Traces of phylogeny and ecology in hippocampal neuron numbers

Abstract

It is not known how selective pressures shape the numbers of interconnected neurons in defined neural circuits during the phylogeny of mammals. Consequently, models of function are without phylogenetic bounds, and species differences in neuronal makeup cannot be linked to ecological factors that generate selective pressures. Based on data from 65 species belonging to 11 orders, we here provide an analysis of five interconnected neuron populations in the circuitry of the hippocampus, the forebrain region encoding episodic memories. Related species tend to share traits in the hippocampal makeup, with distinct differences between clades. Phylogenetic signals result in the clustering of related species according to relative neuron numbers, but signal strengths allow the clusters to overlap. Tree-based methods show that neuron numbers can be explained by a selective mechanism that constrains them close to an across-species mean. Neuron numbers concerned with hippocampal input are more constrained than those providing output. An ancestral state estimate is provided, and species close to this phenotype are identified. Of the ecological factors tested, food, in terms of diet breadth, leaves its trace in many neuron numbers and strongly so in hippocampal input populations. Home range effects are more selective and relate to neuron ratios rather than neuron numbers. Phylogenetic constraints and ecologically guided relations seem necessary for the appropriate function of hippocampal input across a wide range of species.

Significance Statement.

This work provides an extensive dataset on the numbers of five interconnected hippocampal neuron populations, acquired using design-based stereology, in a large number of mammalian species. Species number and diversity along the phylogenetic tree allow formal tests for evolutionary processes at neuronal circuit-level, specifying phylogenetic scenarios and ecological drivers of species diversification, the assessment of phylogenetic signal, and the estimation of an ancestral state at the root of the tree. In the case of hippocampal neuron numbers, analyses point to changes in hippocampal input, driven by the critical resources of habitat size and food, as differentiators between species.

Introduction

A key determinant of the size of the major divisions of the mammalian brain (e.g. brainstem, cerebellum, or hippocampus) is overall brain size. As brain size in terms of its volume or neuron number varies over several orders of magnitude, so do component sizes, forming tight log-linear correlations (1–3). Yet, species may differ in the information about the environment most salient to them, the sensory modality conveying this information, and the niche-dependent, appropriate behavioral response to the information. To accommodate the demand for species-specific processing, a brain can use deviations from the overall correlation between component size and brain size. While small compared with the effect of brain size, such deviations still permit many-fold changes in the relative sizes of brain components (1, 4). These size-independent patterns of change are referred to as mosaic change (4). Patterns of mosaic change are not only common to taxonomic groups but also, across taxonomic groups, to species that use the environment in similar ways (5).

Beyond the size of major brain divisions, datasets that allow structural and functional changes to be assessed in a phylogenetic context are rare. Finlay and Darlington (1) already pointed at the pruning or functional reassignment of subsets of neurons that may allow circuit adjustments independent of the link between the size of a brain region and overall brain size. Yet, the relative roles of linkages and mosaic changes in the shaping of functionally defined, interconnected neuron populations remain unknown. Modeling neuronal networks can define at which bounds morphological changes either deteriorate function or result in no further improvement of function (6, 7). But it is not known whether these bounds coincide with the limits of interspecies variation. The expansion of the septo-hippocampal memory system in insectivores and bats may be related to the perceptual representation of complex niches (5). But are there changes within the system that accompany the change in system size? We do not know.

A stumbling block to finding answers is the number of species required to address these questions. Basic phylogenetic statistics requires data from 50 or more species to generate reasonably robust outcomes (8), which until now was not available on a higher resolution than volumetric data of brain divisions. We here provide a basic phylogenetic comparative analysis of neuron numbers of the five principal neuron populations in the hippocampus (Fig. 1A–E) in 65 mammalian species (418 specimens) belonging to 11 orders (Fig. 1F). These hippocampal neuron populations form an interconnected network (Fig. 2A) and have defined roles in the formation of new memories: dentate granule cells (GC) and hilar neurons (HIL) (10, 11), CA3 pyramidal neurons (CA3) (12), CA1 pyramidal neurons (CA1) (13, 14 ), and subicular neurons (SUB) (15, 16). Subsequent to the phylogenetic analyses, we explore the roles of ecological factors in shaping neuron numbers and their relations to each other.

Fig. 1.

Hippocampal neuron populations (A–E) across a mammalian phylogenetic tree (F) with relative neuron population sizes. A–E) Neuron populations are illustrated in Nissl-stained sections of the mid-dorsoventral hippocampus in species selected from our sample (bold species names in F). GC, granule cells; HIL, hilar neurons; CA3 and CA1, CA3 and CA1 pyramidal neurons; SUB, subicular neurons. White arrows and dashed lines (in red) indicate population boundaries. Scale bars: 0.5 mm. F) Sixty-five species of our sample plotted with traits at tips on a tree pruned from the phylogeny of Álvarez-Carretero et al. (9). Afrotheres and marsupials are both represented by species of two orders. Color coding within orders (Rodentia and Chiroptera) represents groups of interest referred to in the text. Circles represent the relative neuron population sizes (trait values centered across species). Black circles indicate values that are larger than the across-species average for a neuron population, and white circles indicate values that are smaller.

Fig. 2.

Connections of the neuron populations within the hippocampus (A) and the log-linear relation of these populations to brain weight (B). A) A simplified survey of monosynaptic connections between the neuron populations assessed in this work. The entorhinal cortex relays information from other cortical areas to the dentate GC. Subsequently, information is passed to CA3, CA1, and SUB neurons. HIL mainly serve in the control of GC activity. CA3, CA1, and SUB all participate in the transfer of information to extrahippocampal targets, indicated by axonal branches, and the axonal feedback to the entorhinal cortex. B) Log-linear relationship between hippocampal neuron numbers and brain weight. As expected from studies of major brain divisions, hippocampal neuron numbers increase as the size of the brain increases. However, the increase is significantly lower in CA3 than in other populations.

Results

Neuron numbers were estimated stereologically using the design-based optical fractionator (17, 18). Statistics were performed on neuron numbers, ratios of interconnected populations, and z-scores calculated for log-transformed neuron numbers of each species, reflecting the allocation of cellular resources to different functional units within the hippocampus, independent of overall structure size (19).

Log-linear correlations and within-species variance

Log-linear correlations were observed between brain weight and the numbers of all five neuron populations (Figs. 2B and S1, 0.74 < r² < 0.87, all P < 0.001). Slopes for GC, HIL, CA1, and SUB were statistically indistinguishable from each other (all P > 0.23), whereas the slope of CA3 was lower than that of other populations (all P < 0.01).

Within-species variance is the substrate for selective pressures to mediate adaptive change. The biological coefficient of variation (CV), as a measure of within-species variance, can be estimated from the observed mean species variance of neuron numbers and the measurement noise (mean coefficients of error originating from the estimation procedure (18, 20)). The estimated biological CV across species (variance as % of the mean) was remarkably similar for HIL, CA3, CA1, and SUB (15–16%), but larger for GC (23%).

Species differentiation and phylogenetic signal

A phylogenetic principal component analysis (pPCA) was applied to define how neuron number z-scores (Fig. 3A) and ratios (source over target, Fig. 3C) differentiate between species and clades. Neuron number z-scores loaded with similar strength on principal component 1 (pPC1) but with opposite signs in GC and HIL neurons versus CA3, CA1, and SUB neurons (Fig. 3B). These two groups of neuron populations roughly represent hippocampal input, intrahippocampal processing, and output. Their relative sizes differentiated input-size-dominant lagomorphs, primates, and artiodactyls (Fig. 3A; large GC and HIL numbers) from processing/output-size-dominant rodents and true insectivores (Fig. 3A; large CA3, CA1, and SUB numbers). pPC2 was dominated by CA3 and SUB loads (Fig. 3B) and separated small echolocating bats (Yangochiroptera; large CA3 and small SUB numbers) from other species, including large fruit bat species (Yinpterochiroptera), which were located among rodent species. Humans were separated from other species along pPC3 (Fig. 3A) driven by the load of CA1, which is large in humans (Fig. 3B).

Fig. 3.

The pPCA of hippocampal neuron numbers (A, B) and ratios (C, D). A) This pPCA evaluates the z-scores of the hippocampal neuron populations sizes. The first three axis of the pPCA accounted for 97.9% of the variance in the data. B) The first principal component, pPC1, was dominated by negative loads of GC and HIL, while CA3, CA1, and SUB neurons pulled in the opposite direction. Along pPC1, clades with dominant hippocampal input (large GC and HIL populations), such as primates, lagomorphs, and artiodactyls, were separated from rodents and true insectivores (Eulipotyphla), in which the processing/output structures of the hippocampal circuitry (CA3, CA1, and SUB) were dominant. Along pPC2, CA3, and SUB sizes separated small echolocating bats (Yangochiroptera) from other species. pPC3 was dominated by CA1 and separated humans (black arrow) from other species. C) This pPCA evaluates the ratios of interconnected neuron populations. pPC1 explained most of the species’ variance (95.5%). True insectivores and rodents dispersed along pPC1, artiodactyls, primates (black arrow: human) and lagomorphs along pPC2, and small echolocating bats along pPC3. D) Loadings identified the GC:HIL and GC:CA3 ratios, again the input side of the hippocampal circuitry, as the strongest contributors to clade variance. The GC:CA3 ratio was the main differentiator along pPC2, while the CA1:SUB ratio was responsible for the dispersion of small echolocating bats along pPC3. For 3D graphs with species identification, see interactive graphs 1 and 2 at Dryad.

In the pPCA of cell ratios, the loads of the ratio of GC to HIL (GC:HIL) and GC:CA3 dominated pPC1 (Fig. 3D) and placed true insectivores (few HIL relative to GC) and small echolocating bats (many CA3 neurons relative to GC) at the extremes of the species distribution, with rodents holding an intermediate position (Fig. 3C). The load of the GC:CA3 dominated pPC2, again separating lagomorphs, primates, and artiodactyls (few CA3 neurons relative to GC) from other species (many CA3 neurons relative to GC, Fig. 3C). The load of CA1:SUB on pPC3 (Fig. 3D) separated small echolocating bats from other species due to their small subiculum (Fig. 3C). Interactive graphs 1 (pPCA of neuron numbers) and 2 (pPCA of ratios) are available at Dryad and provide species identifications and associated pPCA ordinate values.

The clustering of related species observed in the pPCAs suggests that the quantitative composition of the hippocampus is at least partly reflecting the phylogenetic histories of the species. Phylogenetic signal, taking into account the within-species variability (21, 22), was present across all populations (multivariate Blomberg's κ = 0.61, P = 0.0002; individual populations: 0.54 < κ < 0.97, all P < 0.002) with the exception of only a trend for CA1 (κ = 0.39, P = 0.059). Phylogenetic signal was also observed for all ratios (multivariate κ = 0.72, P = 0.0008). Signal was moderate for most individual ratios (0.66 < κ < 0.77, all P < 0.002), but strong for the ratio GC:CA3 (κ = 1.11, P = 0.0006). Statistical outcomes for cell numbers and ratios using κ were congruent with those using Pagel's λ as a test statistic (0.77 < λ < 1; P < 0.0001), including the weak signal for CA1 neuron numbers, which was not significant (λ = 0.50, P = 0.39). Detailed outcomes are presented in Table S1.

Evolutionary scenarios and ancestral state

Next, we tested three evolutionary scenarios of neuron numbers and ratios. First, unconstrained hippocampal evolution is captured by a Brownian motion (BM) model, where lineages diverged from each other constantly as time accumulated. Second, hippocampal configurations evolved but under constraints captured by an Ornstein–Uhlenbeck (OU) model. Under this model, traits could also evolve by BM processes. But as traits increasingly diverged from a mean, restraining forces increased, favoring motion that returned the trait to the mean. Third, a scenario where most of the hippocampal evolution condensed early during the mammalian radiation, as major splits took place, with subsequent evolution slowing down. This scenario is represented by the early burst (EB) model. The three models were compared using the Akaike information criterion (AIC), including the within-species variability in the modeling (23). Detailed outcomes are presented in Table S2. Across all comparisons and with conservative assessments of relative model likelihoods (ΔAIC weights), the constrained OU model was the only dominant one, and it was always a viable alternative when the BM model provided a better fit. The EB model never provided the best fit, and its probabilities were typically lowest (24). For individual populations, the OU model was the best fit for GC and CA1 neuron numbers (relative model likelihoods >0.99). For SUB, OU provided the best fit, but BM and, much weaker, EB models remained viable alternatives (relative model likelihoods: OU = 0.66, BM = 0.26, EB = 0.08). The BM model provided the best fit for HIL and CA3, but the OU and EB models remained viable alternatives (relative model likelihoods HIL: BM = 0.51, OU = 0.33 EB = 0.17; CA3: BM = 0.6, OU = 0.2, EB = 0.21). None of the models provided an unequivocal best fit for any of the neuron ratios. The BM model provided the best fit for most ratios (0.50< relative model likelihood <0.57), while the OU model performed best for GC:HIL (relative model likelihood = 0.46). The EB model performed worst for all ratios.

How may an ancestral quantitative morphotype have looked like? We used ancestral state modeling considering both within-species variance and the most likely evolutionary model found above (BM or OU, see Table S3). The z-scores of the ancestral phenotype at the deepest node of our species sample, marking the split of eutherian (placental mammals) from metatherian mammals (marsupials), were estimated to be 1.36 (GC), −1.18 (HIL), −0.15 (CA3), 0.37 (CA1), and −0.39 (SUB). Mean species neuron numbers of our tree (log₁₀) and their variances were used to transform ancestral z-scores to neuron number estimates: 1,690,000 (59%) GC, 93,000 (3%) HIL, 302,000 (10%) CA3, 549,000 (19%) CA1, and 230,000 (8%) SUB. The ancestral state is illustrated by open squares in Fig. 4. The species deviating least from the ancestral pattern (Table S4) are wood mice, followed by European moles and common brushtail possums. Furthest from the ancestral state are all mole-rats (rodent family Bathyergidae).

Fig. 4.

Scatter plots of z-scored neuron numbers (A–C) and the impact of CA1 and SUB numbers on z-scores of the input neuron populations (D–F). A) Hippocampal input populations, i.e. GC, HIL, and CA3 pyramidal neurons (CA3), were confined to a banana-shaped cloud, with Bathyergidae (small GC and HIL and large CA3) defining one end of the spectrum and primates, artiodactyls, and the eastern rock sengi, the only species of Macroscelidea in our sample, positioned at the other end (large GC and HIL and small CA3). Most of the species in which a reflected blade of CA3 (black asterisk) can be seen are located in this area. B) Populations of the classic trisynaptic loop (GC, CA3, and CA1) contained the interdependence of GC and CA3, but CA1 neuron numbers spread the species in our sample, with no discernible cross-clade pattern. C) Output neuron populations CA3, CA1, and SUB appeared least constrained. For interactive 3D graphs with species identification, see interactive graphs 3–5 at Dryad. D–F) Human, dog, and domesticated rabbit data points stand out in the input scatter plots when z-scores are calculated based on all five neuron populations (D). When CA1 is excluded from all species z-scores (E), the rabbit returns to the position of its wild type. Human and dog shift towards the remaining species. When SUB is excluded (F), the dog data point shifts strongly towards the remaining species. The positions of human, dog, and domesticated rabbit in the input plots are due to their large CA1 (human, rabbit) and SUB populations (human, dog).

Neuron population relations along the hippocampal circuit

The limited dimensionality of our data (five for neuron populations) allowed the direct visualization of relations between neuron numbers in scatter plots of the z-scores. Within the input (GC → HIL and GC → CA3; Fig. 4A, for an interactive 3D plot, see Dryad interactive graph 3), relations between neuron number z-scores were mostly confined to a cloud curving along the borders of the plot. Even though increases in relative GC numbers were accompanied by increases in HIL numbers, 10-fold differences in the ratios between GC and HIL were observed (Fig. S2A). Relative CA3 numbers decreased as GC numbers increased, resulting in a >20-fold difference in the ratio between the two populations (Fig. S2B). Marking the presence of a qualitative trait, the reflected blade of CA3 (Fig. 4A) showed that it was invariably present in species with the largest relative GC numbers. Focusing on the classic trisynaptic loop (GC → CA3 → CA1 neurons; Fig. 4B, for an interactive 3D plot, see Dryad interactive graph 4), the interdependence of granule and CA3 neuron numbers was still a distinctive feature, but shifts in CA1 numbers expanded in the plot. In the output plot (CA3 → CA1 → SUB; Fig. 4C, for an interactive 3D plot, see Dryad interactive graph 5), ratio differences were limited to <6-fold differences across all species (Fig. S2C and D). Species that showed weak dependencies between CA3 and CA1 within the trisynaptic loop were further dispersed by their SUB numbers.

In initial visualizations of the input neuron numbers, two species stood out—human and domestic dog. To explore the role of domestication in shaping hippocampal circuitry, we included data from domesticated forms of pig, house mouse, and rabbit in the scatter plots. These groups were not part of any other analysis. Domesticated pigs and house mice (laboratory C57BL6 and DBA strains) largely retained the positions of their wild types, while the domestic rabbit diverged from other species and its wild type in a manner similar to humans and dogs (Fig. 4D). Excluding either CA1 or SUB numbers from the calculations of the z-scores moved the human, dog, and domesticated rabbit data points back (partly or completely) towards the cloud of the other species (Fig. 4E and F). Their distinct positions in the input scatter plot based on z-scores over all five neuron populations (Fig. 4D) are due to their large CA1 and/or SUB numbers.

Correlation of ecological factors with hippocampal cell numbers and ratios

The selection of ecological factors was based on reports of factors that impact hippocampal volumes in primates (25, 26) and their broad applicability to our species sample. We assessed diet breadth, habitat breadth, home range, social group size, and trophic level. Vetted factor values from an extensive collection of mammalian ecological data (PanTHERIA (26)) were available for 56 species (referred to as “56 data”). Estimates were supplemented for the remaining nine species (total of 65 species; “65 data”). After selecting the best-fitting phylogenetic model, we compared the ecological model with the null model and step-fitted the model to identify the ecological factors that improved the fit. Main outcomes were similar for both datasets (Fig. 5). Effect sizes are provided as partial r² values, i.e. the proportion of the variance in neuron numbers (Fig. 5A) or ratios (Fig. 5B) explained by the ecological factor. Effect sizes and P-values are reported as ranges of values when both the 56-data and the 65-data generated significant outcomes. For all outcomes, consult Table S5.

Fig. 5.

Effect sizes of ecological factors on neuron numbers (A) and ratios (B). Plots provide full model PGLS effect sizes in terms of the slopes of the regression line (β) and confidence intervals. Effect sizes in terms of the variance explained by ecological variables, i.e. partial r², are reported in the main text. (A) Diet breadth was a significant factor for relative GC and HIL neuron numbers. Mammals relying on many different food categories tend to have higher relative GC and HIL numbers. Although β values appear numerically small, relative neuron numbers approximately double across the range of diet breadth from 1 to 7. There were no major differences between testing the 56 data (gray) and 65 data (black, estimates of ecological factors supplemented for nine species). Effect size for CA3 also identifies diet breadth as a contributor, but in this comparison, the ecological model does not explain neuron numbers significantly better than the null model. B) Home range size was a significant factor for the ratios GC:HIL and CA3:CA1. Species with small GC:HIL or CA3:CA1 ratios tend to have larger home ranges. Across the home ranges, β values allow for ∼2.5-fold changes in the ratios.

For neuron numbers, diet breadth, that is, the number of food categories (ranging from 1 to 8) consumed, had the strongest and most consistent impact (Fig. 5A). Significant improvements of fit by the full ecological model were seen on the hippocampal input side for GC and HIL (ecological vs. null models: 0.001 < P < 0.03). Diet breadth contributed significantly to the improved fit for GC and HIL (0.29> effect size >0.18, 0.001 < P < 0.002). A trend for an ecological impact on CA3 neuron number (P = 0.068) was seen in the 56 data. Diet breadth was the responsible factor (effect size = 0.09, P = 0.024). Ecological factor impact on CA1 neuron numbers was not significant, although diet breadth again improved the fit. There was a trend for an ecological impact on SUB neuron numbers (0.067 < P < 0.08) with, again, an impact of diet breadth (0.13> effect size >0.09, 0.02 < P < 0.03) and a weaker contribution by social group size (number of individuals in the social group). In keeping with the effects of diet breadth on individual neuron populations, diet breadth also explained a substantial amount of the variance along pPC1 for neurons numbers, which is the strongest differentiator between species. The ecological model improved the fit (0.001 < P < 0.025), and diet breadth was the only factor identified to significantly improve the fit (0.38> effect size >0.16, 0.001 < P < 0.006).

For ratios between neuron populations (Fig. 5B), an improvement of fit by the full ecological model was seen for GC:HIL and GC:CA3 (0.002 < P < 0.023). For GC:HIL, home range (in km²) had the strongest effect (0.23> effect size >0.18, 0.002 < P < 0.008) with a lesser contribution by trophic level (herbivore, omnivore, or carnivore). Diet breadth impacted on the GC:CA3 ratio (0.19> effect size >0.17, 0.002 < P < 0.003). Outcomes for CA3:CA1 only showed a trend for the 65 data (P = 0.06), with an impact of home range in both datasets (0.09> effect size >0.07, 0.02 < P < 0.04). No effects were observed in any test of CA1:SUB. When the analysis was run on ordinates along pPC1 for ratios, the ecological model improved the fit (0.001 < P < 0.004). Home range (0.28> effect size >0.23, 0.001 < P < 0.003) and trophic level (0.28> effect size >0.08, 0.02 < P < 0.04) improved the fit. In contrast to the widely distributed effect of diet breadth on neuron numbers, home range impacted ratios more selectively. Along pPC1, the effect of home range was mainly due to GC:HIL.

Discussion

Hippocampal input neuron numbers and ecologically driven species differentiation

A consistent narrative develops from the phylogenetic assessment of hippocampal input neuron numbers and their ecological correlates. GC have the highest biological variability within species. GC are also strongly associated with species differentiation along pPC1 of the pPCA, which should be least dependent on the phylogenetic history of the species but instead emphasizes the relationship between phenotype and associated function (27). A strong load on pPC1 is also found for HIL. The good performance of the constraining scenario (OU) for GC:HIL is biologically plausible in that one would not expect HIL numbers to vary independently from the number of GC that they control (10). The relative numbers of GC and HIL correlate with an ecological factor, diet breadth, that defines how many types of food are consumed by a species.

A parsimonious explanation of these findings is selective change of a quantitatively variable population in response to access to food as a critical ecological factor. This explanation is also in line with the idea that neuron populations forming late and over extended periods mediate neocortical expansion and diversification (28). GC are the last principal neurons forming during ontogenetic development, and their formation is extended into adulthood in many mammals (29). Also, we previously found that the latest, adult-formed GC were strong differentiators between closely related species (19). They shared this role with HIL, which can be both across-clade (present results) and within-clade differentiators (19).

HIL deserve special attention in this narrative. In all species, the hilar population encompasses “true” HIL at the GC–hilar interface, the subgranular layer, and the hilar polymorphic cell layer. In the species with the largest GC numbers, HIL are supplemented by a subset of CA3 pyramidal cells located in the reflected blade of CA3. The reflected blade is a specialized section of CA3 that integrates with the granule and hilar populations anatomically, physiologically, and functionally (30, 31). We cannot tell whether this repurposing of CA3 neurons is due to computational advances in pattern separation afforded by the involvement of CA3 neurons in GC control (32, 33), or whether GC numbers simply outpaced the potential for change in the number of true HIL in species with a reflected blade.

Why may diet breadth impact neuron numbers but habitat size neuron ratios? Neural network modeling of hippocampal function typically emphasizes connectivity between nodes (neurons), which determines the number of items that can be stored/remembered by the network. Node number instead relates to the information content/richness of the item (34). An increase in diet breadth may increase, in any one location (item), the number of salient cues (information content) that provide access to a critical resource, food. In contrast, with an increasing home range, there will be an increasing number of places to discriminate and remember. As home range increases, the ratio GC:HIL decreases, which may facilitate a key function of the dentate gyrus, pattern separation (10, 11), in the discrimination between increasing numbers of places. For single ecological factors, diet breadth and home range explain sizable amounts of the variation in relative neuron numbers and ratios, but, of course, only part of the total observed variation. There is room for other ecological factors that may correlate with the number of items and/or their information content to be effectively encoded in memories.

From the slower increase of CA3 pyramidal neuron number with body weight, the presence of a reflected blade in many large-bodied species, and the correlation of home range and body size (35), one may be tempted to identify the preceding narrative as a large brain–small brain story. This is not so. The eastern rock sengi, the only member of the Macroscelidea in our sample, weighs only ∼50 g but has very large GC and HIL numbers and a very small CA3 neuron number. Consequently, it holds some of the most extreme positions in both the pPCAs for neuron numbers and ratios and in scatter plot visualizations. Also, small-bodied members of true insectivores (European mole) and rodents (rock mouse and guinea pig) are located among the large-bodied artiodactyla and primates, sharing their input-dominated hippocampi and the presence of a reflected blade. While there is no very large-bodied and output-dominant species in our sample, they do exist. California sea lions (>100 kg body weight) have a very low GC:CA3 ratio of 1.57 (36), which is in the range of bats or the lower range of rodents. Also, California sea lions, like African elephants (37), lack a CA3 reflected blade.

Hippocampal output neuron numbers

Output neuron population measures of related species clustered in various analyses, but in contrast to input neuron populations, no clear pattern of intrahippocampal change across clades emerged. Diet breadth was identified as a factor that had an impact several times. But ecological models generally did not perform better than null models, and some output neuron numbers or ratios were not affected by any of the ecological factors selected for testing. While there are several interesting findings, a narrative consistent across our findings does not evolve.

Noteworthy is the size of CA1 relative to CA3. It has been suggested that information transfer from CA3 to CA1 deteriorates quickly if the CA3:CA1 ratio drops below 1 and becomes asymptotic if the ratio increases above 2 (6, 7). The ∼2:1 ratio for the ancestral CA3:CA1 phenotype corresponds well to the optimum for information transfer. The lower bound is respected by all clades in our sample except for the rodent mole-rats. The ratio is, however, exceeded by many clades. The observation may be resolved by CA3 serving multiple, functionally distinct CA1 pyramidal cell populations that are known to be present, but that we have not resolved morphologically in the present work (see “Caveats, summary, and perspectives” section).

Changes in relative neuron numbers may, of course, also be related to interactions of output neuron populations with their targets in other brain divisions. pPCA outcomes contain some evidence to this effect. With the caveat of small species numbers, true insectivores, primates, and bats distributed largely orthogonally in the analysis of cell ratios. This distribution, including the separation and relative positions of small echolocating bats and fruit bats, resembled that in a PCA of the sizes of their major brain divisions (5) and may reflect intrahippocampal correlates of changes in macroscopic brain organization. But the primate, true insectivore, and bat directions were not exclusive to these clades. Artiodactyls and lagomorphs shared direction with primates, and rodents with true insectivores.

A surprising observation is the weakness of the phylogenetic signal for the CA1 pyramidal neurons, because the volumetric expansion of CA1, being largest relative to body weight in humans (38), is often quoted as an example of phylogenetic change within the hippocampus. Both in our sample and a sample of primates (25), other species have a CA1 that is relatively larger than in humans in terms of relative neuron numbers (small echolocating bats) or volume (other primates). Changes in the relative size of CA1 should be interpreted as species-specific adaptations that can occur in multiple clades, but that, in the present sample, do not characterize an entire clade.

The positions of human, dog, and rabbit did not initially fit with the remainder of the species in the plot of hippocampal input structures. But positions returned to the species cloud when the influence of unusually large SUB or CA1 neuron numbers were removed. In essence, GC, HIL, and CA3 typical of input-dominant hippocampi have been merged with oversized outputs in these species. Humans and dogs share a common habitat and a history of domestication (39, 40) and social interactions that left traces in their brain structure (41) and function (42). Domestication-related genomic changes have also been found in rabbits, in which they frequently targeted genes regulating nervous system development (43).

Caveats, summary, and perspectives

Although we present hippocampal neuron populations as homogenous units, they are not. Function and anatomy differentiate along each hippocampal axis (dorsoventral (44), proximodistal (31), and radial (45, 46)), and these differences are gradually coalescing into concepts of functional streams within the hippocampus that consist of subsets of the principal neuron populations (31, 47). Selective pressures may also shape these streams, and our data only represent the population sum totals. Also, while our sample appears large, it only comprises ∼1% of the extant mammals. Additional species may expand the range of relations, bridge the spaces between outliers and other species in the present sample, and allow for the refinement of phylogenetic analyses. Lastly, neuron numbers only represent one of many hippocampal traits that differ between species.

In summary, circuit-level phylogenetic analyses can reveal the evolutionary scenarios and possible ecological drivers of change that result in different quantitative morphological phenotypes. In our case of hippocampal neuron numbers, it points to changes in hippocampal input, driven by critical resources, as differentiators between species, while differences in output may reflect changes in macroscopic quantitative organization of the brain. Additional species data can be easily integrated into our dataset to provide phylogenetic information on species specialization and to permit the refinement of the phylogenetic analyses across clades. Our observations also allow the identification of species with quantitatively interesting hippocampi that may serve to address special-interest questions. The ancestral state estimate may permit theoretical neural network evaluations that define basic hippocampal computational properties, what may be gained (and lost) by changes that are observed in clades, and which computational constraints may be responsible for the phylogenetic morphological constraints that were observed.

Materials and methods

Animals and tissue preparation

All data are included in the specimen-based primary dataset available at Dryad. The file contains the species included in this study, taxonomic information, sources, and, if applicable, permits. Animal work was conducted in accordance with the laws and regulations of the Swiss Federal Veterinary Office and institutional requirements. References are provided for species presented previously. The file also contains sex, brain weight, body weight, and age, if known. For a species to be included in the analysis, the minimum sample size was 3. The average sample size was 5.4 (excluding humans, n = 73). Samples were sex balanced when possible. Criteria for the inclusion of data published by other groups were design-based stereological estimates, compatible anatomical definitions, and reporting of the data for individual specimens. The file also contains data from domesticated mice, rabbits, and pigs, which were only used for scatter plot visualizations. Additional species are included in the dataset but were not analyzed due to either n < 3 or their absence from the phylogenetic tree used.

Brains were extracted postmortem and immersion fixed in 4% phosphate-buffered paraformaldehyde containing 15% picric acid. Small brains were processed in toto, whereas the hippocampi and entorhinal cortex of large brains were dissected. One hemisphere or hippocampal region was cryoprotected, frozen, and stored at −80 °C. The second was embedded in methacrylate, serially sectioned at 20 μm, and Giemsa stained (30). Sections perpendicular to the long axis of the temporal hippocampus facilitate the definition of neuron populations. We therefore used horizontal sections in species in which the disposition of the hippocampus resembles that of laboratory mice and coronal sections in species in which the temporal hippocampus turns rostrally.

Stereology

Unilateral neuron number estimates were performed using the optical fractionator (17, 18) as implemented in StereoInvestigator Version 10 (MBF Bioscience, Willington, Vermont) using ×63 (NA 1.4) or ×100 (NA 1.3) oil immersion objectives. Series spanning the entire dorsoventral (septotemporal) extent of the hippocampus were analyzed. The optical fractionator belongs to the design-based stereological methods, which provide estimates approaching the true number as the sampling intensity increases, independent of assumptions about the geometry of the objects counted (their size, shape, or distribution) and independent of presectioning tissue shrinkage. By default, we used number-weighted section thickness to estimate the total number (48). All details of the sampling schemes used for each neuron population in each species are contained in the primary dataset.

The precision of the neuron number estimates was assessed by coefficient of error (CE) estimates (18, 20) calculated with a smoothness factor of m = 0, which provides conservative CE estimates (49, 50). We aimed for CEs of 0.1 or less, i.e. repeated estimates should have a standard deviation amounting to 10% or less of their mean. Mean CEs were compared with observed CV (CV = SD/mean) of each neuron population in each species. The variance of neuron numbers should originate mainly from biological variability and not from poor estimates, i.e. CV²/CE² < 0.5. This was the case for most of the estimates. Both CV and CE values are statistics that are based on small samples. Occasional values >0.5 must therefore be expected and did occur unclustered and infrequently. In all cases, they were due to small CVs rather than large CEs. CE estimates for m = 0 (conservative estimate) and m = 1 (optimistic estimate) are contained in the primary dataset of hippocampal neuron numbers at Dryad. CE estimates can also be used, together with the observed CV, to estimate the natural variance of neuron numbers: CV_observed² = CV_biol² + CE². We estimated the biological variance using CE estimates based on both m = 0 and m = 1. The two estimates did not differ by more than 0.5–1.3%, and their means were used to calculate the coefficients of variation of each neuron population in each species. Finally, the mean CV across all species was calculated to obtain a measure of the variability of the neuron populations.

Definitions of neuron populations

In all regions, the differentiation between neuronal and glial nuclei is straightforward and based on the smaller size, more irregular shape, finely granular chromatin, and absence of a distinct nucleolus of glial cell nuclei, and the sparse and weakly or unstained glial cytoplasm (51, 52). The definition of neuron populations relied on traits visible in Nissl-stained sections. We previously published a descriptive summary of identification criteria (30). The absence of mossy fibers as one defining characteristic of CA1 was not used in the hippocampi of marsupials and the European hedgehog. In these species (and sheep and Siamese cats (53, 54)), a thin band of mossy fibers extended for a variable distance into, otherwise histologically typical, temporal CA1 (55, 56). We have estimated neuron numbers in the reflected blade of CA3 separately whenever it was possible. These estimates, although not used in this study, are included in the primary dataset of hippocampal neuron numbers and were added to the HIL numbers for the present analysis (30).

For most species, we also verified neuron population boundaries by using one or more additional stains, both classical (e.g. acetylcholinesterase or Timm) or immunocytochemical (e.g. calbindin, parvalbumin, calretinin, NECAB1-3; for examples of methods and outcomes, see (30, 57–62)). Population boundaries along the hippocampal dorsoventral axis have been illustrated for several species (49, 56, 63–67). Population boundaries may be most difficult to define in the human hippocampus. However, human neuron numbers, which were included in the data we analyzed, were published by several independent groups and largely or completely overlapped (19).

Phylogenetic statistics

Tests were performed on neuron number estimates and two measures derived from them. First, number estimates were log₁₀-transformed and z-scored, which preserves the relative sizes of the number estimates while providing a metric independent of absolute size. Second, ratios between source and target neuron numbers were calculated as measures of the convergence or divergence of information flow in each pair of interconnected neuron populations. We used the phylogeny compiled by Álvarez-Carretero et al. (9), which is resolved at the species level and provided the best overlap with our data. The plot of the tree with trait values at tips in Fig. 1F was prepared using the R adephylo package (68). Analyses were repeated for both neuron numbers and ratios. All tests were conducted in R version 4.3.2.

Phylogenetic principal component analysis

To explore evolutionary trends in the multivariate space defined by the neuron populations and visualize patterns beyond the effect of shared evolutionary history, we used pPCA (27, 69). We employed the phyl.pca function from the R phytools package (70). The pPCA finds a rotation of the variable space that minimizes the impact of phylogenetic signal on the first principal component (pPC1). Phylogenetic signal is redistributed to lower pPCs. For 2D visualization, we used ggplot2 (71). 3D interactive graphs provided at Dryad were generated using the plotly package (72).

Phylogenetic signal

To test whether species tend to resemble each other more or less through time, we used two different methods for measuring the phylogenetic signal. Tests for phylogenetic signal differ in their sensitivity to the size of the phylogeny, the strength of the phylogenetic signal, how reliably the lengths of evolutionary paths reflect the true underlying relationships among species in a phylogenetic tree, and the complexity of the underlying evolutionary processes (8, 73). While Bloomberg's κ (74) is potentially most informative, it has higher type I and II error rates in small phylogenies with inaccuracies in path lengths (8). We therefore calculated both Bloomberg's κ and Pagel's λ (75). Multivariate κ and λ were estimated on the combined dataset for neuron population sizes and combined ratios using the geomorph package (76, 77). κ and λ of individual neuron populations and ratios were estimated using the phytools package (78).

Evolutionary processes

To explore the evolutionary processes behind the observed pattern of hippocampal neuron numbers, we performed a set of evolutionary modeling tests. Tests were performed for BM, OU, and EB models. Brief model descriptions can be found in the Results section on evolutionary scenarios. Because of the z-scoring of the data, which generates perfectly constrained multivariate data, we restricted analysis to univariate testing. We performed these tests using Geiger v2.0 package (79) and included within-species variation (SEs of the species-level data) in the analysis. The best model was deemed to be the one with the lowest AIC.

Ancestral state estimation

The ancestral state at the root of our tree was extracted from the phylopars function of the Rphylopars package (80). Within-species variation was included, and models of evolution were specified for each neuron population based on the outcomes of the previous section. Estimated ancestral state scores were subtracted from the mean scores of each neuron population and species to obtain the cumulative deviation (Δz-score) of each species from the ancestral score (Table S4). Z-scores of the ancestral state were back-transformed into putative neuron numbers by, first, multiplying each population z-score of the ancestral state with the across-species mean SD of centered data (0.4966), second, adding the across-species mean of the log-transformed data (5.5548) and, finally, inverse log transformation.

Phylogenetic generalized least square method

To test the association between the hippocampal morphology and several ecological factors (diet breadth, habitat breadth, home range, social group size, and trophic level), we used phylogenetic generalized least squares (PGLS) models. Ecological factor data were originally taken from PanTHERIA (81) and extensively vetted and, when necessary, adjusted. Ecological data, mostly on social group size and home range, could not be found for eight species. Instead, they were estimated based on related evidence, e.g. flight distance in bats for home range estimates. All adjustments and estimations were finalized prior to any statistical analysis. Data on human social group size and home range in the literature span such wide ranges that any value selected for analysis may be considered an estimate. All tests were therefore performed twice: once on available ecological data (56 species) and once on the ecological data supplemented by estimates (65 species, including humans). Ecological factor data and sources are available on Dryad.

We ran PGLS using the gls function in the R packages nlme (82) and ape 5.8 (83). In PGLS, three candidate BM models were created, investigating correlations across different Pagel's λ values (0, 0.5, and 1). Initially, we also included OU models in this analysis, but α values that distinguished BM and OU models were so small (<0.001) as to be irrelevant in the context of the present study. Subsequently, the best-fitting candidate models were tested against null models to keep type I errors at the nominal rate (84). To identify traits that effectively contribute to significant candidate models, we conducted AIC-based stepwise model selection across all traits.

Supplementary Material

Supplementary material is available at PNAS Nexus online.

Funding

I.A. received funding for this study from the EMDO Foundation, Zurich, Switzerland, and the Prof. Dr. med. Karl and Rena Theiler Haag Foundation, Zurich, Switzerland.

Author Contributions

Jovana Maliković (Conceptualization, Data curation, Formal analysis, Investigation, Methodology, Resources, Visualization, Writing—review & editing), Juan L. Cantalapiedra (Formal analysis, Methodology, Writing—review & editing), Lorenzo Vinciguerra (Project administration, Resources, Writing—review & editing), Katja Schönbächler (Project administration, Resources, Writing—review & editing), Ana Luiza F. Destro (Resources, Writing—review & editing), Jennifer Rodger (Project administration, Resources, Writing—review & editing), Marielle Jörimann (Investigation, Resources, Writing—review & editing), Liora Las (Resources, Writing—review & editing), Stephen G. Hörpel (Resources, Writing—review & editing), David P. Wolfer (Funding acquisition, Project administration, Resources, Supervision, Writing—review & editing), Lutz Slomianka (Conceptualization, Formal analysis, Investigation, Methodology, Supervision, Validation, Visualization, Writing—original draft, Writing—review & editing), Irmgard Amrein (Conceptualization, Data curation, Formal analysis, Funding acquisition, Investigation, Methodology, Project administration, Resources, Supervision, Validation, Visualization, Writing—review & editing)

Data Availability

The primary data of neuron numbers (EcoEvoHippo_Neuron_numbers_Hippocampus.xlsx) and ecological factors (EcoEvoHippo_Ecological_Factors.xlsx), interactive 3D graphs, and R code are available at https://doi.org/10.5061/dryad.kd51c5bht.