Transitions in Parental Care and Feeding Mode

Buccal Cavity Morphology

Data Collection and Landmarking

For morphological analysis, we sampled 41 species across 20 genera representing the major radiations of Neotropical cichlids, mostly from the University of Michigan Museum of Zoology and the Field Museum of Natural History. Minimally, we sampled to the resolution of the node representing the most recent common ancestor between mouthbrooding and nonmouthbrooding taxa for each clade of mouthbrooding species, so that we sampled the immediate substrate-brooding sister taxon for each mouthbrooding taxon. Where possible, we sampled three to five individuals per species, except in cases where fewer than three individuals were available in collections ( $n=17$); see the supplemental PDF for accession numbers and sample sizes for each species. We selected for analysis only specimens that were fixed in closed-mouth, neutral positions (i.e., no opercular flaring or hyoid depression). Because most mouthbrooding Neotropical cichlids are biparental and exhibit little to no sexual dimorphism, we did not distinguish between males and females in our sampling. The exception is Gymnogeophagus, which exhibits maternal-only care and extreme sexual dimorphism; we still sampled males and females from mouthbrooding Gymnogeophagus to be consistent with the rest of the study.

Each specimen was dissected using a modified version of the procedure described by Ridewood (1904) to remove the right suspensorium and oral jaw elements, exposing the buccal cavity and branchial arches (fig. 2). First, we made a midline cut on an anterior-posterior axis separating the premaxillary bones and continued this cut by bisecting the parasphenoid along the midline. We separated the circumorbital bones from around the orbit before cutting the levator operculi. Ventrally, we separated the basihyal from its articulation with the urohyal and basibranchials, then cut the branchiostegal membrane from the urohyal through the mandibular symphysis. We were then able to remove the entire right cheek intact, including the opercular series, suspensorium, adductor mandibulae, nasal and circumorbital bones, and oral jaws.

We photographed specimens before and after dissection on a Leica or Zeiss stereomicroscope to capture both external and buccal cranial morphologies. We based our external landmarks on those in Weller et al. (2017), including curves for the rostrum, eye, lower jaw, operculum, and adductor mandibulae complex (fig. 2A). For the buccal cavity, we chose landmarks describing structures that make up the borders of the cavity, including the vomer, parasphenoid, basibranchial series, hyoid, and lower jaw. A full list of fixed and sliding landmarks is given in the supplemental PDF.

Photographs were scaled and landmarked using StereoMorph (Olsen and Westneat 2015), then imported into geomorph using the readland.shapes function, which converts StereoMorph curves into sliding semilandmarks (Adams and Otárola-Castillo 2013).

Mouthbrooding Transition Rates among Winnowers and Nonwinnowers

Using Pagel’s (1994) method for testing the correlated evolution of discrete characters, we found the highest support (72% by AIC weight) for a model where parental care type depends on feeding mode but feeding mode does not depend on parental care type. The next best fit, mutual dependence of parental care type and feeding mode, had reasonable support ( $\mathrm{\Delta }\mathrm{AIC}=+2.2$ compared with the best model, 24% AIC weight), followed by the mutual independence model ( $\mathrm{\Delta }\mathrm{AIC}=+6.0$, 3% AIC weight) and feeding mode depending on parental care type ( $\mathrm{\Delta }\mathrm{AIC}=+8.4$, 1% AIC weight).

Of the three different rate assumption Mk models we fit for the four discrete behavioral categories, the symmetrical rates model was best supported (86.9% AIC weight), followed by the all-rates-different model (12.8%); the equal rates model had very little support (0.3% AIC weight). We used these AIC weights to calculate the weighted average of the transition rate matrices (table 1). The mouthbrooding transition rate was 15 times higher in winnowers than in nonwinnowers and 2.5 times higher than the transition rate to or from winnowing itself.

For the 10,000 simulated stochastic character maps for each rate model, we randomly sampled a proportion equal to the AIC weight for each model (8,690 symmetrical rate maps, 1,280 all-rates-different maps, and three equal rates maps) for a combined total of 10,000 stochastic character maps. Across these simulations, mouthbrooding evolved from substrate brooding and winnowing a median value of three times, compared with only once from substrate brooding and nonwinnowing (fig. 3). The high probability density (HPD) interval for winnowing to winnowing and mouthbrooding was also considerably wider than it was for any other transition (zero to five times). Winnowing itself was gained a median of seven times (HPD interval of six to nine), more than double the number of times that mouthbrooding was gained. Winnowing never evolved within a mouthbrooding clade. Mouthbrooding was also rarely lost: we found no transitions from mouthbrooding to substrate brooding in nonwinnowers and a median of zero transitions from mouthbrooding to substrate brooding in winnowers, with an HPD of zero to two transitions.

We ran the methods described above for two alternate versions of the data set: one where we classified four species with ambiguous reports of mouthbrooding as mouthbrooders (inclusive) and one where we excluded these species entirely (exclusive). Detailed results are provided in the supplemental PDF, but the results in these cases were qualitatively similar to those we report here: (1) we found the best support for Pagel’s (1994) test for a model where mouthbrooding depends on winnowing; (2) we found the highest support for a symmetrical rates Mk model, followed by equal rates and all-rates-different models; and (3) we found that the highest transition rate in the weighted average transition rate matrix was from winnowing to winnowing and mouthbrooding (see sec. S2 of the supplemental PDF).

Buccal Cavity Area Differences

Mouthbrooders and winnowers had larger buccal cavities than species that exhibited neither behavior (fig. 4A). In nonwinnowing species, mouthbrooders had considerably larger buccal cavities compared with substrate brooders (nearly 60% as a proportion of head area). All species of Bujurquina, a genus of mouthbrooding cichlasomatines, had buccal cavities between 11% and 17% of head area; Andinoacara, the nearest substrate-brooding sister taxon to Bujurquina, had buccal cavities between 6% and 9% of head area. Among winnowing species, both mouthbrooders and substrate brooders had large buccal cavities (12%–24% of head area for mouthbrooding winnowers and 10%–21% of head area for substrate-brooding winnowers).

We used a phylogenetic ANOVA (Adams and Collyer 2018) to test whether buccal cavity areas depended on feeding mode and parental care strategy, fitting buccal cavity area as a function of parental care mode and feeding mode, with squared standard length included as a covariate to control for body size. We found that only feeding mode had a significant effect at the $\alpha =.05$ threshold ( $P=.015$), while the effect of parental care mode was considerably weaker ( $P=.14$), as was their interaction term ( $P=.15$). The effect of parental care was significant when we performed a nonphylogenetic ANOVA ( $P=.0002$), as was the interaction of parental care and feeding mode ( $P=.003$). We found the same results using the phylogenetic residuals from the phyl.resid function as an alternative method of correcting for body size. This indicated that the magnitude of difference between buccal cavity areas was substantial but that because we have a single clade of nonwinnowing mouthbrooders in our sample, we lack sufficient phylogenetic resolution to attribute that difference to mouthbrooding.

To test whether these buccal cavity area differences were largely driven by differences in head area across taxa, we also fit a PGLS model of buccal area as a function of head area, again with squared standard length as a covariate. We found no significant relationship between buccal cavity area and head area ( $P=.19$; fig. 4B). We also tested whether head area differed significantly across behavioral categories by fitting a PGLS model of head area as a function of parental care and feeding mode, again with standard length as a covariate. We found no significant relationship between head area and any of the variables ( $P>.35$ for feeding, parental care, and their interaction).

Buccal Cavity Shape Differences

Both mouthbrooders and winnowers had similar buccal cavity shapes, even after controlling for size. In a principal components analysis, PC1 (47% of shape variance) corresponded largely to buccal cavity depth (fig. 5A), with both mouthbrooding and winnowing species having deeper buccal cavities. Differences in parasphenoid curvature accounted for most of the change in buccal cavity depth. More highly curved parasphenoids result in a greater distance between the roof and the floor of the buccal cavity, producing a larger overall volume even when the mouth is fully closed. PC2 (24% of shape variance) largely separated different substrate-brooding nonwinnowing taxa from each other, while winnowers and mouthbrooders spanned a relatively narrow range of PC2 scores.

We found that only feeding mode, not parental care or their interaction, had a significant effect on buccal cavity shape in a PGLS model ( $P<.01$), although this result can be partially attributed to group aggregation on the phylogeny (see “Discussion”). We also used the PGLS model to predict buccal cavity shape for each of the four behavioral categories (fig. 5C) and calculated Procrustes distances between each pair of shapes to quantify shape differences. The highest Procrustes distance (least similar) between any two predicted shapes was between the mouthbrooding winnower shape and the buccal cavity shape for neither behavior. The lowest distance (most similar) was between the mouthbrooding and substrate-brooding winnower shapes, followed by the mouthbrooder and winnower shapes.

Mouthbrooding Evolves More Frequently in Winnowing Clades

Each discrete modeling method supported the same conclusion: mouthbrooding is gained and lost more frequently in winnowing than in nonwinnowing clades. The results of Pagel’s (1994) correlation test strongly support that transitions to or from mouthbrooding occur much more often in winnowing lineages but that transitions to or from winnowing do not depend on parental care. Our Mk model and stochastic character map results also generally favored mouthbrooding being gained within winnowers more often than it was lost (fig. 3), although we did recover some ambiguity around the directionality of this change (table 1).

This ambiguity is partly a consequence of our model design, in which we created four discrete character states as a combination of parental care and feeding states. It makes more biological sense that winnowing and mouthbrooding would not be gained simultaneously, but this transition is not penalized by any of our Mk models or stochastic character maps. Incorporating this assumption into future models could clarify the most likely direction of this transition, but even without making an assumption either way, our analyses are highly consistent overall: mouthbrooding transitions occur more often in winnowers.

We also noted that the combination of winnowing and mouthbrooding has only evolved within the Geophagini tribe (in Gymnogeophagus, Geophagus sensu stricto, the “Geophagus” steindachneri group, and Satanoperca). This could be a function of clade age: geophagines are much older as a group than the Heroini and evolved winnowing much earlier. It could also be a simple matter of body size, since many of the winnowing geophagines get fairly large (for cichlids), so the potential reduction in total brood volume for mouthbrooding is not as severe on an absolute scale (Steele and López-Fernández 2014). Still, geophagines make up about half of Neotropical cichlid diversity and are substrate brooding in the majority, making our observations generally informative for mouthbrooding evolution in the group.

Both Mouthbrooding and Winnowing Are Associated with Larger Buccal Cavities

We found that both winnowers and mouthbrooders tended to have larger, more curved buccal cavities than nonwinnowers, even relative to their overall head area. This is especially apparent in Bujurquina when compared with its nearest substrate-brooding sister taxon (Andinoacara): despite extremely similar external cranial morphologies, Bujurquina buccal cavities were about twice the size of those in Andinoacara, a difference that was mostly achieved via a steeper angle of the ceratobranchial and a higher arch of the parasphenoid, creating more space even when the mouth was closed. While winnowers and mouthbrooders showed clear departures from the buccal cavities of fishes that exhibit neither behavior, mouthbrooding winnowers differed little from their substrate-brooding counterparts (fig. 5). In contrast with Bujurquina and Andinoacara, mouthbrooding winnowers were nearly indistinguishable from their nearest substrate-brooding sister taxa in each case (compare Satanoperca jurupari and S. daemon in fig. 5A).

One explanation for these observations is that the mouthbrooding Bujurquina + Tahuantinsuyoa clade evolved larger buccal cavities because these were advantageous for mouthbrooding, while within the winnowing lineages buccal cavities underwent no change when mouthbrooding evolved because winnowing adaptations for winnowing had already resulted in a morphology advantageous for mouthbrooding. This could make sense from a functional morphological perspective. Winnowing involves sifting mouthfuls of substrate for edible material, probably by controlling fluid flow in the mouth to manipulate substrate particles, a process punctuated by cycles of premaxillary protrusion, mouth opening, and opercular flaring (Weller et al. 2017). Mouthbrooding, at least in Neotropical cichlids, is a superficially similar process: eggs (and sometimes fry) are tumbled in the mouth via repeated cycles of premaxillary protrusion, hyoid depression, and opercular flaring (Van Wassenbergh et al. 2016).

Both processes require holding, sensing, and manipulating many particles in the buccal cavity and having fine control over whether those particles are swallowed, with the meaningful difference that in winnowing the goal is to swallow those particles whereas in mouthbrooding the goal is to keep the offspring tumbling in the mouth while the parent abstains from feeding. Mouthbrooders swallowing their brood (filial cannibalism) is frequently reported in other mouthbrooding groups (Okuda 2000), including in African cichlids, where it is apparently accidental (females often swallow their first brood before successfully rearing their second, per Taborsky 2006). To our knowledge, filial cannibalism has never been reported for a Neotropical mouthbrooding cichlid, although there is comparatively less research on Neotropical mouthbrooding generally. Based on our conjecture here, we would hypothesize that it occurs less frequently in winnowing than in nonwinnowing species.

Having a larger, more rounded buccal cavity not only provides more room for substrate (or eggs) but may also provide some hydrodynamic advantage in generating turbulence and associated water flow around the particles (Brooks et al. 2018). In their experiments with mouthbrooder egg development, Shaw and Aronson (1954) found that eggs not exposed to flowing water developed slowly or not at all, which they attributed to a lack of adequate oxygen diffusion; eggs that were exposed to flowing water but not tumbled (i.e., not freely rotated and translated) were still susceptible to fungal growths. Greater curvature in the parasphenoid produces a buccal cavity shape that resembles a pipe that expands into a sphere in the center: the rapid change in diameter should disrupt laminar flow and produce turbulence, which could allow eggs to spend more time tumbling and less time settled at the bottom of the mouth (Van Wassenbergh et al. 2015; Vogel 2020). Substrate-brooding cichlids generate a constant water flow around their eggs by “fanning” with their fins and bodies, supporting the idea that those conditions would need to be re-created in the mouth cavity of mouthbrooders.

While the results we present here have a number of clear functional implications, a major caveat is that several of our inferences rest on the single clade of confirmed nonwinnowing mouthbrooders among Neotropical cichlids. We can say with some confidence that mouthbrooding and substrate-brooding winnowers have similarly large, curved buccal cavities, but we have only one phylogenetically independent observation of the buccal cavity of a nonwinnowing mouthbrooder in this clade. Bujurquina’s larger buccal cavity may be an adaptation to mouthbrooding, but it may be due to another selective pressure (like respiration or feeding) or even the product of a constructional constraint (Barel 1983). This could potentially be addressed by expanding our data set to include putative Neotropical mouthbrooders, such as Apistogramma megastoma (Römer et al. 2017); the large mouth for which this species is named may have a correspondingly large buccal cavity. Expanding the scope of our phylogenetic analysis to include mouthbrooding cichlid clades from Africa should also reveal whether the morphological attributes of winnowing mouthbrooders and the rates of transition between them are similar to those we found in the Neotropics.

However, even if we can attribute this increased buccal cavity area to mouthbrooding, convergent morphologies between two behaviors are not conclusive evidence of an evolutionary path of least resistance leading from one to the other. Both winnowing and mouthbrooding are dynamic processes, involving constant shape changes to the notoriously kinetic fish head; a wide variety of buccal cavity shapes could be achieved by adjustments to posture alone, and cranial expansion (as during suction feeding) probably contributes more to increased buccal cavity volume than any change to static morphology possibly could. The morphologies of these fishes are at best indirect proxies for their behaviors. In that sense, it is somewhat remarkable that we recovered the magnitude of differences that we present here and that these differences appear to be specific to the buccal cavity (as opposed to the overall head area). A more complete understanding of the relationship between feeding and mouthbrooding will require a detailed study of the behavioral, kinematic, and mechanical components of each process and of how and where they might overlap or conflict.

When and How Much Does Morphology Matter for Mouthbrooding?

The results we present here support the hypothesis that, at least in Neotropical cichlids, mouthbrooding evolves more often from winnowing than it does from other feeding modes because winnowing produces a feeding phenotype that is advantageous for mouthbrooding. We can construct a plausible evolutionary sequence of events from this interpretation: a lineage evolves winnowing, a feeding mode that selects for, among other traits, a large, curved buccal cavity and the necessary motor control and sensation for manipulating particles intraorally. These traits are adaptive for winnowing but happen to be advantageous for mouthbrooding as well. If a fish in this lineage begins mouthbrooding, even facultatively, it is likely to have much more success than a fish from a nonwinnowing lineage would because it can mouthbrood a larger brood volume for longer periods of time. By comparison, a nonwinnowing fish without these feeding traits must have either a smaller brood (fewer or smaller eggs) or a higher rate of offspring loss to physiological limits, such as inadequate oxygenation to the eggs in the buccal cavity. In this context, mouthbrooding is advantageous for the winnower but not the nonwinnower, and it is selected for only in the winnowing population.

Such a model would explain the higher transition rate to or from mouthbrooding in winnowing lineages, but it is an incomplete explanation. Mouthbrooding clearly evolves even in the absence of a phenotypic alignment with an existing feeding mode. Among Neotropical cichlids, the Bujurquina + Tahuantinsuyoa clade is the best-documented example of nonwinnowing mouthbrooders, but there are reports from the aquarium hobby of facultative mouthbrooding in Heros liberifer (Staeck and Schindler 2015) and at least two species of Apistogramma, A. barlowi (Römer and Hahn 2008) and A. megastoma (Römer et al. 2017), both of which have remarkably large heads relative to other members of the group. And among the Central American Heroini cichlids, winnowing has evolved at least twice (fig. 3), while mouthbrooding has apparently never evolved in this group. Still, this is further complicated in non-Neotropical cichlids, where mouthbrooding is extremely widespread (especially in the African lakes) among taxa that occupy a wide range of feeding niches and morphologies (Liem 1973; Goodwin et al. 1998; McGee et al. 2020). These African mouthbrooders lay fewer eggs than substrate brooders (sometimes by an order of magnitude), and these eggs are considerably larger, with effective diameters averaging twice those of substrate brooders (Coleman 1991, 1998)—meaning that they have even fewer offspring for the same brood volume. Data on egg size and number is much more scarce for Neotropical cichlids, but reports from the aquarium hobby suggest that egg size differences are negligible, although we would require empirical data to draw more reliable conclusions (Coleman and Galvani 1998).

We suggest a conceptual life history model to reconcile these seemingly contradictory observations in which the relative selective advantage of mouthbrooding depends on both feeding morphology and environmental risks that impact offspring survival (fig. 6). Mouthbrooding is hypothesized to be advantageous for specific environmental correlates of offspring mortality, such as high predation rates, low nest site availability, and hypoxia risk (Östlund-Nilsson and Nilsson 2004; Rüber et al. 2004; Taborsky and Foerster 2004; Duponchelle et al. 2008). As these factors increase, so does offspring mortality, even for mouthbrooders (Vranken et al. 2019). The rate of offspring loss, however, should depend on parental care strategy, with substrate brooders losing proportionally more offspring than mouthbrooders (fig. 6, compare slopes of gray and yellow lines), while mouthbrooders have a lower maximum brood volume. This would suggest a tipping point where mouthbrooding yields more surviving offspring than substrate brooding, but only past some threshold of environmental risk (fig. 6, yellow dashed line).

In this model, feeding phenotype—insofar as it influences cranial morphology and behavior—matters because it determines maximum brood volume. A winnowing species (or one with a similarly congruent feeding mode) should be able to brood a larger volume of offspring than an otherwise identical nonwinnowing species, meaning that mouthbrooding is advantageous for winnowers across a wider range of environments (fig. 6, blue line and blue dashed line). Essentially, when environmental drivers of offspring mortality are extremely low, substrate brooding is advantageous for all feeding modes, and when they are extremely high, mouthbrooding is advantageous for all feeding modes; but for some range of environments, mouthbrooding is advantageous only for species with congruent feeding modes.

Although simple, this conceptual model does yield some testable predictions. First, we would predict that the selection pressure for mouthbrooding is strong in environments where we observe mouthbrooders with conflicting feeding modes (African lakes), intermediate in environments where we mostly observe mouthbrooders with congruent feeding modes (South America), and weak in environments where we observe winnowers but never mouthbrooders (Central America). The relative rates of offspring loss under different environmental conditions would probably be extremely difficult to measure, partly because many of these parameters are unwieldy to collect for a wide range of species and partly because of the number of ecological, biogeographical, and phylogenetic factors influencing life history evolution across these disparate environments.

Much more tractable are two predictions about the costs of mouthbrooding: (1) the reduction in brood size for mouthbrooders (relative to substrate brooders) should be different for winnowers and nonwinnowers and (2) winnowers should have better mouthbrooding performance or lower energetic costs of mouthbrooding than nonwinnowers. The parallel evolution of smaller broods and larger eggs has been well documented in the evolution of mouthbrooding in the African Great Lakes (Duponchelle et al. 2008); to our knowledge, this pattern has never been examined in the Neotropical cichlids. Assessing mouthbrooding performance would first require a biologically reasonable definition of performance in the context of parental care, but measuring this performance in winnowing and nonwinnowing mouthbrooders could provide a fascinating basis for testing whether winnowers really make for superior mouthbrooders.

Constraint, Opportunity, and Multifunctionality

The multifunctional perspective views individual traits as compromises between conflicting functional pressures (Pigliucci 2007). Bird beak shape, for example, must balance foraging, thermoregulation, and song production (Friedman et al. 2019), and Stayton (2019), in his study of hard-shelled turtles, was even able to determine the relative importance of hydrodynamic performance, strength, and self-righting on the evolution of shell shape. Mouthbrooding, too, could be interpreted as a straightforward example of a multifunctional constraint, where conflicting reproductive and feeding pressures reduced cranial morphological diversity (Hoey et al. 2012). Instead, we show that mouthbrooding may be most likely to evolve precisely when it does not introduce additional constraints to the head, possibly in an example of co-optation (or exaptation, per Gould and Vrba 1982). While this does not necessarily contradict the multifunctional perspective, we think it raises a larger evolutionary question: when multifunctionality increases for one trait, what happens to evolution at the level of the whole organism? We predict that by further constraining one trait, organisms can access new phenotypic space in the evolution of another trait. Mouthbrooding, for example, may release formerly constrictive reproductive constraints, allowing parents to reproduce under a wider range of environmental conditions. In the context of co-optation, increasing multifunctionality may generate more opportunities than constraints—but this is obvious only when studies account for how multifunctional traits affect evolution at the level of the whole organism.