CoChat
Sign Up
PreprintCited by 33Neural dynamics and brain functionVisual perception and processing mechanismsNeurobiology and Insect Physiology Research

Connectome-constrained deep mechanistic networks predict neural responses across the fly visual system at single-neuron resolution

Open access
Lappalainen, Janne K., Tschopp, Fabian D., Prakhya, Sridhama, McGill, Mason, Nern, Aljoscha, Shinomiya, Kazunori, Takemura, Shin-ya, Gruntman, Eyal +2 more·2023·bioRxiv (Cold Spring Harbor Laboratory)
DOI 10.1101/2023.03.11.532232
Open this paper in CoChat — free accountOpen PDF

Abstract

We can now measure the connectivity of every neuron in a neural circuit, but we are still blind to other biological details, including the dynamical characteristics of each neuron. The degree to which connectivity measurements alone can inform understanding of neural computation is an open question. Here we show that with only measurements of the connectivity of a biological neural network, we can predict the neural activity underlying neural computation. We constructed a model neural network with the experimentally determined connectivity for 64 cell types in the motion pathways of the fruit fly optic lobe but with unknown parameters for the single neuron and single synapse properties. We then optimized the values of these unknown parameters using techniques from deep learning, to allow the model network to detect visual motion. Our mechanistic model makes detailed experimentally testable predictions for each neuron in the connectome. We found that model predictions agreed with experimental measurements of neural activity across 24 studies. Our work demonstrates a strategy for generating detailed hypotheses about the mechanisms of neural circuit function from connectivity measurements. We show that this strategy is more likely to be successful when neurons are sparsely connected—a universally observed feature of biological neural networks across species and brain regions.

AI snapshot

Written from the paper’s full text.

TL;DR

The authors built a connectome-constrained, task-optimized deep mechanistic network of the fly optic lobe motion pathways using only measured connectivity and the task of optic flow computation. The model predicted neural tuning across many cell types and 24 experimental studies, and suggested TmY3 as a novel candidate motion detector.

Key findings

  • A connectome-constrained DMN with 45,669 neurons, 1,513,231 connections, 65 cell types, and 721 hexagonal columns reduced the model to 734 free parameters; it was trained on optic flow from Sintel videos.
  • The ensemble predicted ON/OFF segregation for 31 of 31 experimentally established cell types (median FRI, p = 4.7 × 10⁻¹⁰, binomial test); the best model correctly assigned 29 of 31 (p = 2.3 × 10⁻⁷).
  • Task-optimized models predicted known direction selectivity and flash response indices better than random-parameter models (p = 2 × 10⁻¹¹ and p = 5.3 × 10⁻⁵, Mann-Whitney-U), and better task performance correlated with better T4/T5 direction selectivity predictions (p = 5.6 × 10⁻⁶, Wald test).
  • The model proposed TmY3 as an ON-motion-selective neuron independent of T4/T5; this prediction has not yet been experimentally tested. In synthetic MNIST networks, sparse connectivity enabled accurate prediction from connectivity alone (median Pearson’s correlation 0.856 for 10% average connectivity), while dense networks were essentially uncorrelated unless noisy connection strengths were also known.

Methods

  • The connectome was assembled from FIBSEM FIB-25 and FIB-19 datasets containing 1,801 neurons, with 830 hand-annotated column positions; a probabilistic expectation-maximization algorithm assigned remaining neurons and fused datasets into a coherent optic lobe connectome.
  • The DMN used threshold-linear point-neuron dynamics with graded chemical synapses; free parameters were 65 time constants, 65 resting potentials, and 604 unitary synapse scaling factors, with CT1 modeled as two compartments, CT1(M10) and CT1(Lo1).
  • Models were implemented in PyTorch and trained end-to-end with backpropagation through time on the Sintel optic flow task; 50 models were generated, and predictions were compared with experimental measurements from 24 studies. Synthetic feedforward MNIST networks with varying connectivity sparsity were used to test when connectomic constraints predict neural responses.

Limitations

  • The paper states the reductionist model cannot account for electrical synapses, complex chemical synapse dynamics, or neuromodulation, and only considers motion detection among many visual computations.
  • The paper notes that connectomic reconstructions can be noisy, some cell types lacked transmitter-phenotype data and used guessed signs, and the TmY3 motion prediction remains experimentally untested.
  • Evident from the text: the model assumes perfect periodicity/translation invariance, uses simplified single-compartment point neurons except CT1, ignores neuronal superposition, covers only the central visual field of the right eye, and relies on an incomplete optic lobe reconstruction with some manually adjusted filters.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1101/2023.03.11.532232

Cites

123 of 123 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Claims the paper makes

0 of 53 cited sources checked · 0 supported · 0 flagged

  • Findingown data

    The model accurately predicts the direction selectivity of the T4 and T5 motion detector neurons.

    “as well as the direction selectivity of the well-known T4 and T5 motion detector neurons”
    Quote in the paperSubstantial claim
  • Findingown data

    Model predictions agreed with experimental measurements of neural activity across 24 studies.

    “We found that model predictions agreed with experimental measurements of neural activity across 24 studies.”
    Quote in the paperSubstantial claim
  • Findingown data

    The model accurately predicts the segregation of the visual system into light increment (ON) and light decrement (OFF) channels.

    “We found that our connectomeconstrained and task-optimized DMN accurately predicts the well-known segregation of the visual system into light increment (ON) and light decrement (OFF) channels”
    Quote in the paperSubstantial claim
  • Findingown data

    In the fourth cluster, TmY3 is not direction selective.

    “In the fourth cluster (star marker), TmY3 is, again, not direction selective.”
    Quote in the paperNot a substantial claim
  • Findingown data

    Densely connected synthetic networks, where each neuron in a given layer connects to every neuron in the next layer, are essentially uncorrelated with their task- and connectome-constrained models constructed with known connectivity but unknown strength.

    “We found that densely connected synthetic networks, where each neuron in a given layer connects to every neuron in the next layer, are essentially uncorrelated with their task-and connectome-constrained models constructed with known connectivity but unknown strength (Fig. 6b, Extended Data Fig. 9a).”
    Quote in the paperSubstantial claim0 of 2 sources checked
    • cites Fig. 6b

      Not checked in the free run.

    • cites Extended Data Fig. 9a

      Not checked in the free run.

  • May combine two claims. 44
  • Classified as probably not a substantial claim. 23
  • The quote only partly states this claim. 2

  • May combine two claims.

    Findingcitation

    ON-impulses on the leading side of the receptive field of T4c lead to fast depolarization, whereas impulses on the trailing side lead to delayed hyperpolarization, matching experimental findings.

    “ON-impulses on the leading side of the receptive field of the ON-contrast, upwards direction selective T4c cell lead to its fast depolarization, whereas impulses on the trailing side of the receptive field lead to a delayed hyperpolarization, again matching experimental findings 54 .”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [54] 2018, Simple integration of fast excitation and offset, delayed inhibition computes directional selectivity in Drosophila, Nature Neuroscience source

      Not checked in the free run.

  • May combine two claims.

    Findingnone given

    The connectome-constrained model comprises 45,669 neurons and 1,513,231 connections, across 65 cell types arranged in a hexagonal lattice consisting of 721 columns.

    “In total, the connectome-constrained model comprises 45,669 neurons and 1,513,231 connections, across 65 cell types arranged in a hexagonal lattice consisting of 721 columns”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Synthetic networks that perform the same handwritten digit recognition task with near perfect accuracy use dramatically different connectivity and different mechanisms.

    “While the synthetic networks all perform the same task with near perfect accuracy, they use dramatically different connectivity, and so use different mechanisms to perform the same task.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    In the best task performing model, the four subtypes of T4 respond strongly to dark edges, and the four subtypes of T5 neurons to bright edges, moving in the four cardinal directions, in agreement with previous experimental findings.

    “First, we found that in the best task performing model, the four subtypes of the T4 respond strongly to dark edges, and the four subtypes of the T5 neurons to bright edges, moving in the four cardinal directions, in agreement with previous experimental findings”
    Quote in the paperSubstantial claim0 of 4 sources checked
    • cites [54] 2018, Simple integration of fast excitation and offset, delayed inhibition computes directional selectivity in Drosophila, Nature Neuroscience source

      Not checked in the free run.

    • cites [55] 2019, The computation of directional selectivity in the drosophila off motion pathway, Elife source

      Not checked in the free run.

    • cites [62] source

      Not checked in the free run.

    • cites [64] source

      Not checked in the free run.

  • May combine two claims.

    Findingown data

    Models with the correct direction selectivity for T4 neurons also predict the correct contrast selectivity for Mi4 and Mi9 neurons, and vice versa.

    “Models with the correct direction selectivity for T4 neurons also predict the correct contrast selectivity for Mi4 and M9 neurons, and vice versa (Fig. 4f).”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingreasoning

    T5c OFF-impulse responses are inverted, resembling the T4c spatiotemporal receptive field.

    “Because T5c is OFF-selective, its OFF-impulse responses are inverted, resembling the T4c spatiotemporal receptive field (Extended Data Fig. 7a).”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    The model correctly predicts the spatial scale of the spatial receptive field for all major inputs to T4 and T5 neurons, and correctly predicts the preferred contrast for all but one of the major inputs to T4 and T5 neurons.

    “We find that the model correctly predicts the spatial scale of the spatial receptive field for all major inputs to T4 and T5 neurons, and correctly predicts the preferred contrast for all but one of the major inputs to T4 and T5 neurons.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Temporal response properties are correctly predicted for all except the Tm4 cell, which is incorrectly predicted to be ON-selective by its temporal receptive field.

    “These temporal response properties are correctly predicted for all except the Tm4 cell in this model, which is incorrectly predicted to be ON-selective by its temporal receptive field.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    For T4c neurons, there are three clusters corresponding to qualitatively distinct responses: two clusters contain direction selective T4c cells with up- and down-selective cardinal tuning, and the third cluster is not direction tuned.

    “For T4c neurons for example, we found three clusters corresponding to qualitatively distinct responses of this cell type for naturalistic inputs: Two clusters contain models with direction selective T4c cells (Fig. 4b and c) with up-and down-selective cardinal tuning, respectively, whereas neurons in the third cluster are not direction tuned.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingcitation

    Tm3 and Tm4 have broad spatial receptive fields (two column radius, 11.6°), while Mi1, Mi4, Mi9, Tm1, Tm2, Tm9, and CT1 compartments in both medulla and lobula have narrow spatial receptive fields (single column radius, 5.8°).

    “In agreement with experimental findings 70,93 , we find that Tm3 and Tm4 have broad spatial receptive fields (two column radius, 11.6 • ), while Mi1, Mi4, Mi9, Tm1, Tm2, Tm9, and CT1 compartments in both medulla and lobula have narrow spatial receptive fields (single column radius, 5.8 • ).”
    Quote in the paperSubstantial claim0 of 2 sources checked
    • cites [70] source

      Not checked in the free run.

    • cites [93] source

      Not checked in the free run.

  • May combine two claims.

    Findingown data

    T4c cells in DMNs in the first and second clusters are ON-motion direction selective, whereas those in the third cluster are not.

    “T4c cells in DMNs in the first and second clusters (circular and triangular marker) are ON-motion direction selective, whereas those in the third cluster (square marker) are not.”
    Quote in the paperSubstantial claim
    • The authors state this as an assumption: Synapse strength is proportional to synapse count from the connectome times a scalar for each filter.

  • May combine two claims.

    Findingcitation

    Temporal receptive fields for inputs are in agreement with experimental measurements, with the exception of Tm4.

    “Temporal receptive fields for inputs also are in agreement with experimental measurements 60,93 , with the exception of Tm4 (red cross).”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [60] source

      Not checked in the free run.

    • cites [93] source

      Not checked in the free run.

  • May combine two claims.

    Findingown data

    The DMN correctly predicts narrow spatial receptive fields for Tm1, Tm2, Tm9, and CT1(Lo1), and a wide receptive field for Tm4.

    “For major T5 inputs, the DMN correctly predicts narrow spatial receptive fields for Tm1, Tm2, Tm9, and CT1(Lo1), and a wide receptive field for Tm4.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    Simulations constrained only with connectivity measurements correlate well only for low connectivities, while simulations constrained by measurements of both connectivity and connection strength correlate well across all average connectivity percentages.

    “Simulations constrained only with connectivity measurements only correlate well for low connectivities (dark blue), while simulations constrained by measurements of both connectivity and connection strength correlate well across all average connectivity percentages (orange).”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    For ON-motion stimuli, in cluster 2 DMNs T4c receives central depolarizing input from Mi1 and Tm3, ventral hyperpolarizing input from Mi9 and dorsal hyperpolarizing input from CT1(M10).

    “For ON-motion stimuli, in these DMNs, T4c receives central depolarizing input from Mi1 and Tm3, ventral hyperpolarizing input from Mi9 and dorsal hyperpolarizing input from CT1(M10).”
    Quote in the paperSubstantial claim
    • The authors state this as an assumption: Synapse strength is proportional to synapse count from the connectome times a scalar for each filter.

  • May combine two claims.

    Findingreasoning

    For ON-motion stimuli, in cluster 3 DMNs T4c receives central depolarizing input from Mi1 and Tm3, dorsal hyperpolarizing input from Mi4 and CT1(M10), and ventral hyperpolarizing input from Mi9.

    “For ON-motion stimuli, in these DMNs, T4c receives central depolarizing input from Mi1 and Tm3, dorsal hyperpolarizing input from Mi4 and CT1(M10), and ventral hyperpolarizing input from Mi9.”
    Quote in the paperSubstantial claim
    • The authors state this as an assumption: Synapse strength is proportional to synapse count from the connectome times a scalar for each filter.

  • May combine two claims.

    Findingown data

    For clusters in which TmY3 is motion selective, cluster 1 indicates ON-selectivity for Tm3, Mi1, and Mi4 cells, and OFF-selectivity for L4, Tm2, and Mi9 cells, in agreement with known selectivities.

    “For clusters in which TmY3 is motion selective, cluster 1 (TmY3 tuning to downwards/front-to-back motion, circular marker) indicates ONselectivity for Tm3, Mi1, and Mi4 cells, and OFF-selectivity for L4, Tm2, and Mi9 cells, in agreement to the known selectivities.”
    Quote in the paperSubstantial claim
    • The authors state this as an assumption: Synapse strength is proportional to synapse count from the connectome times a scalar for each filter.

  • May combine two claims.

    Findingown data

    Half of the 65 cell types receive input from 21 up to 200 cells, while the other half receives input from less than 21 cells.

    “(b) (left) Half of the 65 cell types receive input from 21 up to 200 cells, while the other half receives input from less than 21 cells.”
    Quote in the paperNot a substantial claim
    • The authors state this as an assumption: For columnar, spatially repeated neurons, only one neuron per position is present.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Findingown data

    Half of the 65 cell types project onto more than six other cell types, while the other half projects onto less than six.

    “(right) Half of the 65 cell types project onto more then six other cell types, while the other half projects onto less than six.”
    Quote in the paperNot a substantial claim
    • The authors state this as an assumption: For columnar, spatially repeated neurons, only one neuron per position is present.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingown data

    Models generalized on optic flow computation after round about 250,000 iterations.

    “Models generalized on optic flow computation after round about 250,000 iterations, yielding functional candidates for our fruit fly visual system models that we analyzed with respect to their tuning properties.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    Cluster 2 (T4c downwards tuned) indicates Mi1, Tm3, Mi9, and CT1(M10) are ON-selective and Mi4 is OFF-selective.

    “For cluster 2 (T4c in DMN downwards tuned, triangle), Mi1, Tm3, Mi9, and CT1(M10) are ON-selective and Mi4 is OFF-selective.”
    Quote in the paperSubstantial claim
    • The authors state this as an assumption: Synapse strength is proportional to synapse count from the connectome times a scalar for each filter.

  • May combine two claims.

    Findingown data

    An integration time step smaller than 20 ms (i.e., a framerate of 50 Hz after resampling) did not yield qualitatively superior task performance nor more realistic tuning predictions.

    “We did not find that an integration time step smaller than 20 ms, i.e. a framerate of 50 Hz after resampling, yielded qualitatively superior task performance nor more realistic tuning predictions.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Findingown data

    Direction selectivity index from neural responses to moving edges for the same 10 best models.

    “(c) Direction selectivity index (DSI) from neural responses to moving edges (Methods, Equation 4) for the same 10 best models as in (b).”
    Quote in the paperSubstantial claim
    • The authors state this as an assumption: Synapse strength is proportional to synapse count from the connectome times a scalar for each filter.

  • May combine two claims.

    Findingown data

    Half of the 65 cell types project output onto 20 up to 200 cells, while the other half projects output onto less than 20 cells.

    “(right) Half of the 65 cell types project output onto 20 up to 200 cells, while the other half projects output onto less than 20 cells.”
    Quote in the paperNot a substantial claim
    • The authors state this as an assumption: For columnar, spatially repeated neurons, only one neuron per position is present.

  • May combine two claims.

    Findingown data

    Cluster 1 (T4c upwards tuned) indicates ON-selectivity for Mi1, Tm3, Mi4, and CT1(M10), and OFF-selectivity for Mi9.

    “Cluster 1 (T4c in DMN upwards tuned, circle) indicates ON-selectivity for Mi1, Tm3, Mi4, and CT1(M10), and OFF-selectivity for Mi9.”
    Quote in the paperSubstantial claim
    • The authors state this as an assumption: Synapse strength is proportional to synapse count from the connectome times a scalar for each filter.

  • May combine two claims.

    Findingown data

    Half of the connections are characterized by less than 1.6 synapses while the other half are characterized by 1.6 up to hundreds of synapses.

    “(c) Half of the connections are characterized by less than 1.6 synapses while the other half are characterized by 1.6 up to hundreds of synapses.”
    Quote in the paperNot a substantial claim
    • The authors state this as an assumption: For columnar, spatially repeated neurons, only one neuron per position is present.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingown data

    ON/OFF contrast selectivity indices for each cell type based on peak transient responses for flash stimuli for 10 best models.

    “(b) ON-and OFF-contrast selectivity indices for each cell type based on peak transient responses for flash stimuli (Methods, Equation 3) for 10 models with best task performance (10 worst-performing models in Extended Data Fig. 4).”
    Quote in the paperNot a substantial claim
    • The authors state this as an assumption: Synapse strength is proportional to synapse count from the connectome times a scalar for each filter.

  • May combine two claims.

    Findingown data

    Even just 10% (83 positions) of available ground truth was sufficient to robustly position the majority of the neurons (64.3%, 534 positions) into the correct columns, and annotate 48% (865 neurons) perfectly consistent.

    “We found that even just 10% (83 positions) of the available ground truth was sufficient to robustly position the majority of the neurons (64.3%, 534 positions) into the correct columns, and annotate 48% (865 neurons) perfectly consistent.”
    Quote in the paperSubstantial claim
    • The authors state this as an assumption: For columnar, spatially repeated neurons, only one neuron per position is present.

  • May combine two claims. The quote only partly states this claim.

    Findingown data

    The best-performing DMN correctly predicts narrow spatial receptive fields for Mi1, Mi4, Mi9, and CT1(M10), and a wide receptive field for Tm3.

    “For major T4 inputs, the best-performing DMN correctly predicts narrow spatial receptive fields for Mi1, Mi4, Mi9, and CT1(M10), and a wide receptive field for Tm3.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Findingown data

    The EM-datasets contain 1801 neurons (702 from FIB-25 and 1099 from FIB-19), with hand-annotated positions available for 830 of these neurons.

    “In total, they contain 1801 neurons (702 from FIB-25 and 1099 from FIB-19), with hand-annotated positions available for 830 of these neurons (SI Figure 1).”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites SI Figure 1

      Not checked in the free run.

    • The authors state this as an assumption: For columnar, spatially repeated neurons, only one neuron per position is present.

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Findingown data

    Half of the 65 cell types receive input from more than ten other cell types, while the other half receives input from less than ten.

    “(a) (left) Half of the 65 cell types receive input from more than ten other cell types, while the other half receives input from less than ten.”
    Quote in the paperNot a substantial claim
    • The authors state this as an assumption: For columnar, spatially repeated neurons, only one neuron per position is present.

  • May combine two claims.

    Findingcitation

    Fast excitation and delayed, offset inhibition enable T4c to detect motion, in agreement with experimental measurements.

    “Fast excitation and delayed, offset inhibition enable T4c to detect motion, in agreement with experimental measurements 54 .”
    Quote in the paperSubstantial claim0 of 1 sources checked
    • cites [54] 2018, Simple integration of fast excitation and offset, delayed inhibition computes directional selectivity in Drosophila, Nature Neuroscience source

      Not checked in the free run.

  • May combine two claims.

    Findingown data

    A pair of presynaptic and postsynaptic cell type is connected by 5.5 synapses in half of the cases and by more than 5.5 up to hundreds in the other half of the cases.

    “(d) A pair of presynaptic and postsynaptic cell type is connected by 5.5 synapses in half of the cases and by more than 5.5 up to hundreds in the other half of the cases.”
    Quote in the paperNot a substantial claim
    • The authors state this as an assumption: For columnar, spatially repeated neurons, only one neuron per position is present.

  • May combine two claims.

    Findingown data

    Half of the 65 cell types receive inhibitory inputs from less than 39.2 synapses and the other half from 39.2 to hundreds.

    “Half of the 65 cell types receive inhibitory inputs from less than 39.2 synapses and the other half from 39.2 to hundreds.”
    Quote in the paperNot a substantial claim
    • The authors state this as an assumption: For columnar, spatially repeated neurons, only one neuron per position is present.

  • May combine two claims.

    Findingown data

    Half of the 65 cell types receive input from less than 93.6 synapses and the other half between 93.6 to 400 synapses.

    “(e) (left) Half of the 65 cell types receive input from less than 93.6 synapses and the other half between 93.6 to 400 synapses.”
    Quote in the paperNot a substantial claim
    • The authors state this as an assumption: For columnar, spatially repeated neurons, only one neuron per position is present.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingnone given

    Weak autapses in the hexagonal model did not affect the tuning predictions.

    “this allowed for weak autapses in our hexagonal model that did not affect the tuning predictions.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Findingown data

    Half of the 65 cell types project less than 55.9 synapses and the other half projects between 55.9 to 600 synapses.

    “(right) Half of the 65 cell types project less than 55.9 synapses and the other half projects between 55.9 to 600 synapses.”
    Quote in the paperNot a substantial claim
    • The authors state this as an assumption: For columnar, spatially repeated neurons, only one neuron per position is present.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingcitation

    The reconstruction did not fully capture the asymmetry reported in [48] of the T5 anatomical receptive field of Tm9.

    “We found that the reconstruction did not fully capture the asymmetry reported in 48 of the T5 anatomical receptive field of Tm9”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [48] 2019, Comparisons between the ON-and OFF-edge motion pathways in the Drosophila brain, Elife

      Not checked in the free run.

  • May combine two claims.

    Findingown data

    At least half of the 65 cell types project no inhibitory synapses and the rest project between zero to hundreds.

    “At least half of the 65 cell types project no inhibitory synapses and the rest project between zero to hundreds.”
    Quote in the paperNot a substantial claim
    • The authors state this as an assumption: For columnar, spatially repeated neurons, only one neuron per position is present.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Findingown data

    For few T4 and T5 inputs, the number of input synapses reported in the literature [48] slightly deviated from our reconstruction.

    “For few T4 and T5 inputs the number of input synapses reported in the literature 48 slightly deviated from our reconstruction.”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [48] 2019, Comparisons between the ON-and OFF-edge motion pathways in the Drosophila brain, Elife

      Not checked in the free run.

  • May combine two claims.

    Findingown data

    Half of the 65 cell types receive excitatory inputs from less than 52.3 synapses and the other half from 52.3 to hundreds.

    “(f) Separating (e) into excitatory and inihibitory synapses, (left) we see that half of the 65 cell types receive excitatory inputs from less than 52.3 synapses and the other half from 52.3 to hundreds.”
    Quote in the paperNot a substantial claim
    • The authors state this as an assumption: For columnar, spatially repeated neurons, only one neuron per position is present.

  • The quote only partly states this claim.

    Interpretationown data

    A connectome-constrained, task-optimized deep mechanistic network constructed with experimentally determined connectivity for 64 cell types in the fly optic lobe motion pathways and unknown single-neuron and single-synapse parameters can predict neural responses at single-neuron resolution.

    “We constructed a model neural network with the experimentally determined connectivity for 64 cell types in the motion pathways of the fruit fly optic lobe but with unknown parameters for the single neuron and single synapse properties.”
    • Also stated in: § Abstract · plain
      “Our mechanistic model makes detailed experimentally testable predictions for each neuron in the connectome.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation

    In neuroscience, there have been competing proposals by theorists for neural circuit mechanisms of visual motion and for integration of eye velocity commands into eye position signals, with each proposal suggesting different neural connectivity.

    “Similarly in neuroscience, there have been competing proposals by theorists for example, for the neural circuit mechanisms of the computation of visual motion 32,33 , and for the integration of eye velocity commands into eye position signals 34,35 , with each proposal suggesting different neural connectivity.”
    Quote in the paperNot a substantial claim0 of 4 sources checked
    • cites [32] Reichardt, W. Autocorrelation, a principle for evaluation of sensory information by the central nervous system. Principles of sensory communications (1961).

      Not checked in the free run.

    • cites [33] source

      Not checked in the free run.

    • cites [34] Seung, H. S. How the brain keeps the eyes still. Proc. Natl. Acad. Sci. USA 13339–13344 (1996). source

      Not checked in the free run.

    • cites [35] source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Interpretationnone given

    The model suggests that TmY3 might also detect motion.

    “Our model further suggests that TmY3 might also detect motion, a prediction that has yet to be experimentally tested.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Study design and conductnone given

    The authors required that at the macroscopic scale, the collective neural activity dynamics across the entire network result in an ethologically relevant computation.

    “We also required that at the macroscopic scale, the collective neural activity dynamics across the entire network result in an ethologically relevant computation.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Interpretationreasoning

    TmY4 and TmY18 cells receive inputs from T4 cells, potentially inheriting their motion tuning.

    “In contrast, TmY4 and TmY18 cells receive inputs from T4 cells, potentially inheriting their motion tuning.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationown data

    The ensemble suggests ON-motion sensitivity for TmY3, but different clusters disagree in their predictions for direction and contrast selectivity.

    “Together, the ensemble suggests ON-motion sensitivity for TmY3, but different clusters disagree in their predictions for direction and contrast selectivity.”
    Quote in the paperSubstantial claim

  • May combine two claims.

    Study design and conductreasoning

    The decoding network cannot by itself detect motion and must rely on the temporal dynamics of the connectome-constrained network to compute motion-selective visual features.

    “Importantly, the decoding network cannot by itself detect motion, which requires the comparison of current and past visual stimuli, but must instead rely on the temporal dynamics of the connectome-constrained network to compute motionselective visual features.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Whole nervous system connectome projects are nearing completion for the adult fruit fly.

    “Whole nervous system connectome projects are nearing completion for the adult fruit fly 2,13”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [2] source

      Not checked in the free run.

    • cites [13] 2022, FlyWire: online community for whole-brain connectomics, Nature Methods source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    There has been early evidence, in small scale models and larger models of the fly motion pathways, that combining connectome and task constraints in a single model might prove successful.

    “More recently, there has been tantalizing early evidence, in small scale models 46 and larger models 119 of the fly motion pathways, that our approach to combining connectome and taskconstraints in a single model might prove successful.”
    Quote in the paperNot a substantial claim0 of 2 sources checked
    • cites [46] 2021, Predicting individual neuron responses with anatomically constrained task optimization, Current Biology source

      Not checked in the free run.

    • cites [119] Tschopp, F. D. , Reiser, M. B. & Turaga, S. C. A Connectome Based Hexagonal Lattice Convolutional Network Model of the Drosophila Visual System. arXiv preprint arXiv:1806.04793 (2018). URL http://arxiv.org/abs/1806.04793.1806.04793.

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductnone given

    The authors ensured that subsamples from the same scene do not distribute across training and validation sets.

    “We ensured that such subsamples from the same scene do not distribute across training and validation sets.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Study design and conductnone given

    Creating 50 DMNs with random parameters and task-optimizing only their decoding network yields an upper bound for the task error of the task-optimized DMN and a lower bound for the accuracy of tuning predictions without task-optimization.

    “We created 50 DMNs with random parameters (all sampled from Gaussians of different means and standard deviations), and task-optimized only their decoding network, yielding an upper bound for the task error of the task-optimized DMN and a lower bound for the accuracy of tuning predictions without task-optimization.”
    Quote in the paperSubstantial claim
    • The authors state this as an assumption: Synapse strength is proportional to synapse count from the connectome times a scalar for each filter.

  • May combine two claims.

    Interpretationreasoning

    The fly visual system likely falls in the region between the two curves, since measured synapse counts inform relative connection strengths between pairs of neurons for the same pair of cell types, but not absolute connection strength.

    “The fly visual system likely falls in the region between the two curves, since measured synapse counts inform relative connection strengths between pairs of neurons for the same pair of cell types, but not absolute connection strength.”
    • Also stated in: § Measuring · hedged
      “The fly visual system likely falls in the region between the two curves, since measured synapse counts inform relative connection strengths between pairs of neurons for the same pair of cell types, but not absolute connection strength.”
    Quote in the paperSubstantial claim
    • The authors state this as an assumption: The individual units in the SCNs are interpreted as neurons with voltage.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    Whole mouse brain connectome projects are now being discussed.

    “and whole mouse brain connectome projects are now being discussed 121 .”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [121] source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing), S4 (based).

    Study design and conductnone given

    Estimates of the spatial asymmetry of connectivity from existing connectomic reconstructions can be noisy.

    “We note, however, that estimates of the spatial asymmetry of connectivity from existing connectomic reconstructions can be noisy.”
    Quote in the paperNot a substantial claim

  • Classified as probably not a substantial claim.

    Study design and conductnone given

    The authors verified the quality of their reconstruction as described in Supplementary Note 1.

    “We verified the quality of our reconstruction as described in Supplementary Note 1”
    Quote in the paperSubstantial claim

  • Classified as probably not a substantial claim.

    Study design and conductreasoning

    Only neurons consistently annotated with both 100% and 90% of reference positions used were counted to estimate the number of synapses between cell types and columns, in order to prune neuron offsets with low confidences.

    “Only the neurons consistently annotated with both 100% and 90% of reference positions used were counted to estimate the number of synapses between cell types and columns, in order to prune neuron offsets with low confidences.”
    Quote in the paperSubstantial claim

  • Probably not a substantial claim: may not pass S2 (checkable).

    Interpretationreasoning

    Our modeling approach provides a discovery tool aimed at using connectomic measurements to generate detailed, experimentally testable hypotheses for the computational role of individual neurons.

    “Our modeling approach provides a discovery tool, aimed at using connectomic measurements to generate detailed, experimentally testable hypotheses for the computational role of individual neurons.”
    Quote in the paperNot a substantial claim

  • May combine two claims.

    Interpretationreasoning

    Cluster 3 indicates ON-selectivity for Mi9 in contradiction to known selectivities and hence rules out the third TmY3 tuning solution.

    “In contrast, cluster 3 (TmY3 tuning to upwards/back-to-front motion, square marker) indicates ONselectivity for Mi9 in contradiction to the known selectivities and hence ruling out the third TmY3 tuning solution.”
    Quote in the paperSubstantial claim
    • The authors state this as an assumption: Synapse strength is proportional to synapse count from the connectome times a scalar for each filter.

  • May combine two claims.

    Interpretationown data

    Cluster analysis of DMN ensembles enables hypotheses generation and suggests experimental tests.

    “Figure 4: Cluster analysis of DMN ensembles enables hypotheses generation and suggests experimental tests.”
    Quote in the paperNot a substantial claim
    • The authors state this as an assumption: Synapse strength is proportional to synapse count from the connectome times a scalar for each filter.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Interpretationown data

    Sparse networks have larger local minima.

    “Sparse networks have larger local minima.”
    • Also stated in: § Additional information · hedged
      “we believe that 1032 the size of the simulations' local minima increase”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.

    Premisecitation

    There are approximately 140 Lawf2 neurons per optic lobe, and each of approximately 700 columns is innervated by approximately 5 Lawf2 cells.

    “(there are approx.140 Lawf2 neurons per optic lobe, and each of approx.700 columns is innervated by approx.5 Lawf2 cells 124 )”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [124] source

      Not checked in the free run.

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Study design and conductnone given

    The authors manually proofread filters on the hexagonal lattice and compared them to reported filters in the literature to ensure overall correspondence.

    “We manually proofread filters on the hexagonal lattice and compared them to the reported filters in the literature to ensure overall correspondence.”
    Quote in the paperNot a substantial claim

  • Probably not a substantial claim: may not pass S3 (load-bearing).

    Premisecitation

    The ones filled in the convex hull are considered mostly false positives from the connectome reconstruction.

    “Although these are considered mostly false positives from the connectome reconstruction 12”
    Quote in the paperNot a substantial claim0 of 1 sources checked
    • cites [12] source

      Not checked in the free run.

64 of 64

  • The authors state this as an assumption.

    Neurons of the same cell type share the same neuron time constant and resting membrane potential.

    “We assumed that neurons of the same cell type shared the same neuron time constant and resting membrane potential.”

  • The authors state this as an assumption.

    The circuitry across the entire visual field is perfectly periodic (spatial homogeneity/translation invariance across retinotopic space).

    “We approximated the circuitry across the entire visual field as perfectly periodic 66,68 , and tiled this local connectivity architecture in hexagonal lattice across retinotopic space”
    “our assumption of spatial homogeneity (i.e., the hexagonally convolutional structure of the network)”

  • The authors state this as an assumption.

    The unitary synapse scaling and the sign of each synapse are the same across all pairs of neurons with the same pre- and postsynaptic cell type.

    “The unitary synapse scaling and the sign of each synapse was the same across all pairs of neurons with the same pre-and postsynaptic cell type.”

  • The authors state this as an assumption.

    The synapse count between each pair of neurons is the same across all pairs of neurons with the same pre- and postsynaptic cell type and their relative location in retinotopic space.

    “Likewise, the synapse count between each pair of neurons was the same across all pairs of neurons with the same pre-and postsynaptic cell type, and their relative location in retinotopic space.”

  • The authors state this as an assumption.

    Synapse strength is proportional to synapse count from the connectome times a scalar for each filter.

    “Instead of parametrizing each individual synapse strength, we assume that synapse strength is proportional to synapse count from the connectome times a scalar for each filter, reducing the number of parameters to Q + 2T = 734 while providing enough capacity for the DMNs to yield realistic tuning to solve the task.”

  • The authors state this as an assumption.

    A convolutional RNN with shared filters between cells of the same postsynaptic type, shared time constants, and resting potential.

    “Assuming a convolutional RNN with shared filters between cells of the same postsynaptic type, shared time constants, and resting potential, the amount of parameters reduces drastically to T 2 C + 2T = 3, 046, 355.”

  • The authors state this as an assumption.

    The same convolutional RNN but additionally convolutional filters are constrained to F = 5 visual columns.

    “Further assuming the same convolutional RNN but additionally convolutional filters are constrained to F = 5 visual columns, i.e. the number of presynaptic input columns in hexagonal lattice is P = 3F (F + 1) + 1, the amount of parameters reduces to T 2 P + 2T = 384, 605.”

  • The authors state this as an assumption.

    Only Q = 604 connections between cell types exist in our connectome.

    “Assuming as in our connectome only Q = 604 connections between cell types exist, this reduces the number of parameters further to QP + 2T = 55, 185.”

  • The authors state this as an assumption.

    For columnar, spatially repeated neurons, only one neuron per position is present.

    “For columnar, spatially repeated neurons, we can also 978 assume that only one neuron per position is present.”

  • The authors state this as an assumption.

    The individual units in the SCNs are interpreted as neurons with voltage.

    “We interpret the individual units in the SCN's as neurons with voltage”

  • The paper relies on this without support.

    A multi-compartment model with bidirectional electrical synapses often resulted in oscillatory dynamics in earlier modeling attempts.

    “Because a multi-compartment model with bidirectional electrical synapses resulted often in oscillatory dynamics in earlier modeling attempts”
11 of 11

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

  • An absence of information across columns has necessitated treating each column as identical in simulations of the optic lobe.

    “An absence of information across columns, has necessitated treating each column as identical in simulations of the optic lobe115.”

    From Neuronal wiring diagram of an adult brain (2023) · cites this paper as [115]

  • A neural network simulation of visual motion detection based on the wiring diagram of columnar circuits in the optic lobe has been created.

    “A neural network simulation of visual motion detection based on the wiring diagram of columnar circuits in the optic lobe has been created115.”

    From Neuronal wiring diagram of an adult brain (2023) · cites this paper as [115]

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

  • Flash-stimulus baseline duration differs

    Minor

    The methods specify one second of grey-screen stimulation before the flashes, whereas the Extended Data figure describes the flash-response measurements as following two seconds of grey input.

  • T4 and T5 preferred edge contrasts are reversed

    Minor

    The results say T4 responds to dark edges and T5 to bright edges, but the figure caption identifies T4 as the ON-motion detector and T5 as the OFF-motion detector.

2 of 2

Cites

Flagged first, then by citations

Research works printed but not linked: 16

  1. 10.1101/2020.12.08.413955v3 · DOI
  2. Montufar, G. F. , Pascanu, R. , Cho, K. & Bengio, Y. On the number of linear regions of deep neural networks. Advances in neural information processing systems 27 (2014).
  3. Goodfellow, I. , Bengio, Y. & Courville, A. Deep learning (MIT press, 2016).
  4. Kornblith, S. , Norouzi, M. , Lee, H. & Hinton, G. Similarity of neural network representations revisited. In International Conference on Machine Learning, 3519–3529 (PMLR, 2019).
  5. Kingma, D. & Ba, J. Adam: A method for stochastic optimization. 3rd International Conference on Learning Representations (ICRL) (2014).
  6. Sandbrink, K. J. et al. Task-driven hierarchical deep neural network models of the proprioceptive pathway. bioRxiv (2020).
  7. Maheswaranathan, N. , Williams, A. , Golub, M. , Ganguli, S. & Sussillo, D. Universality and individuality in neural dynamics across large populations of recurrent networks. In Wallach, H. et al. (eds.) Advances in Neural Information Processing Systems, vol. 32 (Curran Associates, Inc., 2019). URL https://proceedings.neurips.cc/paper/2019/file/5f5d472067f77b5c88f69f1bcfda1e08-Paper.pdf.
  8. 10.1038/nn.4050 · DOI
  9. 2001, Approximation with artificial neural networks, Faculty of Sciences, Etvs Lornd University, Hungary, 24, 7
  10. Li, Y. , Yosinski, J. , Clune, J. , Lipson, H. & Hopcroft, J. E. Convergent learning: Do different neural networks learn the same representations? In International Conference on Learning Representations (ICLR) (2016).
  11. Raghu, M. , Gilmer, J. , Yosinski, J. & Sohl-Dickstein, J. Svcca: Singular vector canonical correlation analysis for deep learning dynamics and interpretability. Advances in neural information processing systems 30 (2017).
  12. 10.1016/j.cub.2015.07.014 · DOI · Paper page
  13. Mi, L. et al. Connectome-constrained latent variable model of whole-brain neural activity. In International Conference on Learning Representations (2022). URL https://openreview.net/forum?id=CJzi3dRlJE-.
  14. Chen, R. T. Q. , Rubanova, Y. , Bettencourt, J. & Duvenaud, D. K. Neural ordinary differential equa-tions. In Bengio, S. et al. (eds.) Advances in Neural Information Processing Systems, vol. 31 (Curran Associates, Inc., 2018). URL https://proceedings.neurips.cc/paper/2018/file/69386f6bb1dfed68692a24c8686939b9-Paper.pdf.
  15. Krizhevsky, A. , Sutskever, I. & Hinton, G. E. Imagenet classification with deep convolutional neural networks. In Advances in neural information processing systems, 1097–1105 (2012).
  16. Butler, D. J. , Wulff, J. , Stanley, G. B. & Black, M. J. A naturalistic open source movie for optical flow evaluation. In European conference on computer vision, 611–625 (Springer, 2012). · DOI · Paper page

Cited by

Sorted by citations

Related

Authors
Lappalainen, J. K., Tschopp, F. D., Prakhya, S., McGill, M., Nern, A., Shinomiya, K., Takemura, S.-Y., Gruntman, E., Macke, J. H., & Turaga, S. C.
Affiliations
  • Howard Hughes Medical Institute (US)
  • California Institute of Technology (US)
  • Max Planck Institute for Intelligent Systems (DE)
Corresponding author
Turaga, Srinivas C.
ORCID
Venue
bioRxiv (Cold Spring Harbor Laboratory)
Publisher
openRxiv
Published
2023-03-13
DOI
10.1101/2023.03.11.532232
OpenAlex
W6940891779
Semantic Scholar
e60c7fc662981feee2a3ebbc4a9dca4d700cadb1
Cited by
31
References
117
Citations by year
2026: 2 · 2025: 6 · 2024: 12 · 2023: 8 · 2022: 1 · 2021: 1 · 2020: 1
Type
posted content
Language
en
Open access
Licence
cc-by
Locations
  • bioRxiv (Cold Spring Harbor Laboratory) (acceptedVersion, cc-by)
  • German Neuroinformatics Node
Funders
  • Howard Hughes Medical Institute
  • International Max Planck Research School for Environmental, Cellular and Molecular Microbiology
  • Deutsche Forschungsgemeinschaft (206/4-1, EXC number 2064/1, 390727645)
  • Bundesministerium für Bildung und Forschung (Tübingen AI Center, FKZ: 01IS18039A, 01IS18039A, FKZ: 01IS18039A)
  • International Max Planck Research School for Advanced Methods in Process and Systems Engineering
Topics
  • Life Sciences › Neuroscience › Cognitive Neuroscience › Neural dynamics and brain function
  • Life Sciences › Neuroscience › Cognitive Neuroscience › Visual perception and processing mechanisms
  • Life Sciences › Neuroscience › Cellular and Molecular Neuroscience › Neurobiology and Insect Physiology Research
Keywords
connectome, neural response prediction, Drosophila visual system, single-neuron resolution

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Claims read by deepseek-v4.1-flash under definition version 3 (extract-v3) on 2026-10-07.

Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-07.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review