Spiders did not repeatedly gain, but repeatedly lost, foraging webs

Much genomic-scale, especially transcriptomic, data on spider phylogeny has accumulated in the last few years. These data have recently been used to investigate the diverse architectures and the origin of spider webs, concluding that the ancestral spider spun no foraging web, that spider webs evolved de novo 10–14 times, and that the orb web evolved at least three times. These findings in fact result from a particular phylogenetic character coding strategy, specifically coding the absence of webs as logically equivalent, and homologous to, 10 other observable (i.e., not absent) web architectures. “Absence” of webs should be regarded as inapplicable data. To be analyzed properly by character optimization algorithms, it should be coded as “?” because these codes—or their equivalent—are handled differently by such algorithms. Additional problems include critical misspellings of taxon names from one analysis to the next (misspellings cause some optimization algorithms to drop terminals, which affects taxon sampling and results), and mistakes in spider natural history. In sum, the method causes character optimization algorithms to produce counter-intuitive results, and does not distinguish absence from secondary loss. Proper treatment of missing entries and corrected data instead imply that foraging webs are primitive for spiders and that webs have been lost ∼5–7 times, not gained 10–14 times. The orb web, specifically, may be homologous (originated only once) although lost 2–6 times.

the last few years. These data have recently been used to investigate the diverse architectures 23 and the origin of spider webs, concluding that the ancestral spider spun no foraging web, that 24 spider webs evolved de novo 10-14 times, and that the orb web evolved at least three times. In 25 fact, these findings principally result from inappropriate phylogenetic methodology, specifically 26 coding the absence of webs as logically equivalent, and homologous to, 10 other observable 27 (i.e. not absent) web architectures. "Absence" of webs is simply inapplicable data. To be 28 analyzed properly by character optimization algorithms, it must be coded as "?" or "-" because 29 these codes, and these alone, are handled differently by such algorithms. Additional problems 30 include critical misspellings of taxon names from one analysis to the next (dropping even one 31 taxon affects taxon sampling and results), and mistakes in spider natural history. In sum, 32 methodological error: 1) causes character optimization algorithms to produce illogical results, 33 and 2) does not distinguish absence from secondary loss. Proper methodology and corrected 34 data instead imply that foraging webs are primitive for spiders and that webs have been lost ~5-35 7 times, not gained 10-14 times. The orb web, specifically, may be homologous (originated only 36 once) although lost 2-6 times. 37

38
"Not all living spiders spin webs, but since 1950 web-building species have been found in 39 almost all the families of spiders once thought of as wandering hunters. It now seems likely that 40 all spiders who actively hunt their prey, or use little or no silk in prey capture, are descendants of 41 web builders." Shear (1994). 42 A The evolution of silk use and web architectures, in particular the origin, modification, and/or 43 loss of the orb web is one of the more fundamental questions in spider biology. Although the 44 ancestral spider has always been presumed to employ silk in prey capture, modern spiders do 45 spin two rather different kinds of orb webs whose homology is hotly debated: the Uloboridae 46 and Deinopidae with mechanically adhesive micro-threads produced by the cribellum-a special 47 spinning plate-and the araneoids with viscid glue produced by two pairs of specialized silk 48 spigots. In the decades preceding the 1980's, arachnologists generally hypothesized that the 49 cribellate and viscid silk orb weavers were reciprocally monophyletic and only distantly related; 50 under this scenario the two kinds of orb webs were convergent. 51 Lehtinen (1967) and Forster (1970) most prominently argued that the cribellum was primitive 52 for araneomorph spiders, and that ecribellate lineages, including araneoids, evolved from 53 cribellate ancestors. If true, they jointly refuted the best argument for orb web convergence and 54 orb weaver polyphyly. Orb weavers--the Orbiculariae-were arguably monophyletic, (reviewed 55 in Coddington, 1986b;Coddington & Levi, 1991, Miller et al., 2010. This 'single origin' 56 hypothesis found evidence in silk gland and spigot morphology as well as observations of 57 similar web-building behaviorsoften identical stereotypical means of laying down similar 58 threads (Eberhard, 1982;Coddington, 1986aCoddington, , 1986bCoddington, , 1986c; reviewed in Eberhard, 1990). 59 Skeptics countered that web architecture was strongly selected, that the orb itself was adaptive 60 and likely convergent, and that morphological and behavioral similarities could be explained 61 away (e.g. Kovoor & Peters, 1988). 62 Although the behavioral and morphological evidence seemed compatible with monophyly, 63 molecular evidence repeatedly questioned orb weaver monophyly. Early targeted gene 64 analyses were largely discounted due to sparse taxon sampling and the perceived inadequacy 65 of the genes sampled (e.g., rRNA and mtDNA genes, Agnarsson, Coddington, Kuntner, 2013 When closely examined these inferences of convergence, either of webs in general or orbs, 93 derive more from what we regard as inappropriatein phylogenetic methods than from the data 94 (although we do dispute some facts, see below). All character optimization algorithms assume 95 that the digits representing character states code for observable, real phenomena. The only 96 exception is missing or inapplicable data, both by convention coded as "?". F&al, however, 97 coded absence of webs using digits (states "6" and "8"), and, given the taxon sample, "no 98 foraging web" optimizes as the ancestral spider condition. But "absence" of webs is not the 99 presence of anything. Absence codes in the data matrix could mean missing, if the taxon is 100 thought to possess a character state, or inapplicable if it is known to lack one, as in this case. 101 Optimization algorithms purposely treat "?" as a special case, different from digits. Given how 102 optimization algorithms work, coding web absence as an observed state (a digit) rather than "?" 103 affects the results. 104 To disentangle the effects of this practice, as well as the relevant shifting data, figures, and 105 publications from April until June, we attempt to duplicate their results to investigate the effect of 106 this methodological choice (composite coding, Strong & Lipscomb, 1999) (and a few empirical 107 mistakes), and to reanalyze their emended data (and our emendations to that) to assess 108 whether web building in spiders evolved 10-14 times, and the orb web 3-6 times. 109 The objectives of this paper are threefold. First, we justify in more detail below why coding 110 absence as the presence of something homologous and coordinate to observable web 111 architectures yields illogical results in this case, as well as disputing some empirical details. 112 Second, we reanalyze the F&al June dataset (with their altered data but methodological 113 problems fixed) to show that webs are primitive and homologous for spiders. Finally, we show 114 that the orb web single origin hypothesis is still reasonable and supported by data, and 115 recommend future improvements to homology hypotheses on web evolution. 116 Properly analyzed, the evidence suggests that prey capture webs are an ancient trait of all 117 spiders. They did not independently evolve 10-14 times. Orb webs may be homologous as orb 118 webs. 119 120

121
To test for the effect of methodological errors on results, we attempted to replicate F&al's 122 results for their transcriptomic data using their corrected web codings. Their original web 123 character was (0) orb; (1) brush sheet; (2) irregular aerial sheet; (3) irregular ground sheet; (4) 124 stereotyped aerial sheet; (5) cobweb; (6) no foraging web; (7) aerial (above ground) silk tube; 125 (8) tubular silk-lined burrows; (9) irregular tangle (not sheet-like). After F&al first published in 126 April, 2018, they published an erratum in June (thanking colleagues for bringing errors to their 127 attention), as well as a new version of the data (supplementary material) and of the publication. 128 The originals are no longer publicly available, although we will provide them upon request. The 129 June version changed 14% (23 of 159 for the transcriptomic matrix) of character states for 130 webs, changed the meaning of state 8 to "no foraging silk-lined burrows," and added a 11 th 131 state, (10) terminal line (see Table S1). 132 Our emended data set (Table S1) recodes 58 F&al "no foraging web" terminals as "?". To 133 capture the homology of all spider webs as webs, we include an additional character "webs: 134 present; absent," a method known as reductive coding (Strong & Lipscomb, 1999). We also take 135 issue with an additional 27 codings that we think are factually wrong (Table S1), but generally 136 accepted F&al's re-codings (e.g. adding an 11 th state "terminal line" to code Segestria) in order 137 to test fairly the effect of reductive versus composite coding (Strong & Lipscomb, 1999). Most of 138 these changes do not affect our two main points (webs are ancestral for spiders and orb webs 139 may be homologous). Examples are that Hypochilus spins a "stereotyped aerial sheet," not an 140 "irregular ground sheet," Scytodes spins an "irregular aerial sheet," not "no foraging web," and 141 Cicurina with our tree and their downloaded data matrix (Fig. S1). 152 We prefer to use the R package corHMM (Beaulieu et al. 2013) on an ultrametric tree to infer 153 ancestral character states. The 'rayDISC' package specifically accommodates character 154 polymorphisms and missing data. Character optimizations using equal (ER), symmetric (SYM), 155 and all rates different (ARD) models were explored for these data using 'rayDISC' (corHMM).

156
The ape package 'ace,' on the other hand, does not handle such data natively, but requires 157 modification to the package code itself and in our experience often passed errors for complex 158 character optimizations with many states and missing/inapplicable data (e.g. the ER, SYM, and 159 ARD models failed at times). Because the results differed moderately based on the model used, 160 AICc scores were employed to select a statistically preferred model. 161

163
We were unable to replicate exactly the online corrected results reported by F&al using the 164 'ace' character reconstruction. Although aspects of the ancestral state reconstruction in our 165 analysis match approximately, orbs evolve four, not three times independently (Fig. S1 versus  166 F&al Figure 3A). 167 This disparity apparently arises because their web optimization (F&al Figure 3A) includes 168 only 158 terminals-omitting Pararchaeawhich is present in F&al Figure 1A (thus 159 169 terminals). F&al misspelled Pararchaea as "Pararchea." If taxon tree and matrix labels in 'ace' 170 do not match exactly, the tip is dropped, which of course affects optimizations. F&al scored 171 Pararchaea as "no foraging web" and it fell in the clade sister to tetragnathids. Rather than 172 maintaining a larger probability of an orb web weaving ancestor (with subsequent loss further up 173 the tree) the webless Pararchaea shifts the ancestral reconstruction for tetragnathids more 174 towards a webless ancestor. As such misspellings are easy to miss in such a large tree, it is 175 possible that more of the orb web optimizations reported by F&al may be short by one additional 176 origin (F&al Table 1). F&al's corrected figures and data still contain spelling errors (e.g. 177 "Euryops" instead of Euryopis); other results may need to be checked. Taxon sampling strongly 178 affects inferred ancestral states. 179 Optimizing web presence/absence on these data ( Fig. 1A and S2) shows that webs are 180 ancestral with ~6 subsequent losses (ARD AICc = 135.7375). This optimization contradicts 181 F&al, whose results inferred no foraging web as the ancestral spider condition. Irregular aerial 182 sheet webs derived independently three times from "no foraging web," and brushed sheets 183 twice. Cob webs and stereotypical aerial sheets are the only architectures clearly derived from a 184 web building ancestor in F&al's optimization.

185
"Non-foraging silk-lined burrows" is a webless condition and consequently mygalomorphs 186 coded as such should have been scored as "no foraging web" under F&al's corrected character 187 state scheme. If the F&al scoring is modified to reflect that change (all taxa with state '8' receive 188 the webless score '6') spiders optimize as having no foraging web plesiomorphically (i.e., spider 189 are unequivocally primitively webless; Fig. S3, based on an ER model (AICc = 375.491). 190 Using our corrected character matrix ( architectures with four independent origins of the orb web (Fig. S5). 198 199

200
The most common current use of morphological characters is to map them on molecular trees 201 rather combining them with molecules to infer phylogeny. When 'absence' conflates primitive absence and secondary loss, as F&al did (e.g., "no foraging 218 web," "non foraging silk lined burrow," and "no web," all "not" hypotheses), such conflations, F&al emphasize that the orb web evolved three times. We show above that their own 258 character hypothesis, properly analyzed (and assuming the inclusion of Pararchaea), implies 259 four origins (taxon sampling matters!). Both the three or four origin results depend on coding "no 260 foraging web" as a real, observable state ("6") rather than as inapplicable data ("?"). When "no 261 foraging web" is coded as inapplicable data (Fig. 1B), the ancient origin hypothesis is sustained. 262 The webs of all extant orb weavers may be homologous as orbs. So much for orbs evolving 3-6 263 times. 264 Their more startling result is that ancestral spiders spun no webs and used no silk to 265 improve foraging success. Although their June Fig. 3A shows some probability of "non-foraging 266 silk lined burrow," (modified from "no foraging web" in April) and "irregular ground sheet," the 267 former is the same, scarcely disguised, methodological choice, and the latter includes factual 268 errors. 269 F&al apply "non-foraging silk lined burrow" exclusively to liphistiomorphs and 270 mygalomorphs. What does this mean? Both "no foraging web" and "non-foraging silk lined 271 burrow," share the notion of "non-foraging," presumably intentionally. If, therefore, all are coded 272 as "no foraging web," it persists as the ancestral spider condition, as in their April publication 273 (see Fig. S3

296
Reconstructing the evolution of spider webs remains an exciting yet unstable field of study: 297 not only the origin and evolution of webs, as such, but the origin of the iconic orb web. Given the 298 sensitivity of optimization algorithms to adjacent nodes, taxon sampling will always bedevil 299 conclusions. Other factors include the optimization algorithm used (especially the proper 300 handling of inapplicable/missing data), maximum likelihood rates of change among states, and 301 subjective disagreements about which conceptual state to apply to which observed web 302 architecture. That said, the notion that the single ancient origin hypothesis "crumbles under the 303 weight of additional transcriptomic data coupled with a significantly increased taxon sampling" is 304 premature, especially if based on a publication with as many irregularities as Fernández et al. 305 (2018). 306   Spider genomics and NGS sequencing technologies may presage stable phylogenetic trees  307  for spiders, but they are just beginning to influence fundamental questions about web  308 construction, its underlying genetics, and the emergent phenotype of web architecture. Rather 309 than homologizing whole web architectures, we recommend a more reductionist approach more 310 likely to accommodate new taxa and data. For example: 1) silk use in prey capture; presence or 311 absence of 2) ampullate, 3) piriform, 4) aggregate, 5) flagelliform and 6) cribellate silks; 7) web 312 location and attachment points; 8) prey locomotion (such as web 'designs' focused towards 313 aerial vs pedestrian prey); 9) refugium location; 10) architectural elements (such as disordered 314 vs patterned, ordered or stereotypical); 11) pattern type (for example 2D vs 3D), and more. This 315 approach avoids arbitrary coding of whole webs as loosely defined conglomerate homology 316 hypotheses, and could allow hypotheses of web architectures to emerge from nuclei of 317 concordant, more objective homology hypotheses. 318 Regardless, our best efforts to reanalyze data on web architecture variation in spiders, 319 including careful attention to the treatment of "absence" or inapplicable/missing data, suggests 320 that the ancient single origin of the orb web is feasible. Orb did not originate 3-6 times, and 321 spider webs did not originate 14 times. Their ancestor spun a web. These results, after all, just 322 reinforce prevailing views regarding the evolutionary history of spider webs. They do illustrate 323 the pitfalls of disregarding long accepted rules for coding homology and mis-coding of "absence" 324 characters, in particular. While we do not make the claim that a multiple origin hypothesis is 325 false, we strongly disagree with assertion that a single origin hypothesis has been falsified, let 326 alone that it has 'crumbled' under the force of evidence.