Today, few would argue that all phenotypes conceivably derived from a given ancestral form are equally likely. The evolutionary synthesis of the 20th Century emphasized "random" mutation, with evolutionary directionality and diversification imposed exclusively by selection (albeit, for some workers, modified by epistasis and pleiotropy), but it is now clear that developmental systems impose uneven probability distributions on the phenotypes accessible from a given evolutionary starting point (Arthur, 2004;Maynard-Smith et al., 1985;Sears, 2014;Uller, Moczek, Watson, Brakefield & Laland, 2018), a phenomenon now termed developmental bias. Such intrinsic factors, governing the production of variation before extrinsic sorting by the biotic and abiotic environment, are likely to be a key component of phenotypic evolution, and this probabilistic view is a significant advance over many conceptualizations of developmental constraints. Here, I present some paleobiological and macroevolutionary perspectives on developmental bias, and the potential for empirical assessment of the expression and consequences of developmental bias across large spatial and temporal scales, and hierarchically above the species level. This enterprise has just begun, but its potential can be seen in the diversity of studies discussed below, which are scattered across the tree of life and through geologic time. Those studies were conducted on different groups using different methods and hierarchical levels, so that generalization is premature in many instances, as I will note at various points. Some will probably not stand up, indeed some seem mutually contradictory, but collectively they provide the seeds of an integrative research program.
In the macroevolutionary arena, over large spans of evolutionary time or broad swaths of phylogeny, only phenotypes are usually accessible for study, with little recourse to the array of experimental methods outlined by Uller et al. (2018). Comparisons between direct observations on developmental biases and large-scale evolutionary patterns are the strongest lines of evidence (e.g., Kavanagh et al., 2013 on vertebrate digits, Houle, Bolstad, van der Linde, & Hansen, 2017 on Drosophila wings, Watanabe, 2018 on bird limbs), but such studies are rare, and most developmental model systems are taxa lacking the kind of fossil record required to address macroevolutionary hypotheses by integrating paleontological and neontological data. Macroevolutionary analyses therefore often begin with data on phenotypic variation at a series of inferred evolutionary starting points within a clade. Analytical strategies range from quantifying the suite of discrete character state transformations downstream of a node in a phylogeny, to ordinating continuous variation within an empirical multivariate morphospace, to occupation patterns within a theoretical morphospace, to calculating the patterns of maximum phenotypic variance in a variance-covariance matrix (a P-matrix) that define an "evolutionary line of least resistance," a phrase that has become a useful shorthand for the positive dimension of developmental bias (versus a strict definition of constraints as barriers). These approaches essentially aim to make predictions across scales and hierarchical levels. That is, to test whether inhomogeneity, that is anisotropy, of variation within and among populations or species is manifest in longer-term evolutionary trajectories of clades or higher taxa, and how such anisotropy interacts with the dynamics of clades to shape their deployment in morphospace. This phenotypic approach is less than ideal, as measures of observed variation can only partially reflect the full variability of the underlying developmental systems (G. P. Wagner, Pavlicev, & Cheverud, 2007;Uller et al., 2018), and phenotypes can be absent or rare within clades for extrinsic as well as intrinsic reasons. As discussed below, a remarkable number and variety of purely phenotypic studies have detected significant patterns consistent with a role for developmental bias, suggesting that morphology alone can indeed capture key aspects of relation between development and macroevolution. However, the conceptual contrast between the inhomogeneous, potentially discontinuous, accessibility of phenotypes that stems from the nonlinear relation between genotypes and phenotypes (e.g., Alberch, 1980;Polly, 2008), and the observed phenotypic covariance structure, which may say little about the malleability or discontinuities of that structure (see Polly, 2008Polly, , 2017)), must always be kept in mind, and deserves far more empirical exploration.
2.1 | Observations on extant and fossil populations Some phenotypic analyses have been strikingly successful in predicting macroevolutionary patterns from data on intraspecific variation, or from short-term observations on ontogenetic allometry. For example, within ungulates, the direction and magnitude of evolutionary change in skull morphology in bovids and cervids significantly aligns with intraspecific (adult) variation for each clade (Haber, 2016). Bovids have diversified in a wider range of directions in morphospace than cervids and have achieved greater overall disparity, consistent with the relative strengths of intraspecific covariation in the two clades. This point is important: tests for the role of developmental bias should address not just the orientation of the main axis of (co-)variation (again, sometimes termed an evolutionary line of least resistance) but also the volume and eccentricity of the multidimensional envelope around that axis, that is the concentration of variation around the main axis, with the expectation that weaker eccentricity imposes a less strongly biased probability distribution on future evolutionary directions. In ungulates at least, both the direction of variation and its eccentricity had predictive power for among-species differences.
Using a model more directly tied to quantitative genetics, Marroig and Cheverud (2005) similarly found that primate genera tend to evolve along an allometric axis of covariation, and simulations showed that evolutionary responses should be 3-4 times weaker when selection is not aligned with the line of least resistance. Hansen (2012;219) lists other examples, at a variety of taxonomic levels, where patterns of species divergence align with axes of maximum phenotypic or quantitativegenetic covariation (and see Bolstad et al., 2014). However, no one has really come to grips with the finding that deviations from apparent lines of least resistance are not definitive evidence for the absence of developmental bias (Hansen & Voje, 2011), so that negative results may be difficult to interpret.
Analyses of extant populations within a phylogeny are clearly valuable, but make strong assumptions regarding ancestral character states that may be distant from branch-tip phenotypes. Inferences of ancestral states from extant taxa have repeatedly proved to be positively misleading when tested against fossil data or simulations. Such errors probably arise, at least in part, because these methods assume that morphological transitions are homogeneous and isotropic (Gerber, 2014); greater accuracy-often with different estimated ancestral states-comes with the inclusion of even sparse fossil data (reviewed by Hunt & Slater, 2016, see also Melstrom & Irmis, 2019;Mitchell, 2015;Mitchell et al., 2019). Of course, paleontological analyses are especially powerful in systems that permit large sample sizes and fine temporal resolution. Although necessarily accessing phenotypes, the ability to place phenotypic divergences closer in time to source populations and their properties allows powerful comparisons with expectations from the hypothesized role of developmental bias (Figure 1a). In an exemplary study, evolutionary changes in the Eocene-Recent ostracode Poseidonamicus tended to occur, and were generally faster, in phenotypic directions close to the axis of maximum variation (Hunt, 2007). Similar results in Neogene and Pleistocene rodent lineages (Firmat et al., 2014;Renaud, Auffray, & Michaux, 2006) underscore the potential for morphometric studies of well-sampled fossil populations to make rigorous predictions for the evolutionary dynamics that shape those populations and their deployment in morphospace in the recent geologic past. The next step will be to probe the extant descendants experimentally to confirm the developmental underpinnings of the evolutionary pattern.
We get a step closer to developmental factors by analyzing ontogenetic patterns, although the basic caveat regarding realized variation versus potential variability still applies for nonexperimental data. But again, despite such caveats, macroevolutionary patterns often correspond to ontogenetic trajectories, suggesting that the probability of phenotypic change indeed tends to be biased towards those trajectories in morphospace. For example, the major directions of avian limb diversification are correlated with clade-specific ontogenetic patterns, corroborating predictions from a developmental-bias hypothesis, but also showing once again that developmental patterns can be reorganized, apparently with significant evolutionary consequences (Watanabe, 2018).
A much larger set of examples might be drawn from situations where the probability of evolutionary change has been biased along existing developmental trajectories. These can involve alterations in developmental rate or timing, whether (a) global heterochrony where overall somatic development shifts relative to reproduction, (b) sequence heterochrony where discrete development events are re-ordered, or (c) local heterochrony, where the rate or timing of development of a trait or suite of traits is altered relative to the rest of the body. The potential link between phenotypic heterochrony and developmental bias has been somewhat obscured by the fact that heterochrony can occur independently at organismic and molecular levels, leading to considerable definitional confusion. For example, the repeated origins of paedomorphic, permanently aquatic salamanders, do not involve changes in the timing of gene expression, but the severing of a hormonal pathway at various points (Johnson & Voss, 2013), and conversely a heterochronic delay in scale-cell maturation can produce a black, nonpaedomorphic butterfly wing phenotype (Koch, Lorenz, Brakefield, & ffrench-Constant, 2000). Thus, parsing heterochronic transitions across hierarchical levels can clarify the nature of the evolutionary changes, and potential examples are plentiful (see Jablonski, 2017a for a sampling).
An intriguing macroevolutionary indicator of the channeling of evolutionary change along ontogenetic lines is seen in paedomorphoclines and peramorphoclines recorded in the fossil record. The Cenozoic brachiopod Tegulorhynchia appears to have invaded increasingly shallow-water habitats in stepwise fashion via successively more paedomorphic species, and the Cretaceous hearturchin Micraster appears to have burrowed deeper into sediments by extending ontogenetic trends via successively more peramorphic species (McNamara, 2012 andSolovjev, 2015, who both provide additional examples, see also Gerber, 2011;Monnet, Klug, & De Baets, 2015;Echevarria, 2016). Many of these studies require more formal phylogenetic and morphometric analysis, but the phylogenetic breadth of the potential examples is impressive. The temporal data for these fossil occurrences appear to corroborate the largely stepwise, directional nature of these changes, and many involve coordinated changes of multiple traits along ontogenetic lines as expected under global heterochrony. These properties suggest evolutionary systems whose uneven distribution of transition probabilities played out over a series of speciation events over tens of millions of years.
Most evolutionary transitions, even those that are clearly heterochronic in the sum of their traits, also involve autapomophies or other indicators of mosaic evolution (Webster & Zelditch, 2005). This complexity does not falsify an evolutionary role for slowing or extending somatic development relative to reproductive maturity, but it does make detection less straightforward,
and different modes of developmental change may actually be more frequent. Tinkering at finer scales, for example by changes in the developmental rates of individual structures or developmental modules, or by shuffling developmental stages, or by heterotopic expression of gene networks at new locations within the body, all likely represent evolutionary changes more probable than the de novo production of complex traits. Local heterochrony is often manifest postembryonically and so can often be detected paleontologically (see examples in Jablonski, 2017a;Urdy, Wilson, Haug, & Sánchez-Villagra, 2013), and some have occurred repeatedly. For example, local paedomorphic reduction of ambulacral rows in echinoderms was evidently instrumental in the origin of many of the peculiar body plans of the early Paleozoic by providing an easy means of altering feeding and respiratory structures (Sumrall & Wray, 2007). In contrast, sequence heterochronies are well-documented (e.g., Bininda-Emonds, Jeffery, & Richardson, 2003;Bininda-Emonds, Jeffery, Coates, & Richardson, 2002;Carril & Tambussi, 2017;Ito, Matsumoto, & Hirata, 2019;Koyabu et al., 2014;K. K. Smith, 2001), but are most often embryonic and so are difficult to detect paleontologically. Nonetheless, macroevolutionary consequences can still be evaluated in a comparative framework when changes in the sequence of developmental events can be pinpointed phylogenetically. For example, evolutionary changes in ossification sequences may have limited or promoted access to particular regions of morphospace in marsupials versus eutherians (Weisbecker, Goswami, Wroe, & Sánchez-Villagra, 2008) or in parrots versus passerines (Carril & Tambussi, 2017).
Local and sequence heterochronies are such accessible pathways for evolutionary change because development is modular, and modularity, with covariation strong within modules and weaker among them, has long been recognized as a likely source of developmental bias. It is often treated as the converse of integration, although some care is needed in that an organism can have weak integration but lack discrete modules. For a given structure or region of the body, stronger developmental integration and weakly defined modules are generally held to impede and channel evolutionary change, whereas weaker integration and more sharply defined modules appear to afford greater freedom to evolving lineages, in the sense that among-module changes are facilitated (while within-module ones will be hindered; see Goswami, Binder, Meachen, & O'Keefe, 2015;Goswami, Smaers, Soligo, & Polly, 2014;Porto, Shirai, de Oliveira, & Marroig, 2013;Sears, 2014;Uller et al., 2018;and Armbruster, Pélabon, Bolstad, & Hansen, 2014, who view allometry as a special case of developmental integration). Weakening of integration and/or strengthening of modularity should thus tend to reduce the anisotropy of developmental bias by allowing evolution to occur in more directions, by promoting rate differentials among traits and modules, and perhaps by facilitating cooption of regulatory networks to build novel structures (Raff, 1996;G. P. Wagner et al., 2007;Espinosa-Soto & Wagner, 2010;Jablonski, 2017a;Larouche, Zelditch, & Cloutier, 2018).
The macroevolutionary consequences of among-clade differences in integration have mainly been explored theoretically and through simulation (e.g., Goswami et al., 2014;Marroig, Shirai, Porto, de Oliveira, & De Conto, 2009;Felice, Randau & Goswami, 2018), although considerable progress has been made in understanding the mammalian skull in particular. However, there is little consensus on the pattern or strength of integration and modularity in most groups, with seemingly contradictory findings that remain unresolved (Felice et al., 2018). These difficulties probably arise, not just from differences in analytical approaches, but because the developmental modularity of an organism need not correspond precisely with the modularity of the gene regulatory networks that orchestrate development, with the phenotypic modules defined by the covariation structure among conspecific adults, or with the evolutionary modularity observed across phylogeny (Hallgrímsson et al., 2009;Klingenberg, 2010Klingenberg, , 2014)).
Measurement of evolutionary change and its correspondence to within-population variation. Ancestor and descendent populations are represented by ellipses having specific volume, orientation, and eccentricity in morphospace, with evolutionary change occurring along a vector forming an angle (θ) with the axis of maximum phenotypic variance (p max ). After Hunt (2007), who found that evolutionary changes in the fossil ostracode Poseidonamicus were significantly aligned with P max in ancestor-descendent transitions. (b,c) Evolution of post-Paleozoic echinoids within a phylomorphospace based on principal coordinates analysis of a character matrix, from Hopkins and Smith (2015), with speculations on the orientation and eccentricity of developmental biases in the major clades, taking the evolutionary trajectories at face value. Black lines show phylogeny for the entire history of the clade, with morphospace occupation during the Triassic and Jurassic (b), and among living echinoids (c) indicated by colored circles. Cool colors mark the "regular" echinoids: cyan, Cidaroida; blue, regular euechinoids. Warm colors mark irregular echinoids: red, Neognathostomata (sand dollars and related forms); orange, Atelostomata (heart urchins and related forms). Note differences in the magnitude of change and tightness of trajectory of disparity through time and across phylogeny; (b,c) by kind permission of the authors; inset photos from A. B. Smith and Kroh (2011), © The Trustees of the Natural History Museum, London, licensed under the Open Government Licence [Color figure can be viewed at wileyonlinelibrary.com] Empirically, strong integration shows the expected association with both low disparity and low evolutionary rates in the avian skull, but only with low disparity in the mammalian skull (Bright, Marugán-Lobón, Cobb, & Rayfield, 2016;Goswami et al., 2014;Felice & Goswami, 2018). The rate differences are not surprising: even closely related lineages can differ in their tendency for unconfined Brownian motion versus equally rapid change that careens within a restricted portion of morphospace, although the relative roles of intrinsic and extrinsic factors may vary (e.g., Hopkins & Smith, 2015;Santos, Perrard, & Brady, 2019;Sidlauskas, 2008). The inverse association of integration and disparity does suggests a large-scale, macroevolutionary role for integration in at least some instances despite evidence that patterns of developmental integration are often unstable over longer timescales and across broad phylogenetic distances.
At present, we do not understand that interplay between the conservation of patterns of integration and modularity, as seen in some of the best-studied systems such as mammal skulls and limbs (Goswami et al., 2014), and the repeated disruption of that integration in, for example, taxa adopting novel locomotory models such as bipedalism and flight (Young, 2017;Young, Wagner, & Hallgrímsson, 2010). More work is needed on determining how resistant any given system of modules has actually been to selection, and therefore how often it actually impeded and channeled evolutionary trends. Even if relatively fluid on long timescales, modularity should still facilitate coordinated evolution among traits within a module and so promote evolution in particular directions at the species level; even the Bicyclus experiments of Beldade, Koops, and Brakefield (2002) found a less linear and more irregular response to selection in the direction of phenotypes absent from natural populations (Melo, Porto, Cheverud, & Marroig, 2016).
A complementary approach tracks the direction and magnitude of morphological evolution in a multidimensional space (Erwin, 2007(Erwin, , 2017;;Foote, 1997Foote, , 1999;;Polly, 2008;Wagner, 2010;Chartier et al., 2014). Empirical morphospaces are defined from the sample of existing forms, living or fossil, included in the analysis. Incorporating a phylogenetic hypothesis to create a phylomorphospace provides an even more effective tool for comparative analyses of macroevolutionary pathways and differential limitations on clade behavior (see Gerber, 2019 andLloyd, 2018 for reviews). For example, the major post-Paleozoic echinoid clades differ strikingly in their evolutionary histories. The two "regular" groups, closer to the basal condition, are taxonomically rich but tightly limited within the morphospace, whereas the two more derived groups of irregular urchins depart from the ancestral portion of the space in two distinct directions: the heart urchins (Atelostomata) follow a narrow path along a well-defined trajectory to produce forms, whereas the sand dollars (Neognathostomata) generate comparable disparity but are less linear in their deployment (Hopkins, 2016;Hopkins & Smith, 2015; Figure 1b,c). This study presents a wealth of conceptual and analytical opportunities: do these contrasting movements through morphospace correspond to intraspecific (co)variation or developmental patterns (as in the hypothetical cartoons in Figure 1), or do they break the patterns exhibited by the ancestral forms? The patterns described here are derived from discrete-character matrices, would the same patterns emerge from continuous characters? Do developmental differences, for example relative eccentricities around the major axes of phenotypic variation, allow us to predict the seemingly isotropic evolutionary transitions in the two "regular" clades, and more directional, divergent evolutionary paths of the two irregular clades? The shift from the "regular" body plan is thought to have been instigated by a fundamental change in growth strategy, from one involving plate addition to one involving plate enlargement that facilitated regional differentiation of test morphology (A. B. Smith, 2005), including a shift from radial to bilateral symmetry and from surface-dwelling to burrowing lifestyles. Despite decades of superb work on Strongylocentrotus (Peter & Davidson, 2015), we know almost nothing about the developmental underpinnings of that key transition, or their downstream macroevolutionary consequences. Similar questions could be asked of virtually any wellsampled large-scale phylomorphospace, particularly those that are time-stratified, so that character states at nodes are more reliably characterized using a rich fossil record. As with taxon-based phylogenies, phylomorphospaces founded entirely on extant taxa face the uncertainty and error inherent in estimating ancestral character states, that is internal nodes of the phylogeny, from branch tips.
Phylomorphospaces provide a framework for assessing parallel evolution, which has also been viewed as a macroevolutionary expression of developmental bias, with the parallel diversifications of East African cichlids often cited as the prime neontological example (e.g., Brakefield, 2011;Uller et al., 2018). Parallelism can be defined as the evolution of similar phenotypes from similar ancestral forms (whereas convergence is the evolution of similar phenotypes from disparate ancestral forms; see Pearce, 2012;R. Powell, 2012;Stayton, 2015). As with heterochrony, parallelism and convergence are best treated hierarchically: although genetic and phenotypic transitions often coincide (e.g., Conte, Arnegard, Peichel, & Schluter, 2012;A. Martin & Orgogozo, 2013;Reno, 2014), phenotypically disparate lineages can converge by similar (i.e. parallel) genetic changes, and phenotypically identical populations can use different (i.e. nonparallel) genetic changes to generate similar descendant phenotypes in parallel (Elmer & Meyer, 2011;Sturm & Duffy, 2012;Agrawal, 2017;A. Martin & Orgogozo, 2013;Rosenblum, Parent, & Brandt, 2014). These contrasting routes at different hierarchical levels, along with the empirical difficulty of determining similarity in starting points and trajectories, have prompted some to reject the distinction between convergence and parallelism (Arendt & Reznick, 2008;Manceau, Domingues, Mallarino & Hoekstra, 2011;Blount, Lenski, & Losos, 2018). However, the conceptual distinctions are clear in a hierarchical framework, and phylomorphospaces, particularly those incorporating fossil data close to evolutionary branchpoints, provide a vehicle for distinguishing the phenotype-level categories (e.g., Jablonski in Pearce, 2012).
Similar to Hunt's approach to species-level transitions (Figure 1a), Bolnick, Barrett, Oke, Rennison, and Stuart (2018) urge an approach to parallelism and convergence based on the angle between evolutionary vectors in phylomorphospace, quantifying what must be a spectrum rather than a simple binary distinction (see also Oke, Rolshausen, LeBlond, & Hendry, 2017). Findings may vary among situations and clades, but there are a few hints on how that variation might play out. For example, parallel phenotypic evolution is less frequent among stickleback traits with many-to-one mapping of form to function than for traits with one-toone mapping (Thompson et al., 2017). Here, selection can operate on emergent functional properties while the underlying traits are relatively free to vary (Bolnick et al., 2018). In contrast, extant damselfishes have apparently shuttled repeatedly between three main ecomorphological types without branching into novel feeding strategies such as durophagy (Cooper & Westneat, 2009); here the issue is the relation between biomechanical factors and developmental ones, making the group a potential focus for closer analysis (and see Muñoz, Hu, Anderson, & Patek, 2018 for a more general exploration of how mechanical sensitivity might bias morphological evolution). As noted in several overviews, nonparallel responses by similar phenotypes to similar selective pressures require serious consideration (Bolnick et al., 2018;Langerhans & Riesch, 2013;Oke et al., 2017)-several explanations are possible, and an exploration of these apparent counter-examples to developmental bias hypotheses will sharpen understanding of bias and, perhaps, its limits or loopholes.
In contrast to empirical morphospaces, theoretical morphospaces are derived directly from geometric or mathematical models of form, providing a more powerful framework for visualizing and interpreting the uneven distribution of realized forms through time and across phylogeny (reviewed by Erwin, 2007Erwin, , 2017;;McGhee, 2007McGhee, , 2015;;Polly & Motz, 2016). Such morphospaces are valuable tools, but the parameters that define the space may only distantly resemble the developmental mechanics that generate the forms under study (as in the seminal work on shell coiling reviewed by McGhee, 2007). Further, axes may have different units and may not be strictly orthogonal, so that spatial relationships within the space can be non-Euclidean and thus nonintuitive (Erwin, 2017;Gavrilets, 2003;Gerber, 2017;Mitteroecker & Hutteger, 2009;Polly & Motz, 2016). Nonetheless, theoretical morphospaces present rich opportunities to assess variation in the frequency of realized forms within the context of morphologies never attained by a focal clade or attained by just one among several clades, especially when considered from an explicitly developmental perspective.
Developmental morphospaces (Eble, 2003) use more realistic generative parameters to build the space and track evolutionary patterns. These morphospaces are far more difficult to produce and have yet to be constructed for many clades. Successes have largely come from taxa with accretionary skeletons, such as mollusks, brachiopods, and echinoids (e.g., Chakra & Stone, 2011;Chirat, Moulton, & Goreily, 2013;Erlich, Moulton, Goriely, & Chirat, 2016;McGhee, 2007;Saucède et al., 2015). However, the most productive effort has been the computational model for mammalian teeth based on the genetic and cellular interactions that underlie development (Salazar-Ciudad & Jernvall, 2010). This Inhibitory Cascade model successfully predicts-with a few lapses-diversification patterns in tooth size and shape, including extinct taxa with dental patterns absent in the modern biota, and it predicts the developmental changes likely to be responsible for observed patterns (Jernvall, 2000;Kavanagh, Evans & Jernvall, 2007;Polly, 2007;Halliday & Goswami, 2013;Harjunmaa et al., 2014;Burroughs, 2019; see also Evans et al., 2016 andOrtiz, Bailey, Schwartz, Hublin &Skinner, 2018 for applications to hominid evolution, with a quantitative genetic extension by Hlusko, Schmitt, Monson, Brasil, & Mahaney, 2016;and Renvoisé et al., 2009 for a different model that accounts for the anomalous molar proportions of arvicoline rodents). Because the model successfully explains dental patterns of stem mammals and early crown mammals from the Mesozoic (Halliday & Goswami, 2013), its failure to predict certain derived forms such as horses is less likely a failure of the model than an indication of genuine changes in the developmental system. The model also shows that some derived morphologies can be produced by a variety of different developmental changes, which may explain the repeated evolution of, for example, the hypocone in several clades (Jernvall, 2000), another case of parallel phenotypic evolution without parallel genetic changes but still mediated by developmental bias. And it drives home the more fundamental point that incremental changes in developmental parameters can produce discontinuous or nonlinear transformations in morphospace, as when cusps are gained or lost (see also Polly, 2017).
The explanatory power of the Inhibitory Cascade model is impressive, and demonstrates how fossils, and non-model organisms generally, can play a direct role in evaluating the macroevolutionary utility of developmental models. Similar opportunities for modeling may exist in birds, where beak shape in at least some modern passerines is set by interactions among a few key genes in a surprisingly simple coordinate system (Fritz et al., 2014;Mallarino et al., 2011). Developmental models for the gene networks rooted in that system may help to explain the uneven distribution and gaps in present-day beak morphospace shown by Cooney et al. (2017); but see Mallarino et al. (2012), who find a different regulatory system in bullfinches.
Quantitative genetic approaches have been less explored from this perspective, but an analysis of primate dentitions, informed by the Inhibitory Cascade model, detected what is in a sense the converse of developmental bias-two complex dental traits, each separately patterned by gene expression cascades, that despite close proximity within the jaw are effectively free to evolve independently (Hlusko et al., 2016). This evolutionary freedom was confirmed morphometrically, with fossil hominoids falling into regions of morphospace not occupied today. G-matrices, the workhorse of evolutionary quantitative genetics, may lack the necessary stability at macroevolutionary scales for such analyses to be generally applicable (for a variety of perspectives see, e.g., Björklund, Husby, & Gustafsson, 2013;Eroukhmanoff, 2009;Jones, Bürger, Arnold, Hohenlohe, & Uyeda, 2012), and they do not incorporate the nonadditive genetic interactions inherent to developmental processes (e.g., Milocco & Salazar-Ciudad, 2019;Polly, 2008;Rice, 2004). As with the operational substitution of variation for variability, the predictive power of the dentition study, despite these uncertainties, is promising.
Developmental models can provide fresh insights even when the molecular basis for development is unknown. In an intriguing hybrid approach that should be emulated for other clades, Gerber (2014) started with an ancestral fossil species, with a known ontogenetic trajectory, and generated a full array of hypothetical descendants from different local alterations of ancestral morphology, in effect exploring a relatively simple generative model for trilobite head evolution. He predicted that heterochrony would define an evolutionary line of least resistancethat is the most accessible or probable evolutionary pathways-such that new species would fall along the ancestral ontogenetic trajectory, and found just that in the trilobite Zacanthopsis. A much larger suite of species within this clade is needed to test this hypothesis definitively, but the approach would be worthwhile for any taxon where the ancestral ontogeny is known.
These and many more morphospace analyses underscore two neglected considerations. First and most obviously, proximity in morphospace need not indicate evolutionary accessibility-developmental mechanisms impose anisotropy (and the geometric pitfalls noted above must also be understood). Second, the density of species in morphospace need not reflect the peaks and valleys of an adaptive landscape, because differential origination can occur for many reasons, from developmental bias to the hitchhiking of traits on factors that promote or damp the origination or loss of taxa. For example, in the largest family of marine bivalves (Veneridae), the evolutionary trajectories of genera from their oldest species to the present day were unrelated to the overall density distribution of species within the morphospace: the most heavily occupied neighborhood was not an evolutionary attractor (Huang, Roy, & Jablonski, 2015, see Roy, Jablonski, & Martien, 2000 on similar dynamics in molluscan body sizes). Instead, the density of species within the morphospace was evidently the product of differential extinction. (Hulsey, Alfaro, Zheng, Meyer, & Holzman, 2019 also found little evidence that cichlid jaw morphology evolved towards inferred optima in morphospace.)
There is much we do not know about the relation between clade behavior in morphospace and developmental bias. We need comparative analyses, conducted in a common framework, as initiated by pioneers such as Raup (1966) and Foote (1997). One approach would be to partition clades using diversity-disparity plots, which are time series plotting the morphospace occupied by a clade against its standing taxonomic diversity (Jablonski, 2017b, also Wright, 2017). Such a partition would permit testing, for example, whether rapid expansion in morphospace in concert with a rapid rise in species numbers involves the same kinds of evolutionary changes and developmental influences as genuine early bursts in morphology where disparity greatly outstrips the proliferation of species and so appears to involve greater evolutionary invention per cladogenetic event. At the same time, unexplained regularities in the origin of higher taxa, for example preferentially in nearshore habitats and in the tropics for marine invertebrate orders (Jablonski, 2005b(Jablonski, , 2010; P. R. Martin, Bonier, & Tewksbury, 2007) and early vertebrate clades (Sallan, Friedman, Sansom, Bird, & Sansom, 2018, who unfortunately exaggerate differences with the invertebrate patterns), have yet to be convincingly linked to ecological or geographic patterns in either developmental or selective factors.
The present-day biota, astonishing in its taxonomic and phenotypic diversity, almost always shapes expectations on the phenotypes evolutionarily accessible to a clade or body plan. However, at least 95% of the multicellular species that ever lived are extinct (Sepkoski, 1992), and among those extinct taxa are many unexpected forms. These extinct taxa inform us on ancestral states of developmental systems, and shift our understanding of what lineages can do (Marshall, 2017;Organ, Cooper, & Hieronymus, 2015;Raff, 2007;Sánchez, 2012;Urdy et al., 2013;Wilson, 2013). The fossil menagerie is almost endless but includes echinoids with periscope-like test extensions, tightly coiled and helical oysters, uncoiled, helical, and spiny nautiloids, sharks with coiled tooth batteries, six-fingered tetrapods (basally and with a second derivation in ichthyosaurs), crocodilians with bulbous shell-crushing teeth, and giant ground sloths including aquatic ones (for sources see Amson, Argot, McDonald, & de Muizon, 2015;Caldwell, 2002;Coates & Clack, 1990;Gale & Smith, 1982;Oyston, Hughes, Wagner, Gerber, & Wills, 2015;Seilacher & Gishlick, 2014;Tapanila, Pruitt, Wilga, & Pradel, 2018;Vermeij, 2015;Westermann, 1998).
Bivalve mollusks have become a model system for macroecology and macroevolution but are barely touched in terms of evolutionary developmental biology. To take just one direction, the evolutionary dissociation of right and left valves to break bilateral symmetry has produced a wide array of extreme morphologies, many of which are extinct. The Cambrian Bivalvia were equivalved, and most species over the past half-billion years have retained that state. However, the first inequivalved taxa appeared in the Early Ordovician during the initial diversification of the group (Cope & Babin, 1999), and this dissociation has occurred repeatedly in multiple lineages since then, with shifts in shell geometry not seen in extant forms (Seilacher & Gishlick, 2014). Many modern oysters today are inequivalve, but the inequality tends to be in the relative convexity of the valves, quite unlike the planispiral or helical oysters seen in the Mesozoic (e.g., Gryphaea and Exogyra, respectively), in which both the coiling geometry and relative size of the two valves also differ dramatically. The most striking example of this leftright dissociation, accompanied by extensive repatterning of other parts of the shell and internal anatomy, may be the Jurassic-Cretaceous rudists, a phenotypically diverse tropical group that included lineages with a conical lower valve and a cap-shaped upper one that was raised vertically like an elevator (by just a mm or two) rather than opening like a hinged trap door as in virtually all other bivalves (Skelton, 1979; Figure 2). (But see also the bizarre Cretaceous clavagellid bivalve Ascaulocardium, another candidate for "the ultimate variation on the bivalve paradigm" [Pojeta & Sohl, 1987].) The developmental biases set by breaking bilateral symmetry in the rudists and other bivalves almost certainly channeled their subsequent evolution, and it may be no coincidence that many inequivalve clades show higher per-taxon extinction rates than the overall group average. Also within mollusks, the density distribution of gastropod shell form in a Raupian (or any other geometric) morphospace has changed significantly through the Phanerozoic, probably driven at least in part by extrinsic factors such as changes in the nature and intensity of durophagous predation (Vermeij, 1987(Vermeij, , 2013)). Little is known of the role of developmental factors in this long-term shift in morphospace occupation, but among major clades, vetigastropods (Trochus, Haliotis) have evidently never explored the range of forms seen today in co-occurring "neogastropods" (Murex, Conus; e.g., Vermeij, 2015), presenting an ideal system for a comparative analysis of the history and developmental basis of shell geometries. As noted above, increasingly realistic generative models for shell deposition and growth are available for exploring observed patterns in their developmental context.
The fossil record is far more than a bestiary of putatively impossible or improbable forms. It provides character states at the base and internal nodes of phylogenies, fossil ontogenies and even in some cases fossil embryos, and, when sampling and other uncertainties are taken into account, a wealth of data on the rate and direction of evolution in form and in underlying developmental processes. Contouring the density of morphospace occupation over the entire history of a clade may permit quantification of transition probabilities in considerable detail, even when outliers are extinct and so not subject to neontological inventory or experiment. And of course even clades with the richest fossil records show perpetually vacant regions of morphospace. The challenging question remains as to whether those vacancies represent absolute developmental constraints, Macroevolution-evolution above the species level and, for present purposes, the deployment of clades through morphospace and how it has changed over timeintegrates the origin of genealogical units and their differential survival and proliferation. One serious shortcoming of many recent treatments of the "extended synthesis" has been the neglect or omission of clade dynamics, that is the multilevel processes in sorting variation at, above, and below, the species level, and their effects at different temporal and spatial scales. Biodiversity in all its dimensions or currencies cannot be understood without integrating the generation of raw material with its subsequent sorting (Gould, 2002;Jablonski, 2007Jablonski, , 2017a,b),b). Synthesis requires that we understand not only how a complex mammal tooth or bivalve hinge was generated from a simple one, and why some complex forms are developmentally more accessible than others, but also how such forms are associated, causally or passively, with the dynamics of the species and clades that bear them. We have scattered clues and a few examples, mostly from the fossil record, but the synthesis between evo-devo and multilevel clade dynamics has barely begun.
Clades can accumulate, lose, and shift morphospace occupancy in several ways (Gould, 1982(Gould, , 2002;;Jablonski, 2017b;Stanley, 1979). Clade dynamics may often involve mixtures of these processes, but it is conceptually useful to recognize three basic modes: speciation bias, extinction bias, and direction bias (Figure 3).
A clade's ability to occupy and move through morphospace over long timescales depends not just on the variability of its constituent species-the suite of phenotypes it can most readily expose to selection and other forces-but on its inherent origination and extinction rates, and how those rates vary across a given phylogeny. Short-lived lineages, or lineages that fail to branch, are likely to generate less disparity through time, or to capitalize on an advantageous novelty, than lineages having longer durations and high rates of cladogenesis (e.g., Futuyma, 2015;Jablonski, 2017b;Rabosky et al., 2013; note that this observation is divorced from the relation between present-day species richness and disparity, which can be decoupled for many reasons). This is not a circular statement: over geologic time most species tend to be static or to follow nondirectional random walks (Hunt & Rabosky, 2014), with speciation evidently helping to partition and stabilize existing, transient variation, thereby allowing those variants to find their own independent trajectories (Futuyma, 1987(Futuyma, , 2015;;Gould, 2002). Thus the evolutionary consequences of intrinsic, anisotropic variation at the organismal level depend in part on other intrinsic properties of species that set origination and extinction rates, from body size to genetic population structure and geographic range size (Jablonski, 2008a(Jablonski, , 2017b)). Even a strongly directional tendency in evolutionary transitions will have little cumulative effect, and may be undetectable on macroevolutionary scales, if origination rates are low or extinction rates are high in the relevant clade.
Of course, extinction risk is set by both organismal and species-level factors. Intuitively, greater intraspecific variation can confer extinction-resistance, directly (more raw material for response to selection) or indirectly (broader and thus more extinction-resistant geographic range, which might promote or be promoted by greater variation). Both paleontological and neontological evidence corroborates this association between intraspecific variation and taxon survival (Forsman & Wennersten, 2016;Kolbe, Lockwood, & Hunt, 2011;Liow, 2007), but potential counterexamples exist (Hopkins, 2011, but see Webster, 2019). Still unknown is whether the shape of the variational envelope around an ancestral phenotype-its eccentricity or direction, that is developmental bias in morphospace-influences species or clade survivorship.
We also need to know whether clades that expand along the axes of developmental biases (an occurrence that might represent a fortuitous alignment of available variation and selection) incur some macroevolutionary advantage: for example produce more species, resist extinction, shift more rapidly through morphospace, encompass a greater volume of morphospace, or encompass a greater geographic or environmental range. Equally, we need to know whether clades pushed by selection in directions oblique or orthogonal to developmental biases produce fewer species, become more extinction-prone, and so forth, perhaps suffering a greater evolutionary cost, or put at a greater disadvantage, because favorable variation is less readily forthcoming. Alternatively, are clades where variation is more isotropic able to respond to a greater variety of environmental challenges and so have lower extinction rates and higher diversification rates-and is this a reasonable gauge of evolvability? The association, at least, is testable in the fossil record.
Such intrinsic factors are often likely to be important, but extrinsic factors also enter into differential origination and extinction. For example, a phenotypic trait within a tropical clade, where taxonomic diversification can be prolific, will have a different future in terms of further exploration of morphospace-including the potential for macroevolutionary expression of developmental bias-than the same trait captured by a boreal clade, where diversification tends to be more modest. Timing is also important in clade dynamics: traits that originate immediately before a regional or global extinction event are more likely to disappear even if not directly selected against; many features have almost certainly been lost not because they were maladaptive but because they happened to be associated with narrow geographic ranges or other features elevating extinction risk d) Clade originates near a lower limit, so that diffusion yields an increase in maximum and mean value to the right. (e) Unbounded, unbiased speciation without a persistent association between origination/extinction and morphology, producing an increase in the maximum value and a decrease in the minimum. This is simple Brownian motion; bounded but unbiased speciation would conform to a simple Ornstein-Uhlenbeck model. F, Narrowing of variation by a decrease in the maximum value, and an increase in the minimum. From Jablonski (2017b), inspired by Gould (1982) and Stanley (1979) JABLONSKI | 113 (Jablonski, 2005a(Jablonski, , 2009(Jablonski, , 2017b)). As noted above, hitchhiking of traits on other features that promote differential origination and extinction is always a possibility: high or low frequencies of specific traits or trait combinations need not directly reflect the probability of the generation or persistence of those traits (Jablonski, 2017b).
Direction bias (Gould, 1982, directed speciation of Stanley, 1979:184-186) is the most intriguing macroevolutionary pattern from the standpoint of developmental bias. We have little direct evidence that repeated, potentially cumulative steps in a single direction in morphospace are underlain by developmental bias, as opposed to sustained selection on "random" variation (as in Grant's [1963Grant's [ , 1989] ] speciational trend and Futuyma's [1987] anagenesis by speciation), but the examples cited above, from Haber's ungulates through Hunt's ostracodes to McNamara's paedo-and peramorphoclines, are certainly suggestive. Iterative evolutionary patterns in the fossil record, where a source clade repeatedly generates similar derived phenotypes over a long interval of geologic time (potentially viewed as parallelism on a grand scale), as in, for example, planktic foraminifera (Norris, 1991;Tyszka, 2006;Caromel, Schmidt & Rayfield, 2017), ammonites (Monnet et al., 2015), and carnivoran mammals (Slater, 2015;Van Valkenburgh, 2007), have usually been interpreted in terms of selection and persistent adaptive peaks, perhaps following repeated extinction that restricts surviving clades to similar regions of morphospace as in forams and ammonites. However, a role for developmental bias should be investigated more thoroughly. It would be especially interesting to know whether macroevolutionary ratchets, in which reversals are extremely rare (Balisi, Casey, & Van Valkenburgh, 2018;Van Valkenburgh, 2007), are developmental in origin, or positive selection at the organismic level is sufficient to lock clades into trends whose direction-for example towards hypercarnivoryactually increases extinction risk, and/or damps origination, at the higher level. Numerous examples where selection for short-term advantage heightens extinction risk (references in Jablonski, 2017b) should be investigated for the relative role of intrinsic and extrinsic factors. Here too, explicit tests will be needed to separate the evolutionary role of developmental bias from macroevolutionary hitchhiking.
One way in which developmental factors could play a key macroevolutionary role would be if sustained, multispecies trends or iterations exhibit phylogenetic "heritability" of the relevant developmental bias. The limits to this heritability are poorly understood and should be the focus of a concerted research effort. At the species level, the fine structure of phenotypic (co) variation seems to be fairly labile. It can change rapidly in response to directional selection, even in natural populations (e.g., Assis, Patton, Hubbe, & Marroig, 2016); the variation signal in fossil ostracods fades over a few million years (Hunt, 2007), with a similar timeframe reported by Schluter (1996) from estimated divergence times for extant taxa; it is disrupted in planktic foraminifera by climate shifts (Brombacher, Wilson, Bailey, & Ezard, 2017); and static allometry-that is intraspecific variation among adults-while stable within species generally does not seem to persist among related species (Voje, Hansen, Egset, Bolstad, & Pélabon, 2014). At higher levels within larger clades, the direction and eccentricity of larger-scale multivariate variation is stable -and predictive of evolutionary patterns-within ungulate families and bird clades but differs among them (Haber, 2016;Watanabe, 2018). Perhaps this macroevolutionary lability is why, as Simpson and Müller (2012) argue, sustained macroevolutionary trends are relatively uncommon, as is iterative evolution at macroevolutionary scales; weak heritability of directional bias is one potential explanation, but inconstant abiotic and biotic environments is another.
Correlated progression is another mechanism that has been invoked for directional trends in the fossil record, but it posits stepwise changes via selection, as one change creates a selective environment for the next step (Kemp, 2016;Thomson, 1988Thomson, , 1992)). The implication is that variation per se is undirected, and the reciprocal interactions of a suite of structurally or functionally interrelated characters provide the vehicle for directional change (Hansen, 2014 also underscores the role of external directional selection in this concept). Separating this extrinsic driver from potential intrinsic, developmental factors will require careful work, but Kemp (2016) adamant rejection of modularity's role in macroevolution seems misplaced, and I have seen no clear reason that correlated progression could not operate in a modular system (and vice versa).
We now know that developmental bias can change in orientation and eccentricity on macroevolutionary scales, but several hypotheses are available for how bias actually evolves, each with different evolutionary implications (Figure 4). Drivers can be broadly divided into intrinsic and extrinsic factors.
A) Clades might begin with relatively isotropic variation, and evolve stronger developmental bias through time. This idea has been proposed at a wide range of temporal and phylogenetic scales, and at different hierarchical levels, reflecting our weak understanding of developmental evolution at large scales, and the different aims of the authors. It is possible that developmental systems have not approached isotropy since the Proterozoic assembly of the metazoan toolkit (see Davidson & Erwin, 2009;Erwin & Valentine, 2013, and see Borenstein & Krakauer, 2008, who find that early, low-dimensionality developmental systems should more fully and evenly occupy morphospace than more derived systems). This view is difficult to test directly, but a start would be to compare variation patterns in a clade in the midst of the Cambrian Explosion with patternslater in the clade's history. The overall amount of intraspecific variation in Cambrian trilobites is significantly higher than in later forms, with the more extreme variation in the early Cambrian and more basal taxa (Webster, 2007), and so drawn mainly from the explosive interval (see Paterson, Edgecombe, & Lee, 2019). However, we do not yet know whether that Cambrian variation is more isotropic; comparable analyses on Cambrian echinoderms and stem-arthropod groups would be extremely interesting. The bursts of disparity and novel morphologies that have evolved within clades since the Cambrian (e.g., Hughes, Gerber, & Wills, 2013;Vermeij, 2015;P. J. Wagner, 2018), while not as dramatic as the origin of metazoan bodyplans, seem to argue against any simple monotonic trend in developmental bias. Alternatively, loosening or reorganizing developmental bias may lie at the root of lower-level clades of any geologic age, such as genera or families, and may at least partly account for their diversification and later slowdown as bias becomes stronger (e.g., Uller et al., 2018;P. J. Wagner, 2018). This view poses a hypothesis about the nature of variability (potentially observed as variation) and covariation among traits at the base versus the tips of any evolutionary tree, greatly expanding the scope for potential tests. The literature somewhat chaotically ranges from very general claims about overall form to specific expectations about discrete traits. For example, one of the few stem mammals for which we have fossil juveniles apparently underwent isometric cranial growth (Hoffman & Rowe, 2018), perhaps suggesting that the allometric ontogenies often viewed in terms of developmental bias came later in the clade's history;
those later changes at the very least imposed new biases, particularly if the evolution of enlarged brains was associated with stronger cranial allometry and smaller clutch sizes. Intriguingly, early Mesozoic echinoids are also more confined but isotropically distributed in morphospace than later, more derived irregular lineages (Hopkins & Smith, 2015; Figure 1). The strongest support for the hypothesis that the origin of clades of any geologic age is associated with a reorganization of trait covariation comes from an impressive analysis of 257 character matrices for fossil taxa spread throughout the Phanerozoic, in which clades produce disparity early in their histories (regardless of geologic age) by breaking up characterstate correlations and forming new ones (P. J. Wagner, 2018). No developmental data lie behind this "correlated change-breakup-relinkage" model, and alternative explanations are possible, but as Wagner points out, these results are consistent with the view that an evolutionary novelty captured within the framework of one set of developmental biases might produce an ecological shift or opportunity that selects for the reorganization of trait covariation and a resulting evolutionary pulse in morphospace. This view posits an interaction between development and function that recalls Van Valen's (1973) dictum that evolution is the control of development by ecology, and thus provides a broader perspective on developmental bias. Comparative developmental or variational analyses of clades that conform to the relinkage model, versus those that do not, coupled with functional analyses of the traits associated with such shifts, are much needed. These encouraging findings do not mean that diversification can be taken, in itself, as evidence for relaxed, and then strengthened, developmental bias. For example, the diversification of beaks in the early Cenozoic radiation of modern birds might be viewed in this way (Uller et al., 2018). However, this diversification occurred perhaps 50 Myr after edentulous bird beaks originated in the modern-bird lineage, implying that these were long-established rather than novel structures and regulatory networks (Meredith, Zhang, Gilbert, Jarvis & Springer, 2014;Bhullar et al., 2015). Thus, alternative mechanisms must also be considered, including release from interactions with pterosaurs, nonavian dinosaurs, and Mesozoic bird lineages outside the crown, all of which were lost at the end of the Mesozoic Era. This release was followed by subsequent ecological crowding, as in the Valentine-Walker model discussed by Jablonski (2017b), with the early Cenozoic diversification perhaps enhanced by, but not initiated by, the simple set of regulatory interactions underlying beak formation in at least some modern passerine birds, noted above. An intermediate, intuitively appealing, possibility is that the loosening and later tightening of developmental bias occurs only at the origin of clades that are founded via an evolutionary novelty sensu G. P. Wagner (2014), that is a trait lacking homology in the ancestor, or radically and irreversibly altered from the ancestral state. This hypothesis, like the others in this section, might be tested against diversifications that were more clearly triggered by ecological opportunity, as discussed below. Reanalysis of P. J. Wagner's (2018) inventory of character-state evolution across paleontological phylogenies from this perspective would also be valuable. B) Clades might evolve to reduce or circumvent developmental bias. This might be the expectation if modularity tends to increase and phenotypic integration tends to decrease over clade histories, as suggested by Goswami et al. (2014). It may also provide a developmental basis for Vermeij's (2015) observation that younger branches within major animal and plant clades explore a greater portion of morphospace than older ones, although such observations overlap with the hypothesis that novel clades succeed in part by restructuring trait covariation. Many clade pairs exemplify Vermeij's pattern, from leaf shapes and gastropod shells to fish skulls and the contrasting echinoid clades in Figure 1. The issue is whether these phylogenetically patterned contrasts are shaped by selection for alleviating energetic tradeoffs, as Vermeij (2015) proposes, by differences in developmental bias, or by some combination that can only be evaluated by a deeper understanding of the developmental basis of variation within and among these clades. C) Developmental biases may be present throughout metazoan history, but reorganization of trait covariation, or shifts in the orientation or eccentricity of continuous variation, might occur idiosyncratically through time and among subclades. Under this model, bias has important local effects within the phylogeny but does not necessarily channel shifts in large-scale evolutionary trajectories or dramatic changes in disparity: the more rapidly a given developmental bias fades over geologic time, or can be reoriented by extrinsic drivers, the more it effectively becomes another source of "random" variation rather than a macroevolutionary factor. As noted in Section 4.1, the evidence for this view is mixed, and may depend on the clade or the traits involved. For example, in Plio-Pleistocene planktic foraminifera, evolutionary changes tend to be aligned with allometric patterns, but shifts in climate state drive clades in different directions (Brombacher et al., 2017). Broader but less direct evidence, subject to further testing and, more important, integration into a general theory of developmental bias, includes the decline in the probability of gene reuse with increasing age (Conte et al., 2012), a decline in the frequency of convergence with phylogenetic distance (Ord & Summers, 2015), a decline in the association of parallelism at the organismal and genetic levels with phylogenetic distance (Elmer & Meyer, 2011, who note many counterexamples), and variation in modularity across species phylogenies within lower taxa such as families or genera (e.g., Sanger, Mahler, Abzhanov & Losos, 2012 on the Anolis skull, Haber, 2016 on ungulate species within families).
Whether evolutionary novelties entail the origin of traits lacking homologs in ancestors, or radical, irreversible evolutionary transformation of existing features (G. P. Wagner, 2014, see also Erwin, 2015;Moczek, 2008), they likely involve breaking some developmental biases and the establishment of new ones. However, some developmental biases remain firmly in place even across these transitions, and may be insurmountable. For example, the origin of the stereom endoskeleton and the water-vascular system enabled the echinoderms to do extraordinary things with their body plans, integument, and plumbing, as seen in sea lilies, sea cucumbers, sand dollars, and the bizarre array of Paleozoic taxa afforded class-level rank, but they never managed to assemble a phenotype that got them onto land, unlike mollusks, arthropods, and chordates, all of which originated at about the same time and in similar marine environments. table biases, or nearly so, may be scarce (Vermeij, 2015), but the best-documented examples appear to have arisen where development is highly integrated, so that a web of pleiotropic effects resist evolutionary change (Galis, Metz, & van Alphen, 2018).
Regardless of how developmental biases evolve within clades, their effects might vary through time owing to changing circumstances. Much is known about how taxa, and increasingly about how phenotypes, respond to ecological opportunities presented by the invasion of new region (e.g., Donoghue & Sanderson, 2015) or survival through a regional or global extinction event (e.g., Jablonski, 2005a,b), but we know little about how those events influence, and are influenced by, developmental factors. In our current state of knowledge, open opportunities might be more forgiving of the selective cost of trajectories that are oblique or orthogonal to developmental bias, so that evolutionary directions become more isotropic and biases less evident (perhaps allowing the establishment of new orientations and eccentricities of the biases). Alternatively, open opportunities might promote more prolific evolutionary change, specifically along lines promoted by developmental biases so that such biases will be most manifest at these times.
The taxonomic diversification pulses seen after mass extinctions are often thought to signal enlarged ecological opportunity with the removal of incumbent species and clades (Foote, 2010;Jablonski, 2008b;Miller & Sepkoski, 1988). A few case studies shed light on the phenotypic effects of those events. For crinoid echinoderms, the end-Permian extinction is indeed followed by accelerated taxonomic and morphological diversification, but this post-Paleozoic radiation is not nearly as prolific as the Paleozoic one, and the two radiations are largely nonoverlapping in morphospace. This contrast could indicate that later crinoids were more heavily constrained by developmental factors (plausible given that post-Paleozoic crinoids show a higher degree of integration than the Paleozoic forms, Gerber, 2013), that ecological opportunities were not as available as might be expected given the magnitude of the extinction (also plausible, see Edie, Jablonski, & Valentine, 2018;Erwin, Valentine, & Sepkoski, 1987;Foster & Twitchett, 2014), or that the specific crinoid lineage(s) that survived the extinction bore developmental biases that differed from precursor clades and strongly influenced the post-Paleozoic history of the clade (Foote, 1999). Analyses using different methods and data show a similar effect for the end-Paleozoic extinction in crinoids but not across the other two, somewhat milder within-Paleozoic mass extinctions (Ciampaglio, 2002). Articulated brachiopods do show an early post-Paleozoic pulse, and manage to achieve Paleozoic-like disparity by measures (though not by others), albeit in a macroevolutionary slog rather than a burst (Ciampaglio, 2004). In contrast to post-Paleozoic crinoids and brachiopods, mammal body-size evolution after the end-Cretaceous extinction of the nonavian dinosaurs shifted from confined, Ornstein-Uhlenbeck evolution to unconfined Brownian Motion without a change in absolute rate (Slater, 2013). Morphological disparity in Early Cenozoic placental mammals actually lagged behind taxonomic diversification (Halliday & Goswami, 2016), but this study excluded autapomorphic traits (and see Wesley-Hunt, 2005, showing more dramatic disparification within the ecological carnivore category, as opposed to individual carnivorous clades, and the expansion of placental jaw disparity across the same interval, despite near-constancy in the mammals overall [Benevento, Benson, & Friedman, 2019]). Further data are needed to test, for example, whether the range of phenotypes in morphospace derived from earliest Cenozoic mammals shows a weakening of eccentricity sensu Haber (2016). The same holds for any of the potential "disparity-first" diversifications (Jablonski, 2017b) seen in the fossil record.
The failure to diversify in response to ecological opportunities also requires explanation. Many hypotheses exist for this dead-clade-walking phenomenon (survival of a mass extinction without subsequent diversification, Jablonski, 2002), and most of them are ecological or biogeographic. However, no comparative studies have been conducted on the phenotypic variation available to clades that follow contrasting postextinction trajectories. We thus lack tests for whether the unsuccessful survivors represent clades whose developmental biases were conserved rather than reorganized across mass extinctions, or happened to be particularly narrow in size and eccentricity, relative to successful survivors.
There are many potential approaches to assessing the macroevolutionary role of developmental bias (Table 1), but our knowledge is scattered across systems and hierarchical levels. We have a wealth of tantalizing clues and several convincing case-studies, but no synthesis that establishes general rules for the operation of bias and its limits. Such a synthesis requires a two-way exchange between the molecular and developmental approaches available to neontologists and the strictly phenotypic but richly historical and dynamic data available to paleontologists, and could take predictive form. Analyses starting in extant populations could aim to predict macroevolution in the fossil record from observed developmental bias, and analyses starting in the fossil record, particularly the fossil record of extant species and clades, could aim to predict developmental bias from macroevolutionary patterns. This approach could be seen as a welter of consistency arguments, but the types of data involved are so disparate that the accumulation of positive results from carefully designed analyses will be richly informative-and prediction failures will be just as interesting if we can learn to account for them, for example by building alternative factors into evolutionary models from the start. Such a macroevolutionary project underscores the need for experimental work and developmental modeling focused on clades with rich fossil records.
Also key to any large-scale evolutionary synthesis will be a more rigorous and comprehensive analysis understanding of the relation between variability-that is the probability distribution of forms generated by developmental systems-and variation, the observed density distribution of adult morphologies and postnatal juveniles within a species at a given point in its evolutionary history. We know they are not identical, and we can account for complete nonequivalence when it occurs. Nonetheless, species-level (co)variation, in both discrete and continuous traits, has shown sufficient macroevolutionary explanatory power in the available studies that some degree of congruity between variability and variation must occur; we need to understand its limits, which probably vary among traits, clades, and perhaps among times in the history of clades.
Comparison of variability (density distribution of phenotypes potentially generated developmentally from a given population) with variation (realized phenotypic differences among individuals) observable in collections of living and fossil taxa. Key to macroevolutionary evaluation of developmental bias-when does it break down (traits, clades, etc.)?
Establishment of new developmental model systems that have a rich fossil record.
Fuller integration of fossil records-even sparse or patchy ones -into phylogenies to inform models for ancestral character states and both rate and direction of change.
Generation of quantitative models of trait development that can predict probability distributions around ancestral phenotypes.
Evaluation of the relative roles, and interaction, of clade dynamics and developmental bias in determining the breadth and pattern of morphospace occupation of clades.
Research Directions, Evaluating:
Whether the most accessible phenotypes predicted from models of development correspond to realized evolutionary trajectories.
Stability of developmental bias among successive or coeval populations, among species, across clades, and whether its strength is associated with constancy in evolutionary tempo and mode.
Temporal trends in strength and orientation of developmental bias.
Whether the most successful clades, in morphologic disparity and/or taxonomic diversity, conform to, or break, ancestral patterns of developmental bias.
Associations between the strength and orientation of developmental bias and:
• Permanent vacancies within a clade's morphospace.
• Differential origination and extinction rates, to generate among-clade and across-time differences in morphospace occupation.
• Markedly different patterns or degrees of phenotypic diversification or contrasting diversity-disparity relationships among clades.
• Parallel evolution, and convergence.
• The directional generation of species over the history of a clade to produce a macroevolutionary trend.
• Macroevolutionary ratchets (i.e., trends lacking reversals).
• The origin of evolutionary novelty, in terms of magnitudes and discontinuities of phentotypic change.
• The temporal, environmental, and biogeographic patterns evident in the origin of evolutionary novelty and subsequent diversification.
These studies will not be sufficient in themselves if the ultimate aim is to understand the diversity of life in its many currencies-taxonomic, phenotypic, and functional. Clade dynamics, specifically multilevel selection, must be more fully integrated with evolutionary developmental biology. Multilevel selection without evo-devo puts the origin of its raw material-variation among species-in a black box; evo-devo without multilevel selection is an assemblage of localized case-studies divorced from the evolutionary dynamics that play such a large part in setting the range, density, and trajectories of morphospace occupation, the differential waxing and waning of clades, and the effective exploitation and elaboration of novel features once they are acquired. Integration of the two branches of macroevolutionary study will produce a richly extended synthesis.