The Convergent Evolution of Hummingbird Pollination Results in Repeated Floral Scent Loss Through Gene Downregulation

Abstract

The repeated evolution of the same trait in different lineages provides powerful natural experiments to study the phenotypic and genotypic predictability of trait gain and loss. A fascinating example is the repeated evolution of hummingbird pollination in plant lineages in the Americas, a widespread and often unidirectional phenomenon. The spiral gingers in the genus Costus are ancestrally bee pollinated, and hummingbird pollination has evolved multiple times independently in the tropical Americas. These pollinator transitions are accompanied by predictable morphological and color changes, but the changes in floral scent have not been described. In this study, we describe the floral scent composition of 30 species of Costus sampled across the phylogeny to understand how floral scent has evolved across the genus with respect to pollinator transitions. We then combine transcriptomics and genomics to identify gene expression differences and gene family evolution associated with pollinator transitions. We show that hummingbird-pollinated species have mostly lost their floral scent, whereas bee-pollinated species exhibit either floral scent maintenance or, in some cases, gains of more diverse scent profiles. We find the floral scent loss appears to be due to gene downregulation rather than pseudogenization. The remarkable consistency of scent loss in hummingbird-pollinated species highlights the shared strong selection pressures experienced by these lineages. Even species with more recent transitions from bee to hummingbird pollination exhibit scent loss, highlighting the rapid breakdown of scent emission following pollinator transitions. This research highlights the capacity for rapid changes when selection pressures are strong through downregulation of floral scent genes.

Introduction

The repeated evolution of the same traits in different lineages is thought to reflect shared selection pressures. These cases of repeated evolution apply not only to the gain of traits, such as the evolution of antifreeze proteins in fish (Chen et al. 1997) or carnivory in plants (Albert et al. 1992), but also the loss of traits, as seen in the eye regression of cavefish (Sifuentes-Romero et al. 2023), or reduction of attractive traits such as flower size or floral scent in self-fertilizing plants (Tsuchimatsu and Fujii 2022). Repeated evolution provides us with systems where we can study the process of adaptation and ask questions about the predictability of evolution and the extent of constraint, or flexibility, of the underlying molecular mechanisms (Orr 2005; Losos 2011; Sackton and Clark 2019).

One such example of repeated evolution is pollinator transitions in flowering plants. For example, the evolution of hummingbird pollination has occurred many times across flowering plants and often involves convergence in a suite of traits. Flowers pollinated by hummingbirds are characterized by traits thought to attract and fit hummingbirds (such as a narrow shape and large amounts of nectar) and, in addition, traits thought to deter bees (red coloration, lack of nectar guides, and lack of scent) (Schemske and Bradshaw 1999; Castellanos et al. 2004; Bergamo et al. 2016). Interestingly, the evolution of hummingbird pollination has primarily been documented in one direction, from ancestral bee pollination to derived hummingbird pollination (Rosas-Guerrero et al. 2014; Abrahamczyk and Renner 2015; Kay and Grossenbacher 2022). However, there is clade-specific variation in this directionality, with the reasons for pollinator transitions remaining unclear (Tripp and Manos 2008; Rosas-Guerrero et al. 2014; Abrahamczyk and Renner 2015; Kay and Grossenbacher 2022; Barreto et al. 2024).

Hummingbird-pollinated flowers often lack floral scent and therefore provide an excellent system to study repeated trait loss (Knudsen et al. 2004; Wessinger 2024). Genomic studies have shown that, even for birds, hummingbirds have a low number of olfactory receptors (Driver and Balakrishnan 2021; Driver 2022), suggesting at least a reduced investment in this sensory modality. Behavioral experiments have also suggested that hummingbirds rely more on visual than olfactory cues when foraging (Goldsmith and Goldsmith 1982). In fact, floral scent analyses of hummingbird-pollinated flowers have shown either complete lack of floral scent or low emission rates of floral scent compounds (Knudsen et al. 2004; Byers et al. 2014; Amrad et al. 2016). The lack of floral scent in hummingbird-pollinated species suggested that there is a cost to emitting scent for these species. This could be due to energetic costs of compound production, but this is predicted to be minimal (Raguso 2016; Pichersky and Raguso 2018). Alternatively, there could be an ecological cost for hummingbird-pollinated species attracting bees, also called the bee avoidance hypothesis (Castellanos et al. 2004; Bergamo et al. 2016). Finally, floral scent reduction could also evolve as an adaptation to prevent the attraction of florivores or nectar robbers, which could be attracted by the scent (Galen et al. 2011; Theis and Adler 2012).

A fascinating group in which to study the evolution of hummingbird pollination and potential trait loss is the genus Costus (Costaceae). The genus contains over 80 species of perennial understory herbaceous monocots, the majority of which have diverged rapidly within the last 3 million years in Central and South America (Maas 1972, 1977; Vargas et al. 2020). Since establishing in the American tropics, hummingbird pollination has evolved repeatedly and independently in Costus from an ancestral state of bee pollination (Kay and Grossenbacher 2022). Specifically, bee-pollinated species are pollinated primarily by orchid bees (Apidae: Euglossini), with females visiting flowers for nectar and pollen and males visiting for nectar (Kay and Schemske 2003). Hummingbird-pollinated species are primarily pollinated by nonterritorial hermit hummingbirds (Kay and Schemske 2003). While different species were initially classified as hummingbird or bee pollinated (Maas 1972, 1977), many species do not fit the classic paradigm of long tubular red flowers associated with hummingbird pollination (Grant and Grant 1968). For example, many bee-pollinated flowers have some red coloration (Kay and Schemske 2003), and hummingbird-pollinated flowers are short (Yost and Kay 2009). Despite this variation, a single phenotypic optimum is shared across most of the independent transitions to hummingbird pollination in the clade (Kay and Grossenbacher 2022). Traits including small flower size, the absence of nectar guides, and brightly colored floral bracts are all predictive of hummingbird pollination (Kay and Grossenbacher 2022). Furthermore, Costus species show extensive sympatry between different pollination syndromes (Kay and Schemske 2003; Vargas et al. 2020; Kay and Grossenbacher 2022), and even the specialist herbivores are shared across Costus species (García-Robledo et al. 2013), allowing us to isolate the effects of pollination syndrome from geographic location or herbivory. The combination of recent genomic resources, a well-resolved phylogeny, documented pollinators, and the possibility of collecting samples both in the field and in greenhouses, makes Costus a powerful system to study the evolution of convergence on hummingbird pollination (Vargas et al. 2020; Kay and Grossenbacher 2022; Harenčár et al. 2023).

While the morphological and color changes involved in the evolution of hummingbird pollination have been well studied in Costus, the expected scent changes have not been characterized. In this study, we describe floral scent in 30 species of Costus sampled across the phylogeny and including eight independent origins of hummingbird pollination (as determined in Kay and Grossenbacher 2022). We test for differences in floral scent between bee-pollinated and hummingbird-pollinated species and investigate how floral scent has evolved across the genus. We then focus on one clade of Costus with variation in floral scent emission and generate floral transcriptomic data. We use this data to identify gene expression differences associated with floral scent production to better understand the genetic changes underlying floral scent evolution in this group. Finally, we annotated a family of genes important for the production of terpenoid floral scent compounds (the terpene synthase [TPS] genes) in three previously available Costus genomes, and a high-quality genome assembly for Costus allenii generated in this study, encompassing two independent transitions to hummingbird pollination. High-quality annotations allow us to analyze the evidence of gene loss or pseudogenization following transition to hummingbird pollination in this gene family that plays an important role in floral scent evolution. Our study, therefore, combines multiple approaches to characterize repeated phenotypic evolution and the molecular mechanisms of these changes with shifts to hummingbird pollination in Costus.

Results

Bee-Pollinated Species Emit More Compounds and More Diverse Floral Scents

We collected 100 samples from 30 Costus species including species previously described as both bee and hummingbird pollinated from the field and the greenhouse (Fig. 1a; supplementary table S1, Supplementary Material online). Using multivariate techniques, we visualized the floral scent profiles to identify the broad patterns in our data set. Before filtering our data, we found that samples from hummingbird-pollinated species did not differ from ambient control samples (Fig. 1b). In contrast, samples from bee-pollinated species spread across a wider section of chemospace, with many samples visually separated from both ambient and hummingbird-pollinated samples (Fig. 1b). We found that both pollination group and species (nested within pollination group) explained variation in floral scent (permutational multivariate analysis of variance [PERMANOVA], pollination group: F_2,167 = 8.0, P < 0.001; species: F_28,167 = 2.5, P < 0.001). Interestingly, bee-pollinated species were significantly different from both hummingbird-pollinated species (pairwise PERMANOVA, P = 0.003) and ambient control samples (pairwise PERMANOVA, P = 0.003), whereas hummingbird-pollinated species did not differ from ambient control samples (pairwise PERMANOVA, P = 0.07).

Fig. 1.

a) Representative images of two pairs of closely related bee- and hummingbird-pollinated species: C. allenii and C. lasius and C. erythrophyllus and C. spiralis. b) NMDS plot illustrating in the variation in the unfiltered data set (stress = 0.12). Points are labeled as bee-pollinated, hummingbird-pollinated, or ambient samples. Ambient samples are control samples of the air taken in the same area as the focal plant individual. Ellipses are plotted for species with more than two samples to show distributions. c) Differences in bee-pollinated and hummingbird-pollinated Costus species in terms of Compound richness (ANOVA, pollination group: F_1,70 = 60.4, P < 0.0001; species: F_28,70 = 9.4, P < 0.0001). d) Functional Hill Diversity Index (ANOVA, pollination group: F_1,52 = 6.8, P = 0.012; species: F_26,52 = 2.9, P < 0.001). e) Total compound amount emitted (measured by proxy of total ion abundance) (ANOVA, pollination group: F_1,70 = 109, P = <0.0001; species: F_28,70 = 2.8, P < 0.001). Fifteen individuals with the highest compound emission were removed from E for visualization.

After filtering likely contaminants from the ambient air, we obtained a final data set with 60 floral scent compounds (supplementary fig. S1 and table S2, Supplementary Material online). The majority of these compounds were terpenoid (87%, 52/60), with a few aromatics and compounds of unknown class. The data set included both monoterpenes and sesquiterpenes. No one compound class was restricted to a specific pollinator group. After removing the ambient compounds, we found that one bee-pollinated sample and 18 hummingbird-pollinated samples had no detectable floral scent, confirming the similarity of the hummingbird-pollinated samples to the ambient air samples. The unscented hummingbird-pollinated samples were from six species. For two species, Costus montanus and Costus stenophyllus, all individuals emitted no detectable floral scent (4/4 individuals). Furthermore, most Costus chartaceus (4/5), Costus lasius (4/7), and Costus wilsonii (5/8) and a single Costus productus individual (1/5) lacked floral scent. To facilitate visualization of our final filtered data set using nonmetric multidimensional scaling (NMDS), we removed the nineteen samples with no detectable floral scent. When this filtered data set was plotted, we observed more overlap between the hummingbird- and bee-pollinated samples (supplementary fig. S2, Supplementary Material online). This suggests that there is little difference in the scent composition of bee-pollinated samples and those samples from hummingbird-pollinated species, which did emit floral scent, although they did still show a statistically significant difference.

We found that bee-pollinated flowers emitted a richer floral scent, averaging 9 (SD 10) compounds in comparison to only 2 (SD 4) for hummingbird-pollinated flowers (Fig. 1c; supplementary fig. S3, Supplementary Material online). We confirmed that this was the case for both plants raised in a greenhouse and those sampled in the field (supplementary fig. S4, Supplementary Material online). Similar to the compound richness, we used a measure of biochemical and structural diversity of the compounds present in a sample to calculate the overall chemical diversity. The Functional Hill Diversity Index considers not only the total number of compounds (richness) but also their biosynthetic complexity. We found that the bee-pollinated floral scents had a higher chemical diversity (Fig. 1d). Bee-pollinated flowers also emitted a higher abundance of floral scent, as measured by total ion abundance in each sample, a proxy for compound amount (Fig. 1e). We validated this proxy using absolute amounts in nanograms for a subset of the data set, again finding that the bee-pollinated species (C. allenii) has significantly higher compound emission when compared to the hummingbird-pollinated species (C. lasius) (supplementary fig. S5, Supplementary Material online).

To see how individual compounds varied between bee- and hummingbird-pollinated species, we calculated the percentage of individuals in each pollination group that emitted the most common floral scent compounds in the data set. We found that these compounds are present in a higher proportion of bee-pollinated flowers than hummingbird-pollinated flowers (supplementary fig. S6, Supplementary Material online). In particular, (E)-beta-ocimene and beta-caryophyllene show the largest percentage differences between the two pollination groups (supplementary fig. S6, Supplementary Material online). We used generalized linear models (GLMs) with a binomial error distribution and logit link function to identify compounds, which significantly differed between pollinator groups (compound presence ∼ pollinator group/species). Significance was determined with likelihood ratio tests. Of the 60 compounds in our data set, 28 were significantly different between pollinator groups (supplementary table S2, Supplementary Material online), including all of the most abundant compounds, except alpha- and beta-pinene (supplementary fig. S6 and table S2, Supplementary Material online). Some of these compounds do not vary independently between samples. We find evidence for six modules: an aromatic module, a terpenoid module, a monoterpene module, two sesquiterpene modules, and a large sesquiterpene-dominated module (supplementary fig. S7, Supplementary Material online). We found that the compound that showed the biggest difference between pollinator groups, (E)-beta-ocimene, was not part of any these large modules of compounds.

Phylogenetic Analyses Show That Hummingbird-Pollinated Species Have Lost Floral Scent

To understand the evolution of floral scent in Costus, we carried out phylogenetic analyses using a previously published phylogeny (Vargas et al. 2020). To examine scent differences between bee- and hummingbird-pollinated species while accounting for phylogenetic relationships, we used a method of randomizing residuals in a permutation procedure (Adams and Collyer 2018). We found that compound richness, compound diversity, compound amount, and (E)-beta-ocimene emission were all significantly different between bee-pollinated and hummingbird-pollinated species (compound richness, F_1,26 = 7.2, P = 0.002; compound diversity, F_1,26 = 5.8, P = 0.008; compound amount, F_1,26 = 627.5, P = 0.003; (E)-beta-ocimene production, F_1,26 = 4.7, P = 0.03). (E)-beta-caryophyllene production did not differ between pollinator groups (F_1,26 = 1.0, P = 0.4). Although phylogenetic ANOVA with simulations based on Brownian motion is a more common approach to these types of comparisons, it was inappropriate for our data set. We found a strong correlation between pollination group and the phylogenetic covariance matrix using two-block partial least squares (PLS) analysis (mean correlation [rPLS] = 0.56; P = 0.01). This reduces the statistical power of testing group differences using phylogenetic ANOVA due to the high aggregation of pollinator groups on the phylogeny (Adams and Collyer 2018). Overall, our comparative analyses complement our previous nonphylogenetic analyses, again showing that bee-pollinated species emit higher amounts of more diverse and complex floral scents and have higher (E)-beta-ocimene emission.

We found no evidence of phylogenetic signal in compound richness (λ = 7.4e⁻⁰⁵, P = 1), compound diversity (λ = 4.15e⁻⁰⁵, P = 1), compound amount (total area of floral scent peaks) (λ = 7.4e⁻⁰⁵, P = 1), (E)-beta-ocimene emission (λ = 7.3e⁻⁰⁵, P = 1), and (E)-beta-caryophyllene emission (λ = 7.4e⁻⁰⁵, P = 1). We also tested for evidence of phylogenetic signal in the multivariate floral scent data set, detecting nonsignificant almost zero phylogenetic signal (K_mult = 0.43, P = 0.68).

To further investigate how floral scent has evolved across the Costus phylogeny, we carried out maximum likelihood ancestral state reconstruction of mean compound emission. We find the ancestral state of all Costus species found in tropical America is to emit more compounds than average hummingbird-pollinated species (4.5 vs. 2). In this reconstruction, the ancestral Costus population emitted an average floral scent—more compounds than hummingbird-pollinated species but less than the average bee-pollinated species (Fig. 2). Many hummingbird-pollinated species have since completely lost floral scent, as shown by the white circles (Fig. 2). We see evidence for low or no scent emission in representatives of all eight origins of hummingbird pollination included in our data set. In contrast, some bee-pollinated species, such as C. allenii, have evolved much stronger floral scents than the ancestral state.

Fig. 2.

Ancestral state reconstruction of floral scent in the genus Costus. Pollinator for each species is illustrated by the right-hand dots. Mean compound richness emitted by each species is shown by the left-hand dots with each shade of gray representing a range of means (where 0 to 2 includes means up to 2.99, 3 to 6 up to 6.99, etc.). Darker dots indicate more compounds are emitted. Predicted compound emission at each node is shown following the same color scale. Tip labels following (Vargas et al. 2020).

Loss of Floral Scent Is Correlated with Downregulation of Floral TPS Genes

To analyze the genetic basis of floral scent evolution, we first chose a clade of closely related species with variation in floral scent emission and pollination systems. We chose the clade containing C. lasius due to the variation in floral scent emission and the availability of a high-quality genome. In this clade, C. lasius (hummingbird-pollinated) emits on average three floral scent compounds, whereas C. allenii (bee-pollinated) emits 24 (one of the highest in our data set), and Costus villosissimus (bee-pollinated) emits seven compounds (Fig. 3a).

Fig. 3.

Differences in floral scent and gene expression between C. allenii, C. lasius, and C. villosissimus. Species differ significantly in a) compound richness (ANOVA, F_2,19 = 20.9, P < 0.001) and b) expression of TPS transcripts (ANOVA, F_2,16 = 508.4, P < 0.001). In b), TPS transcripts are summed for all TPS loci and then the average expression calculated by dividing by the number of TPSs expressed in that species (20 for C. allenii, 5 for C. villosissimus, and 1 for C. lasius).

We first produced floral transcriptomes of all three species. After clustering transcripts with over 95% similarity, we found 112,998 transcripts present in C. allenii, 80,208 in C. lasius, and 96,886 in C. villosissimus. Remapping rates were high in all three species: 99.2% in C. allenii, 99% in C. lasius and 99.1% in C. villosissimus. All three also had Benchmarking Universal Single-Copy Orthologs (BUSCO) scores of more than 80% (supplementary table S3, Supplementary Material online), suggesting high-quality transcriptomes with most single-copy orthologs from the Embryophyta database present in all species.

Due to the dominance of terpenoid compounds in the floral scent, we focused on the expression of TPS genes. TPS genes are responsible for producing terpenoid compounds. We found that 20 TPS genes were expressed in the floral transcriptome of C. allenii, compared to five in C. villosissimus and only one in C. lasius, when expression was defined as one transcript per million (TPM) in at least one sample of the species. Furthermore, we found that counts of TPS expression per TPS gene expressed were much higher in both C. allenii and C. villosissimus when compared with C. lasius (Fig. 3b). In fact, in C. allenii, a TPS is the highest expressed gene in the floral transcriptome, and TPSs are also the fourth and sixth most expressed genes in the transcriptome. On average, 12% of total expression can be attributed to TPSs in C. allenii, 0.17% in C. villosissimus, and 0.00014% in C. lasius, highlighting the orders of magnitude differences between species.

Floral Scent Loss Is Not Driven by Gene Loss in Costus

While the transcriptome assemblies can give us insight into levels of TPS expression, it is hard to determine gene copy number due to the redundancies in the assemblies. For this, we need genome data to allow us to determine the number of TPS genes in each species. There are currently three publicly available Costus genomes: Costus bracteatus (bee), C. lasius (hummingbird), and Costus pulverulentus (hummingbird) (Valderrama et al. 2022; Harenčár et al. 2023). Costus lasius and C. pulverulentus represent two independent transitions to hummingbird pollination (Kay and Grossenbacher 2022). To generate a data set of two hummingbird- and two bee-pollinated species, we assembled an additional high-quality reference genome for bee-pollinated C. allenii with a BUSCO score of 98.8% (supplementary table S4, Supplementary Material online for assembly statistics; supplementary fig. S8, Supplementary Material online to see alignment with C. lasius genome). We generated high-quality whole-genome annotations for each species with high completeness and consistency with all BUSCO scores over 90% (supplementary table S5, Supplementary Material online).

We annotated between 63% and 75% of each genome as repeat elements (supplementary tables S6 to S9, Supplementary Material online). We found that, of the elements that could be classified, retrotransposable elements represented the largest percentage of the genome. These are the most abundant form of repeat elements found in angiosperm genomes (Oliver et al. 2013). Many repeats were not classified into previously known families, perhaps suggesting the existence of new repeat elements.

We annotated a total of 64 TPSs in C. allenii, 44 in C. bracteatus, 40 in C. lasius, and 43 in C. pulverulentus. The TPSs belonged to previously described TPS families TPS-a, TPS-b, TPS-c, TPS-e/f, and TPS-g (Fig. 4; supplementary fig. S9, Supplementary Material online). Those TPSs responsible for floral scent biosynthesis are primarily found in the TPS-a, TPS-b, and TPS-g families. TPS-a members tend to produce monoterpenes, TPS-b members sesquiterpenes, and TPS-g members either monoterpenes or sesquiterpenes (Chen et al. 2011).

Fig. 4.

Phylogenetic analysis of TPS gene families. We found the highest rates of gene gain and lowest rates of gene loss for C. allenii with the other three species exhibiting mixes of loss and gains across the gene tree. The phylogeny was constructed in IQ-TREE with 1,000 bootstraps and model JTT + F + R4. TPS family classifications based on A. thaliana are illustrated.

We used computational analysis of gene family evolution (CAFE5) to model gene gain and loss across the species phylogeny for each TPS clade identified in Fig. 4 (Mendes et al. 2020). For the bee-pollinated species, we found that C. allenii has gains in four clades with losses in one while C. bracteatus shows gains in one clade and losses in four clades. For the hummingbird-pollinated species, we found that C. lasius has gains in two clades and losses in three, while C. pulverulentus exhibits no gains and losses in two clades. There is no clear correlation between pollination group and gene losses or gains with some evidence that C. allenii shows enhanced gene duplication. We also estimated gene duplications and losses for each species using a gene tree-species tree reconciliation algorithm, reconcILS (Mishra et al. 2024). The algorithm reconciles the gene tree and species tree considering not only duplication and loss events but also incomplete lineage sorting (ILS), which has been documented in Costus (Uckele et al. 2024). Using this approach, we identified the greatest number of lineage-specific duplication events for C. allenii (30), with 9 in C. bracteatus, 11 in C. lasius, and 5 in C. pulverulentus. We found similar numbers of lineage-specific loss events in each species with 7 in C. bracteatus, 8 in C. allenii, 9 in C. pulverulentus, and 14 in C. lasius. The lineage-specific duplications in C. allenii correspond to tandem duplicates in close proximity within the genome. We find support for increased lineage-specific duplications in C. allenii; however, our gene family analyses do not find evidence for increased gene losses in hummingbird-pollinated species.

We mapped the RNA-seq reads both to the C. allenii genome and the C. lasius genome. In C. allenii, we found a consistent percentage of successfully mapped reads attributed to TPSs as with the transcriptome (12% transcriptome, 13% mapped to either genome). We found some differences for both C. lasius and C. villosissimus. For the C. villosissimus transcriptome mapping, we found 0.17% of reads were attributed to TPSs, in comparison to 0.009% when mapped to either genome. Due to a lack of genome for C. villosissimus, it is unclear whether this could be due to genetic changes or rearrangements not captured by the available genomes. For C. lasius, we found only 0.00014% of reads mapped to the transcriptome were due to TPSs, in contrast to 0.011% of the genome-mapped reads.

When C. allenii reads were mapped to the C. allenii genome, we found high levels of multimapping. This is due to recent TPS duplications in C. allenii, which are highly expressed. To illustrate this, the highest gene expressed in C. allenii is TPS49, and the third most expressed “gene” consists of reads that multimap to TPS49 and TPS52. This clade is represented by only one gene in each of the other species. Other recent lineage-specific duplications include TPS2, TPS3, TPS4, and TPS5 where the majority of reads multimap due to sequence similarity between duplicates. Both of these clades have only one member in each of the other species with an available genome sequence (supplementary fig. S9, Supplementary Material online). This highlights the complexity of mapping to gene families with recent duplications, especially when they are highly expressed, meaning that the typical choice of ignoring multimapped reads will lose large percentages of mapped reads from these duplications. We, therefore, decided to carry out our differential expression analysis using the C. lasius genome.

By mapping all reads to the C. lasius genome, we identified 4,790 genes differentially expressed between C. allenii and C. lasius, 7,678 between C. allenii and C. villosissimus, and 6,191 between C. lasius and C. villosissimus. Of the 40 TPS genes found in the C. lasius genome, 15 were differentially expressed between C. allenii and C. lasius, with 13 of these upregulated in C. allenii. This includes TPS33, orthologous to the TPS49/52 clade, the most highly expressed genes in C. allenii. We found 19 TPSs that were differentially expressed between C. allenii and C. villosissimus, with 14 upregulated in C. allenii and 5 in C. villosissimus. Ten TPS genes were differentially expressed between C. lasius and C. villosissimus, with 7 upregulated in C. villosissimus and 3 in C. lasius. Overall, more TPSs were upregulated in C. allenii when compared with either C. lasius or C. villosissimus.

Discussion

Repeated evolutionary transitions provide unique opportunities to study the extent of convergence both at the phenotypic and molecular level, shedding light on the predictability of evolution. Transitions to hummingbird pollination are widespread in the Americas and are predicted to be associated with loss of floral scent; however, this has not been studied in a comparative framework across multiple independent transitions within the same genus. In this study, we combine comparative chemical ecology with genomics and transcriptomics and demonstrate that floral scent in Costus is dominated by terpenes and its evolution is highly associated with pollinator type. We found that hummingbird-pollinated Costus species have lost or reduced their floral scent across multiple independent evolutionary origins, whereas bee-pollinated species in the Americas exhibit either maintenance of floral scent or gain of a more diverse scent profile. We find that this is related to downregulation of genes associated with floral scent emission (in this case TPSs) in floral tissues of hummingbird-pollinated species but is not accompanied by loss of TPS genes.

Overall, we found that bee-pollinated species emitted richer and more diverse floral scents at higher abundances. We did not find any unique compounds associated with either pollinator group, suggesting that the hummingbird-pollinated species emitted a subset of that emitted by bee-pollinated species, at lower amounts. We found the floral scent was mainly composed of terpenoids, with a few aromatic compounds. The most prevalent compounds detected in our data set, such as beta-ocimene, are typical floral-scent compounds found in insect-pollinated flowers (Farré-Armengol et al. 2017). Interestingly, bee-pollinated Costus species showed much greater variation in their floral scents than hummingbird-pollinated species, with some having very diverse floral scents, yet others emitting far fewer compounds. Bee-pollinated species also show greater morphological variation, possibly reflecting variation in orchid bee diversity (Kay and Grossenbacher 2022). Furthermore, not all species in our data set follow the expected patterns. For example, Costus arabicus and Costus vinosus (bee-pollinated) emit little scent. It is unclear whether the variation in floral scent between different bee-pollinated species is related to pollinator preference, other abiotic or biotic factors, or historical contingency.

There are multiple hypotheses for why floral scent would be lost in hummingbird-pollinated species. Firstly, it could be because there is simply no selection to maintain the production due to poor sense of smell, or lack of preference shown for floral scent, by hummingbirds (Goldsmith and Goldsmith 1982; Knudsen et al. 2004; Driver and Balakrishnan 2021). If this were the case, we might expect to find a correlation between length of time following pollinator transition and extent of floral scent loss. However, our data show that downregulation of floral scent can occur on relatively quick evolutionary timescales. The Latin American radiation of Costus is dated at 3 million years (95% confidence interval [CI], 1.50 to 4.87) and so all transitions to hummingbird-pollination have occurred within this timeframe (Vargas et al. 2020). Costus wilsonii, C. lasius, and Costus spiralis are all hummingbird-pollinated species that are most closely related to bee-pollinated species, representing more recent evolutionary shifts to hummingbird pollination. Yet all three species emit little to no floral scent, suggesting there can be quick breakdown of floral scent emission following pollinator transitions. This observation is consistent with previous studies that have shown that floral scent often evolves rapidly in response to selection (Ramos and Schiestl 2019; Zu et al. 2020; Liu et al. 2024).

The rapid and repeated loss of scent in hummingbird-pollinated species suggests a shared strong selection pressure against floral scent emission. One potential mechanism is the energetic cost of terpene production. A similar argument has been made in the case of eye regression in cavefish. If vision is energetically costly to maintain, then it should be selected against when no longer useful (Moran et al. 2015; Sifuentes-Romero et al. 2023). This is supported by evidence of the energetic costs of maintaining a visual system (Moran et al. 2015). The metabolic cost of terpene production is unknown but is thought to be minimal (Raguso 2016; Pichersky and Raguso 2018). In fact, it has been argued that the complex network of terpene metabolism, in which individual enzymes produce multiple products, makes terpene pathways more efficient, compared with a linear pathway with separate enzymes and regulators for each step (Lanier et al. 2023).

An alternative scenario to the energetic cost is that terpene production has ecological costs in hummingbird-pollinated species. Floral scent may attract florivores and floral larcenists in some systems (Galen et al. 2011). In Costus, most floral larceny is by small hummingbirds (e.g. Phaethornis striigularis and Eupherusa eximia) that pierce the base of the flower under the tough floral bract to obtain nectar, and they rob bee- and hummingbird-pollinated Costus at similar rates (KMK, unpublished data). Weevils (likely Cholus sp.) have also been observed occasionally consuming flowers of both pollination syndromes (although consumption rates were not recorded). Thus, it is unlikely that floral scent loss functions to deter florivory. We propose that the selection for floral scent reduction is most likely the result of strong selection against bee attraction (Castellanos et al. 2004; Bergamo et al. 2016). Birds do not consume pollen, leading to less pollen loss, and birds can travel longer distances between plants and promote higher gene flow, which is beneficial to combat inbreeding depression (Thomson and Wilson 2008; Krauss et al. 2017; Dellinger et al. 2022). One caveat is that comparisons in pollination effectiveness between bees and hummingbirds have not been carried out in orchid bees. Both male and female orchid bees tend to forage over long distances, perhaps resulting in a more “bird-like” pollen transport (Janzen 1971, 1981; Wikelski et al. 2010; Opedal et al. 2017; Kay and Grossenbacher 2022; Gamba and Muchhala 2023). While pollinator efficiency is unknown in Costus, rapid floral scent loss is congruent with the hypothesis of “bee avoidance” following the establishment of hummingbird pollination.

A final piece of evidence to support bee repellence is the large reduction in beta-ocimene emission in hummingbird-pollinated flowers relative to bee-pollinated flowers, as this compound is a common insect pollinator attractant (Farré-Armengol et al. 2017). In Mimulus, the loss of monoterpenes, including beta-ocimene, maintains reproductive isolation between bee- and hummingbird-pollinated species (Byers et al. 2014). Interestingly, we found that α- and β-pinene were emitted by both bee- and hummingbird-pollinated species. One potential reason is that pinenes do not repel orchid bees but do repel other types of bees. Terpenoid-dominated scents have been found to repel honey bees (Larue et al. 2016), and in particular, α-pinene has a repellent effect on honey bees (Fernandes et al. 2019). Furthermore, α-pinene has been found in other hummingbird-pollinated species with a suggested role as a moth pollinator deterrent (Bischoff et al. 2014). We speculate that bee-pollinated species emit a mix of compounds, some of which potentially attract orchid bees, and some of which potentially repel other pollinators to narrow attraction to orchid bees, specifically. Then, when floral scent is reduced in hummingbird-pollinated flowers, the compounds that are attractive to orchid bees are lost but not those that are general bee repellants. Further experimental work is needed to test these hypotheses.

We found that the floral scents of Costus are dominated by terpene compounds. Although few Zingiberales have described floral scents, our results are consistent with previous reports from the genus Hedychium that also found terpene-dominated floral scents (Báez et al. 2011; Zhou et al. 2022). Terpene diversity in a given organism is mainly determined by the TPS gene family, which carries out crucial steps in terpene formation (Tholl 2006). We did not find evidence that hummingbird-pollinated species with an available genome sequence (C. lasius and C. pulverulentus) exhibited increased rates of TPS gene loss than the bee-pollinated species. There are multiple reasons why trait loss would be associated with changes in gene regulation rather than gene loss. While C. lasius represents a more recent transition to hummingbird pollination than C. pulverulentus, both transitions have occurred within the last ∼3 million years (95% CI, 1.50 to 4.87) as Costus radiated in Latin America (Vargas et al. 2020). There may not have been enough time for pseudogenization to occur, a process estimated to take 0.5 to 6 million years (Marshall et al. 1994; Esfeld et al. 2018). Potentially, the genes will eventually be lost but are currently in the first phase, reduction in expression, which according to the “step-wise” model to pseudogenization is then followed by relaxed selection and then loss-of-function mutations (Graham et al. 2023).

Alternatively, we might not expect any TPS genes to be lost even in the long term. Terpenes, and therefore TPSs, are important for many other roles in plants. Plants use terpenes to deter herbivores, to attract parasitoids or predators of herbivores, and as antifungal and antibacterial compounds to prevent disease (Gershenzon and Dudareva 2007). These pleiotropic functions of TPSs could explain why we observe downregulation in floral tissues specifically rather than any pseudogenization. One argument against this is that in other systems TPS pseudogenization does occur following pollinator transitions. In Mimulus, loss of monoterpene production (specifically beta-ocimene) is not due to expression differences but loss-of-function mutations (Peng et al. 2017). It is unclear whether pseudogenization will occur at some point in Costus.

Although we did not observe gene loss in hummingbird-pollinated lineages, we did observe extensive lineage-specific duplications in the bee-pollinated species C. allenii. Costus allenii was one of the species with the highest scent diversity in our data set, and we found multiple TPS genes to exhibit several recent lineage-specific duplication events. We also found extremely high TPS expression, accounting for approximately 12% of expression in floral tissue. The gene with the highest expression in floral tissue was found to be a TPS-a family member, with high expression also found for several TPS-b members. These two families produce sesquiterpenes and monoterpenes, respectively, both of which are found in the floral scent of C. allenii. In C. lasius and C. villosissimus, we found reduced TPS expression compared to C. allenii, as expected based off their reduced floral scent emission. We also saw high levels of phenotypic variation in C. villosissimus, however, there is no publicly available genome, making interpretation of the results more difficult. We speculate that C. allenii has higher floral scent emission than closely related C. villosissimus because C. allenii grows as isolated individuals in deeply shaded, perennially wet forest, whereas C. villosissimus grows at forest edges and large treefall gaps and often has conspecifics nearby (Chen and Schemske 2015). Thus, C. allenii may have evolved the emission of stronger floral scent to attract pollinators in dense forests where flowers are less visible and pollinators are less abundant. While the overall trends are clear, we hesitate to overinterpret the gene expression data as we have only sampled at one time point and it is possible that there is floral scent emission happening earlier that is not captured in our data.

We observe more consistency and predictability in the phenotypic patterns than the underlying molecular mechanisms. In many systems, parallel changes in floral scent have accompanied independent pollinator shifts, including the current study (Knudsen et al. 2004; Byers et al. 2014). This can also be seen in other reproductive shifts, such as transitions to self-fertilization (Woźniak et al. 2022). In some systems, the underlying molecular mechanisms of floral scent loss are also convergent, having been shown to involve the same loci in multiple independent evolutionary events. For example, loss of floral scent in different Mimulus lineages is due to changes at the same loci, including frameshifts, deletions, and potentially posttranscriptional regulation (Peng et al. 2017). A similar pattern was found in two different systems that have lost benzaldehyde floral scent production due to either evolution of self-fertilization (Sas et al. 2016) or transition to hummingbird pollination (Amrad et al. 2016). The same locus contains loss-of-function mutations in both cases (Raguso 2016). However, this is not universally true. In Capsella, the loss of beta-ocimene following transition to self-fertilization in one lineage is due to changes in subcellular enzyme localization, but genetic mapping results suggest another locus is responsible for scent loss in a related species (Woźniak et al. 2022). In this study, we saw no evidence for pseudogenization. Instead, we found evidence that gene expression differences are responsible for scent differences between Costus species, with scent loss likely due to changes in, potentially tissue-specific, gene expression.

In summary, we have shown that Costus species exhibit repeated loss of floral scent following transitions from bee to hummingbird pollination, suggesting a shared strong selection pressure favoring scent loss. We also find bee-pollinated species that have recently gained more diverse scents, which, at least in C. allenii, is associated with lineage-specific TPS duplications. We show that scent loss is not associated with loss of TPS genes but rather appears to be due to gene downregulation. This demonstrates the capacity for rapid metabolic changes in response to selection following pollinator transitions. This system provides us with the exciting opportunity to identify the cis- or trans-regulatory changes driving gene expression differences in multiple independent transitions to hummingbird pollination, allowing us to study the extent of genetic convergence.

Materials and Methods

Floral Scent Sample Collection

We collected 100 samples of 30 species of Costus between 2019 and 2023 (supplementary table S1, Supplementary Material online). We sampled an average of 4 individuals per species for bee-pollinated species and 3 for hummingbird-pollinated species. We, therefore, expect that missing compounds due to undersampling will occur to a similar extent for both bee- and hummingbird-pollinated species. All samples for Costus glaucus, Costus laevis, C. montanus, C. stenophyllus, and C. wilsonii were collected in Costa Rica (La Gamba, Las Cruces, Las Alturas, and Monteverde). All other species were sampled in the University of California, Santa Cruz (UCSC) greenhouses in Santa Cruz, California, USA from plants composing the UCSC living Costus collection. For each day of sampling, both floral scent samples and ambient samples were collected (with one exception: a Costus scaber sample collected on 2019 April 8 was missing a control and so we used a control from the nearest date available, March 7). Scent was sampled from a single flower using dynamic headspace sampling from approximately 9 AM until 3 PM (6 h). This time point was chosen as it is the window of most pollinator activity during the lifespan of the flower (Kay and Schemske 2003). Flowers from living plants were placed inside oven bags (Reynold's). A Spectrex PAS-500 air pump with Tygon tubing was used to pull scented air through a volatile trap for 6 h at a flow rate of 200 mL air/min. The volatile trap contained a filter of glass wool and 50 mg Porapak. We conditioned scent traps by passing 5 mL of hexane before usage. Volatiles were eluted from the trap using 400 μL of hexane and stored at −20 °C.

Chemical Analysis

Samples were analyzed using Agilent model 5977A mass-selective detector with an Agilent GC model 7890B. Samples were run on a HP-5 Ultra Inert column (Agilent, 30 m × 0.25 mm, 0.25 μm). For analysis, 1 μL of each sample was injected in splitless mode using the ALS 7694 autosampler with helium as the carrier gas (250 °C injector temperature). The program started at 40 °C for 3 min and then rose at 5 °C/min to 210 °C. The temperature was held at 210 °C for 1 min, before rising at 20 °C/min to 300 °C where it was held for 2.5 min, and 315 °C for a final minute.

Compounds were identified by comparing mass spectra and retention indices with reference libraries. The identity of some compounds was confirmed by comparison to authentic reference standards (see supplementary table S2, Supplementary Material online). Compounds that were not found in at least two samples were removed. A two-step process was used to filter data sets for floral scent components. First, compounds were filtered to include only those that are found in five times higher amounts in floral than ambient samples using the “filter_ambient_ratio” function from the bouquet package (Powers et al. 2023). This resulted in a list of ambient compounds that were not considered floral scent compounds. Second, using the raw data matrices, we subtracted the ambient compound amounts from the corresponding floral scent amount collected on the same date. We then removed the ambient compounds identified by the package bouquet from this subtracted matrix to create the final data matrix of floral scent.

Statistical Analyses of Floral Scent

To visualize divergence in floral scent among samples, we used NMDS using the “metaMDS” function in vegan (Oksanen et al. 2020). To test for differences both between pollination groups, and between floral and ambient samples, we carried out a PERMANOVA using the “adonis2” function in vegan (Bray–Curtis distance matrix, 1,000 permutations). To identify which groups were significantly different, we carried out post hoc pairwise testing using the “pairwise.perm.MANOVA” function in the RVAideMemoire package with a Bonferroni correction (Hervé 2021). We did this for both filtered and unfiltered data sets using the raw peak areas for each compound included in the analysis.

We tested for differences in compound richness emitted by bee-pollinated and hummingbird-pollinated species using ANOVA with a nested model including species within pollinator group. To test for differences in the diversity of compounds emitted by bee-pollinated and hummingbird-pollinated species, we used the package chemodiv (Petrén et al. 2023). This package considers the biochemical and structural diversity of the compounds present in a sample to calculate the overall chemodiversity per sample. First, we used the “calcDiv” function to calculate the functional Hill diversity index for each sample. Then, we carried out an ANOVA test for the diversity indices, using a nested model of species nested within pollinator group to test for differences in chemodiversity. We also tested for differences in compound emission using ANOVA. In this case, a proxy for compound amount was used: the sum of the total area of floral scent peaks in the sample or the total ion abundance in the sample. In addition to this proxy, we also added 82.9 μg of 2-undecanone (10 μL of a 1:100 dilution in hexane) to all C. allenii and C. lasius samples to calculate absolute abundances of each compound for all 15 samples collected for these species.

To test for difference in compound presence between bee- and hummingbird-pollinated samples, we used GLMs with a binomial error distribution and logit link function. The response variable in our model was “presence” or “absence” of a specific compound. The explanatory variables were pollinator group and species with species nested within pollinator group (compound presence ∼ pollinator group/species). Significance was determined with likelihood ratio tests. We adjusted the P-values using the “p.adjust” function in R. To identify groups, or modules, or covarying compounds, we used the R package corrplot (Wei and Simko 2021). We performed hierarchical clustering based on Pearson correlation coefficients.

Phylogenetic Analyses of Floral Scent

To investigate the evolution of floral scent in a phylogenetic context, we used a previously published phylogeny (Vargas et al. 2020). First, we trimmed the phylogeny to include only those species for which we have scent data using the “drop.tip” function in ape (Paradis and Schliep 2019). To test for correlation between pollination group and the phylogenetic covariance matrix, we used a two-block partial least squares analysis using the “two.b.pls” function in the geomorph package (Adams and Otárola-Castillo 2013; Adams and Collyer 2018). A significant result means that our power to test for differences between groups using phylogenetic simulation-based is weakened due to group aggregation on the phylogeny. We then used the “phylosig” function in the package phytools to test for phylogenetic signal in multiple traits (compound richness, compound diversity, compound amount, (E)-beta-ocimene emission, and (E)-beta-caryophyllene emission) (Revell 2012). We also tested if floral scent data overall show phylogenetic signal using the multivariate generalization of Bloomberg's K (K_mult) (Adams 2014). Under Brownian motion, K_mult is expected to be 1. If K_mult > 1, higher phylogenetic signal is detected than the null expectation under Brownian motion. If K_mult < 1, lower phylogenetic signal is detected than the null expectation under Brownian motion. To calculate the K_mult of Costus floral scent, we used the filtered data set with 1,000 iterations to test significance. We carried out the analysis using the “physignal” function in the geomorph package (Adams and Otárola-Castillo 2013). Due to the weak phylogenetic signal detected in both multivariate and univariate analyses, we used the “procD.pgls” function in the RRPP package to test for differences in traits between pollinator groups rather than a simulation-based ANOVA (Adams and Collyer 2018; Collyer and Adams 2018). This function uses a method of randomizing residuals in a permutation procedure to evaluate the significance of phenotypic differences between groups considering the phylogeny and has higher statistical power than methods using phylogenetic simulations (Adams and Collyer 2018). We set lambda as calculated for the univariate variables. We corrected for multiple testing using the “p.adjust” function in R.

To better understand how floral scent has evolved in the genus Costus, we carried out an ancestral state reconstruction of the compound richness. The species-level phylogeny from (Vargas et al. 2020) contains 24 hummingbird-pollinated species, 25 bee-pollinated species, and 3 bee-pollinated African outgroups. Our subsampled phylogeny contains samples in proportion to the full phylogeny (13 hummingbird-pollinated species, 13 bee-pollinated species, and 1 bee-pollinated African outgroup). We note that there are limitations to this subsampled phylogeny but we find value in using it to infer general trends and not values at specific nodes. To do this, we used the function “ace” in the ape package for ancestral character estimation (Paradis and Schliep 2019). We divided the compound richness trait into 12 categories, each of which were assigned a different shade of blue. We plotted both the compound richness for each phylogenetic tip label and internal node to illustrate how floral scent has evolved in Costus.

Sample Collection for RNA-seq

Live individuals of C. allenii and C. villosissimus were collected in Soberanía National Park, Colón Province, Panama. Vouchers are deposited at University of Panama (C. allenii: 0132264 and 0132265; C. villosissimus: 0132270). Live individuals of C. lasius were collected in El Valle de Antón, Cocle Province, Panama. Voucher is deposited in Michigan State University herbarium (Kay 0321). For all three species, the individuals used for RNA-seq analyses were grown in greenhouses at UCSC. Freshly opened flowers of each species were removed at approximately 9 AM and flash frozen immediately in liquid nitrogen before being stored at −80 °C until extraction. Flowers of Costus open early in the morning and fall off by mid-afternoon.

RNA Extraction and Sequencing

Frozen flowers were removed and dissected on dry ice. Tissue from the petals and labella was broken down using a TissueLyser (Qiagen). RNA was then extracted using an RNeasy Mini Kit (Qiagen) with Qiashredder spin columns (Qiagen) following standard protocols. Samples were sent to Novogene for quality control, library preparation, and sequencing. Our final data set consisted of 19 samples: 8 C. allenii (from 4 individual plants), 6 C. lasius (from 2 individual plants), and 5 C. villosissimus (from 1 individual plant). Samples were sequenced on a NovaSeq 6000 using 150-bp paired-end reads. This generated approximately 53 million reads per library (mean = 52.74 million, SD = 12.01 million, N = 19). The raw sequence data were deposited in the SRA under the BioProject accession PRJNA1010186.

Transcriptome Assembly

Before assembling transcriptomes, we trimmed the raw reads using TrimGalore! (Martin 2011). We then assembled transcriptomes for each species using Trinity version 2.5.1 with default settings (Grabherr et al. 2011; Haas et al. 2013). To reduce redundancy, we clustered contigs with a minimum sequence identity of 95% using CD-HIT version 4.6.8 (Fu et al. 2012). We assessed transcriptome completeness and quality using BUSCO (v5.6.1) (embryophyta_odb10 data set) (Simão et al. 2015; Manni et al. 2021; Zdobnov et al. 2021). We also checked assembly quality by mapping RNA-seq reads back to the assembly using Bowtie (v2.4.1) to calculate mapping rates (Langmead and Salzberg 2012).

TPS Gene Expression Analysis

We extracted the longest isoforms for each Trinity “gene” using the Trinity script “get_longest_isoform_seq_per_trinity_gene.pl” and translated these sequences. We then searched for potential TPSs using hmmsearch two HMM profiles: Terpene_syth_C (PF03936) and TPS N-terminal domain (PF01397). We used the R package rhmmer to parse the output from hmmsearch resulting in a list of putative TPS sequences for each species. To reduce the TPSs to only those expressed in the data set, we used a threshold of TPM > 1 (transcripts per million) in at least one individual of a species (Eddy 2011; Arendsee 2017). To calculate expression of each putative TPS, we normalized counts using gene length corrected trimmed mean of M-values (GeTMM), an approach that improves intersample and intrasample comparisons (Smid et al. 2018). To do this, we first calculated fragments per kilobase of transcript per million fragments mapped for each transcript and carried out TMM normalization using the package edgeR (Robinson et al. 2010). We then combined GeTMM values for transcripts that were identified by Trinity to be the same gene. This does not normalize counts between species but is used to determine whether genes are lowly or highly expressed within a species.

Costus allenii Genome Assembly

We selected a wild-collected individual C. allenii from Pipeline Road, Soberanía National Park, Colón Province, Panama for de novo genome assembly. A voucher for this population of C. allenii is deposited at the University of Panama Herbarium (PMA: 0132264). This individual was grown to flowering in the UCSC greenhouses from a cutting, and we harvested 1.5 g fresh tissue from leaf meristems and flower buds. We isolated cell nuclei following Workman (Workman et al. 2019) and then used the Circulomics Nanobind Plant Nuclei Big DNA kit to extract HMW DNA. We assessed DNA concentration with a Qubit Fluorometer (Thermo Fisher Scientific) and purity with 260/280 and 260/230 absorbance ratios from a NanoDrop Spectrophotometer (Thermo Fisher Scientific) and ran a 7% agarose gel at 70 V for over 8 h to assess DNA fragment integrity and size. HMW DNA was sent to the UC Davis Genomics Center for additional cleaning, Pacific Biosciences HiFi library preparation, and sequencing on one PacBio Revio SMRT cell. This produced 74.8 Gb of HiFi yield and 7.9 M HiFi reads. We assembled the PacBio data into a primary and alternative assembly using HiFiasm (v0.19.8) (Cheng et al. 2021). We then used RagTag (v2.1.0) (Alonge et al. 2022) to scaffold C. allenii contigs using the closely related C. lasius as a reference. We used D-GENIES (Cabanettes and Klopp 2018) with Minimap2 (Li 2018) to generate a dotplot comparing the two assemblies both before and after scaffolding. The genome assembly quality was assessed using BUSCO v5.6.1 (Simão et al. 2015). Raw sequence data and the final genome assembly were submitted to the National Center for Biotechnology Information (NCBI) under BioProject PRJNA1091680.

Genome Annotations

Of the three species for which we generated RNA-seq data, an unannotated genome was previously available for C. lasius (Harenčár et al. 2023). In this study, we also generated a genome assembly for C. allenii. In addition, previously published genomes are available for both C. bracteatus and the more distantly related C. pulverulentus (Valderrama et al. 2022; Harenčár et al. 2023). We started by processing RNA-seq data as evidence for annotations. Trimmed reads from C. lasius were mapped using 2pass mapping in STAR to the C. lasius genome and the C. pulverulentus genomes (Martin 2011; Dobin et al. 2013). Bam files were concatenated for use as described below. Trimmed reads from C. allenii were mapped using 2pass mapping in STAR to both the C. allenii and the closely related C. bracteatus genomes and bam files concatenated per species (Martin 2011; Dobin et al. 2013).

We annotated the genomes using BRAKER3 (v3.0.7), which combines RNA-seq and protein data in an automated pipeline (Lomsadze et al. 2005, 2014; Stanke et al. 2006, 2008; Gotoh 2008; Iwata and Gotoh 2012; Buchfink et al. 2015; Hoff et al. 2016, 2019; Kovaka et al. 2019; Brůna et al. 2020, 2021; Pertea and Pertea 2020; Gabriel et al. 2023). This pipeline requires a masked genome. We identified and annotated tandem repeats and transposable elements using RepeatModeler (Flynn et al. 2020) (v2.0.4 for C. lasius and v2.0.5 for the other genomes) and RepeatMasker v4.1.5 (Tarailo-Graovac and Chen 2009). We then used RepeatMasker with the outputs of RepeatModeler to generate softmasked genomes. The masked genomes were then provided to BRAKER along with plant proteins downloaded from OrthoDB (odb10_plants) (Zdobnov et al. 2021). For all species, BRAKER was run in ETP mode that combines evidence from both proteins and RNA-seq (Gabriel et al. 2021). BUSCO v5.6.1 (Simão et al. 2015) and OMArk (Nevers et al. 2025) were used to evaluate the completeness and consistency of the sets of gene models. During the project, the C. lasius genome was updated (direction of arms on chromosome 1 was altered) and so we rescaffolded chromosome 1 of C. allenii and transferred our annotation using liftoff (Shumate and Salzberg 2021).

TPS Gene Family Annotation

Following whole-genome annotation, we manually curated the gene models for TPS genes in the annotation. To do this, we used bitacora (Vizueta et al. 2020), a pipeline that curates an existing annotation and finds additional gene family members. In this pipeline, BLASTP and hmmer are used to search for gene family members of interest within the existing annotation (Altschul et al. 1990; Eddy 2011). Additional regions of the genome are then searched using TBLASTN and are annotated using GeMoMa (Keilwagen et al. 2019). hmmer is used to validate the gene models. As input, we used two HMM profiles: Terpene_syth_C (PF03936) and TPS N-terminal domain (PF01397) downloaded from Pfam. In addition, we provided previously annotated genes from Arabidopsis thaliana and Oryza sativa (Chen et al. 2011; Yu et al. 2020; Jia et al. 2022). Gene models from bitacora and the genome-wide annotation from BRAKER were compared and TPS genes were manually curated using IGV (Robinson et al. 2011). We also searched for additional TPSs by carrying out exonerate searches using the curated Costus gene models (Slater and Birney 2005). We checked all annotated TPSs for complete protein domains using the NCBI conserved domain search (Marchler-Bauer et al. 2015). We excluded proteins with <250 amino acids.

Gene Family Analysis

To classify these TPSs in previously described TPS families, we constructed a phylogenetic tree based on amino acid sequences with TPSs from A. thaliana. The Costus TPS protein sequences, combined with those from A. thaliana, were aligned with MAFFT v7.508 (-maxiterate 1000, using L-INS-I algorithm) (Katoh and Standley 2013). We then trimmed the alignment using trimAl (-gt0.6, sites only included when present in 60% of sequences) (Capella-Gutierrez et al. 2009). We used this alignment to construct a gene tree with IQ-TREE and the ModelFinder function to determine the best-fit model (Nguyen et al. 2015; Kalyaanamoorthy et al. 2017; Hoang et al. 2018). We rooted the tree using the phytools package in R (“midpoint.root” function) (Revell 2012; R Core Team 2023). We plotted the resulting gene trees using the following packages in R: ape (Paradis and Schliep 2019), evobiR (Blackmon and Adams 2015), ggnewscale (Campitelli 2023), ggstar (Xu 2022), ggtree (Yu et al. 2017), ggtreeExtra (Xu et al. 2021), and tidytree (Yu 2022).

To investigate patterns of gene family evolution, we used computational analysis of gene family evolution CAFE (v5.1), which uses a birth–death model of gene family evolution to model gene gain and loss across a species tree (Mendes et al. 2020). The required input files are a species tree which we adapted from (Vargas et al. 2020), and gene counts for each species in each clade (as identified in Fig. 4).

We also took an alternative approach by using reconcILS to reconcile our gene tree and species tree to estimate gene duplications and losses for each species (Mishra et al. 2024). The algorithm considers not only duplication and loss events but also ILS. To create a gene tree, we aligned all Costus TPS nucleotide sequences with MAFFT v7.508 (-maxiterate 1000, using L-INS-I algorithm) (Katoh and Standley 2013). The alignment was trimmed using trimAl (-gt0.6, sites only included when present in 60% of sequences) (Capella-Gutierrez et al. 2009) and used to construct a gene tree with IQ-TREE and the ModelFinder function to determine the best-fit model (Nguyen et al. 2015; Kalyaanamoorthy et al. 2017; Hoang et al. 2018). We rooted the tree using the phytools package in R (“midpoint.root” function) (Revell 2012; R Core Team 2023). The species tree was adapted from (Vargas et al. 2020). We used the default parameter settings.

Differential Expression Analysis

First, we trimmed the reads using TrimGalore! (Martin 2011). We then mapped the reads from all three species to both the C. allenii and C. lasius genome (Harenčár et al. 2023). We ran 2pass mapping using STAR (Dobin et al. 2013). In general, we found higher levels of multimapped C. allenii reads when mapped to C. allenii. To investigate this, we used mmquant (with merge, -m, set to 10) to count the reads, a tool that can account for multimapping reads (Zytnicki 2017). The output from mmquant was read into R for further analysis. To calculate the percentage of gene expression due to TPS expression, we normalized reads by library size using TMM (Robinson and Oshlack 2010). Due to the consistency in TPS expression as calculated by mapping to both genomes, we decided to use the C. lasius genome assembly to identify genes differentially expressed between the three species to remove the complication of multimapping. To do this, we first counted reads using featureCounts (Liao et al. 2014) and again read the output into R. We filtered genes to include those with at least one count per million in at least one library and normalized reads by library size using TMM. Differential gene expression was evaluated using the nonparametric “noiseqbio” function from the NOISeq package (Tarazona et al. 2011, 2015). We used the default cut-off of q = 0.95 in the “degenes” function, which is equivalent to an adjusted P-value of 0.05.

Plotting and Data Manipulation

Additional packages used for general plotting: cowplot (Wilke 2020), ggplot2 (Wickham 2009), ggtree (Yu et al. 2017), gplots (Warnes et al. 2024), scico (Pedersen and Crameri 2023), and tidytree (Yu 2022). Packages used for data transformation and manipulation include data.table (Barrett et al. 2024), dplyr (Wickham et al. 2021), knitr (Xie 2023), magrittr (Bache and Wickham 2023) readr (Wickham et al. 2023), reshape2 (Wickham 2007), and tibble (Müller and Wickham 2022). Analyses were carried out in R version 4.3.2 (R Core Team 2023).

Supplementary Material

Supplementary material is available at Molecular Biology and Evolution online.

Author Contributions

Conceptualization: K.D., K.M.K., and S.R.R. Investigation: K.D. and K.M.K. Formal analysis: K.D. Visualization: K.D. Resources: K.M.K. and S.R.R. Writing—original draft: K.D. Writing—review and editing: K.D., K.M.K., and S.R.R. Supervision: K.M.K. and S.R.R. Project administration: K.M.K. and S.R.R. Funding acquisition: K.M.K. and S.R.R.

Funding

This work was funded by the National Science Foundation by a Dimensions of Biodiversity grant awarded to K.M.K. (DEB-1737889) and S.R.R. (DEB-1737771) and the Jean H. Langenheim Chair in Plant Ecology and Evolution held by K.M.K.

Data Availability

Data and R scripts used for analysis are available from Open Science Framework: https://osf.io/2ap4k/. The raw sequence reads and final genome assembly for Costus allenii are deposited in the SRA (BioProject PRJNA1091680). The raw RNA-seq data are deposited in the SRA (BioProject PRJNA1010186).