{"id":"71c9ff9b-b58a-4a77-9a5e-8c55d97c01ba","arxiv_id":"2507.09317","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A response-effect embedding model infers directed (asymmetric) species associations from community and environmental data, with validation mainly on simulations.","lead":"This paper introduces a machine-learning framework that models species associations as directed links, with each species having separate effect and response latent profiles. The authors test it on simulated communities and an Alpine plant dataset, and compare it with standard joint species distribution models and Markov random fields.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Direction identifiability is not established: Eq. (1)'s same-site conditional regressions can fit symmetric co-abundance structure with directed edges, and the paper's own food-web results (Sec. 5.1.2, Supp. 3.3.1) report directed associations confused with symmetric ones.","rationale":"The reader's weakest assumption and my concern coincide: the directed recovery claim rests on the ability to separate direction from symmetric residual co-variation using cross-sectional data, and that ability is neither proven nor demonstrated. I focus on this rather than on secondary issues such as the missing code link or the inconsistent hierarchical-filter equations, because it is the condition that must be true for the paper's central novelty, asymmetric association recovery, to have content. The paper does have genuine strengths: the response/effect embedding formulation is modular, the two simulation experiments are nontrivial, and the positive/negative association classification results are plausible in several configurations, especially small species pools. But those strengths do not establish direction recovery. The proposed check is deliberately simple: reuse the existing asymmetric simulated data and evaluate with a direction-aware metric against a symmetrized baseline. If the check passes, the central claim is much better supported and the paper can move toward acceptance; if it fails, the abstract's superiority claim must be tempered or the model must be shown to need temporal or spatial information beyond a single cross-sectional abundance matrix. Because the reader already conditioned acceptance on this class of concern, I keep the verdict unchanged rather than escalating to rejection.","tokens_in":33425,"tokens_out":12961,"duration_ms":169888,"concrete_test":"Re-run the Experiment-1 asymmetric simulations (+/0 and -/0, several pool sizes and densities) and score EA with a directed-edge metric: for each pair, a true positive requires sign(hat_a_ij) = sign(A_ij) and hat_a_ji classified neutral when A_ji = 0. Compare this directed F1 against a symmetrized baseline (hat_a_ij = hat_a_ji) from the same fit. On the symmetric ground-truth runs, measure the rate at which EA declares one-way links (hat_a_ij significant, hat_a_ji neutral). If the directed F1 is not clearly above the symmetrized baseline, or if a substantial fraction of symmetric-truth links are declared one-way, the model has not demonstrated directed recovery. This directly tests the confusion reported in Section 5.1.2 and Supp. Section 3.3.1 using existing simulated data.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Eq. (1c) models eta_B,ki = o_i + sum_{j in C_ki} y_kj <rho_i, alpha_j>, i.e., each species' mean is regressed on abundances of all other species at the same site. This is a simultaneous-equation dependency network, not a generative DAG: the predictors y_kj are themselves outcomes of the same process, and no compatibility or identifiability condition is given for the system of conditional distributions. For symmetric co-abundance data (e.g., from an MRF or residual-covariance JSDM), the same conditional regressions are well defined and yield regression coefficients, so an asymmetric A = P Q^T can be fitted even when the true association structure is symmetric. The low-rank factorization does not remove this ambiguity: any symmetric matrix can be written as P Q^T with P != Q. The paper's own results supply the failure mode: Section 5.1.2 states that the model detected symmetric dependencies when abiotic niches overlapped strongly, especially in trophic chains and when the predator had no alternative prey, and Supp. Section 3.3.1 identifies confusion of directed associations with symmetric reciprocal associations as the main source of error. In addition, Experiment 2 compares EA only against itself, and the Experiment-1 association-type metrics classify edges by sign rather than by whether the inferred direction is correct. Consequently, the abstract's claim of superior capacity at retrieving asymmetric interactions is not currently supported by a direction-identifiable estimator or a direction-aware benchmark.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces the Ecological Association Network (EA) framework, which models pairwise species associations as directed influences with a low-rank factorization A = P Q^T, using species-specific response embeddings rho_i and effect embeddings alpha_i. These associations enter a conditional exponential-family abundance model together with environmental covariates, under three aggregation modes (additive, multiplicative, hierarchical). The model is fitted by stochastic gradient descent with elastic-net regularization, and model selection is performed by cross-validation or information criteria. The framework is evaluated on two simulation experiments — a Virtualcom-based community assembly experiment compared against HMSC, EcoCopula, EMtree, MRFcov, and PLN, and a predator-prey food web simulation with a multiplicative interaction mode — followed by an empirical case study on an Alpine plant community. The abstract claims that the framework recovers known symmetric and asymmetric associations and is superior to existing JSDMs and probabilistic graphical models.","tokens_in":33739,"tokens_out":5580,"duration_ms":70237,"significance":"If the recovery claim were established, the framework would fill a genuine gap: most JSDM and MRF approaches produce symmetric association matrices, while many ecological interactions are directed. The manuscript has several strengths: the model is modular (three aggregation functions, flexible biotic-context definitions including temporal and spatial extensions), it handles count and presence/absence data, the source code is made available, and the simulation benchmarks use external generative models (Virtualcom and a trophic simulation) rather than the inference model itself. The empirical case study is ecologically interpretable and connects inferred structure to known Alpine plant facilitation and competition patterns. However, the central claim of recovering asymmetric (directed) associations is not currently supported by the reported analyses, for reasons detailed in the major comments. The paper would still be valuable as a flexible dependency-network model for exploratory analysis, but the abstract and conclusions overstate the evidence for direction recovery and for superiority over existing methods.","major_comments":[{"comment":"Eq. (1c) specifies eta^B_ki = o_i + sum_{j in C_ki} y_kj <rho_i, alpha_j>, i.e., each species' mean is regressed on the abundances of all other species at the same site. These predictors are themselves outcomes of the same community process, making the model a simultaneous-equation dependency network rather than a generative DAG. The paper gives no compatibility or identifiability condition under which directed edge weights are uniquely determined from the joint distribution of Y_k. In particular, data generated from a symmetric joint model (e.g., an MRF or a residual-covariance JSDM) still yield well-defined conditional regressions with nonzero coefficients, so an asymmetric matrix A = P Q^T can be fitted to symmetric ground truth; the low-rank factorization does not remove this ambiguity because any symmetric matrix can be written as P Q^T with P != Q. The manuscript's own results illustrate the failure mode: Supp. §3.3.1 states that the main source of error is the confusion of directed associations with symmetric reciprocal associations, and §5.1.2 reports that symmetric dependencies are detected when abiotic niches overlap strongly, especially in trophic chains and when the predator has no alternative prey. Since the central claim is recovery of asymmetric associations, this identifiability gap is load-bearing and needs to be addressed, for instance by proving identifiability under stated assumptions, by using temporal or spatial structure that breaks symmetry, or by substantially reframing the claims as conditional dependency summaries rather than causal directed interactions.","section":"§2.2.1, Eq. (1c)"},{"comment":"Experiment 1's evaluation does not measure direction recovery. The inferred association matrices are discretized and compared by sign class (positive, negative, neutral) using precision, recall, and F1 per type, and the supplementary results report 'association type inference' rather than edge orientation. Thus an inferred edge with the correct sign but the wrong direction is scored as correct, and the comparison cannot establish that EA retrieves asymmetric (directed) associations better than symmetric baselines. The abstract's claim of 'superior capacity at retrieving symmetric and asymmetric interactions' is therefore not supported by the reported metrics. I request direction-aware evaluation metrics (e.g., orientation accuracy, directed precision/recall, or Hamming distance on the directed adjacency matrix) in addition to sign-based metrics.","section":"§3.1.3, Supp. §2.4.2"},{"comment":"Experiment 2 is a feasibility study rather than a comparative test: the multiplicative aggregation setting is not supported by the alternative methods, so EA is evaluated only against itself (with and without embedding sharing) and against potential vs. realized food webs. Consequently, it cannot support the comparative 'superior capacity' claim for asymmetric interactions. Moreover, the dominant error in this experiment is the confusion of directed with symmetric reciprocal associations (Supp. §3.3.1), so the experiment does not demonstrate reliable recovery of edge direction even within EA. At minimum, the paper should report direction-oriented accuracy separately from undirected edge detection and should temper the abstract accordingly.","section":"§3.2, §3.2.4"}],"minor_comments":[{"comment":"The text says that when ski = 0 the abundance is deterministically set to zero, but the displayed equation places the Dirac mass at zero in the 'otherwise' branch and the count distribution F in the ski = 0 branch. The two branches appear to be reversed and should be corrected.","section":"§2.2.2, Eqs. (5)–(6)"},{"comment":"The factorial design described in the text appears to imply more than 33 simulation datasets; the manuscript should clarify how the 33 datasets were obtained and which combinations were excluded or merged.","section":"§3.1.1"},{"comment":"The caption refers to 'snow duration' while the main text and §4.3 refer to 'snowmelt date' or 'snow melting date'; these should be made consistent.","section":"Fig. 6 caption"},{"comment":"The supplementary text cites 'Gallien and Münkemüller 2015' while the main text cites 'Münkemüller and Gallien [2015]'; the reference should be unified.","section":"Supp. §2.1"},{"comment":"Several entries in Table 1 are cryptic, notably 'Support: 2/pool_size' and 'Covariance mode: full'; these should be defined in the table caption or in the methods section.","section":"Table 1"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is earnest and technically detailed, and the limitations in §5.1.2 and Supp. §3.3.1 are openly acknowledged. The main issue is that the central claim — superior recovery of asymmetric associations — is not supported by the current evaluation, and the identifiability gap is fundamental to the model specification. I would not reject outright because the framework may be salvageable with a reframing as a dependency-network model, direction-aware evaluation, and a clear statement of the assumptions under which directed edges can be identified. However, the abstract and conclusions should not claim what the evidence does not show."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read arXiv:2507.09317. Bottom line: the model is a real step forward for directed association inference in community ecology, but the paper's headline claim—superior recovery of asymmetric associations—is not yet backed by the evidence. The identifiability issue is the load-bearing gap, and their own simulations show the failure mode.\n\nNew here is the response-effect embedding factorization A = P Q^T within a conditional dependency network, with three aggregation modes (additive, multiplicative, hierarchical). That's a genuine extension of exponential family embeddings to directed species associations, and the food web experiment tests a multiplicative coupling that JSDMs and MRFs simply cannot express. The comparison set (HMSC, EcoCopula, EMtree, MRFcov, PLN) is appropriate, and the Virtualcom simulations ground the evaluation in a process-based generative model. The empirical Alpine case study is interpretively sensible, and the embeddings capture traits in a plausible way.\n\nThe central weakness is identifiability. In Eq. (1c), each species' mean is regressed on the abundances of all other species at the same site—a simultaneous-equation dependency network. No compatibility or identifiability condition is given, and any symmetric association matrix can be written as P Q^T with P ≠ Q. So an asymmetric estimate does not imply an asymmetric generating process. The paper's own food-web results (Sec. 5.1.2, Supp. 3.3.1) confirm this: directed associations are frequently confused with symmetric reciprocal ones, especially in trophic chains and under niche overlap. That is an honest limitation in the discussion, but it contradicts the abstract's 'superior capacity at retrieving symmetric and asymmetric interactions.'\n\nExperiment 1 doesn't rescue the claim either: no method is consistently better for symmetric or dense networks; EA's advantage is limited to positive associations in small pools. The evaluation classifies edges by sign, not by whether the inferred direction is correct. Experiment 2 has no JSDM/MRF baseline at all.\n\nMinor issues: the hierarchical filter equations in Section 2.2.2 are inconsistent with the prose (Eq. 5–6 appear to have the zero case flipped), and the code link is a placeholder despite the text saying the code is available.\n\nWho is this for? Ecologists wanting directed association networks from cross-sectional community data, and methodologists working on dependency networks. It deserves a serious referee; the framework is novel, modular, and worth engaging. But the revision needs to either establish identifiability conditions under which directions can be recovered, add a benchmark that measures direction errors, and temper the abstract accordingly. I'd send it to review with a request for heavy revision.","headline":"A novel directed-association framework whose abstract overclaims: identifiability of edge direction is not established and the paper's own simulations show the failure mode.","tokens_in":34294,"tokens_out":2459,"would_cite":true,"duration_ms":27252,"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":"A model recovers directed species links from abundance data","keywords":["network inference","representation learning","probabilistic graphical models","species embeddings","latent variable models","species association networks","asymmetric associations","community data"],"falsifier":"Simulate a community with a known directed predator–prey link between two species whose abiotic niches overlap strongly, fit the model under the multiplicative filter, and count how often the true directed edge is recovered as a symmetric pair; the paper's own Experiment 2 already reports this confusion, and a quantitative curve of direction-recovery error versus niche overlap would settle whether the directed parameterization is identifiable in the regimes ecologists actually sample.","tokens_in":33165,"feed_emoji":"🕸️","tokens_out":6386,"duration_ms":71253,"temperature":0.7,"pith_summary":"Most tools for inferring species associations from co-occurrence or co-abundance data assume that any link between two species is symmetric: species A's effect on B equals B's effect on A. This paper proposes a framework in which each directed association $a_{ij}$ (the influence of a source species $j$ on a target species $i$) is instead parameterized as the dot product of two low-dimensional latent vectors, an effect embedding $\\alpha_j$ and a response embedding $\\rho_i$, and is fitted jointly with species-environment responses inside a conditional generative model for abundances. The authors claim that this recovers both symmetric and asymmetric associations from community data plus environmental covariates, and that it outperforms joint species distribution models and probabilistic graphical models on simulated data while producing ecologically interpretable association networks for Alpine plants.","feed_headline":"Directed species links recovered from abundance data","feed_subtitle":"Effect and response embeddings let a single framework retrieve asymmetric and symmetric associations from community and environmental data.","key_machinery":"The central object is the factored association matrix $A = P Q^\\top$, built from two species-specific latent embeddings: an effect embedding $\\alpha_j \\in \\mathbb{R}^d$ and a response embedding $\\rho_i \\in \\mathbb{R}^d$. The biotic context at a site is the abundance-weighted average of the effect embeddings of co-occurring species, $z_{ki} = \\frac{1}{|C_{ki}|}\\sum_{j \\in C_{ki}} y_{kj}\\alpha_j$, so the biotic contribution to species $i$'s linear predictor is $\\rho_i \\cdot z_{ki} = \\sum_j y_{kj} a_{ij}$. This factorization is what makes the network directed and low-rank; fitting it jointly with habitat suitability $h_i(x_k)$ and a chosen aggregation function (additive, multiplicative, or hierarchical zero-inflated) defines the conditional generative model that is trained by stochastic gradient descent with elastic-net regularization. The same embeddings also provide the structure for co-clustering species into response groups and effect groups.","core_discovery":"The paper's central claim is that the response–effect duality long used in functional ecology can be turned into an identifiable statistical parameterization of directed species associations. Writing the association matrix as $A = P Q^\\top$, with rows $\\rho_i$ (how species $i$ responds to the community) and columns $\\alpha_j$ (how species $j$ affects the community), every pairwise influence is directional by construction: $a_{ij} = \\langle \\rho_i, \\alpha_j \\rangle$ is generally not equal to $a_{ji}$. These associations enter a conditional model $y_{ki} \\sim F(o_i + h_i(x_k) + \\sum_{j \\in C_{ki}} y_{kj} a_{ij}, \\phi_i)$, with additive, multiplicative, or hierarchical ways of combining the abiotic response $h_i(x_k)$ with the biotic context. The paper reports that the fitted model recovers known positive and negative associations in process-based simulations, detects predator–prey links under a multiplicative filter when pairs co-occur, and yields modules and structural roles in an Alpine plant dataset that align with snow-melt gradients and facilitation–competition ecology. The authors' bottom line is that asymmetric associations are retrievable from spatial community data without imposing a symmetry assumption, provided the aggregation of abiotic and biotic filters is chosen appropriately.","pith_inferences":["The identification step is the part most worth probing: the paper's own Experiment 2 shows that strongly overlapping abiotic niches make directed predator–prey links come out as symmetric reciprocal associations, which suggests that cross-sectional co-occurrence data alone may not guarantee edge-direction identifiability; the temporal and spatial biotic-context extensions sketched in the supplemen","Edges recovered by this conditional regression are best read as net spatial associations conditional on measured environment, not as established causal ecological interactions; linking them to independently known interaction data would quantify how much of the inferred network is true interaction signal versus shared missing-environment response.","If the low-rank factorization is the bottleneck for representing arbitrary directed acyclic graphs, as the discussion suggests, constraining a few known edges or trophic levels as semi-supervised information could improve direction recovery more than increasing data alone.","A practical extension offered by the framework is to use learned effect embeddings as community ordination axes and response embeddings as species loadings, yielding a direct ordination of communities in interaction space."],"forward_implications":["Each species pair receives two directed association values, so asymmetric interactions such as amensalism, parasitism, or predator–prey dependence are represented without assuming $a_{ij} = a_{ji}$.","Because associations are low-rank factorized, the number of parameters grows with the embedding dimension rather than with the square of the species pool, making large communities tractable and allowing species to share embeddings within functional groups.","The three aggregation modes extend association inference beyond additive environmental-plus-biotic structure; the multiplicative mode is necessary for obligate dependencies such as a predator requiring prey presence.","On the Alpine plant data, the inferred network separates modules tied to snow-melt timing and highlights asymmetric facilitation of forbs and grasses by dominant graminoids in stressful early-melting sites.","Compared with HMSC, EcoCopula, EMTree, MRFcov, and PLN on simulated communities, the paper reports superior recovery of both symmetric and asymmetric associations, with a particular edge on negative associations."],"supporting_citations":[{"why":"Supplies the process-based community simulator (Virtualcom) used to generate ground-truth communities in Experiment 1.","marker":"Münkemüller and Gallien [2015]"},{"why":"Defines the joint species distribution model framework whose residual-correlation approach the paper contrasts with its directed associations.","marker":"Warton et al. [2015]"},{"why":"Provides the HMSC baseline that infers residual covariance and is one of the five competitor methods in the comparison.","marker":"Ovaskainen et al. [2017]"},{"why":"Provides the MRFcov baseline, a Markov random field approach to species association inference.","marker":"Clark et al. [2018]"},{"why":"Provides the Poisson log-normal network (PLN) baseline used in the comparison.","marker":"Chiquet et al. [2018]"},{"why":"Provides the EcoCopula copula-based graphical model baseline used in the comparison.","marker":"Popovic et al. [2019]"},{"why":"Provides the EMTree tree-based graphical model baseline used in the comparison.","marker":"Momal et al. [2019]"},{"why":"Supplies the Alpine plant abundance dataset with environmental variables used in the empirical case study.","marker":"Choler [2005]"},{"why":"Frames the dependency-network formalism on which the node-wise conditional regression model is built.","marker":"Heckerman et al. [2000]"}],"fun_headline_variants":["Asymmetric species links emerge from abundance data","Directed species interactions retrieved from community data","Bidirectional species associations from environmental data","Model recovers directed species relationships","Unmasking asymmetric species associations"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the direction of a species-to-species association can be separated from symmetric residual correlation using only a cross-sectional abundance matrix together with environmental covariates; the paper's own predator–prey simulation shows that when abiotic niches overlap strongly, directed links are frequently inferred as symmetric reciprocal ones, so if that identifiability fails in a given data regime the central claim about recovering asymmetric associations collapses.","fun_headline_variants_meta":{"raw":{"variants":["Asymmetric species links emerge from abundance data","Directed species interactions retrieved from community data","Bidirectional species associations from environmental data","Model recovers directed species relationships","Unmasking asymmetric species associations"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00039,"raw_usage":{"total_tokens":2106,"prompt_tokens":1048,"completion_tokens":1058,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":664,"completion_tokens_details":{"reasoning_tokens":998}},"tokens_in":664,"tokens_out":1058,"duration_ms":9808,"temperature":1.0,"reasoning_tokens":998,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T17:59:02.250318+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate a community with a known directed predator–prey link between two species whose abiotic niches overlap strongly, fit the model under the multiplicative filter, and count how often the true directed edge is recovered as a symmetric pair; the paper's own Experiment 2 already reports this confusion, and a quantitative curve of direction-recovery error versus niche overlap would settle whether the directed parameterization is identifiable in the regimes ecologists actually sample.","supporting_citations":[{"cited_title":"u nkem \\","cited_arxiv_id":null,"evidence_quote":"Supplies the process-based community simulator (Virtualcom) used to generate ground-truth communities in Experiment 1."},{"cited_title":"So many variables: joint modeling in community ecology","cited_arxiv_id":null,"evidence_quote":"Defines the joint species distribution model framework whose residual-correlation approach the paper contrasts with its directed associations."},{"cited_title":"How to make more out of community data? a conceptual framework and its implementation as models and software","cited_arxiv_id":null,"evidence_quote":"Provides the HMSC baseline that infers residual covariance and is one of the five competitor methods in the comparison."},{"cited_title":"Unravelling changing interspecific interactions across environmental gradients using markov random fields","cited_arxiv_id":null,"evidence_quote":"Provides the MRFcov baseline, a Markov random field approach to species association inference."},{"cited_title":"Variational inference for sparse network reconstruction from count data","cited_arxiv_id":"1806.03120","evidence_quote":"Provides the Poisson log-normal network (PLN) baseline used in the comparison."},{"cited_title":"Untangling direct species associations from indirect mediator species effects with graphical models","cited_arxiv_id":null,"evidence_quote":"Provides the EcoCopula copula-based graphical model baseline used in the comparison."},{"cited_title":"Dependency networks for inference, collaborative filtering, and data visualization","cited_arxiv_id":null,"evidence_quote":"Frames the dependency-network formalism on which the node-wise conditional regression model is built."}],"review_version":1}