Building functional and mechanistic models of cortical computation based on canonical cell type connectivity
Authors: Arno Granier, Katharina A Wilmes, Mihai A Petrovici, Walter Senn
Organizations: 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, Switzerland · Center for Artificial Intelligence in Medicine, University of Bern, Switzerland
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.
Figures & tables
Figure 1: Cortical neuron types. The extent of the apical dendritic trees is represented in dark colors, and patterns of local axonal projections in light colors. As sketches, these have some degree of arbitrariness and do not represent the variability of arborization within one class. Marker genes defining a cell type are indicated in italics. Excitatory cells can be named classically based on their laminar position (L2IT - L6b; IT: intratelencephalic, s: stellate cells, p: pyramidal, PT: pyramidal tract, CT: corticothalamic) and inhibitory cells based on expressed molecular markers (Pvalb, Sst, VIP, Lamp5). Excitatory cells are pyramidal in shape, except for L4s and L6b which have different morphologies. Interneuron subtypes depicted here are representative subtypes from the Allen Cell Types Database, notably as reported by Patiño et al. (2024) and Wu et al. (2023) . Pyramidal subtypes depicted are the major classes while more minor classes, such as near-projecting layer 5 pyramidal cells and Car3 expressing layer 6 pyramidal cells, are not represented individually. Good entry points for a catalog of those genes are the work of Matho et al. (2021) or the Allen Cell Types Database. Note that gene expression varies greatly across areas for excitatory but not inhibitory cells ( Tasic et al., 2018 ) . Consequently, some genes presented here are area-specific, such as Adamts2, Agmat, and Rrad, which characterize layer 2/3 in primary visual cortex but not in anterior lateral motor cortex ( Tasic et al., 2018 ) .
Figure 2: Canonical local cortical circuits. (a) The main information flow in a cortical column of the mouse barrel cortex as reported by Sievers et al. (2024) . The thickness of the lines indicates the strength of the projection. Dotted lines indicate the dendritic location of projections from L3IT and L4IT to L5PT. Recurrent projections from cells of one type to other cells of the same type are present throughout (not drawn). (b) The logic of intercolumnar feedforward inhibition as reported by Sievers et al. (2024) . One population of dendrite-targeting (L3d) and one population of soma-targeting (L3s) interneurons per excitatory class/layer mediate feedforward inhibition. These are controlled by local populations via direct connections and disinhibitory connections relaying through a population of bipolar cells (VIP). L4IN: layer 4 interneurons. (c) Cell-type-specific targeting of Sst interneurons subtypes as reported by Wu et al. (2023) . Expressed genes are reported in italics. (d) Local outputs of L5PT (morphologically-defined thick-tufted layer 5 pyramidal cells) as reported by Bodor et al. (2023) .
Figure 3: Specific structure of interareal projections. (a,b) Cell types of origin and laminar target patterns of main interareal feedforward (a) and feedback (b) projections, based on integrated data from Markov et al. (2014) and Harris et al. (2019) . Note that this mixes data from monkeys and mice. Harris et al. (2019) additionally report interareal projections from pyramidal cells of layer 4, not represented here. The separation of pyramidal cells from layers 2 and 3 stems from Markov et al. (2014) (in primates), while Harris et al. (2019) (in mouse) represent both of these as a homogeneous class ( Cux2 ), rendering their laminar target patterns unclear. In both studies, the granularity of layer 6 pyramidal cells could not be accounted for. Finally, Harris et al. (2019) report that interareal projections from layer 5 PT and layer 6 CT pyramidal cells are less numerous than from IT pyramidal cells. More data is needed on these matters. (c) A sketch of a typical interareal graph of connectivity, with only the strongest connections represented. Each square is an area, and colors represent broad groupings of areas, typically lobes. The spatial arrangement of squares does not reflect the physical placement of areas but is based solely on connectivity (e.g. following a force-directed graph drawing algorithm). This type of graph but with real data is presented e.g. by Gămănuţ et al. (2018) , their fig. 8. (d) Single pyramidal cells in mouse primary visual cortex project to non-random sets of other areas ( Han et al., 2018 ) . Notably, pyramidal cells in the primary visual cortex projecting to anterolateral (AL) and posteromedial (PM) visual areas are segregated populations with little local interactions ( Kim et al., 2018 ) . (e) Pyramidal cells preferentially receive feedback from areas that they themselves project to, in particular, for pyramidal cells of layer 5, at their apical dendrites ( Kim et al., 2020 ; Siu et al., 2021 ; Young et al., 2021 ) . V2L: lateral visual areas, V2M: medial visual areas.
Figure 4: What do we need for a functional and mechanistic model of cortical computation? This figure sketches model elements: the specifics are simply illustrative. (a) Defining cortical cell types. Different shapes of neuron sketches represent different neuron models (details pending), while different colors represent different cell types. Additionally to cortical cell types, we also sketch a thalamic cell type. (b) Defining prototypical synapse types. On top of the arrows we sketch the plasticity kernels, on the right the possible subcellular targets, and on the bottom the connectivity patterns. (c) Sketches of cell type connectivity blueprints matrices. Colors refer to the colors of synapse types as in b. (d) A sketch of an interareal connectivity matrix. Pictograms represent different connectivity blueprints.
Figure 5: A circuit implementing backpropagation. More precisely, the variation of Kolen and Pollack (2002) avoiding weight transport, on a multilayer perceptron with ReLU nonlinearities. (a) Two cell types using neuron models with two (red) and three (purple) compartments. s=somatic activity, d=dendritic activity, t=target. (b) Two projection types: a one-to-one non-plastic projection setting postsynaptic dendritic activity (pink) and an all-to-all projection type with learnable parameter W. (c) Canonical cell type connectivity. (d) Interareal connectivity organized in a strict hierarchy. (e) A sketch of the resulting network.
Figure 6: Canonical cortical circuit search. Given neuron and projection models, cell type connectivity and interareal connectivity matrices encode all the information necessary to construct actual networks. Here, cell type connectivity and interareal connectivity matrices are equivalently represented flattened and concatenated. Sets of connectivity matrices can be seen as genomes, actual networks as individuals, and the process that transforms the former into the latter as development. These networks can be trained and tested on functional tasks, and the underlying genomes filtered based on functional performance. 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.
Figure 7: Finding a circuitry for backpropagation through search. (a) The YinYang dataset ( Kriener et al., 2022 ) . Inputs are the (x,y) coordinates; target labels are represented by the colors. (b) A sketch of a small network with one hidden layer comprising 20 units. Arrows marked with ‘?’ indicate existing connections wih unknown (varying) circuitry (cell-type-specific connectivity). (c) Accuracy of the best found circuitry (applied to a network as depicted in panel b) during the search process, that is, across different generations (populations with varying circuitries). Vertical lines correspond to the generations for which the best found circuits are sketched in panel g. (d) Similar to panel c but reporting the normalized ‘metabolic’ cost of the best found circuitry across generations. This cost scales with the number of existing projections between cell-type-specific populations. (e) Circuitries or canonical cell type connectivities with best performances in the population for different generations during the search process. (f) Compressed binary encodings of the cell type connectivity matrices reported in panel e. (g) Sketches of circuitries for corresponding subpanels i, iii, v, vii of panel e.
How the wiring and functional organization of cortex shape recurrent computation remains a central question in both neuroscience and machine learning. Here, we leverage data released through the Machine Intelligence from Cortical Networks (MICrONS) program--a functional connectomics resource spanning multiple areas of mouse visual cortex, in which dense calcium imaging is co-registered with high-resolution electron microscopy reconstruction from the same animal--to build biologically grounded recurrent neural networks. Using neuronal spatial coordinates, anatomical connectivity, and function-derived relationships from nearly 12,000 coregistered excitatory neurons, we initialize recurrent weights and impose communication-aware spatial constraints during learning. Across three cognitive decision-making tasks, networks constrained by cortical structure and function consistently outperform baseline and partially constrained models. Functional weight initialization provides the largest gain, while real spatial embedding yields robust additional improvements across conditions. These biologically grounded networks also develop low-entropy, modular, and small-world organization, and retain strong performance even when recurrence is restricted to positive weights. Together, our results show that the machinery of cortex--its geometry, wiring, and functional structure--can be harnessed as a powerful inductive basis for building recurrent networks that learn more effectively while converging toward key organizational principles of biological computation.
Mo Shakiba, Rana Rokni, Mohammad Mohammadi +1
Neuromatch Academy, Neuromatch, Inc., USA · McGovern Institute for Brain Research, Massachusetts Institute of Technology (MIT)
This work simulates the developmental process of cortical neurogenesis, initiating from a single stem cell and governed by gene regulatory rules derived from mouse single-cell transcriptomic data. The developmental process spontaneously generates a heterogeneous population of 5,000 cells, yet yields only 85 mature neurons - merely 1.7% of the total population. These 85 neurons form a densely interconnected core of 200,400 synapses, corresponding to an average degree of 4,715 per neuron. At iteration zero, this minimal circuit performs at chance level on MNIST. However, after a single epoch of standard training, accuracy surges to over 90% - a gain exceeding 80 percentage points - with typical runs falling in the 89-94% range depending on developmental stochasticity. The identical circuit, without any architectural modification or data augmentation, achieves 40.53% on CIFAR-10 after one epoch. These findings demonstrate that developmental rules sculpt a domain-general topological substrate exceptionally amenable to rapid learning, suggesting that biological developmental processes inherently encode powerful structural priors for efficient computation.
Connectomes, graph-level maps of neurons and their synaptic connections, provide a structural basis for understanding how brain circuits support function and computation. However, mapping connectome structure to computation remains difficult because these graphs are high-dimensional, sparse, and sensitive to local structural variation. Existing approaches often depend on hand-crafted structural descriptors or task-specific predictors, which limits their ability to represent connectomes in a form that is both generative and functionally meaningful. We propose a conditional generative latent framework that encodes connectome graphs into a compact structural space while using available node-level conditions to guide reconstruction and generation. From this space, the model can reconstruct observed connectivity with a mean edge-reconstruction AUC up to 0.910 and generate new candidate connectomes, enabling a unified analysis of graph structure and computational behavior. Using connectome-derived graphs as recurrent computational substrates, we found that the learned latent space captures functional variation across reservoir-computing experiments, with cross-validated R2 values up to approximately 0.87. Interpretability analysis further revealed task-specific structural mechanisms: in our examples, memory performance is associated with reciprocal recurrent connectivity, whereas prediction and classification are more strongly associated with spectral properties of the recurrent network. These findings suggest an AI-for-science approach to linking neural connectivity to computation and provide a generative and interpretable basis for studying how distinct structural mechanisms shape computational capacity.
Zhuolin Yu, Xingyu Liu, Yuanhao Jia +4
National Key Laboratory for Multimedia Information Processing, School of Computer Science, Peking University, Beijing, China · School of Electronics Engineering and Computer Science, Peking University, Beijing, China · School of Computer Science, Beijing University of Posts and Telecommunications, Beijing, China +2