Anti-predator defences are typically regarded as static signals that conceal prey or advertise their unprofitability. However, startle displays are performed by prey when attacked and can include a spectacular array of movements, colours, and sounds. Here we present the first phylogenetically-controlled comparative analyses of startle displays including behaviour, using praying mantises as a test case. For 58 species, with a dated phylogeny, we estimate the strength of phylogenetic signal in the presence and ‘complexity’ (number of display components) of displays and their components and test hypotheses on their evolutionary correlates including primary defence and body size. We report strong phylogenetic signal in display presence and complexity, and strong lability in behavioural, but not morphological, components. Body size correlates with display presence and complexity independently of phylogeny, but not in phylogenetically-controlled analyses. Finally, species in species-rich clades are more likely to have a display, and a more complex one, suggesting support for ecological displacement via behavioural traits. To further elucidate the conditions under which startle display evolve, future work should include quantitative descriptions of display components, habitat type, and predator communities. Understanding the evolution of startle displays enriches our overall understanding of predator-prey dynamics and provides scaffolding for the development of new theory.

Startle display and primary defense data

Studies on mantis displays can be found in publications from several decades (Varley, 1939; Crane, 1952; Maldonado, 1970; Edmunds, 1972, 1976; Pita, 1972; Loxton, 1979; Grandcolas & Desutter-Grandcolas, 1998; O’Hanlon et al., 2018).To generate our dataset, we searched Web of Science and Google Scholar for all published descriptions of praying mantis startle displays. Search terms were: defensive display, deimatic, startle, antipredator, frightening attitude, praying mantis, praying mantid, and Mantodea. Six papers and the observations of scientists therein provided sufficiently detailed accounts of the displays or lack thereof. Nine species were scored based on personal observations of scientists currently working in the field of praying mantis behaviour (G. Holwell, K. Barry and J. O’Hanlon). In total, we gathered behavioural descriptions for 58 praying mantis species, with each species representing a different genus across the Mantodea. Our data therefore represent approximately 13% of mantis diversity at the generic level. We scored the description of the species’ primary defense, i.e. camouflage, as crypsis or masquerade, based on the descriptions of ‘general resemblance’ and ‘special resemblance’ respectively by Edmunds (Edmunds, 1972, 1976) and the overall shape of the mantis (e.g. ‘mantis shaped’, ‘dead leaf shaped’, etc.) in accordance with the conceptual framework outlined by Skelhorn et al (2010b).

We scored the presence of a display as whether the display behaviour had been witnessed first-hand by a scientist and/or a description published. We scored absence of display if the species had been explicitly reported to lack a display despite being exposed to the same stimuli as those reported to have a display (Edmunds, 1972, 1976).

For species with a display (N = 31), we scored the presence and absence of seven components, four behaviours and three colour patterns: (1) wings display – wings raised during the display, (2) arms display – arms raised during the display, (3) mouth display – if the mouth is opened during the display, (4) sound – display includes sound (e.g. rustling wings not related solely to raising the wings), (5) wing colours – wings have contrasting colour markings or ‘eyespots’ that are revealed or highlighted during the display, (6) arm colours – raptorial forelimbs have conspicuous colours or ‘eyespots’ that are revealed or highlighted during the display, (7) abdomen colours – part of the abdomen revealed during the display has a contrasting colour. Display complexity was the unweighted sum of these traits with a maximum possible complexity score of seven.

Body size and shape data

To compile the dataset of mantis body size, we took size data from seven publications and in addition, directly measured 294 specimens from 49 species kept in the collections of the Cleveland Museum of Natural History and the National Museum of Natural History (Smithsonian Institution). We took three measurements of body size in both males and females for each species in the data set where specimens were available: body length, pronotum length, and forewing length (Supplementary Figure 1). We took an average for each measurement from a variety of sample numbers per species. From these measures we created four further variables: ‘flight capacity’ (forewing length divided by body length, larger values indicate larger wings relative to body length and thus greater likelihood of flight), ‘size dimorphism’ (female body length minus male body length, positive values indicate females larger than males), and ‘flight dimorphism’ (male flight capacity minus females flight capacity, positive values indicate male have larger wings per unit body size than females).

Mantis phylogeny and timetree estimation

A total of 94 praying mantis taxa were selected (Supplementary Table 1) from tree topologies recovered in prior studies [1–8]. Four outgroup taxa were selected from Blattodea [3]. The selected taxa provided broad taxonomic and clade specific representation that also allowed collection of attribute data used to test the hypotheses of this study. A total of 10 gene fragments were included, all of which have proven to be phylogenetically informative at multiple levels of the Mantodea phylogeny [1–8]. These genes are 12S rRNA, 16S rRNA, 18S rRNA, 28S rRNA, Cytochrome Oxidase I (COI), Cytochrome Oxidase II (COII), NADH dehydrogenase subunit 4 (ND4), Histone 2A (H2A), Histone 3 (H3), and Wingless (wnt-1) (see Supplementary Table 1 for GenBank accession numbers).

For all included taxa, but one with newly sampled data, sequences were downloaded from GenBank (NCBI) and imported into Geneious v7.1.4 for 10 loci (Supplementary Table 1). The new sequence data (Negromantis sp.) was generated in the Cleveland Museum of Natural History DNA Laboratory. Thoracic and coxal muscle tissue was excised from specimens and extracted using the Qiagen DNeasy protocol for animal tissue. Specimen and DNA vouchers and template are deposited in the Department of Invertebrate Zoology at the Cleveland Museum of Natural History (CMNH). New sequence data from four gene loci were generated previously, but not published until this study (GenBank accession numbers in Supplementary Table 1), using well-established protocols for amplification, gel verification, amplicon purification, and sequencing [2,3,7]. Gene regions were sequenced with complements and sufficient overlap with adjacent regions to ensure the accuracy of sequence data.

Gene fragments were aligned in Geneious v7.1.4 using MAFFT v7 [9] and reading frame was determined for coding genes. We used SequenceMatrix v1.7.8 [10] to concatenate the matrix for a dataset including 10207 characters. We used PartitionFinder v2.1.1 [11,12] to determine partitioning strategy and models. A total of 22 partitions were input based on gene and codon position with the BIC and greedy settings for model selection and search strategy, respectively. Four partitions were recovered, each using the GTR+I+G model: 1) 12S, 16S, ND4_pos1, ND4_pos3, 2) 18S, 28S, H2A_pos1, H2A_pos2, H2A_pos3, H3_pos1, H3_pos2, H3_pos3, wnt_pos1, wnt_pos2, wnt_pos3, 3) COII_pos3, COI_pos3, and 4) COII_pos1, COII_pos2, COI_pos1, COI_pos2, ND4_pos2.

We conducted a mixed-model maximum likelihood (ML) analysis using RAxML v8 [13] with partitions corresponding to PartitionFinder results. One thousand non-parametric bootstrap (BS) pseudoreplicates were performed under a GTRGAMMA. We also conducted a mixed-model MrBayes v3.2.3 [14] analysis using PartitionFinder results by implementing four independent runs (four chains each) for 40 million generations. Each Bayesian run was started from a random tree and subsequently monitored for convergence using the program Tracer v1.7.1 [15]. Specifically, the plots for each run for sampled generations (every 1000) were compared using mean likelihoods, standard deviations, and distribution plots to ensure they converged on the same distribution after the burn-in. Bayesian analyses were performed on the Cyberinfrastructure for Phylogenetic Research (CIPRES) Science Gateway[16]Sampled trees were used to calculate a 50% majority rule tree to determine posterior probabilities (PP) [17]. FigTree v1.4.4 [18] was used to visualize topologies and produce figures for both ML and Bayesian analyses.

We used the Bayesian output tree as a start tree to estimate divergence times using a lognormal uncorrelated relaxed-clock model of among-lineage rate variation in BEAST v1.8.3 [19]. Like the Bayesian analysis, we utilized the CIPRES Science Gateway to run BEAST. We applied an exact root height with the mean set at 197 and the standard deviation set to 20 based on root dates recovered consistently in prior studies, which used significantly expanded taxon sampling or datasets along with multiple fossil calibrations [3,20,21]. We chose this method of calibration because of the lack of an adequate fossil record capable of calibrating later diverging nodes in the phylogeny [22–24] and it simply recovered a consistent phylogenetic tree scaled to time (timetree) that can address our evolutionary questions rather than addressing fundamental questions about the temporal origins of lineages [8]. Based on PartitionFinder results, we assigned all four partitions the GTR+I+G model. We executed two independent runs using the Yule process [25] for the tree prior. Trees were sampled every 1,000 generations over 10 million generations. Tracer was used to monitor convergence across runs. We used LogCombiner v1.8.3 to process log and tree files. TreeAnnotator v1.8.3 was used to produce a maximum clade credibility tree (25% burn-in) with median node heights with upper and lower confidence-interval (CI) values.