{"id":"65bcb3e3-dce8-4c98-aa9b-bb576cff4077","arxiv_id":"2505.13011","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A VAE with SHAP-based interpretability learns low-dimensional codes for Drosophila connectome subgraphs, and the codes can be steered to generate graph subgraphs with targeted edge counts and other properties.","lead":"This paper trains a variational autoencoder on 500 subgraphs sampled from the fruit fly connectome to learn a compact code for each neural circuit. The authors show that a few dimensions of this code track structural features like edge count, and that adjusting the code can generate graphs with desired properties.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"SHAP-guided DP steering is the load-bearing route for 'predefined properties,' yet Appendix H admits it fails for three of four target statistics; the abstract's controlled-generation claim is therefore broader than the evidence.","rationale":"The strongest claim is about controlled generation of subgraphs with predefined properties. The paper's white-box route for this is SHAP + DP, and this route is the only one that directly connects 'interpretable latent variables' to 'predefined properties'; CMA-ES is a black-box optimizer over the full latent space and does not by itself establish interpretability. Appendix H concedes that DP fails for three of the four target statistics because of coupling between dimensions. This is not just poor performance: it invalidates the additive assumption built into Algorithm 1, which treats SHAP values as fixed per-dimension knapsack weights. SHAP values satisfy local efficiency, not global additivity, and no test of the additivity assumption is provided. The single successful edge-count result could be an artifact of a nearly linear relationship rather than evidence for the general method. The RQ3 CMA-ES experiment has no baselines or error bars, so it cannot independently carry the general claim without risking a trivial or overfit result. The sampling concern raised by the reader is real but secondary: even if the cylindrical samples are not perfect columnar units, the steering claim concerns the learned generative model, whereas sampling validity mainly affects the biological inferences in the title. The DP/SHAP concern directly attacks the operational meaning of 'predefined properties.' The VAE reconstruction results and surrogate correlations are useful internal validations, but they do not test the steering mechanism. A conditional verdict is appropriate: the core demonstration may be salvageable with a narrowed claim and a validation of the additive assumption, but as written the headline is broader than the evidence supports.","tokens_in":14283,"tokens_out":8729,"duration_ms":101715,"concrete_test":"Reproduce Algorithm 1 for edge-count, reciprocity, betweenness, and non-neuronal on 500 held-out decoded graphs. For each target, compute SHAP values over a grid of k and test the additivity assumption y(z) - y_ref = sum_i shap_i(z_i) on 1000 random latent vectors; report residual distributions and the fraction of target values for which DP's traced-back z achieves the target within tolerance. If three of four features fail, as Appendix H suggests, the abstract's 'predefined properties' claim must be narrowed. Separately, run CMA-ES with the same budget but on randomly permuted latent coordinates; if AUC/accuracy matches the reported values, RQ3 does not demonstrate interpretable steering.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that latent codes can be manipulated to controllably generate connectome subgraphs with predefined properties. The paper's interpretable, white-box steering mechanism is SHAP-guided dynamic programming (Section 3.5, Algorithm 1). Appendix H states explicitly that DP 'yields suboptimal results' for reciprocity, betweenness, and non-neuronal, attributing this to 'strong coupling between dimensions' that induces 'excessive variance in SHAP values.' This is not a minor performance gap: Algorithm 1 assumes each latent dimension contributes additively to the target statistic through a fixed SHAP value shap[i][k] independent of other coordinates (line 13: dp[i][j + shap[i][k]] <- k). SHAP values are local, interaction-dependent attributions; the efficiency property (Eq. 24) holds only at a single input point and does not license a global additive decomposition. The paper provides no validation of this additivity assumption, and its own Appendix H shows the assumption fails outside edge-count. Consequently, the demonstration of 'controllable generation with predefined properties' rests on one successful edge-count curve plus a black-box CMA-ES experiment that lacks baselines, error bars, and comparison to non-interpretable latent optimization. The abstract's broad claim is therefore not supported by the evidence presented.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript proposes an end-to-end framework, applied to the FlyWire Drosophila connectome, for learning interpretable low-dimensional representations of neural-circuit subgraphs and using those codes for controlled graph generation. The pipeline consists of (i) adaptive cylindrical sampling of 81-100-neuron subgraphs along the visual-stream axis, padded to 100 nodes; (ii) a graph VAE with a GAT-based edge encoder, a latent bottleneck, and a staged '1+2n' training protocol; (iii) a differentiable surrogate per graph statistic (edge count, reciprocity, betweenness, non-neuronal count) that enables SHAP attribution of latent dimensions; and (iv) two steering mechanisms, a SHAP-guided dynamic program (Algorithm 1) and CMA-ES. The experiments report competitive graph-generation MMD versus EDGE, DisCo, GDSS, and GruM; surrogate correlations between 0.92 and 0.998; a monotonic DP steering curve for edge count; and AUC of 0.919/0.852 in RQ3 when generating a target graph from full versus statistical information. Appendix H concedes that DP 'yields suboptimal results' for the other three statistics, attributing this to coupling between dimensions.","tokens_in":14515,"tokens_out":15402,"duration_ms":142033,"significance":"If the results hold, the framework is a novel and potentially useful combination of connectomics subgraph sampling, information-bottleneck representation learning, and SHAP-based white-box steering, and several parts of the manuscript are genuine strengths: the surrogate is validated against the ground-truth statistics of decoded graphs rather than only against its own training target, the experiments use the real FlyWire dataset, and the authors candidly report where DP fails. That candor, however, reveals that the headline claim--'controllably generate connectome subgraphs with predefined properties'--is demonstrated for one of four statistics via DP, while the complementary CMA-ES experiments lack baselines and variance estimates. The contribution to connectomics is also conditional on the unvalidated assumption that the cylinder-sampled subgraphs correspond to functionally meaningful columnar units. With a narrowed central claim, added statistical support, and validation of the steering assumptions, this could be a solid contribution to interpretable graph representation learning; in its current form the abstract overstates the demonstrated capability.","major_comments":[{"comment":"The manuscript's central load-bearing claim, stated in the abstract and conclusion, is that latent codes can be manipulated to 'controllably generate connectome subgraphs with predefined properties.' Appendix H reports that for reciprocity, betweenness, and non-neuronal count, dynamic programming 'yields suboptimal results,' so the DP-based controlled generation is demonstrated only for edge count (Figure 3c). The CMA-ES experiments in Sec. 4.4 report AUC of 0.919 and 0.852, but with no baselines, no comparison against optimizing latent vectors in a non-interpretable space of equal dimension, and no per-seed variability; they therefore cannot carry the full weight of the 'predefined properties' claim. The claims should be narrowed accordingly, or the missing baselines and variance information should be supplied.","section":"Abstract; Sec. 4.3; Appendix H"},{"comment":"Algorithm 1 assumes a global additive decomposition of each target statistic in terms of per-dimension SHAP values, expressed by the transition dp[i][j + shap[i][k]] <- k. SHAP values are local attributions: the efficiency property in Eq. (24) holds at a single input point and does not license treating shap[i][k] as a coordinate-independent contribution of setting z_i = k. The paper never tests this additivity assumption, and Appendix H's own diagnosis--that coupling between dimensions induces excessive variance in SHAP values--is direct internal evidence that the assumption fails for three of four statistics. The authors should either establish conditions under which Algorithm 1 is valid, validate additivity for the statistics it is applied to, or replace the DP with a steering method that accounts for interactions.","section":"Sec. 3.5, Algorithm 1; Appendix E"},{"comment":"No reported metric carries an error bar, confidence interval, or significance test. Table 1 supports comparative statements ('matching' GruM on degree MMD, 'particularly strong performance in clustering coefficient'), yet each cell appears to be a single value, and the baseline procedure of 'selecting the best-performing configuration' from 10 random hyperparameter samples is not an acceptable substitute for a repeated-seed protocol with reported variability. The surrogate correlations in Table 2 motivate the entire SHAP analysis; for reciprocity (0.9192) and betweenness (0.9622) the unexplained variance is substantial enough to matter for the downstream steering failure, so confidence intervals should be reported here as well.","section":"Sec. 4.1; Tables 1-2; Sec. 4.4"},{"comment":"The biological claims of the paper depend on the assertion that cylindrical sampling 'captures columnar functional units characteristic of neural circuitry,' but this is not validated: there is no comparison with anatomically defined columnar units, no sensitivity analysis for the cylinder radius or the 81-100 neuron window, and the centers are drawn randomly in the XZ-plane. Relatedly, the 'non-neuronal' count used as a target statistic is, by the paper's own definition in Eq. (21), the number of padding nodes introduced to standardize variable-size subgraphs to 100 nodes; it is an artifact of the sampling protocol rather than a structural property of the circuit. This should be acknowledged when the paper claims that latent dimensions 'govern important structural properties' of neural circuits.","section":"Sec. 3.1; Appendix D, Eq. (21)"},{"comment":"The RQ2 analysis is internally inconsistent about its principal result: the text identifies six influential dimensions (1, 9, 11, 13, 24, and 28) and two sentences later refers to 'these 8 key dimensions.' Moreover, the paper provides no quantitative criterion for declaring a dimension 'key' (no |SHAP| threshold, no permutation or ablation test), so the claim that a few dimensions govern structural properties is not precisely defined. The number, identity, and selection rule for key dimensions should be stated consistently and justified.","section":"Sec. 4.3"}],"minor_comments":[{"comment":"The reference list contains exact-duplicate entries: [7] and [8] (Chen et al., 2023), [18] and [19] (Hou et al., 2024), [30] and [31] (Loshchilov and Hutter, 2016), and [50] and [51] (Xu et al., 2024); these should be merged.","section":"References"},{"comment":"The reciprocity definition is internally inconsistent: the left-hand side sum_{i<j} A_ij A_ji / sum_{i!=j} A_ij (1 - A_ji) equals (number of mutual pairs)/(total edges - 2*(mutual pairs)), whereas the right-hand side tr(A^2)/(||A||_0 - tr(A^2)) equals 2*(mutual pairs)/(total edges - 2*(mutual pairs)); the numerator differs by a factor of two, and the 1000x scaling mentioned in the appendix is not defined in the main text.","section":"Appendix D, Eq. (18)"},{"comment":"The text 'Figure 3 demonstrates our model's graph reconstruction capability' is a mis-reference: Figure 3 shows SHAP and DP results for edge count, while the reconstruction panels are Figure 9.","section":"Appendix I"},{"comment":"Several quantities needed to reproduce the experiments are missing: the latent dimensionality d used in the main runs (Appendix G sweeps 4-32), the binarization threshold kappa, the DP bounds l and r, the discretization of the target statistic y, and the distribution whose standard deviation sigma is used to set the search range in Algorithm 1; the phrase 'possible eigen value bounds (l <= lambda <= r)' does not match the latent variable notation used elsewhere.","section":"Sec. 4.1; Algorithm 1"},{"comment":"The opening equality of the VAE derivation, log p(x;theta) = integral of q(z|x) log p(x) dz, is incorrect as written (for normalized q the integral equals log p(x)), and the rest of the derivation would benefit from a clean restatement.","section":"Appendix B, Eq. (8)"},{"comment":"The training-loss notation is inconsistent: L1, Lrec_2, Lkl_2, and L3 = Ledge_rec_3 + Lnode_rec_3 + Lkl_3 are introduced without defining the three components of L3; this should be aligned across Sec. 3.3 and Appendix F.","section":"Sec. 3.3"},{"comment":"The affiliation line contains 'Sichuang, China,' which appears to be a typo for 'Sichuan, China.'","section":"Author affiliations"},{"comment":"Appendix I itself notes that the reconstruction results 'would benefit from comparison with standard reconstruction baselines in future work'; because the decoder's output distribution feeds all downstream SHAP and steering analyses, this missing comparison should be acknowledged as a limitation in the main text as well.","section":"Appendix I"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is an unreviewed preprint whose own appendices candidly concede the main gaps I have raised in the major comments; the decision hinges on whether the authors can recalibrate the headline claims to match the demonstrated results and add the missing statistical support. I would ask the authors to provide per-seed results for Tables 1-3 and for the RQ3 AUC/accuracy figures, and to state explicitly how many trained models underpin each figure. The duplicate references and the figure mis-reference suggest the manuscript needs careful copy-editing. I see no indication of an integrity problem; the issue is calibration of claims to evidence, which a major revision can address."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a credible engineering paper that assembles familiar pieces—GAT-VAE, SHAP, DP, CMA-ES—into a pipeline for Drosophila connectome subgraphs, and it reports its failures honestly. The headline claim, however, is broader than the evidence: the only target statistic that the SHAP-guided DP actually steers is edge count. Appendix H says so explicitly for reciprocity, betweenness, and non-neuronal. That matters more than the abstract suggests.\n\nWhat's genuinely new is the unusual application of graph VAEs to spatially sampled connectome subgraphs with a SHAP-based interpretability overlay, and the demonstration that a small set of latent dimensions strongly controls edge count. Reconstruction is good (node acc 0.982, edge acc 0.978, edge AUC 0.959), and the RQ1 MMD table shows the model is competitive with diffusion baselines on degree and clustering. The surrogate models have high Pearson correlations, and the paper is admirably candid about the DP limitations in Appendix H.\n\nSoft spots, in order of seriousness:\n\n1. The DP procedure assumes SHAP values give a global additive decomposition, but SHAP's efficiency property holds only at a single input point. The paper presents no validation of the additivity assumption, and its own Appendix H shows the assumption breaks down.\n\n2. No error bars or significance tests anywhere. MMD differences of 0.002 vs 0.004 may be noise; the table has no variance. This is fixable and necessary.\n\n3. The sampling step—random XZ center, cylinder radius fit to get 81–100 neurons—gets no representativeness check. The latent structure may reflect the sampling rule rather than connectome organization.\n\n4. Minor but real: '6 key dimensions' in one paragraph, '8 key dimensions' in the next; Appendix G's dimensionality analysis stops at 6, so why 8? Also, no code/data release, which hurts reproducibility.\n\nThe CMA-ES experiments are a reasonable workaround but lack baselines and error bars, so they don't rescue the uncontrolled claims.\n\nWho is this for? Someone working in connectome representation learning or graph VAE interpretability will find useful ideas and an honest evaluation. It's not a foundational result. My take: it deserves a serious referee, not a desk reject, but only with major revisions—the abstract's controlled-generation claim needs to be scoped to edge count, or all four statistics need working DP, and the code should be released. I would not cite it for the 'genetic bottleneck' interpretation, but I might cite it as an application of interpretable GVAEs to connectomes.","headline":"A credible engineering pipeline for interpretable connectome latent spaces, but the 'predefined properties' claim only holds for edge count—still worth a serious referee.","tokens_in":15039,"tokens_out":2542,"would_cite":false,"duration_ms":26876,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper argues that a compact set of latent variables can both encode and controllably steer the structure of fruit fly connectome subgraphs.","keywords":["connectome","latent representation","variational autoencoder","SHAP","graph generation","Drosophila","FlyWire","interpretability"],"falsifier":"Train the same VAE and SHAP pipeline on 100-neuron subgraphs sampled randomly within the visual region without aligning to the visual-stream axis; if the set of influential latent dimensions and their mappings to edge count, reciprocity, betweenness, and non-neuronal ratio do not persist, the claimed organizational principles are an artifact of the cylindrical sampling. A second check is to compare subgraphs generated by the steering procedure against anatomically identified columnar units in the FlyWire dataset to see whether the controlled generations match real functional columns.","tokens_in":14063,"feed_emoji":"🧠","tokens_out":6017,"duration_ms":57891,"temperature":0.7,"pith_summary":"The paper tries to establish that the local wiring of the fruit fly brain is organized around a small number of low-dimensional principles, and that these principles can be captured by a variational autoencoder trained on subgraphs sampled from the FlyWire connectome. It claims that only a few latent dimensions carry most of the information about structural features such as edge count, reciprocity, betweenness centrality, and the fraction of non-neuronal cells, with SHAP analysis identifying exactly which dimensions matter. It further claims that by adjusting these latent codes, one can generate new connectome subgraphs with targeted structural properties, and it supports this with graph reconstruction results and controlled generation experiments. If correct, this gives neuroscience a tool for testing hypotheses about circuit design rules and offers a route toward bio-inspired artificial neural networks.","feed_headline":"Six latent dimensions steer the fly brain's circuit structure","feed_subtitle":"A VAE plus SHAP turns connectome subgraphs into compact codes that can be adjusted to generate circuits with target properties.","key_machinery":"The central machinery is the combination of an information-bottleneck variational autoencoder with a SHAP-based explainability module. The encoder uses multi-head graph attention networks to embed the directed adjacency matrix, concatenates these edge embeddings with one-hot node-type features, and maps them to a low-dimensional Gaussian latent $z$; the decoder reconstructs the adjacency matrix and node categories from $z$. A differentiable surrogate model approximates each graph statistic so that SHAP values can attribute the statistics to latent dimensions, and the additive property of SHAP values turns the task of hitting a target statistic into a knapsack-like dynamic programming problem. This pipeline is what carries the argument from \"the latent space reconstructs the data\" to \"the latent space can be steered.\"","core_discovery":"On the paper's own terms, the discovery is that connectome subgraphs from the fly visual system can be compressed into low-dimensional latent codes that preserve essential topology, and that individual latent dimensions are interpretably tied to measurable structural statistics. Using a VAE with an information bottleneck, the model reconstructs both the adjacency structure and the neurotransmitter-type labels of padded 100-node subgraphs, and SHAP analysis shows that a handful of dimensions (notably dimensions 1, 9, 11, 13, 24, and 28) dominate four structural readouts: edge count, reciprocity, betweenness, and non-neuronal ratio. The paper then demonstrates that manipulating these dimensions through SHAP-guided dynamic programming or CMA-ES produces graphs whose statistics move toward the requested targets, with an AUC of 0.919 when the full target graph is available and 0.852 when only its statistics are used.","pith_inferences":["My inference: the cylindrical sampling assumption could be tested directly by training the same VAE on randomly selected 100-neuron subgraphs that ignore the visual-stream axis; if the influential latent dimensions change, the discovered principles are sampling artifacts rather than columnar biology.","My inference: the genetic-bottleneck framing suggests a quantitative prediction the paper does not make: the information content of the six influential dimensions should scale sublinearly with subgraph neuron count, and measuring that scaling across subgraph sizes would test whether the bottleneck truly mirrors genome-scale compression.","My inference: because the SHAP-guided DP search treats dimensions additively, explicitly modeling pairwise interactions between the six influential dimensions could either improve steering accuracy or reveal that the paper's linear attribution masks coupled effects.","My inference: if the same six dimensions control the same statistics in other brain regions of the fly, that would support a universal \"wiring code\" rather than region-specific organization; this is a natural transfer experiment the paper does not run."],"forward_implications":["If the claim holds, local connectome structure can be analyzed and compared through compact latent codes rather than through full graph statistics, making cross-region and cross-species comparisons tractable.","The identified steering dimensions provide a direct way to test causal hypotheses: generate subgraphs with perturbed statistics and probe how those perturbations affect circuit function.","The same pipeline can be applied to other whole-brain connectomes as they are mapped, to test whether low-dimensional organizational principles generalize beyond the fly.","Controlled subgraph generation could supply synthetic neural-network training data with specified topology, bridging connectomics and artificial neural network design.","The reconstruction metrics quantify how much information survives the bottleneck, giving an empirical bound on the compressibility of local connectome structure."],"supporting_citations":[{"why":"supplies the FlyWire whole-brain connectome dataset and the biological context of the fruit fly visual system","marker":"[9]"},{"why":"provides the variational autoencoder framework that implements the information bottleneck","marker":"[25]"},{"why":"supplies the graph attention network architecture used to encode the adjacency structure","marker":"[49]"},{"why":"provides SHAP values, the game-theoretic attribution method that links latent dimensions to structural statistics","marker":"[32]"},{"why":"motivates the genetic bottleneck hypothesis that low-dimensional generative rules underlie connectome complexity","marker":"[45]"},{"why":"supplies the dynamic programming technique used to search latent codes for target graph statistics","marker":"[10]"},{"why":"supplies CMA-ES, the black-box optimizer used as a complementary steering mechanism","marker":"[14]"},{"why":"supplies the MMD evaluation metrics and protocol used to compare generated graphs in RQ1","marker":"[18]"},{"why":"supplies the GruM baseline that the model matches on degree-distribution MMD in RQ1","marker":"[20]"}],"fun_headline_variants":["Latent codes unlock and steer fly brain circuit wiring","Interpretable AI maps connectome to six steerable levers","Fly brain connectome compressed to controllable latent dimensions","Generative model and SHAP reveal how to sculpt neural circuits","Six latent knobs tune the fly visual system's circuitry"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is that the cylinder-sampled subgraphs are functionally meaningful columnar units of the fly visual system; if the sampling is arbitrary or biased, the learned latent variables and their apparent ties to structural features could reflect sampling artifacts rather than genuine brain organization.","fun_headline_variants_meta":{"raw":{"variants":["Latent codes unlock and steer fly brain circuit wiring","Interpretable AI maps connectome to six steerable levers","Fly brain connectome compressed to controllable latent dimensions","Generative model and SHAP reveal how to sculpt neural circuits","Six latent knobs tune the fly visual system's circuitry"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000174,"raw_usage":{"total_tokens":1238,"prompt_tokens":857,"completion_tokens":381,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":473,"completion_tokens_details":{"reasoning_tokens":311}},"tokens_in":473,"tokens_out":381,"duration_ms":4593,"temperature":1.0,"reasoning_tokens":311,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:21:28.263807+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the same VAE and SHAP pipeline on 100-neuron subgraphs sampled randomly within the visual region without aligning to the visual-stream axis; if the set of influential latent dimensions and their mappings to edge count, reciprocity, betweenness, and non-neuronal ratio do not persist, the claimed organizational principles are an artifact of the cylindrical sampling. A second check is to compare subgraphs generated by the steering procedure against anatomically identified columnar units in the FlyWire dataset to see whether the controlled generations match real functional columns.","supporting_citations":[{"cited_title":"Flywire: online community for whole-brain connectomics","cited_arxiv_id":null,"evidence_quote":"supplies the FlyWire whole-brain connectome dataset and the biological context of the fruit fly visual system"},{"cited_title":"Encoding innate ability through a genomic bottleneck","cited_arxiv_id":null,"evidence_quote":"motivates the genetic bottleneck hypothesis that low-dimensional generative rules underlie connectome complexity"},{"cited_title":"What is dynamic programming? Nature biotechnology, 22(7):909–910, 2004","cited_arxiv_id":null,"evidence_quote":"supplies the dynamic programming technique used to search latent codes for target graph statistics"}],"review_version":1}