Taxon sampling and morphological matrix

We included 24 fossils that we have recently described or revised (Table 1) [7,28,37], with the taxonomic placements at the time supported by character evidence only. Specimen numbers, provenances, epochs, and deposits are given in Table 1. All specimens described in this study are publicly deposited and accessible (exception: Glypta transversalis, the holotype of which we could not locate; we instead studied the drawing in the original description). No permits were required for the described study, which complied with all relevant regulations. Specimens from the Green River Formation were obtained from the Paleobiology Department of the Smithsonian National Museum of Natural History, USA (specimen numbers starting with ‘USNM’), the Paleontology Section of the University of Colorado Museum of Natural History, USA (‘UCM’), and the Museum of Comparative Zoology, Harvard University, USA (‘PALE’). Fossils from the Messel Pit were studied at the Senckenberg Research Station Messel, Senckenberg Research Institute Frankfurt, Germany (‘SF MeI’). Finally, the specimens from the Tadushi Formation are deposited at the Arthropoda Laboratory, Paleontological Institute, Russian Academy of Sciences, Moscow, Russia (‘PIN’).

Download:

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

Table 1. Fossil taxa examined.

https://doi.org/10.1371/journal.pone.0212942.t001

For most species, we studied the holotype and sometimes paratype directly, but in a few cases (listed in [7]), we only obtained high-resolution photographs of the holotypes. Morphology of all fossils was studied in detail and characters were coded into a morphological matrix (S1 File). Morphological nomenclature follows Townes [25], except for wing venation characters, which are based on Kopylov [36] with few modifications (Fig 1).

Download:

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

Fig 1. Wing vein and forewing cell nomenclature used in the current study.

https://doi.org/10.1371/journal.pone.0212942.g001

The morphological matrix consists of 222 morphological characters of the adult insects scored for 103 extant and 24 fossil taxa. To this day, it represents the largest matrix of this kind composed for ichneumonids (S2 File). We aimed to get a good representation of the ichneumonid subfamilies (24 out of 42) while focusing on Pimpliformes, where we included 80 extant genera. In cases of heterogeneous and potentially non-monophyletic genera, we included representatives of several species groups. The full list of included taxa with specimen numbers and collection details is given in S1 Table. Our matrix is partly based on the morphological matrix from Gauld et al. [39] (~50% of the characters), which had been designed to capture the morphological diversity within a single ichneumonid subfamily, the Pimplinae. It was complemented with characters from Bennett et al. (unpublished) (~15% of the characters), which had a broad but sparse taxon sampling across ichneumonids. In order to cover most of the morphological diversity of the taxa scored in this study, we have re-defined and expanded many of the character definitions, and added specific characters for previously underrepresented subfamilies. To maximise the morphological information coming from fossils, we added some characters that are not traditionally used in ichneumonid taxonomy but are often well (and sometimes exclusively) preserved in fossils, such as wing venation characters. The full list of characters and states including detailed descriptions is given in S2 File.

Out of the 222 morphological characters, a majority are multistate characters, while 34% are binary. The 13 continuous characters represent ratios of length measurements designed to capture the shape of structures (except of the fore wing length which was included as a measure of size). In order to be able to analyse them along with the remaining characters we eventually transformed them to ordered six-state discrete characters following the gap weighting approach as defined in Thiele [40]. Six states is currently the upper limit for ordered characters in the used phylogenetic software (see below). We took special considerations when scoring fossils (but also extant taxa) and coded uncertain interpretations as polymorphisms; initially, many questionable coding were noted with “?” (e.g., “1?”), and preliminary analyses with either keeping the states (“liberal approach”) or keeping the question marks (“conservative”) resulted in better support values when the liberal approach was applied, while not creating any supported conflict in tree topology compared to the conservative approach. We thus decided to continue with the more liberal interpretation.

Bayesian phylogenetic analyses

Phylogenetic analyses of the morphological matrix were conducted in MrBayes 3.2 [41]. To model the evolution of the morphological data, we used the approach described in the seminal work by Lewis [15]. In its most basic version, the ‘Mk model’ assumes that all states of a multi-state character are equally likely, as are the respective transition rates between them (we refer to this as the ‘unordered Mk’ model). An extension of this model, which was adopted from parsimony approaches, allows considering some of the characters as ‘ordered’, i.e., allows transitions only between adjacent states [42]. In a character with three states represented by labels ‘0’, ‘1’ and ‘2’, only the transitions between ‘0’ and ‘1’ and between ‘1’ and ‘2’ are allowed in an ordered character, but never directly between ‘0’ and ‘2’. As an additional extension, one could think of a model that account for unobserved states, i.e., states that are possible but that are not found in any of the taxa included in a particular analysis. This might be a reasonable approach when a morphological matrix is reduced to include only part of the taxa; in such a situation, we know of possible states that a character can adopt, even though they are not realized among the taxa in our matrix. We call this the ‘full-state Mk’ model. To account for different characters evolving at different rates, we in all three of these cases included gamma-distributed among-character rate variation (ACRV) in the same way it is done for molecular data (among-site rate variation or ASRV, [43]), with the alpha parameter governing the shape of the distribution and thus extent of ACRV estimated from the data. To assess the impact of ignoring ACRV, in the case of the ordered model, we also tried running without ACRV.

We tested the four models described above on our morphological matrix by estimating their marginal likelihoods under stepping-stone sampling [44] in MrBayes. For the ‘unordered Mk’ analysis, we made sure that the highest number used as a state label for each character corresponded to the number of states minus one (as ‘0’ is also used as a label). For the ‘ordered Mk’ analysis, we chose a priori those characters that can be considered to evolve in an ordered fashion; this included both characters that were originally continuous and transformed by us into discrete states (see above) and other characters that have been conceptualized in a way that suggests that they would likely adhere to a step-wise pattern of evolution (58 characters in total, see S2 File). For the ‘full-state Mk’ model, we obtained the maximum number of realized states in each character from a larger matrix at hand, which includes about twice as many ichneumonid taxa. Our data was then re-coded so that the largest state label reflects the thus estimated true size of the state space. This procedure changed the state space in 52 out of the 222 characters, in most cases only by a single state difference, but sometimes by up to three states. The R scripts to perform these matrix manipulations and the resulting morphological matrices are available as S3 File, and can be downloaded as an R package from Github (‘devtools::install_github(‘seraklop/RoguePlots/rogue.plot‘). For the best-scoring model (the ‘ordered’ model), we also ran stepping-stone sampling with equal rates among sites instead of gamma-distributed ACRV (‘no ACRV’). Stepping-stone sampling was conducted by running 50 million generations each in four independent runs under all four models, with one million generations set as an initial burn-in and then 49 steps from posterior to prior [44]. Each step consisted of one million generations, half of which were discarded as a within-step burn-in. The alpha value, which determines the skewness of the sampling distribution used for the stepping-stone sampling, was set to 0.4 (the default value; not to be confused with the alpha parameter of the ACRV, which was estimated from the data).

The preferred model as identified by the marginal likelihoods was then used for a Bayesian analysis with four independent runs, each of which with one cold and three heated chains, for 50 million generations. The ‘variable’ coding bias was invoked, as no constant characters have been scored for our dataset, but some characters showed autapomorphic states. We assessed convergence by examining plots of log likelihoods over time, ascertaining that the potential scale reduction factor (PSRF) for all scalar parameters of the model had dropped below 1.02, and recording the average standard deviation of split frequencies (ASDSF) between the four independent runs. ASDSF was below 0.015 after 50 million generations; given the large phylogenetic uncertainty (and thus large parameter space to cover by the MCMC) introduced by the incompletely scored fossils, we consider this as acceptable topology convergence. Half of the generations were then excluded as a conservative burn-in. All calculations were performed on UBELIX (http://www.id.unibe.ch/hpc), the HPC cluster at the University of Bern, Switzerland.

RoguePlots to illustrate taxon placement uncertainty

To illustrate the uncertainty in the placement of fossil taxa, we developed R scripts to generate what we call ‘RoguePlots’. These plots show the posterior probabilities of the placement of a fossil on a partially resolved phylogeny, such as a consensus tree. Especially when there is a lot of phylogenetic uncertainty in a dataset, which is rather the rule than the exception in morphological data, fossils might with a non-negligible probability attach to branches that are not represented in a consensus tree (which can happen even if it is fully resolved, unless it is a strict consensus). Fig 2 illustrates how RoguePlots solve this issue: in order to represent such cases and thus ascertain that the posterior probabilities of attachments shown on a tree sum to one, the branches that are parallel to the direction of a rectangular phylogram of a RoguePlot represent fossil placements on this very branch, while the perpendicular branches stand for attachments to any unrepresented branch of which the corresponding node is the most recent common ancestor (MRCA).

Download:

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

Fig 2. Illustration of the generation of a ‘RoguePlot’.

Panel (a) shows three different topologies (I–III) as they might be found during a Bayesian MCMC analysis or based on a set of bootstrapped matrices. The frequency with which each topology is encountered is shown above. Fossil X attaches to one of three different branches in each tree. Two of these branches are also present in the majority-rule consensus tree shown in panel (b), while the third branch, respectively, is in conflict between the trees and thus appears as a polytomy in the consensus. Fossil attachment frequencies to branches present in the consensus are reflected by the colour of the branches parallel to the direction of the tree, while those to branches not present are shown on the perpendicular branch that corresponds to its most recent common ancestor. In the current example, the percentages of Fossil X to attach to a branch in the crown group of genus C sums to 93%, while it is 7% for it being a stem-group representative and 0% to be with either of the other genera A and B. Classifying it as belonging to genus C is thus well-supported in this fictitious example. In addition to showing fossil placements, RoguePlots can be used to illustrate the attachment points of any taxon in a phylogeny, such as a species that was identified as a rogue.

https://doi.org/10.1371/journal.pone.0212942.g002

Based on the majority-rule consensus tree from the Bayesian analysis of the morphological matrix and on 1,000 trees sampled evenly from the post-burn-in phase of the four independent runs, we generated RoguePlots to illustrate phylogenetic placements for each of the 24 fossils covered here in order to re-evaluate their genus- and higher-level classifications.

Phylogenetic analyses

The morphological matrix has an overall coverage of 76.0% when counting polymorphisms as informative, or 71.6% otherwise, which corresponds to an average of 53 of the 222 characters missing per taxon. Among the extant taxa, coverage values are much higher (88.9% and 84.5%, respectively, depending on how polymorphisms are counted), while the fossil taxa as expected show much more missing data and thus lower coverage (20.6% and 15.9%, respectively). The fossil with the lowest number of characters, Eclytus lutatus Theobald, had 92.3% of characters missing (only 17 of 222 characters scored), while the most completely scored fossil was Scambus fossilobus Spasojevic et al. [37], with ‘only‘ 67.1% missing data (73 characters scored).

The comparison of the model likelihoods obtained from stepping-stone sampling revealed very large differences between the models, judging from Bayes factor comparisons (Table 2). While the variance in the marginal likelihoods between the four independent runs was always below six log units, the differences between the models were in the order of several hundreds. The ‘ordered Mk‘ model, which treated 58 characters as ordered (S2 File), very clearly outperformed the other models and was used for all further analyses. The ‘full-state Mk‘ model, which increased the state space of each character to that observed in other, not sampled ichneumonid taxa, reached the second-best marginal likelihood, followed by the ‘unordered Mk’ model. The model that used ordered characters but did not include among-character rate variation (‘no ACRV’), was clearly behind all other models, with a drop in likelihood of almost 850 log units (Table 2).

Download:

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

Table 2. Marginal likelihoods of the different models of morphological evolution as obtained under stepping-stone sampling.

https://doi.org/10.1371/journal.pone.0212942.t002

The majority-rule consensus tree of the analysis under the preferred model has a poorly resolved backbone, as can be expected from a morphology-only dataset, but recovers most of the subfamilies (Fig 3). When rooting with Xorides, as supported by a previous molecular analysis that included Xorides as the only represenative of Xoridinae [45], the remaining subfamilies of Ichneumonidae were recovered as a monophyletic group (or in other words, the two xoridine taxa included here grouped together). We thus rooted the tree between Xoridinae and the remaining subfamilies. Higher-level relationships among subfamily groups were not resolved, despite some of these groupings having been proposed on morphological grounds (Ophioniformes, Ichneumoniformes, Pimpliformes). However, they were based mostly on larval characters, which were not included in our matrix because of extensive missing data even among extant taxa. Some of the subfamilies were not recovered either (Tryphoninae, Cylloceriinae, and Diacritinae), but most of the other subfamilies and the four tribes within Pimplinae received reasonable support. Some groups were only recovered as paraphyletic, such as Campopleginae, Delomeristini and Theronini, and the acaenitine genus Coleocentrus clustered with Rhyssinae. The genera Xanthopimpla and Lissopimpla, for which a separate tribe or even subfamily has been discussed previously [45], render the Theronini paraphyletic in our tree, but on a rather long branch. Overall, the tree is largely congruent with a more sparsely sampled molecular tree published recently [45], even though the latter showed better resolution along the backbone. As the goal of the current study is to re-evaluate the classification of fossil species, which relies heavily on morphology, and combination with a molecular dataset would require more complex and time-consuming analyses, we at this point use the morphological tree only. It allows re-assessment of most subfamily and genus placements of fossils; however, because of the low resolution of relationships between subfamilies, its value is limited when it comes to placing taxa that are backbone representatives within Ichneumonidae.

Download:

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

Fig 3. Tree of extant taxa.

Majority-rule consensus tree of the topologies of the extant taxa sampled during the Bayesian analysis of a morphological dataset of 222 discrete characters. Values near nodes represent posterior probabilities of clades. Informal subfamily groups (Xoridiformes, Ophioniformes, etc.), the subfamily Pimplinae and the Polysphicta genus group (which includes parasitoids of spiders) are indicated by bars on the right. Tips are coloured according to their subfamily placement, and either subfamily or tribal (in the case of Pimplinae) classification is also indicated by the last three letters in the label.

https://doi.org/10.1371/journal.pone.0212942.g003

Phylogenetic placement of 24 fossil ichneumonids

The confidence in the phylogenetic placement of the fossils (Figs 4–9) ranged from very high, with up to 99.9% probability for attachment to a single branch (in Xanthopimpla praeclara, Fig 9D), to two or three competing placements that could be either in closely related (e.g., in Rhyssella vera, Fig 7D) or in unrelated groups (e.g., in Carinibus molestus, Fig 4A), to very low probabilities on any individual branch (e.g., Dolichomitus? saxeus, Fig 4D). Images of the studied fossils, along with partial RoguePlots that include all consensus tree branches that have a higher than 1% attachment probability, are shown in Figs 4–9 (complete RoguePlots are available as S4 File). We here briefly discuss the placement of the individual fossils and, in some cases, revise the genus- and/or subfamily assignments. The fossils are sorted alphabetically according to the genus name in the original description. Taxon names are followed by the percentage of characters scored for the individual fossil species.

Download:

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

Fig 4. Partial RoguePlots of four fossil species.

The plots include all consensus tree branches reaching more than 1% attachment probability. Xoridinae as outgroup and in some cases subfamilies in which the fossil has been placed previously are included regardless of attachment probabilities. Branches parallel to the direction of the phylogram are coloured according to the probability of direct attachment of the fossil to the branch, while perpendicular branches represent the sum on any branch not in the consensus tree, of which the corresponding node is the most recent common ancestor. The fossil drawings represent our interpretation of fossil morphology, with dotted line indicating uncertain interpretations. a) Carinibus molestus, holotype USNM 580881. b) Crusopimpla rediviva, holotype #2156. c) Crusopimpla tadushiensis, holotype #PIN 3364/277. d) Dolichomitus? saxeus, holotype #PIN 3364/31.

https://doi.org/10.1371/journal.pone.0212942.g004

Download:

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

Fig 5. Partial RoguePlots of four fossil species.

The plots include all consensus tree branches reaching more than 1% attachment probability. Xoridinae as outgroup and in some cases subfamilies in which the fossil has been placed previously are included regardless of attachment probabilities. Branches parallel to the direction of the phylogram are coloured according to the probability of direct attachment of the fossil to the branch, while perpendicular branches represent the sum on any branch not in the consensus tree, of which the corresponding node is the most recent common ancestor. The fossil drawings represent our interpretation of fossil morphology, with dotted line indicating uncertain interpretations. a) Eclytus? lutatus, holotype PALE-1418. b) Eopimpla grandis, holotype USNM 66581. c) Glypta transversalis, modified from original drawing (Plate X, Fig 25 in [47]). d) Ichninsum appendicrassum, holotype UCM 39378.

https://doi.org/10.1371/journal.pone.0212942.g005

Download:

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

Fig 6. Partial RoguePlots of four fossil species.

The plots include all consensus tree branches reaching more than 1% attachment probability. Xoridinae as outgroup is included regardless of attachment probabilities. Branches parallel to the direction of the phylogram are coloured according to the probability of direct attachment to the branch, while perpendicular branches represent the sum on any branch not in the consensus tree, of which the corresponding node is the most recent common ancestor. The fossil drawings represent our interpretation of fossil morphology, with dotted line indicating uncertain interpretations. a) Lithotorus cressoni, holotype PALE-4652. b) Mesoclistus? yamataroti, holotype UCM 62725. c) Mesornatus markovici, holotype SF MeI 15245. d) Phygadeuon? petrifactellus, holotypeUSNM 66580.

https://doi.org/10.1371/journal.pone.0212942.g006

Download:

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

Fig 7. Partial RoguePlots of four fossil species.

The plots include all consensus tree branches reaching more than 1% attachment probability. Xoridinae as outgroup is included regardless of attachment probabilities. Branches parallel to the direction of the phylogram are coloured according to the probability of direct attachment to the branch, while perpendicular branches represent the sum on any branch not in the consensus tree, of which the corresponding node is the most recent common ancestor. The fossil drawings represent our interpretation of fossil morphology, with dotted line indicating uncertain interpretations. a) Pimpla? eocenica, holotype USNM 66582. b) Plectiscidea? lanhami, holotype UCM 19167. c) Polyhelictes bipolarus, holotype MeI 16069. d) Rhyssella vera, holotype MeI 8814.

https://doi.org/10.1371/journal.pone.0212942.g007

Download:

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

Fig 8. Partial RoguePlots of four fossil species.

The plots include all consensus tree branches reaching more than 1% attachment probability. Xoridinae as outgroup is included regardless of attachment probabilities. Branches parallel to the direction of the phylogram are coloured according to the probability of direct attachment to the branch, while perpendicular branches represent the sum on any branch not in the consensus tree, of which the corresponding node is the most recent common ancestor. The fossil drawings represent our interpretation of fossil morphology, with dotted line indicating uncertain interpretations. a) Scambus fossilobus, holotype SF MeI 13431. b) Scambus? mandibularis, holotype USNM 501474. c) Scambus? parachuti, holotype 565885. d) Tilgidopsis haesitans, holotype USNM 66931.

https://doi.org/10.1371/journal.pone.0212942.g008

Download:

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

Fig 9. Partial RoguePlots of four fossil species.

The plots include all consensus tree branches reaching more than 1% attachment probability. Xoridinae as outgroup is included regardless of attachment probabilities. Branches parallel to the direction of the phylogram are coloured according to the probability of direct attachment to the branch, while perpendicular branches represent the sum on any branch not in the consensus tree, of which the corresponding node is the most recent common ancestor. The fossil drawings represent our interpretation of fossil morphology, with dotted line indicating uncertain interpretations. a) Trigonator macrocheirus, holotype SF MeI 17304. b) Tryphon amasidis, holotype UCM 15690. c) Xanthopimpla messelensis, holotype SF MeI 16988. d) Xanthopimpla praeclara, holotype SF MeI 17300.

https://doi.org/10.1371/journal.pone.0212942.g009

Carinibus molestus Spasojevic et al. [7]– 22%, Fig 4A. This taxon was described in its own genus with uncertain subfamily placement. We can confirm this notion but can narrow down potential subfamilies: The most likely placement in our analysis is with the acaenitine Mesoclistus (42%), and the total probability for it to be an acaenitine sums to 55%. But there is an alternative placement in the subfamily Metopiinae (total 32%), with a probability so high that it cannot be ignored. Given the unique combination of characters described by the authors [7], the genus might well represent an ancestor or extinct subfamily; we thus leave the genus in incertae subfamiliae.

Crusopimpla? rediviva (Brues) [27,28]– 18%, Fig 4B. This species attaches to various places in the tree, as does the type species of the genus, C. tadushiensis (see next fossil), with the exception that a placement as one of the more ancestral lineages within Pimplinae received some support (22%). The placement on the stem branch of Orthocentrinae was recovered with higher probability, however (25%), but we believe that this reflects an artefact due to model misspecification. The genus Crusopimpla was erected as a stem-lineage representative of the subfamily Pimplinae, with the presence of several carinae on the propodeum as a defining characteristic [28]. None of the extant pimplines has complete carination, which might explain why such a placement was never chosen. However, propodeal carination might evolve in a very non-stationary manner in Ichneumonidae: the earliest branching subfamily, Xoridinae, has complete carination of the propodeum, and the character distribution on the phylogeny [28] suggests that this carination has progressively been reduced in parallel in several of the more derived ichneumonid subfamilies, including Pimplinae. The use of a directional model of character evolution, such as the one used to study the reduction in hymenopteran wing venation and muscles [13], might have led to a different result here, but this model is not currently available for multistate data. Given these considerations, we leave this taxon in its current position.

Crusopimpla tadushiensis Kopylov et al. [28]– 17%, Fig 4C. This species appears all over the tree in our analysis, with placements in seven different subfamilies and on the ancestral branch of Ichneumonidae minus Xoridinae, but never in the subfamily Pimplinae in which it was originally described. Given the possibility that model misspecification caused this difference in interpretation (see previous taxon), and the fact that the preferred placement in Orthocentrinae (22%) seems highly unlikely given the invariably smaller size of its representatives, we refrain from suggesting any changes to its classification.

Dolichomitus? saxeus Kopylov et al. [28]– 14%, Fig 4D. The placement of this taxon was highly uncertain, which might be expected given its poor preservation, but there seems a clear tendency for it to take up a basal position. This might reflect a true ancestral placement of the fossil or could be due to ‘stem-ward slippage’ [46], the phenomenon that incompletely preserved fossils tend to attach closer to the root than would be accurate. The sparse character evidence precludes any clear assessment of this question, and as a placement within Pimplinae could not be ruled out either (17%), and genus assignment was already marked with a question mark by the authors of the original description, we leave it in its current taxonomic position.

Eclytus? lutatus Scudder [47]– 8%, Fig 5A. Even though it was described in a tryphonine genus, this fossil received very high support in our analysis to belong to Orthocentrinae (95.2%), and there-in to the Orthocentrus-group of genera (85.3%). All placements on other branches in the tree, including anywhere near the three Tryphoninae taxa we included, had less than 1% frequency. The fossil is not very well preserved and its placement difficult, as pointed out previously [7,48]. Furthermore, we did not manage to include an extant member of the genus Eclytus in our analysis, and this high placement probability might be due to insufficient taxon sampling and/or size-related homoplasy of the scored morphological characters. We thus agree with the decision taken by Spasojevic et al. [7] and leave it as questionable within its original genus. Future analyses might include more taxa from the morphologically quite diverse Tryphoninae and thus allow for a more meaningful conclusion. However, a placement in Orthocentrinae should definitely be considered, given the similarities to some genera of that subfamily in terms of the small size and shape of the first tergite.

Eopimpla grandis Cockerell [49]– 16%, Fig 5B. As stated previously [7], there is little evidence for placing this fossil in Pimplinae, and the phylogenetic analysis was highly ambiguous about subfamily placement. The fossil clustered with 22.5% with Ortholaba (Diacritinae), 19% with Netelia (Tryphoninae), and 16.5% on the ancestral branch leading to Ichneumonidae without Xoridinae. Removing it from Pimplinae into an uncertain subfamily association [7] thus seems justified.

Glypta? transversalis Scudder [47]– 12%, Fig 5C. Spasojevic et al. [7] suggested that this fossil does not fit very well in Banchinae, but instead could be associated with Lycorininae or with the genus Physotarsus in Tryphoninae. While the fossil was never recovered with the banchine in our dataset (Apophua, a genus closely related to Glypta), it indeed clustered most often with Physotarsus (13%), but there were also numerous alternative placements that could not be ruled out. In lack of a well-supported alternative, we leave the fossil in its original placement, but follow Spasojevic et al. [7] in adding a question mark behind the genus name.

Ichninsum appendicrassum Spasojevic et al. [7]– 26%, Fig 5D. The authors of this genus avoided subfamily placement because of presumably plesiomorphic character combinations. Our results here confirm what they suggested as the most likely subfamily placement (Pimplinae) with quite high cumulative posterior probability (65.3%). Even though the fossil cannot be associated with a single tribe within the subfamily, we move it from incertae subfamiliae to Pimplinae.

Lithotorus cressoni Scudder [47]– 20%, Fig 6A. This fossil was described in the subfamily Diplazontinae, but both Townes [50] and Spasojevic et al. [7] mentioned the possibility that it is instead closely related to the Helictes-group of genera in Orthocentrinae. Our results confirm this notion and place it either as a sister taxon to Entypoma or as a stem-lineage representative of the subfamily (without Hyperacmus). The sum of placement probabilities in Orthocentrinae approaches 98.5%, while the original placement in Diplazontinae is not supported at all. We thus move the genus Lithotorus to Orthocentrinae.

Mesoclistus? yamataroti Spasojevic et al. [7]– 23%, Fig 6B. We can confirm the subfamily placement here (Acaenitinae: 96%), and indeed observed the highest probability of it clustering with Mesoclistus (77%). As we have only included two other genera of the Coleocentrus genus group in this analysis (Coleocentrus and Procinetus), we leave the genus placement as uncertain, as it was already suggested in the original description.

Mesornatus markovici Spasojevic et al. [37]– 27%, Fig 6C. This taxon is placed with 23% probability on the stem lineage leading to all ichneumonids except Xoridinae, and with 20%, 15%, and 14.5%, respectively, with Apophua (Banchinae), Lycorina (Lycorinae), and Xorides (Xoridinae). We can thus confirm the uncertain subfamily placement of this genus; it might even belong to a stem lineage or now extinct subfamily.

Phygadeuon? petrifactellus Cockerell [49]– 14%, Fig 6D. Spasojevic et al. [7] added a question mark to the genus assignment, but left the taxon within the then subfamily Cryptinae. As this subfamily has in the meantime been split by Santos [51], the correct placement would now be in the subfamily Phygadeuontinae. Our results largely agree with this placement, but emphasize even more the large uncertainty: it grouped with Gelis areator 18% of the time, but was also found as a stem-lineage representative of Orthocentrinae in 14% of the trees. We leave the current classification unchanged but emphasize the uncertainty of the placement.

Pimpla? eocenica Cockerell [52]– 16%, Fig 7A. Spasojevic et al. [7] revised this fossil and added a question mark to the genus assignment, citing high uncertainty in this genus placement due to poor preservation of the fossil. Our morphological analysis confirmed the high uncertainty in the placement, which attached on several mostly rather basal branches within Pimplinae, but also with other subfamilies, most of all Diacritinae, Collyriinae, and Cylloceriinae. Summing the probabilities of it associating with pimpline taxa revealed a slight preference of this over any other subfamily (56%); within the subfamily, P. eocenica was most often ending up in the tribes Delomeristini, Pimplini and Theroniini or on the branches close to the ancestor of the subfamily. The highest probability (23%) was for P. eocenica to be sister to Perithous septemcinctorius, a placement not supported by any particular character. We thus leave P. eocenica in Pimplinae and, with a question mark, in the genus Pimpla.

Plectiscidea lanhami Cockerell [53]– 13%, Fig 7B. Spasojevic et al. [7] noted uncertainty in the generic placement of this fossil, mostly due to the pentagonal areolet which does not occur in any of the recent representatives of the genus. But they agreed with the placement in Orthocentrinae. Our analysis now recovers the fossil with 89% with the very small genus Allomacrus, and only with about 7.5% in Orthocenrinae. However, Allomacrus was variously classified within either Cylloceriinae or Orthocentrinae and these subfamilies are certainly closely related. Even though Allomacrus has an open areolet, its shape indicates that it might be pentagonal when closed. As far as visible in the fossil, the wing venation (i.e., proportions of pterostigma and fore wing cells), humped first tergite and proportions of the remaining tergites, and possibly long ovipositor are indeed very similar to Allomacrus, thus we transfer the fossil to this genus: Allomacrus lanhami, comb. nov.

Polyhelictes bipolarus Spasojevic et al. [37]– 26%, Fig 7C. The authors of this genus and species mentioned that the character combination would allow placement in two not very closely related groups: the Helictes genus group in Orthocentrinae, or the spider parasitoids of the Polysphincta genus group in the pimpline tribe Ephialtini. This ambiguity is expressed both in the genus and species names of the taxon. However, our analysis places it very firmly in the Polysphincta group (100%). This somewhat surprising result is probably due to some measurement characters in the fore wing that were not taken into account when the taxon was described, such as the rather elongate radial cell and relatively long vein r-rs in the forewing, which are consistently very short in Orthocentrinae. We thus move the taxon from incertae subfamiliae to Pimplinae and the tribe Ephialtini.

Rhyssella? vera Spasojevic et al. [37]– 27%, Fig 7D. This fossil received highest probability for attaching to the stem lineage of the Rhyssinae taxa we sampled (36%), followed by being sister to the extant Rhyssella furanna (18%) or on the branch leading to the remaining Rhyssinae (17%). There was some probability also that it clustered with Acaenitinae (total 24%), mostly with Coleocentrus (12%), which was recovered here as the sister group to Rhyssinae. Nevertheless, we consider the original subfamily placement as being supported by this analysis (total in Rhyssinae: 72%), although the genus assignment should be considered uncertain. We nevertheless leave the taxon in Rhyssella, which was recovered as the best option; describing a new genus, as might be suggested by the stem-lineage placement, appears unwarranted given the lack of unique autapomorphies of the taxon.

Scambus fossilobus Spasojevic et al. [37]– 33%, Fig 8A. Our analysis recovers this taxon with high certainty within Pimplinae (91%), where the highest probability attaches it to the stem- of the tribe Ephialtini (67%). This confirms the author’s claim of the first unequivocal representative of the subfamily, and the genus placement, even though not directly confirmed here, cannot be refuted, as we did not adequately sample the diversity of the large genus Scambus.

Scambus? mandibularis Spasojevic et al. [7]– 20%, Fig 8B. This fossil also was placed mostly on stem lineages in various subfamilies, but with most of the weight in Pimplinae (sums to 53%), most likely either as a stem lineage representative or ancestral to the tribe Ephialtini. Genus placement is not resolved though, and we thus leave it as uncertain within Scambus.

Scambus? parachuti Spasojevic et al. [7]– 29%, Fig 8C. This taxon received a placement in various subfamilies, but mostly in a very basal position or even as a stem lineage. This is also the case for the subfamily Pimplinae in which it was described. The highest probability is observed for a stem-lineage placement in Orthocentrinae, but this is not decisive (36%). Even though the placement in Pimplinae is not supported here, it cannot be rejected either (sums to 8%). We thus leave it as uncertain in the genus Scambus.

Tilgidopsis haesitans Cockerell [54]– 11%, Fig 8D. Described in Ophioninae but with reported similarities to the unrelated Poemeniinae, Cockerell already made it clear that subfamily placement is not simple in this taxon. Spasojevic et al. [7] formalized this by moving it to incertae subfamiliae. Interestingly, our analyses put it closer to Ophioninae again, but most likely as a stem lineage with 37.5% probability to attach to the branch leading to Leptophion (Ophioninae) and Heteropelma (Anomaloninae). Other placements in various subfamilies (Acaenitinae, Brachycyrtinae, Diacritinae, Ctenopelmatinae, etc.) cannot be fully excluded either.

Trigonator macrocheirus Spasojevic et al. [37]– 23%, Fig 9A. Despite rather sparse sampling of the subfamily Labeninae, this fossil is firmly placed within that subfamily (87%), with closest ties to Poecilocryptus (80%) and Certonotus (6%), the sole representatives of the two tribes that the authors associated it with. However, without a denser sampling of the labenine genera, tribal placement of the genus cannot be made with any certainty.

Trymectus amasidis [55]– 22%, Fig 9B. This fossil was originally described in Tryphon and is rather well preserved; the more surprising was its unresolved placement. It ended up on the branch leading to the Tryphoninae taxa Zagrphus and Thymaris in 24% of the trees, with the rest of the probabilities distributed rather evenly among numerous taxa in subfamilies as different as Metopiinae, Ctenopelmatinae, Banchinae, Pimplinae etc., or on some of the most basal branches in the tree. The placement as a new genus Trymectus with uncertain subfamily placement by Spasojevic et al. (named after the first letters of the first three afore-mentioned subfamilies; [7]) can thus here be confirmed.

Xanthopimpla? messelensis Spasojevic et al. [37]– 30%, Fig 9C. While the similarities of this fossil with recent representatives of the genus seem striking at first glance (see Fig 3 in [37]), our analysis placed it with higher probability in Acaenitinae (50% with Mesoclistus), even though Xanthopimpla and its sister genus Lissopimpla remain a clear possibility (19% in total). We acknowledge that we judged some of the characters typical of Xanthopimpla as not clear enough for scoring in the fossil, even though they are indicated, e.g., the groves on the metasomal tergites, enlarged claws, the potentially twisted mandible and short hypopygium, the very narrow triangular areolet, and the slightly medially down-curved ovipositor. Very similar shapes of the areolet and ovipositor occur in some extant Xanthopimpla species [56], and our failure to include them in the morphological matrix might have hindered the confident placement of the fossil.

Adding these characters and achieving a better coverage of the morphological diversity among extant Xanthopimpla species would probably tip the scale in favour of the original placement. The shape and length of the ovipositor and absence of a large, triangular hypopygium strongly contradict a placement in Acaenitinae. We thus refrain from making any formal subfamily changes here, but put a question mark behind the genus name to express the uncertainty in the phylogenetic placement. Especially, the fossil could also belong to a stem lineage of Xanthopimpla + Lissopimpla or of Theronini.

Xanthopimpla praeclara Spasojevic et al. [37]– 29%, Fig 9D. In contrast to the previous fossil, this species was placed very firmly (99.9%) with the recent Xanthopimpla varimaculata, confirming that this genus dates back to the Early to Mid Eocene.

Evolutionary models for morphological data

The application of model-based approaches to a specific type of character requires the availability of sufficiently realistic models of its evolution, and the discussion whether the simplifying Mk model [15] is suitable to infer phylogenetic relationships from discrete, morphological data is still on-going. It was further sparked by a simulation study suggesting the superiority of the Bayesian approach over parsimony for morphological data [57], but the specifics of the simulation procedure, especially the single set of relative branch lengths used, are still under discussion [58,59].

Even though a lot remains to be done in the field of morphology models, the stochastic framework allows for a direct comparison of the fit of different models, as we have conducted here when comparing the unordered versus ordered Mk models and the expansion of the state space as implied by a larger morphological matrix (Table 2). The very large differences in model likelihoods revealed by stepping-stone sampling suggest that provided with a large-enough data matrix (222 characters times 127 taxa), such comparisons are very instructive. In our case, the largest effect was observed when comparing the ‘ordered Mk’ model with or without gamma-distributed among character rate variation (ACRV), an observation made earlier for molecular data [60]. ACRV will likely constitute a vital part of any morphology model, as was already suggested by Lewis [15], but systematic reviews of empirical matrices are currently missing. The second-largest Bayes factor resulted from the comparison of ‘unordered’ versus ‘ordered’ data, with characters for ordering chosen a priori according to character conceptualization (S2 File). Some of the ‘ordered’ characters (13 out of 58) in fact represent discretised continuous characters, and it remains to be shown whether character ordering is vital also in datasets where continuous characters are analysed as such [61]. Finally, we tested whether increasing the state space of each character by adding unobserved states as they appeared in a larger matrix of ichneumonid taxa (unpublished data). Even though one might argue that such an approach leads to a more realistic morphology model, it received lower marginal likelihoods in our analyses (Table 2). A possible explanation of this is that we used a fully symmetrical model, where transitions between different states and thus also state frequencies are assumed to be equal at stationarity. Adding states that are not actually observed among the data creates additional asymmetry in state frequencies, which might be the reason for the lower marginal likelihood of that model.

We also attempted to run an asymmetric model which allows state frequencies to be unequal following a beta distribution for two-state or Dirichlet distribution for multi-state characters. This extension to the model has already been suggested by Lewis [15] and is implemented in MrBayes [41]. It has been demonstrated recently to outperform the symmetric model in about half of 206 empirical datasets, most of which were rather small [14]. On our data matrix of 222 characters times 127 taxa, even using a fixed level of asymmetry instead of a hyper-prior, we found computation times to be prohibitive (more than 1,100 CPU hours for 10 million generations).

Bayesian phylogenetic fossil placement

Our morphological phylogenetic analysis exemplifies the use of Bayesian inference in the classification of fossil taxa and shows how placement uncertainty can be illustrated using RoguePlots. The advantages of reproducible, stochastic analyses of fossil placement have been discussed previously [16,62], and we here further emphasize the benefits of the direct assessment of alternative placements that comes with Bayesian approaches. Placing fossils in a classification system that is strongly biased towards extant taxa, as it is the case in most groups with a poor (or at least poorly studied) fossil record, triggers difficulties that go beyond incomplete fossil preservation. Representatives of a stem lineage might not yet possess all the synapomorphies of the crown group, requiring an extension of the circumscription of a taxon. And even more importantly, because the divergence times for most taxa remain poorly known or contentious, it is often unclear to what degree a fossil species can be placed at all in a certain taxonomic rank: the genus, tribe, or even family in question might not have existed yet at the time the species fossilized. A phylogenetic perspective on fossil classification is thus vital for a proper evaluation and communication of both the position of a fossil in the taxonomic system and the amount of uncertainty associated with it.

RoguePlots provide a natural way of representing placement uncertainty on a tree summary like a consensus tree, but also on individual trees, such as a maximum likelihood tree or shortest tree under the parsimony criterion. The approach can take sets of trees from a Bayesian analysis, but also those resulting from bootstrapped data matrices; it illustrates the frequencies with which different fossil placements occur among the different trees. One difficulty arising from the need to represent relationships among fossil and extant taxa in a potentially very heterogeneous set of topologies lies in the restrictions imposed by using a single tree plot: even if the tree in question is fully resolved, some of the trees in the set will most likely contain clades not represented in it. RoguePlots circumvent this issue by making use of the distinction in branches that are parallel versus perpendicular to the direction of a rectangular phylogram: parallel branches are used to illustrate the probability of direct attachment to the branch, while perpendicular branches represent the sum on any branch not in the tree, of which the corresponding node is the most recent common ancestor. This approach ascertains that the shown placement probabilities sum to one, which might not be the case otherwise (e.g., see Fig 8 in [23]). While it is possible to produce RoguePlots based on trees obtained from a maximum likelihood or parsimony analysis of a morphological dataset, their interpretation is much more intuitive in the Bayesian context, where the frequency of a clade in the set of trees represents its posterior probability. In contrast, bootstrapping a morphological matrix will necessarily remove some characters, while multiplying others, precluding a straight-forward interpretation of the resulting placement frequencies. A Bayesian analysis can thus be viewed as being closer to the traditional way of placing fossils, which is based on shared synapomorphies.

Morphological phylogenetics versus character evidence for fossil placement

Using a stochastic framework for the phylogenetic analysis of fossil taxa does neither replace nor preclude careful character evaluation, but instead is based on it being made explicit. The step of coding morphological observations into discrete characters and states is crucial for this endeavour and certainly comes with a string of pitfalls, from character conceptualization to the delimitation of character states. In the ideal case, characters that would have been used as arguments to place a fossil can be adequately represented in a morphological matrix, in which case the phylogenetic reconstruction method identifies them as characters that evolve at a low rate (which is nearly equivalent to identifying them as synapomorphies). They will lead to a high probability of the predicted placement of the fossil, or at least to an accurate representation of the uncertainty of this placement (but see below). In our analyses, this was the case for the vast majority of fossils, where the RoguePlots provided matching illustrations for alternative placements that had already been discussed using character evidence in the original descriptions or revisions of the fossils [7,28,37]. In several cases, the morphological phylogenetic analysis helped identify informative characters that have not been used much previously to define higher taxa, and were thus omitted from consideration in the original descriptions. The best example here is Polyhelictes bipolarus, a taxon originally described with uncertain subfamily association [37], as the traditionally used characters would imply a placement either in the Helictes-group of genera in the subfamily Orthocentrinae, or among the spider parasitoids of the Polysphincta-group in Pimplinae. Examination of the morphological matrix showed that the unequivocal placement in the latter group in the morphological phylogeny was driven mostly by measurements of forewing veins and cells, characters not typically used in ichneumonid subfamily diagnoses.

Morphological phylogenetics applied to fossils has repeatedly altered down-stream macroevolutionary interpretations that were based on the fossils being placed through character-evidence alone, for instance in the context of the inference of morphological evolution [63], time-calibrated phylogenies [23] or historical biogeography [64]. However, any phylogenetic analysis is only as good as the underlying morphological matrix, and several issues can result in erroneous phylogenetic placement of fossils, such as limitations in character or taxon sampling, artefacts caused by the fossilization process [46], limited phylogenetic signal in the morphological data [65–67], or misspecification of the evolutionary model [12–15]. The identification of resulting artefacts is crucial when interpreting the results of a morphological phylogenetic analysis and can point to way to improving the representation of the morphological diversity of extant and fossil taxa in the matrix in the future.

In ichneumonids, limited phylogenetic signal in the morphological data might be particularly problematic, as high levels of homoplasy have been reported for the group [38]. Many character states, especially those related to parasitoid life style and host choice, converged in distantly related groups (e.g., long ovipositor in parasitoids of wood boring insects; short and thick legs in parasitoids of concealed hosts, etc.). A potential solution when confronted with limited phylogenetic signal in a morphological matrix might be to perform a combined analysis with molecular data in order to get a better-resolved consensus tree on which fossil placement could then be plotted; however, our morphological tree is not in fact that much less resolved than a molecular tree resulting from a recent 93-gene study[45]. Although this molecular tree recovered three supra-subfamilial groupings, the Ichneumoniformes, the Pimpliformes, and the Ophioniformes, its backbone was largely unresolved within these groups; the signal in the morphological data is thus not even that much weaker. Furthermore, the low backbone resolution might only have influenced the placement of fossil taxa that belong to stem groups or to extinct subfamilies.

Conclusions and future directions

We here exemplified the use of RoguePlots in combination with Bayesian morphological phylogenetics to re-assess the classification of 24 fossil ichneumonids. While numerous fossils of these parasitoid wasps remain to be studied, we hope that the approach will prove useful also in many other groups to study and illustrate alternative fossil placements (as well as the phylogenetic placement of other ‘rogue taxa’, [22,68]) and that it will lead to a more in-depth appreciation of fossil placement uncertainty. This is crucial for many macro-evolutionary inference approaches that rely on the correct interpretation of the fossil record, for instance for deriving calibration information for molecular dating studies [1,69]. Potential artefacts specific to morphological phylogenetics deserve further study, especially the impact of model misspecification and of an inadequate representation of the morphological diversity of a group due to limited taxon sampling. Also, there is ample room for improvement of models of morphological character evolution, an area explored insufficiently in the past (but see, e.g., [12–14,70,71]).

A natural extension of our approach would be to include age information of the fossils to improve their placement, either in the context of morphology-only tip dating [19,20] or combined with molecular data in total-evidence dating [23,72]. This might have a considerable impact; for instance, knowing that a fossil is older than the inferred splits in the crown group would enforce its placement as a stem-lineage representative. However, including age information might interact with the morphology model to produce artefacts, especially if a character evolves far from stationarity on the time scale covered by the phylogeny [13]. Carful model testing would expose such artefacts and improve our understanding of character evolution. In any case, using the context of a calibrated phylogeny to further test the placement of ichneumonid fossils would represent a promising next step and lead to a better understanding of the evolutionary history of this highly species-rich taxon.