MicroRNA-mediated network redundancy is constrained by purifying selection and contributes to expression robustness in Drosophila melanogaster

Abstract

MicroRNAs (miRNAs) are post-transcriptional, non-coding regulatory RNAs that function coordinately with transcription factors (TFs) in gene regulatory networks. TFs and their targets are often co-regulated by miRNAs, forming composite feedforward circuits (cFFCs) with varying degrees of redundancy, primarily mediated by miRNAs. However, the maintenance of miRNA-mediated regulatory redundancy and its impact on gene expression evolution remain elusive. By integrating ChIP-seq data from ENCODE and miRNA targeting from TargetScanFly, we quantified miRNA-mediated cFFC redundancy in Drosophila melanogaster embryos and larvae, revealing more than three quarters of miRNA targets are involved in redundant cFFCs. Higher cFFC redundancy, where more miRNAs target the same gene within a cFFC, is correlated with stronger purifying selection, reduced expression divergence between species, and increased expression stability under heat shock stress. Redundant cFFCs primarily regulate older or broadly expressed young genes. These findings highlight the role of miRNA-mediated cFFC redundancy in enhancing gene expression robustness through natural selection.

Network analysis reveals miRNAs mediate high redundancy of composite feedforward circuits (cFFCs) in Drosophila, correlating with strong purifying selection, reduced interspecies expression divergence, and increased stability under stress.

Introduction

In animals, a single miRNA can target hundreds of mRNAs, typically exerting modest effects on these target genes at the post-transcriptional level¹. MiRNAs are highly abundant in animals, with an organism’s genome containing hundreds to thousands of miRNA genes². Approximately 30~60% of coding genes are targeted by miRNAs, forming intricate gene regulation networks^3,4. The high conservation of most miRNAs and their target sites underscores their significant biological roles⁴. However, miRNA mediated repression is generally subtle. For instance, even highly expressed miRNAs typically reduce the expression of many targets by less than 50% at both the transcriptome and protein levels^5–7. These observations raise questions about the true functional roles of miRNA targeting. Accumulated evidence indicates that the subtle repression by miRNAs might cumulatively be effective or play crucial roles by regulating multiple functionally relevant targets^8–10. Solid evidence also supports the notion that subtle repression is vital for the canalization of phenotypes and the entire transcriptome^9,11. Therefore, the functionality of miRNA targeting should be comprehensively understood, particularly in the context of gene regulatory networks or loops.

Transcription factors (TFs) constitute another extensively studied class of gene regulators within animal networks. Similar to miRNAs, TFs can regulate a broad spectrum of target genes. The primary distinction lies in their mode of operation: TFs function at the transcriptional level, modulating the activities of target genes through either activation or repression¹². It is widely acknowledged that many TFs and miRNAs exhibit substantial conservation across extensive evolutionary distance^13,14. However, within the animal kingdom, miRNA binding sites are typically less conserved than the miRNAs themselves over these prolonged evolutionary periods; for example, some target sites undergo rapid turnover across species^15–17. Similarly, TF binding sites, identifiable through comprehensive ChIP-seq analysis, also display patterns of high variability^18,19. These findings collectively underscore the dynamic nature of gene regulation mechanisms, characterized by the swift acquisition and loss of both miRNA and TF binding sites throughout evolution.

In complex multicellular organisms, TFs and miRNAs do not operate in isolation; rather, they collaborate to form extensive networks of synergistic and interactive regulation^12,20. Network motifs composed by TFs and miRNAs, such as feedforward and feedback circuits, frequently recur in gene regulatory networks (GRNs) to regulate the expression of target genes^12,21–24. A common form of miRNA-TF interplay motif consists of one or more miRNAs, an intermediary TF, and a target gene²⁵. In this motif, the miRNAs regulate the intermediary TF, while both the TF and the miRNAs jointly regulate the target gene, forming a feedforward circuit (Fig. 1a, b). Known as a composite feedforward circuit (cFFC), this structure has garnered significant scientific interest for its integration of both transcriptional and post-transcriptional regulation, contributing to gene regulatory networks. While both miRNAs and TFs can influence cFFC redundancy, previous studies indicate that miRNA-derived cFFC redundancy contributes more significantly than TF-derived cFFC redundancy to overall cFFC redundancy²⁵. Therefore, our study focuses exclusively on miRNA-derived cFFC redundancy. A cFFC with only one miRNA is termed non-redundant (Fig. 1a), while one with multiple miRNAs constitutes a redundant cFFC (Fig. 1b). It is important to note that multiple miRNAs from a conserved family are treated as one miRNA when calculating redundancy due to their identical seed sequences and presumed targeting of the same genes. These types of miRNA-mediated cFFC redundancies, which are overrepresented in GRNs, underscore their potential importance in systems stability^25,26.

Fig. 1. Schematic of miRNA-mediated non-redundant and redundant composite feed-forward circuits (cFFC).

a MiRNA-derived non-redundant cFFCs, a single miRNA regulates the transcription factor (TF), which in turn regulates the same target gene (TG). This can involve either one TF (A) or multiple TFs (B). b MiRNA-mediated redundant cFFCs, where multiple miRNAs co-regulate the same TF-TG interaction (C), or multiple miRNA-TF pairs co-regulated the same TG (D). c Table showing the number of targets, miRNAs and cFFCs identified in wandering third instar larva (L3) and embryonic GRN, using conserved miRNAs (families) that co-expressed with the TF expression window. For detailed information, refer to the Methods section.

In a redundant cFFC, multiple miRNAs targeting the same gene represent functional redundancy, a form of genetic redundancy particularly related to genetic or mutational robustness^27,28. Essentially, miRNA-mediated redundancy confers robustness in cFFC regulation against loss-of-function mutations that disrupt the targeting of individual miRNAs. This robustness may arise either as a byproduct of neutral evolution or evolve directly in response to natural selection²⁹. In the former scenario, the robustness serves merely as a byproduct of system architecture, allowing the regulation of multiple miRNAs on the same gene to relax selective constraints and reduce the evolutionary conservation of redundant cFFCs. In the latter scenario, robustness conferred by miRNA-mediated redundancy provides an evolutionary advantage against environmental or genetic perturbations, a phenomenon known as canalization³⁰; consequently, the loss of redundancy leads to decreased fitness^27,28. Redundant cFFCs are expected to exhibit lower levels of expression variation in response to perturbations compared to non-redundant cFFCs. The role of miRNAs in introducing redundancy to the GRN suggests that the redundancy observed in cFFCs is biologically significant, indicating that such redundancy may contribute to the evolutionary conservation to these circuits²⁵. This apparent paradox raises the fundamental question of whether redundant cFFCs primarily serve to stabilize or modulate gene expression. Therefore, the broader implications of this redundancy within regulatory circuits, both in terms of their functional roles and evolutionary significance, remain unclear.

In this study, we constructed two GRNs in Drosophila melanogaster embryos and larvae, utilizing ChIP-seq and miRNA target prediction data to identify miRNA-mediated cFFC network motifs. To understand the regulatory role of network redundancy in gene expression, we examined selective constraints by analyzing the conservation and polymorphisms of cFFCs with varying degrees of redundancy. Specifically, we explored how cFFC redundancies contribute to the stabilization and interspecies divergence of target gene expression. To gain deeper insights into the roles of highly redundant cFFCs in system stability and long-term evolution in gene expression regulation, we analyzed the age distribution of target genes associated with varying levels of cFFC redundancy, as well as their expression breadth across different tissues. Our findings suggest that maintaining a certain degree of miRNA-mediated cFFC redundancies in GRNs may enhance system stability and provide evolutionary buffering for gene expression.

Results

Genome-wide identification of redundant and non-redundant miRNA-mediated cFFCs in Drosophila embryos

The spatiotemporal regulation of miRNAs during development have been widely and precisely studied^31–33. Redundant functions of miRNAs often arises from multiples tissues/organs or different developmental stages⁸. Therefore, it is crucial to confine the regulation to specific developmental stages when investigating the functions and evolutionary trajectories of miRNA-mediated redundancies within feedforward loops. In this study, we constructed two GRNs in Drosophila embryos and wandering third instar larvae (L3) using ChIP-seq data for well-characterized TFs available from the ENCODE database^34–38. We collected ChIP-seq data for 499 TFs from 536 experiments performed at embryonic stages (Supplementary Data 1) and data for 40 TFs from 48 experiments at the wandering third instar larval stage (Supplementary Data 2). We then cross-referenced the ChIP-seq data with RNA-seq data available from the GEO database (https://www.ncbi.nlm.nih.gov/geo/) to ensure the active expression of these TFs during their respective developmental stages (See Methods; Supplementary Table 1 and Text 1). This approach allowed us to identify 341 TFs expressed in embryos and 31 TFs expressed in larvae, which were then used to construct the embryonic and larval transcriptional regulatory networks. A TF-target interaction was identified when the binding peak of the TF was located within 1.5 kb upstream to 0.5 kb downstream of a gene body.

Next, we obtained miRNA regulatory data from TargetScanFly (Release 7.2), incorporating predictions from 91 conserved miRNA families^39,40. Additionally, we collected small RNA-seq data in embryos and third instar larvae from the GEO database (Supplementary Table 2) to identify miRNAs that co-expressed with TFs during the embryonic or L3 stages¹⁵ (see Methods and Supplementary Text 1). Supplementary Figs. 1 and 2 illustrate the expression profiles of TFs and miRNAs, respectively. We treated each conserved miRNA family as a single targeting entity, calculating its total expressions as the sum of its members, as family members are often co-expressed and share the same seed sequence, targeting the same sites. By integrating these TF regulations with miRNA targeting, we constructed gene regulatory networks (GRNs) for embryos and wandering third instar larvae (Fig. 1). In embryos, we identified 2144 targets regulated by non-redundant cFFCs, referred to as non-redundant cFFC targets, and 8408 targets regulated by redundant cFFCs, referred to as redundant cFFC targets (Fig. 1c). A similar ratio of non-redundant cFFC targets to redundant cFFC targets was observed in the larval GRN (Fig. 1c). Our analysis indicated an average of 7.24 miRNAs per redundant cFFC target in the embryonic GRN and 8.00 miRNA per redundant cFFC target in the larval GRN (Fig. 1). The maintenance of such redundancy, albeit costly, suggests that a moderate level of redundancy is typical in an evolutionarily stable system. This observation highlights the significance of redundancy in target gene regulation through cFFC formation and its integration with TF networks. Additionally, we identified cFFCs for non-conserved miRNA families (Supplementary Table 3 and Supplementary Fig. 3) and found that non-conserved miRNAs exhibit significantly lower redundancy compared to conserved miRNAs (Supplementary Table 4), despite their higher abundance (Supplementary Text 2).

Purifying selection on miRNA target sites in both redundant and non-redundant cFFCs

Previous analyses suggested that cFFC redundancy may be evolutionarily unstable if one circuit simply serves as a backup for another; however, it may be stable if the “redundant” circuits confer robustness that is selective advantageous^27,28. The former scenario represents functional redundancy likely to be relaxed from purifying selection, while the latter scenario represents genetic canalization subject to selective constraints. By detecting signals of nature selection, it may be possible to determine whether and how natural selection maintains such miRNA-mediated cFFC redundancy for evolutionary advantages. Given that cFFC motifs were identified based on conserved miRNA families, mutations leading to altered cFFCs predominantly arise from changes in miRNA targeting sites. Therefore, we performed evolutionary genetics analyses on miRNA target sites involved in cFFCs with varying degrees of redundancy to detect natural selection acting on these redundancies. Target sites with higher miRNA-mediated redundancy are expected to experience stronger purifying selection. We categorized the miRNA target sites into cFFC and non-cFFC sites based on their involvement in cFFC regulation. The cFFC sites were further classified into non-redundant, low-redundant (<7.24 miRNAs per target in the embryonic GRN and <8.00 in the larval GRN), and high-redundant (≥7.24 miRNAs per target in the embryonic GRN and ≥8.00 in the larval GRN) categories, depending on the varying levels of miRNA-mediated redundancy.

We detected purifying selection on miRNA-mediated cFFC redundancies at both the between-species and within-species levels. Between-species conservation was assessed using probability of preferentially conserved targeting (P_CT) values from TargetScan⁴⁰. Sites subject to stronger purifying selection are expected to be conserved across species and show higher P_CT values, reflecting the selective maintenance of miRNA targeting⁴. cFFC sites exhibited higher P_CT values (median: 0.28 in the embryonic GRN and 0.27 in the larval GRN) than non-cFFC sites (median: 0.18 in the embryonic GRN and 0.15 in the larval GRN) in both the embryonic and larval GRNs (Mann–Whitney U tests, both P-value < 0.05 s; Fig. 2a, b). Additionally, we found a significant association between levels of cFFC redundancy and P_CT values (ANOVA tests, all P-values < 0.05), with higher P_CT values correlating with increased cFFC redundancy in both the embryonic GRN (Fig. 2a) and the larval GRN (Fig. 2b). These findings suggest that the cFFC redundancy is maintained by natural selection for certain biological functions.

Fig. 2. Sequence evolution of miRNA targeting sites with different levels of miRNA-mediated cFFC redundancy.

a, b Conservations (P_CT values by TargetScan 7.2) of miRNA sites with different regulation contexts in embryonic (a) and wandering third instar larval GRN (b). Comparisons were performed between sites not involved in cFFC regulations (non-cFFC sites) and those within cFFC regulations (cFFC sites), or within cFFC sites with different levels of miRNA-derived redundancy (non-redundant, low-redundant and high-redundant cFFC sites). Tajima’s D values of miRNA targeting sites with different regulation contexts in embryonic (c) and wandering third instar larval GRN (d). Population data was adopted from Dai et al.⁴² and method was from Tajima⁴¹. Fourfold sites were used as the neutral control. Error bars show the 1000 bootstrapping values. Comparisons were performed as in a, with Mann–Whitney U tests performed. ANOVAs were performed within cFFC sites. All significance levels are showed as follows: P-value < 0.001***, < 0.01**, <0.05* after Benjamini & Hochberg correction⁵⁷ for multiple testing when applicable.

Within-species selective constrains were inferred from Tajima’s D values⁴¹. The stronger the purifying selection, the lower Tajima’s D is expected to be. Using D. simulans as an outgroup, we calculated the derived allele frequencies and Tajima’s D for miRNA target sites from different strains of D. melanogaster using data and methods from Dai et al.⁴². Compared to the whole-genome fourfold sites (median Tajima’s D value: −0.418), which are typically used as neutral controls, all miRNA target sites exhibited lower Tajima’s D values, indicating purifying selection (Mann–Whitney U tests, all P-values < 0.05; Fig. 2c, d). The cFFC sites exhibited lower median Tajima’s D values (embryonic: −1.276; larval: −1.271) compared to non-cFFC sites (embryonic −1.152; larval: −1.079) (Mann–Whitney U tests, all P-values < 0.05; Fig. 2c, d). Additionally, we observed a general decline in Tajima’s D values with increasing levels of redundancy when comparing high-redundant, low-redundant, and non-redundant sites (embryonic: −1.356 vs. −1.172 vs. −0.920; larval: −1.365 vs. −1.126 vs. −0.922). These findings suggest a positive correlation between the degree of selective constraint and cFFC redundancies (one-way ANOVA, all P-values < 0.05; Fig. 2c, d) and show strong consistency across the two developmental GRNs, indicating enhanced purifying selection on cFFC sites (Fig. 2c, d). Taken together, our results suggest that cFFC redundancy is subject to selective constraints, and the degree of these constraints correlates with redundancy levels. A similar pattern was also observed for non-conserved miRNAs (Supplementary Text 2 and Supplementary Fig. 4).

We further analyzed the effects of miRNA targeting efficacy while examining purifying selection on target sites across different cFFC classes. Strong binding and effective miRNA repression depend on perfect or near-perfect base pairing of the seed region (positions 2–8 from the 5’ end of the miRNA). The targeting efficacy of miRNAs was inferred from various target types categorized by TargetScanFly 7.2⁴⁰, specifically the 7mer-a1, 7mer-m8, and 8mer site types. Our analysis revealed no enrichment of specific site types at any cFFC redundancy level in either the embryonic or larval GRNs (Chi-square tests, all P-values > 0.05; Supplementary Fig. 5a, b). ANOVA analysis indicated that all three factors — target site type (T), cFFC redundancy (R), and their interactions (T*R) — significantly influenced Tajima’s D values (ANOVA tests, all P-value < 0.05; Supplementary Fig. 5c, d), while only target site type (T) and cFFC redundancy (R) significantly affected P_CT values in both the embryonic and larval GRNs (ANOVA tests, both P-value < 0.05; Supplementary Fig. 5e, f). High-redundant 8mer sites exhibited the strongest selective constraint among all site type classes, as indicated by the lowest Tajimas’ D values and the highest P_CT values (Mann–Whitney U tests, all P-values < 0.05; Supplementary Fig. 5c–f). These results suggest that targeting efficacy may interact with cFFC redundancy to influence the strength of purifying selection on miRNA target sites.

The association between miRNA-mediated cFFC redundancy and expression stability

It is thought that miRNAs canalize gene expression through widespread, weak regulation^8,9,17. We hypothesize that miRNA-mediated cFFC redundancies may also contribute to system stability. To test this hypothesis, we first checked the distribution of 3’UTR length for targets with different redundancy levels, as cis-regulatory elements within the 3’UTR may influence mRNA stability^38–40. Genes with longer 3’UTRs are more likely to harbor additional cis-regulatory elements, such as miRNA target sites, thereby increasing the likelihood of their involvement in high-redundant cFFCs. Therefore, to minimize the potential impact of other unknown cis-regulatory elements, it is crucial to compare target genes with similar 3’UTR lengths but differing levels of cFFC redundancy. To achieve this, we selected a subset of targets with reasonable range of 3’UTR lengths and controlled for effects on gene expression stability by analyzing the 3’UTR length distribution of targets with varying cFFC redundancy in the larval GRN. We found that the high-redundant targets have the longest 3’UTRs, peaking at 1024~2048 bp, followed by low-redundant (peaking at 256~512 bp), non-redundant, and non-cFFC targets (peaking at 128~256 bp; Supplementary Fig. 6a). To further investigate the effects of 3’UTR length and cFFC redundancy on time-series expression changes in response to heat shock stress, we conducted a multi-way ANOVA. By restricting the 3’UTR lengths to a range of 128–2048 bp, which encompasses the most frequent peak lengths of cFFC targets, we found that the interaction effects between 3’UTR length and cFFC redundancy disappeared (Supplementary Fig. 6b). This suggests minimal co-effects on expression changes. Therefore, we used this subset of genes (2300 high-redundant, 4918 low-redundant, 1116 non-redundant, and 267 non-cFFC targets) for the subsequent analysis of gene expression robustness.

Next, we sought to investigate the functional consequences of cFFC redundancy on expression divergence between species. To do so, we sequenced the transcriptome of 0.5-h and 2-h embryos, which encompass the majority of embryonic ChIP-seq data, each with two biological replicates for both D. melanogaster and D. simulans. After mapping the raw reads to their respective genomes, we quantified gene expression levels as Transcripts Per Million (TPM). Expression divergence between species was calculated as one minus the Spearman’ correlation coefficient (1 - ρ) of the TPM values for one-to-one orthologous gene pairs between D. melanogaster and D. simulans, as described previously⁴³. To account for the impact of 3’UTR length, we classified target genes with varying levels of cFFC redundancy into three groups within a controlled range of 3’UTR length (128~2048 bp). During early embryogenesis in Drosophila, the maternal-to-zygotic transition involves the gradual replacement of maternally contributed mRNAs and proteins by zygotic gene products. To address the potential impact of maternal effects on TF regulation, we obtained a list of maternal genes from Ibarra-Morales et al.⁴⁴ and excluded them from our analysis of gene expression divergence. A general trend of decreased expression divergence with increasing target redundancy was observed in both embryonic stages when maternal genes were excluded (Fig. 3a; Mann–Whitney U tests, P-values < 0.05). High-redundant targets consistently exhibited the lowest expression divergence, whereas targets not involved in any cFFCs (non-cFFC targets) showed the highest expression divergence (ANOVA, P-values < 0.05). This pattern was also observed in 4- to 8-h embryos, suggesting a consistent conclusion (Fig. 3b). Our results indicate that cFFC redundancy negatively correlates with gene expression divergence between species.

Fig. 3. The association between miRNA-mediated cFFC redundancy and expression stability.

a Comparison of target gene expression divergence (1 - ρ) between D. melanogaster and D. simulans embryos under different regulation contexts: non-cFFC (targets that not in any regulation context of cFFCs), non-redundant cFFC, low-redundant cFFC and high-redundant cFFC in embryonic GRN. The maternal genes adopted from Ibarra-Morales et al.⁴⁴ were excluded from analyses using transcriptomes from 0.5-h and 2-h embryos. The samples are biologically independent, with n = 2. Mann–Whitney U tests were performed. b Same as a) but using transcriptomes from 4 to 8 h embryos. The samples are biologically independent, with n = 4 for D. melanogaster and n = 2 for D. simulans. c Time course analysis of expression changes of target genes identified in larval GRN in D. melanogaster larvae post heat shock stress (30 °C versus 25 °C, n = 3 biologically independent samples). The expression difference for each gene was calculated as the difference in mean TPM values in contrast between the treated and control samples, normalized by the overall average of TPM values across all samples in both conditions. Dashed lines represent simulated steady states following heat shock, modeled using damped oscillation equations (see Supplementary Methods). The boxplots show the distributions of expression differences, and the points represent the mean values. Mann–Whitney U tests were conducted between each pair of target types at every time point. These analyses were focused on genes with 3’UTR lengths ranging from 128 to 2048 bp. Significance levels are denoted as follows: *P-value < 0.05, **P-value < 0.01, ***P-value < 0.001, P-value > 0.05, n.s. (not significant) after Benjamini & Hochberg correction⁵⁷ for multiple testing when applicable.

Finally, the impact of cFFC redundancy on system stability can also be assessed under perturbation. In this scenario, targets with cFFC redundancy are expected to exhibit smaller variation compared to those without it, and are also likely to return to a steady state more rapidly once the perturbations are removed. We subjected D. melanogaster larvae to heat shock at 30 °C, and conducted RNA-seq analysis on both treated and control samples, each with three biological replicates, at 0, 3, 6, 9, and 12 h post-treatment (see Methods). The expression difference for each gene was calculated as the difference in mean TPM values in contrast between the treated and control samples, normalized by the overall average of TPM values across all samples in both conditions. We indeed observed that the high-redundant cFFC targets exhibited minimal expression differences at 0 hours (−0.050 for high-redundant vs. −0.143 for low-redundant, −0.207 for non-redundant, and −0.203 for non-cFFC targets; Mann–Whitney U tests, all P-values < 0.05 except for the comparison between the non-cFFC and non-redundant group; Fig. 3c). This time point reflects the extent of transcriptomic changes in response to heat shock stress and marks the beginning of the system’s recovery. The steady state of the system rebound were modeled using damped oscillation equations across different cFFC regulation contexts (see Supplementary Method and Text 3; Supplementary Fig. 7). While the precise time for these cFFC redundant targets to return to a steady state could not be determined, their expression fluctuated around the steady state at the 3, 6, 9, and 12-h time points post-treatment (Fig. 3c; Supplementary Fig. 7), suggesting a rebound time of less than 3 h.

MiRNA-mediated redundant cFFCs tend to target old genes or broadly expressed young genes

If miRNA-mediated cFFC redundancy is evolutionarily significant, we anticipate that the associated targets are more likely to be evolutionarily conserved and exhibit widespread expression. We first assessed the correlation between miRNA-mediated cFFC redundancy and gene ages. Genes were categorized into three age groups: Drosophilid (youngest), pre-Drosophilied (middle-aged), and pre-Bilateria (oldest), following the classification defined by Witt et al. and Kondo et al.^45,46. The proportion of genes in each age group across different levels of cFFC redundancy was calculated (Fig. 4a). When compared to the entire gene set as a background, non-redundant and non-cFFC targets were found to be enriched in the Drosophilid group (43.76% for non-redundant targets and 52.80% for non-cFFC targets; Fisher’s exact tests, P-value < 0.05; Fig. 4a). Conversely, targets in the low- and high-redundant targets were notably enriched in the pre-Bilateria group (accounting for 57.95% in low-redundant and 69.49% in high-redundant targets; Fisher’s exact tests, P-values < 0.05; Fig. 4a). Additionally, targets in high-redundant cFFCs had the highest proportion of pre-Bilateria genes compared to other target groups (69.49% vs. 57.95% for low-redundant, 45.68% for non-redundant, and 37.38% for non-cFFC targets; Fisher’s exact tests, P-value < 0.05). The trend of an increasing proportion of pre-Bilateria genes and a decreasing proportion of Drosophilid genes along with miRNA-mediate cFFC redundancies (Fig. 4a, ANOVA, P-values < 0.05) suggests that this type of redundancy leads to a preference in miRNA targeting towards older genes.

Fig. 4. The preference of redundant cFFCs for targeting old genes or broadly expressed young genes.

a Distribution of genes across different ages, grouped by their different regulation contexts. Fisher exact tests were performed between two adjacent redundancy groups, using the sum of the other two age groups as the control. b Boxplots displaying the tissue specificity (τ) values of genes with different regulation contexts. Mann-Whitney U tests were performed between every two groups. c Similar to (b), but genes grouped by ages, with one-way ANOVA P-values shown. d Two-way ANOVA analysis of the impact of cFFC redundancy, gene age, and their interaction effects on tissue specificity (τ). e Proportion of genes regulated by different cFFC contexts, grouped by tissue specificity. Fisher exact tests were performed between two adjacent τ groups, using the sum of the other three redundancy groups as the control. All significance levels are denoted as follows: *P < 0.05, **P < 0.01, ***P < 0.001, and n.s. (not significant) for P-value > 0.05 after Benjamini & Hochberg correction⁵⁷ for multiple testing when applicable. The regulatory contexts of target genes were identified in embryonic GRN.

Next, we used τ, as defined by Witt et al. to measure tissue expression specificity⁴⁶ and compared τ values across genes with varying levels of cFFC redundancy (Fig. 4b). Our results indicated that non-cFFC targets exhibited higher tissue specificity (median, τ = 0.798) compared to those regulated by cFFCs (median, τ = 0.545; Mann-Whitney U test, P-value < 0.05; Fig. 4b). Furthermore, targets in non-redundant cFFCs (median, τ = 0.747) demonstrated higher tissue specificity than those in redundant cFFCs (median, τ = 0.517; Mann-Whitney U test, P-value < 0.05; Fig. 4b). A significant difference was also observed between low-redundant (median, τ = 0.539) and high-redundant target gene groups (median, τ = 0.492; Mann-Whitney U test, P-value < 0.05; Fig. 4b). We also examined the impact of cFFC redundancies on gene expression specificity within different age groups, revealing significant differences between non-cFFC and cFFC targets in the pre-Bilateria and pre-Drosophilid groups (P-values < 0.05, Mann-Whitney U Test; Fig. 4c), and among various levels of cFFC redundancy within the Drosophilid group (ANOVA, P-value < 0.05; Fig. 4c and Supplementary Fig. 8, 9). This suggests a potential role of cFFC redundancy in influencing gene expression breadth, particularly in young genes. After accounting for the effects of gene ages and the interaction between cFFC redundancy and gene ages, we found that cFFC redundancy significantly affects gene expression breadth (ANOVA, P-value < 0.05; Fig. 4d). Additionally, we observed a notable enrichment of highly redundant genes among those most widely expressed (τ > 0.9; Fisher’s exact tests, P-value < 0.05, Fig. 4e). In conclusion, miRNA-mediated cFFC redundancy appears to be associated with widespread gene expression across tissues.

Discussion

The role of miRNA in fine-tuning or buffering gene expression has been extensively investigated^{8,11,17,26,33,47}. However, the evolutionary significance of miRNA-composed network motifs in gene regulation remains relatively unexplored^25,48. Various motifs containing miRNAs exist in gene regulatory networks (GRNs)^20,49, particularly those integrating with TFs, such as miRNA-mediated cFFCs²⁵. Due to the extensive targeting by miRNAs, redundancies in cFFCs are often maintained in GRNs^25,48. In our study, we observed an average cFFC redundancy value of 7.24–8.00 across embryonic and larval GRN, with 75.07–77.24% of target genes being regulated by redundant cFFCs. This value significantly deviates from the value of 17.9 reported in human by Iwama et al.²⁵, indicating potential variation of cFFC redundancy across different organisms. Factors contributing to this variation may include organism complexity, such as the number of targets, miRNAs, and TFs, as well as the developmental stages during which the GRN is analyzed and the methodologies used for network construction.

The retention of miRNA-mediated cFFC redundancy in GRNs has profound implications for the evolutionary dynamics of gene expression. The observed selective advantages in miRNA target sites within these redundant cFFCs imply a functional importance that transcends immediate regulatory processes, suggesting a strategic role for redundant cFFCs in conferring adaptability and enhancing the fitness of GRNs. This may contribute to the overall robustness and evolutionary success of biological systems. Further investigation into the specific molecular mechanisms underlying these advantages is crucial for a comprehensive understanding of the complex interaction between miRNAs and cFFCs in shaping the gene regulation dynamics. The correlation between miRNA-mediated cFFC redundancy and network robustness, along with reduced divergence between species (Fig. 3), underscore its role as a stabilizing element in the evolution of gene regulatory architectures. The observed conservation of redundant cFFCs across species indicates that their existence is not merely a result of neutral evolution (Fig. 2) but suggests a significant contribution to the adaptability and resilience of regulatory networks under evolutionary pressures.

Gene expression breadth, defined as the range of tissues, cell types, or conditions in which a gene is expressed, is crucial for shaping the adaptability and functional diversity of organisms across evolutionary timescales^50,51. Understanding the evolutionary dynamics of gene expression breadth, which is closely linked to protein evolution rates⁵¹, can provide insights into how organisms adapt to varied environments and lifestyles. Several factors, such as pleiotropy, significantly influence the modulation of gene expression breadth⁵². Genetic redundancies, such as the miRNA-mediated cFFC redundancy shown in our study, are suggested to be pleiotropic to some extent²⁷. These redundancies play distinct roles in regulating the expression of old genes and diversifying tissue expression patterns in young genes (Fig. 4c). Genes with broader expression profiles enhance the adaptive flexibility of the network, enabling effective responses to diverse conditions and evolutionary pressures. The interplay between miRNA-mediated cFFC redundancy, target expression breadth, and system stability provides fertile ground for exploring the temporal and spatial dynamics of gene expression during evolutionary processes. This knowledge is pivotal for unraveling how miRNAs and TFs collaboratively shape the molecular landscapes that dictate the evolution of biological complexity.

Methods

Constructing the miRNA mediated gene regulatory network

TFs and miRNAs represent the two primary layers of the gene regulatory network. To investigate TF regulations, we downloaded 536 embryonic and 48 wandering third instar larval ChIP-seq experiments of D. melanogaster from ENCODE^34–38 (https://www.encodeproject.org/, Supplementary Data 1 and 2), encompassing 499 TFs and 48 TFs, respectively. We then filtered the TFs based on their expression in the corresponding developmental stages. Eight embryonic RNA-seq datasets, including two embryonic cell lines (S2 and KC167) and six embryonic stages ranging from 0 to 24 h, were downloaded from the GEO database (https://www.ncbi.nlm.nih.gov/geo/, Supplementary Table 1). Additionally, ten RNA-seq samples at L3 stage were derived from heat shock experiments. TFs were retained if their expression was detected in more than three samples with transcripts per million (TPM) value greater than 10. Using IDR (Irreproducibility Discovery Rate) thresholded narrow peaks, we identified reproducible TF binding peaks located within 1.5 kb upstream to 500 bp downstream of the gene body (dm6) as indicative of TF-to-target regulation.

For miRNA regulations, a dataset containing all predicted targets of conserved miRNA families was obtained from TargetScanFly (Release 7.2)⁴⁰. Targets displaying only three seed match types (8mer, 7mer-a1, and 7mer-m8) were retained for the analysis. To ensure effective miRNA regulations, we only considered the miRNAs that are expressed in their corresponding developmental stages in alignment with TFs. Eleven embryonic small RNA-seq datasets, including two embryonic cell lines (S2 and KC167) and seven embryonic samples covering from 0 to 24 h (including samples at stage 1–5), as well as one third instar larval dataset, were downloaded from the GEO database (https://www.ncbi.nlm.nih.gov/geo/, Supplementary Table 2). Embryonic miRNAs were identified if their expression had reads per million (RPM) values greater than 10 in both the 2–10 h and 12–24 h embryos. The larval miRNAs were those with RPM values greater than 10 in the larval sample. By combining these TF and miRNA regulations, we generated two gene regulatory networks for embryos and third instar larvae, respectively.

Measuring miRNA expression during the embryonic stage

MiRNA expression data in embryos were collected from the GEO database, covering stages from 0 to 24 hours (Supplementary Table 2). Initially, adapter sequences were removed from the raw FASTQ reads using Cutadapt (version 2.10, -a CTGTAGGCACCATCAATC -O 3 -m 18). Subsequently, the reads were mapped to tRNA sequences (tRNAscan, http://gtrnadb.ucsc.edu/genomes/eukaryota/Dmela6/Dmela6-summary-all.html) and transposable element (TE) consensus sequences (Repbase r22.02, https://www.girinst.org/repbase/.) using Bowtie⁵³ (version 1.2.3, with default parameters to filter out tRNA-derived and TE-derived small RNA reads. Next, unmapped reads were processed through the miRDeep2 pipeline to assess miRNA expression levels⁵⁴. The FASTQ files were then converted into FASTA files using fastq2fasta.pl, followed by mapping reads to the D. melanogaster genome (flybase r6.34) with mapper.pl (parameters: -d -c -i -j -l 18 -m -o 8 -p genome_index -s reads_collapsed.fa -t reads_collapsed_vs_genome.arf). Subsequently, quantifier.pl was employed to quantify miRNA expression levels using the parameters: -p hairpin.fa -m mature.fa -r reads_collapsed.fa -y now -t dme -W. The miRNA hairpins and mature sequences were obtained from miRBase (release 22, http://www.mirbase.org). The expression level of each miRNA family was measured as the sum of RPM values of its respective family members.

Defining miRNA-mediated cFFC redundancies of target genes and their miRNA targeting sites

In cFFCs, the target genes are co-regulated by miRNA-TF regulations. As illustrated by Iwama et al.²⁵ in their Fig. 3, miRNA-mediated cFFC encompass four types of motifs: unitary cFFC, multiple TF-miRNA-mediated cFFCs, multiple miRNA-derived cFFCs, and multiple TF-derived cFFCs. Target genes can be concurrently regulated by various combinations of these cFFC motifs. Hence, the regulatory context of a target can be classified into unique-miRNA regulation and multiple-miRNA regulation based on the number of miRNAs involved in the cFFCs. Unique-miRNA regulation is classified as miRNA-mediated non-redundant cFFC, as mutations in the miRNA targeting site can results in the loss of cFFC regulation. In contrast, multiple-miRNA regulation is classified as miRNA-mediated redundant cFFCs. Consequently, in our study, targets in unitary cFFC (type A in Fig. 1a) and multiple TF-derived cFFCs (type B in Fig. 1a) were identified as non-redundant, while those in multiple miRNA-derived cFFCs (type C in Fig. 1b), or multiple TF-miRNA-derived cFFCs (type D in Fig. 1b), or combinations of these motifs were considered redundant (also classified as type D). Network redundancy was measured as the average number of miRNAs involved in redundant cFFCs per target²⁵. Genes exhibiting cFFC redundancies above the network redundancy threshold were classified as high-redundant, while those below this threshold were classified as low-redundant. In our analysis, miRNA target sites not involved in any cFFCs were categorized as non-cFFC sites.

Evolutionary genetic analysis of miRNA target sites in cFFC regulations

We assessed the conservation of miRNA target sites using their P_CT scores from TargetScan⁴, which determine the probability of preferentially conserved targeting. For population genetic analysis, we identified polymorphisms in miRNA targeting sites following the methods outlined by Dai et al.⁴², and calculated Tajima’s D values according to Tajima⁴¹, using four-fold sites as neutral controls.

Heat shock perturbation experiment

The heat shock perturbation experiment was conducted following the protocol outlined by Lu et al.¹¹. Briefly, flies (D. melanogaster w¹¹¹⁸) were reared at 25 °C on standard sugar-yeast-agar medium under 12:12 h light/dark cycles. To collect embryos, 6- to 9-day-old flies were placed in bottles containing grape juice agar plates for 3 days, and eggs laid on a fresh juice agar plate were collected within a 2-hour window. The heat shock treatment was administrated to first instar larva (1.5 days post egg-laying), maintaining them in an incubator at 30 °C for 2 days. Larvae (about 30 for each sample) from both the control and treatment groups were collected for total RNA extraction at 0, 3, 6, 9, and 12 h after perturbations. Total RNA was extracted using the TRIzol® Reagent, and ribo-depleted total-RNA libraries were then constructed and sequenced on an Illumina HiSeq2500 at BGI (http://www.genomics.cn/index). Three replicates were performed for each group.

Embryo collection and RNA-seq

Young male and virgin female of D. melanogaster (w¹¹¹⁸) and D. simulans (w⁵⁰¹) flies were collected and reared separately on fresh medium for 2–3 days at 25 °C. To collect early embryos, males and females of the same species were placed in small plastic bottles, each sealed with a grape-juice-agar plate. These agar plates, prepared to enhance egg-laying, consisted of 1000 ml of 100% grape juice, 34 grams of agar, and 5 ml of propionic acid. Fresh yeast, blended with distilled water to create a paste, was spread on the surface of each agar plate as a food source. Eggs (about 300 for each sample) deposited on the plate surfaces within 0.5 hours were collected and immediately lysed using TRIzol Reagent (Invitrogen) for RNA extraction (labeled as 0.5 h); alternatively, they were left at 25 °C for 1.5 hours before RNA extraction with TRIzol Reagent (Invitrogen) (labeled as 2 h). Then, ribo-depleted total-RNA libraries were evaluated using an Agilent 21100 Bioanalyzer (Agilent Technologies). Transcriptome libraries were prepared using the standard protocols of the Illumina TruSeq RNA Sample Prep kit and sequenced using an Illumina Genome Analyzer (Illumina, San Diego, CA, USA) at BGI (Shenzhen, China). Two replicates were performed for each group.

RNA-seq analysis, assessment of variation and divergence in gene expression

All RNA-seq data were mapped to their respective reference genomes (Ensembl BDGP6 for D. melanogaster and FlyBase r2.02 for D. simulans), using HISAT2 (version 2.1.0)⁵⁵. Gene expression levels were quantified as transcripts per million (TPM) using StringTie (version 1.3.3b, -e -A)⁵⁶. The gene expression were compared between the heat shock group (30 °C) and the control group (25 °C) at five time points (0, 3, 6, 9, and 12 h after perturbation). The expression difference for each gene was calculated as the difference in mean TPM values in contrast between the treated and control samples, normalized by the overall average of TPM values across all samples in both conditions.

To assess the expression divergence of targets under different cFFC regulatory context, RNA-seq was performed on 0.5-h and 2-h embryos of D. melanogaster and D. simulans. Following gene expression analysis, 11,063 one-to-one orthologous protein-coding genes between D. melanogaster and D. simulans (Flybase 2017_06, http://flybase.org) were identified. The expression divergence of these genes was estimated by 1 - ρ (Spearman’s correlation) according to the methods described by Coolon et al.⁴³. To estimate the distributions of expression divergence for each gene group, 1000 bootstrapping values were used.

Gene ages and expression breadth

The ages of target genes were obtained from Witt et al. and Kondo et al.^45,46 and categorized into three groups: Drosophilid (youngest), pre-Drosophilid (middle-aged), and pre-Bilateria (oldest), based on their evolutionary timeline. The gene expression breadth, quantified as τ, was obtained from Witt el al.⁴⁶. Comparative analyses were subsequently conducted among genes with varying levels of cFFC redundancies. ANOVA was performed to examine the interaction effects between cFFC redundancies and gene ages on expression breadth.

Statistics and reproducibility

Two-sided Mann-Whitney U tests were applied to assess differences between miRNA targeting site groups with varying cFFC redundancies (Fig. 2). Sample sizes were as follows: in the embryonic GRN, 19,153 for non-cFFC, 2195 for non-redundant, 22,862 for low-redundant, and 51,537 for high-redundant sites; in the larval GRN, 9543 for non-cFFC, 2077 for non-redundant, 26,185 for low-redundant, and 57,942 for high-redundant sites. The same test was also applied to compare target genes under different levels of cFFC redundancy: in embryonic GRN, 346 non-cFFC (16 maternal genes), 1232 non-redundant (153 maternal genes), 5113 low-redundant (674 maternal genes), and 1910 high-redundant cFFC targets (199 maternal genes) (Fig. 3a, b); and in the larval GRN, 267 non-cFFC, 1116 non-redundant, 4918 low-redundant, and 2300 high-redundant cFFC targets (Fig. 3c). In the heat shock experiment, approximately 30 larvae per sample were collected in triplicate, and about 300 eggs per sample were collected for each sample in duplicate for measuring gene expression divergence. The Benjamini & Hochberg correction⁵⁷ were used for multiple testing when applicable.

Two-sided Fisher’s exact tests were used to analyze the enrichment of gene age groups in specific cFFC redundancy groups (Fig. 4a), as well as the enrichment of cFFC redundancy in gene groups with tissue specificity (Fig. 4e). Two-way ANOVAs were performed to evaluate the effects of cFFC redundancy and gene age on tissue-specific gene expression (Fig. 4d), while one-way ANOVAs were used to examine the effects of cFFC redundancy on P_CT, Tajima’s D, and tissue specificity (Figs. 2 and 4b, c). Supplementary Figs. follow the same sample sizes.

Reporting summary

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

Author contributions

A.D., Y.L., T.T., and Z.L. conceived the idea and supervised the study. A.D. performed the analysis. W.L., X.Z. and X.M. did the experiments. A.D., T.T., and Z.L. interpreted the results and wrote the manuscript. All authors read and approved the final manuscript.

Peer review

Peer review information

Communications Biology thanks the anonymous reviewers for their contribution to the peer review of this work. Primary Handling Editors: Eirini Trompouki and David Favero. A peer review file is available.

Data availability

All raw RNA-sequencing data have been deposited in the National Genomics Data Center (NGDC; https://ngdc.cncb.ac.cn/), under the accession number PRJCA022918. This dataset includes transcriptomes from both D. melanogaster and D. simulans, specifically from embryos at 0.5 hours and 2 hours post-fertilization. Additionally, it encompasses data from D. melanogaster larvae that subjected to a 30 °C heat shock treatment, with both treated and control samples at 0, 3, 6, 9, and 12 hours post-treatment. Source data for the figures are available at 10.5281/zenodo.13999706⁵⁸.

Code availability

In-house shell and perl scripts for downloading data, constructing GRNs, as well as R scripts for statistical analyses and figure generation, are available at 10.5281/zenodo.13999706⁵⁸.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at 10.1038/s42003-024-07162-w.