KurisutinaResearch date: 29 September 2026.
Citation and status. Janne K. Lappalainen, Fabian D. Tschopp, Sridhama Prakhya, Mason McGill, Aljoscha Nern, Kazunori Shinomiya, Shin-ya Takemura, Eyal Gruntman, Jakob H. Macke, and Srinivas C. Turaga. Connectome-constrained networks predict neural activity across the fly visual system. Nature 634, 1132–1140. Published 11 September 2024; issue 31 October 2024. Peer-reviewed primary computational study, version of record, DOI 10.1038/s41586-024-07939-3, CC BY 4.0 subject to credited third-party exceptions.
Complete reading record. Read the entire publisher article: abstract, main narrative, Methods, equations, all five main-figure captions, all nine Extended Data captions, bibliography, and availability/declaration material. Read all seventeen pages of the official supplement, including Notes 1–5, Figures 1–9, Tables 1–2, equations, and references. Read the two-page reporting summary, using OCR checked against both page images. Visually inspected all five main figures, all nine Extended Data figures, all nine supplementary figures, and both supplementary tables. The main PDF has 26 pages, including the reporting summary as its final two pages; the separate reporting PDF is also retained. Reading is complete for these scientific texts. No figures were digitized and no simulation or statistical analysis was rerun. The supplementary ZIP of data and the peer-review correspondence were identified but not obtained or read; consequently the complete catalogue of predictions for every cell type and every source-data entry was not audited. Cited original studies were not opened as additional papers.
Local main PDF, publisher HTML, readable paragraphs, equations, supplement, reporting summary, and provenance with hashes and exact scope.
Factual core. A partial consensus fly visual-system connectome was tiled across retinotopic space, yielding 45,669 modeled neurons, 1,513,231 directed connections, and 64 cell types. Shared physiology and synapse-count constraints reduced the mechanistic network to 734 adjustable parameters. These were optimized for optic-flow estimation, without fitting neural-activity recordings. Among fifty models, the ten with best task performance had an ensemble-median contrast preference matching all 32 characterized cell-type/compartment entries; the best single model matched 30. Models also reproduced major T4/T5 direction-selective properties, while alternative parameter solutions and some incorrect responses remained. Ablations showed different contributions of synaptic signs, counts, spatial connectivity, and task training. In separate synthetic feedforward digit-classification networks, median corresponding-neuron response correlation was .85 at 10% connectivity versus .38 at 80% when only connectivity and signs were supplied. Adding noisy strength estimates yielded correlations above .9 across tested connectivities. No new animal recordings were collected. The work establishes useful task-conditioned predictions and mechanistic hypotheses, rather than reconstruction of a particular animal’s full dynamics or consciousness. Main results, Figures 1–5.
What information actually enters the model.
The connectome is assembled from separate local reconstructions and earlier anatomical work, rather than being an unmodified whole-brain wiring diagram of one fly. An expectation-maximization procedure estimates retinotopic positions from synapse statistics, synapse locations, and supplied annotations. The circuit is then treated as periodic across 721 columns. Lamina components are manually constructed from prior data. Supplementary Note 2 describes replacing an insufficiently asymmetric Tm9-to-T5 filter with a Gaussian at the previously reported location, adjusting some input counts to prior anatomical measurements, and filling filter convex hulls, which introduces weak autapses. These are disclosed anatomical modeling choices; they matter when specifying what a literal raw-connectome-only reproduction would require.
The model additionally receives cell-type labels, transmitter/receptor information, and biological assumptions. Most connection signs follow transmitter expression; some transmitter phenotypes were guessed where measurements were unavailable. CT1 is represented as separate medulla/lobula compartments with locally repeated units, reflecting prior evidence about compartmentalization. A cell type shares a membrane time constant and resting potential; a presynaptic/postsynaptic type pair shares one nonnegative strength multiplier. Synapse count supplies relative strength within that pair. The 734 parameters comprise 65 time constants, 65 resting potentials, and 604 strength multipliers. There are 65 modeled entries because CT1 is counted twice, despite 64 biological cell types.
In condensed notation the dynamics are tau[type(i)] dV_i/dt = -V_i + sum_j w_ij f(V_j) + Vrest[type(i)] + input_i, with w_ij = alpha[type(i),type(j)] × sign[type(i),type(j)] × synapse_count[type(i),type(j),offset]. The nonlinearity describes graded transmitter release in a non-spiking approximation. Thus topology, synaptic efficacy, cellular dynamics, input, and current state remain distinct ingredients even in this deliberately simple system.
The task is an additional powerful constraint: predict optic flow from greyscale Sintel film sequences. A separate two-layer decoder reads instantaneous responses of 34 types; its lack of temporal memory makes the recurrent visual network supply the motion computation. Extended Data Figure 2 lists 7,427 decoder parameters in addition to the 734 mechanistic parameters. Accordingly, “734 parameters” describes the visual-system model, not every trained parameter in the end-to-end pipeline. Training and validation avoid putting spatial subsamples of the same scene on both sides. The physiological reference literature is used for comparison, but this is not a prospectively blinded experimental prediction exercise.
What the results distinguish.
| Result or analysis | Supported inference | Boundary |
|---|---|---|
| Accurate contrast preferences, including in randomly parameterized models with signs supplied | Signed anatomy is already strongly informative about this response feature | Contrast preference alone is not recovery of the full voltage trajectory or the motion mechanism |
| Better motion predictions after task training and with detailed anatomy | Structure and task jointly constrain some individual-cell functions | They do not uniquely identify all parameters or dynamics |
| T4c model clusters with upward, downward, or absent tuning | Different mechanistic solutions remain despite the same broad constraints | Ensemble agreement is not a calibrated probability of biological truth |
| Preferred-direction/contrast differences propagate to upstream cells | Sparse activity measurements could eliminate whole families of alternatives | This is partly demonstrated by conditioning on already known responses, rather than by new recordings |
| Synthetic sparsity experiment | In the tested feedforward model family, fewer unknown connections make matching internal responses easier | This is not a theorem covering recurrent mammalian brains, plasticity, or consciousness |
Several quantitative details prevent overstatement. The reported relationship between task error and direction-selectivity agreement is r = −.60, p = 2.6 × 10⁻⁶ across fifty models (Extended Data Figure 2); the main narrative gives the corresponding positive association with task performance. No comparable significant task-error association is reported for contrast preference. The T4c clusters have average task errors 5.297, 5.316, and 5.357, despite qualitatively different tuning. A small improvement on the chosen task can favor a more realistic solution without making task performance an identity certificate.
The main text’s 32 characterized contrast entries include CT1’s two compartments. Other Methods passages use 31 cell types. Preserve that distinction rather than claiming 32 independent animals or 32 independent experiments. The fifty simulations are computational realizations, not biological replicates. The Methods state that the main fifty models start from the same parameter values; stochastic training still yields different solutions. The reporting form’s generic description of random initialization should not be substituted for that more specific statement.
Validation and selection audit.
The strongest headline comparison uses the ten models selected by task performance, with recordings excluded from the training objective. Some further illustrations have a different evidential status. The explicit “Model selection” Methods subsection selects models for Extended Data Figures 4, 7, and 8 whose target and input contrast preferences, direction-selectivity threshold, and preferred direction match empirical knowledge; it assumes a 225-degree preference for the candidate TmY3 analysis. Those illustrations are mechanistic hypotheses conditional on response agreement. Their agreement with the selection features cannot count as an independent validation of those same features.
For the moving-edge correlation benchmark, the paper takes the maximum correlation across six simulated speeds against a single experimentally measured tuning curve. This tolerates different preferred speeds, but means the score is not accuracy at a prespecified matched speed. Tm4 is incorrectly predicted to depolarize to the single-ommatidium light increment in the illustrated temporal-response analysis. Supplementary Figure 7 shows that the trained decoder relies primarily on ON-selective T4 types. These observations identify the scope of success without negating it.
The model excludes several mechanisms the authors explicitly discuss: electrical synapses, richer chemical-synapse nonlinearities, neuromodulation, plasticity, and many non-motion computations. A successful prediction within the simplified regime is evidence that all omitted detail is unnecessary for that particular prediction; it does not show that omitted detail is unnecessary for every behavior or state.
Interpretation for consciousness, connectomics, and person reconstruction — our reasoning.
This paper is strong evidence against dismissing connectomes as functionally uninformative. Combining structural constraints with a defensible task and simple dynamics can predict much more than an aggregate output: particular identified cell types acquire recognizable response properties. This provides a tractable middle step between mapping anatomy and reconstructing computation.
It is also a concrete example of why “the connectome is sufficient” needs a specified meaning. Sufficiency relative to a constrained model class, supplied task, cell-type priors, and a selected sensory assay is different from sufficiency to infer an individual’s memories, current state, learning rules, or conscious experience. Nothing here measures phenomenal consciousness or tests a theory-specific consciousness criterion. Anatomical correspondence and useful neural prediction cannot by themselves adjudicate that question.
The orientation literature sharpens the distinction. The existing Nabavi summary separates changing pathway efficacy from identifying the entire memory store; Roy distinguishes distributed causal participation from locally recoverable content. Rose and Wolff caution against treating failure of one readout as absence of retained state. In this paper, a fixed anatomical model also supports more than one dynamical solution. These are compatible observations at different scales, not a proof that one biological substrate must contain every relevant variable.
Proposed next discrimination. Freeze the architecture, priors, task, preprocessing, ensemble selection, and primary metrics before obtaining new recordings. Compare predictions for previously uncharacterized cells, novel stimuli, and causal perturbations. Report all ensemble members alongside task-ranked predictions, and test whether additional recordings eliminate alternatives prospectively. A second axis should test behavioral-state shifts with fixed anatomy, asking whether measured neuromodulatory or dynamical variables explain the missing changes. This would estimate the extra information needed beyond a connectome for a defined function. A claim about consciousness would require an independently justified consciousness endpoint in addition to those physiological tests. No such experiments were run here.