Source-linked AI summary
Connectome-Based Modelling Reveals Orientation Maps in the Drosophila Optic Lobe
Jia-Nuo Liew, Shenghan Lin, Bowen Chen, Wei Zhang, Xiaowei Zhu, Wei Zhang, Xiaolin Hu
TL;DR
The paper examines whether orientation maps can emerge in Drosophila despite its non-layered architecture and limited cortical-like circuitry. Using connectome-constrained spiking simulations, it finds spatially organised orientation tuning, including coherent maps in the medulla. The authors conclude that similar visual structures may arise from shared computational principles across species.
Problem
Whether Drosophila can support coherent orientation maps without cortical lamination or large-scale recurrent loops remains unknown.
Method
The study drives a connectome-constrained Drosophila visual-system model with oriented stimuli and analyses downstream spiking responses.
Results
The simulations reveal spatially organised orientation tuning, coherent medulla maps, topological singularities, and inter-layer columnar alignment.
Takeaways & Limitations
The findings suggest orientation maps can arise from shared computational principles across species even without cortical lamination.
Takeaways & Limitations
The study lacks in vivo validation and uses simplified leaky integrate-and-fire dynamics that omit nonlinear firing and neurotransmitter effects.
Abstract
from arXiv · showhide
The ability to extract oriented edges from visual input is a core computation across animal vision systems. Orientation maps, long associated with the layered architecture of the mammalian visual cortex, systematically organise neurons by their preferred edge orientation. Despite lacking cortical structures, the Drosophila melanogaster brain contains feature-selective neurons and exhibits complex visual detection capacity, raising the question of whether map-like vision representations can emerge without cortical infrastructure. We integrate a complete fruit fly brain connectome with biologically grounded spiking neuron models to simulate neuroprocessing in the fly visual system. By driving the network with oriented stimuli and analysing downstream responses, we show that coherent orientation maps can emerge from purely connectome-constrained dynamics. These results suggest that species of independent origin could evolve similar visual structures.
1 Introduction
The paper asks whether coherent orientation maps can emerge in Drosophila despite its compact, non-layered brain and limited cortical-like circuitry. It addresses this question with connectome-constrained simulations and reports spatially organised orientation tuning in the optic lobe.
- Motivation: Orientation selectivity is a core visual computation, classically organised into columnar and laminar maps in mammalian visual cortex.These responses are linked to spatially organised feedforward inputs and recurrent cortical dynamics.
- Motivation: Drosophila neurons, including T4 and T5, exhibit orientation tuning, but whether such tuning forms coherent maps across the visual system remains unknown.Earlier work also hypothesised orientation selectivity in medulla neurons from synaptic connectivity patterns alone.
- Approach: The study simulates L1-L3 responses to bar-like stimuli and propagates activity through connectome-constrained leaky integrate-and-fire dynamics.The resulting population activity is analysed for spatially organised orientation tuning.
- Contributions: The authors computationally demonstrate spatially coherent orientation maps in the medulla of an invertebrate visual system.They identify this as the first such computational demonstration in the medulla of an invertebrate visual system.
- Contributions: The study identifies topological singularities and inter-layer columnar alignment in orientation preference across distal and proximal medulla regions.These structures are reported in the Dm and Pm regions.
- Implications: The findings suggest canonical orientation maps can arise from shared computational principles across species without cortical lamination.This frames orientation maps as potentially convergent computational motifs rather than an exclusively neocortical feature.
3 Methods
The study combines connectome-constrained LIF simulations, structured visual stimuli, circular-Gaussian tuning analysis, and spatial analyses to examine orientation selectivity and organisation in the Drosophila optic lobe.
- Network model: The network uses the complete adult fly connectome, comprising 138,639 neurons and 1,508,983 synapses, with LIF dynamics constrained by anatomical synapses.The model focuses on known connectivity while simulating spiking activity in the visual system.
- Stimuli: Poisson spike trains representing bar-like or grating stimuli were applied to retinotopically organised L1-L3 neurons arranged on the compound eye’s hexagonal ommatidial lattice.L1-L3 were stimulated directly rather than their upstream photoreceptors, and receptive-field overlap produced responses across neighbouring units.
- Orientation tuning: Neuron orientation preferences were estimated by fitting circular Gaussian functions to firing rates across stimulus orientations, requiring R2 ≥0.7 and RSSnorm ≤0.4 for a good fit.The circular model accounts for the 180° periodicity of orientation space.
- Structural prediction: Structural prediction estimated preferred orientation from direct upstream columnar neurons by mapping their compound-eye columns and fitting an ellipse to their spatial locations.Predicted orientations were compared with simulation-derived preferences using absolute differences.
- Structural prediction: Across analysed neurons, the mean absolute difference between simulated and structurally predicted preferred orientations was 13.7°.This comparison tested whether dendritic input geometry corresponded to orientation preference.
- Spatial organisation: Local circular-mean smoothing within a radius of approximately 5 × 10^3 nm was used to visualise mesoscale orientation maps and identify pinwheel-like singularities.Spatial position angle α was compared with preferred orientation β through Δ= |α −β|, distinguishing tangential from radial alignment.
- Validation analyses: Ablation and connectivity analyses tested network contributions to tuning, while within-column circular standard deviation was compared with a randomised baseline to assess orientation-column coherence.Silencing Mi neurons abolished tuning across Dm and Pm, whereas Tm silencing selectively affected Pm; excitatory connections preferentially linked similarly tuned neurons.
5 Conclusion and Discussion
The study finds spatially organised orientation selectivity and a rudimentary orientation map in Drosophila, with mechanisms involving recurrent connectivity and lateral inhibition. These findings support structured visual maps arising without mammalian cortical architecture, while remaining predictions requiring experimental validation.
- Drosophila exhibits spatially organised orientation selectivity and a rudimentary orientation map, although the maps are less discrete than mammalian examples.The authors relate this reduced discreteness to lower neuronal density.
- Structured orientation preferences in Drosophila resemble patterns reported in non-mammalian species and may reflect convergent evolution driven by shared computational demands.The paper identifies Drosophila as the only invertebrate in which such spatial structure has been computationally demonstrated.
- Neurons with similar orientation preferences are more likely to form recurrent connections, consistent with roles for dendritic integration, closed-loop motifs, and lateral inhibition.
- The orientation columns are spatially adjacent and arranged in a primitive pinwheel-like map, supporting efficient responses to lines and edges.
- The findings lack in vivo validation, and the simplified LIF model omits nonlinear firing dynamics and neurotransmitter effects.The authors call for experimental follow-up to confirm the predictions.
NeurIPS Paper Checklist
The checklist assessment states that the abstract and introduction accurately present the paper’s question and contributions. It also judges the reported results to support the stated scope.
- The abstract and introduction clearly state that the study tests whether structured orientation maps can emerge in Drosophila from biological connectivity.
- The connectome-driven simulations and analyses are judged to support the claims made in the abstract and introduction.
- The checklist emphasizes that claims should match theoretical and experimental results and reflect how broadly the findings can generalize.
2. Limitations
The checklist materials state that the paper discusses limitations and is an experimental, simulation-driven study rather than a theoretical work with formal proofs. They also state that experimental details are disclosed for reproducibility.
- The paper includes a dedicated Limitations section acknowledging simulation-based findings without in vivo validation and discussing modelling simplifications.
- The checklist recommends stating strong assumptions and explaining robustness to violations when theoretical results are presented.
- The study provides no formal theoretical results, theorems, or proofs because its contribution is experimental and simulation-driven.
- The paper is judged to disclose model parameters, connectome sources, stimulus design, simulation details, and analysis equations needed to reproduce its main results.
5. Open access to data and code
The checklist states that the paper provides open access to the simulation and analysis code with instructions for reproducing the main experimental results. It also reports that necessary experimental details are specified.
- The paper answers Yes to providing open access to data and code with sufficient reproduction instructions.
- Full simulation and analysis code is appended through a GitHub link in Appendix A, along with instructions for reproducing the main results.
- The paper specifies LIF parameters, input stimulus construction, simulation setup, and the neuron types studied.
- The checklist states that experimental details may be provided in the paper, code, appendix, or supplemental material at the level needed to interpret results.
7. Experiment statistical significance
The paper reports statistical robustness through population-level visual summaries and shuffled controls, while noting that traditional error bars are not uniformly presented. It also describes the computational resources used for reproducibility.
- Orientation-tuning distributions are summarized with histograms, and shuffled control data are used to assess baseline significance.
- Traditional error bars are not uniformly presented, but population-wide distributions and randomized baselines convey statistical reliability for the reported claims.
- The submission states that its research conforms to the NeurIPS Code of Ethics and uses non-invasive simulations with publicly available biological data.
10. Broader impacts
The paper describes potential positive societal impacts from understanding minimal biological circuits, while reporting no high-risk released assets or new datasets and models. The supplied discussion does not provide a concrete negative societal-impact analysis.
- The paper suggests that minimalistic biological circuits could inspire lightweight, sustainable AI systems.
- The supplied broader-impact materials emphasize considering possible harms, including malicious use, fairness, privacy, and security concerns.
- The work does not release pretrained models, datasets, or tools identified as posing misuse or dual-use risks.
- The paper uses and credits existing publicly available connectome and Drosophila datasets rather than introducing new datasets or models.
14. Crowdsourcing and research with human subjects
The study uses simulated neural models and biological datasets rather than crowdsourcing, surveys, or human participants. Accordingly, participant instructions, compensation, risks, and IRB approval are described as not applicable.
- Participant instructions, screenshots, and compensation details are therefore not applicable to this study.
- The study does not involve crowdsourcing, surveys, or human-subject research because all experiments use simulated neural models based on biological datasets.
- Risks to participants and Institutional Review Board approval are not applicable because the research contains no human subjects or participant-based studies.
- The core research methods do not involve large language models as important, original, or non-standard components.
A Code and data availability
The paper provides code and data availability information, describes the conductance-based LIF model and its synaptic parameters, and reports parameter-sensitivity results. It also includes supplemental validation of LIF parameter perturbations.
- Full code and data are available online, and simulations used an Intel CPU with 20 parallel threads and approximately 72 hours per angle sweep.
- The membrane-potential equations define the conductance-based LIF dynamics used to simulate the visual network.
- Spikes occur when membrane potential reaches threshold, followed by reset and a refractory period, with synaptic transmission delayed by fixed latency.
- The synaptic weight wsyn scales excitatory and inhibitory postsynaptic potentials according to anatomical connectivity between upstream and downstream neurons.
- At 10% parameter variation, preferred orientations changed minimally, whereas 50% variation noticeably degraded orientation selectivity without causing ODE divergence.
B.2 Stimulus designs and validation
The model used biologically informed bar stimuli to drive L1–L3 activity and validated that lamina and downstream neurons responded consistently with OFF-bar and ON-cell visual roles.
- L3 showed the strongest response variation to light intensity, followed by L1, while L2 showed the weakest modulation.The model used this sensitivity ordering to set layer-specific gain parameters.
- The firing-rate model treated distance from the OFF-bar as a proxy for local light-intensity variation and used layer-specific gains and baselines.The maximum firing rate was set to 200Hz, with fitted parameters A1 = 7, A2 = 5, A3 = 10 and Bx = 20Hz.
- Sensitivity analysis varied the gain parameters A1, A2 and A3 by 10% to assess robustness of the simulated responses.The corresponding orientation-map changes were examined in Figure S2c–e.
- OFF-responsive downstream neurons peaked when the bar aligned with their receptive fields and weakened with distance, whereas ON-cells showed stronger inhibition nearby.The tested downstream types were Tm1, Tm2, Tm4, Mi1, Mi4 and Tm3.
- The downstream activation patterns mirrored OFF-bar locations on the compound eye, consistent with the model’s visual topology.Tm1, Tm2 and Tm4 were evaluated as OFF-cells, while Mi1, Mi4 and Tm3 were evaluated as ON-cells.
B.3 Circular Gaussian function
Orientation preferences were fitted with a circular Gaussian, while circular standard deviation quantified how tightly preferred orientations clustered within columns.
- The circular Gaussian fit predicts each neuron’s firing rate across orientations using preferred orientation, tuning width, amplitude, baseline, and circular distance.The input orientation ranges from 0° to below 180°, and circular distance uses the smaller separation around the orientation domain.
- Circular standard deviation was used to quantify the sharpness of orientation tuning within each column.The measure was computed with the scipy Python library.
- The mean resultant length R was computed from the summed cosine and sine components of neurons’ preferred orientations.The calculation uses each neuron’s preferred orientation αi and the number of neurons n in the column.
- A circular standard deviation of s = 0 indicates perfect alignment, whereas larger values indicate broader orientation distributions.Lower values therefore represent stronger orientation coherence within a column.
C Targeted ablation analysis
Targeted silencing and connectivity analyses identified pathway-specific contributions to orientation selectivity and revealed orientation-dependent network structure across optic-lobe neurons.
- Silencing Tm neurons reduced well-fit orientation-selective neurons by 89.6% in Pm but by only 9.3% in Dm.The ablation compared the number of good fits remaining with the original result.
- Silencing Mi neurons eliminated tuning across both layers, leaving fewer than 5% of well-fit neurons in Dm or Pm.The analysis therefore identified Mi neurons as essential upstream sources of orientation selectivity throughout the optic lobe.
- Excitatory synapses clustered near the diagonal of pre- versus postsynaptic preferred orientations, indicating preferential connectivity between similarly tuned neurons.The corresponding inhibitory network showed an off-diagonal structure consistent with cross-orientation suppression.
- Gaussian fitting quantified orientation selectivity, while supplementary figures documented T4/T5 tuning profiles and orientation-preference structure across neuron types and layers.These analyses supported evaluation of model fit quality and map organization beyond the main results.