Evolution and genetics of accessory gland transcriptome divergence between Drosophila melanogaster and D. simulans

Abstract

Studies of allele-specific expression in interspecific hybrids have provided important insights into gene-regulatory divergence and hybrid incompatibilities. Many such investigations in Drosophila have used transcriptome data from complex mixtures of many tissues or from gonads, however, regulatory divergence may vary widely among species, sexes, and tissues. Thus, we lack sufficiently broad sampling to be confident about the general biological principles of regulatory divergence. Here, we seek to fill some of these gaps in the literature by characterizing regulatory evolution and hybrid misexpression in a somatic male sex organ, the accessory gland, in F1 hybrids between Drosophila melanogaster and D. simulans. The accessory gland produces seminal fluid proteins, which play an important role in male and female fertility and may be subject to adaptive divergence due to male–male or male–female interactions. We find that trans differences are relatively more abundant than cis, in contrast to most of the interspecific hybrid literature, though large effect-size trans differences are rare. Seminal fluid protein genes have significantly elevated levels of expression divergence and tend to be regulated through both cis and trans divergence. We find limited misexpression (over- or underexpression relative to both parents) in this organ compared to most other Drosophila studies. As in previous studies, male-biased genes are overrepresented among misexpressed genes and are much more likely to be underexpressed. ATAC-Seq data show that chromatin accessibility is correlated with expression differences among species and hybrid allele-specific expression. This work identifies unique regulatory evolution and hybrid misexpression properties of the accessory gland and suggests the importance of tissue-specific allele-specific expression studies.

Introduction

Gene expression phenotypes such as transcript abundance, splicing, tissue, or developmental stage of expression, and environment-induced plasticity, may be related to many downstream organismal phenotypes and may be acted upon by stabilizing, directional, or diversifying selection. The gene regulatory variants shaping these expression phenotypes, which may be influenced by selection or drift, can be broadly classified into cis-acting (such as promoters and enhancers) and trans-acting (such as transcription factors) (Rabinow and Dickinson 1981; Dickinson et al. 1984; Wittkopp et al. 2004; Gibson et al. 2004; Ronald et al. 2005; Wittkopp 2005; Ronald and Akey 2007). Because the genetic control of gene expression can involve several sites spanning both cis- and trans-acting factors, selection could plausibly have many potential substrates on which to act. Thus, understanding the relative importance of these factors in regulatory evolution is critical for achieving a comprehensive understanding of expression divergence.

A common approach for detecting, and estimating the magnitudes of, cis and trans effects, is to measure allele-specific expression (ASE) in hybrids and their parents. ASE can be used to classify genes into regulatory types based on the presence and directionality of cis and trans components (McManus et al. 2010). It has been broadly applied to both intraspecific and interspecific hybrids to study the genetics of expression variation within and between species (reviewed in Gaur et al. 2013; Signor and Nuzhdin 2018). The comparison of within- and between-species regulatory genetics may inform our understanding of evolutionary mechanisms because the concordance or discordance of phenomena on these two timescales can narrow the range of evolutionary explanations for the variation. A major conclusion from accumulated ASE research is that intraspecific gene expression evolution in animals is mediated predominantly through trans effects, while interspecific evolution proceeds predominantly via cis effects (reviewed in Signor and Nuzhdin 2018; Hill et al. 2021). A plausible explanation for this pattern invokes variation in mutational opportunity and modes of selection. In this worldview, trans regulation has a broader mutational target, but trans-effect mutations tend to have greater deleterious pleiotropic effects (Wittkopp 2005; Lemos et al. 2008; Gruber et al. 2012). Thus, such variants tend to be common as low frequency polymorphisms, generating population level trans effects, but rarely fix to generate interspecific differences. Alternatively, cis-acting variants may be less pleiotropic and therefore, more likely to fix and make greater contributions to interspecific expression divergence. However, this pattern is not always observed. For example, Sánchez-Ramírez et al. (2021) found cis-regulatory divergence was more frequent than trans-regulatory divergence in male, but not female Caenorhabditis species. Some studies of flies have found relatively more trans effects between species (McManus et al. 2010; Coolon et al. 2014). Thus, the literature is still unsettled.

There is no reason to expect that the genetics of regulatory variation will be homogeneous across tissues in multicellular organisms, as the cell and developmental biology, as well as the influence of mutation and selection on expression phenotypes may vary across cell types, tissues, and organs. Indeed, as one would expect, tissues and cell types exhibit varying rates of expression divergence (Gu and Su 2007; Brawand et al. 2011; Romero et al. 2012; Kryuchkova-Mostacci and Robinson-Rechavi 2015; Liang et al. 2018; Chen, Swofford et al. 2019; Pal et al. 2021). Intraspecific studies of mice (Babak et al. 2015; Andergassen et al. 2017; St Pierre et al. 2022), humans (Babak et al. 2015; Leung et al. 2015; Castel et al. 2020), birds (Wang et al. 2017), and Drosophila (Combs et al. 2018), have revealed tissue-specific variance in cis-effects such that genes may exhibit ASE in some tissues but not others; the total number and magnitude of cis-effects also varies across tissues. These studies did not identify trans-effects, however, limiting insight into how regulatory variation and evolution may vary among tissues. Few studies of interspecific ASE in animals have investigated single somatic tissues and characterized both cis and trans regulatory divergence (Goncalves et al. 2012; Davidson and Balakrishnan 2016; Mugal et al. 2020; Ranz et al. 2023).

Much of the influential Drosophila work on these topics analyzes parental and hybrid expression in whole animals or heads, with a particular focus on the melanogaster subgroup (Wittkopp et al. 2004, 2008; Landry et al. 2005; Graze et al. 2009; McManus et al. 2010; Coolon et al. 2014; Wei et al. 2014). However, as these studies investigated complex mixtures of many tissue and cell types, the properties of individual organs are impossible to identify. Another line of investigation has sought to investigate Drosophila hybrid male sterility phenotypes in nonmelanogaster species through analysis of gene mis-expression in the testis of partially sterile males (Haerty and Singh 2006; Lu et al. 2010; Llopart 2012; Gomes and Civetta 2014; Gomes and Civetta 2015; Brill et al. 2016; Banho et al. 2021), though the causal relationship between mis-expression and fertility remains unclear (Civetta 2016). Recent work comparing the genetics of gene expression divergence in different reproductive organs of D. willistoni subspecies (Ranz et al. 2023) provided evidence that the genetics of regulatory divergence varies across organs; however, the low levels of expression divergence and extremely low sequence divergence of these subspecies reduced power for some key inferences and left unclear the extent to which the observed properties reflect within species vs between species variation.

The central role of D. melanogaster as a model system implies that the investigation of the genetics of expression divergence between this species and its close relatives will play an important role in determining how the genetics of regulatory divergence varies across tissues and biological functions, though as noted above, there are currently no tissue- or organ-based investigations of this topic in the species pair. Drosophila melanogaster and its sibling species, D. simulans, shared a most recent common ancestor roughly 2–3 million years ago (Obbard et al. 2012) and exhibit little shared ancestral polymorphism (Langley et al. 2012). Hybrid males derived from crosses between D. melanogaster males and D. simulans females are generally completely sterile or inviable (Sturtevant 1920), with severely atrophied or absent testes. The reciprocal cross exhibits nearly complete sexual isolation (Welbergen et al. 1987, 1992) and the few hybrids produced are sterile females (Sturtevant 1920, 1929). Despite the male sterility generally observed in hybrids between male D. melanogaster and female D. simulans, a previous investigation reported that such males have morphologically normal accessory glands that produce seminal fluid competent to induce female postmating responses (Stumm-Zollinger and Chen 1988), which include increased egg laying, facilitation of sperm storage, immune system responses, increased feeding rates, increased activity level and decreased sleep, and decreased receptivity to remating (reviewed in Ravi Ram and Wolfner 2007; Avila et al. 2011; Sirot et al. 2014; Wigby et al. 2020). Our goal here is to contribute to the literature on the genetics of interspecific regulatory divergence between D. melanogaster and D. simulans using the accessory gland (AG) as a model.

The accessory gland, ejaculatory duct, and ejaculatory bulb synthesize several seminal fluid proteins (Sfps), which are transferred to females along with sperm during mating and are essential for fertilization, similar to the seminal fluid of the mammalian prostate (reviewed in Poiani (2006) and Wilson et al. (2017)). Drosophila has a polyandrous mating system featuring competition between males for matings, and between ejaculates for access to eggs (Boorman and Parker 1976; Imhof et al. 1998; Clark et al. 1999). Sfps of Drosophila and many other insects induce the female postmating response (PMR), which may evolve in response to sperm competition and sexual conflict (Hollis et al. 2019). Sfps play a key role in mediating sperm competition; genetic variation in Sfp loci is linked to competitiveness (Clark et al. 1995; Chapman et al. 2000; Fiumera et al. 2005), and males respond to perceived level of competition through differential allocation of Sfps to the ejaculate (Sirot et al. 2011; Hopkins et al. 2019). Sfp protein sequences evolve at an especially rapid rate, often under the influence of recurrent directional selection (Tsaur et al. 1998; Aguadé 1999; Begun et al. 2000; Holloway and Begun 2004; Begun et al. 2006; Schully and Hellberg 2006; Wong et al. 2008; Majane et al. 2022), perhaps partially due to sexual conflict (Swanson and Vacquier 2002; Haerty et al. 2007). Reduced selective constraint may also contribute to their rapid divergence (Dapper and Wade 2020; Patlar et al. 2021). Sfp genes have rapid rates of turnover (Mueller et al. 2005; Wagstaff and Begun 2005), exhibiting gene gain and loss even among closely related species (Begun and Lindfors 2005).

While there is a long history of work on sequence evolution and turnover in Sfps (reviewed in Hurtado et al. 2022), less is known about gene expression evolution among Sfps or in the accessory gland more broadly. RNAi knockdowns demonstrate that PMR phenotypes are sensitive to expression level of multiple Sfps (Ravi Ram and Wolfner 2007; Patlar and Civetta 2022), suggesting that Sfp expression is a plausible target of selection. Consistent with the observation that male-biased genes tend to have higher levels of interspecific expression divergence (Meiklejohn et al. 2003; Parisi et al. 2004; Ellegren and Parsch 2007; Brawand et al. 2011; Graveley et al. 2011; Assis et al. 2012; Whittle and Extavour 2019; Pal et al. 2021), we recently reported rapid expression divergence as well as the evolution of novel genes and expression phenotype in the accessory gland (Cridland et al. 2020). However, Cridland et al. did not focus on the general properties of accessory gland transcriptome divergence, did not compare Sfp expression to expression divergence of other gene classes expressed in the accessory gland, and did not address the genetics of accessory gland expression divergence between species.

In this study, we have several goals. First, we characterize expression divergence between D. melanogaster and D. simulans in this key somatic organ of male reproduction, including contrasts between evolutionary properties of Sfps and other genes expressed in the gland. Second, we use ASE analyses derived from measures of gene expression in accessory glands of species and their hybrids to estimate cis and trans expression effects in the accessory gland and ejaculatory duct; we quantify inheritance of expression phenotypes and characterize misexpressed genes that may be related to hybrid incompatibilities. Third, we investigate connections between regulatory evolution and divergence in upstream noncoding regions and protein sequence evolution. Finally, we integrate ATAC-seq data to link changes in chromatin state with expression divergence. Given the relatively low cell type diversity in these tissues and the numerical dominance of the accessory gland main cell (Majane et al. 2022), many of the inferences from these data are likely due to main cell expression phenotypes.

Methods

RNA-Seq

We performed RNA-Seq on each of three genotypes: D. melanogaster (RAL 517, Mackay et al. 2012), D. simulans (w⁵⁰¹), and a D. melanogaster X D. simulans interspecific hybrid, with three biological replicates per sample. We raised all flies on cornmeal-molasses-agar medium at 25°C and 60% relative humidity, on a 12:12 light/dark cycle. We crossed the parents to produce interspecific hybrids by pairing five D. simulans females with 25 D. melanogaster males; the excess of males helps overcome behavioral isolation between species. The reciprocal cross shows extraordinarily strong behavioral isolation, and in rare cases where the cross is successful, males die during development. We collected virgin male adult F1 animals and kept them in groups of five males per vial. We aged flies for 2 days before dissection. On the day of the experiment, we anesthetized flies with light CO₂, dissected their accessory glands and anterior ejaculatory duct in cold 1X PBS, and collected the tissue in TRIzol (Thermo-Fisher 15596026) on ice. We confirmed that the hybrid organs appeared morphologically normal with seminal fluid production. The cell diversity of the dissected tissue is low, with roughly 15% of cells deriving from the ejaculatory duct and 85% from the accessory gland, which is composed primarily of a single cell type, the main cell (Majane et al. 2022). After we collected tissue from 20 males we flash-froze the TRIzol tubes containing tissue in liquid nitrogen and stored the material at −80°C. We extracted RNA using the standard TRIzol protocol followed by DNAse digestion (Invitrogen AM1907) and clean-up with AMPure beads (Beckman-Coulter A63881). Novogene performed RNA-Seq library preparation and 150 bp paired-end Illumina sequencing; reads can be found at PRJNA913156.

Assigning species-of-origin

To determine the species-of-origin for each allele in the hybrid, we used an alignment-based approach relying on differences in the number of mismatches between the reads and each reference. We aligned each hybrid sample to each of two references, D. melanogaster (custom reference based on Flybase release 6.04 with SNPs included for strain RAL 517), and D. simulans (Princeton University, release 3.0), using HiSat2 and requiring a MAPQ score ≥30. We sorted reads using custom perl, bash, and R scripts (github.com/alexmajane/hybridASE) into groups that mapped either to one reference uniquely or mapped to both references. For read pairs where at least one mate aligned uniquely, we assigned both reads to that species. For the remaining reads, we analyzed the number of nucleotide mismatches algorithmically to assign species-of-origin. Reads that aligned to one species with at least six fewer mismatches were assigned to the species with fewer mismatches. We also subjected our D. melanogaster and D. simulans samples to the same workflow, to (A) account for artifactual effects of the procedure on expression analysis, (B) establish a ground-truth false-positive rate for species-assignment, and (C) identify problematic gene regions with high rates of erroneous species-assignment. To address the latter point, we calculated the fold change of counts (see below) per gene with and without inclusion of misassigned parental reads. If the absolute value of log₂(fold change) was greater than 0.025 in either species, we removed that gene from our downstream analyses, a total of 382 genes.

Quantification of gene expression

After we assigned hybrid reads to species-of-origin, we quantified gene expression, using reads with a confident species assignment, with Salmon (Patro et al. 2017). We used Salmon's alignment-free approach because it accounts for differences in transcript length and GC content across samples, which in our case vary among orthologs across species due to evolution and/or variation in gene annotations. Following quantification we used tximport (Soneson et al. 2015) to estimate counts per gene. We limited our analysis to D. melanogaster–D. simulans 1-to-1 orthologs based on the FlyBase annotation (02/2020 release) with additional manually curated Sfp orthologs (Majane et al. 2022).

Differential expression

We used estimated counts from tximport as the basis of all downstream analyses. We performed independent analyses of autosomal and X-linked genes because of their different inheritance in the hybrid. We used DESeq2 (Love et al. 2014) to normalize count data with the median-of-ratios method (Anders and Huber 2010), identified DE genes using Wald tests, and estimated moderated log-fold changes with the apeglm model (Zhu et al. 2019).

Regulatory and inheritance classifications

We refer to counts from D. melanogaster as P_mel, D. simulans as P_sim, and allele-specific hybrid counts as F1_mel and F1_sim. We calculated total F1 expression (F1_total) as F1_mel + F1_sim. We classified genes into regulatory and inheritance categories using the algorithm following McManus et al. (2010). For this purpose, we define DE as a significant Wald test (Bonferonni adjusted P < 0.05) and make comparisons between (A) parental expression: P_mel and P_sim (P), (B) ASE within the hybrid: F1_mel and F1_sim (H), and (C) between parent expression and expression of parental-specific alleles in the hybrid: P_mel and F1_mel or P_sim and F1_sim (T). We define regulatory classes as follows:

1. conserved: no significant P, H, or T.

2. cis: significant P and H, no significant T.

3. trans: significant P and T, no significant H.

4. cis + trans: significant P, H, and T, same directionality between the parental contrast and hybrid ASE. cis and trans effects favor expression of the same allele.

5. cis by trans: significant P, H, and T, opposite directionality between the parental contrast and hybrid ASE. cis and trans effects favor expression of different alleles.

6. compensatory: significant H and T, but no significant P. cis and trans effects complement one another such that there has been no evolved expression difference among species.

7. ambiguous: all other patterns of expression without classical interpretations, such as no P or H, but significant T (evidence of trans effects that appear only in the hybrid).

We classified genes by inheritance comparing overall hybrid expression (F1_total) to parental expression. Note that throughout, the term dominance is used in the phenotypic sense; we do not make inferences about the genetic basis of inheritance of expression phenotypes in the hybrid. We define the following inheritance classes:

1. conserved: no DE between F1_total and P_mel or P_sim.

2. additive: DE between F1_total and both P_mel and P_sim. F1_total has an intermediate expression level.

3. mel dominant: no DE between F1_total and P_mel. DE between F1_total and P_sim.

4. sim dominant: no DE between F1_total and P_sim. DE between F1_total and P_mel.

5. overdominant: DE between F1_total and both P_mel and P_sim. F1_total is expressed at a higher level than both parents.

6. underdominant: DE between F1_total and both P_mel and P_sim. F1_total is expressed at a lower level than both parents.

SFPs and AG-biased genes

We defined sets of Sfps and accessory gland (AG)-biased genes to investigate patterns of DE, regulation, and inheritance in these gene classes. We refer to Wigby et al. (2020) for annotation of Sfps; 208 Sfps were expressed in our dataset. To annotate AG-biased genes we obtained expression data from FlyAtlas2 (Leader et al. 2018), and calculated the index of tissue specificity, τ (Yanai et al. 2005), for all genes. We defined AG-biased genes as those that are more highly expressed in the accessory gland than all other tissues and have τ ≥ 0.8. There are 378 AG-biased genes expressed in our dataset, 238 of which are non-Sfps, which we used for further analysis. Note that the categorization of Sfps and AG-biased genes rely on D. melanogaster data, while the true sets of Sfps and AG-biased genes among orthologs in the two species are likely highly similar but not identical (Begun and Lindfors 2005; Findlay et al. 2008; Majane et al. 2022).

Gene ontology analysis

We performed gene ontology (GO) enrichment analyses with Enrichr (Kuleshov et al. 2016). We defined background gene sets as all genes expressed in the data (log₂(counts) ≥ 1). We used the Bioconductor D. melanogaster annotation (Carlson 2021) and queried terms from all three sub-ontologies (Biological Process, Molecular Function, and Cellular Component). We used Fisher's Exact Test with Benjamani–Hochberg corrected P value <0.05 to identify GO terms significantly overrepresented in inheritance and regulatory groups vs the background gene set. For overrepresented GO terms, we calculate an enrichment score as the ratio of term frequency in the test gene sets vs the background gene set.

Upstream sequence analysis

We obtained sequences spanning 1000, 750, and 500 bp upstream of each D. melanogaster TSS (Flybase annotation 6.41) and removed overlapping coding sequence. We then used BLAST (gap open penalty: 2; gap extension penalty: 1) to identify orthologous sequence in the D. simulans genome (Princeton University, release 3.0). We discarded sequences with more than one BLAST hit or overlapping alignments. Next we aligned D. melanogaster and D. simulans sequences using MUSCLE (Edgar 2004). We estimated the Kimura 2-parameter nucleotide substitution rate (Kimura 1980) for each upstream region using EMBOSS distmat (Rice et al. 2000). We also trimmed the 500 bp sequences to shorter lengths of 100, 200, and 300 bp upstream of the TSS and repeated estimation of substitution rate. The small amount of sequence divergence between any two cosmopolitan genotypes (such as the reference sequence strain and RAL 517) relative to the divergence to D. simulans (Langley et al. 2012) makes it extremely unlikely that use of the reference sequence rather than RAL 517 to estimate divergence affects any general conclusions.

Protein sequence evolutionary analysis

We analyzed sequence divergence by estimating synonymous (dS) and nonsynonymous (dN) substitution rates. We obtained the longest open reading frame per gene from FlyBase annotations (D. melanogaster 6.41, D. simulans 2.02), translated nucleotide sequences with EMBOSS transeq (Rice et al. 2000), and aligned amino acid sequences with MUSCLE (Edgar 2004). We back-translated to codon alignments with gaps removed using PAL2NAL (Suyama et al. 2006). For each gene we estimated dN and dS using Goldman and Yang's maximum likelihood codon-based substitution model (codeml; Goldman and Yang 1994). We performed this analysis with PAML (Yang 1997) implemented in BioPython (Cock et al. 2009).

We additionally analyzed adaptive protein evolution in D. melanogaster using available population genomics data (Fraïsse et al. 2019) with precomputed McDonald–Krietman tests (McDonald and Kreitman 1991), as in our previous work (Majane et al. 2022- see supplement for detail). We used the summary statistic α, which estimates for each gene the proportion of amino acid substitutions due to positive selection. A positive value is consistent with directional selection; larger values suggest a greater proportion of adaptive substitutions.

Chromatin state integration

We performed ATAC-seq experiments on 25 individual accessory glands, with three biological replicates each for RAL 517 (D. melanogaster) and w⁵⁰¹ (D. simulans). Virgin males were transferred to vials 0–2 h posteclosion, then dissected at 48 ± 2 h following transfer. Tissues were lysed in 200 µl of ATAC-seq lysis buffer (10 mM Tris-HCl, 10 mM NaCl, 3 mM MgCl₂, 0.1% IGEPAL CA-630) and manually homogenized 25 times with a plastic pestle, followed by a 1-min incubation on ice, and repeated twice. Following homogenization, samples were pelleted at 4 °C (100 g for 10 min) to recover nuclei. We removed the supernatant and washed the pellet in 200 µl of the lysis buffer. The nuclei preparation was filtered through a 40-µm cell strainer and washed with another 200 µl of lysis buffer. The purified nuclei were isolated by centrifugation at 1,000 × g for 10 min at 4 °C. After removing the supernatant, 12.5 µl Nextera Tagment DNA Buffer (Illumina), 11.25 µl ddH2O and 1.25 µl Tn5 Transposase were added to the purified nuclei to tagment the DNA. Libraries were then processed following Buenrostro et al. (2015) with a final upper and lower size-selection using SPRI beads with bead-to-sample ratios of 0.4 and 1.7, respectively. An aliquot of the purified library was analyzed on a Bioanalyzer (Agilent) to ensure the characteristic nucleosome periodicity of ATAC-seq libraries. Sequencing was performed at the UC Davis Genome Center using an Illumina Hiseq4000 to generate 150 bp paired-end reads, which can be found at PRJNA993436.

Raw ATAC-seq reads were trimmed using cutadapt v1.15. Reads from D. melanogaster and D. simulans were aligned to the dm6 (v6.41) and r2.02 reference genomes, respectively, using bowtie2 (v2.3.3.1) option -X 2000 –local –very-sensitive-local. Next, duplicate reads were marked using picard tools (v2.17.0) and filtered using samtools (v1.7) with option -q 30 -f 2 -F 1804 to remove unmapped reads and unmapped mates. Peak calling was performed on each species separately with Genrich v.0.5 206 (https://github.com/jsh58/Genrich) using option -j -q 0.001 -d 200.

We used reciprocal BLAST with gap open penalty = 2 and gap extension penalty = 1 on peak sequences between species and verified orthology of 1-to-1 best hits using synteny (nearest upstream and downstream annotated exons). We defined conserved peaks as those with a single reciprocal best hit in each species and shared synteny as inferred from the nearest upstream and downstream exons. We defined orphan peaks as those with no BLAST hits to the other species (e.g. melanogaster peaks that did not BLAST to the unique orthologous location in simulans). We additionally BLASTed orphan peak sequences to the reciprocal species’ genome, and filtered out peaks with (A) no hits to the genome, (B) a hit within 100 bp of any annotated peak, (C) multiple hits. It is important to compare truly orthologous regions when we quantify peak accessibility by counting reads. We defined orthologous regions of orphan peaks in reciprocal species as the span of each BLAST hit. We reannotated conserved peaks by reciprocal BLAST of peaks to each other species’ genome and extended the boundaries of each peak to the span of BLAST hits intersecting the original annotation.

To quantify chromatin accessibility in each peak, we counted the number of aligned ATAC-seq reads intersecting each peak with HTSeq (Anders et al. 2015). We analyzed count data using DESeq2 similarly to RNA-Seq analysis. We then annotated peaks with the nearest TSS. For each gene we used our Salmon quantification to select the transcript with highest expression in our sample and chose its annotated TSS. We then selected 1-to-1 peak-to-TSS pairs with the closest or overlapping peak per TSS, removing any duplicate matches from further analysis. We also filtered the data to include only pairs where each orthologous peak (or region in the case of orphans) matched the same gene in both species.

Results

Alignment and identification of allele-specific reads

We performed RNA-Seq on D. melanogaster, D. simulans, and an interspecific hybrid (P_mel, P_sim, and F1). We obtained between 25.4 and 30.8 M reads per RNA-Seq sample. We aligned each sample to each of two references, D. melanogaster and D. simulans. Each species aligns at a rate of 94–97% to the matching reference (Supplementary Table 1). A slightly better alignment rate for P_sim is expected, since the D. simulans strain we used in our experiment matches the reference strain. Alternatively, our D. melanogaster strain matches the updated reference sequence for SNPs but not for the small number of indels in the genic regions used here.

For the reads from hybrids, a higher alignment rate to D. simulans is expected given that the hybrid inherits a D. simulans X chromosome. Reads from the F1 aligned to the D. melanogaster reference at a rate of 61–62% and to the D. simulans reference at a rate of 69–71%. Among F1 reads, ∼20% aligned uniquely to D. melanogaster, 33–34% aligned uniquely to D. simulans, and 46–47% aligned to both references (Supplementary Table 2). We used a mismatch-based approach to compare alignments and assign nonuniquely aligned reads to each species (Supplementary Table 3). We were able to assign ∼25% of nonuniquely aligned reads (∼12% of total aligned reads) to D. melanogaster, and 35–36% of nonuniquely aligned reads (16–17% of total aligned reads) to D. simulans, leaving 18–19% of total aligned reads of indeterminable origin and removed from our analysis.

We passed parental pure species reads through the same analysis pipeline used for the F1, to (1) account for artifactual effects of the procedure on expression analysis, (2) establish a ground-truth false-positive rate for species-assignment, and (3) identify problematic gene regions with high rates of erroneous species-assignment. ∼1.2% of P_mel reads uniquely aligned to the D. simulans genome, and 0.6–0.7% of P_sim reads uniquely aligned to the D. melanogaster genome (Supplementary Table 2). Among nonuniquely aligning parental reads, our assignment algorithm assigned 0.16–0.17% of P_mel reads to D. simulans, and 0.031–0.035% of P_sim reads to D. melanogaster (Supplementary Table 3). We used incorrectly assigned parental reads to identify gene regions with elevated levels of misassignment. Most misassigned reads do not overlap genes (Supplementary Table 4). We identified 382 genes where misassigned reads significantly impacted estimates of gene expression (Supplementary Data); we removed these from downstream analysis. We used Salmon (Patro et al. 2017) to quantify gene expression on reads that passed our filtering and species-assignment. Salmon mapping rates are as follows: 89–90% of F1 reads of D. melanogaster origin (hereafter F1_mel), 84–85% of F1 reads of D. simulans origin (hereafter F1_sim), ∼94% of P_mel reads, and 86–87% of P_sim reads.

Transcriptome-wide view

We converted quantified expression values from Salmon to estimated counts with tximport (Soneson et al. 2015) and used these counts as the basis of all downstream analyses. Given the different inheritance patterns of the autosomes and X chromosome in the hybrid, we analyze the corresponding genes separately. The hemizygous X is omitted from the ASE analysis. We refer to total F1 expression (sum of both alleles) as F1_total, and ASE measures as F1_mel and F1_sim.

Principal component analysis of transcriptome-wide gene expression shows a high level of shared variance between replicates (Fig. 1). Among autosomal genes, P_mel groups with F1_mel and P_sim groups with F1_sim, while F1_total groups away from these two clusters. PC1 appears to explain differences between F1 samples and P_mel/sim, with allele-specific samples lying between F1_total and P_mel/sim, though much closer to P_mel/sim. PC2 appears to explain expression differences between D. melanogaster and D. simulans, with F1_total lying roughly midway between the species. Among X-linked genes, all of which derive from the D. simulans parent, both PC1 and PC2 appear to explain differences between D. melanogaster and D. simulans allele-derived expression, and X-linked F1 expression is tightly grouped with P_sim expression. We did not find appreciable differences in distributions of expression variance between samples or allele-specific expression (Supplementary Fig. 1), aside from a very modest elevation in standard deviation among X-linked genes relative to autosomal genes.

Fig. 1.

PCA of transcriptome-wide expression (log-transformed counts) showing the first two PCs. a) Autosomal-linked genes expression. Parental expression is overall similar to parent-specific alleles in the hybrid, with total hybrid expression clustering away from parental and allele-specific expression. b) X-linked gene expression. Drosophila simulans expression is similar to hybrid expression.

We characterize transcriptome-wide divergence using the correlations of average gene expression, which are high, as expected for two closely related species. Among autosomal genes, P_mel and P_sim have a Pearson correlation coefficient r = 0.934 (Fig. 2a). F1_total expression is more similar to each parent; expression profiles between F1_total and P_mel have an r = 0.965 (Fig. 2b), while F1_total and P_mel have an r = 0.967 (Fig. 2c). Correlations between allele-specific expression and the same-species parent are strongest: F1_mel and P_mel have an r = 0.982 (Fig. 2d); F1_sim and P_sim have an r = 0.980 (Fig. 2e). Allele-specific expression within hybrids is somewhat more correlated than parental expression profiles are to one another; F1_mel and F1_sim have an r = 0.945 (Fig. 2f). Among X-linked genes, P_mel and P_sim are similarly correlated as with autosomal genes (r = 0.930, Fig. 2g). X-linked gene expression in hybrids is overall very similar to D. simulans: F1_total and P_sim have an r = 0.983 (Fig. 2h). Comparing F1_total and P_mel is very similar to the parental X-linked expression contrast with an r = 0.929 (Fig. 2i). Overall, the data strongly support the existence of widespread additivity for autosomal genes, and D. simulans-like expression on the hybrid X chromosome, suggesting strong influence of cis effects. The strong similarity in expression between the hybrid and the parents, and between hybrid ASE and parents, suggests that the accessory glands of this hybrid are not subject to particularly widespread misexpression.

Fig. 2.

Regulatory classification by cis and trans mechanisms. a) Fraction of genes classified into each regulatory type, with Sfps and AG-biased genes shown separately. b) log₂(fold change) of ASE and parental expression are shown with regulatory types highlighted. Ambiguous genes are removed and scale limited for clarity. For full data visualization, see Supplementary Fig. 3.

Differential gene expression

Here, we define differential expression as genes with a significant difference in normalized counts (Wald test, adjusted P < 0.01) and an absolute value of moderated log₂(fold change) >1 (Table 1). We also include counts of DE genes without imposing a log₂(fold change) cutoff in Table 1. While many more genes are DE without a cutoff, in general we find that the patterns among these two criteria are similar, and given that DE genes with a fold-change cutoff are potentially more biologically relevant, we discuss them further below.

Table 1.

Differentially expressed genes. DE is defined as a significant Wald test (Bonferroni adjusted P < 0.01), and a log₂(fold change) value greater than 1. The rightmost three columns additionally show the number and fraction of significant DE genes without an imposed log₂(fold change) cutoff.

Gene group	Contrast	DE: log2(FC) > 1	Non-DE: log2(FC) > 1	Fraction DE: log2(FC) > 1	DE	Non-DE	Fraction DE
Autosomes	Pmel Psim	1608	7615	0.174	3005	6218	0.326
Autosomes	F1total Pmel	883	8340	0.096	2290	6933	0.248
Autosomes	F1total Psim	956	8267	0.104	2459	6764	0.267
Autosomes	F1mel Pmel	430	8793	0.047	1585	7638	0.172
Autosomes	F1sim Psim	561	8662	0.061	1790	7433	0.194
Autosomes	F1mel F1sim	1103	8120	0.120	2132	7091	0.231
X	Pmel Psim	284	1415	0.167	564	1135	0.332
X	F1sim Pmel	303	1396	0.178	631	1068	0.371
X	F1sim Pmel	66	1633	0.039	310	1389	0.182

We find that 17% of 9,223 total expressed genes are DE between P_mel and P_sim (Table 1). The list of DE genes and the resulting GO enrichments are provided in the Supplementary Data. The two most significantly enriched terms were GO:0005887 and GO:0031226, integral and intrinsic components of plasma membrane, respectively. Among the significant terms with at least two-fold enrichment, we find neurotransmitter:sodium symporter activity (GO:0005328), solute:sodium symporter activity (GO:0015370), solute:cation symporter activity (GO:0015294), extracellular ligand-gated ion channel activity (GO:0005230), and neurotransmitter receptor activity (GO:0030594). Among the genes associated with such functions are multiple serotonin and octpamine receptors, as well as nAchRB3. The DE serotonin receptors, 5-HT1A, 5-HT1B, and 5-HT2A, are all expressed at substantially higher levels in D. simulans, and while the possible significance for AG function is unclear, such variation could plausibly be related to reproductive behavioral divergence between the two species, such as copulation duration (Grant 1983; Welbergen et al. 1987; Price et al. 2001). Several gustatory and odorant receptors were DE, as were several odorant binding protein genes. A number of DNA damage response genes were also differentially expressed. As expected, fewer genes are DE in comparison to the hybrid: 10% of expressed genes are DE between F1_total and P_mel, and 10% between F1_total and P_sim.

DE is less frequent between hybrid allele-specific expression and parents relative to total hybrid expression (Table 1): 5% of genes are DE between F1_mel and P_mel, and 6% of genes are DE between F1_sim and P_sim. 12% of genes are DE between F1_mel and F1_sim (allele-specific expression within the hybrid, indicative of cis-regulatory effects). Among expressed X-linked genes, 17% are DE between P_mel and P_sim, similar to autosomal genes. DE is more common between F1 and P_mel with 18% of X-linked genes DE. Just 4% of X-linked genes are DE between F1 and P_sim, suggesting that trans-regulatory differences associated with large shifts in X-linked gene expression are rare.

DE among Sfps and accessory gland-biased genes

Sfps are known to have very high rates of amino acid substitutions (e.g. Haerty et al. 2007; Swanson et al. 2001) as well as gene turnover between species (Begun and Lindfors 2005; Mueller et al. 2005; Wagstaff and Begun 2005; Hurtado et al. 2022). Given the observation that rates of protein evolution are often correlated with gene expression evolution (Makova and Li 2003; Nuzhdin et al. 2004; Jordan et al. 2005; Khaitovich et al. 2005; Lemos et al. 2005; Liao and Zhang 2006; Sartor et al. 2006; Hunt et al. 2013; Warnefors and Kaessmann 2013; Hodgins et al. 2016; Zhong, Lundberg, and Råberg 2021), we asked whether expressed Sfps (208 total) were more likely than non-Sfps to be DE between D. melanogaster and D. simulans (P_mel vs P_sim) and between hybrid alleles (ASE: F1_mel vs F1_sim). Indeed, 28% of Sfps are DE between D. melanogaster and D. simulans, compared to 17% of non-Sfps (G test, P < 0.001, Table 2), and 22% of Sfps are DE between hybrid alleles (ASE), compared to 12% of non-Sfps (G test, P < 0.001).

Table 2.

Number of DE genes among gene classes. AG-biased and nonbiased gene sets exclude Sfps. DE is defined as a significant Wald test (Bonferroni adjusted P < 0.01), and a log₂(fold change) value greater than 1. The rightmost three columns additionally show the number and fraction of significant DE genes without an imposed log₂(fold change) cutoff.

Gene class	Contrast	DE: log2(FC) > 1	Non-DE: log2(FC) > 1	Fraction DE: log2(FC) > 1	DE	Non-DE	Fraction DE
Non-Sfp	Pmel Psim	1549	7466	0.172	2883	6132	0.320
Sfp	Pmel Psim	59	149	0.284	122	86	0.587
Nonbiased	Pmel Psim	1461	7316	0.166	2733	6044	0.311
AG-biased	Pmel Psim	88	150	0.370	150	88	0.630
Non-Sfp	F1mel F1sim	1057	7958	0.117	2025	6990	0.225
Sfp	F1mel F1sim	46	162	0.221	107	101	0.514
Nonbiased	F1mel F1sim	989	7788	0.113	1889	6888	0.215
AG-biased	F1mel F1sim	68	170	0.286	136	102	0.571

Sfps are a highly expressed class of genes, however, (Supplementary Fig. 2a–c), and so may be more likely to be DE because of greater statistical power to detect expression differences. The median log₂(counts) of Sfps is 8.46, while the median of non-Sfps is 4.31. To account for the effect of expression level on the likelihood of DE, we used a multiple logistic regression with average expression and Sfp status as independent variables, and DE as the dependent variable (DE ∼ log₂(counts) + Sfp). Interestingly, highly expressed genes are not much more likely to have large effect-size DE (log₂(fold change) > 1); average expression level has a weak relationship with DE between P_mel and P_sim (β = 0.04 ± 0.01, P < 0.001). Sfp status does predict large-effect size DE (β = 0.45 ± 0.16, P = 0.006). Average expression does not have a relationship with DE between F1_mel and F1_sim (β = 0.01 ± 0.01, P = 0.08), and Sfp status predicts DE (β = 0.53 ± 0.17, P = 0.002). We therefore conclude that Sfps are significantly enriched for large-effect size DE events compared to non-Sfps. If we consider genes to be DE without a log₂(fold change) cutoff however, we observe a strikingly different result. In the parental contrast, average expression significantly predicts DE between P_mel and P_sim (β = 0.22 ± 0.01, P < 0.001); however, Sfp status does not predict DE (β = −0.04 ± 0.16, P = 0.792). Considering ASE, average expression significantly predicts DE between F1_mel and F1_sim (β = 0.21 ± 0.007, P < 0.001), but Sfp status does not predict DE (β = 0.09 ± 0.16, P = 0.554). Therefore, we conclude that in our experiment, Sfps are not any more likely to be DE without a cutoff than non-Sfps when accounting for expression level.

Sfps have at least two obvious attributes; they are transferred to females during mating and they tend to exhibit AG-biased expression. To investigate whether the enrichment of Sfps amongst DE genes is more likely related to their status as Sfps or to their AG-biased expression, we also characterized DE in AG-biased genes. There are 238 genes that are AG-biased (τ > 0.8) but are not Sfps; 17.2% of nonbiased genes are DE between P_mel–P_sim, and 11.7% are DE between hybrid alleles. The proportion of AG-biased genes exhibiting DE is dramatically greater; 37% are DE between P_mel–P_sim, and 28.6% are DE between hybrid alleles (Table 2). AG-biased genes are more highly expressed than non-AG-biased genes, but not as highly expressed as Sfps (Supplementary Fig. 2d–f).

To ask whether AG-biased genes are more likely to be DE, we used a multiple logistic regression on non-Sfps, with average expression and AG-bias as independent variables, and DE as the dependent variable (DE ∼ log₂(counts) + AG-bias). In the parental contrast, average expression only very weakly predicts DE between P_mel and P_sim (β = 0.03 ± 0.01, P < 0.001), while AG-bias very strongly predicts DE (β = 0.95 ± 0.14, P < 0.001). For ASE, average expression does not predict DE between F1_mel and F1_sim (β = 0.01 ± 0.01, P = 0.40), and AG-bias strongly predicts DE (β = 0.96 ± 0.15, P < 0.001). It is therefore apparent that AG-biased genes are much more likely to be DE than more broadly expressed genes in the accessory gland. Thus, there are two general conclusions. First, from the expression divergence perspective, Sfps are not obviously different from other AG-biased genes, thereby providing no direct support for the idea that high Sfp expression divergence is a direct result of their function in the female reproductive tract (though our observations do not rule out this possibility). Second, while the role of directional selection in large interspecific expression differences cannot be inferred with these data, the enrichment pattern is consistent with the idea that AG expression phenotypes are more likely to be the result of selection on gland-related phenotypes rather than as a pleiotropic effect of selection on expression in other organs.

Gene regulatory divergence classification

We characterized cis- and trans-regulatory effects for each autosomal gene by comparing ASE in hybrids to expression in each parent species. A total of 2764 genes have evidence of cis effects (30% of expressed genes), 3338 have evidence of trans effects (36%), and 1601 have evidence of both cis and trans effects (17%). While there are more genes with significant trans effects (G test, P < 0.001), the median cis effect is significantly larger (log₂(fold change) = 0.92) than the median trans effect (log₂(fold change) = 0.64, Wilcoxon rank sum test, P < 0.001).

We further classified the regulatory basis of each gene following McManus et al. (2010) (Fig. 3, Supplementary Tables 5 and 6). The biggest category (4062 genes, or 44%) is conserved, with no significant cis or trans effects. Roughly equal numbers of genes exhibit pure cis or pure trans regulation; 933 = 10.1%, and 912 = 9.9%, respectively. Genes with both cis and trans effects are classified into three groups. The largest group by far is cis ± trans regulation, where cis and trans effects have the same directionality (e.g. P_mel > P_sim and F1_mel > F1_sim), which contains 1116 (12.1%) genes. Only 104 (1.1%) genes have cis by trans regulation—cis and trans effects with opposite directionality (e.g. P_mel > P_sim and F1_mel < F1_sim). Finally, 381 genes (4.1%) exhibit compensatory regulation, such that there is no DE between P_mel and P_sim despite evidence of cis and trans regulatory evolution. There are 1715 genes (18.6%) that cannot be classified into any of the above categories, and are labeled ambiguous.

Fig. 3.

Inheritance classification gene expression phenotypes in hybrid offspring. a) Fraction of inheritance types for autosomal, X-linked, Sfps, and AG-biased genes are each shown independently. b) log₂(fold change) of hybrid expression relative to each parent for autosomal genes. c) X-linked genes. Scale is limited in b and c for clarity. For full visualization, see Supplementary Fig. 3. d) Inheritance types among each of the regulatory classifications identified through cis and trans mechanisms.

Sfps and AG-biased genes are much less likely to be conserved (Fig. 3a, Supplementary Table 5), which is expected given their higher rates of expression divergence. Removing conserved and ambiguous classifications from consideration more clearly reveals differences in how Sfps and accessory gland-biased genes are regulated relative to all autosomal genes (Supplementary Table 6); both have smaller proportions of cis-regulation. There are about half as many cis-regulated genes among Sfps and a third fewer among accessory gland-biased genes. cis + trans regulation is particularly common among Sfps and accessory gland-biased genes. The large cis + trans component for such genes would support the idea that multiple regulatory substitutions are aligned with the direction of expression divergence for Spfs and AG-biased genes, as would be expected under the hypothesis that directional selection contributes to expression divergence. Rates of pure trans, cis by trans, and compensatory regulation are roughly equal among gene sets.

Inheritance classification, misexpression, gain- and loss-of-function phenotypes

We characterized patterns of inheritance of expression phenotypes for autosomal genes by comparing F1_total to each parent (Fig. 4, Supplementary Tables 7 and 8). Conserved genes exhibit no DE in any comparison, comprising 4886 genes (53% of all expressed genes). Seven hundred and thirty-one genes (7.9%) are additive, where P_mel and P_sim are DE and F1_total has an intermediate expression phenotype. Genes with parental divergence and with F1_total expression levels that were not DE relative to either parent are classified as either mel dominant or sim dominant. Rates are similar: 1403 (15.2%) are mel dominant, and 1208 (13.1%) are sim dominant. Genes that are overexpressed in F1_total relative to both parents are overdominant, and underexpressed genes underdominant. There are 448 (4.9%) overdominant and 547 (5.9%) underdominant genes.

Fig. 4.

GO terms associated with regulatory and inheritance types. Enrichment is the ratio of terms in the test gene set compared to the background gene set. a) Regulatory types; b) inheritance types in autosomes; c) over- or underexpressed X-linked genes. Terms associated with cis and cis by trans genes are weakly significant and not shown here (see Supplementary Data).

We also classified X-linked genes according to phenotypic inheritance patterns (Fig. 4a and c, Supplementary Tables 7 and 8). Compared to autosomal genes, X-linked have a similar percentage of conserved genes. As expected given the hemizygous sim X chromosome and lack of cis effects in the hybrid, there is little additivity (3%) or mel dominant inheritance (7%); X-linked genes have a strong excess of sim-dominant phenotypes (28%). X-linked genes are also more likely to be underdominant or overdominant compared to autosomal genes: 14.5% of X-linked genes are misexpressed compared to 10.8% of autosomal genes (G-test, P < 0.001), consistent with the faster-X hypothesis (Vicoso and Charlesworth 2006).

As with regulatory classes, Sfps and accessory gland-biased genes are much less likely to be conserved than all genes. Looking at the distributions of nonconserved classes (Supplementary Table 8), both gene classes are more likely to be additive than mel or sim dominant. Sfps and accessory gland-biased genes additionally have significantly higher levels of underdominance than overdominance, a departure from trends among all autosomal genes.

Beyond misexpression, we also classified genes that are not DE between parents and which have a gain-of-function (GOF, overexpression in hybrids) or loss-of-function (LOF, underexpression in hybrids) expression phenotype (Supplementary Data). There are 58 genes with significant GOF (12% of all overexpressed genes), and 40 with significant LOF (10% of all underexpressed genes)—representing relatively rare events. Further, restricting GOF to cases with insignificant expression in parents (log₂(counts) < 1 in P_mel and P_sim, log₂(counts) > 1 in F1_total) leaves only five instances, including prolyl-4-hydroxylase-α MP and four uncharacterized genes (Supplementary Table 9). There are two cases of LOF with insignificant hybrid expression (log₂(counts) > 1 in P_mel and P_sim, log₂(counts) < 1 in F1_total): β-Tubulin at 85D and Pendulin (Supplementary Table 10).

Next, we examined the relationship between regulatory and inheritance classes (Fig. 4d, Supplementary Table 11). We expect that genes with stronger cis components would be more likely to have an additive inheritance pattern, on the basis of the relative contributions of each species’ allele to total hybrid expression (Lemos et al. 2008; McManus et al. 2010). We found that genes with strong cis regulatory components had the highest levels of additivity: 31.8% of cis and 31.9% of cis + trans regulated genes had additive inheritance, compared to just 5.9% of trans regulated genes. We expected that genes with antagonistic cis and trans components would be more likely to lead to misexpression in hybrids, highlighting potential incompatibilities between species. Indeed, we find that cis-by-trans and compensatory gene classes are more likely than others to be associated with underdominant/overdominant inheritance. Cis-by-trans regulated genes have an excess of overdominance (22%) relative to underdominance (7%), but this may be attributable to the small sample size (n = 104 genes). Finally, we observe a strong trend towards trans regulated genes being inherited in a mel-dominant fashion (46% of trans regulated genes, compared to just 27% being sim-dominant; mel-dominant genes have a significantly greater proportion of trans regulation: G-test, P < 0.001).

GO enrichment analysis

To investigate possible biological correlates of regulatory and inheritance classes we investigated the associated GO enrichments (Supplementary Fig. 4, Supplementary Data). To increase the sample of X-linked genes, we used genes overexpressed or underexpressed in the hybrid relative to D. simulans, rather than genes strictly classified as overdominant or underdominant. Purely cis-regulated, cis-by-trans, and conserved genes are generally associated with larger P values and/or weakly enriched GO terms. Among our more significant results are 73 terms associated with translation in purely trans-regulated genes, driven by several ribosomal subunit and eukaryotic elongation factor genes (Supplementary Fig. 4a, Supplementary Data). Translation-related genes are also significantly enriched among mel-dominant, and particularly for overdominant inheritance. Among the 40 translation proteins that are trans and mel-dominant, 32 are more highly expressed in D. melanogaster (chi-square test, P = 0.005). The remaining 33 genes exhibiting other modes of inheritance show no bias towards either parent (chi-square test, P = 0.82). There are 61 overdominant translation-related genes on the autosomes. This gene set only partially overlaps with the trans-regulated gene set—18 of overdominant translation-related genes are trans-regulated, but eight are cis + trans, six are cis by trans, 12 are compensatory, and 17 are ambiguous. Translation-related GO terms are also enriched in overexpressed genes on the X chromosome (Supplementary Fig. 4b and c, Supplementary Data). On the X chromosome, there are an additional 52 overexpressed genes associated with translation. Taken together, the data suggest that translation-related genes are especially likely to be both trans-regulated and overdominant, but that overdominance may be the result of diverse regulatory mechanisms.

Underdominant inheritance/underexpression is strongly associated with golgi/endoplasmic reticulum vesicle transport GO terms on both the autosomes (137 genes, Supplementary Fig. 4b) and X chromosome (36 genes, Supplementary Fig. 4c). Of 137 underdominant transport-related genes on the autosomes, 86 have an ambiguous regulatory classification, while 20 are trans, 18 are cis + trans, and 13 are compensatory. Of the ambiguous terms, all are non-DE between parents, and also non-DE between hybrid alleles. Therefore, there is no evidence of cis effects in these genes. Underdominance is indicative of trans factors, however these effects have not led to divergence between D. melanogaster and D. simulans, suggesting trans effects that occur specifically in the hybrid.

Upstream sequence divergence

To investigate the possible sequence basis of expression variation, we characterized divergence in the regions upstream of AG-expressed genes. We analyzed distributions of substitution rate for various upstream sequence lengths (Supplementary Fig. 5); the 300 bp region captured the highest overall levels of divergence, so we chose this set for further analysis. We observe significant variation in upstream sequence evolution among regulatory and inheritance classes (Kruskal tests, P < 0.01). Among regulatory classes, conserved genes have the lowest rate of upstream sequence divergence with a median of 0.070 substitutions/bp (Fig. 5a). All other classes except cis by trans have significantly greater divergence rates (Wilcoxon rank sum tests, P < 0.01). Ambiguous and compensatory genes have the greatest rates at 0.088 substitutions/bp. Among inheritance classes, additive and conserved genes are very similar; median = 0.074 per bp and 0.073 per bp, respectively (Wilcoxon rank sum test P > 0.05). Alternatively, underdominant genes have much higher rates of upstream sequence divergence with a median = 0.104/bp (pairwise Wilcoxon rank sum tests, P < 0.001 vs all other classes). Given the enrichment of underdominant genes for golgi/protein transport-related GO terms, we asked whether those genes were confounded with the elevated level of upstream sequence divergence. Of 451 underdominant genes with upstream sequence information, 109 are associated with golgi/protein transport-related GO terms. If we remove these from the analysis, underdominant genes still have significantly greater upstream sequence divergence than all other classes (median = 0.101/bp, pairwise Wilcoxon rank sum tests, P < 0.001). Genes with mel or sim dominance exhibit intermediate levels of upstream sequence divergence (medians = 0.086 and 0.080/bp, respectively). Since cis-regulatory evolution could proceed through mutations in promoter regions, we asked whether the magnitude of ASE or parental divergence is correlated with upstream sequence divergence, however, we observed no relationship (Supplementary Fig. 6). Upstream sequence divergence in Sfps or AG-biased genes does not differ significantly from non-Sfps/non-AG-biased genes (Wilcoxon rank sum tests, P = 0.16, P = 0.26, respectively). Overall, then, the most obvious correlate of upstream sequence divergence is hybrid misexpression.

Fig. 5.

a) Distributions of Kimura-2-parameter estimated substitution rates among regulatory and inheritance classes. b) Distributions of nonsynonomous substitution rate (dN) among regulatory and inheritance classes. K2P distance and dN vary significantly across (Kruskal tests, P < 0.001). Alongside the median, significant differences by pairwise Wilcoxon rank sum tests (Holm–Bonferroni adjusted P < 0.05) are indicated by different numbers across gene sets.

Protein sequence evolution

Previous studies have found positive correlations between gene expression divergence and protein sequence evolution (Makova and Li 2003; Nuzhdin et al. 2004; Jordan et al. 2005; Khaitovich et al. 2005; Lemos et al. 2005; Liao and Zhang 2006; Sartor et al. 2006; Hunt et al. 2013; Warnefors and Kaessmann 2013; Hodgins et al. 2016; Zhong et al. 2021). We investigated whether this was the case for AG-expressed genes. Additionally, we investigated whether rates of protein sequence evolution differ among regulatory and inheritance classes.

We observed no association between protein sequence evolution (dN) and expression divergence between parents or ASE (Supplementary Fig. 7a, b). Additionally, we did not observe a strong difference between dN and regulatory or inheritance classes (Fig. 5b). A multivariate regression of dN by expression level and parental expression divergence suggests that genes with parental conservation have greater dN than DE genes, though effect sizes are very small (average expression: β = 1.3 × 10⁻³ ± 8.1 × 10⁻⁵, P < 0.001; parental expression DE: β = 1.7 × 10⁻³ ± 6.4 × 10⁻⁴, P = 0.007). Notably, we observe that genes with higher expression levels tend to have lower dN, consistent with the literature (Pál et al. 2001; reviewed in Drummond et al. 2005).

Median dN is 3.6 times greater for Sfps than non-Sfps (Wilcoxon rank sum tests, P < 0.001). Among AG-biased non-Sfps, dN is modestly elevated, 1.3 times higher, but still significantly different from nonaccessory gland-biased genes (Wilcoxon rank sum tests, P = 0.00101, P < 0.001) (Supplementary Fig. 7c, d). Increased protein divergence could be explained by directional selection (Tsaur et al. 1998; Aguadé 1999; Begun et al. 2000; Holloway and Begun 2004; Begun et al. 2006; Schully and Hellberg 2006; Wong et al. 2008; Majane et al. 2022) or reduced constraint (Dapper and Wade 2020; Patlar et al. 2021). To seek evidence bearing on these alternatives we used McDonald–Kreitman tests and compared the summary statistic α (higher α suggests a greater overall proportion of adaptive amino acid substitutions) among gene classes. The resulting patterns are similar to—though weaker than—patterns in dN (Supplementary Fig. 8a). In both regulatory and inheritance classes, conserved genes have significantly higher median α than some other types, but we do not observe significant differences among other classifications. As with dN, we observe significantly elevated α among Sfps (Supplementary Fig. 8b): median Sfp α = 0.256; non-Sfp median α = −0.375 (Wilcoxon rank sum test, P < 0.001). Unlike dN, α does not differ significantly among AG-biased and nonbiased genes (Supplementary Fig. 8c, Wilcoxon rank sum test, P = 0.52). Thus, while Sfps and AG-biased non-Sfps tend to exhibit similar patterns of regulatory divergence, recurrent adaptive protein divergence appears to be more prevalent among Sfps.

Chromatin state integration

We used ATAC-Seq data from D. melanogaster and D. simulans to investigate connections between chromatin accessibility and DE or ASE. We annotated ATAC-Seq peaks as conserved, mel orphans, or sim orphans. Conserved peaks are called in orthologous regions of both species, whereas orphan peaks are called only in one species.

In total, we annotated 7,416 conserved, 2,370 orphan sim, and 1,680 orphan mel peaks. We made peak-to-gene associations annotating peaks to the closest/overlapping transcription start site (TSS), which left us with 2,898 conserved, 1,627 orphan sim, and 1,232 orphan mel peaks with 1-to-1 gene annotations. Among annotated conserved peaks, 88% overlap the TSS. Roughly 6% of nonoverlapping peaks are upstream of the TSS, and 6% are downstream. The median width of conserved peaks is 644 bp in both D. melanogaster and D. simulans, while the mean is 759.5 bp in melanogaster and 758 bp in simulans. Orphan peaks are much less likely to overlap with the TSS: 16% of orphan mel and 19% of orphan sim have overlap. Orphan peaks are also more likely to be upstream than downstream. In D. melanogaster, 57% of orphan peaks are upstream while 26% are downstream; in D. simulans, 50% are upstream and 30% are downstream. Orphan peaks are also smaller than conserved peaks: the median width of mel orphans is 399 bp, while the median width of sim orphans is 255 bp. PCA of log₂(counts) shows that replicates cluster together by species (Supplementary Fig. 9a–c), though we note that clustering is not as strong as RNA-Seq data, which is expected due to the background and variance inherent to ATAC-Seq data.

Relative accessibility between species in conserved peaks is normally distributed (Supplementary Fig. 10a; 25% percentile log₂(mel/sim) = −0.205; 75% percentile = 0.212), suggesting there is no systematic directionality in chromatin accessibility between species. The distribution of log₂(mel/sim) for orphan peaks is highly skewed towards each respective species [Supplementary Fig. 10b, and c; mel orphans: 25% percentile log₂(mel/sim) = 0.48; 75% percentile = 1.50; sim orphans: 25% percentile log₂(mel/sim) = −1.48; 75% percentile = −0.48], in line with the expectation that the species with a peak called will have higher accessibility. There are a small number of cases where the species without a peak has higher accessibility (2.3% of mel orphans, 4.9% of sim orphans). We removed from further consideration (A) orphan peaks that are less accessible in the species with a peak present and (B) orphan peaks that are not DA, leaving 74% of mel and 72% of sim orphans (Supplementary Fig. 10d, e).

There is a weak positive relationship between accessibility and expression (Fig. 6a and b), suggesting that chromatin state and expression are indeed correlated (Spearman rank correlations of gene expression with conserved chromatin peak accessibility: mel: ρ = 0.296; sim: ρ = 0.312). Given this relationship, we asked whether the presence of chromatin peaks was associated with the likelihood of DE in nearby genes. In both parental and ASE contrasts, DE genes are enriched for the presence of nearby orphan peaks relative to non-DE genes (Fig. 6c, d, Supplementary Tables 12 and 13). We used a multiple logistic regression with average expression and peak status as independent variables and DE as the dependent variable (DE ∼ log₂(counts) + peak). In the parental contrast, presence of a conserved peak negatively predict DE (β = −0.17 ± 0.08, P = 0.046), and orphan peaks are strong predictors of DE (mel orphan: β = 0.78 ± 0.11, P < 0.001, sim orphan: β = 0.86 ± 0.09, P < 0.001). Similarly, conserved peaks are negatively correlated with DE in ASE (β = −0.30 ± 0.10, P = 0.002), and orphan peaks strongly predict DE (mel orphan: β = 0.85 ± 0.11, P < 0.001, sim orphan: β = 0.77 ± 0.10, P < 0.001). Average gene expression also predicts DE in both contrasts, but the regression coefficients are notably smaller in comparison with orphan peak presence (parental: β = 0.22 ± 0.01, P < 0.001; ASE: β = 0.21 ± 0.01, P < 0.001).

Fig. 6.

Interfacing promoter region accessibility estimated from ATAC-Seq data with gene expression. Gene expression and conserved peak accessibility have a positive relationship in a) P_mel and b) P_sim. Spearman's rank coefficient ρ is displayed. DE genes are more likely to be associated with orphan peaks in both the c) parental and d) ASE contrast. e) Parental expression divergence and f) ASE (y-axis) for are plotted against accessibility differences for DA conserved chromatin peaks called by ATAC-Seq (x-axis). g) Distributions of log₂(fold change) of parental expression difference differ among peak types. M: median, Q1: 1st quartile, Q3: 3rd quartile, S: skewness. Genes associated with conserved peaks have a broader 1st–3rd interquartile range and significantly larger median absolute value than genes without an annotated promoter peak. Orphan peaks are skewed towards greater expression values in the species with a peak. h) Distributions of log₂(fold change) of ASE show similar patterns to parental divergence.

We further investigated correlates of gene expression divergence with presence or absence of nearby 1-to-1 peaks by comparing log₂(fold changes) (Fig. 6g). In the parental contrast (P_mel−P_sim), genes without a peak annotation have a narrower distribution of log₂(P_mel/P_sim) relative to peaks with a conserved peak nearby, but both gene sets have medians near 0. The median absolute value of log₂(P_mel/P_sim) in genes with a conserved peak is 0.37, significantly higher than genes with no peak, median = 0.29 (Wilcoxon rank sum test, P < 0.001). Genes with a mel orphan peak nearby are biased towards positive values of log₂(P_mel/P_sim) with median = 0.252, and genes with a sim orphan peak nearby are biased towards negative values with median = −0.221. The median absolute values of log₂(P_mel/P_sim) of genes near orphan peaks are significantly greater than genes near a conserved peak or no peak (median mel orphan = 0.61; sim orphan = 0.59; Kruskal test, P < 0.001, pairwise Wilcoxon rank sum tests, P < 0.001 in each case). Medians associated with mel peaks and sim peaks are not significantly different (Wilcoxon rank sum test, P = 0.30). We observe the same patterns in ASE among different classes of peaks, but the magnitude of expression differences is smaller compared to parental DE (Fig. 6h).

Next, we asked whether the magnitude of expression differences between species or ASE was correlated with the magnitude of peak accessibility differences among DA peaks. We find a positive relationship between ranks of these measures (Fig. 6e and f; conserved peaks, parental divergence: ρ = 0.35; ASE: ρ = 0.22). These correlations are, however, weaker for orphan peaks (Supplementary Fig. 11), consistent with a less linear relationship between chromatin conformation and expression for species-specific peaks. In summary, it appears that the presence of accessible chromatin is correlated with differential expression of nearby genes, and that quantitative differences in conserved peak accessibility are correlated with concordant differences in gene expression. However, there is no strong quantitative relationship for orphan peaks.

To investigate the relationship between chromatin accessibility and cis/trans regulatory divergence, we compared Spearman rank correlations between expression divergence and accessibility divergence among conserved, pure cis, pure trans, and cis + trans regulated genes (Supplementary Table 14). As expected, conserved genes exhibited no correlation of accessibility divergence with either parental expression (ρ = −0.01) or ASE divergence (ρ = −0.08). Genes with trans-regulatory components exhibit a moderately stronger correlation of accessibility divergence with parental expression divergence (ρ = 0.52 for both trans and cis + trans) than pure cis-regulated genes (ρ = 0.41). Genes with trans-regulatory components have a weak correlation of accessibility divergence with ASE divergence (ρ = 0.14), while cis (ρ = 0.42) and cis + trans (ρ = 0.44) genes have relatively strong correlations. It therefore appears that only cis-regulated genes show strong correlations between accessibility divergence and ASE divergence.

Regulatory and inheritance classes are associated with different proportions of chromatin peak types (Supplementary Fig. 12, Supplementary Table 15). As expected, genes that are annotated as conserved in regulation or inheritance are much less likely to be associated with nearby peaks. The set of genes with purely cis regulation have a smaller proportion of conserved peaks than genes with significant trans factors. Genes with cis + trans regulation, and genes with additive inheritance, have the highest respective shares of orphan peaks. Underdominant genes have the highest share of conserved peaks among inheritance classes.

Focusing on genes likely to play a specific role in accessory gland function, we observe an enrichment of overall cis-effects (cis-only effects and cis + trans effects) driving expression divergence for Sfps and AG-biased genes relative to all autosomal genes (Supplementary Tables 5 and 6), potentially under directional selection, which led us to investigate the properties of chromatin accessibility for these genes (Fig. 7). As was observed for all expressed genes, there are positive correlations between expression divergence and peak accessibility for Sfps and AG-biased genes (Fig. 7a–g). Notably, genes associated with mel orphan peaks among both SFPs and AG-biased genes had very strong associations with increased expression in mel relative to sim (Fig. 7i–l). Another remarkable difference between all AG-expressed genes vs Sfps and AG-biased genes is the stronger positive association between mel orphan peaks and expression divergence for the latter (Fig. 7g and h) than for the former (Fig. 7e and f). This supports the idea that gain or loss of cis-regulatory elements plays a greater role in the expression divergence of genes with specific AG function than it does in the more general transcriptome divergence of the gland, a phenomenon which may contribute to their general faster rate of expression evolution and which is consistent with other evidence that expression novelty is relatively common in this tissue (Cridland et al. 2020).

Fig. 7.

Interfacing promoter region accessibility estimated from ATAC-Seq data with gene expression, subset to SFPs and AG-biased non-SFPs (AGB). Gene expression and conserved peak accessibility have a positive relationship among SFPs in a) P_mel and b) P_sim, and among AG-biased non-SFPs in c) P_mel and d) P_sim. e) Parental expression divergence and f) ASE (y-axis) in SFPs are plotted against accessibility differences for DA conserved chromatin peaks called by ATAC-Seq (x-axis). g, h) AG-biased non-SFPs show a similar, stronger result. i, j) Distributions of log₂(fold change) of parental expression difference differ among peak types. M: median, Q1: 1st quartile, Q3: 3rd quartile, S: skewness. Genes associated with conserved peaks have a similar 1st–3rd interquartile range and do not differ significantly in median absolute value from genes without an annotated promoter peak. Orphan mel peaks are skewed towards greater expression values in the species with a peak. k, l) Distributions of log₂(fold change) of ASE show similar patterns to parental divergence.

Discussion

Given that D. melanogaster and D. simulans are sister species that share a most recent common ancestor only 2–3 million years ago (Obbard et al. 2012), we expect strongly correlated phenotypes for most traits. In agreement with this expectation, autosomal gene expression profiles between D. melanogaster and D. simulans were very similar, as evidenced by transcriptome-wide expression correlations (r = 0.934). This correlation is somewhat stronger than that observed in our recent bulk RNA-Seq (Cridland et al. 2020) and single-cell RNA-Seq studies (Majane et al. 2022) of the same tissues. Given the myriad technical differences between studies, it is difficult to interpret variation in the magnitudes of these correlations. Similarly, the lack of similar datasets for several organs in these two species, subjected to the same analysis pipeline, makes it difficult to draw strong conclusions about whether rates of transcriptome divergence in the AG are unusually large (Cridland et al. 2020).

Hybrid expression profiles are overall more similar to each parent than the parents are to each other and are roughly midpoint between parental expression in PCA. ASE profiles within the hybrid are more similar than parents are to one another, as expected, given that parental divergence is the result of cis and trans effects, while divergence between hybrid alleles is driven only by cis effects. Similarly, we observed a higher proportion of DE genes between parents than between parental alleles in the F1. As we see with transcriptomic correlations, large effect-size DE between hybrid alleles and parent-of-origin is rare (4% of genes), suggesting that trans-effects of large size are uncommon.

X-linked and autosomal gene expression profiles have very similar correlations between parents and also exhibit similar levels of DE. These results are unexpected since previous whole-animal transcriptome data shows a strong faster-X effect on expression divergence, even among nonmale biased genes (Kayserili et al. 2012; Meisel et al. 2012). We observed a modest rate of DE between the hybrid and D. simulans X chromosomes, suggesting that while trans effects on the X are not very rare, the effect-sizes are generally small.

Since Sfps exhibit rapid protein divergence (Tsaur et al. 1998; Aguadé 1999; Begun et al. 2000; Holloway and Begun 2004; Begun et al. 2006; Schully and Hellberg 2006; Wong et al. 2008; Majane et al. 2022) and genomic turnover (Wagstaff and Begun 2005; Mueller et al. 2005; Begun et al. 2006; Hurtado et al. 2022), we formally investigated their interspecific expression divergence. Sfps are much more likely to have large effect-size DE, consistent with previous work on major expression divergence in this tissue (Cridland et al. 2020). However, accessory gland-biased non-Sfps exhibit similar patterns. Thus, genes that make the largest contribution to the unique transcriptome of the accessory gland, regardless of whether they are Sfps, tend to exhibit greater expression divergence. Determining whether or not this property of the AG transcriptome is driven by directional selection is an important unsolved problem.

The median cis effect is 43% larger than the median trans effect, consistent with strong correlations of expression between hybrid alleles and their parents of origin. In contrast, McManus et al. (2010) observed significantly larger trans effects in whole female Drosophila hybrids. However, despite their smaller mean effect-size, trans effects are more common than cis effects. Some evidence suggests that cis effects contribute more to species expression divergence than trans-effects, perhaps as a consequence of their lower average pleiotropy (reviewed in Signor and Nuzhdin 2018; Hill et al. 2021). Our results do not support this model, in line with results from some other studies (McManus et al. 2010; Coolon et al. 2014; Sánchez-Ramírez et al. 2021). Additional investigation of diverse somatic organs to understand the distribution of cis and trans effects across tissues, which could help answer a broader question of which factors influence the variation observed in relative levels of interspecific cis and trans divergence.

We classified genes into regulatory and inheritance classes as originally outlined in McManus et al. (2010). Equal proportions of genes have evolved through purely cis or trans regulation, with a larger proportion exhibiting evidence of both mechanisms, consistent with a complex basis of regulatory divergence. Opposing directionality of cis and trans evolution (cis by trans and compensatory classes) is rare in our study. Other studies in Drosophila found that opposing cis and trans effects were much more common (Ranz et al. 2004; Landry et al. 2005; Graze et al. 2009; McManus et al. 2010; Coolon et al. 2014). Studies of mouse liver (Goncalves et al. 2012) and testis (Mack et al. 2016) also have relatively high levels of opposing cis and trans effects. Fraser (2019) found that errors in estimation of ASE can lead to inflated estimates of cis by trans effects; however, suggesting that comparisons of these effects across studies may be complicated by technical issues. Relative to all AG-expressed genes, Sfps and accessory gland-biased genes are more likely to exhibit additive inheritance and accumulate substantially higher levels of cis + trans regulation divergence, consistent with their greater expression evolution, as expected under directional selection. Sfps and accessory gland-biased genes are also more likely than other genes to exhibit elevated misexpression, also consistent with an enrichment of functionally significant regulatory substitutions generating incompatibilities (see below).

In general, misexpression in the hybrid AG is relatively rare, with just 10.8% of autosomal genes overdominant or underdominant, consistent with Ranz et al. (2023). Further, just a handful of genes have complete GOF or LOF expression phenotypes. These results suggest that the accessory gland is not prone to widespread dysgenesis between these species, in contrast to results from Drosophila testis or female whole-animal data (Ranz et al. 2004; Haerty and Singh 2006; Moehring et al. 2007; McManus et al. 2010; Coolon et al. 2014; Cartwright and Lott 2020; Go and Civetta, 2020; Go and Civetta 2022; Ranz et al. 2023), Caenorhabditis data (Sánchez-Ramírez et al. 2021), mouse liver (Goncalves et al. 2012), and testis (Mack et al. 2016) data. Thus, our transcriptomic data are consistent with the observations of Stumm-Zollinger and Chen (1988), that hybrid accessory glands have relatively normal morphology, seminal fluid, and ability to induce the female PMR. Other studies have found limited levels of misexpression in Drosophila larva (Moehring et al. 2007; Wei et al. 2014), female heads (Graze et al. 2009), Hawai’ian Drosophila testes (Brill et al. 2016), and avian brains (Davidson and Balakrishnan 2016). Clearly, the level of hybrid dysgenesis in gene expression is highly variable among species and tissues; more work on tissue-specific ASE is needed to reveal the proximate and ultimate explanation for this variation.

Several Drosophila studies have reported that male-biased genes are prone to misexpression in hybrids and are especially likely to be underdominant in whole animals or testis (Haerty and Singh 2006; Michalak and Noor 2003; Moehring et al. 2007; McManus et al. 2010), a pattern observed in most, but not all Drosophila crosses (Banho et al. 2021). Underdominant male-biased genes are also linked to male sterility (Michalak and Noor 2004). Importantly, we found higher levels of misexpression among AG-biased genes, with a particular enrichment for underdominance. Autosomal genes have similar rates of overdominance and underdominance. Sfps have 2.7 times as many misexpressed genes compared to non-Sfps, with a ratio of underdominant to overdominant expression of 2.3; accessory gland-biased non-Sfps are have 2.9 times as many misexpressed genes, with a ratio of underdominant to overdominant expression of 3.4. These data suggest that the widely reported observation of male biased underexpression is not limited just to the testis. Whether this pattern holds for male-biased genes across multiple somatic tissues or only those related to reproduction is an important question for future studies. Future work making more explicit connections between directional selection on regulatory divergence and hybrid misexpression is desirable.

The faster-X hypothesis (Vicoso and Charlesworth 2006) predicts not only elevated expression divergence among X-linked genes, but also increased X-linked misexpression in hybrids. Faster-X divergence has been observed in flies (Kayserili et al. 2012; Meisel et al. 2012), but Drosophila hybrids actually have a lower rates of X misexpression in the testes (Lu et al. 2010; Llopart 2012) or no difference from autosomal misexpression in larvae (Wei et al. 2014). In contrast, our data reveal no evidence of faster-X divergence but do show faster-X misexpression, similar to patterns observed in mice (Good et al. 2010; Larson et al. 2016). Faster-X gene expression patterns may therefore vary among tissues in Drosophila, and future studies of tissue-specific hybrid gene expression are needed to determine the extent and biology underlying these potential differences.

Genes that are autosomal overdominant or X-linked overexpressed are both highly enriched for translation-related genes, including numerous elongation factors and ribosomal subunits. Overexpressed proteins in D. melanogaster–D. simulans hybrid embryos were enriched for “translation initiation” genes (Bamberger et al. 2018), and misexpressed genes in hybrid house mice testes were also significantly enriched for translation-related GO terms (Mack et al. 2016), suggesting that misexpression of translation-related genes could contribute to hybrid incompatibilities across species and developmental stages. Notably, hybrid male sterility evolves very quickly and the regulation of Drosophila spermatogenesis occurs primarily at the translational level (Schäfer et al. 1995). Whether reduced hybrid male fertility and misexpression of translational machinery are functionally linked remains a speculative matter. Translation is also enriched among trans-regulated and mel-dominant genes in our data, but these gene sets are only partially overlapping with one another, suggesting that translation-related genes may be regulated and inherited via diverse mechanisms in the accessory gland, and that overdominant translation-related genes are not regulated through a unifying mechanism. Underdominant autosomal and underexpressed X-linked genes are highly enriched for genes related to protein transport, golgi, and endoplasmic reticulum. Notably, ambiguous regulated genes are also enriched for these GO terms, and this gene list substantially overlaps with underdominance. Ambiguous regulation may occur in many ways and is difficult to put into biological context. However, in this case, the vast majority of these genes are not DE between the parents but do have evidence of trans-effects leading to underdominance. This suggests that emergent properties of trans factors active specifically in hybrid cells leads to underexpression, potentially indicative of hybrid incompatibilities related to protein transport in the golgi and endoplasmic reticulum.

We analyzed levels of nucleotide divergence in regions upstream of the TSS, which could plausibly affect expression divergence. While we do not find a quantitative relationship between expression divergence and upstream sequence divergence, genes that are conserved in their regulation and inheritance tend to have a lower level of divergence. Compensatory, ambiguous, and particularly underdominant genes have elevated levels of upstream sequence divergence, suggesting that underdominance might be arising from incompatibilities between rapidly evolving cis-acting sequences and trans regulatory factors.

As expected, Sfps have much higher dN than non-Sfps. Sfps also have elevated median α, suggesting overall greater levels of adaptive substitutions in these genes. While accessory gland-biased non-Sfps also have elevated dN, they do not exhibit significant elevation in α. Thus, from a regulatory perspective, Sfps are not obviously different from other AG-biased genes. However, Sfps appear to experience much more recurrent adaptive protein divergence than other AG-biased genes. This supports the idea that proteins that interact directly with the external environment are more likely to evolve adaptively (Gillespie and Langley 1974). Somewhat unexpectedly, we found that among AG-expressed genes, those in the conserved regulatory category have the highest overall levels of dN, ω, and α. This suggests that for most genes expressed in the accessory gland, there is a modest decoupling between expression evolution in the gland and rapid protein evolution driven by selection.

We associated ATAC-seq peaks with nearest TSS to identify chromatin tied to putative regulatory regions. We found that orphan peaks were highly biased in accessibility towards one species and observed a modest but significant correlation between gene expression and accessibility among conserved peaks, similar to results of some studies (Nair et al. 2021), but weaker than that observed in others (Starks et al. 2019). We observed weaker correlations of expression divergence with orphan peaks, which may be at least partially explained by the greater median distance of orphan peaks to the TSS. However, orphan peaks physically associated with Sfps or AG-biased genes are associated with greater expression divergence, suggestive of differential importance of changes in chromatin conformation for expression evolution for these genes compared to the whole gland transcriptome. We further identified a weakly positive quantitative relationship between differential accessibility of conserved peaks and the log(fold-change) of DE, in contrast to some other studies that found stronger associations (Racioppi et al. 2019; Gontarz et al. 2020; Nair et al. 2021; Sanghi et al. 2021). However, limiting the analysis to genes with evidence of cis or trans regulatory divergence, reveals a relatively strong correlation between species differences in peak accessibility divergence and gene expression divergence. While orphan peaks do not appear to have this quantitative relationship in general, the presence of an orphan peak strongly predicts both small- and large-effect size DE in nearby genes. Taken together, the data suggest that presence/absence of chromatin peaks (either by evolutionary gain or loss, which is impossible to determine with this data) likely contributes to gene expression differences between D. melanogaster and D. simulans—if a peak appears in one species, there is a better chance that the nearest gene will be DE, and more often than not in the direction of the species with the peak—but with the exception of Sfps and AG-biased genes, our data provide no evidence of a straightforward quantitative relationship. To overcome some of the technical and biological variables that complicate this analysis, an allele-specific multiomic gene expression and ATAC-seq experiment on single cells (Cao et al. 2018; S. Chen, Lake, et al. 2019) from F1 animals would provide stronger insights into both the genetics of regulatory divergence across cell types and organs, and the relationships between chromatin state, expression, and regulation.

Data availability

Sequence data can be found at NCBI under BioProject numbers PRJNA913156 (RNA-seq reads, https://www.ncbi.nlm.nih.gov/bioproject/?term=PRJNA913156) and PRJNA993436 (ATAC-seq reads, https://www.ncbi.nlm.nih.gov/bioproject/?term=PRJNA993436). Supplemental data and analysis scripts can be found at https://github.com/alexmajane/hybridASE.

Supplemental material available at GENETICS online.

Funding

This work was supported by National Institutes of Health grant NIGMS R35GM134930 to DJB and a National Science Foundation Graduate Research Fellowship Program grant to ACM. The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health.