Message interaction dynamics covary with brain volume across mammalian connectomes
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Genomics Radio. Generated commentary on the latest computational biology and genomics papers.
Ines: We're looking at the paper: "Message interaction dynamics covary with brain volume across mammalian connectomes".
Marcus: Here's the gist of the paper. Brain network communication models typically assume that signals propagate independently, despite the high network density and small diameter of mammalian connectomes, where interactions among simultaneously propagating messages are likely Message survival accounts for over half of the variance in brain volume across mammal species and outperforms a broad range of established graph-theoretic measures and an alternative communication model with interacting messages The study investigates the copy–spread–annihilate (CSA) model, a synchronous Markovian messagepassing process in which messages spread through the binarized network while undergoing collision-dependent deletion Pruning redundant reasoning tokens The paper investigates how interactions among simultaneously propagating messages influence communication on unweighted, undirected, non-spatially embedded mammalian connectomes Message survival distributions under the CSA model are overwhelmingly best described by a lognormal distribution across 102 mammal species How the training signal is built The CSA model is defined as a discrete Markovian process which combines spreading of duplicated messages with destructive collisions of messages simultaneously arriving at nodes The first component of the system’s dynamics models message broadcasting where messages arriving singly at a node replicate themselves to all nearest neighbors The second component is deletion where biological mechanisms for deletion of simultaneously arriving messages could include classical inhibition, as well as XOR-type interactions Where the gains come from The CSA model produces results in line with past studies of static network structure, specifically that message survival distributions generate robust lognormal message survival distributions across species despite relying only on binary network topology Message survival accounts for over half the variance in mammalian brain volume, which is shown by the Pearson correlation between mean CSA message survival and mean log10 brain volume across 102 species Structural Determinants of Message Survival The study examines which structural properties are associated with message survival under CSA dynamics, finding that neutral degree assortativity graphs maximize the trade-off between message creation and destruction engendered by different assortativity regimes leading to maximal message lifetime The mean degree of hubs (nodes above the 90th percentile in each connectome, averaged within species) was strongly associated with message survival but was a substantially weaker predictor of brain volume than message survival itself Comparison with Alternative Measures CSA message survival exhibited the strongest association with brain volume among all measures examined, significantly outperforming the colliding random-walk model and various graph-theoretic measures after false-discovery-rate correction The empirical correlation of message survival with log10 brain volume for the 96 species was r = −0.731, while the randomized networks yielded a mean correlation of r = −0.689 The present work suggests that communication in mammalian brain networks cannot be understood solely from static topology considerations because interactions among propagating messages introduce emergent dynamics that shape communication over the connectome and scale systematically with brain organization across species The results indicate that message interactions deserve consideration as a fundamental component of future theories of large-scale neural communication Supplementary Materials Queueing models are discussed in relation to message interactions, suggesting that the study can take inspiration from engineered networks like the Internet which is extremely efficient at dealing with congestion The paper concludes by predicting that measures defining a message can be specified will produce distributions that resemble those of message survival under the CSA model Page 1 --- Message interaction dynamics covary with brain volume across mammalian connectomes Page 2 --- SHORT TITLE: MESSAGE INTERACTIONS ON BRAIN NETWORKS In contrast to our increasing understanding of representation in the brain, the principles governing intercommunication across the brain’s structural network remain poorly understood due in part to the fact that mapping a network and following signals across it in the same experiment remains difficult despite recent advances e.g., Dorkenwald et al., 2024; MICrONS Consortium, 2025 Existing communication models have largely assumed that propagating messages do not interact with one another this assumption is surprising because mammalian connectomes are densely interconnected with characteristic path lengths of only a few synaptic traversals Knoblauch et al., 2016; Lin et al., 2024 Under such conditions, simultaneous messages should encounter one another frequently making interactions a potentially important determinant of network communication Graham, 2014; 2021; 2023; Hao and Graham, 2020; Luppi et al., 2024; Hao et al., 26 Here we describe experiments aimed at elucidating the role of message interactions on brain networks and their potential for shaping message passing strategies and evolutionary scaling of brains across mammal species We study the copy-spread-annihilate (CSA) model Hao and Graham, 2020; Hao et al., 26, a discrete Markovian process which combines spreading of duplicated messages with destructive collisions of messages simultaneously arriving at nodes Page 3 --- SHORT TITLE: MESSAGE INTERACTIONS ON BRAIN NETWORKS The first component of the system’s dynamics models message broadcasting where messages arriving singly at a node replicate themselves to all nearest neighbors Broadcasting processes, also termed spreading or cascade processes, are emerging as a fundamental component of candidate models of whole-brain function Fakhar et al., 2025; Betzel et al., 26 and have long been a foundational idea in global workspace models Baars, 1997; Paz, 2022 The second component is deletion where biological mechanisms for deletion of simultaneously arriving messages could include classical inhibition as well as XOR-type interactions Gidon et al., 2020; 2022 and other gating mechanisms see Olshausen et al., 1993; Steriade and Paré, 207 Exponential Distance Rule In both the mouse and macaque monkey tracer-based studies have shown that the distribution of white matter axon lengths follows a decaying exponential distribution Markov et al., 2011; Ercsey-Ravasz et al., 2013 Across pairs of regions the distribution of interareal distances is lognormally distributed and positively skewed Thus, situating the exponential distribution of wires in a physical brain with lognormally-distributed interareal distances yields a heavy-tailed lognormal-like distribution of tract lengths between areas an observation known as the exponential distance rule EDR Pruning redundant reasoning tokens EDR-like relationships have been observed in imaging-based connectomes in 125 mammal species Faskowitz et al., 2023; at the single cell level within regions of mammal cortical gray matter Horvat et al., 2016; and in the Drosophila projectome Jozsa et al., 2024 How the training signal is built The EDR implies that most pairs of brain areas are separated by relatively short tracts but the distribution has a heavy tail As with EDR distance relationships nearby areas are likely to be connected by stronger links i.e., more fibers and strength also falls off with a heavy tail over which connection strength remains fairly constant Markov et al., 2011 Where the gains come from The EDR has been invoked to suggest that graph topology alone cannot explain brain organization either in terms of the brain’s communication strategies or its evolution Wang and Kennedy, 2016 Pruning redundant reasoning tokens However, EDR-like relationships have so far not been investigated in the context of dynamic message interactions This fact is surprising given that the EDR implies a high potential for signal interaction Detailed mechanistic models of brain network communication that invoke or replicate EDR have generally used static network statistics and, in the case of spatial embedding, generative models that recapitulate constraints like the EDR It is not clear that wire length should be the only or the most important constraint on message-passing and routing in brain networks Neuronal signals have tunable delays e.g., via variations in axon diameter Innocenti et al., 2016 could offset constraints imposed by wire length How the training signal is built Here we seek to understand the extent to which brain network topology by itself constrains message passing in the presence of message interactions without inclusion of spatial embeddings or edge weights Because EDR influences network topology we hypothesized that a well-chosen model of message interactions should give rise to EDR-like behavior for free because firing rates, synaptic weights, synchronous firing, numbers of synaptic contacts, and dendritic bouton sizes follow similar lognormal-like distributions leading Buzsáki and Mizuseki 2014 to conclude that skewed (typically lognormal) distributions are fundamental to structural and functional brain organization Where the gains come from We apply the CSA model to a large set of mammalian connectomes to show that the model produces results in line with past studies of static network structure Then we present novel results regarding effects of evolutionary scaling of mammal brains Pruning redundant reasoning tokens We investigate CSA model relationships with brain size and compare these to alternative models We also perform degreepreserving switch randomization to gauge the relative importance of structural regularities and degree distributions in our results How the training signal is built Under the CSA model, average message survival decreased strongly with increasing log10 brain volume accounting for over half of the variance across the 102 mammal species Where the gains come from CSA message survival exhibited the strongest association with brain volume among all measures examined and significantly outperformed alternative measures like the colliding random-walk model after false-discovery-rate correction Page 4 --- SHORT TITLE: MESSAGE INTERACTIONS ON BRAIN NETWORKS Message Survival Distributions We first asked whether CSA dynamics generate message survival distributions consistent with established large-scale regularities of mammalian brain organization particularly the heavy-tailed lognormal-like distributions associated with the exponential distance rule Throughout this experiment we express message lifetime as message survival measured in edge traversals It differs in two respects from message lifetime as defined by Hao et al., 26 first it implies a network distance traveled rather than a time period Second message survival is not weighted by the number of messages that survive the maximum number of timesteps from injection Pruning redundant reasoning tokens In all other respects, it is identical to message lifetime and all results are very similar when survival is weighted by the number of messages surviving the maximal time from injection However when considered as distances it should be noted that messages can retrace their steps and/or travel in loops meaning that they are not necessarily traveling on geodesic paths between injection and final deletion How the training signal is built As such message survival gauges the maximum impact of a message on the network Considering time as distance allows comparison to canonical results in brain network design and scaling such as the EDR Where the gains come from Under the CSA model, average message survival decreased strongly with increasing log10 brain volume accounting for over half of the variance across the 102 mammal species Pearson r = −0.7647, p = 8.6 × 10−21 Comparison with Alternative Measures CSA message survival exhibited the strongest association with brain volume among all measures examined and significantly outperformed other measures after false-discovery-rate correction Pruning redundant reasoning tokens Degree-preserving randomization weakens this relationship suggesting that the CSA model’s emergent dynamics in concert with the higher-order network organization contributes to the observed effects Structural Determinants of Message Survival We next examined which structural properties are associated with message survival under CSA dynamics finding that neutral degree assortativity graphs maximize the trade-off between message creation and destruction engendered by different assortativity regimes leading to maximal message lifetime Degree assortativity shows a similarly modest correlation with brain volume r = 0.4012, p = 2.9 × 10−5 Where the gains come from A structural feature that strongly predicts the CSA dynamics nevertheless predicts brain volume substantially less well than the dynamics themselves, as CSA message lifetime explained 58.5% of the variance in brain volume compared with 26.9% for mean hub degree Discussion Principal Findings The three principal findings emerged First CSA dynamics generate robust lognormal message survival distributions across species despite relying only on binary network topology Second message survival accounts for over half the variance in mammalian brain volume while structural features are not as strongly correlated with brain volume compared to the dynamical measure itself Third degree-preserving randomization weakens this relationship suggesting that the CSA model’s emergent dynamics in concert with the higher-order network organization contributes to the observed effects Pruning redundant reasoning tokens CSA-like strategies may be useful not only in mesoscale white matter networks but also brains in orders outside Mammalia as well as micro-scale neuronal networks which show similar topology The CSA model represents a stochastic communication framework in which routing emerges from local interactions rather than centralized control or pre-specified paths Biological Plausibility Although the CSA model is not a biophysical model discrete messaging in the brain is biologically plausible given single-unit electrophysiological evidence for message packets being passed from neuron to neuron over the course of tens or hundreds of milliseconds We suggest that a core network communication task that brains perform is the creation and crucially the deletion of polysynaptic messages on the network of white matter tracts which is almost entirely excitatory Coincidence Detection We hypothesize that message deletion serves functionally as the sensing of coincidences i.e., detecting events that impinge on that area together The results suggest a testable prediction if a quantity defining a message can be specified one would look across neuronal populations for all copies of a given message and ask how many hops those messages take Page 5 --- SHORT TITLE: MESSAGE INTERACTIONS ON BRAIN NETWORKS Copy-Spread-Annihilate (CSA) Model Network dynamics are modeled as a Markovian agent-based process of collisions and spreading activation on empirical connectomes where nodes serve as synchronous-time message passing agents Under this model a message that meets one or more messages at a node on a given timestep results in all colliding messages being deleted from the system If a message arrives singly at a node it survives to the next time step at which point it passes copies of itself to all the node’s nearest neighbors Each time step new messages are injected into the system at a fixed fraction of randomly-chosen unoccupied nodes The quantity of message injections is termed the injection rate See Figure 1 Pruning redundant reasoning tokens Message injections represent intrinsic activity as well as external inputs In past work we have shown that the CSA model maintains sparse activity on mouse and monkey tracer-based connectomes as injection rate increases Hao and Graham, 20201 Collisions occur only at nodes rather than along edges meaning that potential ephaptic interactions are not modeled see e.g., Sheheitli and Jirsa, 2020 For each injected message we track all descendant copies throughout the simulation Message lifetime is defined as the survival time of the longest-lived descendant following Hao et al., 26 i.e., the number of timesteps from injection through deletion (inclusive) yielding positive integers with a minimum value of one Page 6 --- SHORT TITLE: MESSAGE INTERACTIONS ON BRAIN NETWORKS Figure 1.
Ines: Copy-spread-annihilate model on the mammalian connectome.
Marcus: On a given timestep messages in the system spread to all nearest network neighbors If a message meets one or more other messages at a node all are deleted Messages that do not encounter another message survive to the next timestep On the next timestep a fraction of open nodes given by the injection rate are randomly selected to receive a new message Then all messages spread and collide as in timestep 1 Figure 2.
Ines: A.
Paper summary: Ines: So just to recap, the paper "Message interaction dynamics covary with brain volume across mammalian connectomes" investigates whether messages interacting dynamically can explain the observed differences in brain size across mammals.
Marcus: The core thesis is that communication models usually assume signals are independent, but this paper uses a copy-spread-annihilate model to show that message collisions and deletions shape survival in a way that correlates strongly with brain volume.
Yuki: They tested this on a large set of one hundred two mammalian connectomes and found the copy-spread-annihilate dynamics produce robust, positively skewed lognormal distributions of message survival across species.
Ines: What’s important is that this dynamical approach shows that message survival accounts for over half the variance in mammalian brain volume, which is a major finding when you compare it to structural measures alone.
Marcus: They also examined structural properties like degree assortativity graphs and found those neutral degree assortativity graphs maximize the trade-off between message creation and destruction, leading to maximal message lifetime in their simulation.
Yuki: The paper suggests that these emergent dynamics are key because they link the physical constraints of connectivity to the evolutionary scaling of brains across mammals.
Ines: It’s about showing that communication strategies aren't just about the fixed wiring pattern; it’s about how messages move and interact over time on that network.
Marcus: And they found that even when you look at structural features like mean hub degree, they don't predict brain volume as well as the message survival itself does.
Yuki: So, if you only look at the static structure of a brain, it doesn't fully explain why some brains are much larger than others; the dynamic interaction is what matters.
Ines: It shifts the focus from just mapping connections to understanding how information actually travels and persists within those connections.
Marcus: And this points toward a model where routing emerges from local interactions rather than relying on some kind of centralized control or pre-specified paths.
Conclusion: Ines: So, looking at the title "Message interaction dynamics covary with brain volume across mammalian connectomes," what we’re really seeing is that the way messages move and interact within a brain network has a measurable relationship with how big that brain actually is.
Marcus: The authors argue that communication in mammalian brains can’t be understood just by looking at the static structure of the connections; instead, interactions among propagating messages introduce emergent dynamics that shape communication and scale systematically with brain organization across species.
Yuki: From a population geneticist's view, this means that evolutionary pressures must favor network structures where these message interaction effects are optimized to allow for the scaling of brains we see in evolution.
Ines: It suggests a core network communication task is the creation and crucially the deletion of polysynaptic messages on the network of white matter tracts, which is almost entirely excitatory.
Marcus: They suggest that message interactions deserve serious consideration as a fundamental component in any future theories about large-scale neural communication, because it’s what links structure to function across different brain sizes.
Yuki: So, if we can figure out how a quantity defining a message is specified, we could predict the survival distributions, which would give us new tools to test these dynamic ideas against real biological data.
Yan Hao, Hannah Lax, Marc-Thorsten Hütt, Daniel J. Graham
Department of Mathematics, Hobart and William Smith Colleges, Geneva, New York, USA · U.S. Department of Veterans Affairs, Boston, Massachusetts, USA · Department of Life Sciences and Chemistry, Constructor University, Bremen Germany · Department of Psychological Science, Hobart and William Smith Colleges, Geneva New York
q-bio.NC
Submitted: 2025-05-21
Updated: 2026-10-30
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 100/100
The gist: Brain network communication models typically assume that signals propagate independently, despite the high network density and small diameter of mammalian connectomes…
Key concepts
- Copy–Spread–Annihilate (CSA) Model
- This is a mathematical process simulating neural communication where messages spread to neighbors (copy), but if two or more messages arrive at the same node simultaneously, they are deleted (annihilated). This models both message broadcasting and destructive interactions within the network.
- Message Survival
- This measures how long a single message lasts as it travels across the network before being deleted. The study found that this survival metric is a powerful predictor of brain volume, correlating strongly with physical size across different mammals.
- Degree Assortativity
- This describes how nodes (brain regions) connect to other nodes based on their own connectivity levels. The research found that 'neutral' assortativity—where connections are not strongly biased toward hubs or low-degree nodes—maximizes the message lifetime in the CSA model.
Terminology
Summary
Brain network communication models typically assume that signals propagate independently, despite the high network density and small diameter of mammalian connectomes, where interactions among simultaneously propagating messages are likely
Message survival accounts for over half of the variance in brain volume across mammal species and outperforms a broad range of established graph-theoretic measures and an alternative communication model with interacting messages
The study investigates the copy–spread–annihilate (CSA) model, a synchronous Markovian messagepassing process in which messages spread through the binarized network while undergoing collision-dependent deletion
Pruning redundant reasoning tokens
The paper investigates how interactions among simultaneously propagating messages influence communication on unweighted, undirected, non-spatially embedded mammalian connectomes
Message survival distributions under the CSA model are overwhelmingly best described by a lognormal distribution across 102 mammal species
How the training signal is built
The CSA model is defined as a discrete Markovian process which combines spreading of duplicated messages with destructive collisions of messages simultaneously arriving at nodes
The first component of the system’s dynamics models message broadcasting where messages arriving singly at a node replicate themselves to all nearest neighbors
The second component is deletion where biological mechanisms for deletion of simultaneously arriving messages could include classical inhibition, as well as XOR-type interactions
Where the gains come from
The CSA model produces results in line with past studies of static network structure, specifically that message survival distributions generate robust lognormal message survival distributions across species despite relying only on binary network topology
Message survival accounts for over half the variance in mammalian brain volume, which is shown by the Pearson correlation between mean CSA message survival and mean log10 brain volume across 102 species
Structural Determinants of Message Survival
The study examines which structural properties are associated with message survival under CSA dynamics, finding that neutral degree assortativity graphs maximize the trade-off between message creation and destruction engendered by different assortativity regimes leading to maximal message lifetime
The mean degree of hubs (nodes above the 90th percentile in each connectome, averaged within species) was strongly associated with message survival but was a substantially weaker predictor of brain volume than message survival itself
Comparison with Alternative Measures
CSA message survival exhibited the strongest association with brain volume among all measures examined, significantly outperforming the colliding random-walk model and various graph-theoretic measures after false-discovery-rate correction
The empirical correlation of message survival with log10 brain volume for the 96 species was r = −0.731, while the randomized networks yielded a mean correlation of r = −0. p. 9 SHORT TITLE: MESSAGE INTERACTIONS ON BRAIN NETWORKS
10
Figure 4. Relationship between mean CSA message survival and mean log₁₀ brain volume
across 102 species (total of 185 mammalian connectomes analyzed; error bars indicate SEM).Where the paper says it
Conclusion
The present work suggests that communication in mammalian brain networks cannot be understood solely from static topology considerations because interactions among propagating messages introduce emergent dynamics that shape communication over the connectome and scale systematically with brain organization across species
The results indicate that message interactions deserve consideration as a fundamental component of future theories of large-scale neural communication
Supplementary Materials
Queueing models are discussed in relation to message interactions, suggesting that the study can take inspiration from engineered networks like the Internet which is extremely efficient at dealing with congestion
The paper concludes by predicting that measures defining a message can be specified will produce distributions that resemble those of message survival under the CSA model
--- Page 1 ---
Message interaction dynamics covary with brain volume across mammalian connectomes
--- Page 2 ---
SHORT TITLE: MESSAGE INTERACTIONS ON BRAIN NETWORKS
INTRODUCTION
In contrast to our increasing understanding of representation in the brain, the principles governing intercommunication across the brain’s structural network remain poorly understood due in part to the fact that mapping a network and following signals across it in the same experiment remains difficult despite recent advances e.g., Dorkenwald et al. p. 1 This
theoretical lacuna is due in part to the fact that mapping a network and following signals across it
in the same experiment remains difficult, despite recent advances (e. , Dorkenwald et al.Where the paper says it
Existing communication models have largely assumed that propagating messages do not interact with one another this assumption is surprising because mammalian connectomes are densely interconnected with characteristic path lengths of only a few synaptic traversals Knoblauch et al., 2016; Lin et al. p. 1 This assumption is surprising because mammalian connectomes are densely
interconnected, with characteristic path lengths of only a few synaptic traversals (Knoblauch et al. ,
2016; Lin et al.Where the paper says it
Under such conditions, simultaneous messages should encounter one another frequently making interactions a potentially important determinant of network communication Graham, 2014; 2021; 2023; Hao and Graham, 2020; Luppi et al., 2024; Hao et al. p. 1 Under such conditions, simultaneous messages should encounter one
another frequently, making interactions a potentially important determinant of network
communication (Graham, 2014; 2021; 2023; Hao and Graham, 2020; Luppi et al. , 2024; Hao et
al.Where the paper says it
Here we describe experiments aimed at elucidating the role of message interactions on brain networks and their potential for shaping message passing strategies and evolutionary scaling of brains across mammal species
We study the copy-spread-annihilate (CSA) model Hao and Graham, 2020; Hao et al., 26, a discrete Markovian process which combines spreading of duplicated messages with destructive collisions of messages simultaneously arriving at nodes
--- Page 3 ---
SHORT TITLE: MESSAGE INTERACTIONS ON BRAIN NETWORKS
The first component of the system’s dynamics models message broadcasting where messages arriving singly at a node replicate themselves to all nearest neighbors
Broadcasting processes, also termed spreading or cascade processes, are emerging as a fundamental component of candidate models of whole-brain function Fakhar et al., 2025; Betzel et al. p. 2 Broadcasting processes, also termed
spreading or cascade processes, are emerging as a fundamental component of candidate models of
whole-brain function (Fakhar et al. , 2025, Betzel et al.Where the paper says it
The second component is deletion where biological mechanisms for deletion of simultaneously arriving messages could include classical inhibition as well as XOR-type interactions Gidon et al., 2020; 2022 and other gating mechanisms see Olshausen et al. p. 2 Biological mechanisms for deletion of simultaneously arriving
messages could include classical inhibition, as well as XOR-type interactions (Gidon et al. , 2020;
2022) and other gating mechanisms (see Olshausen et al.Where the paper says it
Exponential Distance Rule
In both the mouse and macaque monkey tracer-based studies have shown that the distribution of white matter axon lengths follows a decaying exponential distribution Markov et al., 2011; Ercsey-Ravasz et al. p. 2 Exponential Distance Rule
In both the mouse and macaque monkey, tracer-based studies have shown that the distribution of
white matter axon lengths follows a decaying exponential distribution (Markov et al. , 2011;
Ercsey-Ravasz et al.Where the paper says it
Across pairs of regions the distribution of interareal distances is lognormally distributed and positively skewed Thus, situating the exponential distribution of wires in a physical brain with lognormally-distributed interareal distances yields a heavy-tailed lognormal-like distribution of tract lengths between areas an observation known as the exponential distance rule EDR
EDR-like relationships have been observed in imaging-based connectomes in 125 mammal species Faskowitz et al., 2023; at the single cell level within regions of mammal cortical gray matter Horvat et al. p. 2 EDR-like relationships have been
observed in imaging-based connectomes in 125 mammal species (Faskowitz et al. , 2023); at the
single cell level within regions of mammal cortical gray matter (Horvat et al. , 2023); at the
single cell level within regions of mammal cortical gray matter (Horvat et al. , 2016); and in the
Drosophila projectome (Jozsa et al.Where the paper says it
Where the paper says it
The EDR implies that most pairs of brain areas are separated by relatively short tracts but the distribution has a heavy tail As with EDR distance relationships nearby areas are likely to be connected by stronger links i.e., more fibers and strength also falls off with a heavy tail over which connection strength remains fairly constant Markov et al. p. 2 As with EDR distance relationships, nearby areas are likely to be
connected by stronger links (i. , more fibers), and strength also falls off with a heavy tail, over
which connection strength remains fairly constant (Markov et al.Where the paper says it
The EDR has been invoked to suggest that graph topology alone cannot explain brain organization either in terms of the brain’s communication strategies or its evolution Wang and Kennedy, 2016
However, EDR-like relationships have so far not been investigated in the context of dynamic message interactions This fact is surprising given that the EDR implies a high potential for signal interaction Detailed mechanistic models of brain network communication that invoke or replicate EDR have generally used static network statistics and, in the case of spatial embedding, generative models that recapitulate constraints like the EDR
It is not clear that wire length should be the only or the most important constraint on message-passing and routing in brain networks Neuronal signals have tunable delays e.g., via variations in axon diameter Innocenti et al. p. 3 Neuronal signals have tunable delays (e. , via variations in axon diameter; Innocenti et
al.Where the paper says it
Here we seek to understand the extent to which brain network topology by itself constrains message passing in the presence of message interactions without inclusion of spatial embeddings or edge weights
Because EDR influences network topology we hypothesized that a well-chosen model of message interactions should give rise to EDR-like behavior for free because firing rates, synaptic weights, synchronous firing, numbers of synaptic contacts, and dendritic bouton sizes follow similar lognormal-like distributions leading Buzsáki and Mizuseki 2014 to conclude that skewed (typically lognormal) distributions are fundamental to structural and functional brain organization
We apply the CSA model to a large set of mammalian connectomes to show that the model produces results in line with past studies of static network structure Then we present novel results regarding effects of evolutionary scaling of mammal brains
We investigate CSA model relationships with brain size and compare these to alternative models We also perform degreepreserving switch randomization to gauge the relative importance of structural regularities and degree distributions in our results
Under the CSA model, average message survival decreased strongly with increasing log10 brain volume accounting for over half of the variance across the 102 mammal species
CSA message survival exhibited the strongest association with brain volume among all measures examined and significantly outperformed alternative measures like the colliding random-walk model after false-discovery-rate correction
--- Page 4 ---
SHORT TITLE: MESSAGE INTERACTIONS ON BRAIN NETWORKS
Message Survival Distributions
We first asked whether CSA dynamics generate message survival distributions consistent with established large-scale regularities of mammalian brain organization particularly the heavy-tailed lognormal-like distributions associated with the exponential distance rule
Throughout this experiment we express message lifetime as message survival measured in edge traversals It differs in two respects from message lifetime as defined by Hao et al., 26 first it implies a network distance traveled rather than a time period Second message survival is not weighted by the number of messages that survive the maximum number of timesteps from injection
In all other respects, it is identical to message lifetime and all results are very similar when survival is weighted by the number of messages surviving the maximal time from injection However when considered as distances it should be noted that messages can retrace their steps and/or travel in loops meaning that they are not necessarily traveling on geodesic paths between injection and final deletion
As such message survival gauges the maximum impact of a message on the network Considering time as distance allows comparison to canonical results in brain network design and scaling such as the EDR
Under the CSA model, average message survival decreased strongly with increasing log10 brain volume accounting for over half of the variance across the 102 mammal species Pearson r = −0.7647, p = 8. p. 9 Under the CSA model, average message survival decreased strongly with increasing log10 brain
volume (Pearson r = −0. 7647, p = 8.Where the paper says it
CSA message survival exhibited the strongest association with brain volume among all measures examined and significantly outperformed other measures after false-discovery-rate correction
Degree-preserving randomization weakens this relationship suggesting that the CSA model’s emergent dynamics in concert with the higher-order network organization contributes to the observed effects
We next examined which structural properties are associated with message survival under CSA dynamics finding that neutral degree assortativity graphs maximize the trade-off between message creation and destruction engendered by different assortativity regimes leading to maximal message lifetime
Degree assortativity shows a similarly modest correlation with brain volume r = 0.4012, p = 2.9 × 10−5
A structural feature that strongly predicts the CSA dynamics nevertheless predicts brain volume substantially less well than the dynamics themselves, as CSA message lifetime explained 58.5% of the variance in brain volume compared with 26. p. 0 Simulations on a large comparative dataset of mammalian
connectomes show that CSA dynamics produce robust positively-skewed lognormal distributions
of message survival across species. Despite using only binary network topology without spatial
embedding, message survival accounts for over half of the variance in brain volume across
mammal species and outperforms a broad range of established graph-theoretic measures and an
alternative communication model with interacting messages.Where the paper says it
Discussion
Principal Findings The three principal findings emerged First CSA dynamics generate robust lognormal message survival distributions across species despite relying only on binary network topology
Second message survival accounts for over half the variance in mammalian brain volume while structural features are not as strongly correlated with brain volume compared to the dynamical measure itself
Third degree-preserving randomization weakens this relationship suggesting that the CSA model’s emergent dynamics in concert with the higher-order network organization contributes to the observed effects
CSA-like strategies may be useful not only in mesoscale white matter networks but also brains in orders outside Mammalia as well as micro-scale neuronal networks which show similar topology
The CSA model represents a stochastic communication framework in which routing emerges from local interactions rather than centralized control or pre-specified paths
Biological Plausibility
Although the CSA model is not a biophysical model discrete messaging in the brain is biologically plausible given single-unit electrophysiological evidence for message packets being passed from neuron to neuron over the course of tens or hundreds of milliseconds
We suggest that a core network communication task that brains perform is the creation and crucially the deletion of polysynaptic messages on the network of white matter tracts which is almost entirely excitatory
Coincidence Detection We hypothesize that message deletion serves functionally as the sensing of coincidences i.e., detecting events that impinge on that area together
The results suggest a testable prediction if a quantity defining a message can be specified one would look across neuronal populations for all copies of a given message and ask how many hops those messages take
--- Page 5 ---
SHORT TITLE: MESSAGE INTERACTIONS ON BRAIN NETWORKS
Copy-Spread-Annihilate (CSA) Model
Network dynamics are modeled as a Markovian agent-based process of collisions and spreading activation on empirical connectomes where nodes serve as synchronous-time message passing agents
Under this model a message that meets one or more messages at a node on a given timestep results in all colliding messages being deleted from the system If a message arrives singly at a node it survives to the next time step at which point it passes copies of itself to all the node’s nearest neighbors
Each time step new messages are injected into the system at a fixed fraction of randomly-chosen unoccupied nodes The quantity of message injections is termed the injection rate See Figure 1
Message injections represent intrinsic activity as well as external inputs In past work we have shown that the CSA model maintains sparse activity on mouse and monkey tracer-based connectomes as injection rate increases Hao and Graham, 20201
Collisions occur only at nodes rather than along edges meaning that…
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Improved message propagation modeling for complex systems: The CSA model, which incorporates
collision-dependent deletion,
can be used to simulate emergent communication dynamics in large-scale neural networks without needing explicit spatial embedding or edge weights. This allows AI systems to predict how message survival distributions across species scale with brain volume and how higher-order structural organization contributes to these dynamics. -
Enhanced network efficiency prediction: By comparing the CSA model results with alternative measures like
Clustering
andPath Transitivity,
AI systems can better assess which topological features, such asClustering 0.6747
orPath Transitivity 0.6682,
are most predictive of message survival compared to static graph-theoretic measures. -
Evolutionary scaling prediction: AI models can predict the scaling relationship between communication dynamics and brain size by utilizing the finding that
message survival accounts for over half of the variance in brain volume across mammal species.
This allows for better inference about evolutionary constraints on neural communication strategies. -
Structural determinant identification: AI systems can determine if observed scaling relationships are explained by simple properties of the degree distribution, as
degree sequence alone is insufficient to explain the observed relationship between message survival and brain volume.
This directs researchers toward studyinghigher-order features of empirical connectome organization
for better predictive power. -
Routing strategy optimization: AI can guide the design of communication protocols by understanding that
network topology is also sufficient to bias the dynamics toward increasingly local message passing in larger brains under the CSA model, but not under the colliding-RW model.
This informs whether a system should prioritize spreading activation or collision management based on desired localization.
Abstract
Brain network communication models typically assume that signals propagate independently, despite the high network density and small diameter of mammalian connectomes, where interactions among simultaneously propagating messages are likely. We investigate these interactions using the copy-spread-annihilate (CSA) model, a synchronous Markovian message-passing process in which messages spread through the binarized network while undergoing collision-dependent deletion. Simulations on a large comparative dataset of mammalian connectomes show that CSA dynamics produce robust positively-skewed lognormal distributions of message survival across species. Despite using only binary network topology without spatial embedding, message survival accounts for over half of the variance in brain volume across mammal species and outperforms a broad range of established graph-theoretic measures and an alternative communication model with interacting messages. These relationships persisted when measured paths could not revisit nodes. Degree-preserving switch randomization weakens the relationship between message survival and brain volume, indicating that higher-order structural organization contributes to the observed scaling law. We describe a mechanistic explanation of model dynamics that suggests message creation differences among networks shapes message survival. Together, these findings suggest that interactions among propagating messages constitute an important and underexplored determinant of communication dynamics in mammalian brain networks.
Sources
- Copy-Spread-Annihilate Dynamics in Degree-Assortative Networks
- A Spatiotemporal Perspective on Dynamical Computation in Neural Information Processing Systems
Related papers
- Self-Supervised Learning of Local Predictive Directions as a Neural Computation
- Synchronized Hearts, Desynchronized Brains: A Nonlinear Dynamics Perspective for Function and Failure
- Planning as Dynamics Relaxation: Hippocampal Recurrent Network Realizes Optimal Goal-Directed Navigation
- The Platonic brain bridge hypothesis: brain networks as architectural priors for multimodal large language models
- BrainPFN: Amortized Brain-Behaviour Prediction via Prior-Data Fitted Networks
- CTM-AI: Evolving Multi-Agent Intelligence Inspired by a Model of Consciousness