{"id":"9830dea5-c765-4aec-afd9-ce7299af92d8","arxiv_id":"2605.25581","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":5.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":0,"one_line_summary":"Proposes CITE-VAE, a latent dynamical causal VAE with identifiability analysis for single-cell perturbation prediction, claiming better OOD generalization on CRISPR data.","lead":"The paper introduces CITE-VAE, a latent dynamical causal generative model that aims to recover unobserved cellular programs and their time-evolving responses to perturbations from single-cell data. A smart generalist might read it to understand new computational approaches for predicting cell behavior under genetic or drug interventions without exhaustive experiments.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.3","headline":"Identifiability theorem relies on unverified assumptions about perturbation mechanisms and latent dynamics that may fail in real single-cell data","rationale":"The reader's weakest assumption correctly flags the unspecified conditions; the full-text identifiability section is the single place where the claim could break even if the architecture and experiments look sound. This is an internal correctness risk rather than an external-consensus issue.","tokens_in":1786,"tokens_out":337,"duration_ms":12172,"concrete_test":"Extract the precise statement of the identifiability theorem (including all listed assumptions) from the paper; then check whether the real CRISPR perturbation dataset satisfies the key conditions (e.g., by testing if learned latents remain stable under small changes to the assumed dynamics or by verifying injectivity of perturbation effects via held-out interventional predictions). If any core assumption is violated, recompute the OOD metrics with a deliberately misspecified model variant.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim rests on an identifiability analysis showing latent causal variables are recoverable under 'suitable conditions.' These conditions (likely including specific forms of the dynamical system, independent causal mechanisms, and that perturbations act injectively through the latent programs) are not automatically satisfied by single-cell expression data. If the true biological process violates the assumed functional class or the temporal evolution is not Markovian in the chosen latent space, the recovered variables are not guaranteed to be the causal ones, weakening both the OOD generalization argument and the interpretation of CITE-VAE outputs. The Causal-3DIdent experiments only validate under synthetic conditions that match the assumptions by construction.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The paper proposes a latent dynamical causal generative model for single-cell perturbation data that jointly captures latent cellular programs, perturbation-conditioned mechanisms, and temporal evolution. It provides an identifiability analysis claiming that, under suitable conditions, the latent causal variables are recoverable up to standard equivalence classes. Guided by this, the authors develop CITE-VAE and report improved OOD generalization to unseen perturbations on synthetic Causal-3DIdent data and real CRISPR single-cell perturbation datasets compared to baselines.","tokens_in":1929,"tokens_out":535,"duration_ms":17596,"significance":"If the identifiability result is rigorously established and the generalization gains are robust to realistic violations of the modeling assumptions, the work could advance causal modeling for temporal single-cell data by providing a framework that links latent programs to perturbation-driven dynamics. The explicit combination of identifiability analysis with a practical VAE-based implementation is a positive feature.","major_comments":[{"comment":"Identifiability Analysis (likely §3 or equivalent): The claim that latent causal variables are recoverable rests on 'suitable conditions' whose precise statement (e.g., functional form of the dynamical system, independence of causal mechanisms, injectivity of perturbation effects through the latent programs) is not visible in the abstract and must be checked against whether they are plausibly satisfied by single-cell expression data. If these conditions are not verified or shown to be approximately satisfied, the recovery guarantee does not transfer to the real-data experiments and weakens the OOD generalization interpretation.","section":"Identifiability Analysis"},{"comment":"Experiments on Causal-3DIdent (synthetic validation section): The synthetic data is generated under conditions that match the model assumptions by construction; this validates internal consistency but does not address the skeptic concern that the assumed dynamical class or Markovian evolution may fail on real single-cell data. A load-bearing test would require either sensitivity analysis under controlled violations or additional synthetic regimes that break the assumptions while remaining biologically plausible.","section":"Experiments"}],"minor_comments":[{"comment":"The abstract states an identifiability result and improved generalization but contains no equations, proof sketches, or error bars; the full manuscript should ensure these are clearly presented with explicit statements of all assumptions.","section":"Abstract"},{"comment":"Notation for the latent variables, perturbation conditioning, and temporal evolution should be introduced consistently and early to aid readability for readers outside the immediate subfield.","section":"Model Definition"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for the constructive comments, which help clarify the scope of our identifiability claims and the role of the synthetic experiments. We respond point-by-point below and indicate where revisions will be made.","responses":[{"response":"The precise conditions (first-order Markov dynamics, independent causal mechanisms, and injective perturbation mapping through latent programs) are stated in Section 3. They are omitted from the abstract for brevity. We do not claim exact satisfaction on real single-cell data; instead, the CRISPR results demonstrate that the learned model yields improved OOD generalization even when assumptions hold only approximately. We will add a dedicated paragraph in the Discussion section on the biological plausibility of these conditions and the limits of the identifiability transfer.","revision_made":"partial","referee_comment":"[Identifiability Analysis] Identifiability Analysis (likely §3 or equivalent): The claim that latent causal variables are recoverable rests on 'suitable conditions' whose precise statement (e.g., functional form of the dynamical system, independence of causal mechanisms, injectivity of perturbation effects through the latent programs) is not visible in the abstract and must be checked against whether they are plausibly satisfied by single-cell expression data. If these conditions are not verified or shown to be approximately satisfied, the recovery guarantee does not transfer to the real-data experiments and weakens the OOD generalization interpretation."},{"response":"The Causal-3DIdent experiments are intended to confirm that the identifiability result holds when the data-generating process matches the model class. We agree that this leaves open the question of robustness to violations. In the revision we will add a sensitivity analysis that introduces controlled, biologically motivated violations (e.g., mild non-Markovian noise or dependent mechanisms) and report the resulting degradation in OOD performance.","revision_made":"yes","referee_comment":"[Experiments] Experiments on Causal-3DIdent (synthetic validation section): The synthetic data is generated under conditions that match the model assumptions by construction; this validates internal consistency but does not address the skeptic concern that the assumed dynamical class or Markovian evolution may fail on real single-cell data. A load-bearing test would require either sensitivity analysis under controlled violations or additional synthetic regimes that break the assumptions while remaining biologically plausible."}],"tokens_in":1457,"tokens_out":491,"duration_ms":21388,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The main contribution here is treating perturbation effects as acting through unobserved cellular programs whose states change over time, then recovering those programs with an identifiability guarantee. This sits between static latent causal models and ordinary temporal predictors, and they implement it as CITE-VAE.\n\nThe synthetic experiments on Causal-3DIdent line up with the theory because the data is generated to match the assumptions. On the CRISPR perturbation datasets the method improves generalization to unseen interventions over the baselines they compare against.\n\nThe soft spot is the identifiability analysis. It only holds under suitable conditions on the dynamics and the way perturbations enter the latent space, and those conditions are not shown to be realistic for actual single-cell expression profiles. If the true process is not Markovian in the chosen latent space or the perturbation mapping is not injective, the recovered variables lose their causal interpretation and the OOD claim weakens. The real-data gains could come from other modeling choices rather than the causal structure.\n\nThis is aimed at people building generative models for biological perturbation data who want both prediction and some level of causal grounding. A reader already working on identifiability in dynamical systems or single-cell causal inference would get the most out of it.\n\nSend it to peer review. The framing is coherent and the experiments are pointed in the right direction, but the assumptions and the source of the real-data improvement need checking.","headline":"The paper frames single-cell perturbation responses through recoverable latent causal programs that evolve over time, but the identifiability result depends on conditions that may not hold in real data.","tokens_in":2455,"tokens_out":360,"would_cite":false,"duration_ms":20579,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"A latent dynamical causal generative model recovers cellular programs and their perturbation-driven dynamics from single-cell data up to standard equivalence classes.","keywords":["single-cell perturbation","causal generative model","latent dynamical processes","identifiability","perturbation prediction","CRISPR","temporal evolution","out-of-distribution generalization"],"falsifier":"A dataset in which the same observed expression trajectories arise from multiple distinct latent program trajectories that cannot be distinguished even with full knowledge of the perturbation schedule and time points.","tokens_in":2688,"feed_emoji":"🧬","tokens_out":702,"duration_ms":13346,"temperature":0.7,"pith_summary":"The paper argues that perturbation effects in single-cell experiments act through unobserved cellular programs whose states evolve over time and produce the observed gene expression profiles. It proposes a generative model that jointly represents these latent programs, the mechanisms conditioned on specific perturbations, and their temporal evolution, backed by an identifiability result that the latent causal variables can be recovered under suitable conditions. The authors then introduce CITE-VAE, a learning framework guided by that analysis, and show it improves generalization to unseen perturbations on both controlled simulations and real CRISPR perturbation datasets. A sympathetic reader would care because accurate recovery of these latent processes could support mechanistic interpretation and reliable prediction of how cells respond to new interventions without requiring exhaustive experimental testing.","feed_headline":"Latent causal model recovers cell programs from perturbation responses","feed_subtitle":"Identifiability result and CITE-VAE framework improve prediction of unseen interventions by tracking evolving unobserved states.","key_machinery":"The latent dynamical causal generative model, which encodes unobserved cellular programs evolving under perturbation-conditioned mechanisms to produce observed expression profiles.","core_discovery":"We propose a latent dynamical causal generative model for single-cell perturbation data that jointly captures latent cellular programs, perturbation-conditioned mechanisms, and temporal evolution. We further provide an identifiability analysis showing that, under suitable conditions, the latent causal variables are recoverable up to standard equivalence classes. Guided by this analysis, we develop CITE-VAE, a learning framework for recovering latent cellular programs and their perturbation-driven dynamics from single-cell sequencing data.","pith_inferences":["If the latent programs correspond to biologically meaningful modules, the recovered dynamics could guide hypothesis generation for follow-up wet-lab experiments on specific pathways.","The approach could extend to multi-perturbation or combinatorial intervention settings where interactions between perturbations need to be modeled through shared latent states.","Failure of identifiability in practice would likely manifest as inconsistent predictions across different random seeds or data subsamples, offering a diagnostic for when the modeling assumptions are violated."],"forward_implications":["The model supports out-of-distribution generalization to unseen perturbations by recovering the underlying causal mechanisms rather than fitting static associations.","Temporal evolution of latent programs can be tracked explicitly, allowing predictions of how cellular responses unfold after an intervention.","Identifiability up to equivalence classes provides a theoretical basis for interpreting the recovered variables as representations of cellular programs.","CITE-VAE can be applied to CRISPR-based single-cell data to achieve better predictive performance than existing baselines on real perturbation responses."],"fun_headline_variants":["Dynamical causal model recovers latent cell programs","CITE-VAE tracks evolving causal states in perturbed cells","Identifiable latent dynamics predict single-cell OOD responses","Latent causal generative model for temporal perturbation data","Identifiability result guides CITE-VAE on cell perturbation dynamics"],"cache_read_input_tokens":2112,"weakest_assumption_plain":"Perturbation effects act through unobserved cellular programs whose states evolve over time in a manner recoverable from observed expression profiles.","fun_headline_variants_meta":{"raw":{"variants":["Dynamical causal model recovers latent cell programs","CITE-VAE tracks evolving causal states in perturbed cells","Identifiable latent dynamics predict single-cell OOD responses","Latent causal generative model for temporal perturbation data","Identifiability result guides CITE-VAE on cell perturbation dynamics"]},"model":"grok-4.3","cost_usd":0.006165,"raw_usage":{"total_tokens":2929,"prompt_tokens":711,"num_sources_used":0,"completion_tokens":68,"cost_in_usd_ticks":61649500,"prompt_tokens_details":{"text_tokens":711,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":2150,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":711,"tokens_out":68,"duration_ms":20104,"temperature":1.0,"reasoning_tokens":2150,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-06-29T23:15:45.892833+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"A dataset in which the same observed expression trajectories arise from multiple distinct latent program trajectories that cannot be distinguished even with full knowledge of the perturbation schedule and time points.","supporting_citations":[],"review_version":1}