Department of Zoology, University of Cambridge, Downing Street, Cambridge, UK

Evolutionary Anthropology Research Group, Department of Anthropology, Durham University, Dawson Building, South Road, Durham, DH1 3LE, UK

Abstract

Background

Brain size is a key adaptive trait. It is often assumed that increasing brain size was a general evolutionary trend in primates, yet recent fossil discoveries have documented brain size decreases in some lineages, raising the question of how general a trend there was for brains to increase in mass over evolutionary time. We present the first systematic phylogenetic analysis designed to answer this question.

Results

We performed ancestral state reconstructions of three traits (absolute brain mass, absolute body mass, relative brain mass) using 37 extant and 23 extinct primate species and three approaches to ancestral state reconstruction: parsimony, maximum likelihood and Bayesian Markov-chain Monte Carlo. Both absolute and relative brain mass generally increased over evolutionary time, but body mass did not. Nevertheless both absolute and relative brain mass decreased along several branches. Applying these results to the contentious case of

Conclusions

Our results confirm that brain expansion began early in primate evolution and show that increases occurred in all major clades. Only in terms of an increase in absolute mass does the human lineage appear particularly striking, with both the rate of proportional change in mass and relative brain size having episodes of greater expansion elsewhere on the primate phylogeny. However, decreases in brain mass also occurred along branches in all major clades, and we conclude that, while selection has acted to enlarge primate brains, in some lineages this trend has been reversed. Further analyses of the phylogenetic position of

Background

Phylogenetic comparative methods and ancestral state reconstruction play important roles in evolutionary biology. They enable historical evolutionary processes, and the function and evolution of specific traits, to be inferred from patterns of diversity in extant species

Recent studies, however, indicate that brain size, measured either in volume or mass, may have decreased in some vertebrate lineages

Although many studies have investigated the possible selective advantages and disadvantages of increased brain size in primates

The importance of understanding the evolution of reduced brain size in primates has recently been brought into sharp focus with the discovery of a small-brained hominin,

Phylogeny of primates with extinct primates

**Phylogeny of primates with extinct primates**. a) Phylogeny used for main reconstruction analysis. Extinct primates are denoted with an asterisk (*); b) and c) Phylogenies off

Reconstructing ancestral brain and body sizes provides a means of testing the generality of the trend of increasing brain and body size through primate evolution. It also provides estimates of brain and body sizes at key points along the primate phylogeny allowing inferences to be made about the ecology of the ancestors of key clades, based on what we know about the relationship between body size, ecology and life history traits in living primates

Here we investigate the evolution of brain and body mass in Primates and assess whether brain mass and body mass show evidence of directional trends. Given the strong allometric relationship between brain and body mass

First, we adopted three approaches to reconstructing the evolutionary history of these traits: weighted squared-change parsimony, maximum likelihood (ML) and Bayesian Markov-chain Monte Carlo (MCMC)

We discuss the implications of our results for hypotheses on the adaptive origins of modern primates, and identify branches along which brain mass has increased greatly or at a high rate, or along which brain mass has decreased in either absolute or relative terms. Finally we use our results to evaluate alternative scenarios about the origin of

Results and discussion

Ancestral reconstructions: congruence between estimates with and without fossil data

Previous studies have used either volume or mass as a measure of brain size. Here we used log_{10}-transformed brain and body mass in all analyses. Brain and body mass estimates were collected for 37 extant and 23 extinct primate species (Additional file

**Supplementary tables and figures**. 1. Table S1: Brain and body mass of primates used in the analyses. 2. Table S2: Posterior distribution of the scaling parameters to identify the best model before reconstructing ancestral states in Bayesian analysis. 3. Figure S1: Correlations between estimates made using directional constant variance random walk and non-directional constant variance random walk models in BayesTraits. 4. Table S3: Ancestral state estimates using most supported models. 5. Table S4: Change in absolute brain and body mass and relative brain mass along each branch. 6. Additional analyses in relation to

Click here for file

Phylogeny of extant primate genera

**Phylogeny of extant primate genera**. Branches are drawn proportional to time. This figure was prepared in Mesquite

Ancestral values for all nodes of the tree with and without the inclusion of fossil data were highly correlated for absolute brain mass (parsimony, Spearman's correlation coefficient (_{s}) = 0.932; ML, _{s }= 0.932; Bayesian MCMC, _{s }= 0.993, all _{s }= 0.939; ML, _{s }= 0.941; Bayesian MCMC, _{s }= 0.960, all _{s }= 1.000, _{s }= 1.000, _{s }values in the parsimony and ML analyses are caused by increased disparity between the estimates at deeper nodes. In particular, estimates of log(brain mass) using fossils are 10-15% lower for the root (Figure

Correlations between estimates of absolute brain mass in log(grams)

**Correlations between estimates of absolute brain mass in log(grams)**. **a) **Correlations are shown with and without fossil data using ML; **b) **with and without fossil data using Bayesian MCMC; **c) **without fossil data between ML and Bayesian MCMC results; **d) **with fossil data between ML and bayesian MCMC results. Numbers indicate nodes in figure 2.

The results from the Bayesian analyses agree more strongly with ML when fossil data are included than when they are excluded for both brain mass (with fossils: _{s }= 0.995, _{s }= 0.923, _{s }= 0.981, _{s }= 0.926,

To measure relative brain mass for the species in the tree we performed a phylogenetically controlled GLS regression analysis between log(brain mass) and log(body mass) using ML in BayesTraits (see Methods), that returned the following fit line: log(brain mass) = 2.18 + 0.684 [log(body mass)]. We then reconstructed ancestral character states for relative brain size with two alternative approaches. With the first approach, which we term _{s }= 0.979,

The level of congruence between ancestral state estimates made with and without fossil data was much lower for relative brain mass than for absolute brain mass. Spearman's rank correlations for all three methods were highly significant although _{s }values were much lower for ML (_{s }= 0.743, _{s }= 0.743, _{s }= 0.835, _{s }= 1.000, _{s }= 0.994 with fossils, _{s }= 0.968 without; Figure

Correlations between estimates of relative brain mass

**Correlations between estimates of relative brain mass**. **a) **Correlations are shown with and without fossil data using ML; **b) **with and without fossil data using Bayesian MCMC; **c) **without fossil data between ML and Bayesian MCMC results; **d) **with fossil data between ML and bayesian MCMC results. Numbers indicate nodes in Figure 2.

It is interesting to note that for all three approaches the estimated brain mass of the last common ancestor of humans and chimpanzees is larger when fossil data are not included. For example, the average estimate of brain mass for the

To summarise, parsimony and ML produced ancestral state estimates that were more discrepant between analyses with and without fossils when compared to estimates obtained with Organ

One possible reason for lower consistency in estimates with and without fossils, particularly for ML and parsimony, might be the presence of directional trends

Evolutionary trends in body and brain mass evolution

We tested for evolutionary trends by comparing a directional random-walk model to the non-directional random-walk model in BayesTraits. The implementation of the directional model requires variation in root-to-tip branch length

We found no evidence for a directional trend in absolute body mass, as the directional model did not provide a better fit to the data when compared to the non-directional model (Table

Posterior distributions of log-likelihoods for the non-directional and directional models

**Posterior distributions of log-likelihoods for the non-directional and directional models**. Figure **a) **shows body mass; **b) **brain mass; **c) **relative brain size. The log-likelihood of the directional model is shown in red, the non-directional model in blue. The posterior distributions of ancestral state estimates were obtained using uniform priors, two million iterations and a sampling interval of 100 (see Methods). The harmonic means and Bayes Factors of the posterior distributions are given in Table 1.

Tests for evolutionary trends^{1}.

**Phenotype**

**Harmonic mean Log(Lh): Constant-variance model**

**Harmonic mean Log(Lh): Directional model**

**Bayes Factor**

Absolute body size

-44.688

-45.275

-1.174

Absolute brain size

-30.282

-27.087

6.390

Relative brain size

1.647

8.576

13.857

^{1 }The model with the highest log-likelihood is the best-fitting model.

To assess how the presence of a directional trend affects the accuracy of estimates made with a non-directional model we performed correlations between the results obtained from the directional and non-directional models. Correlations between the estimates made for absolute brain mass under the directional and non-directional models suggest no nodes are estimated less accurately than others under the non-directional constant-variance model (_{s }= 0.995, _{s }between estimates under directional vs. non-directional model is lower (r_{s }= 0.943,

Taken together our results suggest that the ancestral state reconstruction procedure implemented in Bayesian framework following Organ

Having obtained the most reliable estimates of ancestral states at each node in the tree for each phenotype it is then possible to use these to make evolutionary inferences. For example our most supported estimate of the body mass at the root of the primate tree using Bayesian analysis is largely consistent with some previous qualitative estimates. Martin

Posterior distributions of trait estimates for the LCA of living primates for a) body mass and b) brain mass

**Posterior distributions of trait estimates for the LCA of living primates for a) body mass and b) brain mass**. Histograms are plotted from a posterior distribution of ancestral state estimates obtained using uniform priors (prior range: -100 to +100) acceptance rates were within 20 to 40% (see methods). To ensure the chain fully explored the parameter space, we extended the MCMC run to 25 million iterations with a sampling interval of 1500.

Body mass variation is associated with a number of behavioural, ecological and life history traits

Our reconstructions suggest the ancestral primate had a small brain (120.23 mg, 95% CI: 114.42 mg to 126.33 mg) which, in relative terms, was much smaller than in any living primate. This result is consistent with a study of a virtual endocast of ^{3}, 95% CI: 336.61 cm^{3 }to 357.17 cm^{3}) is similar to estimates for the two earliest hominids known from the fossil record, the 7.7 million year old ^{3};) ^{3})

Increases in brain mass in particular lineages

We next examined the amount of change along different branches of the tree, both as total change along the branches and as rate of change accounting for differences in branch lengths (see below). We first calculated the means of the posterior distribution of the ancestral states for each node, using the same posterior predictive model developed for brain and body mass, and we computed the change in absolute brain and body mass, and relative brain mass (using the

Evolutionary trajectories of brain and body mass

**Evolutionary trajectories of brain and body mass**. Evolution of brain (red) and body (blue) mass from the ancestral primate to **a) ****b) ****c) ****d) **

Changes in absolute brain mass along each branch of the phylogeny can be considered in two ways: a proportional increase (as % of increase relative to the ancestor) and an absolute increase in mass (described above). The average proportional change in absolute brain mass along a branch is 0.243 (i.e. a 24.3% increase), with changes greater than 0.344 being in the upper quartile, which includes branches from all the major clades of the phylogeny. Notably, three of the top four proportional increases are along the deepest branches (ancestral primate to ancestral strepsirrhine (node 38-node 65; see Figure

The average proportional change in relative brain mass is 0.201 (that is, 20.1% increase), with changes above 0.278 falling in the upper quartile, which again includes branches from all major primate groups. The five branches which show the largest increase in relative brain mass are the terminal

We next examined evolutionary changes along branches controlling for branch length (change relative to time). The average rates were an increase of 0.025/million years (that is, a 2.5% increase/million years) for a proportional change in brain mass, 5,640 mg/million years for an absolute change in brain mass, and 0.020/million years for a change in relative brain mass confirming that most change in relative brain mass was due to brain rather than body mass. The branch with the highest rate of change in absolute brain mass is the terminal human branch (140,000 mg/million years). However for rate of proportional change in absolute brain mass the human branch comes only fourth, below the branches between the last common ancestor of Macaques and other

It is also notable that the estimated brain size of the last common ancestor of modern primates is smaller relative to body size than any living species and that the expansion of the primate brain began early, with the deepest branches (for example, 38 to 39; 38 to 65; 39 to 40) ranking in the upper quartile in terms of both increases in absolute and relative brain mass (Additional file

Decreases in brain mass and evolutionary scenarios for

Despite both absolute and relative brain mass showing strong and significant evolutionary trends to increase, we find several branches go against this trend (examples shown in Figure

To assess whether the proposed evolution of

Under a number of scenarios the evolution of

Evolution of brain size during the evolution of

**Ancestor**

**Ratio of change in log(absolute brain mass) & log(body mass) ^{2}**

**Change in log(brain mass)**

**Change in log(relative brain mass)**

16

0.720*

-0.398

-0.020*

24

1.058

-0.398

-0.141

32

1.586

-0.398

-0.226

Ngandong

16

0.784

-0.450

-0.058

24

1.131

-0.450

-0.178

32

1.649

-0.450

-0.264

Dmanisi

16

0.437*

-0.216*

0.122*

24

0.678*

-0.216*

0.002*

32

1.116

-0.216*

-0.084

16

0.420*

-0.137*

0.059*

24

0.908

-0.137*

-0.034

32

5.219

-0.137*

-0.147

^{1 }computed as: [(brain descendant - brain ancestor)/(body descendant-body ancestor)]

* indicates a result which falls within the range of decreases in brain size for Primates estimated in this study.

Thus under the insular dwarfism model, if

Next we performed several analyses to test whether the evolution of the

Evolution of brain size during the evolution of ^{1}.

**Ancestor**

**Ratio of change in**

**log(absolute brain mass) & log(body mass) ^{2}**

**Change in**

**log(brain mass)**

**Change in**

**log(relative brain mass)**

Argue Tree 1

16

0.400*

-0.171*

0.121*

24

0.639*

-0.171*

0.012*

32

1.116

-0.171*

-0.066

Argue Tree 2

16

0.428*

-0.173*

0.104*

24

0.709*

-0.173*

-0.006*

32

1.336

-0.173*

-0.084

Both trees

16

0.418*

-0.176*

0.110*

24

0.679*

-0.176*

-0.001*

32

1.225

-0.176*

-0.076

^{1 }These are the two most parsimonious topologies obtained by Argue

^{2 }computed as: [(brain descendant - brain ancestor)/(body descendant-body ancestor)]

* indicates a result which falls within the range of decreases in brain size for Primates estimated in this study.

To further study the selective pressures and anatomical changes associated with decreases in brain mass we suggest

Conclusions

By reconstructing ancestral states of brain and body mass in primates we have shown that Organ's et al

Our results provide robust confirmation for the suggestion that strong evolutionary trends have governed the expansion of the primate brain. In contrast body size evolution has not tended to increase in primates, implying brain and body mass have been subject to separate selection pressures and supporting the findings of previous studies in other taxonomic groups that these two highly correlated traits can show differences in their patterns of evolution

Methods

Brain and body mass data

Data for body and brain mass were obtained from previously published datasets

Through a literature search we obtained data for fossils where cranial remains were sufficiently intact to make reliable estimates of cranial capacity (N = 23, Additional file

To calculate relative brain mass we performed a phylogenetically-controlled regression analysis (see below) between log(brain mass) and log(body mass) in BayesTraits _{59 }= 14.53, R^{2 }= 0.858,

Relative brain mass on body mass for each species (extant or extinct) was calculated as residual values using the regression equation (see below). These residuals were used to test for an evolutionary trend to increase relative brain mass and to reconstruct ancestral states (

Phylogeny

It is important to incorporate both topology and branch length information during reconstruction analyses as species are part of a hierarchically structured phylogeny, therefore not statistically independent, and differences in time since divergence from the common ancestors determines differential potential for evolutionary change

Where fossil data were included we follow Finarelli & Flynn

We also use the two most parsimonious

Reconstruction methods

Ancestral state reconstructions of absolute brain and body mass and relative brain mass at each node of the phylogeny were estimated using three methods: weighted squared-change parsimony in the Mesquite

Weighted squared-change parsimony infers ancestral states by minimising the square-change along branches

ML reconstruction is based on a Brownian motion model to estimate transitions at any node along the phylogeny. The advantages of this method are that the probability of change at any point in the tree is not dependent on a prior state change or on changes on other branches

Finally, the ancestral state reconstructions of brain and body size were performed in Bayesian framework with MCMC in BayesTraits

Results of the Bayesian analysis obtained using 2 or 10 million runs were qualitatively similar, therefore we performed all analyses with two million runs. All analyses were performed using uniform priors (prior range: -100 to +100), with 2,000,000 MCMC runs after a burn-in of 500,000, sampling every 100 runs, and repeated multiple times to test the stability of the harmonic means. Rate deviation was adjusted to obtain an acceptance of the proposed model parameters (above) between 20% and 40%. Ancestral state reconstructions were then simultaneously estimated using the best evolutionary model for the data; data deviation was adjusted to obtain an acceptance rate for each node's estimate between 20 to 40%.

Next we tested if a directional-change random walk model improved the fit to the data relative to the best non-directional random walk (Brownian motion) model obtained as described above. While the non-directional random walk model has one parameter - alpha, the variance of evolution - the directional random-walk model has an additional parameter that captures the directional change using a regression between trait values and the total path length (beta)

ANCML (ML analysis) provide standard errors for each nodal value reconstruction. However, some authors consider standard errors to be underestimated and difficult to compare across methods ^{i}, Hemel Hempstead, UK). For the Bayesian analysis we thus calculated the mean of the posterior distribution of the ancestral states at each node of the tree. Because some sets of estimates made with different methods were not normally distributed, we used Spearman's rank correlation for all tests to allow correlation coefficients (_{s}) to be fully comparable throughout the analysis.

The change in brain and body mass along each branch was calculated by taking the difference between consecutive nodes. As the estimates for each node using absolute values of log brain and body mass are log values, subtracting consecutive node values gives a proportional change in mass. We therefore also converted log values into absolute numbers before calculating differences to get the absolute change in mass. Estimates of ancestral relative brain mass are based on residual values from a regression analysis of two log values. We therefore simply subtracted successive nodes to calculate change in relative brain mass. Finally, to control for differential potential in divergence due to longer time since the last splitting event we repeated the analysis and calculated the rate of change by dividing the change along a branch by the branch length, for each measure of brain mass.

Abbreviations

CI: Confidence Interval; GLS: Generalised Least Squares; MCMC: Markov-Chain Monte Carlo; ML: Maximum Likelihood.

Authors' contributions

SHM, NIM, IC and RAB designed the study, SHM performed the study, SHM, NIM, IC and RAB wrote the manuscript.

Acknowledgements

We thank Chris Venditti for advice and help with the analysis in BayesTraits and for helpful comments on the manuscript, and Brenda Bradley, Adrian Friday, Dieter Lukas, Jeremy Niven and four anonymous reviewers for helpful comments on the manuscript. SHM & NIM thank BBSRC, the Leverhulme Trust and Murray Edwards College for financial support. IC & RAB thank BBSRC/NERC for financial support (grant number BB/E014593/1).