Building functional and mechanistic models of cortical computation based on canonical cell type connectivity
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: "Building functional and mechanistic models of cortical computation based on canonical cell type connectivity".
Marcus: Here's the gist of the paper. Neuronal circuits of the cerebral cortex are the structural basis of mammalian cognition Here, we propose a perspective on current knowledge of cortical structure, from which we extract two core principles for computational modeling Core Principles for Modeling The paper establishes two foundational principles to guide the construction of functional and mechanistic models of cortical circuits The first principle is that cortical cell types fulfill distinct computational roles The second principle is that cortical connectivity can be efficiently characterized by only a few canonical blueprints of connectivity between cell types Starting with these two foundational principles, we outline a general framework for building functional and mechanistic models of cortical circuits centered on the notion of canonical connectivity among cortical cell types.
Ines: Functional Specialization of Cell Types Cortical cell types are functionally specialized Excitatory (glutamatergic) neuron types, typically pyramidal in shape, have primarily been characterized by their laminar position, morphology, and the brain structures targeted by their axons.
Marcus: Recent efforts to develop precise and complete taxonomies focus on the unbiased definition of cell types as clusters of cells with similar gene expression profiles.
Ines: The functional specialization of broad interneuron classes defined by key molecular markers (Pvalb, Sst, VIP) is well established.
Title and authors: Ines: The paper, "Building functional and mechanistic models of cortical computation based on canonical cell type connectivity," is really about using a few core ideas to build functional and mechanistic models for cortical circuits. It suggests that we should focus on the specific patterns of connections between different types of brain cells.
Marcus: Exactly. They are proposing two main principles here that guide how you construct these models: first, that cortical cell types fulfill distinct computational roles, and second, that connectivity can be efficiently characterized by only a few canonical blueprints between those cell types.
Yuki: So the goal is to find those few blueprints—those basic patterns of connection—that capture the essence of what those different cells are actually doing in the cortex. That’s a big step toward understanding how complex functions emerge from simpler components.
The paper's summary: Ines: They summarize their approach by first establishing that excitatory neurons, like pyramidal cells, have been studied mainly by where they are located in the layers and what structures their axons target. But they point out that recent work is moving toward defining cell types more robustly as clusters of cells with similar gene expression profiles.
Marcus: And they highlight that even within those broad classes, like inhibitory interneurons defined by markers such as Pvalb or Sst, there are specific cell types carrying unique signals. For instance, one specific type in the Sst class in the mouse posterior parietal cortex carries a strong error-correction signal.
Yuki: That suggests that functional specialization isn't just about location anymore; it’s about what those cells are doing computationally, which ties into how different populations handle things like decision-making or motor control.
The paper's improvements: Ines: The authors suggest several ways to improve the modeling process. They start by defining single neuron models that have a set of variables and a parameter space, plus a function that describes the neuronal dynamics, which helps constrain simulations by biological principles rather than just fitting data.
Marcus: And then they move into how to define connectivity. They say you first define cell type connectivity blueprints as matrices where rows are presynaptic cell types and columns are postsynaptic ones, and the value is the synapse type.
Yuki: So you have these blueprints for local connections, and then for interareal connections too—which they suggest should be characterized by just two classes: feedforward or feedback. That simplifies the way we think about how information flows across different areas.
Ines: And then they propose using blackbox optimization methods or evolutionary strategies to fill in those connectivity matrices between cell types until the resulting architecture becomes functional. It's like a form of neural architecture search where you are searching for an architecture that works based on fitness selection.
Conclusion: Marcus: So, to wrap up, the paper "Building functional and mechanistic models of cortical computation based on canonical cell type connectivity" suggests focusing on the two principles we talked about: distinct computational roles for cell types and a limited set of canonical blueprints for their connections.
Ines: The authors show how you can define these blueprints—matrices describing presynaptic to postsynaptic cell type synapse types—and then use optimization methods to find architectures that are functional. This is about moving from just observing structure to building models that actually work computationally.
Yuki: For us, this means we can start mapping different cortical cell types directly onto specific computational roles, which gives us a much stronger basis for understanding how complex behaviors arise in the brain across species.
Marcus: And they point toward future work involving indirect encoding of connectivity, like using connection probabilities or distance rules between cell types to guide the search process. That lets you explore more complex ways of defining those blueprints.
Ines: It's a framework that extends beyond just the cortex, so it could potentially be used for modeling other brain structures too, both within and between them. We need to keep looking at how these specific connectivity motifs repeat across different areas and species.
Arno Graniera, Katharina A Wilmesa, Mihai A Petrovicia, Walter Senna
Department of Physiology, University of Bern, Switzerland · Graduate School for Cellular and Biomedical Sciences, University of Bern, Switzerland · Institute of Neuroinformatics, University of Zurich and ETH Zurich Zürich · Center for Artificial Intelligence in Medicine, University of Bern
q-bio.NC, cs.NE
Submitted: 2025-04-03
Updated: 2026-10-30
Code: https://github.com/arnogranier/neuroevo-bp
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 85/100
The gist: Neuronal circuits of the cerebral cortex are the structural basis of mammalian cognition Here, we propose a perspective on current knowledge of cortical structure…
Key concepts
- Functional Specialization of Cell Types
- Cortical neurons perform different computational tasks based on their type. While morphology is important, recent work suggests that gene expression profiles define these specialized roles. For example, specific interneurons carry error-correction signals, indicating that cell identity dictates what a neuron is functionally capable of doing in the circuit.
- Structure of Local Connectivity
- Local connections within the cortex are not random; they suggest parallel and recurrent pathways for information processing. The key takeaway is that specific cell types connect with each other in predictable ways, rather than just being physically close. This emphasizes that cell type identity is a more important factor than mere spatial proximity when determining local circuit behavior.
- Canonical Blueprints of Connectivity
- Instead of modeling every possible connection, the paper suggests that cortical connectivity can be simplified into a small set of standard connection patterns between cell types. This allows researchers to focus on finding these essential blueprints rather than searching through an impossibly large space of all possible connections.
- Neural Architecture Search
- This involves using optimization techniques, like evolutionary strategies, to automatically search for the best circuit structure. The search space is limited by the defined rules for cell-type connectivity. The goal is to find a specific arrangement of connections (a functional architecture) that achieves a desired computational outcome.
Terminology
Summary
Neuronal circuits of the cerebral cortex are the structural basis of mammalian cognition Here, we propose a perspective on current knowledge of cortical structure, from which we extract two core principles for computational modeling
Core Principles for Modeling
The paper establishes two foundational principles to guide the construction of functional and mechanistic models of cortical circuits The first principle is that cortical cell types fulfill distinct computational roles The second principle is that cortical connectivity can be efficiently characterized by only a few canonical blueprints of connectivity between cell types Starting with these two foundational principles, we outline a general framework for building functional and mechanistic models of cortical circuits centered on the notion of canonical connectivity among cortical cell types.
Functional Specialization of Cell Types
Cortical cell types are functionally specialized Excitatory (glutamatergic) neuron types, typically pyramidal in shape, have primarily been characterized by their laminar position, morphology, and the brain structures targeted by their axons. Recent efforts to develop precise and complete taxonomies focus on the unbiased definition of cell types as clusters of cells with similar gene expression profiles. p. 1 Excitatory (glutamatergic) neuron types, typically pyramidal
in shape, have primarily been characterized by their laminar position, morphology, and the brain structures
targeted by their axons, while inhibitory (GABAergic) interneuron types have been characterized by key molecular
markers, morphology, and patterns of local connectivity. In addition to these classical accounts, recent efforts to
develop precise and complete taxonomies focus on the unbiased definition of cell types as clusters of cells with
similar gene expression profiles (Brain Initiative Cell Census Network 2021; Chen et al. 2023;
Di Bella et al. The functional specialization of broad interneuron classes defined by key molecular markers (Pvalb, Sst, VIP)
is well established (Tremblay, Lee, and Rudy 2016). 2022; Green et al. Notably, Green
et al. Notably, Green
et al. (2023) report that one specific cell type within the Sst class in the mouse posterior parietal cortex carries
a strong error-correction signal. Different excitatory cell types also show clear differences in functional response patterns, likely underlying specialized computational roles. For example, pyramidal cells of cortical layer 5 that project subcortically (pyramidal
tract, PT) and those that do not (intratelencephalic, IT) exhibit distinct activity profiles during decision-making
(Musall et al.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Structure of Local Connectivity
The structure of local cortical connectivity suggests parallel and recurrent, yet specific information processing pathways Recent contributions include reconstructions of a complete ‘column’ of mouse primary somatosensory cortex and a cubic millimeter of mouse visual cortex spanning multiple cortical areas. The spatial proximity of axons and dendrites is not sufficient to predict connectivity; rather, cell-type-specific connectivity is the norm. p. 3 (2023) further show
that proximal projections from subcortically-projecting pyramidal cells of layer 5 preferentially target specific
interneuron types that in turn preferentially target back the same type of pyramidal cells, see fig. Both for excitatory and inhibitory projections, it is clear that the spatial proximity of axons and dendrites
is not sufficient to predict connectivity; rather, cell-type-specific connectivity is the norm, supported by the
differential expression of synaptic molecules depending on pre- and post-synaptic cell types (Favuzzi et al. 2019;
Bernard et al. This highlights again the cell type as a central unit for deciphering cortical connectivity
and computation.Where the paper says it
Where the paper says it
Structure of Interareal Connectivity
Interareal projections originating from specific cell types might form largely segregated streams of computation across the cortex Connections between cortical areas can be characterized on the basis of which specific cell types send and receive projections in the source and target areas respectively. The notion of a cortical hierarchy can be constructed by assigning each interareal connection to one of only two classes (or a continuum between them), traditionally deemed feedforward and feedback. p. 3 Similar to local connectivity (see section 3), interareal connectivity can be efficiently characterized
by only a few blueprints of connectivity between cortical cell types. For example, the notion of a cortical hierarchy
can be constructed by assigning each interareal connection to one of only two classes (or a continuum between
them), traditionally deemed feedforward and feedback (Rockland and Pandya 1979; Felleman and Van Essen
1991; Markov et al. Still, two areas will often be connected reciprocally (bidirectionally)
and asymmetrically (one direction is feedforward and the other feedback). This notion of reciprocal asymmetric
connectivity seems to us better equipped to deal with realities of the cortex compared to the notion of a hierarchy.Where the paper says it
Where the paper says it
General Framework for Model Construction
A general framework for constructing models involves defining single neuron models, multiple cell types, and several prototypical projection types. Connectivity is defined in two steps: first, define cell type connectivity blueprints as matrices where the row index represents presynaptic cell type, the column index represents postsynaptic cell type, and the value represents synapse type. p. 5 This is the heart of the matter and can be done in two steps. First, define
cell type connectivity blueprints as matrices where the row index represents presynaptic cell type, the column
index represents postsynaptic cell type, and the value represents synapse type, see fig. 2024 for the diversity of feedback target patterns). Second, construct the matrix of interareal
connectivity where the row index represents the source area, the column index represents the target area, and
the value represents the connectivity blueprint that this connection follows, see fig.Where the paper says it
Where the paper says it
Canonical Circuit Search
The paper proposes using blackbox optimization methods or evolutionary strategies to fill connectivity matrices (‘blueprints’) between cell types such that the resulting architecture becomes functional. This constitutes a form of neural architecture search where the search space is a subspace of the space of models defined above. p. 7 That
is, a form of neural architecture search where the search space is a subspace of the space of models defined above. This constitutes a potentially complementary approach to the mapping of those matrices through experiments and
analysis of actual cortical data. Furthermore, the key
assumption that the connectivity between two cortical areas necessarily follows one of b cell type connectivity
blueprints greatly reduces the scaling with the number of areas, down to log2
(p)bt2 + log2
(b)a
2 = 2000 bits. The
encoding of connectivity could even potentially be reduced further into a few rules of connectivity applied to a
combinatorial code specifying cell type identity (Haber et al. This is akin to natural selection. Evolutionary search methods would then build a new set of genomes by applying crossover and mutation and repeat again the same
steps.Where the paper says it
Where the paper says it
Where the paper says it
Discussion on Future Directions
Future work should investigate different, likely less discrete indirect encoding of connectivity, notably in the context of neuroevolution. This includes the encoding of pre- and post-cell-type-specific connection probability and/or distance rule in [0, 1]t2, with t the number of cell types. p. 9 Future work should investigate different, likely less discrete indirect encoding of connectivity, notably in the
context of neuroevoution. This includes the encoding of pre- and post-cell-type-specific connection probability
and/or distance rule in [0, 1]t
2
, with t the number of cell types. Here, we sketch a general framework for constructing such models. A first step is to define single neuron models, characterized by a set of variables, a parameter space, and
a function describing neuronal dynamics. A projection type is defined by a parameter
space, a function describing its dynamics, and a function describing parameter learning. The dynamics of a
projection type typically describes how the variables of the postsynaptic neurons change as a function of the
variables of the presynaptic neurons and the projection parameters. Perhaps, another clue
lies in the observed parallels in the structure and computation of thalamus and cortical layer 6b (Zolnik et al. Our modelling framework very naturally extends to the definition of other brain structures beyond cortex
and their cell type connectivity, both intra- and inter- structures.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Acknowledgments
We thank Anna Vasilevskaya, Roman Doronin, and Georg Keller for helpful discussions of the reviewed data and feedback on the structure of the manuscript. This work was funded by the European Union’s Horizon Europe Programme under the Specific Grant Agreement No. p. 9 Future work should investigate different, likely less discrete indirect encoding of connectivity, notably in the
context of neuroevoution. This includes the encoding of pre- and post-cell-type-specific connection probability
and/or distance rule in [0, 1]t
2
, with t the number of cell types. Different excitatory cell types also show clear differences in functional response patterns, likely underlying specialized computational roles. For example, pyramidal cells of cortical layer 5 that project subcortically (pyramidal
tract, PT) and those that do not (intratelencephalic, IT) exhibit distinct activity profiles during decision-making
(Musall et al.Where the paper says it
Where the paper says it
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
-
Better model generalization through canonical blueprints: AI systems can be trained to search for functional architectures by using
cell type connectivity blueprints as matrices where the row index represents presynaptic cell type, the column index represents postsynaptic cell type, and the value represents synapse type.
This allows models to find functional circuits even when specific biophysical details are unknown. -
Task-specific function mapping: Improved AI can map computational roles by identifying
different cortical cell types map to different functions and computations
based on their specialized connectivity, enabling systems that perform tasks like decision-making or motor control with greater accuracy. -
Mechanistic model construction: AI can build mechanistic models by defining
single neuron models, characterized by a set of variables, a parameter space, and a function describing neuronal dynamics,
which allows for the creation of simulations that are constrained by biological principles rather than just data fitting. -
Automated neural architecture search: Systems can perform
neural architecture search where the search space is a subspace of the space of models defined above
to find architectures that are functional, such as findinga neural circuit that implements backpropagation
when given single neuron and projection models. -
Evolutionary optimization for connectivity: AI can use
blackbox optimization methods or evolutionary strategies to fill connectivity matrices between cell types,
which allows the system to search for architectures based on fitness selection, mimicking natural selection on connectivity rules.
Abstract
Neuronal circuits of the cerebral cortex are the structural basis of mammalian cognition. The same qualitative components and connectivity motifs are repeated across functionally specialized cortical areas and mammalian species, suggesting a single underlying algorithmic motif. Here, we propose a perspective on current knowledge of the cortical structure, from which we extract two core principles for computational modeling. The first principle is that cortical cell types fulfill distinct computational roles. The second principle is that cortical connectivity can be efficiently characterized by only a few canonical blueprints of connectivity between cell types. Starting with these two foundational principles, we outline a general framework for building functional and mechanistic models of cortical circuits. Finally, we show that the compressed and generative representations of connectivity entailed by this framework are particularly amenable to automatic search through gradient-free optimization, providing a concrete first step towards modelling the evolution of functional neural circuits.
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
- Message interaction dynamics covary with brain volume across mammalian connectomes
- 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