{"id":"bbd98945-b8c3-4097-998e-220ba965cec6","arxiv_id":"2509.01916","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"GRACE-VAE couples a graph neural network encoder with a causal discrepancy VAE decoder so that known pathway and protein networks improve learning of latent causal programs and prediction of CRISPR perturbation effects.","lead":"This paper introduces GRACE-VAE, a graph-aware causal VAE that uses known biological network structure to learn latent causal programs from CRISPR perturbation data. It claims to inherit prior identifiability guarantees and to improve prediction of single and unseen double perturbation effects.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Lemma 1's proof is invalid: polynomial automorphisms need not be affine, so the identifiability guarantee as presented is not established.","rationale":"The reader's weakest assumption was the gap between polynomial-decoder theory and the MLP implementation. My pass found a sharper, internal problem: the appendix's own Lemma 1 is refuted by a concrete polynomial automorphism. This is more load-bearing than the mismatch because it attacks the proof actually supplied for the identifiability claim, not merely the applicability of a cited theorem. The inheritance argument itself (GNN only in encoder, likelihood unchanged) is structurally sound, and the empirical study is substantial; however the false lemma means the paper has not demonstrated the theoretical claim as written. I do not see this as requiring rejection: the authors can remove the lemma, defer to Zhang et al. (2023), and clarify the decoder assumption; the empirical contributions and double-intervention evaluation remain useful. Hence I keep the reader's CONDITIONAL verdict. I also note the NORMAN single R2 overclaim, but it is secondary and does not affect the verdict.","tokens_in":16715,"tokens_out":9343,"duration_ms":105718,"concrete_test":"Directly check Lemma 1 with U=(U1,U2), f(U)=U, hat U=(U1+U2^2,U2), hat f(x1,x2)=(x1-x2^2,x2). Both decoders are full-row-rank polynomials and P_{hat f(hat U)}=P_U=P_{f(U)}, while the conclusion hat U=Lambda U+b fails. If the authors respond that Lemma 1 is not needed, require them to mark the appendix proof as merely descriptive, cite Zhang et al. for the theorem, and either (a) state the guarantee only for polynomial decoders or (b) show the actual MLP decoder satisfies Assumption 1 on the data support.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing point is in the Appendix's Lemma 1 and its use in Theorem 1. Lemma 1 asserts that any alternative latent representation matching the observational distribution must be affine in the true latent. The proof's final step invokes a 'fundamental theorem of algebra for polynomial automorphisms' to conclude that two polynomial maps inverse to each other must have degree 1. That theorem is false. F(x,y)=(x+y^2,y) is a polynomial automorphism of R^2 with polynomial inverse (x-y^2,y), non-linear and full-rank everywhere. Concretely, with f=id and hat f=F^{-1}, both full-row-rank polynomial decoders, take hat U=F(U). Then P_{hat f(hat U)}=P_U=P_{f(U)}, but hat U is not Lambda U+b. Thus Lemma 1 as stated is false, and the proof of Theorem 1 collapses at its first step. To retain the theoretical claim, the authors must either add assumptions that exclude nonlinear polynomial automorphisms and prove a corrected lemma, or explicitly defer the CD-equivalence theorem to Zhang et al. (2023) and verify its assumptions. The latter is in tension with the implementation, whose decoder is a DAG layer plus leaky-ReLU MLPs, not a polynomial map; Assumptions 2 and 3 are not checked on biological data. Secondary empirical overclaim: in Table 1 VGAE has higher NORMAN single R2 (0.9493) than GRACE-VAE (0.9452), contradicting the text's 'outperforms all baselines on R2'.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes GRACE-VAE, a variational autoencoder that injects structured biological network context (pathway-gene, pathway-pathway, and protein-protein interactions) through a GNN encoder, while keeping a causal SCM-based decoder. It claims that, because the GNN is placed only in the variational encoder, the model inherits the identifiability guarantees of causal discrepancy VAEs and can recover the latent causal graph and intervention targets up to CD-equivalence. The authors evaluate GRACE-VAE on three CRISPR perturbation datasets, including unseen double interventions, and compare against CMVAE, CMVAE-multihot, and VGAE, with ablations over GNN architecture and context levels. The central claims are: (1) theoretical identifiability of the latent causal structure, and (2) empirical improvement in interventional outcome prediction from using structured context.","tokens_in":17021,"tokens_out":5873,"duration_ms":58554,"significance":"If the theoretical claim were correct and the empirical improvements robust, this would be a useful contribution to causal representation learning in structured biological settings. The empirical study is fairly comprehensive: three datasets, 10 seeds, multiple baselines, ablations, and hyperparameter reporting. The idea of using a GNN encoder to provide structural context while leaving the causal decoder unchanged is principled and could be practically valuable. However, the theoretical guarantee as presented is not established: the appendix contains a false lemma, and the implemented model does not satisfy the polynomial-decoder assumption of the theorem. The empirical claim that GRACE-VAE outperforms all baselines on R2 is also contradicted by the paper's own Table 1. The core idea has potential, but the paper needs substantial correction before the claims are supportable.","major_comments":[{"comment":"Lemma 1 is false as stated, and its proof collapses. The proof's final step invokes a 'fundamental theorem of algebra for polynomial automorphisms' to conclude that two mutually inverse polynomial maps must have degree 1. That theorem is false: F(x,y)=(x+y^2,y) is a polynomial automorphism of R^2 with polynomial inverse (x-y^2,y), is full-rank everywhere, and is not affine. Concretely, take f=identity and \\f=F^{-1}, and \\U=F(U). Then \\f(\\U)=U, so the observational distribution is identical, but \\U is not ΛU+b. Since Theorem 1's proof begins with Lemma 1 to establish an affine reparameterization, the CD-equivalence theorem is not established as written. The authors must either prove a corrected lemma under additional hypotheses or explicitly defer to Zhang et al. (2023) and verify that all of that paper's assumptions hold for the present setting.","section":"Appendix, Lemma 1 / Theorem 1"},{"comment":"Assumption 1 assumes a full-row-rank polynomial decoder, but the implemented decoder is a DAG layer plus two fully connected layers with leaky ReLU activations, which is not a polynomial map. Assumptions 2 and 3 (linear interventional faithfulness and total separation) are also never verified on the CRISPR datasets. Therefore the statement that 'GRACE-VAE inherits the identifiability guarantee and identifies the latent causal graph and intervention targets up to CD-equivalence' does not apply to the architecture and data actually evaluated. This is a load-bearing gap between theory and experiment.","section":"Identifiability guarantee / Implementation details"},{"comment":"The text states: 'Across the three datasets, GRACE-VAE outperforms all baselines on R2 (Table 1).' This is contradicted by Table 1: on NORMAN single-intervention R2, VGAE achieves 0.9493±0.0123 versus GRACE-VAE's 0.9452±0.0108. The claim needs to be qualified: GRACE-VAE outperforms CMVAE and CMVAE-multihot on this metric, and it has the best or comparable RMSE/MMD, but it is not strictly the best on single R2. This overstatement affects the paper's primary empirical contribution.","section":"Results, R1 / Table 1"},{"comment":"The conclusion claims that 'GRACE-VAE yields faithful latent DAGs' and that the experiments demonstrate recovery of the true latent causal graph. However, no structure-learning metric (e.g., SHD, edge AUROC, target recovery accuracy) is reported against a ground-truth latent DAG. The three datasets are real perturbation screens without known latent SCMs, and the reported metrics evaluate prediction of interventional outcomes only. To support the 'recovery' claim, the authors should include synthetic experiments with known latent graphs and report structural accuracy.","section":"Conclusion and Experimental Setup"}],"minor_comments":[{"comment":"There is a typo in the heading ('guanrantee') and in the text: 'leaves the observational distribution PX changed' should read 'leaves the observational distribution PX unchanged' (or 'does not change').","section":"Identifiability guarantee"},{"comment":"The citation Fine and Rosenberger (1997) is a textbook on the fundamental theorem of algebra; it does not support the invoked statement about polynomial automorphisms. If the authors believe such a theorem exists, they should cite a proper source and state the exact conditions.","section":"References"},{"comment":"The hyperparameter selection is described as based on 'R2, MMD, and RMSE' collectively, which is ambiguous. Please specify the selection rule (e.g., a primary metric or a rank-based aggregate).","section":"Implementation details"},{"comment":"The learned DAG in Figure 2 would benefit from a description of how edges are thresholded (if at all) and how edge weights are normalized, so that the visualization is reproducible.","section":"Figure 2"}],"recommendation":"major_revision","confidential_remarks":"The paper's core idea is interesting and the empirical comparison is substantial, but the theoretical appendix contains a clear mathematical error and the main empirical claim is overstated. I recommend major revision. The overlap between the present first author and the cited Zhang et al. (2023) is worth editorial attention: the paper should state explicitly which theoretical results are imported, verify their assumptions, and avoid reproducing an incorrect proof in the appendix. Adding synthetic experiments with known latent graphs would help support the structure-recovery claim."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The short version: the paper's practical contribution is real, but the theoretical guarantee as presented doesn't hold. The proof of Lemma 1 in the appendix invokes a 'fundamental theorem of algebra for polynomial automorphisms' to conclude that polynomial automorphisms of R^p are affine, which is false—(x+y^2,y) is a counterexample. So the CD-equivalence claim is not established here. The architecture also uses an MLP decoder with leaky ReLU, which is not a polynomial map, so Assumption 1 doesn't even apply to the system being evaluated.\n\nWhat's new: the idea of putting a GNN in the variational encoder while keeping the decoder as a latent SCM is a sensible and under-explored combination. The experiments on NORMAN, Replogle-Small, and Replogle-Large are substantial, with held-out double perturbations on NORMAN serving as a genuine generalization test. The ablations (GNN architecture, edge types) are useful, and the parameter comparison against multihot is a fair point.\n\nThe empirical overclaim: Table 1 shows VGAE has higher NORMAN single-intervention R2 (0.9493 vs 0.9452), so the text's claim that GRACE-VAE 'outperforms all baselines on R2' is wrong. It does beat CMVAE and CMVAE-multihot on that metric, and does better on RMSE/MMD, but the sentence should be corrected. The double-intervention result is more convincing.\n\nAnother soft spot: the identifiability section is largely inherited from Zhang et al. 2023, and the first author overlaps; that's not inherently damning, but given the broken lemma reproduction and the untested Assumptions 2 and 3, the theory section needs a serious repair. The authors should either prove a corrected lemma under assumptions that exclude nonlinear polynomial automorphisms, or explicitly state the guarantee only applies under the conditions of Zhang et al. and verify them.\n\nWho this is for: researchers working on causal representation learning with biological network priors will find the empirical recipe interesting. If the theory is fixed, it could be a solid contribution. As it stands, the load-bearing theoretical claim does not hold.\n\nI'd send it to peer review, but the review should insist on the theory being repaired or dropped, the overclaim fixed, and code released. It's a conditional accept at best, not a rejection of the empirical direction.","headline":"GRACE-VAE is a useful empirical architecture for incorporating network context into causal representation learning, but its identifiability proof rests on a false algebraic claim, and the headline empirical claim overstates a baseline comparison.","tokens_in":17567,"tokens_out":2175,"would_cite":false,"duration_ms":21148,"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 graph-aware VAE inherits causal identifiability and improves CRISPR intervention prediction.","keywords":["causal representation learning","latent causal graph","graph neural network","variational autoencoder","soft interventions","identifiability","CRISPR perturbation prediction","network-structured data"],"falsifier":"Run GRACE-VAE on synthetic data generated from a known latent SCM using an MLP decoder with leaky ReLU activations and the same training loss; if the recovered DAG and intervention targets are not equivalent up to permutation, scaling, and shift, the implemented model does not inherit the identifiability guarantee.","tokens_in":16531,"feed_emoji":"🧬","tokens_out":6739,"duration_ms":67455,"temperature":0.7,"pith_summary":"GRACE-VAE is a graph-aware causal discrepancy variational autoencoder for learning latent causal structure from data that come with an observed interaction network, such as pathway and protein-protein associations around gene-expression measurements. The paper's central claim is that inserting a graph neural network into the variational encoder—while leaving the generative decoder as a latent structural causal model—preserves the existing identifiability guarantee: from observational and single-node soft-interventional data, the latent DAG and intervention targets are recoverable up to permutation, scaling, and shift. The practical motivation is CRISPR perturbation screens, where structured biological context improves prediction of gene-expression outcomes after interventions, including double perturbations never seen in training. A sympathetic reader would care because most causal representation learning assumes flat i.i.d. data, whereas real scientific datasets are relational; this work argues the relational structure can be exploited without sacrificing theoretical guarantees.","feed_headline":"Graph-aware VAE recovers latent causal programs from CRISPR data","feed_subtitle":"Adding pathway and protein network context improves prediction of intervention outcomes, including unseen combos.","key_machinery":"The load-bearing design choice is the placement of the graph neural network: GRACE-VAE inserts the GNN in the variational encoder only, so the generative likelihood p(X|U) and the latent SCM over U are identical to the model in Zhang et al. (2023). Identifiability then follows from the existing three-step proof: Lemma 1 identifies the latent vector up to an affine map from the observational distribution; Assumption 2 (linear interventional faithfulness) uses the changed marginal distributions under each intervention to pin down the coordinate permutation; Assumption 3 (total separation) orients every edge of the DAG. The remaining freedom is the CD-equivalence class. The intervention encoder","core_discovery":"The paper's central claim is that causal representation learning from soft interventions remains identifiable when the observed data carry network structure, provided the graph information is injected only into the variational encoder. GRACE-VAE instantiates this idea: a GraphSAGE (or other GNN) encoder processes the observed entities and their auxiliary group nodes; the decoder is an unchanged latent structural causal model with a soft intervention encoder; training couples an ELBO reconstruction term with an MMD alignment term that matches generated counterfactuals to observed interventional distributions. Because the GNN never enters the likelihood, the identifiability theorem from Zhang","pith_inferences":["An implication left implicit is that, since the guarantee concerns only the generative side, the GNN could be replaced by any sufficiently expressive inference network over the network and the identifiability result would still hold.","The paper does not verify that real CRISPR data satisfy the faithfulness and separation assumptions or that the implemented MLP decoder is a polynomial map; a synthetic benchmark with known ground-truth SCM and the exact architecture would reveal whether the implemented model's recoveries are actually CD-equivalent.","A testable extension suggested by the framing: adding graph context that is causally irrelevant (for example, permuted pathway assignments) should not improve intervention prediction; if it does, the gain is likely regularization rather than recovered causal structure.","The same architecture should transfer to other domains with latent interventions and observed relational structure, such as social influence, ecology, or drug combination screens, where network context could act as weak supervision for latent mechanisms."],"forward_implications":["If the claim holds, any observed network with observational plus soft-interventional data can be used to recover latent causal variables and intervention targets without knowing those targets, up to relabeling and scaling.","Structured context yields measurable gains: on NORMAN, single-intervention R² rises from 0.902 with the flat baseline to 0.945 with GRACE-VAE, and double-intervention R² rises from 0.738 to 0.785.","Ablations show that both components matter: removing the causal decoder or removing the graph encoder degrades performance, so the benefit comes from combining network-aware inference with an explicit causal generative model.","The identifiability guarantee is insensitive to the choice of GNN encoder, so future encoder architectures can be swapped in without re-proving identifiability.","Richer pathway and gene topology helps single interventions but shows diminishing or slightly negative returns for unseen double interventions, suggesting context level should be chosen per task."],"supporting_citations":[{"why":"Supplies the causal discrepancy VAE framework, the identifiability theorem with Assumptions 1-3, and the CD-equivalence definition that GRACE-VAE inherits.","marker":"(Zhang et al. 2023)"},{"why":"Provides the GraphSAGE message-passing layer used as the default GNN encoder that injects network structure into the variational inference.","marker":"(Hamilton, Ying, and Leskovec 2017)"},{"why":"Provides the maximum mean discrepancy used in the alignment loss that matches generated counterfactual samples to observed interventional distributions.","marker":"(Gretton et al. 2012)"},{"why":"Source of the REPLOGLE CRISPR perturbation datasets used to evaluate single-intervention prediction.","marker":"(Replogle et al. 2022)"},{"why":"Supplies Reactome pathway knowledge used to build the pathway-gene and pathway-pathway structured context network.","marker":"(Jassal et al. 2020)"},{"why":"Supplies protein-protein interaction data used as structured context in the biological networks.","marker":"(Li et al. 2024)"},{"why":"Justifies the assumption that each latent variable is subject to at least one intervention, needed for identifiability when intervention targets are unknown.","marker":"(Squires et al. 2023)"}],"fun_headline_variants":["Graph-aware VAE recovers causal programs from CRISPR","Network context improves causal variational inference","Predicting unseen intervention combos from network data","Causal VAE with graph structure: identifiability boost"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The identifiability guarantee holds only if the decoder is a polynomial map with full rank and non-empty latent support, and the intervention data satisfy the paper's faithfulness and separation conditions; the implemented MLP decoder and the real biological datasets are not verified against these conditions.","fun_headline_variants_meta":{"raw":{"variants":["Graph-aware VAE recovers causal programs from CRISPR","Network context improves causal variational inference","Predicting unseen intervention combos from network data","Causal VAE with graph structure: identifiability boost"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000304,"raw_usage":{"total_tokens":1552,"prompt_tokens":679,"completion_tokens":873,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":423,"completion_tokens_details":{"reasoning_tokens":822}},"tokens_in":423,"tokens_out":873,"duration_ms":9357,"temperature":1.0,"reasoning_tokens":822,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T12:03:57.529232+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run GRACE-VAE on synthetic data generated from a known latent SCM using an MLP decoder with leaky ReLU activations and the same training loss; if the recovered DAG and intervention targets are not equivalent up to permutation, scaling, and shift, the implemented model does not inherit the identifiability guarantee.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the causal discrepancy VAE framework, the identifiability theorem with Assumptions 1-3, and the CD-equivalence definition that GRACE-VAE inherits."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the GraphSAGE message-passing layer used as the default GNN encoder that injects network structure into the variational inference."},{"cited_title":"M.; Rasch, M","cited_arxiv_id":null,"evidence_quote":"Provides the maximum mean discrepancy used in the alignment loss that matches generated counterfactual samples to observed interventional distributions."},{"cited_title":"M.; Saunders, R","cited_arxiv_id":null,"evidence_quote":"Source of the REPLOGLE CRISPR perturbation datasets used to evaluate single-intervention prediction."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies Reactome pathway knowledge used to build the pathway-gene and pathway-pathway structured context network."},{"cited_title":"S.; and Uhler, C","cited_arxiv_id":null,"evidence_quote":"Justifies the assumption that each latent variable is subject to at least one intervention, needed for identifiability when intervention targets are unknown."}],"review_version":1}