A synergistic core for human brain evolution and cognition

A fundamental question in neuroscience is how brain organisation gives rise to humans’ unique cognitive abilities. Although complex cognition is widely assumed to rely on frontal and parietal brain regions, the underlying mechanisms remain elusive: current approaches are unable to disentangle different forms of information processing in the brain. Here, we introduce a powerful framework to identify synergistic and redundant contributions to neural information processing and cognition. Leveraging multimodal data including functional MRI, PET, cytoarchitectonics and genetics, we reveal that synergistic interactions are the fundamental drivers of complex human cognition. Whereas redundant information dominates sensorimotor areas, synergistic activity is closely associated with the brain’s prefrontal-parietal and default networks; furthermore, meta-analytic results demonstrate a close relationship between high-level cognitive tasks and synergistic information. From an evolutionary perspective, the human brain exhibits higher prevalence of synergistic information than non-human primates. At the macroscale, we demonstrate that high-synergy regions underwent the highest degree of evolutionary cortical expansion. At the microscale, human-accelerated genes promote synergistic interactions by enhancing synaptic transmission. These convergent results provide critical insights that synergistic neural interactions underlie the evolution and functioning of humans’ sophisticated cognitive abilities, and demonstrate the power of our widely applicable information decomposition framework.

a powerful framework to identify synergistic and redundant contributions to neural information 23 processing and cognition. Leveraging multimodal data including functional MRI, PET, 24 cytoarchitectonics and genetics, we reveal that synergistic interactions are the fundamental 25 drivers of complex human cognition. Whereas redundant information dominates sensorimotor 26 areas, synergistic activity is closely associated with the brain's prefrontal-parietal and default 27 networks; furthermore, meta-analytic results demonstrate a close relationship between high-28 level cognitive tasks and synergistic information. From an evolutionary perspective, the human 29 brain exhibits higher prevalence of synergistic information than non-human primates. At the 30 macroscale, we demonstrate that high-synergy regions underwent the highest degree of  Synergistic and redundant interactions identify brain networks with distinct 37 neurocognitive profiles 38 In theoretical and cognitive neuroscience, considering the human brain as a distributed 39 information-processing system has proven to be a powerful framework to understand the neural 40 basis of cognition 1 . Crucially, a deeper understanding of any information-processing 41 architecture calls for a more nuanced account of the information that is being processed. 42 As an example, let us consider humans' two main sources of information about the world: the 43 eyes. The information that we still have when we close either eye is called "redundant 44 information"because it is information that can be conveyed by either source (for instance, 45 information about colour is largely redundant between the two eyes). Redundancy provides 46 robustness: we can still see with one eye closed. However, closing one eye also deprives us of 47 stereoscopic information about depth. This information does not come from either eye alone: 48 ones needs both, in order to perceive the third dimension. This is called the "synergistic 49 information" between two sources -the extra advantage that we derive from combining them, 50 which makes them complementary 2,3 . 51 Thus, in addition to their own unique information, when multiple sources are considered 52 together their information contribution can be identified as synergistic (only available when 53 both sources are considered together) or redundant (available from either source 54 independently). Every information-processing systemincluding the human brainneeds 55 to strike a balance between these mutually exclusive kinds of information, and the advantages 56 they provide: robustness and integration, respectively 4-7 . Being fundamentally different, 57 synergistic and redundant information cannot be adequately captured by traditional measures 58 of macroscale information exchange ("functional connectivity") in the human brain, which 59 instead simply quantify the similarity between regional activity 2,8 . 60 Here, we reveal the distinct contributions of synergistic and redundant interactions to human 61 cognition, and we delineate their large-scale organisation in the human brain. To this end, we 62 leveraged the partial information decomposition (PID) framework 2,3,9 to quantify synergistic 63 and redundant interactions between brain regions ( Figure 1A,B), obtained from resting-state 64 functional MRI data from 100 Human Connectome Project subjects (Methods). We ranked 65 each brain region separately in terms of how synergistic and redundant its interactions with 66 other brain regions are; the difference between these ranks (synergy minus redundancy) 67 determines the relative relevance of a given region for synergistic versus redundant processing, 68 thereby defining a redundancy-to-synergy gradient across brain regions ( Figure 1C).  Schaefer atlas. (C) Brain surface projections of regional redundancy-to-synergy gradient scores, obtained as the 74 difference between each region's rank in terms of synergy and in terms of redundancy; positive scores (red) 75 indicate a bias towards synergy, and negative scores (blue) a bias towards redundancy. (D) Matrix of redundancy-76 to-synergy gradient scores (synergy minus redundancy ranks) for each connection between brain regions. (E)

77
Results of the NeuroSynth term-based meta-analysis, relating the distribution of redundancy-to-synergy gradient 78 across the brain (discretised in 5% increments) to a gradient of cognitive domains, from lower-level sensorimotor 79 processing to higher-level cognitive tasks. These results are robust to the use of different parcellations (cortical-80 only, having lower or higher number of nodes, and obtained from anatomical rather than functional considerations;

81
Figure S1A-C) and are also replicated without deconvolving the hemodynamic response function from the 82 functional data ( Figure S1D).

84
Our results demonstrate that traditional FC mostly captures redundant, rather than synergistic, 85 information exchange in the human brain ( Figure S2). Furthermore, they clearly show that  Figure S3), corresponding to the brain's somatomotor and salience subnetworks 90 ( Figure S4). In contrast, regions with higher relative importance for synergy predominate in 91 higher-order association cortex, and are affiliated with the default mode (DMN) and fronto-92 parietal executive control (FPN) subnetworks 11 ( Figures S3-4). 93 It is noteworthy that synergy, which quantifies the extra information gained by integrating 94 multiple sources 3,12 is most prevalent in regions belonging to the DMN and FPN. Functionally, 95 these regions are recruited by complex tasks that rely on multimodal information, decoupled 96 from immediate sensorimotor contingencies 13,14 ; anatomically, they receive multimodal inputs 97 from across the brain 15 . Therefore, it has been speculated that these networks are devoted to 98 the integration of information 13,15 . Our findings about regional prevalence of synergy in DMN 99 and FPN provide formal information-theoretic evidence to confirm this long-standing 100 hypothesis. Furthermore, by considering a synergy-redundancy gradient in terms of 101 connections instead of regions, we show that the most synergy-dominated connections 102 correspond to links between DMN/FPN and other subnetworks, whereas redundancy-103 dominated connections tend to occur within each subnetwork ( Figure 1C).

104
The distinct cytoarchitectonic profiles and subnetwork affiliations further suggest that 105 redundant and synergistic interactions may be involved with radically different cognitive 106 domains. To empirically validate this hypothesis, we performed a term-based meta-analysis 107 using NeuroSynth. The redundancy-to-synergy gradient identified in terms of regional rank 108 differences was related to 24 terms pertaining to higher cognitive functions (e.g. attention, 109 working memory, social and numerical cognition) and lower sensorimotor functions (such as 110 eye movement, motion, visual and auditory perception) adopted by previous studies 13, 16 .

111
Supporting the inference from neuroanatomy to cognition, our results reveal that the regional 112 gradient from redundancy to synergy corresponds to a gradient from lower to higher cognitive    It is also known that only a subset of regions are directly connected by white matter tracts 18 ; 135 therefore, we reasoned that the more an organism's survival depends on information exchange 136 between regions X and Y, the more one should expect X and Y to be directly connected. Thus, 137 direct physical connections in the brain reveal where the need for robust communication is   High-synergy brain regions are selectively potentiated by human evolution 163 The association between synergistic information processing and higher cognitive functions, 164 raises the intriguing possibility that the human brain may enable humans' uniquely 165 sophisticated cognitive capacities in virtue of its highly synergistic nature. We pursued this 166 hypothesis through three convergent approaches. 167 First, we show that the human brain is especially successful at leveraging synergistic  Figure 3B).

175
The patterns of synergy and redundancy in the macaque brain broadly resemble those observed 176 in humans ( Figure S8 and Supplementary Table 7), demonstrating their evolutionary stability 177 -including the expected high redundancy in sensorimotor regions ( Figure 3C). However, 178 redundancy is more prevalent than synergy in the prefrontal cortex (PFC) of macaques, despite 179 PFC being among the most synergy-dominated cortices in humans ( Figure 3C). Intriguingly, 180 prefrontal cortex underwent substantial cortical expansion in the course of human evolution 19 .

181
These findings suggest that the high synergy observed in human brains may be a specific in humans versus chimpanzees, and the gradient of regional prevalence of synergy previously 187 derived from functional MRI (⍴ = 0.42, p = 0.001; Figure 3D). Thus, these findings suggest 188 that the additional cortical tissue gained through human evolution is primarily dedicated to 189 synergy, rather than redundancy.

190
To provide further support for the evolutionary relevance of synergistic interactions, we 191 capitalised on human adult brain microarray datasets across 57 regions of the left cortical 192 mantle 20 , made available by the Allen Institute for Brain Science (AIBS) 21 . We demonstrate 193 that regional dominance of synergy correlates with regional expression of genes that are both 194 (i) related to brain development and function, including intelligence and synaptic transmission  Figure 3E). Thus, the more important a brain region is in terms of synergy, 197 the more likely it is to express brain genes that are uniquely human.

198
Taken together, these findings provide converging evidence for the hypothesis that

207
The proportion of redundant information exchange across the brain is equivalent in humans and macaques. (C)

208
Surface projection of regional redundancy-to-synergy gradient scores for the macaque brain. (D) Significant 209 correlation between human regional redundancy-to-synergy gradient scores and regional cortical expansion from 210 chimpanzee (Pan troglodytes) to human (both on left hemisphere of DK-114 cortical atlas). (E) Significant 211 correlation between human regional redundancy-to-synergy gradient scores and regional expression of brain- Neurobiological origins of synergy in the human brain 217 These observations raise the question of how such high synergy in the human brain could have 218 been attained. To address this question from a neurobiological perspective, we explored the 219 association between the redundancy-to-synergy gradient and regional expression profiles of 220 20,674 genes from AIBS microarray data 10,22 . Using partial least squares (PLS) regression, we 221 show that the first two PLS components explained 31% of the variance in the regional synergy-  Figure S14). 229 We next sought to identify the role played by overexpressed genes related to brain synergy, for Synapses are the key structures by which neurons exchange information; therefore they 250 constitute a prime candidate for the neurobiological underpinning of synergistic interactions in 251 the human brain, as suggested by our genetic analysis. To provide a more direct link between 252 synaptic density and regional prevalence of synergy, we used positron emission tomography 253 (PET) to estimate in vivo regional synaptic density based on the binding potential of the 254 synapse-specific radioligand [ 11 C]UCB-J 23 . This radioligand has high affinity for the synaptic 255 vesicle glycoprotein 2A (SV2A) 24 , which is ubiquitously expressed in all synapses throughout 256 the brain 25 . Supporting the notion that regional brain synergy is related to underlying synaptic 257 density, we found that an anterior-posterior principal component of synaptic density derived 258 from [ 11 C]UCB-J PET is significantly correlated with the regional gradient from redundancy 259 to synergy (⍴ = 0.26, p = 0.033; Figure 4D).  (Fig 1D, 2B). As the brain's input-output systems, allowing the integration of complementary information from across the brain in the service of 277 higher cognitive functions (Fig 1D): they bridge across different modules (Fig 1C), form a 278 globally efficient network (Fig 2A), and their neuroanatomical organisation coincides with 279 synapse-rich association cortex (Fig 4D and Supplementary Fig 3). 280 We further discovered that synergistic interactions were specifically enhanced in humans as a 281 result of evolutionary pressures, with dedicated cortical real estate and dedicated genes, 282 including those promoting synaptic transmission. This process resulted in a neural architecture 283 that is capable of leveraging synergistic information to a greater extent than other primates.

284
Our findings suggest that regions of the default mode and executive control (sub)networks may 285 be able to support human higher cognition precisely thanks to their extensive involvement with 286 synergistic processing.

287
Intriguingly, the high-synergy DMN is involved in self-related cognitive processes 26,27 , and it 288 is also especially disrupted by loss of consciousness, whether caused by anaesthesia or severe 289 brain injury 28 . Indeed, the global workspace theory of consciousness posits that integration of 290 information within a global workspace is necessary for consciousness 29 -and a formal link has 291 also been established between synergy and the measure of consciousness known as integrated 292 information 3,30 . Therefore, decomposition of information exchange into synergy and 293 redundancy may also shed light on the emergence of consciousness in the human brain -294 providing a framework to discover the information-processing principles that govern how 295 mental phenomena emerge from neurobiology.  Above, Un corresponds to the unique information one source but the other doesn't, Red is the 320 redundancy between both sources, and Syn is their synergy: information that neither X nor Y 321 alone can provide, but that can be obtained by considering X and Y together. It is worth noticing 322 that the unique information is fully determined after synergistic and redundant comments have 323 been accounted for; hence, we focus our analyses on the two latter components.

324
The simplest example of a purely synergistic system is one in which X and Y are independent 325 fair coins, and Z is determined by the exclusive-OR function Z = XOR(X,Y): i.e., Z=0 326 whenever X and Y have the same value, and Z=1 otherwise. It can be shown that X and Y are 327 both statistically independent of Z, which implies that neither of them provide -by themselves 328 -information about Z . However, X and Y together fully determine Z: hence, the relationship 329 between Z with X and Y is purely synergistic. region's redundancy rank from its synergy rank yielded a gradient from negative (i.e. ranking 371 higher in terms of redundancy than synergy) to positive (i.e. having a synergy rank higher than 372 the corresponding redundancy rank); note that the sign is arbitrary.

373
It is important to note that the gradient is based on relative -rather than absolute -differences 374 between regional synergy and redundancy. Consequently, a positive rank difference does not 375 necessarily mean that the region's synergy is greater than its redundancy; rather, it indicates 376 that the balance between its synergy and redundancy relative to the rest of the brain is in favour 377 of synergy -and vice versa for a negative gradient.

378
The same procedure was also repeated for network edges (instead of nodes), using their weights 379 to rank them separately in terms of synergy and redundancy and then calculating their 380 difference. This produced a single connectivity matrix where each edge's weight represents its 381 relative importance, being higher for synergy (positive edges) or redundancy (negative edges).

383
NeuroSynth term-based meta-analysis of redundancy-to-synergy gradient 384 The regional redundancy-to-synergy gradient identified in terms of nodal rank differences was

390
A meta-analysis analogous to the one implemented by previous studies 13, 16 , was conducted to 391 identify topic terms associated with the redundancy-to-synergy gradient. Twenty binary brain

415
Matrices of synergy and redundancy were thresholded proportionally using the same network 416 density as the structural connectivity matrix of the same subject. This procedure was selected 417 in order to ensure that the same number of edges would be present in both matrices, so that the 418 two matrices can be compared. Then, the upper triangular portion of each connectivity matrix 419 (structural and synergy/redundancy) was flattened into a vector, and the Spearman correlation 420 coefficient between these two vectors was computed. We use this correlation as a measure of 421 similarity between synergy or redundancy and structural connectivity.

422
HAR-BRAIN genes. 423 The maps of regional expression of human-accelerated genes for the DK-114 atlas were made identified from this procedure, 1711 were described in the Allen Human Brain Atlas (AHBA) 429 microarray dataset (human.brain-map.org) 21 and were used in the analyses by Wei and 430 colleagues, referred to as HAR genes.

431
HAR genes were subsequently subdivided into HAR-BRAIN and HAR-NonBRAIN genes.

432
BRAIN genes were selected as the set of genes commonly expressed in human brain tissue To explore the association between the redundancy-to-synergy regional gradient and all 20,647 474 genes measured in the AHBA microarrays, at each of 308 regions, we used partial least squares 475 (PLS) as a dimensionality reduction technique 10,22,44,46 . PLS finds components from the predictor 476 variables (308 × 20,647 matrix of regional gene expression scores) that have maximum covariance 477 with the response variables (308 × 1 matrix of regional redundancy-to-synergy gradient). The PLS 478 components (i.e. linear combinations of the weighted gene expression scores) are ranked by 479 covariance between predictor and response variables, so that the first few PLS components provide 480 a low-dimensional representation of the covariance between the higher dimensional data matrices. 481 Goodness of fit of low-dimensional PLS components was tested non-parametrically by repeating 482 the analysis 1000 times after shuffling the regional labels. The error on the PLS weights associated 483 with each gene were tested by resampling with replacement of 308 ROIs (bootstrapping); the ratio 484 of the weight of each gene to its bootstrap standard error was used to Z-score the genes and rank REViGO plots significant GO terms in semantic space, where semantically similar GO terms are 500 represented clustered near one another and labelled in a representative manner. 501 For our hypothesis-driven analysis, testing for enrichment of HAR-Brain genes, we also used non-502 parametric permutation testing. Specifically, we randomly drew 1000 samples of the same number 503 of genes and estimated their PLS weighting, and compared the PLS weights of the HAR-Brain 504 genes to this permutation distribution. This provided an estimate of the probability of HAR-Brain 505 gene enrichment of each PLS component under the null hypothesis 10,22 . We note that this 506 permutation procedure does not take into account the correlation between HAR-Brain genes; more 507 sophisticated null models for permutation testing that controlled for these or other characteristics 508 of candidate genes will be important to develop for computational inference in future studies. 509 510 511

512
In-vivo estimates of regional synaptic density in the human brain were obtained from positron  Principal Components Analysis (PCA) was subsequently employed to derive the principal 545 components that explain most of the variance in regional [ 11 C]UCB-J BPND across volunteers.

546
Components were selected if their associated eigenvalue was greater than unity; two principal 547 components satisfied this criterion, explaining 45% and 16% of the variance, respectively.   571 We express our gratitude to the Primate neuroimaging Data-Exchange (PRIME-DE) initiative, 572 to the organizers and managers of PRIME-DE and to all the institutions that contributed to the 573 PRIME-DE dataset (http://fcon_1000.projects.nitrc.org/indi/indiPRIME.html), with special 574 thanks to the Newcastle team. We are also grateful to Anoine Grigis, Jordy Tasserie and Bechir 575 Jarraya for their help with the Pypreclin code, and Rodrigo Romero-Garcia for generating and 576 sharing the 500mm2 subparcellation of the DK atlas, and the corresponding Von Economo 577 cytoarchitectonics map. We are also grateful to Yongbin Wei and colleagues for generating 578 and making available the data pertaining to HAR genes and cortical expansion. We are grateful 579 to UCB Pharma for providing the precursor for the radioligand used in PET imaging.   Cortical gene expression patterns were taken from the transcriptomic data of the Allen Human 614 Brain Atlas (AHBA, http://human.brain-map.org/static/download).  The Java Information Dynamics Toolbox is freely available online:

631
The code used for NeuroSynth meta-analysis is freely available online:

633
The HRF deconvolution toolbox is freely available online:

637
The code for PLS analysis of gene expression profiles is freely available online: