Multiple drivers and lineage-specific insect extinctions during the Permo–Triassic
Fossil record of insectsWe compiled all species-level fossil occurrences of insects using https://paleobiodb.org/ (PBDB) as a starting point (downloaded October 12, 2021). The dataset obtained from PBDB contained initially 5808 occurrences for a period ranging from the Asselian to the Rhaetian. The dataset was cleaned of synonyms, outdated combinations, nomina dubia, and other erroneous and doubtful records, based on revisions provided in the literature and/or on the expertise of the authors. After correction, including data addition from the literature, our dataset was composed of 3636 species (1784 genera, and 418 families) for 17,250 occurrences resulting from an in-depth study and curation of the entire bibliography of fossil insects, spanning from the Asselian (lowermost Permian) to the Rhaetian (uppermost Triassic). Although most of the taxa included in the datasets are nominal taxa (published and named), a few unnamed taxa (genera or species) that are considered separate from others were also included, although not formally named in the literature or not published yet. These unpublished taxa are identifiable by the notation ‘fam. nov.’ or ‘gen. nov.’ following their names.Occurrences used here are specimens originating from a given stratigraphic horizon assigned to a given taxon. The age of each occurrence is based on data from PBDB, corrected with a more precise age (generally stage, sometimes substage), and the age of each time bin boundaries relies on the stratigraphic framework proposed in the International Chronostratigraphic Chart (updated to correspond with the ICS 2022/0295). Similarly, the ages of some species assigned to the wrong stage were corrected. In fact, some species from the French Permian deposit of Lodève were initially considered to be of Artinskian age in PBDB but most species from this deposit originate from the Merifons member, which is of Kungurian age96.Our data compilation allows a robust integration of data before and after our period of interest (i.e. the lower Permian and all geologic stages after the Carnian) to encompass occurrences of genera that may survive until the Late Triassic and to generate a sufficient background for the model to correctly estimate the extinction events around the P/T boundary. Since we used different datasets, the differences between genus-level or family-level occurrence numbers are explained by the systematic placement of some specimens that can only be placed confidently in a family but not in a genus (Supplementary Table 1). Tentative species identifications originally placed with uncertainty (reported as ‘aff.’ or ‘?’) were always included at a higher taxonomic level. Uncertain generic attributions were integrated as occurrences at the family level (e.g. a fossil initially considered Tupus? is recorded as an occurrence of Meganeuridae). Our total dataset was subdivided into smaller datasets, which represent orders or other subclades of insects (e.g. Mecoptera, Holometabola and Polyneoptera). Note that all the ichnospecies—a species name assigned to trace fossils (e.g. resting trace, nest and leaf damage)— and insect eggs (e.g. Clavapartus latus, Furcapartus exilis and Monilipartus tenuis) were not included in the analyses97. To prevent potential issues regarding the diversification estimates for clades with poor delineation, we refrained from analysing several orders that serve as taxonomic ‘wastebaskets’ (e.g. Grylloblattodea). These groups are poorly defined, likely polyphyletic or paraphyletic, and not supported by apomorphic characters—e.g. the monophyly of the ‘Grylloblattodea’ (Grylloblattida Walker, 1914 plus numerous fossil families and genera of uncertain affinities) is not supported by any synapomorphy, nor the relationships within this group. The occurrences assigned to these orders were rather included in analyses conducted at a higher taxonomic level (at the Polyneoptera level in the case of the ‘Grylloblattodea’). The detail of the composition of all the datasets is given in Supplementary Table 14, and each dataset is available in Supplementary Data 1.Studying extinction should, when possible, rely on species-level diversity to better circumscribe extinction events at this taxonomic rank, which is primarily affected by extinction98,99,100. However, in palaeoentomology, species-level occurrence data may contain less information than genus-level data, mainly because species are most of the time only known from one deposit, resulting in reduced life span, and are also sometimes poorly defined. Insects are also less prone to long-lasting genera or species than other lineages, maybe because of the relatively short time between generations (allowing for rapid evolution) or because morphological characters are better preserved or more diagnostic than in other lineages (i.e. wing venation), allowing easier differentiation. Another argument for the use of genus-level datasets is the possibility to add occurrences represented by fossils that cannot be assigned at the species level because of poor preservation or an insufficient number of specimens/available characters. By extension, the genus life span provides clues as to survivor taxa and times of origination during periods of post-extinction or recovery. A genus encompassing extinction events indicates that at least one species of this genus crossed the extinction. To get the best signal and infer a robust pattern of insect dynamics around the P/T events, we have chosen to analyse our dataset at different taxonomic ranks (e.g. genus, family and order levels) to extract as much evidence as possible.To further support our choice to work at these different levels, most recent works aiming to decipher the diversification and extinction in insect lineages have worked using a combination of analyses21,22,26; this also applies to non-insect clades51,101,102. This multi-level approach should maximise our understanding of the Permo–Triassic events.Assessing optimal parameters and preliminary testsPrior to choosing the settings for the final analyses (see detail in Dynamics of origination and extinction), a series of tests were carried out to better evaluate the convergence of our analyses. First, we analysed our genus-level dataset with PyRate36 running for 10 million generations and sampling every 10,000 generations, on ten randomly replicated datasets using the reversible-jump Markov Chain Monte Carlo (RJMCMC) model37 and the parameters of PyRate set by default. As the convergence was too low, new settings were used, notably increasing the number of generations to 50 million generations and monitoring the MCMC mixing and effective sample size (ESS) each 10 million generations. We modified the minimal interval between two shifts (-min_dt option, testing 0.5, 1.5 and 2), and found no major difference in diversification patterns between our tests. We have opted for 50 million generations with a predefined time frame set for bins corresponding to the Permian and Triassic stages, and a minimum interval between two shifts of two Ma. These parameters allow for maintaining a short bin frame and high convergence values while correctly identifying periods of diversification and extinction. For each analysis, ten datasets were generated using the extract.ages function to randomly resample the age of fossil occurrences within their respective temporal ranges (i.e. resampled ages are randomly drawn between the minimum and the maximum ages of the geological stratum). We monitored chain mixing and ESS by examining the log files in Tracer 1.7.1103 after excluding the first 10% of the samples as a burn-in period. The parameters are considered convergent when their ESS are greater than 200.Dynamics of origination and extinctionWe carried out the analyses of the fossil datasets based on the Bayesian framework implemented in the programme PyRate36. We analysed the fossil datasets under two models: the birth–death model with constrained shifts (BDCS38) and the RJMCMC (-A 4 option37). These models allow for a simultaneous estimate for each taxon: (1) the parameters of the preservation process (Supplementary Fig. 17), (2) the times of origination (Ts) and extinction (Te) of each taxon, (3) the origination and extinction rates and their variation through time for each stage and (4) the number and magnitude of shifts in origination and extinction rates.All analyses were set with the best-fit preservation process after comparing (-PPmodeltest option) the homogeneous Poisson process (-mHPP option), the non-homogeneous Poisson process (default option), and the time-variable Poisson process (-qShift option). The preservation process infers the individual origination and extinction times of each taxon based on all fossil occurrences and on an estimated preservation rate, denoted q, expressed as expected occurrences per taxon per Ma. The time-variable Poisson process assumes that preservation rates are constant within a predefined time frame but may vary over time (here, set for bins corresponding to stages). This model is thus appropriate when rates over time are heterogeneous.We ran PyRate for 50 million MCMC generations and a sampling every 50,000 generations for the BDCS and RJMCMC models with time bins corresponding to Permian and Triassic stages (-fixShift option). All analyses were set with a time-variable Poisson process (-qShift option) of preservation and accounted for varying preservation rates across taxa using the Gamma model (-mG option), that is, with gamma-distributed rate heterogeneity with four rate categories36. As explained above, the minimal interval between two shifts (-min_dt option) was modified and a value of 2 was used. The default prior to the vector of preservation rates is a single gamma distribution with shape = 1.5 and rate = 1.5. We reduced the subjectivity of this parameter, and favoured a better adequation to the data, allowing PyRate to estimate the rate parameter of the prior from the data by setting the rate parameter to 0 (-pP option). Therefore, PyRate assigns a vague exponential hyper-prior to the rate and samples the rate along with all other model parameters. Similarly, because our dataset does not encompass the entire fossil record of insects, we assumed that a possible edge effect may interfere with our analyses, with a strong diversification during the lowermost Permian and, conversely a strong extinction during the uppermost Triassic. Because the RJMCMC and BDCS algorithms look for rate shifts, we constrained the algorithm to only search for shifts (-edgeShift option) within the following time range 295.0 to 204.5 Ma. We monitored chain mixing and ESS by examining the log files in Tracer 1.7.1103 after excluding the first 10% of the samples as a burn-in period. The parameters are considered convergent when their ESS are greater than 200.We then combined the posterior estimates of the origination and extinction rates across all replicates to generate rates through-time plots (origination, extinction, and net diversification). Shifts of diversification were considered significant when log Bayes factors were >6 in the RJMCMC model, while we considered shifts to be significant in the BDCS model when mean rates in a time bin did not overlap with the 95% credibility interval (CI) of the rates of adjacent time bins.We replicated all the analyses on ten randomly generated datasets of each clade and calculated estimates of the Ts and the Te as the average of the posterior samples from each replicate. Thus, we obtained ten posterior estimates of the Ts and Te for all taxa and we used these values to estimate the past diversity dynamics by calculating the number of living taxa at each time point. For all the subsequent analyses, we used the estimated Ts and Te of all taxa to test whether or not the origination and the extinction rate dynamics were correlated with particular abiotic factors, as suggested by the drastic changes in environmental conditions known during the Permo–Triassic. We used proxies for abiotic factors, such as global continental fragmentation or the dynamic of major clades of plants, and for biotic factors via species interaction within and between ecological guilds. This approach avoids re-modelling preservation and re-estimating times of origination and extinction, which reduces drastically the computational burden, while still allowing to account for the preservation process and the uncertainties associated with fossil ages. Similarly, the times of origination and extinction used in all the subsequent analyses were obtained while accounting for the heterogeneity of preservation, origination and extinction rates. To discuss the magnitude of the periods of extinction and diversification, we compared the magnitude of these events to the background origination and extinction rates (i.e. not during extinction or diversification peaks).The PyRate approach has proven to be robust following a series of tests and simulations that reflect commonly observed biases when modelling past diversity dynamics31,38. These simulations were based on datasets simulated under a range of potential biases (i.e. violations of the sampling assumptions, variable preservation rates, and incomplete taxon sampling) and reflecting the limitations of the fossil record. Simulation results showed that PyRate is able to correctly estimate the dynamics of origination and extinction rates, including sudden rate changes and mass extinction, even if the preservation levels are low (down to 1–3 fossil occurrences per species on average), the taxon sampling is partial (up to 80% missing) or if the datasets have a high proportion of singletons (exceeding 30% of the taxa in some cases). The strongest bias in birth–death rate estimates is caused by incomplete data (i.e. missing lineages) altering the distribution of taxa; a pervasive effect often mentioned for phylogeny-based models104,105,106. However, in the case of PyRate, the simulations confirm the absence of consistent biases due to an incomplete fossil record36. Finally, the recently implemented RJMCMC model was shown to be very accurate for estimating origination and extinction rates (i.e. more accurate than the BDCS model, the boundary-crossing and three-time methods) and is able to recover sudden extinction events regardless of the biases in the fossil dataset37.The severity of extinctions and survivorsFor each event—the Roadian–Wordian, the LPME, and the Ladinian–Carnian—we quantified the percentage of extinctions and survivors at the genus level. We used the Te and Ts from our RJMCMC analysis and computed the mean for the Te (Tem) and for the Ts (Tsm) of each genus. We then filtered our dataset to keep only the genera with a Tsm older than the upper boundary of the focal event, i.e., we only kept the genera that appeared before the end of the event. Then, we discarded the genera that have disappeared before the lower boundary of the focal event, i.e. Tem older that the lower boundary of the event. The remaining genera, which corresponds to all the genera (total) present during the crisis (Ttgen), can be classified into two categories, ‘survivor genera’ (Sgen), i.e. those that survived the crisis, and those that died: ‘extinct genera’ (Egen). The survivors have a Tem younger than the upper boundary of the focal event, while the ‘extinct genera’ died out during the event and have a Tem between the lower and upper boundaries of the event of interest. To obtain the percentage of survivors, we used the following formula: (Sgen/Ttgen) × 100. Similarly, the percentage of extinction is calculated as: (Egen/Ttgen) × 100.Age-dependent extinction modelWe assessed the effect of taxon age on the extinction probability by fitting the age-dependent extinction (ADE; -ADE 1 option) model50. This model estimates the probability for a lineage to become extinct as a function of its age, also named longevity, which is the elapsed time since its origination. It is recommended to run the ADE model over time windows with roughly constant origination and extinction rates, as convergence is difficult—but not impossible—to reach in extinction or diversification contexts50. We ran PyRate for 50 million MCMC generations with a sampling every 50,000 generations, with a time-variable Poisson process of preservation (-qShift option), while accounting for varying preservation rates across taxa using the Gamma model (-mG option). We replicated the analyses on ten randomised datasets and combined the posterior estimates across all replicates. We estimated the shape (Φ) and scale (Ψ) parameters of the Weibull distribution, and the taxon longevity in a million years. According to ref. 50, there is no evidence of age-dependent extinction rates if Φ = 1. However, the extinction rate is higher for young species and decreases with species age if Φ 1. Although ADE models are prone to high error rates when origination and extinction rates increase or decrease through time, simulations with PyRate have shown that fossil-based inferences are robust50. We investigated the effect of ADE during three different periods (-filter option) as follows: (1) between 264.28 Ma and 255 Ma (pre-decline), (2) between 254.5 Ma and 251.5 Ma (decline) and (3) between 234 Ma and 212 Ma (post-crisis). We monitored chain mixing and ESS by examining the log files in Tracer 1.7.1103 and considered the convergence of parameters sufficient when their ESS were greater than 200.Selection of abiotic and biotic variablesTo test correlations of insect diversification with environmental changes, we examined the link between a series of environmental variables and origination/extinction rates over a period encompassing the GEE, the LPME and the CPE but also for each extinction event. We focused on the role of nine variables, also called proxies, which have been demonstrated or assumed to be linked to extinctions and changes in insect diversity26,67.The variations in the atmospheric CO2 and O2 concentrations are thought to be correlated with the diversification of several insect lineages, including the charismatic giant Meganeuridae65,66,67. Because the increase of O2 concentration has likely driven the diversification of some insects, its diminution may have resulted in the extinction or decline of some lineages. Therefore, we investigated the potential correlation of the variations of this variable with insect dynamics using data from ref. 55. We extracted the data, with 1-million-year time intervals, spanning the Permo–Triassic.Similarly, the modification of CO2 concentration, notably its increase, is known to promote speciation in some modern insect groups107. Therefore, a similar effect may have occurred during the Permian and Triassic but remains to be tested. We based our analyses on the dataset of ref. 108. We used their cleaned dataset and extracted all verified values for the Permo–Triassic interval. Because the initial data (i.e. independent estimates) were made in various locations for the same age, different values of the CO2 concentration are provided. We incorporated all these values in our analysis, allowing PyRate to search for a correlation for each value of the CO2 concentration. We obtain a final correlation independent of the sampling location, in line with our large-scale analysis.The continental fragmentation, as approximated by plate tectonic change over time, has recently been proposed as a driver of Plecoptera dynamics26. Because the period studied encompasses a major geological event, the fragmentation of the supercontinent Pangea, we investigated the effect of continental fragmentation on insect diversification dynamics. We retrieved the index of continental fragmentation developed by ref. 69 using paleogeographic reconstructions for 1-million-year time intervals. This index approaches 1 when all plates are disjoined (complete plate fragmentation) and approaches 0 when the continental aggregation is maximal.Climate change (variations in warming and cooling periods) is a probable driver of diversification changes over the history of insects21,109. Temperature is likely directly linked with insect dynamics109 but also with their food sources, notably plants110. Because it was demonstrated that modification of temperature impacted floral assemblages110, we tested the correlation between temperature variations and the diversification dynamic of insects. Major trends in global climate change through time are typically estimated from relative proportions of different oxygen isotopes (δ18O) in samples of benthic foraminiferan shells111. We used the data from ref. 112, converted to absolute temperatures following the methodology described in Condamine et al.113 (see their section Global temperature variations through time). The resulting temperature data reflects planetary-scale climatic trends, with time intervals inferior to 1-million-year, which can be expected to have led to temporally coordinated diversification changes in several clades rather than local or seasonal fluctuations.The fluctuation in relative diversity of gymnosperms, non-Polypodiales ferns, Polypodiales ferns, spore-plants, and later the rise of angiosperms has likely driven the diversification of numerous insects57,60,61,114. Close interactions between insects and plants are well-recorded during the Permian and Triassic57,60,61. In fact, herbivorous insects are known to experience high selection pressure from bottom-up forces, resulting from interactions with their hosts or feeding plants30,72. Therefore, it appears crucial to investigate the effect of these modifications on the insects’ past dynamics. We used the data from ref. 38 for the different plant lineages (all with 1-million-year time intervals). All the datasets for these variables are available in the publications cited aside from each variable or in Supplementary Data 1.Multivariate birth–death modelWe used the multivariate birth–death (MBD) model to assess to what extent biotic and abiotic factors can explain temporal variation in origination and extinction rates55. The model is described in ref. 55, where origination and extinction rates can change through time in relation to environmental variables so that origination and extinction rates depend on the temporal variations of each factor. The strength and sign (positive or negative) of the correlations are jointly estimated for each variable. The sign of the correlation parameters indicates the sign of the resulting correlation. When their value is estimated around zero, no correlation is estimated. An MCMC algorithm combined with a horseshoe prior, controlling for over-parameterisation and for the potential effects of multiple testing, jointly estimates the baseline origination (λ0) and extinction (µ0) rates and all correlation parameters (Gλ and Gµ)55. The horseshoe prior is used to discriminate which correlation parameters should be treated as noise (shrunk around 0) and which represent a true signal (i.e. significantly different from 0). In the MBD model, a correlation parameter is estimated to quantify independently the role of each variable on the origination and the extinction.We ran the MBD model using 20 (for short intervals) or 50 million MCMC generations and sampling every 20,000 or 50,000 to approximate the posterior distribution of all parameters (λ0, µ0, nine Gλ, nine Gµ and the shrinkage weights of each correlation parameter, ωG). The MBD analyses used the Ts and the Te derived from our previous analyses under the RJMCMC model. The results of the MBD analyses were summarised by calculating the posterior mean and 95% CI of all correlation parameters and the mean of the respective shrinkage weights (across ten replicates), as well as the mean and 95% CI of the baseline origination and extinction rates. We carried out six analyses, over: (1) the Permo–Triassic (between 298.9 and 201.3 Ma); (2) the Roadian–Wordian (R/W) boundary (between 270 and 265 Ma), (3) the LPME (between 254.5 and 250 Ma), (4) the Ladinian–Carnian (L/C) boundary (between 240 to 234 Ma), (5) the Permian period (between 298.9 and 251.902 Ma) and (6) the Triassic period (between 251.902 and 201.3 Ma). We monitored chain mixing and ESS by examining the log files in Tracer 1.7.1103 and considered the convergence of parameters sufficient when their ESS were greater than 200.Multiple clade diversity-dependence modelTo assess the potential effect of diversity-dependence on the diversity dynamics of three or four insect guilds, we used the multiple clade diversity-dependence (MCDD) model in which origination and extinction rates are correlated with the diversity trajectory of other clades31. This model postulates that competitive interactions linked with an increase in diversity results in decreasing origination rates and/or increasing extinction rates. The MCDD model allows for testing diversity-dependence between genera of a given clade or between genera of distinct clades sharing a similar ecology.We estimated the past diversity dynamics for three (i.e. herbivores, predators, and a guild composed of generalists + detritivores/fungivores dubbed ‘others’) or four insect groups or guilds (i.e. herbivores, predators, generalists and detritivores/fungivores) by calculating the number of living species at every point in time based on the times of origination (Ts) and extinction (Te) estimated under the RJMCMC model (see above) (Supplementary Figs. 19–24). We defined our four insect groups with a cautious approach i.e. insect genera, families or orders for which nothing is known about the ecology or about the ecology of their close relatives were not considered for the analysis. For example, no diet was assigned to Diptera, Mecoptera or Glosselytrodea. The ecology of the Triassic Diptera and Permo–Triassic Mecoptera is difficult to establish because extant Diptera and Mecoptera have a wide diversity of ecology. Fossil Mecoptera are also putatively involved in numerous interactions with plants (species with elongated mouthparts), suggesting a placement in the herbivore group, while other species were likely predators. Therefore, we cannot decide to which group each species belongs. Similarly, nothing is known about the body and mouthparts of the Glosselytrodea, most of the time described based on isolated wings; we did not assign the order to any group. The definition and delineation of insect clades have also challenged the placement of several orders (e.g. ‘Grylloblattodea’) in one of our four groups. The order ‘Grylloblattodea’ is poorly delineated and mostly serves as a taxonomic ‘wastebasket’ to which it is impossible to assign a particular ecology. Finally, genera, species, or families not placed in a higher clade (e.g. Meshemipteron, Perielytridae) were not included in the analysis. Oppositely, the guilds ‘herbivores’ and ‘predators’ are well defined, and their ecology is evidenced by the morphology of their representatives and the principle of actualism. For example, the ecology of Meganeurites gracilipes (Meganeuridae) has been deeply studied, and its enlarged compound eyes, its sturdy mandibles with acute teeth, its tarsi and tibiae bearing strong spines, and the presence of a pronounced thoracic skewness are specialisations today found in dragonflies that capture their prey while in flight115. All Odonatoptera are well-known predator insects. The raptorial forelegs of the representatives of the order Titanoptera and their mouthparts with strong mandibles are linked with predatory habits81. The Palaeodictyopteroidea were herbivorous insects with long, beak-like, piercing mouthparts, and probably a sucking organ81,82. Most Hemiptera are confidently considered herbivorous insects by comparison with their extant representatives. For example, the Cicadomorpha or Sternorrhyncha are known to feed on plants and their fossil representatives likely possessed the same ecology because of similar morphologies116. Some hemipteran families (e.g. Nabidae) are predators and we cautiously distinguished herbivorous and carnivorous taxa among Hemiptera. The detail of the ecological assignations for the 1009 genera included in our analyses can be found in Supplementary Data 1 (Table MCCD).We calculated ten diversity trajectories from the ten replicated analyses under the RJMCMC model. The estimation of past species diversity might be biased by low preservation rates or taxonomic uncertainties. However, such trajectory curves are likely to provide a reasonably accurate representation of the past diversity changes in the studied clades, notably because the preservation during the Permian and Triassic period is relatively good for insects (i.e. no gaps).Our MCDD analyses comprise all the insect genera spanning from the lowermost Permian to the uppermost Triassic and were run and repeated on ten replicates (using the Te and Ts estimated under the RJMCMC model) with 50 million MCMC generations and a sampling frequency of 50,000. For each of the four insect groups, we computed the median and the 95% CI of the baseline origination and extinction rates (λi and µi), the within-group diversity-dependence parameters gλi and gµi, and the between-groups diversity dependence parameters gλij and gµij. The mean of the sampled diversity dependence parameters (e.g. gλij) was used as a measure of the intensity of the negative (if positive) or positive interactions (if negative) between each pair of groups. The interactions were considered significant when their median was different from 0 and the 95% CI did not overlap with 0. We monitored chain mixing and ESS by examining the log files in Tracer 1.7.1103 and considered the convergence of parameters sufficient when their ESS were greater than 200.We cross-validated the result of the MCDD model using the MBD model. The MBD model can be used to run a multiple clade diversity-dependence analysis by providing the diversity trajectories of insect guilds as a continuous variable. These data are directly generated by PyRate using the lineages-through-time generated by the RJMCMC analyses (-ltt option). We ran the MBD model using 50 million MCMC generations and sampling every 50,000 to approximate the posterior distribution of all parameters (λ0, µ0, four Gλ, four Gµ and the shrinkage weights of each correlation parameter, ωG). We carried out three analyses, over the period encompassing the three extinction events (between 275 and 230 Ma): (1) for herbivores; (2) for predators; and (3) for ‘others’. For each analysis, the lineages-through-time data of the two other guilds are used as continuous variables to investigate a diversity dependence effect. We monitored chain mixing and ESS by examining the log files in Tracer 1.7.1103 and considered the convergence of parameters sufficient when their ESS were greater than 200.Reporting summaryFurther information on research design is available in the Nature Portfolio Reporting Summary linked to this article. More
