2.3.1 Embryo Collection and Imaging

We anaesthetized females with embryos in a solution containing 50 mL of synthetic freshwater (Protas et al. 2011) and 20 µL of clove oil. Then, each female was washed twice in synthetic freshwater, and the embryos were removed by opening the brood pouch gently with forceps and flushing out the embryos with a glass pipette. The embryos were moved into a new Petri dish with synthetic freshwater and kept at 12°C. Embryos were monitored daily and photographed using a Leica S8 Apo Microscope. Embryos that were used for RNAseq were harvested at either 70% through embryonic development, in which the embryos have a comma shape and will shortly show eye pigmentation in the surface form, or 90% of the way through embryonic development, when eye and head pigmentation is visible and the body of the embryo is elongated (Mojaddidi et al. 2018).

We collected embryos from gravid CAVE_mw females upon arrival in the lab or females that mated immediately after arrival to the lab and then raised the embryos in vitro as described above. SURF_li females mated during transport to the lab or immediately after arrival at the lab. We then removed the embryos and raised them in vitro. We called these sets of embryos “wild-caught” as they were from wild-caught females that either mated in the wild or immediately upon arrival to the lab. Our initial comparisons of cave and surface juveniles and embryonic development used these wild-caught embryos because we did not know whether the cave animals would breed in the laboratory environment. Fortunately, we were able to raise juveniles to adulthood in the lab and generate gravid females from these lab-reared juveniles for both CAVE_mw and SURF_li populations. We labeled juveniles from these lab-reared embryos as “lab-bred.” F1 hybrid embryos, generated as detailed above, could not be classified as either wild-caught or lab-bred as they were generated from wild-caught cave males and either wild-caught surface females or lab-bred surface females around a month after the animals were brought into the lab from the wild. We examined pigmentation as a presence versus absence trait in all embryos. Ideally, we would also have examined the presence versus absence of eyes (ommatidia) in embryonic development, but this requires visualization methods that often result in embryo death, and our sample sizes were too limited to successfully view enough individuals.

2.3.2 Comparative Juvenile Analysis

For the comparative juvenile analysis, our goal was to harvest juveniles when they were first able to walk and had the appearance of miniature adults. We selected this timepoint as it is easier to flatten individuals for measurements when their appendages can be easily removed. Embryos were observed every day, and once they were able to walk, they were harvested in 100% ethanol and placed at −20°C until mounting on microscopic slides. Juveniles that were selected for mounting had at least one intact antenna II. As individuals were raised in a dish and were harvested right when they were able to walk, antennae were generally intact. To ensure that juveniles could be flattened in ethanol, all pereopods were removed under a Leica S8 Apo stereomicroscope. Some individuals also had to have their mouth parts removed to allow further flattening. After limb removal, the juvenile was washed in 50% ethanol, 100% Milli-Q water, and 50% glycerol, and then placed on a microscope slide with the dorsal side up in 100% glycerol. On the left- and right-hand sides of the juvenile, a coverslip was placed with a single drop of glycerol to create a shallow wall around the juvenile so that when a coverslip was placed over the juvenile, the third coverslip rested on top of the first two coverslips and did not crush the juvenile. Juveniles were measured using LAS Core Software (Leica, Wetzlar, Germany) on a Leica S8 Apo stereomicroscope. Antenna II length (left and right if available), head width, and body length were measured. Five groups of samples were measured CAVE_mw_wild samples, SURF_li_wild samples, CAVE_mw_lab samples, SURF_li_lab samples, and SURF_li_CAVE_mw_F1 hybrid samples (Supporting Information File 1 and Table S2). Each juvenile was photographed and measured separately by two different people using LAS Core Software. The following measures were analyzed: left and right antenna II length, body length, and head width. The measurements from the two different people were averaged. Again, pigmentation was examined as a presence versus absence trait in all embryos.

2.3.3 Statistical Analysis

First, we performed the Shapiro–Wilk test in R 4.1.3 (R Core Team 2022) to assess normality of measurements. Measures which did not deviate significantly from normality included antenna II lengths (left and right), body length, and head width. For these measures, ANOVA and Tukey's Honestly Significant Difference (HSD) tests for pairwise comparison were performed for each group. To address questions of allometry, we also performed principal components analysis on the measurements of body length, head width, and left and right antennae length, followed by an ANOVA and Tukey's HSD test on the resulting first and second principal component scores across groups. The Bonferroni correction for multiple testing was performed for tests of comparative juvenile analysis.

2.4.1 Samples

One brood of CAVE_kl embryos and one brood of SURF_tb embryos were prepared for Iso-seq to obtain full-length transcripts with the goal of combining this sequence ultimately with Illumina sequence to obtain more complete transcriptomes (Table 2). RNA extraction was performed as described (Lomheim et al. 2023), and the sequencing was performed on a PacBio machine by the Functional Genomics Lab, Vincent C. Coates Genomics Sequencing Laboratory, and UC Berkeley. For Illumina sequencing, samples were collected more recently, and individuals at the SURF_tb and CAVE_kl locations were not present, so nearby populations were used instead, CAVE_mw and SURF_li. Three CAVE_mw broods and three SURF_li broods at two different developmental timepoints, that is, mid-stage (ca. 70%) through embryonic development and late stage (ca. 90%) through embryonic development, were harvested. Additionally, three broods of late-stage F1 hybrids (SURF_li_CAVE_mw_F1) were harvested (Table S3). Each of these samples consisted of a single brood of siblings. After harvesting, all broods were homogenized in 200 µL of TRIzol (Thermofisher, Waltham, Massachusetts, the United States). Sequencing libraries were prepared for each of these samples: three CAVE_mw mid-stage, three CAVE_mw late-stage, three SURF_li mid-stage, three SURF_li late-stage, and three late-stage SURF_li_CAVE_mw_F1. RNA was extracted using the RNeasy Plus Universal Mini Kit by the Genetic Epidemiology and Genomics Lab (GEGL), UC Berkeley. PolyA selection was performed, and libraries were constructed using the Nugen Kit low-input protocol. Sequencing was performed on a HiSeq. 4000 machine with 150 bp paired-end reads by the Functional Genomics Lab, Vincent C. Coates Genomics Sequencing Laboratory, and UC Berkeley.

2.4.2 De novo transcriptome assembly and annotation

Two transcriptomes were made, one SURF_li transcriptome and one CAVE_mw transcriptome. First, Iso-seq sequences were run through the in-house Iso-seq pipeline of the Functional Genomics Lab, Vincent C. Coates Genomics Sequencing Laboratory, and UC Berkeley, which included the synthesis of CCS sequences, demultiplexing of sequences, primer polyA and concatamer removal, clustering, and polishing. The final transcripts were combined with the assemblies generated from Illumina reads. We used the same methods as in a previous study by Lomheim et al. (2023). Briefly, for the Illumina samples, the reads were trimmed with Trimmomatic using the following parameters: sliding window 4:24, headcrop 10, avgqual 30, and minlen 30. Then, the transcriptomes were generated using the de-novo-transcriptome-assembly pipeline of the National Center for Genome Analysis Support (NCGAS) (https://github.com/NCGAS/de-novo-transcriptome-assembly-pipeline), which incorporates multiple assemblers: SOAP version 1.03 (kmer 35, 45, 55, 65, 75, and 85; Xie et al. 2014), TransAbyss version 2.0.1 (kmer 35, 45, 55, 65, 75, and 85; Robertson et al. 2010), Trinity 2.11.0 using default parameters (Grabherr et al. 2011), and Velvet 1.2.10 (kmer 35, 45, 55, 65, 75, and 85; Zerbino and Birney 2008). All the assemblies plus the Iso-seq sequences were run through Evidential Gene's tr2aacds pipeline (Gilbert 2013). Recently, a genome was published for an individual of A. aquaticus from Wytham Woods, Great Britian (Thomas and University of Oxford and Wytham Woods Genome Acquisition Lab, Natural History Museum Genome Acquisition Lab, Darwin Tree of Life Barcoding collective, Wellcome Sanger Institute Tree of Life Management, Samples and Laboratory team, Wellcome Sanger Institute Scientific Operations: Sequencing Operations, Wellcome Sanger Institute Tree of Life Core Informatics team, Tree of Life Core Informatics collective and Darwin Tree of Life Consortium 2025). We mapped the de novo transcriptomes to the published genome in quality control (minimap2, Li 2018), but no annotation was published with this genome. As a result, once transcriptomes were generated, we identified homologous sequences to the Tribolium castaneum annotation from 2019 (file entitled GCF_000002335.3_Tcas5.2_protein.faa) https://ftp.ncbi.nlm.nih.gov/genomes/all/GCF/000/002/335/GCF_000002335.3_Tcas5.2/ using Blast2GO (Conesa et al. 2005), allowing us to transfer annotation information from Tribolium to our de novo transcriptomes. Transcriptomes were evaluated using Galaxy Version 2.10.1+galaxy2 with BUSCO version 5.3.2 (Arthropoda selected; Simão et al. 2015) and QUAST version 5.2.0 (eukaryote selected; Gurevich et al. 2013). XSEDE resources were used (Towns et al. 2014).

2.4.3 Differential Expression of Mid- and Late-Stage Embryos

Differentially expressed genes between surface and cave samples at the mid-stage and late-stage timepoints were examined. Differential expression was assayed as previously discussed (Lomheim et al. 2023). Briefly, trimmed reads were quantified against both transcriptomes with the default parameters from Kallisto (Bray et al. 2016). The Kallisto output of estimated counts was combined into four matrices: all mid-stage embryo samples mapped against CAVE_mw, all mid-stage embryos mapped against SURF_li, all late-stage samples mapped against CAVE_mw, and all late-stage samples mapped against SURF_li. Differential expression analysis was performed using DEseq2 (Love et al. 2014). DESeq2's shrinkage estimation and dispersion modeling provide more stable results for analyses with fewer biological replicates by using empirical Bayes estimators. These estimators pool information across genes to gain extra information and provide more accurate estimations of dispersion in cases where there might be artificially high variance due to sample size. DESeq2 also models the likelihood of data capture using a negative binomial distribution model, which also helps combat overdispersion in smaller sample sizes and heterogeneous data such as field samples of non-model species.

Pairwise DESeq2 analysis was conducted on the four matrices described above using Trinity's differential expression perl wrappers (run_DE_analysis.pl and analyze_diff_expr.pl, Haas 2023). This wrapper filters out transcripts with 0 CPM in any replicate, as this is ambiguously missing data, which could result from a lack of capture (technical bias) or a lack of expression (real variation). Normalization was also handled in the DESeq() function, using DESeq2's default mean of ratios method to normalize the count data and adjust for library size effects and compositional biases. Wald tests were used to evaluate the significance of the difference between log-fold changes. Multiple test correction was done with False Discovery Rate via DESeq2's workflow. Signal and noise in the normalized, filtered data were then visualized in a principal component analysis (PCA) of count data (Figure S1). The first two PCs explained 67.38% of the variation in the data and showed reasonably cohesive grouping of samples (in the context of expected noise due to variation in field samples) with distinct distances between the replicates (signal of different expression). Finally, we classified genes as significantly differentially expressed if they had a log2Fold change of > 2 (minimum 4× change in expression) and FDR < 0.05. We chose a moderate log fold change cutoff to guard against false positives and an FDR of 0.05 to account for our more limited statistical power. These cutoffs were further evaluated against the volcano plot to ensure we removed obviously spurious results, such as those with high fold change but no statistical support.

To unify the information from the four pairwise comparisons (mid-stage to both transcriptomes and late-stage to both transcriptomes), we needed to identify homologous loci in cave and surface transcriptomes to compare expression between the cave and surface samples. Each transcriptome had unique names as they were de novo assembled separately; there is no published annotation for this genome, and the gene name annotation transferred from Tribolium was not unique (i.e., multiple genes in our transcriptomes could have the same Tribolium ID). To establish a set of homologous sequences between the cave and surface (cave/surface pairs), a reciprocal blast technique was used. CAVE_mw and SURF_li databases were created using the web-based platform European Galaxy NCBI BLAST+ (Galaxy Version 2.10.1+galaxy2; Camacho et al. 2009; Cock et al. 2015). The CAVE_mw transcriptome was then blasted against the SURF_li database, and vice versa, and we retained the longest and highest hit for each transcript in each blast run. As these are the same species, it is expected that there will be high homology between assembled transcripts. However, due to limitations of de novo assembly and sequencing a subset of transcripts in a sample, not all transcripts will have a complete assembly or a homologous pair retained. As such, transcripts under 400 bp and lacking a reciprocal hit were removed. A database of 15,516 cave/surface pairs resulted, with entries including the CAVE_mw transcript name and sequence, the differential expression results for that sequence (ref: CAVE_mw), the SURF_li transcript name and sequence, and the differential expression results for that sequence (ref: SURF_li).

Further filtering of possible differentially expressed genes was applied. Transcripts that had unique Tribolium IDs were retained as those were clearly unique sequences with clear homology to a related species. Conversely, we eliminated any transcripts that shared a Tribolium annotation with 9 or more other assembled transcripts in that transcriptome, rationalizing that a large number of hits to the same Tribolium gene were more likely to be fragments or chimeras rather than large numbers of unique paralogues. To further filter out any duplicates, we self-blasted the transcripts against their own database and removed any expressed duplicates (identity > 90%, length > 150 bp, and CPM > 1). This setup allowed for the use of VLOOKUP in Excel to identify which high-confidence transcripts showed an LFC > 2 and FDR < 0.05 against both reference transcriptomes.

To perform GO enrichment, we compared this filtered set of differentially expressed transcripts against the background of the full list of genes output from DESeq2, minus any loci that shared the same Tribolium ID with 9 or more transcripts (low confidence assemblies). This was done separately for the output of mid-stage comparisons and late-stage comparisons, generating a mid-stage reference list and a late-stage reference list. Then, we used G-profiler to analyze GO enrichment at each timepoint (Raudvere et al. 2019).

2.4.4 Allele-Specific Expression Analysis

Allele-specific expression was performed as previously by Lomheim et al. (2023). First, cave/surface transcript pairs were generated, as described above, yielding 15,516 genes with cave/surface transcript pairs. The cave and surface sequences were combined (combined transcripts) using merge_pat_mat_fasta.pl script from ASE-Tigar (Nariai et al. 2016). This does not combine the actual sequences but retains two sequences for each gene, and the reads are mapped to each of the sequences.

The trimmed paired-end reads of all SURF_li, CAVE_mw, and SURF_li_CAVE_mw_F1 samples were mapped to the combined transcripts using bowtie2, -X 1000 -k 100 –very-sensitive, and ASE-TIGAR (Langmead and Salzberg 2012). Z values from ASE-Tigar are a measure of transcript abundance. The ratio of Z mapped to each surface transcript to the corresponding cave transcript was compared for each sample. Genes were selected where the surface samples mapped to the surface transcript 3.3× more than the corresponding cave transcript and where the cave samples mapped to the cave transcript 3.3× more than the corresponding surface transcript. Within this subset of genes, we selected genes where all three F1 hybrids showed a bias towards one allele, mapping to either the cave transcript 3.3× more than the surface transcript, or vice versa. Transcripts that were not annotated or where the Tribolium ID was present in 10 or more copies in the transcriptome were removed.

Next, to identify fixed SNPs that were different between cave and surface transcripts, all trimmed, paired-end reads of CAVE_mw, SURF_li, and SURF_li_CAVE_mw_F1 samples were mapped to combined transcripts using bowtie2 (default settings plus very fast end-to-end) and Freebayes (Garrison and Marth 2012). Mapping SURF_li reads to the CAVE_mw transcripts and reciprocally mapping CAVE_mw reads to the SURF_li transcripts identified fixed SNPs between the cave and surface transcripts. All fixed SNPs were retained that were present with at least five observations.

For all the above genes prioritized by ASE-Tigar, genes were further prioritized that also showed differential expression. For this subset of genes, we performed FreeBayes allele counting in the F1 hybrids using the fixed SNPs outlined above (Garrison and Marth 2012). To eliminate bias from the reference used, allele counting was performed both when the cave and the surface transcripts were used as a reference. Five SNPs, if available, were selected across the span of a transcript for allele counting in the F1 hybrids. Most of the time, the location of these SNPs was the same in the cave versus the surface transcript. However, if an indel was present, the location of the SNPs might be offset in the two transcripts, and alignments were performed if it was unclear where the location of the SNPs was.

For each SNP, a binomial test was performed to investigate when the allele counts deviated from the predicted fraction of 0.5. To control for multiple testing, the resulting p values were adjusted using the Benjamini–Yekutieli procedure (Benjamini and Yekutieli 2001). Any gene where at least two of the five SNPs showed adjusted p values < 0.05 for all three F1 samples to both the cave and surface transcript was identified as having allele-specific expression.

3.1.1 Asellus infernus Embryos Were Depigmented Throughout Embryonic Development

We first asked when morphological differences were established in A. infernus as compared to the surface form, embryonically or postembryonically. A. infernus adults collected from wells in Mangalia (cave) were depigmented and eyeless as compared to A. aquaticus adults collected from the Limanu Bridge Stream population (surface) (Figure 1). We compared embryonic development of wild-caught gravid cave females and wild-caught surface females, which became gravid upon transport to the lab or immediately after arriving at the lab. Regarding pigmentation, A. infernus embryos never developed pigmentation, whereas surface embryos developed eye pigmentation at around 80% of the way through embryonic development (Figures 2, S2, and S3) and shortly after, developed head and body pigmentation. Additionally, F1 hybrid embryos from A. infernus males crossed to surface females did develop eye pigmentation, similar to the surface form (Figure 2). Furthermore, adult F1s had both head and body pigmentation and ommatidia (Figure 1D–F). Therefore, we established that pigmentation was never seen embryonically in A. infernus, rather than initially being present and then lost later in development.

3.1.2 Lab-Bred Asellus infernus Juveniles Had Longer Body Length and Longer Antennae II Than Lab-Bred Surface Asellus aquaticus Juveniles

To further address the question of the timing of the establishment of morphological characteristics, we investigated size characteristics between lab-bred A. infernus and surface individuals. For this comparison, we examined lab-bred A. infernus and surface individuals that were raised in the same conditions to eliminate, as much as possible, any environmental variables. Specifically, we compared antenna II length, body length, and head width in lab-bred surface and cave juveniles as soon as they were able to walk and would have been released from the brood pouch (Figure 3 and Table S4). For body length, antenna II length, and head width, cave samples were longer compared to surface samples (p < 0.0005 for all comparisons; Figure 3 and Table S4). Overall, we conclude that lab-bred cave and surface juveniles are different in body size, head width, and antenna size, and that these differences, many of which are also present in comparisons between adult cave and surface individuals, are established by the end of embryonic development.

3.1.3 Population-Environment Interactions (Wild-Caught Vs. Lab-Bred) Influenced Several Measurements' Differences Within Both Cave and Surface Samples

To investigate whether environmental changes also influence these morphological differences, we compared wild-caught and lab-bred juveniles. For the surface wild-caught to surface lab-bred comparisons, two measures showed significant differences: head width and body length (p < 0.0005 for both; Figure 3 and Table S4). For the cave wild-caught to cave lab-bred comparisons, only body length showed a significant difference (p = 0.00073). To ensure that differences in these measurements were not simply reflections of distinctions in overall size, we performed PCAs on the measurements of body length, head width, and left and right antennae length (Figure 4 and Table S5). The first two principal components together explained approximately 95% of the total variation, so further components were not considered. The first principal component (PC1) corresponded roughly to an average of all four body measurements and therefore can be interpreted as a general measure of overall size: large positive PC1 scores are associated simultaneously with longer body lengths, antennae, and wider heads, while more negative PC1 scores are associated with smaller measurements overall. The second principal component (PC2) corresponded to a contrast between antennae length and head and body measurements: highly positive PC2 scores are associated with longer antennae but shorter body lengths overall, and more negative PC2 scores were associated with longer bodies but shorter antennae. Figure 4 displays schematic archetypes of isopods with extreme values in each of these two directions of variation.

We performed ANOVA using juvenile's specific combination of population and environment as a predictor for each of the two principal component scores, which are uncorrelated by construction, ensuring that the differences in contrasts between antennae relative to body measurements (i.e., PC2) are not simply due to differences in size (i.e., PC1). Pairwise comparisons across subgroups were assessed using Tukey's HSD test, and p values were reported after adjusting for multiple testing (Table 1).

Significant differences in overall size (PC1) were detected for all comparisons except between cave lab-bred versus cave wild-caught. A detailed look at all comparisons together suggests that cave specimens were larger overall than surface specimens. However, size differences between lab-bred and wild-caught subgroups were not so straightforward and depended on whether the isopod was a surface or cave-dwelling specimen. For surface individuals, lab-bred subjects were generally larger than their wild-caught counterparts across the board (p = 0.00003), while for cave individuals, wild-caught specimens were on average larger than lab-bred individuals, though this difference was not significant (p = 0.4604).

Significant differences in contrasts between antennae and body length (PC2) were also found in most comparisons, except for those across cave wild versus surface lab-bred subgroups and cave lab-bred versus surface wild subgroups. No marginal cave versus surface trend emerged in this direction; rather, differences in PC2 scores were dependent on the interaction between surface/cave environments and lab-bred/wild breeding. The cave lab-bred and surface wild-caught groups had positive PC2 scores (i.e., longer antennae and shorter body lengths, while controlling for overall size), while cave wild-caught and surface lab-bred groups generally had negative PC2 scores (i.e., shorter antennae and longer body lengths, while controlling for overall size). See Figure 4 for visualization of group morphologies and Table 1 for specific pairwise comparisons and corresponding p values.

3.1.4 F1 Hybrid Samples Were Intermediate and Different in Most Size Measures Between Cave and Surface Individuals

Because cave and surface forms can interbreed, we had the opportunity to compare the above size characteristics in F1 hybrids with the goal of seeing whether the F1 hybrids were similar to A. infernus, to surface, or intermediate between the two. Our expectation was that the F1 hybrids would be intermediate between the two measures. The F1 samples were significantly different from both cave lab-bred and wild-caught samples for the measures of head width, body length, and antenna length (p < 0.0005 for all measures; Figure 3 and Table S4). The F1 individuals were different for both surface lab-bred and wild-caught samples for only antenna length (p = 0.0153 for surface lab and p < 0.0005 for surface wild; Figure 3 and Table S4).

3.2.1 Different mRNA Levels Were Seen in Genes Involved in Phototransduction and Sulfide Detoxification in Cave and Surface Samples

Next, we investigated the question of what genes showed expression differences between A. infernus and surface samples. First, we generated transcriptomes for the A. infernus cave and A. aquaticus Limanu springs surface populations (Table 2). With BUSCO scores of 93%, this is similar to updated transcriptomes for the Subterranean Rak population and Rakov Škocjan of A. aquaticus (Lomheim et al. 2023) and transcriptomes generated from the Molnár János Cave population and its adjacent surface populations, Malom Lake and Soroksár from Hungary (Pérez-Moreno et al. 2018). Cave and surface transcriptomes had a similar number of transcripts from the previously examined populations (Table 2).

Recently, an unannotated genome for A. aquaticus became available (Thomas and University of Oxford and Wytham Woods Genome Acquisition Lab, Natural History Museum Genome Acquisition Lab, Darwin Tree of Life Barcoding collective, Wellcome Sanger Institute Tree of Life Management, Samples and Laboratory team, Wellcome Sanger Institute Scientific Operations: Sequencing Operations, Wellcome Sanger Institute Tree of Life Core Informatics team, Tree of Life Core Informatics collective and Darwin Tree of Life Consortium 2025), and we mapped both the cave and surface transcriptomes to the genome. For the cave transcriptome, all 88,614 transcripts had a primary alignment. Of these, 55,950 transcripts had high-quality, non-chimeric primary alignments, and another 2117 had supplementary alignments across transcripts. For the surface transcriptome, all 80,182 transcripts had a primary alignment. 59,304 transcripts had high-quality, non-chimeric primary alignments, and another 1537 had supplementary alignments across transcripts. In all figures, tables, and files, an asterisk is placed next to any transcript that did not have a primary alignment.

We performed differential expression analysis on wild-caught embryos at two different embryonic timepoints, mid-stage and late-stage, identifying genes that were surface-biased (higher mRNA levels in surface samples as compared to cave samples) and cave-biased (higher mRNA levels in cave samples as compared to surface samples) (Supporting Information Files 2, 3, 4, and 5 and Figures S4 and S5). We found 166 genes that were surface-biased and 73 genes that were cave-biased at the mid-stage timepoint. At the late-stage timepoint, we found 160 genes that were surface-biased and 73 that were cave-biased. We had expected that genes involved in phototransduction would show lower mRNA levels in the cave individuals as compared to the surface samples and found that this was the case for six genes at the mid-stage timepoint and seven genes at the late-stage timepoint; some of the genes were in common between the two timepoints (Table 3). Additionally, several genes involved in melanin synthesis and the diurnal clock showed significantly different mRNA levels in cave versus surface samples.

We had also predicted that some of the genes that were differentially expressed would be in response to the sulfidic environment, such as genes involved in H₂S oxidation. As expected, we saw cave-biased differential expression in sulfide:quinone oxidoreductase and in persulfide dioxygenase ETHE1 at both the mid-stage and late-stage timepoints (Supporting Information Files 3 and 5).

GO enrichment analysis of differentially expressed genes did not highlight any categories involved with eyes, pigmentation, or sulfide metabolism. However, in the mid-stage embryos, the KEGG term of “arachidonic acid metabolism” was enriched in the genes that were surface biased (KEGG:00590, p = 0.042). In the late-stage embryos, the GO term “structural constituent of cuticle” was enriched in cave-biased genes (GO:0042302, p < 0.001). The lack of enrichment data is likely due to the comparatively small number of genes identified as showing statistically significant mRNA levels and the usage of whole embryos rather than tissue-specific RNAseq.

3.2.2 Genes With Cave-Biased and Surface-Biased Allele-Specific Expression Were Identified Including Genes With Possible Roles in Adaptation to Sulfidic Environments

To continue examining the question of what genes showed expression differences between cave and surface samples, allele-specific expression was used. Allele-specific expression is a method of identifying genes that might have cis-regulatory changes. First, we used ASE-Tigar to prioritize genes with allele-specific expression (Supporting Information Files 6 and 7). In our final list of genes with allele-specific expression using FreeBayes allele counting and eliminating genes that did not also show differential expression, we identified 5 genes that were cave-biased and 11 genes that were surface-biased (Figure 5 and Supporting Information File 8).