{"id":"fc60fda4-7307-4c43-8a4a-a68d2626821b","arxiv_id":"2506.12439","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"SENA-discrepancy-VAE combines a pathway-masked encoder with discrepancy-VAE, giving causal latent factors interpretable as biological process combinations with comparable unseen-perturbation prediction.","lead":"This paper adds a biological pathway mask to a causal representation learning model for single-cell gene perturbation data, so each learned latent factor can be read as a combination of biological processes. It reports prediction quality on unseen double gene knockouts similar to the original model while making the latent factors interpretable.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The identifiability guarantee for discrepancy-VAE does not transfer to SENA, because interventions shift Z while U = Z^T (I-A)^{-1} generally couples all factors; the single-target assumption is violated in the implemented architecture.","rationale":"The reader's weakest assumption identifies the same load-bearing spot, and my reading agrees. I add a sharper formulation: because U is a linear transform of Z, even a one-hot Z-shift generically produces a multi-target U-shift. This is not only a biological limitation; it is an architectural mismatch between the intervention mechanism and the theorem's premise. The paper is transparent about the limitation in the Discussion, and the predictive and interpretability claims are empirically supported by Table 1, the Hits@100/DAR analyses, and the GO-based case studies. The issue is the transfer of formal causal guarantees, not the quality of the experiments. I credit the released code and the external gene-set validation; the condition is addressable via synthetic benchmarks or sparsity checks. Therefore I do not move the reader's CONDITIONAL verdict; the manuscript should be accepted only if the single-target assumption is either shown to hold empirically in the trained model or the causal claims are scaled back to 'predictive, biologically coherent' rather than fully identified causal factors.","tokens_in":27351,"tokens_out":8632,"duration_ms":111356,"concrete_test":"For each trained model (MLP and SENA-delta, lambda=0 and 0.1, latent dim 105), compute the empirical intervention shift in causal-factor space: Delta U_p = E[U^I_p] - E[U_ctrl] using the model's interventional encoder and U = Z^T (I-A)^{-1}. Quantify sparsity via the participation ratio PR_p = (sum_i Delta U_{p,i}^2)^2 / sum_i Delta U_{p,i}^4 and the number of components with |Delta U_{p,i}| > 0.1 max_j |Delta U_{p,j}|. If SENA's shifts are as sparse as the MLP's (one dominant component per perturbation), the single-target violation is empirically mild; if they are dense, the identifiability guarantee does not apply to the trained model and the causal interpretation of factors and edges is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3 and Appendix I define the SENA encoder as alpha = sigma((W⊙M)^T x), with mu_j = alpha^T delta_j^(mu), sigma^2_j = alpha^T delta_j^(sigma), and the causal factors as U = Z^T (I-A)^{-1}. The discrepancy-VAE identifiability theorem (Ahuja et al. 2023; Zhang et al. 2024) requires each intervention to target exactly one latent causal factor U_i. In SENA, however, the interventional encoder (Fig. 1) applies a learnable shift to the meta-pathway activities Z, and because U depends on Z through the full matrix L=(I-A)^{-1}, a shift in a single Z_j generically changes several U_i: Delta U = (Delta Z)^T L has nonzeros wherever L has nonzeros in the corresponding column. Thus the intervention target in U-space is multi-dimensional unless L is monomial, which is neither assumed nor enforced. So the model's own intervention mechanism does not satisfy the condition under which U is identifiable up to permutation and scaling. The Section 3 claim that linear combinations 'seamlessly' reconcile multi-BP perturbation effects with the single-target assumption conflates BP-mixing in the encoder with the intervention target in U-space. The Discussion explicitly concedes that the single-target assumption 'does not allow CRL methods to thoroughly disentangle the interplay between BPs, perturbations and latent factors.' The Appendix I validation (r=1.00) is definitional: Eq. (10) follows algebraically from the definitions of mu, Delta, and L, so it verifies arithmetic, not biological interpretability. The central 'causal latent factors' claim is therefore not formally supported by the cited theorems.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SENA-discrepancy-VAE, an extension of discrepancy-VAE in which the MLP encoder is replaced by a two-layer masked encoder (SENA-delta) that maps gene expression to the activity levels of Gene Ontology biological processes and then linearly maps those activities to the mean and variance of latent meta-pathway variables Z, with causal factors defined as U = Z^T (I-A)^{-1}. The model is trained on unperturbed and single-gene Perturb-seq data from Norman2019 and Wessels2023 and evaluated on prediction of double-gene perturbation distributions (MMD) and reconstruction (MSE). The authors report that SENA-discrepancy-VAE with lambda=0.1 matches or slightly improves on the original discrepancy-VAE at latent dimension 105, while providing pathway-based interpretations of latent factors and a causal graph with apparently plausible edges. The paper also presents DA/DAR analyses suggesting that BPs containing the targeted gene are more differentially activated, and an appendix 'validation' of the linear mapping between pathway activities and latent factors.","tokens_in":27725,"tokens_out":10476,"duration_ms":129568,"significance":"The intended contribution, an interpretable causal representation learning method for Perturb-seq that preserves predictive performance, is valuable and the architectural idea is simple and reusable. The paper provides code, uses two real datasets, and includes extensive ablations; the finding that a constrained, pathway-masked encoder can match a free MLP encoder at high latent dimension is informative. However, the formal causal claim is not fully supported: the transfer of identifiability guarantees is not established, and the empirical interpretability evidence is partly definitional. If the authors can clarify or relax the theoretical claims and add controls for the interpretability metrics, the paper would be a solid contribution to interpretable single-cell modeling.","major_comments":[{"comment":"The manuscript does not establish that the trained SENA-discrepancy-VAE satisfies the single-target intervention assumption under which discrepancy-VAE is identifiable. The interventional branch applies a learnable shift to the meta-pathway activities Z, and with U = Z^T (I-A)^{-1}, a shift in one Z_j changes multiple U_i through the generally non-monomial matrix L. If the intended target is the single U_j whose noise is shifted, this should be stated and the sparsity of the learned shift in U-space should be verified; otherwise the guarantees of Ahuja et al. (2023) and Zhang et al. (2024) do not transfer. The Discussion in Section 7 concedes that the single-target assumption prevents full disentanglement of BPs, perturbations and latent factors, which further weakens the 'principled identifiability' claim in the Introduction.","section":"Section 3, Figure 1, Appendix I"},{"comment":"The 'experimental validation' reporting r=1.00 is circular. E(U_j)=E(Z^T)L_j=mu^T L_j=alpha^T Delta^(mu) L_j is an algebraic identity that follows immediately from the definitions of mu, Delta and L, so the perfect correlation only verifies that the code implements the stated linear algebra; it provides no evidence about biological interpretability. This exercise should be reframed as a software sanity check, and the interpretability claims should rest on the DA/DAR analyses instead.","section":"Appendix I, Eq. (10) and Fig. 5"},{"comment":"The Hits@100 and DAR metrics are partly self-referential. A BP is classified as 'affected' by perturbation p exactly when it contains the targeted gene in the mask M, and the SENA layer computes that BP's activity as a weighted sum that includes the targeted gene's expression. Because a KO changes the targeted gene's expression, a positive differential activation for 'affected' BPs is substantially built into the construction, independent of what the model learns. The paper should include a control, e.g., permuted gene-BP masks or a fixed random-weight SENA layer, to demonstrate that the learned weights add biological signal beyond the mask.","section":"Section 4.3 and Section 6.3"},{"comment":"The biological validity of the inferred causal graph is supported only by a small number of manually selected edges (e.g., factor 15 to 69), and Appendix II shows stability across hyperparameters but not correctness. Without a quantitative comparison to known regulatory interactions or a blinded evaluation, the claim that the model 'recapitulat[es] known biological causal relationships' is anecdotal. This is load-bearing for the causal-interpretation claim and should be addressed with a systematic evaluation.","section":"Section 6.2"}],"minor_comments":[{"comment":"In the Introduction, 'yielding talent factors' should read 'yielding latent factors'.","section":"Introduction"},{"comment":"The heading 'CASUAL REPRESENTATION LEARNING' contains a typo and should be 'CAUSAL REPRESENTATION LEARNING'.","section":"Section 2.1"},{"comment":"Equation (4) uses the indicator function I[.] without defining it; please define the notation or use standard indicator notation.","section":"Eq. (4)"},{"comment":"The caption of Table 1 states 'mean ± variance', but the reported quantities appear to be standard deviations; the caption should be corrected.","section":"Table 1"},{"comment":"The caption of Figure 8 refers to 'Eq.??'; the equation number is missing and should be filled in.","section":"Figure 8 caption"},{"comment":"There are typos in gene names: 'JUNE' in Section 6.2 should be 'JUN', and 'TXB3' in Section 6.3 should be 'TBX3'.","section":"Section 6.2 and Section 6.3"},{"comment":"The dataset is called 'Wessels2023' in the text but the cited paper is Wessels et al. (2022); please unify the dataset label with the reference.","section":"Appendix V"}],"recommendation":"major_revision","confidential_remarks":"The paper is a conference paper (ICLR 2025) and appears to be the camera-ready version. For a journal submission, the authors should be encouraged to either substantially strengthen the theoretical transfer argument or reframe the contribution as an interpretable encoder with heuristic biological validation, since the current abstract overclaims formal causal guarantees. The self-referential interpretability metrics and the circular Appendix I validation should be fixed before reconsideration."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a real engineering contribution with a formal hole in the causal claim. If you read it, read the architecture and the ablation carefully; don't cite the identifiability.\n\nWhat's new: SENA-δ, a masked linear encoder that injects GO pathway membership into the discrepancy-VAE encoder, so each latent factor is a linear combination of pathway activities. That's a sensible way to get a biological vocabulary into CRL, and the authors are honest that they trade representational flexibility for interpretability. On Norman2019, λ=0.1 matches or slightly beats the MLP-encoder discrepancy-VAE on MMD and MSE at 105 latent factors, and the Wessels2023 results are there, though they need higher λ to stay comparable. The ablation study with λ sweep is careful. They ship code and data. That's all real.\n\nThe problems. First, the Appendix I 'validation' with r=1.00 is an identity. E(U_j)=α^T Δ L_j follows from the definitions of μ, Δ, L; running cells through the model and checking the arithmetic confirms the code, not the biology. Presenting it as experimental validation is misleading.\n\nSecond, and more important: the identifiability guarantee doesn't transfer. The theorem requires each intervention to target exactly one causal factor U_i. The architecture's interventional encoder applies a learnable shift to Z, and U = Z^T (I-A)^{-1}. Unless (I-A)^{-1} is monomial, a shift in one Z component changes several U components, so the intervention targets a multi-dimensional set in U-space. The Section 3 argument that linear combinations 'seamlessly' reconcile multi-BP effects with single-target interventions conflates the encoder's gene-to-pathway mixing with the intervention target in U-space. The Discussion concedes the single-target limitation, but the model's own mechanism violates it. That means 'inferring causal latent factors' is not backed by the cited theorems. This is not a minor point: it's the difference between 'causal' and 'interpretable linear encoder'.\n\nThe empirical interpretability evidence (DA/DAR, UMAP, known H2O2-endothelial relationship) is softer but not circular—it uses external GO annotations and shows real signal. The 'first effort' framing is overbroad; earlier biologically informed VAEs and post-hoc enrichment analyses exist, even if they lack CRL guarantees.\n\nBottom line: useful paper, honest about some limits, but the load-bearing causal claim is unsupported. Who's it for: people working on interpretable CRL or perturbation prediction who want a concrete architectural variant and a benchmark. I'd send it to a serious referee—the engineering is solid enough. I'd cite the architecture, not the identifiability.","headline":"SENA-δ is a useful interpretable encoder for discrepancy-VAE with solid predictive benchmarks, but the causal identifiability claim doesn't hold because interventions shift Z and U = Z^T (I-A)^{-1} couples factors.","tokens_in":28309,"tokens_out":3387,"would_cite":true,"duration_ms":39893,"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":"SENA-discrepancy-VAE learns causal latent factors that are readable as combinations of biological pathway activities while matching the predictive accuracy of its non-interpretable baseline on unseen double perturbations.","keywords":["interpretable causal representation learning","discrepancy-VAE","SENA-δ encoder","Perturb-seq","pathway activity","single-cell transcriptomics","intervention prediction","Gene Ontology"],"falsifier":"Train SENA-discrepancy-VAE on a synthetic linear-Gaussian causal model with known single-target interventions and compare the inferred intervention-to-factor assignments and adjacency matrix with the ground truth; if the recovered assignments are not equivalent up to permutation and scaling, the claimed causal interpretation fails. On real data, a direct biological test is to knock out a gene that sits in two pathways assigned to two different latent factors and measure whether the expression shift activates both factors, which the model predicts it should not.","tokens_in":27172,"feed_emoji":"🧬","tokens_out":16671,"duration_ms":166007,"temperature":0.7,"pith_summary":"The paper's claim is that a causal representation learner for single-cell perturbation data can be made interpretable in terms of known biological processes without giving up predictive power. Its model, SENA-discrepancy-VAE, replaces the black-box encoder of discrepancy-VAE with a two-layer masked network that first computes activity levels for a curated set of Gene Ontology pathways, then maps those activities linearly onto the latent causal factors. On the task of predicting gene expression under unseen double-gene perturbations in the Norman2019 K562 dataset, the model achieves a maximum mean discrepancy of 1.5849 for predicted double-perturbation distributions and a control-sample reconstruction error of 0.02134 at 105 latent factors with $\\lambda=0.1$, essentially matching or slightly beating the original non-interpretable model (1.5981 and 0.02152). The paper also reports that the inferred causal graph among latent factors recapitulates known biological relationships, such as hydrogen peroxide production influencing endothelial cell morphogenesis, and that pathway activities containing the knocked-out gene show the strongest differential activation. If this holds, biologically grounded causal representations can be extracted from Perturb-seq data and used to predict and explain the effects of unseen genetic and drug perturbations.","feed_headline":"Interpretable causal model matches black-box perturbation prediction","feed_subtitle":"SENA-discrepancy-VAE exposes pathway-level biology while predicting double-gene knockouts as well as its non-interpretable baseline.","key_machinery":"The load-bearing object is the SENA-$\\delta$ encoder, a two-layer masked MLP. Its first layer maps gene expression $x$ to pathway activities $\\alpha$ using a sparsity mask $M$ defined by $M_{i,k}=1$ if gene $i$ belongs to pathway $k$ and $\\lambda$ otherwise, so each pathway activity is a weighted sum of the expression of its genes with a small allowance for unannotated genes; its second layer maps $\\alpha$ linearly to the mean and variance of the Gaussian exogenous variables $Z_j$, making every latent causal factor a linear combination of pathway activities. This construction reconciles the biological fact that a gene knockout shifts many processes with the CRL requirement, inherited from discrepancy-VAE theory, that each intervention target exactly one latent factor: the single target becomes a composite of pathway activities. The decoder stays polynomial, so the identifiability theorem (latent factors recoverable up to permutation and scaling) continues to apply, and the path matrix $(I-A)^{-1}$ of the learned causal DAG converts meta-pathway activities into the causal pathway archetypes $U$.","core_discovery":"The central discovery is that pathway knowledge can be injected directly into the encoder of a causal representation learner while preserving the identifiability guarantees that make the latent factors causal. The SENA-$\\delta$ encoder computes pathway activities $\\alpha = \\sigma((W \\odot M)^\\top x)$, where the mask $M$ encodes gene--pathway membership, and then parameterizes each latent exogenous variable $Z_j$ as a linear combination of those activities: $\\mu_j = \\alpha^\\top \\delta_j^{(\\mu)}$ and $\\sigma_j^2 = \\alpha^\\top \\delta_j^{(\\sigma)}$. Because each causal factor is $U = Z^\\top(I-A)^{-1}$, every latent factor has an expected value that is a linear readout of pathway activities through the learned causal DAG, so the factors can be read as combinations of biological processes rather than anonymous axes. Empirically, the paper shows this design matches or slightly improves on the non-interpretable discrepancy-VAE when predicting unseen double perturbations, and that the learned factors and causal edges are biologically meaningful, stable across hyperparameters, and validated by differential-activation tests on two Perturb-seq datasets.","pith_inferences":["Because $\\lambda>0$ lets genes outside a pathway contribute to its activity, the learned weights on masked edges are a candidate resource for discovering missing gene--pathway memberships; a gene whose masked weight remains large across perturbations would be a testable, unannotated member of that pathway, an extension the paper does not perform.","The single-target intervention assumption forces overlapping pathway sets to share a latent factor, so the framework would become more general if the identifiability theory were relaxed to interventions that shift a sparse set of factors; the paper itself marks this as the next step.","If pathway activity levels are largely conserved across cell types, the fixed gene--pathway mask could allow causal factors to be aligned across tissues by retraining only the second layer of the encoder, enabling transfer of perturbation-response predictions between cell lines; this is a reader inference, not a paper claim.","The generative structure of the model (sample $Z$, compute $U = Z^\\top(I-A)^{-1}$, pass through the polynomial decoder) can produce counterfactual expression profiles for interventions never observed, pointing toward in silico drug-combination screening; the paper demonstrates double-perturbation prediction but does not frame the model as a screening platform."],"forward_implications":["At 105 latent factors and $\\lambda=0.1$, SENA-discrepancy-VAE predicts unseen double-gene perturbations with MMD 1.5849 on the predicted distributions and a control-sample reconstruction MSE of 0.02134, compared with 1.5981 and 0.02152 for the original MLP-encoder discrepancy-VAE; interpretability is therefore not bought at the cost of predictive accuracy on this benchmark.","The inferred causal graph can be read as relationships between biological processes: the model recovers a path from catechol-containing compound biosynthesis through hydrogen peroxide biosynthesis to endothelial cell morphogenesis, matching known regulatory biology in an unsupervised way.","The pathway activity layer is itself informative: for knockout perturbations, pathways containing the targeted gene show significantly higher differential activation than other pathways (DAR > 1 for nearly all of the 37 evaluated knockouts), so the same encoder can be used to flag which biological processes a perturbation engages.","The meta-pathway activities are significantly associated with high-level Gene Ontology categories in permutation tests, indicating that the data-driven second layer groups specific pathways into broader, biologically coherent programs.","On the Wessels2023 dataset, where perturbations have similar transcriptomic effects, both models collapse all single-gene perturbations onto a single latent factor; this shows the limit of the approach when biological signal is weak, a finding the paper itself reports."],"supporting_citations":[{"why":"Supplies the interventional CRL identifiability theorem (polynomial decoder, single-target interventions, non-collapsing encoder) that grounds the causal interpretation.","marker":"Ahuja et al. (2023)"},{"why":"Provides discrepancy-VAE and the stronger identifiability result (latent factors recoverable up to permutation and scaling) whose encoder SENA-$\\delta$ modifies.","marker":"Zhang et al. (2024)"},{"why":"Provides the K562 Perturb-seq dataset with 105 single-gene and 131 double-gene perturbations used for the main benchmark.","marker":"Norman et al. (2019)"},{"why":"Provides the second CaRPool-seq dataset in THP1 cells used to test generalization when perturbations have weak transcriptomic effects.","marker":"Wessels et al. (2022)"},{"why":"Supplies the Gene Ontology biological processes from which the pathway mask $M$ is built.","marker":"Ashburner et al. (2000)"},{"why":"Supplies the pathway filtering procedure (size, overlap, and replicability criteria) used to select the final biological process set.","marker":"Ruiz-Arenas et al. (2024b)"},{"why":"Provides GEARS, the state-of-the-art multigene perturbation prediction baseline that SENA-discrepancy-VAE is compared against.","marker":"Roohani et al. (2024)"},{"why":"Defines maximum mean discrepancy, the metric that measures how well the model predicts unseen double-perturbation distributions.","marker":"Gretton et al. (2012)"}],"fun_headline_variants":["Pathway-aware causal model matches black-box accuracy","Causal AI that explains itself predicts gene knockouts","Interpretable causal learning decodes pathway logic","Causal model with pathway knowledge beats opaque baseline","SENA-discrepancy-VAE: interpretable causal predictions"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that every perturbation changes exactly one hidden causal factor: the single-target intervention assumption inherited from discrepancy-VAE theory, which the paper itself concedes in the Discussion keeps the method from fully disentangling how biological processes, perturbations, and latent factors relate.","fun_headline_variants_meta":{"raw":{"variants":["Pathway-aware causal model matches black-box accuracy","Causal AI that explains itself predicts gene knockouts","Interpretable causal learning decodes pathway logic","Causal model with pathway knowledge beats opaque baseline","SENA-discrepancy-VAE: interpretable causal predictions"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000166,"raw_usage":{"total_tokens":1278,"prompt_tokens":991,"completion_tokens":287,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":607,"completion_tokens_details":{"reasoning_tokens":213}},"tokens_in":607,"tokens_out":287,"duration_ms":3826,"temperature":1.0,"reasoning_tokens":213,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T00:50:52.427899+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train SENA-discrepancy-VAE on a synthetic linear-Gaussian causal model with known single-target interventions and compare the inferred intervention-to-factor assignments and adjacency matrix with the ground truth; if the recovered assignments are not equivalent up to permutation and scaling, the claimed causal interpretation fails. On real data, a direct biological test is to knock out a gene that sits in two pathways assigned to two different latent factors and measure whether the expression shift activates both factors, which the model predicts it should not.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the second CaRPool-seq dataset in THP1 cells used to test generalization when perturbations have weak transcriptomic effects."}],"review_version":1}