{"id":"1d7fd6ce-5989-49c6-b16d-27264c9f2f75","arxiv_id":"1908.05635","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A deep-learning embedding of 2468 Heliconius wing photographs shows statistically significant similarity between co-mimic subspecies and supports mutual convergence in Müllerian mimicry.","lead":"Using 2468 photos of Heliconius butterflies, the authors trained a deep neural network to measure how similar butterfly wing patterns are, then showed that species that mimic each other are more similar than other pairs. This gives a large-scale quantitative test of a 140-year-old evolutionary theory and suggests mimicry itself can create new wing patterns.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The embedding is optimized to separate named subspecies, so the claim that distances quantify 'total visible phenotypic similarity' is not established; a label-free control embedding would test whether the convergence signal is an artifact of the classification objective.","rationale":"The reader's weakest_assumption identified exactly this premise: that Euclidean distance in the 64-dimensional embedding captures 'total visible phenotypic similarity' even though the network is trained only to separate 38 subspecies. I agree that this is the most load-bearing assumption because it underlies every downstream claim, including the convergence result and the 'phenomic' generalization. I considered whether the post hoc selection in the mutual convergence analysis is a stronger concern, but that analysis is presented as a case study and is not the paper's primary quantitative test; the embedding-validity concern is more fundamental because it threatens the interpretation of all distances. The proposed control experiment directly tests whether the convergence signal is an artifact of the classification objective. Since this concern is already reflected in the reader's CONDITIONAL verdict and is addressable rather than fatal, the verdict should remain UNCHANGED.","tokens_in":25706,"tokens_out":8113,"duration_ms":86010,"concrete_test":"Train a control network on the same 2468 images using a reconstruction objective (e.g., a convolutional autoencoder) or a self-supervised contrastive task that never receives subspecies labels; then compute the same co-mimic versus other pairwise distance comparison and the same neighbor-joining tree correlation with color-pattern gene trees. If the convergence signal (Mann-Whitney P near 5.5e-7) and the gene-tree correlation (P <= 2.59e-256) weaken substantially or disappear under the label-free embedding, the headline results depend on label-driven feature selection rather than neutral phenotypic similarity. A complementary check is to compare ButterflyNet distances to independent human perceptual similarity ratings on a held-out set of image pairs.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that ButterflyNet's Euclidean distances quantify 'total visible phenotypic similarity' (Abstract; Discussion) is not supported by the training objective. The network minimizes triplet loss plus categorical cross-entropy to separate 38 named subspecies, so the embedding is optimized to preserve only image variation that discriminates those labels. In this system, subspecies are defined largely by wing-pattern phenotype, and traditional mimicry hypotheses were based on the same visual features, so the significant co-mimic similarity (P = 5.5e-7) may partly reflect the network recovering human taxonomic labels rather than an independent phenomic measure. The validation against color-pattern gene trees is also partly circular: both the subspecies labels and the 'color pattern genes' are tied to wing-pattern variation, and the random-tree null (P <= 2.59e-256) is weak because any tree with geographic or taxonomic signal beats random. Thus the 'phenomic' and 'total phenotypic similarity' claims are overgeneralized from a task-specific representation. The specific convergence result is plausible, but the foundation for the broad claim is not independently validated.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper trains a deep convolutional triplet network ('ButterflyNet') on 2468 dorsal and ventral photographs of 38 subspecies of Heliconius erato and H. melpomene, using subspecies identity as the training signal. It embeds all images into a 64-dimensional Euclidean space and uses pairwise distances to quantify phenotypic similarity. The authors report three main results: (i) same-subspecies images are significantly more similar than different-subspecies images; (ii) traditionally hypothesized sympatric co-mimics are significantly more similar than other subspecies pairs (Mann-Whitney pairwise, P = 5.50e-7), which they interpret as quantitative support for Müllerian mimicry; (iii) neighbor-joining trees from phenotypic distances are more similar to color-pattern gene trees than to random or neutral-gene trees. A case study of four subspecies is used to argue for mutual convergence and for the claim that mutual convergence can generate phenotypic novelty. The paper concludes that deep learning can provide phenomic embeddings enabling quantitative tests of evolutionary hypotheses previously testable only subjectively.","tokens_in":25905,"tokens_out":7038,"duration_ms":69273,"significance":"If the central claims hold, this paper would be a valuable methodological contribution: it applies modern deep metric learning to a classic evolutionary question and provides a large, openly available image dataset, code, and embedding coordinates. The effort to test a quantitative prediction of Müllerian mimicry with a data-driven phenotype representation is commendable, and the authors are careful to validate the embedding against gene-tree comparisons and to check robustness after excluding hybrids. The central observation that co-mimic subspecies are phenotypically closer than other pairs is directionally clear and consistent with the mimicry literature. However, the strength of the conclusions depends critically on whether the embedding is a neutral and comprehensive phenomic measure, and on the statistical handling of non-independent pairwise distances. The current manuscript does not fully establish these prerequisites, and several load-bearing claims require additional validation before the broad conclusions can be accepted.","major_comments":[{"comment":"The Mann-Whitney and Kruskal-Wallis tests treat average pairwise distances between subspecies pairs as independent observations, but these distances are non-independent because each subspecies appears in many pairs (e.g., 684 'other' pairs derived from only 38 subspecies). This pseudoreplication inflates the effective sample size and produces extreme P-values; the reported P = 5.50e-7 for the co-mimic test is therefore not a valid significance level. The authors should use a permutation test that resamples subspecies labels while preserving the pair structure, or a mixed-effects model with random intercepts for subspecies, to account for the dependence. This is load-bearing for the central claim of significant convergence.","section":"Results, 'Testing phenotypic convergence'; Supplementary Methods, 'Statistical analyses of phenotypic distance'"},{"comment":"The claim that the 64-dimensional embedding captures 'total visible phenotypic similarity' is not supported by the training objective. The network is trained to classify 38 named subspecies, so the embedding is optimized to preserve image variation that discriminates those labels. In this system, subspecies are defined largely by wing-pattern phenotype, and the traditional mimicry hypotheses were based on the same visual features. Consequently, the significant co-mimic similarity may partly reflect the network recovering human taxonomic labels rather than measuring an independent phenomic signal. A label-free control embedding (e.g., a convolutional autoencoder, or raw pixel-based distances after alignment) should be compared; if the convergence signal persists under that control, the claim of a general phenotypic measure would be substantially strengthened.","section":"Abstract and Discussion, 'Quantification of phenotypic similarity using deep learning'"},{"comment":"Hyperparameters were tuned to optimize test classification accuracy, as acknowledged in the Supplementary Methods, so the reported 86% test accuracy is an optimistic estimate. More importantly, the pairwise distance analyses use all 2468 images, including the 1500 images on which the network was trained. Distances involving training images are likely overfit (e.g., artificially compact same-subspecies clusters), which could bias the convergence comparisons. The authors should recompute all distance-based statistics using only the 968 held-out test images, or use a cross-validated embedding procedure, to ensure the results are not driven by training-set memorization.","section":"Materials and Methods, 'Deep learning'; Results, 'Deep learning network output and accuracy'"},{"comment":"The comparison of phenotypic neighbor-joining trees to random topologies (P ≤ 2.59e-256) is a very weak null because any tree with taxonomic or geographic signal will beat a random tree. The more meaningful comparison is to neutral-gene trees, but the combined-dataset result is only P ≤ 0.0066 (table S10), which is far more modest. Moreover, the strong similarity to color-pattern gene trees is expected given that both the subspecies labels and the color-pattern genes are tied to wing-pattern variation; this does not independently validate the embedding as a general phenomic measure. The authors should either provide a stronger null (e.g., trees from a label-free embedding or from permuted image sets) or temper the claim of 'objective, phylogenetically informative phenome capture.'","section":"Results, 'Phylogenetic results'; Supplementary Methods, 'Phylogenetic analyses'"},{"comment":"The mutual-convergence case study selects focal taxa and evolutionary polarities using gene phylogenies and phylogeographic reconstructions, then tests distances in the embedding. This selection process can bias the outcome because the data used to define the expected direction of change are not independent of the phenotype being measured. Additionally, the image-level distances within the small set of selected subspecies again raise the non-independence problem described above. The authors should report a permutation analysis over all possible polarity assignments (or a sensitivity analysis with different focal-taxon choices) to show that the mutual-convergence result is not an artifact of the selection procedure.","section":"Discussion, 'Mutual convergence and strict coevolution'; Fig. S7"}],"minor_comments":[{"comment":"The relationship between '2468 butterfly photographs' and '1234 butterfly specimens' is not made explicit until the Materials and Methods; consider stating early that each specimen has both dorsal and ventral photographs.","section":"Abstract and Introduction"},{"comment":"The caption states 'Sample sizes are 38, 19, and 684 subspecies pairs'; please clarify that these are unordered pairs and that the pair-level distances are averages over image pairs.","section":"Fig. 3 caption"},{"comment":"Equation S1 writes the triplet loss as an expectation without a margin or hinge; please specify the exact loss used (e.g., whether a margin is applied) and how the expectation is estimated during training.","section":"Supplementary Methods, Eq. S1"},{"comment":"The statement that hyperparameters were tuned to optimize test classification accuracy appears only in the supplementary; this is an important limitation and should be disclosed in the main text alongside the reported test accuracy.","section":"Supplementary Methods, 'Deep learning'"},{"comment":"The phrase 'total visible phenotypic similarity' is used as if the embedding is a complete representation of the phenotype; given the low input resolution (64 pixels high), the fixed cropping, and the label-driven training, a more cautious phrase such as 'task-specific phenotypic distance' would better match the evidence.","section":"Discussion, 'Quantification of phenotypic similarity using deep learning'"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a high-profile claim about using deep learning to test an iconic evolutionary hypothesis, and the dataset and code are genuinely valuable resources. However, the statistical treatment of pairwise distances is pseudoreplicated, the central 'total phenotypic similarity' claim is not supported by the labeling objective, and several validations use weak or partly circular controls. These are fixable with additional analyses (permutation tests, test-set-only distances, label-free controls), so I recommend major revision rather than rejection. I also note that the paper was already published in Science Advances in 2019; if this is a resubmission or a post-publication review, the editor should weigh whether the requested additional analyses are feasible and whether the claims should be moderated."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things you should know. First, this paper is worth your time: it takes a standard triplet-loss embedding, applies it to 2468 museum photographs of Heliconius, and shows that the learned Euclidean distances cleanly separate co-mimics from non-mimics and correlate with wing-pattern gene trees. Second, the paper's central slogan—that the embedding captures 'total visible phenotypic similarity'—is not established. The network is trained to discriminate 38 named subspecies, so the distances reflect the axes that separate those labels. In a system where subspecies are largely defined by wing pattern, that may be fine for the mimicry test, but it is not a neutral phenome measure.\n\nWhat is genuinely new: the application of a learned metric to museum specimen images as a phenomic distance, at a scale that was previously impractical, plus the idea that mutual convergence can shuffle pattern features between lineages and generate new combinations. The data and code are archived, which counts for a lot: the Dryad repository and Python scripts make the results checkable.\n\nWhere it gets soft. The statistical tests treat pairwise distances as independent samples; with 15 co-mimic pairs and 684 other pairs, the effective degrees of freedom are much lower than the sample sizes imply. The P values are extreme, so the conclusion probably survives, but a permutation or Mantel test would be the right tool. The hyperparameters were tuned to maximize test classification accuracy, so the reported 86% is not an unbiased estimate, and I suspect there is also leakage between dorsal and ventral images of the same specimen. The mutual-convergence case study is the weakest section: 12 focal subspecies are chosen post hoc, and the polarity of evolution is taken from the first author's prior phylogenies. That does not falsify the claim, but it makes the 'novelty generator' conclusion more speculative than the rest.\n\nNone of this sinks the core result. The co-mimic signal is large, consistent across species, and robust to hybrid exclusion. The paper deserves a serious referee, one who will ask for toned-down 'total phenome' language and more careful statistics, not for rejecting the work. I'd bring it to reading group precisely because the gap between what a classification-trained embedding measures and what the authors claim is instructive.","headline":"A serious and checkable application of deep metric learning to butterfly mimicry, whose core convergence result likely holds but whose 'total phenome' claim and mutual-convergence case study need careful qualification.","tokens_in":26446,"tokens_out":3487,"would_cite":true,"duration_ms":34668,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Phenotypic distances learned from 2,468 butterfly photos quantitatively confirm Müllerian mimicry and show that mutual convergence creates new wing patterns.","keywords":["Müllerian mimicry","deep learning","triplet network","phenotypic embedding","Heliconius","evolutionary convergence","coevolution","phenomics"],"falsifier":"Re-photograph a subset of the same specimens under altered lighting or background and measure the embedding distance between photographs of the same individual; if that within-specimen distance is as large as typical distances between co-mimic subspecies, the Euclidean distances are tracking imaging artifacts rather than a neutral phenome measure.","tokens_in":25519,"feed_emoji":"🦋","tokens_out":9297,"duration_ms":82140,"temperature":0.7,"pith_summary":"This paper claims that a deep convolutional triplet network can turn ordinary butterfly photographs into a quantitative measure of total visible phenotypic similarity, and that this measure is precise enough to test evolutionary hypotheses that have previously rested on subjective description. Applying the method to 2,468 dorsal and ventral photographs of 38 subspecies of Heliconius erato and Heliconius melpomene, the authors find that sympatric co-mimics from the two species are significantly more similar to each other than are other subspecies pairs, and that phenotypic trees built from the learned distances significantly resemble wing-pattern gene trees. They further argue that the two species have mutually converged, with each acting to some extent as model and mimic, and that this mutual convergence can recombine pattern features across lineages to generate novel phenotypes. If correct, the method provides an automated, objective phenomic distance for questions in evolutionary biology that have been testable only subjectively since the 1870s.","feed_headline":"2,468 butterfly photos confirm mimicry's oldest model","feed_subtitle":"Photo-learned distances show co-mimics converge and reveal that convergence can invent new wing patterns.","key_machinery":"The central object is ButterflyNet, a 15-layer convolutional neural network trained with a triplet loss together with a subspecies-classification loss. Each training triplet shows two photographs of the same subspecies and one of a different subspecies, and the network learns a 64-dimensional embedding in which Euclidean distance is small for same-subspecies pairs and large for different-subspecies pairs. The learned coordinates of all 2,468 images then supply pairwise phenotypic distances between subspecies, which are tested for mimicry versus non-mimicry similarity and used to build neighbor-joining trees; those trees are compared with gene phylogenies using an unrooted tree-distance metric. The same embedding enables sister-group comparisons in which a focal co-mimic's distance from its nearest conspecific is measured across all 64 axes, giving the quantitative evidence for mutual convergence.","core_discovery":"Across the complete dataset, traditionally hypothesized sympatric co-mimics are significantly more similar to each other than are other subspecies pairs (rank-based pairwise comparison, P = 5.50e-7), and neighbor-joining trees reconstructed from phenotypic distances are significantly more similar to color-pattern gene trees than to random trees (P <= 2.59e-256). For both species separately, phenotypic trees are closer to gene trees for wing-pattern loci than to neutral-locus gene trees. Comparative case studies of twelve subspecies show statistically significant mutual convergence: in the focal example H. melpomene has converged about 1.6 times as far as H. erato, consistent with frequency-dependent Müllerian fitness benefits, and the shared derived pattern features appear to have been exchanged between the lineages rather than inherited from a common ancestor. The paper concludes that mutual convergence, equivalent to reciprocal advergence, can itself generate phenotypic novelty by recombining features from the two lineages, and that deep learning embeddings can serve as quantitative phenomic distances for evolutionary analysis.","pith_inferences":["Beyond the paper's dataset, the same embedding pipeline could be applied to other mimicry rings or to any phenotype-rich museum collection; the key requirement is a set of labels (subspecies, species, morphs) strong enough to train the triplet network.","Because the embedding is optimized for subspecies discrimination, 'total phenotypic similarity' here means similarity along axes that separate the named subspecies; traits that vary within subspecies or are shared by all 38 subspecies may be underrepresented in the distances, so the method is best read as classification-relevant phenomics.","The proposed recombination mechanism, mutual convergence generating novel feature combinations, could be tested in silico by simulating two-lineage frequency-dependent predator learning on two binary pattern traits and checking whether the predicted novel combination arises and spreads.","Comparing individual embedding axes with known wing-pattern loci, such as optix-linked traits, could reveal which learned dimensions correspond to which genetic switches, connecting the phenomic distances to developmental genetics."],"forward_implications":["Co-mimic subspecies of H. erato and H. melpomene are confirmed to be more similar in visible phenotype than non-mimics, quantitatively validating the central prediction of Müllerian mimicry theory.","Phenotypic distances from deep embeddings carry significant phylogenetic information, especially for wing-pattern genes, offering a fully automated and objective route to morphological phylogenetics.","Both species show statistically significant mutual convergence, supporting reciprocal coevolution between model and mimic rather than entirely one-sided mimicry evolution.","Mutual convergence can combine pattern features from different lineages, providing a coevolutionary mechanism for the origin of novel wing patterns without direct gene exchange.","The method generalizes as a phenomic-distance pipeline: any large, consistently photographed collection can be embedded and interrogated for convergence, hybridization, biogeographic, and phylogenetic hypotheses."],"supporting_citations":[{"why":"Supplies the mathematical model of Müllerian mimicry whose convergence prediction is tested.","marker":"(4)"},{"why":"Provides the traditional qualitative hypotheses of which subspecies are sympatric co-mimics, used as the comparison groups.","marker":"(12)"},{"why":"Formulates the one-sided advergence alternative and the controversy over reciprocal coevolution that the mutual-convergence tests address.","marker":"(13)"},{"why":"Introduces the triplet-network embedding method on which ButterflyNet is built.","marker":"(21, 22)"},{"why":"Provides the published gene phylogenies and phylogeographic reconstructions used as ground truth and to assign evolutionary polarity.","marker":"(26)"},{"why":"Provides the color-pattern and neutral gene sequence phylogenies compared against phenotypic trees.","marker":"(27)"},{"why":"Supplies operational definitions and statistical approaches for testing evolutionary convergence with multivariate data.","marker":"(24, 25)"},{"why":"Provides the neighbor-joining algorithm used to build phenotypic trees.","marker":"(39)"},{"why":"Provides the comparative-method framework for sister-group comparisons used in the mutual-convergence case studies.","marker":"(40)"}],"fun_headline_variants":["Deep learning confirms butterfly mimicry convergence","AI validates oldest evolutionary model via butterfly photos","Butterfly photos prove mutual convergence in mimicry","Deep learning tests Müllerian mimicry, finds convergence","AI on 2,468 butterflies backs mimicry's oldest math"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire distance measure rests on the premise that Euclidean distance in the 64-dimensional embedding, trained only to separate the 38 named subspecies, faithfully represents total visible phenotypic similarity rather than the particular image features that distinguish those labels.","fun_headline_variants_meta":{"raw":{"variants":["Deep learning confirms butterfly mimicry convergence","AI validates oldest evolutionary model via butterfly photos","Butterfly photos prove mutual convergence in mimicry","Deep learning tests Müllerian mimicry, finds convergence","AI on 2,468 butterflies backs mimicry's oldest math"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000162,"raw_usage":{"total_tokens":1238,"prompt_tokens":939,"completion_tokens":299,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":555,"completion_tokens_details":{"reasoning_tokens":225}},"tokens_in":555,"tokens_out":299,"duration_ms":3359,"temperature":1.0,"reasoning_tokens":225,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T13:07:47.648339+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-photograph a subset of the same specimens under altered lighting or background and measure the embedding distance between photographs of the same individual; if that within-specimen distance is as large as typical distances between co-mimic subspecies, the Euclidean distances are tracking imaging artifacts rather than a neutral phenome measure.","supporting_citations":[],"review_version":1}