{"id":"9675b041-63db-4b37-86bf-25f6b6992a52","arxiv_id":"2411.14003","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"A generative intervention model maps perturbation features to distributions over atomic interventions in a jointly learned causal model, enabling out-of-distribution prediction with mechanistic insight.","lead":"This paper introduces a causal model that maps drug or perturbation features to a distribution over atomic interventions in a learned causal graph, then predicts the response of unseen perturbations by sampling from the intervened model. If it works, it gives drug-response prediction with mechanistic interpretability, not just a black-box shift.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Mechanism-inference benchmark uses PCA features built from the true interventions: synthetic fully OOD and Target-F1 results may show decoding rather than causal generalization; real-data OOD only holds out dosages, not drugs.","rationale":"The paper introduces a genuinely useful framework, and the prediction side of the central claim is supported by the synthetic experiments and by real-data dose extrapolation. The code is released, and the synthetic benchmark is carefully constructed. In good faith, the central claim should be read as a two-part claim: robust OOD prediction of perturbed densities, and mechanistic inference of the underlying atomic interventions. The prediction part is reasonably supported. The mechanism-inference part is the least secure because the synthetic features gamma are built by PCA-compressing the true (I, psi) pairs, so the fully OOD setting is OOD only with respect to the sampled intervention, not with respect to the information available in gamma. The causal baselines never see gamma, making the Target-F1 comparison favorable to GIM in a way that may not reflect the causal modeling contribution. The paper explicitly acknowledges that generalization depends on informativeness of gamma, so this is a scoping and evidence concern rather than an internal inconsistency. The proposed MLP-decoder check would settle whether the target-inference gains come from the engineered features or from the joint causal model. If the decoder matches GIM, the abstract should be tempered; if not, the claim stands. The real-data claim could also be strengthened by a held-out-drug evaluation, but the one decisive check for the mechanism-inference concern is the decoder comparison. Overall, the reader's CONDITIONAL verdict remains appropriate; the concern sharpens the condition but does not overturn it.","tokens_in":27314,"tokens_out":12987,"duration_ms":138193,"concrete_test":"On the same synthetic data, train an MLP decoder that maps the PCA features gamma directly to the true target vector I (Bernoulli head, same training environments and same fully OOD target pairs), and compute its fully OOD Target-F1 alongside GIM's. If this decoder matches or exceeds GIM's Target-F1, then the mechanism-inference advantage in Figure 2 is explained by the engineered PCA encoding of the ground-truth intervention (Section 5.1), not by the joint causal model; the abstract's \"often better than other causal inference methods\" should be restricted to settings where gamma carries the target information. If the decoder is worse, the causal-modeling component is responsible for the target-inference gains and the concern does not land.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing soft spot is the construction of the synthetic perturbation features. Section 5.1 states: \"We create the features gamma by standardizing the concatenation of I and psi, applying principal component analysis (PCA) across all training environments of a system, and keeping the top 15 principal components.\" So in the fully OOD task, gamma is a lossy code of the very intervention targets and parameters the method claims to infer. GIM therefore receives the ground-truth mechanism (compressed) as input, while the causal baselines (BaCaDI*, UT-IGSP, etc.) never see gamma. The Target-F1 comparisons in Figure 2 are thus not a clean test of causal discovery under unknown interventions; they are largely a test of how well g_phi can invert a PCA encoding. The paper honestly acknowledges the general dependence on gamma informativeness (Section 4.2 and Figure 4), but the headline claim \"effectively inferring the underlying perturbation mechanisms, often better than other causal inference methods\" is only supported when gamma is engineered from those mechanisms. On SciPlex3, moreover, the OOD evaluation holds out only the highest dosage of each of four seen drugs, not a whole unseen drug, so generalization to genuinely novel mechanisms is not demonstrated on real data. None of this invalidates the framework under its stated assumptions, but it means the central two-part claim (predict unseen perturbations and recover mechanisms) is currently validated only in a favorable, partially circular setup.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes generative intervention models (GIMs), a causal perturbation modeling framework that maps observed perturbation features gamma to distributions over atomic interventions I=(I,psi) in a jointly estimated structural causal model. Predictions for unseen perturbations are obtained by marginalizing over the intervention distribution in the causal model (Eqs. 4-6). The authors derive a MAP objective with Monte Carlo gradient estimators, evaluate on synthetic linear/nonlinear SCMs and SciPlex3 scRNA-seq drug perturbation data, and report that GIMs match unstructured baselines on distributional prediction while recovering intervention targets better than other causal inference methods.","tokens_in":27583,"tokens_out":4846,"duration_ms":47003,"significance":"If the central claim holds, the framework fills a real gap: existing causal discovery methods with unknown interventions (BaCaDI, UT-IGSP, GnIES, JCI-PC) cannot make predictions for unseen perturbation features, while unstructured black-box models provide no mechanistic insight. The core modeling contribution is clearly specified: Equations (4)-(6) define a coherent generative process, Appendix B gives explicit gradient derivations, and the experiments on synthetic data use known ground truth, which is a genuine strength. The paper additionally studies the dependence on the information content of gamma in Figure 4 rather than hiding it, and it ships code. However, the central two-part claim is currently supported under favorable evaluation choices: the synthetic perturbation features are constructed by PCA over the true intervention targets and parameters, and the real-data OOD evaluation holds out only dosages of seen drugs, not unseen drugs. These issues are load-bearing for the headline claims about mechanism inference and generalization to novel perturbations.","major_comments":[{"comment":"The synthetic perturbation features are created by standardizing and PCA-compressing the concatenation of the true intervention target mask I and intervention parameters psi across training environments (Section 5.1 and Appendix D.1). In the fully OOD setting, the test features are constructed from the very quantities the model is asked to infer. Consequently, the Target-F1 comparisons in Figure 2 and the information-content analysis in Figure 4 largely measure whether g_phi can invert a PCA-compressed code of the true mechanism, rather than whether the model discovers mechanisms from perturbed samples alone. The causal baselines (BaCaDi*, UT-IGSP, GnIES, JCI-PC) never receive gamma, so the comparison is not an equal test of mechanism inference. This does not invalidate the framework under the stated assumption that gamma is informative, but the abstract's claim of 'effectively inferring the underlying perturbation mechanisms, often better than other causal inference methods' is overstated. I recommend either constructing gamma from observable proxies independent of the true I and psi, or restricting the mechanism-inference claim explicitly to the setting where gamma is informative by construction and acknowledging that Target-F1 then includes a decoding component.","section":"Section 5.1, Appendix D.1"},{"comment":"The only real-data out-of-distribution evaluation holds out the highest dosage (10 uM) of each of the four drugs; all test perturbations share drugs seen in training. The perturbation features include one-hot drug identity and dosage encodings (Appendix D.2), so the model can use drug identity as a strong conditioning signal. This does not demonstrate generalization to unseen drugs, which is the motivating scenario in the introduction ('predicting how cells react to novel molecules'). The abstract's 'unseen perturbation features' should be qualified in the real-data experiments as 'unseen dosages of seen drugs', and the paper should either add a truly held-out drug evaluation or explicitly scope the generalization claim.","section":"Section 5.1, Figure 5"},{"comment":"The text states that GIMs 'significantly outperform' the baselines in nonlinear systems (Section 6, Figure 2B) and 'tend to yield better results for fully OOD interventions' (Section 6, Figure 3). The reported boxplots show medians and IQRs over perturbations/contexts, but no repeated-seed variance, confidence intervals, or hypothesis tests are provided. Without these, the word 'significantly' is not supported, and some apparent differences (e.g., GIM versus MLP on fully OOD W2 in Figure 3) may be within noise. Please add repeated-seed error bars and appropriate statistical tests, or soften the significance language.","section":"Section 6, Figures 2 and 3"}],"minor_comments":[{"comment":"The notation p(I; gamma, phi) = p(I; g_phi(gamma)) p(psi; h_phi(I, gamma)) uses the same symbol p for distributions over different spaces and conditions psi on I while also treating I as the target set; a brief notational clarification would help readers distinguish the target mask from the full atomic intervention.","section":"Section 3.2, Eq. (4)"},{"comment":"The sentence 'Identifiability for unseen perturbations depends on the informativeness of the features gamma, which we evaluate empirically Section 6' is missing a preposition ('in Section 6') and should also cite Figure 4 explicitly, as that figure is the direct empirical study of this dependence.","section":"Section 4.2, last paragraph"},{"comment":"The claim that GIMs achieve robust predictions 'on par with unstructured approaches' is stronger than what Figure 3 shows for in-distribution and partially OOD settings, where the MLP baseline consistently has the lowest Mean Distance; the wording should be adjusted to 'competitive' or the comparison should be qualified.","section":"Section 6, SciPlex3 paragraph"},{"comment":"The initialization statement 'with lambda_0 = 0 and mu_0 = 1-9' appears to contain a typo (presumably 1e-9); please correct the exponent notation.","section":"Appendix D.4"}],"recommendation":"major_revision","confidential_remarks":"This is a well-executed and clearly written paper, and the core modeling idea is a useful contribution. The main issue is that the most impressive-sounding results (mechanism inference on fully OOD synthetic perturbations, and generalization on real data) are obtained under conditions that make the claim look stronger than the evidence supports: synthetic gamma is a PCA code of the true interventions, and real-data OOD is dosage-level, not drug-level. If the authors add an experiment with gamma built from features that do not directly encode the true I and psi, or substantially soften the claims, the paper could be acceptable. The lack of repeated-seed statistics is a secondary but important issue for a journal version."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper delivers something real: a framework that amortizes the mapping from perturbation features to distributions over atomic interventions in a jointly estimated causal model. That closes an actual gap — prior causal methods infer interventions per context without using features, and black-box predictors give no mechanistic account. The MAP objective, Monte Carlo estimators, and gradient derivations are clearly laid out, and code is available. On synthetic data with the correct model class, the central claim holds: GIM matches black-box predictors out of distribution and beats causal baselines on nonlinear systems.\n\nThe soft spots are real but not fatal. The fully OOD benchmark is more favorable than the headline suggests: the synthetic features gamma are defined as the top PCA components of the true intervention targets and parameters. So in that task GIM is partly decoding a compressed ground-truth mechanism, while the causal baselines never see gamma. The Target-F1 comparison is therefore not a clean test of causal discovery under unknown interventions. The paper is honest about the general dependence on gamma informativeness — it states it in Section 4.2 and probes it in Figure 4 — but the abstract's mechanism-inference claim rests on that favorable construction. On SciPlex3, the OOD evaluation holds out only the highest dosage of each of four seen drugs, so genuinely novel mechanisms are not tested on real data. Also, the selection of the most effective drugs and top marker genes makes the real-data evaluation look better than it should, and there are no repeated-seed error bars or significance tests anywhere, which matters when the comparisons are close.\n\nNone of this undermines the framework under its stated assumptions. The amortized intervention mapping is new, the derivations look sound, and the synthetic experiments support the core claim when the model class matches. The paper would benefit from a revision that re-runs the fully OOD benchmark with features that do not encode the true mechanism, adds seed variability, and scopes the real-data claims to dose extrapolation rather than novel-drug prediction.\n\nWho should read this: anyone working on perturbation prediction or causal discovery with unknown interventions. It deserves a serious referee — the idea is good enough that the flattering benchmarks should be fixed rather than ignored.","headline":"GIMs close a real gap by amortizing the mapping from perturbation features to atomic interventions in a jointly learned causal model, but the fully OOD benchmark is partly circular because the features encode the true mechanisms.","tokens_in":28175,"tokens_out":1363,"would_cite":true,"duration_ms":12491,"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":"One model learns to map perturbation features to atomic interventions in a causal model, predicting effects of unseen perturbations and identifying their targets.","keywords":["generative intervention model","causal perturbation modeling","out-of-distribution prediction","intervention target identification","causal discovery","structural causal model","single-cell perturbation response"],"falsifier":"Hold out perturbations whose feature vectors γ are constructed from randomly permuted or heavily PCA-truncated versions of the true intervention targets; if GIM predictions remain above the observational baseline for such uninformative features, the claimed dependence on γ's informativeness is not real.","tokens_in":27062,"feed_emoji":"🧬","tokens_out":7739,"duration_ms":70767,"temperature":0.7,"pith_summary":"The paper tries to establish that a causal model can predict the distribution shift caused by a never-seen perturbation, provided the perturbation is described by observable features. The proposed generative intervention model (GIM) learns a map from those features to a distribution over atomic interventions — which variables are targeted and with what parameters — inside a causal model that is itself learned from data. If correct, this gives practitioners both prediction and mechanism: for a new drug, the model outputs a shifted gene-expression distribution and an explanation in terms of the regulatory mechanisms it hits. On synthetic structural causal models and on scRNA-seq drug-perturbation data, the authors report that GIMs match unstructured black-box predictors on out-of-distribution accuracy while recovering intervention targets and causal graphs at least as well as dedicated causal-inference methods.","feed_headline":"Drug features predict unseen effects and causal targets in one model","feed_subtitle":"A causal model plus an intervention generator matches black-box accuracy and names the targets.","key_machinery":"The central object is the generative intervention model, defined by equations (4)–(5): a learnable conditional distribution over atomic interventions, p(I;γ,φ) = p(I;gφ(γ)) p(ψ;hφ(I,γ)), where gφ and hφ are neural networks outputting Bernoulli target probabilities and Gaussian intervention parameters. It turns the unobserved atomic intervention into an amortized latent variable: instead of inferring I separately for each observed perturbation, the model learns a shared map from features to interventions and integrates I out when computing the likelihood. This is what lets the same causal model predict p(x|D;γ*) for arbitrary unseen features γ*.","core_discovery":"GIMs treat the atomic intervention I (target set and parameters) underlying each perturbation as a latent variable generated from the perturbation features γ by a shared, learned function, and marginalize it out in the likelihood. Concretely, the intervention prior factorizes as p(I;γ,φ) = p(I;gφ(γ)) p(ψ;hφ(I,γ)), and the perturbed density is p(x;M,γ,φ) = ∫ p(x|I;M) p(I;γ,φ) dI. Model M (causal graph and mechanisms) and the generator φ are estimated jointly by MAP, using a continuous relaxation of the graph and Gumbel-sigmoid reparameterization of the discrete targets. The result is that, unlike earlier causal approaches that infer one intervention per training context, a GIM can evaluate any γ, including unseen ones, by sampling its intervention and propagating it through the causal model. The paper's empirical claim is that this yields out-of-distribution predictions on par with the best unstructured baselines while providing mechanistic interpretations that are often more accurate than those of the causal baselines.","pith_inferences":["A practical consequence the authors leave implicit: GIM's posterior over interventions for a candidate γ makes it a natural component in active experiment design — pick the next perturbation to maximize information about the causal model and its intervention generator, then predict its effect before running it.","The atomic-intervention assumption suggests a boundary condition for deployment: perturbations that act by rewiring the graph itself, or by introducing interactions between mechanisms, fall outside the model class; detecting such cases would require a model-selection or residual check on held-out perturbations.","The same 'features → latent interventions → distribution' template could transfer to other domains where interventions are described by context variables — policy changes, device failures, environmental interventions — with the caveat that atomicity and feature informativeness must hold."],"forward_implications":["For a new drug or dosage never seen in training, GIM produces a full predictive distribution over system variables, because sampling an intervention and propagating it through the causal model generates samples from p(x|D;γ*).","The same learned causal model answers mechanistic questions: the inferred intervention targets and parameters indicate which genes or pathways a perturbation acts on, giving interpretability that black-box predictors lack.","Multiple perturbations specified by different feature vectors can be combined in silico by generating each atomic intervention and applying them jointly in the causal model, enabling predictions for combination treatments.","When the causal model is misspecified, as in the scRNA-seq analysis with a zero-inflated log-normal mechanism, the predictive distribution can still be accurate, suggesting the framework degrades gracefully under model mismatch.","Identifiability of unseen perturbations is bounded by the information content of γ: as the paper's PCA analysis shows, predictive accuracy degrades monotonically as γ is made less informative."],"supporting_citations":[{"why":"Supplies the interventional Markov equivalence class and identifiability results that GIMs inherit for recovering graph and targets in the large-sample limit.","marker":"Brouillard et al. (2020)"},{"why":"BaCaDI is the direct per-perturbation baseline and ablation; GIM extends its inference model with an amortized generator from features.","marker":"H¨agele et al. (2023)"},{"why":"DiBS continuous graph representation (latent Z, Bernoulli graph likelihood) is used for differentiable MAP inference of the causal model.","marker":"Lorch et al. (2021)"},{"why":"JCI provides the framework for causal discovery with unknown interventions and the JCI-PC baseline.","marker":"Mooij et al. (2020)"},{"why":"UT-IGSP is the permutation-based baseline for unknown intervention targets and graph discovery.","marker":"Squires et al. (2020)"},{"why":"SciPlex3 single-cell drug perturbation dataset is the real-world evaluation for held-out dosages.","marker":"Srivatsan et al. (2020)"},{"why":"NOTEARS-style continuous acyclicity constraint (with the NO-BEARS variant) is used in the model prior.","marker":"Zheng et al. (2018)"},{"why":"Provides the formalization of atomic interventions and sparse mechanism shifts that the generative process assumes.","marker":"Peters et al. (2017)"}],"fun_headline_variants":["Unseen drug effects predicted via causal interventions","Causal model maps drug features to intervention targets","Predict perturbation outcomes and name their mechanisms","Generative intervention model catches black-box accuracy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that every perturbation's effect can be captured by a sparse atomic intervention on one fixed structural causal model, and that the features describing the perturbation are informative enough for the learned map to generalize to perturbations never seen in training.","fun_headline_variants_meta":{"raw":{"variants":["Unseen drug effects predicted via causal interventions","Causal model maps drug features to intervention targets","Predict perturbation outcomes and name their mechanisms","Generative intervention model catches black-box accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000203,"raw_usage":{"total_tokens":1370,"prompt_tokens":912,"completion_tokens":458,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":528,"completion_tokens_details":{"reasoning_tokens":403}},"tokens_in":528,"tokens_out":458,"duration_ms":4838,"temperature":1.0,"reasoning_tokens":403,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T15:38:43.722301+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Hold out perturbations whose feature vectors γ are constructed from randomly permuted or heavily PCA-truncated versions of the true intervention targets; if GIM predictions remain above the observational baseline for such uninformative features, the claimed dependence on γ's informativeness is not real.","supporting_citations":[{"cited_title":"Differentiable causal discovery from interventional data","cited_arxiv_id":null,"evidence_quote":"Supplies the interventional Markov equivalence class and identifiability results that GIMs inherit for recovering graph and targets in the large-sample limit."},{"cited_title":"DiBS : Differentiable Bayesian structure learning","cited_arxiv_id":null,"evidence_quote":"DiBS continuous graph representation (latent Z, Bernoulli graph likelihood) is used for differentiable MAP inference of the causal model."},{"cited_title":"Permutation-based causal structure learning with unknown intervention targets","cited_arxiv_id":null,"evidence_quote":"UT-IGSP is the permutation-based baseline for unknown intervention targets and graph discovery."},{"cited_title":"R., McFaline-Figueroa, J","cited_arxiv_id":null,"evidence_quote":"SciPlex3 single-cell drug perturbation dataset is the real-world evaluation for held-out dosages."},{"cited_title":"K., and Xing, E","cited_arxiv_id":null,"evidence_quote":"NOTEARS-style continuous acyclicity constraint (with the NO-BEARS variant) is used in the model prior."}],"review_version":1}