In order to assess the phylogenetic distribution of the cecal appendix across mammals and attempt to correlate it with ecological and anatomical variables, a broad literature search of published studies on mammalian gastrointestinal anatomy was conducted. At the species level, anatomical data were compiled for appendix presence and length, cecal size and morphology, colon length, presence of colonic separation mechanisms (CSMs), and stomach histological lining (entirely glandular vs. some squamous epithelium) for a broad sample of 361 species representing all major mammalian clades (SOM 1, 2). For quantitative variables, numeric values were compiled from the literature. For qualitative characters, published descriptions were used. The nomenclature of species names follows the on-line edition of a standard on-line reference (Wilson and Reeder, 2005). To determine whether the presence of an appendix is associated with dietary factors, data on published dietary categories of these species were also compiled (SOM 3). Data on activity pattern (time of day that an animal is typically most active), body mass, and mean social group size for each species were taken from the PanTHERIA database (Jones et al., 2009). We also collated data on whether each species practiced coprophagia, which we define for the purposes of this study as the regular eating of any type of feces in sufficient amounts to make a significant contribution to the nutrition of the animal as reported in the literature.

Following Smith et al. (2009), we defined a cecal appendix as a narrow, close-ended extension off the apex of the cecum with a distinct change in diameter of the lumen between it and the cecum. While previous studies have also used characteristics such as a thickened wall and concentration of lymphoid tissue to define this structure (Fisher, 2000), there were insufficient published data on these characters in most species for their inclusion here.

Given that the presence and size of the cecum and appendix are thought to be related to cellulose richness in the diet (Darwin, 1871), we have scored diet type and ordered the states in a manner which roughly reflects increasing order of cellulose richness: 0, carnivory; 1, insectivory; 2, omnivory; 3, frugivory; 4, granivory; 5, gummivory; 6, folivory. To verify that our results did not depend too heavily on a detailed ordering scheme, we have also lumped the states into a binary character, in which state 0 represents a cellulose-poor diet (including carnivory, insectivory, omnivory, and frugivory) and state 1 represents a cellulose-rich diet (including folivory, granivory, and gummivory). We also examined slightly different coding schemes to assess the sensibility of our results to this factor, but these did not alter the results substantially (probabilities were about the same), so we do not report the results of these additional analyses below. The examined alternative, for the multi-state character, consisted of inverting the position of gummivory (to state 3) and frugivory (to state 5). For the binary character, we also assessed the related issue of impact of food hardness (hard, including insectivory, folivory, granivory, and gummivory; soft, including carnivory, omnivory, and frugivory), and again moving frugivory into the “hard” category. To determine whether a cecal appendix was found more frequently in combination with any particular cecal shape, a qualitative cecal morphology character was created with the following states: 0, cecum absent; 1, appendix-like; 2, spherical; 3, cylindrical; 4, tapering; 5, spiral; 6, paired.

The phylogeny follows a recent, nearly exhaustive, species-level mammalian phylogeny (Bininda-Emonds et al., 2007) for topology and branch lengths. More specifically, we have used the electronic version of the phylogeny (in NEXUS format), in the variant using the best estimates of molecular ages. We then pruned the tree to retain only the species in our sample, and added a few taxa from our sample to the tree using the literature (usually, the taxonomy in Wilson and Reeder [2005]) to determine the best topology and approximate divergence time. These added taxa are the marsupials Chaeropus ecaudatus, Thylacinus cynocephalus, and Antechinmys laniger, the afrotherians Rhynchocyon cirnei and Elephantulus pilicaudus, the phocid Pagophilus groenlandicus, and the wild horse Equus caballus przewalskii. The tree includes 361 terminal taxa representing all main mammalian clades. There are 19 polytomies, each with three to 13 daughter branches. Given that comparative tests require fully dichotomous trees or perform better with such trees, for the analyses using pairwise comparisons (described in more detail below), we used a set of 10 randomly resolved trees produced by Mesquite (Maddison and Maddison, 2011) by the addition of zero-length branches, and averaged the results (probabilities) over the 10 trees. In any case, when the average probability suggests significance, most tests using the individual random resolutions and the test using the tree with polytomies suggest significance, and conversely, when the average probability is over 0.05, in most cases, most other tests gave congruent results. This procedure is not problematic because pairwise comparisons are a non-parametric test that does not use branch lengths (thus, adding length to the randomly added branches would not have altered the results). We used the “most pairs” selection algorithm, which maximizes the number of compared pairs.

To compare the relative size of various parts of the digestive system, we divided linear measurements by the cubic root of body mass estimates that we collected from the literature. The adequacy of this transformation was checked through linear regressions in the Microsoft Excel analytical tools. The data matrix is available in Mesquite (Maddison and Maddison, 2011) NEXUS format (SOM 2).

To select appropriate analytical methods for our data, we performed several exploratory analyses. We determined if there was a phylogenetic signal in the data by comparing the tree length using squared-change parsimony, character by character, to a population of random trees. Because of the presence of polytomies in our master tree, we used the random taxon reshuffling procedure (Laurin, 2004). This procedure has the advantage, over the equiprobable tree algorithm, of retaining the same tree resolution and branch length distribution. The latter is required to deal with continuous characters, if squared-change parsimony is used (Maddison, 1991), because the squared length depends on branch lengths as well as on topology.

For continuous data, we had initially planned to use phylogenetic independent contrasts or PIC, for short (Felsenstein, 1985) to assess character correlation. However, diagnostic checks using the four tests available in the Phenotypic Diversity Analysis Program (PDAP: PDTREE) module of Mesquite (Midford et al., 2008) revealed very highly significant artifact that various data and branch length transformations failed to adequately remove. These artifacts include the four classical ones that Mesquite allows testing for, namely the following four regressions: 1, absolute value of the standardized contrasts vs. their expected standard deviation; 2, absolute value of standardized contrasts vs. nodal value; 3, absolute value of the standardized contrasts vs. node height; and 4, estimated nodal value vs. node height. In an ultrametric tree (as is the case here because only extant and very recently extinct taxa are represented), significant relationships in any of these four characters implies that the characters did not evolve according to Brownian motion over the reference tree, which may mean that the characters followed another evolutionary model, or that the topology and/or branch lengths are wrong. In any case, this means that PIC would yield unreliable results. Because of this, and because we have both discrete and continuous data, we decided to use the pairwise comparison test (Maddison, 2000 and Read and Nee, 1995), which effectively evaluates the association between changes in two variables. All tests of correlations reported in this manuscript were made using pairwise comparisons. This test does not require the assumption that characters evolved according to a Brownian motion model (contrary to PIC), and can handle discrete as well as continuous variables, polymorphism, and even polytomies (at the cost of reduced power). More sophisticated methods are either designed for continuous data only, such as Phylogenetic Generalized Least Squares (PGLS), or disallow polymorphism (more than one state being present in a given taxon, which is abundant in our data), such as the Mesquite implementation of Pagel's (Pagel, 1994) correlation method.

To test the hypothesis that gains and losses of the appendix are equally frequent, we used a binomial distribution to assess the probability of this pattern, assuming that gains are as probable as losses. Binomial probabilities were calculated using GraphPad Software (http://www.graphpad.com/quickcalcs/binomial2.cfm).

To compare the evolutionary rates of the appendix in various clades, we have simply divided the number of inferred transitions (using parsimony) in each clade by that clade's sampled phylogenetic diversity index (Faith, 1992), which was calculated using the Stratigraphic Tools of Mesquite (Josse et al., 2006). To assess the statistical significance of the extreme difference in evolutionary rates between Euarchontoglires (fast evolution of the appendix) and its sister-group Laurasiatheria (no evolution of the appendix), we did a binomial test to assess the probability that all 33 changes observed in the smallest clade that includes Euarchontoglires and Laurasiatheria occur in Euarchontoglires, under the null hypothesis that chances are equally likely to occur in lineages of either clade. The null hypothesis is that the proportion of changes reflects the relative sampled phylogenetic diversity of both clades. The sampled phylogenetic diversity of Euarchontoglires is 4762.9 million years (Ma), and that of Laurasiatheria is 3220.1 Ma; thus, each of the 33 events had a probability of 0.592 of occurring within Euarchontoglires.

The use of parsimony both for phylogenetic signal detection and for inferring character history is justified largely by pragmatic considerations. We are unaware of software implementing maximal likelihood or Bayesian methods with the diversity of data that we have (continuous characters, multi-state discrete characters, both ordered and unordered, binary data, missing data, with a tree incorporating polytomies). In any case, parsimony has the advantage of having been used for such analyses for decades, so its properties are well understood. Bayesian methods are advantageous especially when there is a population of trees in which node frequency reflects support, but this is not the case here (in our randomized trees, node frequencies are essentially random).

Given the number of statistical tests performed below, we corrected for multiple testing using the False Discovery Rate, a procedure that is simple to use, statistically valid, and more powerful than conventional Bonferroni corrections (Benjamini and Hochberg, 1995 and Curran-Everett, 2000).

Of the 361 mammalian species sampled here, 50 were found to possess a cecal appendix (SOM 2). As previously noted by Smith et al. (2009), several different morphotypes of the appendix have been described in mammals (Fig. 1). An appendix was found to be present in the Metatheria and Euarchontoglires, but not Laurasiatheria.

Most characters display a strong phylogenetic signal (Table 1), which confirms that phylogeny-informed analyses (such as pairwise comparisons) are required to assess correlation between characters. This result also indicates that character history can be reliably inferred from optimizations (Cubo et al., 2002 and Laurin, 2004).

Darwin's suggestion that the hominoid appendix is a vestige of a larger cecum was derived in part from the assertion that the appearance of the appendix occurs concomitantly with decreases in cecum size and a shift from folivory to frugivory resulting in decreased dietary cellulose consumption. Parsimony optimization (Swofford and Maddison, 1987), a procedure that minimizes the changes on a tree to account for character state distribution, indeed confirms that the appearance of the appendix in hominoids is associated with a decrease in cecum size (Fig. 2, Fig. 3, Fig. 4, Fig. 5 and Fig. 6) in these mostly frugivorous taxa. However, this may be a coincidence because no statistical test was performed to assess the probability that this association is random, and this is typically very difficult to do on singular events (as is the case here; there is a single appearance of the appendix in hominoids). Furthermore, our tests fail to support the general application of this model to the evolution of the appendix in mammals, since the changes in the state of the appendix were not associated with changes in the state of cellulose richness of diet (Table 2). However, changes in the character state of the cecum size were also not correlated with changes in cellulose digestion (Table 2), a counter-intuitive result that might possibly change with a denser taxonomic sampling.

Parsimony optimization of appendix presence and cecum presence and size (Fig. 2, Fig. 3, Fig. 4, Fig. 5 and Fig. 6) refutes the suggestion that appendix appearance is linked to a decrease in cecum size within mammals. Thus, the ancestral mammal and monotreme are reconstructed as having had a small cecum (Fig. 2), and there is no reason to believe that a large cecum was present in any older ancestor of these taxa, yet, monotremes have an appendix. The same applies to the marsupial Caenolestes fuliginosus. The other marsupials with an appendix (Phascolarctos cinereus, Vombatus usrinus, Lasiorhinus krefftii, and Phalanger carmelitae) are in a part of the tree in which the optimization is ambiguous (Fig. 2), but in which the cecum may be small or mid-sized. In Euarchontoglires (and specifically within primates and glires) we find an appendix often associated with a large cecum (Fig. 5 and Fig. 6). A large cecum is clearly primitive for that clade, but only a small proportion of primates (and no glires) with a reduced cecum have an appendix. This last pattern, the only one consistent with Darwin's hypothesis, occurs only in the strepsirhine Daubentonia madagascariensis, in hominoids, and in some cercopithecids (Fig. 5). In the latter, an appendix is present in Macaca nigra, and the appendix is polymorphic (not being present in all individuals) in several other cercopithecids (Procolobus verus, Pliocolobus badius, Colobus polykomos, Cercopithecus mitis, C. mona, Chlorocebus aethiops, and Papio hamadryas). Among eutherians, the only other example consistent with Darwin's hypothesis is provided by the sirenian Trichechus manatus, in which appendix appearance is associated with a reduction in size of the cecum, from mid-sized to small (Fig. 3). However, a cecum reduction occurs in other clades that lack an appendix (Fig. 5), such as the strepsirhine Cheirogaleus major, the cebids Alouatta and Ateles, and several cercopithecids (Nasalis larvatus, Presbytis femoralis, Colobus angolensis, Allenopithecus nigroviridis, Miopithecus talapoin, Erythrocebus patas, Macaca mulatta, Cercocebus agilis, Mandrillus leucophaeus, and Theropithecus gelada). Thus, among mammals, only Trichechus manatus and some catarrhine primates (especially hominoids) exhibit the predicted combination of an appendix accompanying a small cecum.

A more convincing demonstration that Darwin's hypothesis on the evolution of the cecal appendix is not generally applicable outside of hominoids comes from the strong positive correlation that we have found between changes in the relative lengths of the appendix, cecum, and colon (Table 2). When the appendix is absent, its length is considered to be 0, so this character also captures information about the presence or absence of the appendix, and this correlation is no longer significant if 0-values are excluded, reflecting the small sample size (n = 9) of the other taxa for which comparisons can be made. Darwin's hypothesis predicts a reverse relationship between sizes of the appendix and cecum, but instead, we find a fairly strong positive relationship (P = 0.005).

To evaluate other potential relationships that might provide insight into the evolution of the morphology of the proximal colon, we tested the relationship between appendix and colon size, and between cecum and colon size. Here our results show that all three parts of the digestive system that we studied are positively correlated with each other, and that the correlation between cecum and colon length is the strongest (P < 0.001), whereas the correlation between appendix and colon is the weakest, although still clearly significant (P = 0.017). This cannot be an indirect correlation reflecting body size because we had divided all length measurements by the cubic root of body mass, and linear regressions indicated that this effectively removed the size effect (which was present in the raw measurements). However, these correlations hold partly because most species without an appendix appear to have a smaller colon and cecum than species with an appendix, so this correlation reflects both presence and size of the appendix. Because of sample size, we are not currently able to determine if this correlation holds within species with an appendix, but this will be tackled in a subsequent study with a greater sample size.

The pairwise comparison test revealed no statistically significant correlations between changes in the appendix and changes in colonic separation mechanisms, cellulose in the diet, stomach wall histological composition, or cecal haustrations during the course of evolution. However, the relationship with CSM is less certain, because this character contains quite a bit of missing data. Alternate schemes of binary cellulose richness against group mean size and appendix, colon, and cecum length divided by cubic root of body mass, also resulted in statistically non-significant results. At least one of our multi-state coding schemes had granivory ranking as lower-cellulose than frugivory, but this also yielded non-significant results.

Additionally, appendices were not found to occur significantly more or less frequently in any particular character state of cecal morphology. Rather, the cecal appendix was observed to occur in at least one species in five of the seven categorical cecal shape states. Further, more highly social (those living in larger groups) and/or diurnal animals do not appear to be more likely to have an appendix than species living in smaller groups, or nocturnal or cathemeral/crepuscular species. A post hoc analysis including only taxonomic groups for which some species possess an appendix, the Euarchontoglires and Metatheria, also yielded no significant correlations (P = 0.25–1.0), suggesting that the appendix-less Laurasiatherians are not obscuring an otherwise observable pattern.

Across the entire mammalian phylogeny, the appendix was found to have undergone 38 evolutionary events, including 32 to 38 gains, and a maximum of six losses. The Chi² test shows that there are significantly more gains than expected by chance alone if gains were equally probable as losses (P < 0.00025). A binomial test gives congruent results (P < 0.0001).

The appendix evolves at a global rate of 0.0034 transitions (gains or losses) per lineage and per million years (Ma) for mammals. However, that rate is extremely heterogeneous (Table 3), varying between 0 in Laurasiatheria (no identified transition) to 0.0071 in Euarchontoglires. Marsupials display an intermediate rate of 0.0019 transitions per lineage per Ma. These rates may be low because we have not sampled all species. Further, the rate may be underestimated since it is possible that multiple gains and losses occurred for polymorphic species, whereas we consider the transition to polymorphism as a single step. Given that polymorphism is restricted to Euarchontoglires in our matrix, the evolutionary rate in that taxon is likely more underestimated than in other taxa. The binomial test gives clear results: whether the comparison is between Euarchontoglires and Laurasiatheria, or between the former and all other placental mammals, the probability of getting such an extreme distribution if changes (per lineage and per Ma) are equally probable in all parts of the tree is less than 0.0001.