2.1 Differential gene expression reveals a cohort of consistently ovary-specific genes

We observed several patterns in tissue-specific gene expression that are consistent across all twelve species. First, in all species the main axis of variation separated ovary RNA libraries from head and carcass (Fig D in S1 Text). This axis accounted for at least 50% of variation, and in several species greater than 70% of variation. To test for possible variation due to different runs on the sequencer, we resequenced several libraries and compared them using principle component analysis. We found variation between sequencing runs to be negligible compared to variation across tissues and individuals (Fig G in S1 Text). Second, in all species we observed that there was a larger amount of significantly downregulated transcripts than upregulated in the ovary relative to the carcass (Figs 2A and 2B, E in S1 Text). Across species, we observed an average of 25.4% to be significantly downregulated and 12.9% of transcripts to be significantly upregulated. When comparing the head to the carcass, we observed an average of 7% of transcripts to be significantly upregulated and 7.2% to be significantly downregulated (Fig F in S1 Text). These differences may reflect variation in the complexity and diversity of functions of the body parts being compared.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Identifying a cohort of ovary-specific genes across drosophilid species.

A, Volcano plot for one example species, D. primaeva (Dpri), showing the results of a differential gene expression analysis comparing the ovary to the carcass. The x-axis shows the log₂ fold change of expression across transcripts, and the y-axis shows the adjusted p-value, log₁₀ transformed. Points that are significantly differentially expressed are shown in black. B, Jitter plots showing the results of the same analysis across the twelve species studied here. The x-axis shows the log₂ fold change of expression across transcripts, and points are arranged with random jitter within species on the y-axis. C, The same jitter plots as in B, now colored according to whether or not transcripts belong to a cohort of core ovary genes. These are defined as genes, grouped by BLAST similarity to D. melanogaster transcripts, for which at least one transcript is upregulated in the ovary of ten or more of the twelve species. D, A comparison of mean expression change across Hawaiian species to reported ovary-enrichment values from D. melanogaster, as reported in FlyAtlas2 [29]. Core ovary genes are marked in magenta. E, The boxed region shown in D, magnified and now showing only core ovary genes, annotated with the gene symbol from D. melanogaster.

https://doi.org/10.1371/journal.pgen.1010607.g002

We used the results of our differential gene expression analysis within species to test for the existence of a suite of genes that show consistent ovary-specific expression across species. We found a cohort of 131 genes, grouped according to BLAST sequence similarity to D. melanogaster, for which at least one transcript was significantly upregulated in the ovaries of more than ten species (Fig 2C). Transcripts matching these genes made up on average 24.6% of the significantly ovary-upregulated transcripts across species, meaning roughly one quarter of ovary-specific genes have conserved expression patterns over evolutionary time. When excluding the species S. varia, this average decreased to 17.7%, as this species has the smallest set of ovary-upregulated transcripts, 100% of which correspond to core ovary genes.

We then tested the extent to which these core ovary genes correspond to observations in well-studied laboratory Drosophila models. To accomplish this, we compared expression across Hawaiian species to reported tissue-specific expression levels from D. melanogaster [29]. We found that Hawaiian core ovary-specific genes show nearly universal enrichment in the ovary of D. melanogaster as well, as reported in the FlyAtlas2 dataset [29] (Fig 2D). We likewise observed that genes reported in D. melanogaster to have highest enrichment in the ovary largely correspond to genes that are significantly upregulated in the ovaries of Hawaiian species (Fig H in S1 Text).

The 131 core ovary genes include several well-known members involved in oogenesis and germline stem cell renewal such as nanos [34], swallow [35], and oskar [46](Fig 2E). We found only two genes that were identified as Hawaiian core ovary genes that are not reported in the FlyAtlas2 dataset [29] to be enriched in the ovary of D. melanogaster: the SET domain binding factor sbf, and Rfx, which are reported to be enriched in the heart, brain, and other non-reproductive tissues of D. melanogaster [29].

We used the same approach to identify a core suite of 52 head-specific genes (Fig I in S1 Text). There was no overlap between the sets of core head genes and core ovary genes. To test whether the correspondence between expression observations in Hawaiian flies and D. melanogaster might be due to factors beyond tissue identity, we compared head expression values to ovary enrichment data from D. melanogaster, as we had done for ovary expression values above. We did not observe a correspondence in either direction between expression in the head of Hawaiian species and enrichment in the ovary of D. melanogaster (Fig J, panel a, in S1 Text). In contrast, we did find a correspondence between head-specific expression and genes enriched in the D. melanogaster brain, eye, and head (Fig J, panel b, in S1 Text). Core head genes include Rhodopsin photoreceptor genes and genes such as hikaru genki with involvement in synaptic centers [47].

2.2 Modeling reveals the phylogenetic decay of expression similarity between tissues

Many studies have investigated the question of whether we expect expression to be more similar across the same organ in different species, or across different organs within the same species [13–20]. Recent studies have suggested that the answer to this question will depend on the phylogenetic distance separating the species being compared [21]. Here we used a modeling approach to investigate this question with respect to the ovaries of Hawaiian drosophilids.

First, we determined an appropriate unit of comparison across species, based on an assessment of homologous features between reference transcriptomes. The agalma pipeline provides a method for determining homologous and orthologous sequences using an all-by-all BLAST approach to determine clusters of reciprocally similar sequences (homology groups). These can then be divided into orthology groups by estimating gene trees and identifying maximally inclusive subtrees with no more than one sequence per taxon [48]. We compared the representation of species across homology and orthology groups, and observed that while the representation of homology groups increases with the number of species compared, representation of orthology groups decreases (Fig K in S1 Text). This is a known obstacle in comparative transcriptomics, attributed to many transcripts being artifactually fragmented during reference transcriptome assembly [49]. To reduce the impact of this on our downstream analyses, we averaged TPM values across all transcripts within a homology group for each sequenced RNA library. Principle component analysis of this average expression dataset showed that the first principle component divides ovary libraries from the rest, while the second component separates samples along an axis that largely corresponds to phylogenetic distance between species (Fig L in S1 Text). While this averaging approach reduces noise due to variable mapping affinities of fragments of the same transcript, it comes at the cost of averaging over potential variation between genuine transcripts that fall into the same homology group. Future analyses using improved assemblies for transcriptomes or genomes will likely be able to avoid this trade off and compare transcript counts directly. To test the robustness of results to this averaging, we also performed key analyses over a dataset of the identifiable strict orthologs. In both datasets (based on homology groups or based on strict orthology groups), we averaged expression values across biological replicate individuals per species.

With average expression counts for homologous transcripts across species, we tested the degree to which variation across this dataset could be attributed to tissue-specific variation (here, ovary vs. carcass), species-specific variation, or neither (residual variation). Using the linear modeling approach adapted from Breschi and colleagues (2016) [21], we found the proportion of variance across the dataset attributed to tissue differences decreased with phylogenetic distance, while the proportion attributed to species difference increased (Fig 3A–3C). In addition, we found that, when comparing ovary and carcass tissues, the Hawaiian drosophilid clade encompasses the crossover point where variation across species swamps variation across tissues (crossed lines, Fig 3A). When comparing across the two species from the picture-wing group included in this study, an average of 45.6% of the variation can be attributed to tissue differences. For the same comparison, 960 genes were identified as tissue-variable genes (TVGs), defined as residual variation accounting for <25% and a two-fold increase in variation attributed to tissues than to species (Fig 3B and M in S1 Text). In contrast, when comparing across all twelve Hawaiian drosophilid species studied here, 34.7% of the variation can be attributed to tissue, with 240 TVGs (Fig 3B and M in S1 Text). Across different clades of comparisons, the number of species-variable genes (SVGs) remains relatively stable (from 304 to 260, Fig 3B).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Linear modeling shows the proportion of variance explained by differences across tissues and species.

A, The results of a linear modeling approach to calculate expression variation for each gene, attributed to variation across organs, species, or residual variation, as described in Breschi and colleagues (2016) [21]. The average proportion of variation attributed to tissues is higher than that attributed to species for the two picture-wing species in clade A, while the opposite is true for all twelve species in clade D. Sample sizes are for clade A: 2,918 homology groups, clade B: 2,044, clade C: 1,239, and clade D: 1,143. B, The number of genes, defined by homology group, classified as tissue variable genes (TVGs), species variable genes (SVGs), or neither in each clade comparison. C, The phylogeny of the twelve species studied here, showing the four clades compared in A-B. Scaled evolutionary distance is calculated as the relative distance from the most recent common ancestor of Hawaiian drosophilids to extant species. D, Comparing results of the differential gene expression approach (log₂ fold change) on the x-axis to results of the modeling approach on the y-axis (variation across tissues). Genes are colored according to TVGs and SVGs. The inset box highlights TVGs that are upregulated in the ovary relative to the carcass. E, The same plot, now showing only upregulated TVGs, annotated with the gene symbol from the D. melanogaster sequences in the same homology group.

https://doi.org/10.1371/journal.pgen.1010607.g003

We then leveraged the results of this linear modeling approach across all twelve species to perform an additional screen for genes that are consistently upregulated in ovaries across species. We compared the proportion of variation explained by tissue for each homology group to the average log₂ fold change from the results of our differential gene expression analysis (Fig 3D). This comparison allowed us to identify genes that fall above our threshold for TVGs that are also upregulated in the ovary (Fig 3E). This group of genes includes many of the same members as the core ovary genes (e.g. nanos and swallow), as well as several new candidates (e.g. singed).

To test the importance of tissue identity, we repeated the same analysis comparing variation across species and tissues using the head in place of the ovary. Consistent with what we describe for the ovary and carcass, as phylogenetic distance increases the proportion of variation across tissues decreases while variation across species increases. In contrast to the above findings, however, for the head and carcass far less of the variation in gene expression can be attributed to tissue differences (Fig N in S1 Text). For these tissues, the crossover point between total proportion of variation occurs roughly at the distance separating the two picture-wing species.

To verify these results were not driven by the species S. varia, which had the most distinct expression patterns of all species, we repeated these analyses excluding this species and recovered largely equivalent results (Fig O in S1 Text). To test robustness to homology group averaging, we repeated this analysis over strict orthologs, again recovering the same results (P in S1 Text). Finally, we also compared our findings to those that would be recovered using a more typical pairwise approach, by repeating the linear modeling analysis on ovary and carcass data using every pairwise combination of the twelve species. We recovered the same trend of decreasing contribution of tissue-level variation with increasing phylogenetic distance. While several pairs of species show more variation between species than tissues, we note that not every pair, nor the average across pairs, captures the crossover point where variation across species overwhelms variation across tissues (Fig Q in S1 Text). This reflects the inherent variability between individual pairwise comparisons of species, and highlights the importance of phylogenetic analytical approaches on entire clades.

2.3 Identifying gains and losses of ovary bias across genes and the phylogeny

While many ovary-specific transcripts belong to the cohort of core ovary genes, on average 75.4% of transcripts are upregulated in the ovaries of one or several species, but not consistently across ten or more of the species studied here (Fig 2B and 2C). This is suggestive of many evolutionary gains and losses of ovary-specific expression of genes. We characterized the evolution of these gains and losses using an ancestral state reconstruction approach. First we quantified expression bias between tissues as the ratio of read counts [7], then reconstructed the value of this continuous trait for each gene (defined using homology groups) at each node of the estimated species tree (Fig C in S1 Text). We then calculated the scaled change of expression bias along each branch, which allowed us to describe how relative expression values between tissues had changed the course of evolutionary time (Fig 4A). Visualizing the distribution of scaled changes by genes shows that most scaled changes are small and centered around zero, representing little change in gene expression bias between tissues (Fig 4B–4C).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 4. Identifying genes that have gained and lost ovary-biased expression across the phylogeny.

A, The phylogeny of the twelve species studied here, highlighting one example branch of the 22 for which we inferred the scaled evolutionary change in expression bias. B, The distribution of changes, grouped by gene, for 100 randomly selected genes, defined by homology group. Each point represents one of the 22 branches from A, with the red point corresponding to the highlighted branch from that panel. C, The distribution, log₁₀ transformed, of scaled genes across all branches and all genes. Changes on the highlighted branch in red. D, The phylogeny with all 22 branches numbered. E, The distribution of changes, grouped by branch, with random jitter on the x-axis within each group. Points colored according to the qualitative change in bias, either from more expression in ovary than carcass to less (blue), the reverse (red), or no change in overall bias (gray). F, The distribution of ancestral and descendant values, showing the two quadrants that represent qualitative changes in bias. Points that represent large swings in expression within those quadrants are labeled a-d. G, The four genes with large swings from F, showing the expression bias for each transcript colored according to more expression in the ovary (red) or carcass (blue). Panels annotated with the gene symbol from the D. melanogaster sequences in the same homology group, with the exception of vilya*, which was annotated using a direct BLAST search since no D. melanogaster sequence was present in that group.

https://doi.org/10.1371/journal.pgen.1010607.g004

Using this dataset of scaled changes across genes and branches, we identified branches for which the direction of tissue bias had changed (e.g. from higher expression in the ovary than in the carcass to lower, or vice versa). Visualizing this dataset according to branches reveals that the majority of these changes in bias are located on the root and terminal branches, rather than internal branches (Fig 4D and 4E). This is likely because internal branches for this rapid radiation tend to be very short; even when scaling evolutionary changes to branch length, it is less probable for our analysis to identify a shift to and from ovary-biased expression on a short branch than a long branch. Repeating the same analysis on a dataset of strict orthologs reflects the same pattern, indicating that this result is not an artefact of expression averaging across homology groups (Fig R in S1 Text).

Visualizing the distribution of genes by ancestral and descendant values allows us to identify shifts in bias which represent the largest swings in expression values (Fig 4F, points a-d). Highlighting the top two such shifts in both directions, we identify four example genes which acquired or lost ovary-specific expression in the phylogeny of Hawaiian Drosophilidae. In the case of FMRFaR and GABA, a few Hawaiian species have gained ovary-biased expression of these genes, while most species and the ancestral state indicate non-ovary bias (Fig 4Fa-b). In the case of vilya and the unnamed gene CG9109, each shows a pattern where one species has lost ovary bias from a biased ancestral state (Fig 4Fc-d).

Repeating the same analysis using the head in place of the ovary revealed a set of evolutionary gains and losses in head-specific expression (Fig S in S1 Text). Identifying the top four changes in head expression shows gains and losses of head expression in the genes hiro, stil, Jhe, and, consistent with the ovary, vilya. In the case of the latter, these results may be driven by substantial changes in expression of vilya in carcass tissues across species, resulting in major differences in both ovary and head-biased expression.

2.4 Genes with a strong correlation of expression evolution

We tested the estimated evolutionary changes in expression bias for evidence of correlated expression evolution between genes. For every gene represented across all species, we performed a pairwise comparison of changes in expression bias, using as data points the scaled change in ovary bias on the 22 branches in the phylogenetic tree. This resulted in 1,306,449 pairwise measures of evolutionary correlation between genes. Because the number of gene pairs being compared is much larger than the number of values used to estimate correlation, this method has the potential to produce many spurious correlations [7]. To reduce the number of these correlations that are driven by low-confidence outliers [50], we removed any transcript with an estimated change in expression bias beyond two standard deviations from the mean change across all transcripts, and performed the same pairwise correlation analysis.

To test the degree to which the correlations observed here reflect known biological interactions between genes, we compared these measures to reported protein and genetic interactions between genes, using the database of published genetic experiments in D. melanogaster, available at http://flybase.org. We found that the mean correlation coefficient for genes that are known to physically interact as proteins was higher than for genes with no or unknown interaction (maximum p-value = <0.001 over 100 replicates, Fig 5A). This indicates that even with a relatively small number of observations, there is sufficient information in the matrix to detect biological signal between gene pairs. These results were calculated based on the correlation in expression bias between the ovary and carcass. However, following the same procedure using correlations in changes in head-biased expression showed no significant difference between the two groups (max. p-value = 0.256, Fig T in S1 Text), suggesting the strength of this signal may be dependent on the tissues being compared.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 5. Estimating pairwise correlation coefficients across genes reveals new networks of correlated expression evolution.

A-B Comparison of the distribution of Pearson’s correlation coefficients based on ovary-biased expression evolution between genes. Box plots indicate mean, upper and lower quartiles, and 1.5x interquartile ranges. Asterisks indicate a significant t-test comparison. A, Genes with no or unknown protein-protein interactions compared to those with reported interactions in FlyBase [52] (maximum p-value = <0.001 over 100 replicates). B, Correlation comparison between genes with no or unknown genetic interactions and those reported to have enhancement or suppression interactions in FlyBase (unknown vs. enhancement max. p-value = <0.001; unknown vs. suppression max. p-value = <0.001; enhancement vs. suppression p-value = 0.497). C, Each point represents a scaled change in expression bias, colored by Pearson’s correlation coefficients relative to one example gene-family, the yolk-protein genes (black points), arranged by phylogenetic branch (numbers shown in Fig 4D). Yellow = strong positive correlation, purple = strong negative correlation. D, The network of strong correlation partners (absolute correlation > 0.825) with the yolk-protein genes, colored by the direction of correlation. Stronger correlations are shown by brighter colors, and thicker, shorter lines. Nodes are annotated with the gene symbols from the D. melanogaster sequences from that homology group. E, The correlation between known protein-protein interaction partners [52] with the yolk-protein genes.

https://doi.org/10.1371/journal.pgen.1010607.g005

We also found that genes known to interact genetically have a significantly higher mean correlation than genes with no or unknown genetic interactions (unknown vs. enhancement max. p-value <0.001, unknown vs. suppression max. p-value = <0.001, Fig 5B). Comparing genes with known genetic enhancement and suppression interactions to each other showed no significant difference (p-value = 0.497). However, for genetic interactions, the range of correlation coefficients was higher in the group of no or unknown interactions (Fig 5B). This indicates that, while the average correlation of expression evolution might be higher for interaction partners, stronger positive and negative correlations exist between pairs of genes which do not interact, or for which interactions have not yet been tested. We observed the same differences in correlations after removing low-confidence outliers, though support for the significant difference between physically interacting proteins was reduced (7 of 100 iterations had a p-value >0.001, the maximum p-value was 0.007).

We tested whether the network inferred based on strong correlation of expression evolution was consistent with known interaction partners from D. melanogaster. We selected as an example the gene yolk-protein gene family, which are known to be expressed in the reproductive system, among other tissues [51] (Fig 5C). We found eight distinct homologous gene groups, comprising 12 unique D. melanogaster parent genes, that had a strong evolutionary correlation with yolk-protein genes (absolute coefficient greater than 0.825, Fig 5D). None of these correlated genes correspond to those listed on FlyBase [52] as having known interactions with yolk-protein genes in D. melanogaster (Fig 5E). We consider these strong evolutionary correlations to be a set of new predictions about evolutionary and genetic relationships between genes which can be tested in wild and laboratory model species of Drosophila. The correlation networks for each gene family, estimated using the full dataset or using the more conservative dataset with low-confidence outliers removed, can be interrogated in the accompanying data visualization for this manuscript (https://shchurch.shinyapps.io/hawaiian_fly_dataviz_2021).

4.1 Field collection

Specimens used for transcriptome sampling were caught on the Hawaiian islands between May of 2016 and May of 2017. Specimens were caught using a combination of net sweeping and fermented banana-mushroom baits in various field sites on the Hawaiian islands of Kaua’i and Hawai’i (see Table A in S1 Text for locality data). Field collections were performed under permits issued by the following: Hawai’i Department of Land and Natural Resources, Hawai’i Island Forest Reserves, Kaua’i Island Forest Reserves, Koke’e State Park, and Hawai’i Volcanoes National Park. Adult flies were maintained in the field on vials with sugar media and kept at cool temperatures. They were transported alive back to Cambridge, Massachusetts where they were maintained on standard Drosophila media at 18°C. Samples were processed for RNA extraction between 5 and 31 days after collecting them live in the field (average 10.8 days, see Table A in S1 Text). One species, Scaptomyza varia, was caught in the field before the adult stage by sampling rotting Clermontia sp. flowers (the oviposition substrate). For this species, male and female adult flies emerged in the lab, and were kept together until sampled for RNA extraction.

4.2 Species identification

Species were identified using dichotomous keys [56–60], when possible. Many keys for Hawaiian Drosophilidae are written focusing on male specific characters (e.g. sexually dimorphic features or male genitalia) [58]. Therefore, for species where females could not be unambiguously identified by morphology, we verified their identity using DNA barcoding. When males were caught from the same location, we identified males to species using dichotomous keys and matched their barcode sequences to females included in our study. We also matched barcodes from collected females to sequences previously uploaded to NCBI [61–63].

The following dichotomous keys were used to identify species: for picture-wing males and females, Magnacca and Price (2012) [56]; for antopocerus males, Hardy (1977) [57]; for Scaptomyza, Hackman (1959) [58]; for species in the mimica subgroup of MM, O’Grady and colleagues (2003) [59]; for other miscellaneous species, Hardy (1965) [60].

For DNA barcoding, DNA was extracted from one or two legs from male specimens using the Qiagen DNeasy blood and tissue extraction kit, or from the DNA of females isolated during RNA extraction (see below). We amplified and sequenced the cytochrome oxidase I (COI), II (COII) and 16S rRNA genes using the primers and protocols described in Sarikaya and colleagues (2019) [40].

For barcode matching, we aligned sequences using MAFFT, version v7.475 [64], and assembled gene trees using RAxML, version 8.2.9 [65]. Definitive matches were considered when sequences for females formed a monophyletic clade with reference males or reference sequences from NCBI; see Table B in S1 Text.

Female D. primaeva, D. macrothrix, D. sproati, and D. picticornis could be identified unambiguously using dichotomous keys. Female D. atroscutellata, D. nanella, D. mimica, D. tanythrix, S. cyrtandrae, S. varipicta, and S. varia were identified by matching barcodes to reference sequences from NCBI, reference males, or both. For the female haleakalae fly used in this study, no male flies were caught in the same location as these individuals, and no other sequences for haleakalae males on NCBI were an exact match with this species. Given its similar appearance to Drosophila dives, we are referring to it here as Drosophila cf dives, and we await further molecular and taxonomic studies of this group that will resolve its identity.

4.3 Sampling strategy

The target number of mature, healthy female flies per species was four, with three intended for dissection and species-specific expression libraries and one intended as a whole-body reference library (Fig 1). When four such individuals were not available, a reference library was assembled by combining the tissue-specific libraries from one of the other individuals. This was the case for the following species: D. sproati, which was dissected and had RNA extracted separately from the head, ovaries, and carcass, with RNA combined prior to library preparation; and S. varia, S. cyrtandrae and D. cf dives, for which RNA was extracted and libraries prepared for separate tissues, and raw reads were combined after sequencing.

For the other eight species, sufficient individual females were available such that reads for transcriptome assembly were sequenced from a separate individual. In these cases one entire female fly was dissected and photographed to assess whether vitellogenic eggs were present in the ovary, and all tissues were combined in the same tube and used for RNA extraction. Library preparation failed for one individual D. atroscutellata fly, as well as two tissue-specific libraries: one head sample from D. mimica, and one head sample from D. sproati.

4.4 Dissection and RNA sequencing

Female flies were anesthetized in 100% ethanol and were dissected in a 1x phosphate-buffered saline solution. The ovary was separated from the abdomen, and the head was separated from the carcass. Photographs of each tissue were taken, and tissues were moved to pre-frozen eppendorf tubes, kept in dry ice, and immediately transported to a -80°C freezer. Dissections were performed as quickly as possible to prevent RNA degradation. Samples were stored at -80°C for between 90 and 336 days before RNA extraction (average 281.9 days, see Table A in S1 Text).

RNA was extracted from frozen samples using the standard TRIzol protocol (http://tools.thermofisher.com/content/sfs/manuals/trizol_reagent.pdf). One mL of TRIzol was added to each frozen sample, which were then homogenized using a sterile motorized mortar. The recommended protocol was followed without modifications, using 10 μg of glycogen, and resuspending in 20μL RNAse-free water-EDTA-SDS solution. DNA for subsequent barcoding was also extracted using the phenol-chloroform phase saved from the RNA extraction.

RNA concentration was checked using a Qubit fluorometer, and integrity was assessed with a Agilent TapeStation 4200. RNA libraries were prepared following the PrepX polyA mRNA Isolation kit and the PrepX RNA-Seq for Illumina Library kit, using the 48 sample protocol on an Apollo 324 liquid handling robot in the Harvard University Bauer Core Facilities. Final library concentration and integrity were again assessed using the QUbit and TapeStation protocols.

Samples intended for transcriptome assembly were sequenced on an Illumina HiSeq 2500, using the standard version 4 protocol, at 125 base pairs of paired-end reads. Samples intended for tissue-specific expression analyses were sequenced on an Illumina NextSeq 500, using a high output flow cell, at 75 base pairs of single-end reads. A table of total read counts for each library can be found in Tables C-D in S1 Text. To account for any possible batch effects across separate rounds of sequencing, each sequencing run was performed with one or several overlapping samples. Principle component analysis of these libraries showed variation between sequencing runs to be negligible relative to variation between tissue and individual (see Results and Fig G in S1 Text).

4.5 Transcriptome assembling and expression mapping

Transcriptome assembly and expression mapping was performed using the agalma pipeline, version 2.0.0 [48]. For the twelve reference transcriptomes, reads from separate rounds of sequencing were concatenated and inserted into the agalma catalog. Further details of transcriptome assembly and homology assessment are included in our previous manuscript [44].

Each tissue-specific expression library was mapped to the corresponding reference transcriptome using the ‘expression’ pipeline in agalma, which uses the software RSEM to estimate gene and isoform count levels from RNAseq data [66]. The agalma pipeline also includes steps to catalog the species, tissue type, and run information, which were exported as a single JavaScript object notation (JSON) file. This file is available in the GitHub repository in the directory analysis/data.

4.6 Phylogenetic analysis

The methods for inferring a phylogenetic tree of the species here are the same as those described in our previous manuscript [44]. In summary, we used the agalma pipeline [48] to infer homology, based on an all-by-all BLAST procedure. Gene orthology was inferred based on the topology of gene trees, estimated with RAxML [65], version 8.2.9. The most likely tree was inferred using IQtree [67], version 2.1.1, on a dataset partitioned by transcripts and using the default Model Finder [68]. For this study, gene trees were additionally annotated with the software phyldog [69].

4.7 Annotating transcripts by sequence similarity

We leveraged the close relationship of these species to species of Drosophila with well-annotated genomes to annotate the transcripts considered here. For each transcript in the reference transcriptome, we performed four comparisons of sequence similarity using local BLAST: [1] comparing nucleotide transcript sequences to nucleotide sequences from D. melanogaster (blastn), [2] comparing translated nucleotide sequences to protein sequences of D. melanoagster (blastx), [3] comparing nucleotide sequences to a database of nucleotide sequences from D. melanogaster, D. virilis, and D. grimshawi (blastx), and [4] comparing translated nucleotide sequences to a database of protein sequences from the same three species (blastn). For downstream analyses, we prioritized annotations from the second comparison, but we provide all sequence similarity reports in the GitHub repository under the directory analysis/BLAST.

To annotate homology groups as defined by the homology inference step of agalma, we extracted the name and sequence ID from all D. melanogaster sequences in the group.

4.8 Normalization and differential gene expression

Transcript count tables were imported into R using the agalmar package, version 0.0.0.9000. Differential gene expression analysis was performed using the package DESeq2, version 1.34.0. For these analyses we used only one sequencing run per library, thereby excluding duplicate sequencing runs. Analyses of differential gene expression were calculated using the default approaches in DESeq2 for estimating size factors, dispersions, and calculating log₂ fold-change and adjusted p-values using the Benjamini and Hochberg correction (Fig B, panel a, in S1 Text). Both individual and tissue were considered in the design formula. Transcripts were considered differentially expressed at a significance threshold of 0.01.

We identified a cohort of core ovary-specific genes by first identifying a parent gene for each transcript using a sequence similarity search against D. melanogaster (Fig B, panel a, in S1 Text). We then identified parent genes that had at least one transcript significantly differentially upregulated in the ovary of more than ten of the twelve species. Because multiple transcripts may match to a single parent-gene, core ovary-specific parent genes may include transcripts that are also not differentially upregulated in the ovary, as long as at least one transcript is upregulated for more than ten out of twelve species. This may be the case when transcripts are artificially fragmented during reference transcriptome assembly, or when sequence-similar transcripts have biologically distinct expression levels.

4.9 Comparison of expression to D. melanogaster

We compared our differential gene expression results to a reference database of tissue expression from D. melanogaster, known as the FlyAtlas2 [29]. We downloaded this reference in July of 2021, from http://motif.gla.ac.uk/downloads/FlyAtlas2_21.04.18.sql. This dataset provides data on transcript abundance and tissue enrichment, including for female ovaries. Tissue enrichment is calculated using the same methods as in the FlyAtlas2 web browser, defined as the fragments per kilobase of transcript per million mapped reads (FPKM) for a given tissue divided by that value for the reference tissue (here, female whole body), with a pseudocount of two counts added to empty values to avoid division by zero. We considered a FlyAtlas gene to be enriched in the ovary, comparable to our data, if the ovary was the maximum enrichment value across all tissues excluding the head, brain, and eye tissues, as these were separated in our RNASeq procedure (Fig B, panel a, in S1 Text). We considered a FlyAtlas gene to be head enriched if either the head, brain, or eye were the maximum enrichment value, excluding the ovary.

4.10 Transforming data into comparable measurements of expression across species

Transcript counts are reported in transcripts per million (TPM), but this measurement is known to not be directly comparable across species due to differences in reference transcriptome size [7, 8]. Therefore, we normalized TPM by species using the procedure described by Munro and colleagues (2021) [8], where TPM values are multiplied by the number of genes in the reference, and this value is divided by 10⁴ (Fig B, panel b, in S1 Text). TPM10k values were natural-log transformed.

An additional challenge when working with reference transcriptomes is the presence of fragmented transcripts created during the assembly process [49]. This fragmentation can result in noise in estimating the amount of transcript as reads are differentially mapped to these fragments. To reduce the impact of this noise on our analysis, we undertook a novel approach where transcripts were grouped according to inferred homology as estimated by the agalma pipeline using an all-by-all BLAST approach (Fig B, panel b, in S1 Text). For each sequenced library, we found the average count value across all transcripts from the same homology group (see Table E in S1 Text for statistics on homology group composition). For each species-tissue pair, we then averaged this value across all biological replicates, here replicate individuals.

4.11 Linear modeling

We performed linear modeling to calculate the relative contribution of tissue- and species-level differences to variation in gene expression (Fig B, panel b, in S1 Text), following the approach of Breschi and colleagues (2016) [21]. These analyses were performed separately on datasets of ovary vs. carcass and head vs. carcass expression. Using the ANOVA script provided at https://github.com/abreschi/Rscripts/blob/master/anova.R, we built a linear model for each gene that accounts for the contribution of the organ, species, and any residual error. We then calculated the relative proportion of each factor divided by the total sum of squares for all factors. We identified groups of highly variable genes, using the same metrics defined by Breschi and colleagues (2016) [21], as any gene for which either tissues or species explains at least 75% of the variance. Species variable genes (SVGs) were defined as highly variable genes whose relative variation was two-fold greater across species than tissues (vice-versa for tissue variable genes, TVGs).

We performed these linear model analyses over four nested clades: a clade containing two picture-wing species (D. sproati and D. macrothrix); a clade containing the four picture-wing-Nudidrosophila-Ateledrosophila species in this study; a clade containing the nine Hawaiian Drosophila species in this study; and a clade of all 12 Hawaiian Drosophila and Scaptomyza species in this study. We repeated these analyses excluding the species S. varia, which showed the lowest similarity in expression to the other eleven species. To compare our analysis to the more typical approach, we also performed these analyses on all pairwise combinations of these twelve species.

4.12 Reconstructing evolutionary history of differential expression

We calculated tissue bias as the ratio of counts in TPM10k for each tissue (ovary and head) to the reference tissue [7], here the carcass (Fig C, panel a, in S1 Text). Using ratios is advantageous because any potential variation in transcript counting efficiency across individuals is canceled out, as this will be present in both the numerator and denominator [7]. We subsequently performed the same transformation steps described above, averaging over ratios from the same homology group and across biological replicates, to calculate average expression bias per homology group per library. To avoid division by zero, we added a pseudocount of 0.01 to each TPM10k value. Ratio values were natural-log transformed so that positive values indicate enrichment in the tissue of interest relative to the reference tissue, negative values indicate the opposite, and values of zero indicate equivalent expression.

We reconstructed the evolutionary history of tissue bias for each homology group using the species tree published in Church and Extavour, 2021 [44], based on the same reference transcriptome data (Fig B, panel c, in S1 Text). First, we calibrated the tree estimated using IQtree (Fig 1A of that publication) to be ultrametric using the R function chronos in the package ape, version 5.6.2 (using a correlated model and a lambda value of 1). We then subset this tree to only include tips for which expression data was available, and annotated this tree to be able to identify specific branches and nodes in ancestral state reconstruction analyses.

Ancestral expression bias values were estimated with the R package Rphylopars, version 0.3.9, using the fast ancestral state reconstruction algorithm based on Ho and Ané, 2014 [70] (Fig C, panel a, in S1 Text). Tips for which expression data were not available were dropped from each reconstruction, and ancestral state reconstruction was only performed when more than three tips had data. Following ancestral state reconstruction, we calculated the scaled change as the difference between the value at the ancestral and descendant nodes, divided by the length of the branch. Scaled changes were compared between homology groups by identifying equivalent branches as those that share the same parent and child node, following the procedure described in Munro and colleagues (2021) [8]. We identified qualitative changes in expression bias as changes that resulted in a ratio changing from negative to positive values or vice versa.

4.13 Estimating correlated evolution of expression across genes

For each homology group that had representation across all twelve species, we calculated pairwise Pearson’s correlation coefficients by comparing scaled changes in expression bias across equivalent branches (Fig C, panel b, in S1 Text). For the twelve-species phylogeny, this meant each correlation coefficient was calculated using 22 individual data points (branches). This resulted in a correlation matrix of 1,306,449 pairwise comparisons of evolutionary correlation.

We compared this correlation network to data on protein interactions and genetic interactions downloaded from http://flybase.org in July, 2021. These data include pairwise observations of genetic enhancement and suppression interactions between parent genes in D. melanogaster. These interactions were matched to pairwise correlation coefficients by identifying the corresponding homology group for each D. melanogaster parent gene ID (more than one parent gene may fall into the same homology group).

We tested whether correlation coefficients for known physical and genetic interaction partners were higher than in genes with unknown interactions using two-sample t-tests. The sample size for physical interaction partners was 1,953, for genetic enhancement was 280, and for genetic supression was 497. In each test we compared the coefficients for either enhancement or suppression interactions to a random sample of 5000 coefficients for which interactions are unknown. We repeated these t-tests 100 times using different random samples, and report the maximum p-value observed. We also compared the distribution of enhancement and suppression interaction coefficients to each other using a single t-test.

Strong correlations for the visualization of co-evolutionary networks were selected using a threshold correlation coefficient of 0.825. Because spurious correlations may be driven by the presence of low-confidence outliers [50], we repeated all correlation analyses after removing any transcript with an estimated change in expression bias beyond two standard deviations from the mean change across all transcripts.