), matching the longer time until completion of frog gastrulation (14 hpf compared to 10 hpf in zebrafish).

To directly compare degradation dynamics, we scaled the developmental time axis to each species’ overall pace (Supplementary Fig. S6D; “Materials and methods” section). Using this scaled time, we tested whether maternal genes followed a shared degradation trajectory across species. For each gene, we compared (using a likelihood ratio test) its species-specific degradation model to a simpler unified model with one degradation function fitted to scaled data of two organisms (“Materials and methods” section). Genes that retain the unified model suggest a developmentally scaled decay program (Supplementary Fig. S6E). This analysis revealed both conserved (e.g. btg4 and akap11; Fig. 5D) and divergent (e.g. h1m, dazl, cobll1a, and mtmr7a; Fig. 5D) degradation patterns. Globally, degradation rates of ∼50% of orthologs scaled within 1.5-fold of zebrafish rates (Supplementary Fig. S6F).

Together, these results reveal that maternal mRNA decay follows shared regulatory principles across vertebrates, while the timing of these programs is scaled to each organism’s developmental pace. This scaling allows conservation of regulatory logic even between species that differ by up to 10-fold in their overall developmental pace. However, we also identify a subset of transcripts that do not scale, hinting at species-specific kinetic tuning, and suggesting that scaling of maternal mRNA degradation can be adapted to species-specific needs.

Species 3′UTR motifs reflect conserved and specie-specific regulatory needs

We next examined the activity of 3′UTR regulatory elements across vertebrate species (Supplementary Table S6), uncovering both conserved and divergent sequence elements.

In Xenopus (frog), regulatory elements (Fig. 5E) largely mirrored those observed in zebrafish (Fig. 3). miR-430 seed sequences (GCACUU) and AU-rich elements (e.g. AUUUUA) accelerated maternal mRNA decay in both species. Similarly, CPEs (e.g. UUUUA and UUUUU) were linked to initially longer poly(A) tails (evident by higher initial polyA+ levels and deA rates, delayed degradation onset). Additionally, we identified frog-specific stabilization by GC-rich sequences (e.g. CCCCG), potentially linked to their reported polyadenylation activity in frog [98].

Mammalian 3′UTR elements (Fig. 5F) were primarily associated with mRNA stabilization, and no clear degradation-inducing signals were detected. Both cytoplasmic (e.g. UUUUA) and canonical (e.g. UAAUAA) PASs promoted poly(A) tail extension (evident by higher initial polyA+ levels and deA rates) in both mouse and human embryos, consistent with other species. C-rich motifs (e.g. CCCCA) correlated with shorter poly(A) tails (evident by lower initial polyA+ levels and deA rates), similar to their effects in zebrafish (Supplementary Fig. S5D).

These findings highlight a conserved role for PASs in promoting poly(A) tail extension across vertebrates. In contrast, degradation-inducing elements are conserved primarily in fast-developing species (zebrafish and frog) and appear less prominent in mammals. This divergence may reflect species-specific regulatory needs, where slower mammalian development does not necessitate rapid maternal mRNA clearance. In those species, other mechanisms (e.g. m6A or codon usage) may contribute to maternal mRNA clearance.

Poly(A)-mediated buffering enables intra-species scaling of mRNA degradation dynamics

While inter-species scaling of developmental gene expression programs has been attributed to differences in metabolic [92, 93] and biochemical [96, 97] rates, the mechanisms that support or constrain intra-species scaling are less understood. Within a species, developmental timing can shift due to environmental factors such as growth temperature, but whether gene regulatory programs scale proportionally is not clear. Scaling could be challenged by differences in how regulatory enzymes respond to metabolic changes, potentially disrupting their coordination. As changes in metabolic rates were shown to enhance the sensitivity of developmental programs to disruptions in mRNA decay [94, 95], degradation pathways may be particularly vulnerable, and signal might exhibit metabolic rate dependent activities.

To investigate this, we performed MPRA on zebrafish at different growth temperature. Embryos were grown at 22°C and 34°C (Fig. 6A), conditions that maintain normal development [99] but differ in developmental pace by ∼2-fold (Fig. S8A). We scaled sample times in 22°C or 34°C experiments by matching developmental stages to standard 28°C, and compared temperature-specific degradation models to a unified model that scales between temperatures, as we did between organisms (“Materials and methods” section).

Figure 6.

Analysis of temperature dependent scaling of mRNA decay in zebrafish. (A) One-cell stage zebrafish embryos were collected and injected with MPRA reporters at a standard growth temperature. Following injection, embryos were grown at either 34°C (magenta) or 22°C (cyan) and analyzed over time. High temperature led to ∼2 h faster development, while low temperature led to ∼5 h slower development. To allow a direct comparison, we scaled sample times in 22°C or 34°C experiments to standard 28°C by their developmental stages. (B and D) Left: fold-change of reporter levels between 1 and 9 hpf (scaled times), in 22°C (y-axis, log₂) or 34°C (x-axis, log₂) experiments. For A₀ reporters at 34°C, 1 h expression is the average of 0.5 and 1.5 h samples. For all reporters, 9 h expression is the average of the last two samples. Pearson R and number of reporters (n) are indicated. Red line represents the average fold-difference between axes. Color represents density (blue = low density; yellow = high density). Right: Distribution (y-axis, % of reporters) of fitted reporter onset times (x-axis, h) and degradation rates (x-axis, 1/h) estimated by the “degradation” model of embryos grown at either 22°C (cyan) or 34°C (magenta). Panel B analyzes A₀ reporters (injected without a poly(A) tail), and panel D analyzes A₄₀ reporters (injected with a 40A tail). (C andE) Left: bar plot representing number of reporters that retain the scaled degradation model between temperatures (gray) and those rejecting it in favor of an alternative independent degradation model (yellow). Number of reporters is indicated on bar. Right: examples of temporal (x-axis, scaled time, h) expression levels (y-axis, normalized FPKM, log₂) of individual reporters at 22°C (cyan) or 34°C (magenta) growth temperatures. Top line reporters retain the “scaled” model, while bottom line reporters reject it. Panel C analyzes A₀ reporters, and panel E analyzes A₄₀ reporters. (F) Comparison of reporters that retain the “scaled” degradation model when injected either without (A₀, blue) or with (A₄₀, black) poly(A) tail. Most reporters in the A₀ group overlap the A₄₀ group.

Degradation of our original (non-adenylated, A₀) reporters differed significantly between temperatures. Although fold-decrease in reporter levels from 1 to 9 hpf (scaled time) was highly correlated across temperatures, it was 1.7-fold greater at 34°C than 22°C (Fig. 6B). Kinetic modeling showed that degradation onset occurred earlier at 34°C (average onset 0.9 h versus 5.1 h, Fig. 6B), whereas its rate was faster at 22°C (average half-life 1.5 h versus 1.2 h, Fig. 6B). These differences reflect slower 22°C degradation prior to genome activation and faster after activation, which partially compensate for the early stability. Faster 22°C degradation after genome activation might stem from its longer absolute time. Still, only 8% of reporters retained a unified, scaled kinetic model between temperatures (Fig. 6C), indicating that simple temporal scaling cannot fully explain temperature-dependent degradation kinetics. The scaled reporters had overall slower degradation, which could help minimize differences both before and after genome activation (Supplementary Fig. S8B).

Given that polyadenylation of reporters slows down their degradation [54], we hypothesized that polyadenylated reporters would scale better between temperatures. By stabilizing transcripts, polyadenylation could dampen kinetic variation, and buffer temperature specific effects. We synthesized A₄₀ reporters with an initial 40A tail, matching the upper 2% of endogenous maternal tail lengths [23], and compared them to their A₀ counterparts. As expected, A₄₀ reporters degraded more slowly than A₀ reporters (Fig. 6D), and fold-decrease in their levels remained tightly correlated between temperatures. Notably, A₄₀ degradation at 34°C displayed bimodal onset: 40% of reporters degraded early (average onset: 1.2 h, Fig. 6D), while 60% showed delayed degradation (average onset: 5.8 h, Fig. 6D), which was absent in A₀ reporters. Most A₄₀ reporters (64%) followed a unified, scaled degradation model across temperatures (Fig. 6F), including previously scaled A₀ reporters (Fig. 6F). Thus, adding a poly(A) tail improved scaling and expanded the set of reporters with temperature-independent degradation kinetics. Still, 36% of A₄₀ reporters did not scale, suggesting that other sequence or structural features also contribute. These had on average faster degradation (Supplementary Fig. S8B).

To further investigate the role of tail length in scaling, we used the “polyA+ ratio” as an indirect proxy for poly(A) tail length. We compared this metric across three groups: reporters that scaled with both A₄₀ and A₀ tails (7%), those that scaled only with A₄₀ tails (60%), and those that failed to scale (33%). Reporters that scaled with both A₄₀ and A₀ tails showed a faster and more sustained increase in “polyA+ ratio” (Fig. S8C), suggesting longer tail lengths and greater stability. Reporters that scaled only with A₄₀ tails had intermediate tail dynamics, consistent with their conditional scaling behavior.

These results show that temperature modulates mRNA degradation kinetics in zebrafish embryos. Faster development (34°C) promotes earlier degradation onset, while slower development (22°C) accelerates degradation after genome activation. Differences are most pronounced before genome activation. Polyadenylation could buffer these effects: stable mRNAs with longer poly(A) tails degrade more gradually and scale more consistently between temperatures. This suggests that mRNA stability contributes to developmental robustness under variable environmental conditions.

Tempo-sensitive function of 3′UTR motifs support intra-species degradation scaling

Given that stability improved temperature-dependent scaling, we hypothesized that 3′UTR sequence elements that affect stability will also affect scaling. We therefore analyzed 3′UTR motifs in our reporter library to identify signals associated with successful scaling, and tested their causal roles.

We found that stabilizing polyadenylation elements (e.g. UUUUUU) were significantly enriched in reporters that scaled across temperatures, while destabilizing C-rich motifs (e.g. CUCC) were enriched in non-scaled reporters (Fig. 7A). Further comparison revealed a stronger enrichment of stabilizing poly-U motifs in reporters that scaled with both A₄₀ and A₀ tails relative to those scaled only with A₄₀ tails (Fig. 7B), suggesting that they enhance scaling even in the absence of an initial poly(A) tail. Thus, stabilization by 3′UTR elements also improved scaling, as did stabilization by polyadenylation.

Figure 7.

MPRA analysis of 3′UTR elements that support temperature dependent scaling of mRNA degradation. (A andB) Volcano plots showing the effect size (difference in % of reporters with element in group, x-axis) and P-value (1% FDR, hypergeometric test, y-axis) of different sequence k-mers for their enrichment in groups of reporters. Colored dots represent k-mers that pass the P-value and effect size thresholds. In those cases, colors represent density (blue = low density; yellow = high density). k-mers that do not pass the thresholds are colored in grayscale by density. Representative top k-mers are indicated. (A) Test compares reporters containing a specific k-mer in each group to other analyzed reporters. Reporters were divided into three groups: scaled reporters in both A₄₀ and A₀ data (top left), scaled reporters in only A₄₀ data (bottom left) and reporters that did not scale in either dataset (top right). (B) Test compares scaled reporters in both A₄₀ and A₀ data to scaled reporters in only A₄₀ data. (C) Temporal (x-axis, hpf) relative expression levels (y-axis, log₂) of modified validation reporters relative to the base sequence (unmodified) at different conditions (top). Plots show data for each of four base sequences. Top to bottom: a control sequence, three base sequences containing a single miR-430, poly-C, or poly-U sequence site. Each of these base reporter sequences was modified to either a null version (existing site mutated, red), or addition of an element: miR-430 (blue), poly-C (purple), or poly-U (orange) sequence. (D) Maximal fold-changes in relative expression [as shown in panel (C)] across all sampled times (left; y-axis, log₂) or across times sampled prior to genome activation (4 hpf scaled time, right; y-axis, log₂). Colors and order as in panel (C). (E) Likelihood ratio test P-values for validation reporters (FDR corrected, color-scale) comparing between a unified scaled model and two independent models (rows: base sequence, columns: modifications of base sequence). Colorscale represents log₁₀(P-value): white = not scaled, red = scaled (n.s., P-value). (F) qRT-PCR measurements of temporal (x-axis, scaled time, h) expression levels (y-axis, normalized to qRT-PCR control and then to initial 1 h sample, log₂) of three maternally provided genes at 22°C (cyan) or 34°C (magenta) growth temperature. FDR corrected likelihood ratio test P-value between a unified scaled model and two independent models are noted.

To directly test the causal role of 3′UTR elements on scaling, we designed a set of validation reporters containing one of three regulatory motifs: stabilizing poly-U signals and destabilizing poly-C signals and miR-430 seeds (“Materials and methods” section). We included loss-of-function changes to mutate signals, gain-of-function changes to introduce new signals, and pairwise combinations of signals, and measured their effect on reporters’ kinetics. These confirmed that poly-U signals confer stability, while miR-430 and poly-C elements induce destabilization (Fig. 7C). However, the regulatory effect of these motifs depended on both temperature and poly(A) tail status: destabilizing elements were more potent in A₄₀ reporters and at lower temperatures, whereas stabilizing CPEs had limited effect under these conditions. Interestingly, unlike miR-430, which becomes active only after genome activation, the effects of poly-U and poly-C motifs were detectable before genome activation (Fig. 7D). This earlier activity could explain their contribution to temperature-dependent scaling (Fig. 7A and B).

We then compared how well a unified model that accounts for scaling across temperatures fit the data versus temperature-specific degradation models. As expected, validation reporters with A₀ tails consistently rejected the unified model and did not scale with temperature (Fig. 7E). In contrast, scaling of reporters with A₄₀ tails was influenced by their 3′UTR elements. Specifically, mutating the poly-U signal disrupted scaling, while adding it enhanced scaling (Fig. 7E). Conversely, mutating the poly-C signal improved scaling. These results highlight the causal role of 3′UTR sequences in tuning mRNA degradation to match developmental pace under different temperature conditions.

Together, these findings demonstrate that the activity of degradation-associated 3′UTR signals is modulated by temperature and polyadenylation status. These regulatory elements help couple mRNA degradation to developmental pace, thereby supporting the robustness of maternal transcriptome regulation across variable environmental conditions.

Temperature dependent scaling of maternal mRNA degradation dynamics

Finally, we tested whether degradation of native maternal mRNAs also scale with developmental pace at different growth temperatures, and whether this scaling varies between genes. Therefore, we examined three well-established maternal transcripts: buc, btg4, and bmp15, using qRT-PCR to measure their expression dynamics at 22°C and 34°C (Fig. 7F).

All three genes exhibited short half-lives (<22 min) and short initial poly(A) tails (<8 nt at 0 hpf [23]). Our model predicted that buc undergoes initial poly(A) extension while btg4 and bmp15 do not, consistent with their experimentally measured poly(A) tail length [23].

Kinetic modeling revealed that while degradation of buc transcripts scaled with developmental pace across temperatures, degradation of bmp15 and, to a lesser degree btg4, did not scale (Fig. 7F). These findings align with our reporter assay results, suggesting that mRNA stability and longer poly(A) tails improve temperature-dependent scaling.

Broad analysis of transcriptional shut-off programs by QUANTA

Finally, we demonstrate that QUANTA can be utilized to study transcriptional shut-off dynamics across diverse temporal RNA-seq datasets and in different contexts (Supplementary Table S7).

We first analyzed drug-induced transcriptional inhibition datasets in human cell lines [100] (Fig. 8A) to assess steady-state mRNA kinetics. Transcription was inhibited by actinomycin D, and total-RNA was sequenced following the treatment. Although drug treatment is expected to turn transcription off, some genes still show evidence of low-level transcription, with persistent pre-mRNA levels (e.g. GAPDH). These are classified as “combined” rather than “shut-off” genes. Indeed, different cells were shown to have different sensitivities to transcriptional inhibitor drugs, and cellular feedback loops may even lead to transcriptional activation of some genes following treatment [101]. Functional enrichment analysis revealed that “shut-off” genes were primarily involved in RNA metabolism and cell-cycle processes, while persistent genes were enriched for protein synthesis and ribosomal functions, possibly linked to a drug treatment induced stress response. Kinetic modeling of “shut-off” genes demonstrated a median mRNA half-life of ∼4 h, indicating a relatively high transcript stability. Sequence analysis of 3′UTRs (Fig. 8B) showed that AU-rich elements were associated with slower degradation and higher steady-state mRNA abundance, while GC-rich sequences correlated with faster degradation and lower abundance, patterns that were consistent also in mouse embryonic stem cells [9] (Fig. 8C). These findings are in line with known mechanisms of GC-rich mRNA destabilization by XRN1 in various human cell types [102].

Figure 8.

Broad analysis of transcriptional shut-off programs by QUANTA. (A) Analysis of drug-induced transcriptional inhibition in three human cell lines. Left: classification (x-axis: class) of genes (y-axis, fraction of genes) into shut-off, turn-on and combined expression. Number of genes and their factions are indicated on bars. Right: distribution (y-axis, % of shut-off genes) of per-gene onset/switch time (x-axis, h) and decay rate (x-axis, 1/h) estimated by QUANTA in three human cell lines. Bottom: examples of temporal (x-axis, hpf) pre-mRNA and mature mRNA expression levels (y-axis, normalized FPKM, log₂) of specific genes in all analyzed datasets. Gene names and average parameters of the fitted models are indicated. (B andC) Motif logos representing k-mers associated with positive (right) or negative (left) effect on kinetic parameters. Only motifs that are associated with 10% or more of k-mer positions are shown. (B) For in three human cell lines (C) For mouse embryonic stem cells. (D) Analysis of stimulus-induced cell-state transitions in LPS-stimulated mouse immune DCs (left) and serum-stimulated human T98G glioblastoma cells (right). Left graph shows classification (x-axis: class) of genes (y-axis, fraction of genes) into shut-off, turn-on and combined expression. Number of genes and their factions are indicated on bars. Right graph shows distribution (y-axis, % of shut-off genes) of per-gene mRNA half-lives (x-axis, h) estimated by QUANTA. (E) Motif logos representing k-mers associated with positive (right) or negative (left) effect on kinetic parameters. Only motifs that are associated with 10% or more of k-mer positions are shown. Left: LPS-stimulated mouse DCs. Right: serum-stimulated human T98G glioblastoma.

Next, we investigated transcriptional shut-off during stimulus-induced cell-state transitions (Fig. 8D). Specifically, we analyzed high-resolution total RNA-seq datasets from lipopolysaccharides (LPS)-stimulated mouse immune dendritic cells (DCs) [2, 103], and serum-stimulated human T98G glioblastoma cells (after serum starvation) [104]. In both systems, only a minority of genes were classified as “shut-off” (14% in DCs, 12% in T98G cells), reflecting selective transcriptional downregulation during state transitions. In DCs, “shut-off” genes were enriched for cell-cycle regulators and E2F1 targets, consistent with E2F1’s role in suppressing DC maturation [105]. In glioblastoma cells, “shut-off” genes were associated with organ development and NFYB targets. Estimated decay rates were faster on average than in steady state conditions (mean half-life 2.8 h in T98G cells and 1.8 h in DCs). Interestingly, sequence analysis in T98G cells revealed a reverse trend compared to steady-state decay: AU-rich sequences were now associated with faster degradation, while GC-rich sequences were linked to slower mRNA clearance (Fig. 8E), suggesting context dependent “shut-off” regulation.

These results illustrate the utility of QUANTA across diverse systems and biological processes to investigate transcriptional shut-off. The QUANTA degradation model provides a more focused and nuanced understanding of mRNA degradation patterns compared to global models, thus eliciting more subtle regulatory effects.

Discussion

In this study, we systematically dissect the kinetics and regulatory logic of mRNA degradation. Focusing on maternal mRNA degradation during early embryogenesis, we combine a quantitative temporal RNA-seq analysis framework (QUANTA) with MPRA to uncover how sequence elements, polyadenylation, and mRNA stability shape maternal expression programs across organisms and environmental conditions like temperature.

A scalable strategy to study mRNA decay components within gene regulatory programs

We present QUANTA, a general framework to quantify mRNA degradation and polyadenylation kinetics from temporal RNA-seq data. QUANTA presents important technical advances. Analysis relies on standard RNA-seq time-series data, minimizing both needed technical expertise and possible perturbations such as transcriptional inhibition [9, 38] or metabolic labeling [5, 39–41]. By focusing on transcriptionally silent mRNAs, it directly estimates degradation kinetics, reducing sensitivity to annotation errors, alternative isoform, or changes in splicing rates compared to methods that rely on inference from introns [43, 44] or SNPs [45]. Additionally, comparing total-RNA and polyA+ RNA [46, 47] within its kinetic models enables indirect estimation of poly(A) tail dynamics. QUANTA’s precise kinetic modeling improves sensitivity to identify cis-regulatory elements. Despite its strengths, several limitations should also be recognized. QUANTA relies on transcriptionally silent genes, which could be relatively few in some responses and limit the analysis, or may be affected by residual transcription. Its sensitivity to poly(A) tail changes may also be reduced when tails are uniformly long.

Nevertheless, as the importance of mRNA clearance within gene expression programs becomes increasingly clear [7, 9, 72], QUANTA offers a powerful tool to explore mRNA decay and enhance discovery of its unique roles and biological functionality. With the rise of single-cell total-RNA-seq methods [106], QUANTA principles could be combined with pseudo-time analysis to also uncover cell-type specific regulatory dynamics.

Dissecting 3′UTR regulation with MPRA

We further leverage QUANTA by developing a compatible MPRA to systematically identify cis-elements controlling mRNA stability. This confirmed known elements [14, 18, 56, 107] and revealed novel regulatory features. Of particular interest is the biphasic behavior of U-rich motifs: stabilizing early, then promoting decay. This duality may coordinate the long-term storage of maternal mRNAs in oocytes with their fast degradation in embryos. In addition, A-rich signals within 3′UTRs emerge as early deadenylation signals, and have not been recognized before. These findings open avenues for mechanistic studies into how these elements function during development.

Developmental scaling of mRNA degradation across species and environmental conditions

Our data reveals that mRNA degradation kinetics scales with developmental pace across both species and environmental contexts.

Although species vary widely in the timing of biological processes such as development and cell-cycle [10], we show that maternal mRNA degradation is globally tuned to each organism’s developmental pace. Such scaling allows conservation of regulatory logic even between species that differ by up to 10-fold in their overall developmental pace. Although speculated, scaling of maternal expression programs between species was not previously described, and has conceptual and practical implications for translating developmental findings across species. Inter-species developmental scaling was mechanistically attributed to changes in metabolic [92, 93] and biochemical [96, 97] rates. We further show that scaling also relies on regulatory adaptations, which reflects species-specific needs. In particular, similar PASs are shared in all organisms, but decay elements appear specific to faster-developing species (e.g. zebrafish and frogs) where they drive rapid maternal mRNA clearance. Slower mammalian development might not necessitate such rapid maternal mRNA clearance. These findings align with studies that showed miRNAs are not required in mice [108] although they promote maternal clearance in fast-developing species [18]. Fast maternal clearance may also help to eliminate prevalent (30%–35%) but possibly non-functional [12] maternal deposition in egg-laying animals, while a more selective maternal deposition in mammalians (18%) could tolerate slower degradation. Thus, evolutionary differences in developmental tempo are matched by distinct mRNA regulatory strategies, enabling species to balance speed, precision, and robustness in early development.

Developmental timing can also shift within species due to environmental factors such as growth temperature, nutrition etc. In some cases different developmental outcomes are desired in changing growth conditions, for example allowing annual killifish embryos to enter or escape diapause [109]. But often, developmental outcomes remain robust to such changes. Zebrafish, for example, preserve normal developmental outcomes despite faster pace at higher temperatures. Although changes in metabolic and biochemical rates [92, 93] can proportionally scale gene expression programs, we show that in zebrafish embryos, growth temperature modulates maternal degradation kinetics in a way that is not always scaled with its effect on developmental pace. Elevated temperatures accelerate early degradation, but cooler conditions only partly speed up degradation after genome activation, leading to proportionally slower degradation at low temperatures. Development at high metabolic rates was shown to be more sensitive to mRNA decay [94, 95]. Enhanced degradation might help reduce developmental errors at high rates, while slower development might not necessitate such strong clearance dynamics. These temperature effects could disrupt developmental outcomes unless buffered, and limit its robustness under environmental variability. Both polyadenylation and 3′UTR signals offer mechanisms to buffer these effects, by stabilizing transcripts and improving scaling between temperatures. The activity of 3′UTR decay signals is modulated by temperature and poly(A) tail length, a plasticity that enables adaptive tuning of the maternal transcriptome, helping to ensure developmental robustness despite environmental variability.

Our study establishes QUANTA as a broadly applicable approach to dissect mRNA decay and its regulatory logic. It reveals a quantitative and scalable logic for maternal mRNA clearance, linking sequence-encoded regulatory elements to organismal developmental pace. Our publicly available resource (https://rabanilab.shinyapps.io/MZT_data) offers a platform for further exploration of maternal transcriptome dynamics across species. Together with our MPRA and temperature-scaling analysis, we uncover rules to tune maternal mRNA regulation that are adaptable, and responsive to both genetic and environmental contexts. Our results have broad implications for understanding developmental robustness at changing environments and translating developmental findings across species.

Supplementary data

Supplementary data is available at NAR online.

Conflict of interest

None declared.

Funding

This research was supported by the European Research Council Horizon 2020 (grant 852451 to M.R.) and by the Israel National Science Foundation (grant 1176/21 to M.R.). Funding to pay the Open Access publication charges for this article was provided by the European Research Council Horizon 2020.

Data availability

MPRA sequencing data generated in this study have been deposited in the NCBI Gene Expression Omnibus, under accessions GSE266357. The QUANTA code is openly available at https://github.com/rabanilab/QUANTA. A snapshot of all analysis code at time of submission is additionally archived at Zenodo (https://doi.org/10.5281/zenodo.15619003).