{"id":"7c78b50a-b1cb-4a92-b44c-7aee95db63e9","arxiv_id":"2505.16492","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"scIMF couples transformer attention to a McKean-Vlasov stochastic differential equation to predict cell-population dynamics and quantify cell-cell interactions from temporal scRNA-seq snapshots.","lead":"scIMF is a deep generative model that treats cell populations as interacting stochastic systems, using a transformer attention mechanism to learn how cells influence each other from snapshots of single-cell RNA sequencing data. It reports improved held-out trajectory predictions on two in vivo datasets, while a comparison method wins on the in vitro pancreatic dataset.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No interaction-free ablation is performed, so the central claim that explicitly modeling cell-cell interactions improves prediction accuracy is not established; the reported gains could stem from the transformer's capacity or the MV-SDE formulation instead.","rationale":"The reader's verdict of CONDITIONAL is well-founded, and my analysis independently identifies the same missing piece: the paper never ablates the interaction term. However, I frame the load-bearing concern as an attribution problem rather than a functional-form misspecification. The reader's weakest_assumption emphasizes that real cell-cell interactions may be sparse, local, or higher-order, and thus not captured by the softmax attention approximation of Eq. 4. While that is a valid concern about the interpretability and completeness of the interaction model, the more decisive issue for the central claim is that the empirical comparisons do not establish that the interaction term causes the observed accuracy gains. Even if the attention mechanism were perfectly expressive, the paper would still need an ablation to rule out the confounding effect of extra model capacity. The proposed concrete test directly isolates the interaction term while holding all other components fixed. If the ablation shows no significant drop in held-out W1/W2, the central claim collapses; if it shows a clear drop, the claim is supported at least on the in vivo datasets. The Panc results would still need nuanced treatment, but the ablation is the first necessary check. I therefore recommend keeping the verdict CONDITIONAL (no change from the reader), with the condition being the addition of this interaction-free ablation and, ideally, external validation of attention-derived interactions against known biology.","tokens_in":12257,"tokens_out":6437,"duration_ms":54662,"concrete_test":"Train an interaction-free variant of scIMF on the ZB and MEF datasets: keep the same fintra, diffusion coefficient, and dynamic-OT objective, but set finter = 0 (or replace the attention module with a cell-independent MLP of identical parameter count). Evaluate the same held-out tasks reported in Tables 2 and 3 over 5 seeds and compare W1/W2 against the full model. If the interaction-free variant performs within noise of the full scIMF, the central causal claim that modeling cell-cell interactions improves prediction accuracy fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is causal: explicitly modeling cell-cell interactions via the McKean-Vlasov drift (Eq. 2-4) yields more accurate reconstructions of multicellular dynamics. However, scIMF's design differs from every baseline in multiple dimensions simultaneously: it uses a transformer attention module, a neural SDE solver with a dynamic optimal transport objective, and an interaction-dependent drift. The comparisons in Tables 2 and 3 are not controlled ablations; no variant of scIMF with the interaction term disabled (or replaced by a cell-independent function of matched capacity) is evaluated. Therefore, the reported improvements—e.g., ZB hard task W1 = 12.92 vs. scNODE's 15.66—may be attributable to the transformer's representational power, the SDE regularization, or the optimal transport objective, rather than to the mean-field interaction mechanism itself. The Panc results (Tables 4-5), where PI-SDE outperforms scIMF, further weaken the universal claim and are only explained post hoc as a quasi-equilibrium regime. Without an interaction-free ablation, the paper's headline conclusion is unsupported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript introduces scIMF, a generative model for temporal scRNA-seq data that combines a McKean-Vlasov stochastic differential equation (MV-SDE) with a transformer-based attention mechanism. The drift is decomposed into a cell-intrinsic term and an interaction term, with the interaction term approximated by softmax attention over minibatches. The model is trained under a dynamic optimal transport objective and evaluated by predicting held-out time points on zebrafish embryogenesis (ZB), mouse embryonic fibroblast reprogramming (MEF), and pancreatic beta-cell differentiation (Panc) data. The authors report improved Wasserstein distances on ZB and MEF held-out tasks relative to PRESCIENT, MIOFlow, PI-SDE, and scNODE, and claim that the attention scores reveal directional, non-reciprocal cell-cell interactions, especially in the ZB data. The paper argues that explicitly modeling cell-cell interactions leads to more accurate reconstructions of multicellular dynamics.","tokens_in":12576,"tokens_out":2610,"duration_ms":25613,"significance":"If the central claim were fully supported, scIMF would be a useful addition to the trajectory-inference toolkit: it targets a genuine limitation of existing methods by incorporating distributional dependence into single-cell dynamics, and it provides a per-cell, time-varying interaction readout. The manuscript also has concrete strengths: the held-out evaluation protocol is a real extrapolation test, the comparisons cover four established baselines, the datasets are public, and the code is released. However, the headline conclusion that explicitly modeling cell-cell interactions is what drives the improved accuracy is not established, because the method differs from every baseline in several coupled design choices and no interaction-free ablation is reported. The Panc results, where PI-SDE outperforms scIMF on every held-out time point, further qualify the universal claim and are addressed only by a post hoc explanation. The biological claims based on attention scores are also not externally validated. These issues are load-bearing because the paper's stated contribution is precisely the interaction mechanism, not merely a new neural-SDE architecture.","major_comments":[{"comment":"The central claim that \"explicitly modeling cell-cell interactions leads to more accurate representations of multicellular system dynamics\" is not supported by a controlled comparison. scIMF differs from each baseline simultaneously in the transformer attention module, the SDE formulation, and the dynamic optimal transport objective. No ablation is reported in which the interaction term f_inter is disabled, masked, or replaced by a matched-capacity cell-independent function. Without such an ablation, the gains in Tables 2 and 3 could be attributable to the transformer's capacity or to the SDE/OT objective rather than to the mean-field interaction mechanism. I would request an experiment with an scIMF variant in which the attention-based interaction term is removed or replaced (e.g., by a cell-independent MLP of comparable parameter count), evaluated on the same held-out tasks.","section":"Methods, Eq. (3)-(4); Results, Tables 2-3; Discussion"},{"comment":"The universal claim in the Abstract and Introduction is contradicted by the Panc results: PI-SDE achieves the lowest W1 and W2 for all seven held-out time points, while scIMF is second. The paper acknowledges this but explains it post hoc as a quasi-equilibrium regime inferred from the attention matrices themselves. This is circular insofar as the same model outputs are used both to explain the performance gap and to support the biological interpretation. A concrete test would be to quantify the degree of attention asymmetry on Panc versus ZB data and to compare it with an independent measure of equilibrium behavior, or to evaluate whether a symmetric-interaction variant of scIMF matches PI-SDE on Panc. Without such corroboration, the scope of the central claim should be restricted to the ZB and MEF settings.","section":"Results, Tables 4-5; Discussion"},{"comment":"The biological interpretation of attention scores as cell-cell interaction strengths is not externally validated. The paper uses the model's own attention outputs to claim that interactions are non-reciprocal in ZB data and symmetric in Panc data, but softmax(QK^T/sqrt(d))V is a learned mechanism whose identifiability as a communication strength is not established. I suggest a sanity check: compare attention-based interaction patterns against known receptor-ligand pairs, spatial transcriptomic data, or a perturbation experiment; at minimum, test whether the inferred asymmetry is stable across random seeds and whether a permuted-label or null model produces similar block structure.","section":"Results, Figs. 4-6; Eq. (4)"},{"comment":"The statement that scIMF achieves superior performance in both key evaluation metrics is not uniformly true on MEF data: on the hard task, scIMF's W2 (27.73) is higher than scNODE's (25.98). The text acknowledges this, but the Abstract's phrasing that scIMF \"outperforms existing methods in reconstructing gene expression at held-out time points\" is too strong. Please qualify the claim to the settings and metrics where the advantage actually holds.","section":"Results, Table 3"}],"minor_comments":[{"comment":"The text contains typos: \"scGenIMF\" and \"scGeneIMF\" should be \"scIMF\", and \"reuslts\" should be \"results\".","section":"Results, section heading"},{"comment":"The sentence \"we conduct held-one-timepoint-out prediction tasksby sequentially holding out\" has a missing space in \"tasksby\"; please correct.","section":"Results, Panc section"},{"comment":"The notation x_{t0} = rho_{t0} is used, but it would be clearer to write x_{t0} ~ rho_{t0} to indicate that the initial condition is drawn from the distribution, consistent with the later usage in Eq. (9).","section":"Methods, Eq. (2)"},{"comment":"The hyperparameters lambda, sigma, Sinkhorn scaling, and blur are given in the text but are not justified; reporting a brief sensitivity analysis or citing prior work for these choices would improve reproducibility.","section":"Methods, Optimization"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is potentially publishable after substantial revision. The main concern is not the novelty of the MV-SDE-plus-attention framework, but the mismatch between the causal claim and the evidence: the missing interaction-free ablation and the Panc counterexample weaken the headline conclusion. The biological-interpretation section also needs a sharper validation strategy. I would not reject, because the held-out benchmarks and code release are solid and the interaction-free ablation is a feasible, well-scoped addition."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe paper introduces scIMF, a transformer-encoded McKean-Vlasov SDE for temporal scRNA-seq. That's a genuinely new combination for this problem: previous interaction-aware methods work at cluster level or on binarized expression, and neural MV-SDEs haven't been applied to the time-series benchmarks used here. The architecture is clean: a cell-intrinsic drift plus an interaction drift approximated by masked multi-head attention, trained with a dynamic OT objective. Code and data are public, and the held-out prediction tasks on zebrafish and MEF are real extrapolations. On those two in vivo datasets scIMF does beat the baselines on most tasks, often clearly (e.g., ZB hard W1 12.92 vs scNODE 15.66).\n\nThe soft spot is the causal claim. The paper's title and abstract say that explicitly modeling cell-cell interactions improves accuracy. But the design differs from every baseline in several dimensions at once: transformer attention, MV-SDE drift, and the DOT objective. There's no ablation with the interaction term masked or replaced by a cell-independent function of matched capacity. PI-SDE is the closest baseline—also a neural SDE with least action—but it lacks both the transformer and the interaction term, so the comparison can't isolate the mechanism. The Panc dataset makes this worse: PI-SDE wins most of those tasks, and the paper's quasi-equilibrium explanation is post hoc. The attention-as-interaction interpretation is plausible but unvalidated against external data; the paper itself notes the graph is fully connected, so the biological claims rest on softmax weights that aren't checked.\n\nI'd still send this to review. The method is a real contribution, the benchmarks are well done, and the questions it raises about ablation and attribution are exactly what a referee should push on. The needed revision is clear: add an interaction-free ablation, temper the abstract to dataset-dependent performance, and either validate the attention scores or frame them as hypotheses. A serious reader gets value from the architecture and the careful benchmarking; just not from the strongest causal conclusion.\n\nRecommendation: accept for peer review, with the ablation as a required change.","headline":"A genuinely new combination of MV-SDE and transformer attention for temporal scRNA-seq, with strong benchmarks on ZB and MEF; but the claim that the interaction term drives the gains is not isolated by any ablation, and the Panc results undercut the universal statement.","tokens_in":13023,"tokens_out":2531,"would_cite":true,"duration_ms":20726,"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":"Explicitly modeling cell-cell interactions improves reconstruction of held-out gene-expression time points from temporal scRNA-seq data.","keywords":["single-cell RNA sequencing","cell-cell interactions","attention mechanism","neural SDE","transformer","McKean-Vlasov SDE","trajectory inference","time-series scRNA-seq"],"falsifier":"Run the same held-out tasks with the interaction term ablated, modeling only the cell-intrinsic drift while keeping the architecture and training budget identical; if the full model does not show clearly lower Wasserstein distances than the ablation, the claim that interaction modeling drives the improvement is falsified.","tokens_in":12078,"feed_emoji":"🧬","tokens_out":5421,"duration_ms":45306,"temperature":0.7,"pith_summary":"scIMF is a deep generative model that reconstructs how a population of cells changes over time from snapshot scRNA-seq data, where individual cells are destroyed at each time point. Its central claim is that explicitly modeling how each cell's trajectory is influenced by the rest of the population—rather than treating cells as independent—makes predictions of unseen time points more accurate. The model does this with a neural stochastic differential equation whose drift depends on the population distribution through a transformer self-attention layer, so interaction strengths are learned from gene expression alone. On three published datasets, it reports lower Wasserstein distances than four existing methods on most held-out tasks, with the biggest gains when extrapolating to unobserved late time points.","feed_headline":"Cell interaction modeling sharpens single-cell trajectory forecasts","feed_subtitle":"A transformer-driven neural SDE reconstructs held-out time points better than single-cell baselines on three datasets.","key_machinery":"The central object is the McKean-Vlasov stochastic differential equation $dx_t = f(x_t, \\rho_t)\\,dt + \\sigma\\,dW_t$, where the drift $f(x_t, \\rho_t)$ depends on the entire cell-state distribution $\\rho_t = \\mathrm{Law}(x_t)$. The drift is decomposed into an intrinsic single-cell term $f_{\\mathrm{intra}}(x_t)$ and a mean-field interaction term $f_{\\mathrm{inter}}(x_t, \\rho_t) = \\int \\phi(x_t, y_t)\\,\\rho_t(dy_t)$, which is approximated by softmax self-attention over a minibatch of cells. This attention mechanism simultaneously supplies the nonlinear interaction term and yields interpretable, asymmetric cell-to-cell influence scores. The model is trained by optimizing a dynamic optimal transport objective, with a least-action cost regularized by Wasserstein penalties at observed time points, and solved with a NeuralSDE framework.","core_discovery":"The central claim is that adding an explicit, learnable population-dependence term to a single-cell stochastic dynamics model improves the reconstruction of multicellular behavior from temporal scRNA-seq data. scIMF models the cell-state distribution with an interacting mean-field stochastic differential equation, where each cell's drift is the sum of a cell-intrinsic term and an interaction term obtained by integrating pairwise influences over the current population distribution. The interaction term is parameterized by transformer self-attention, so the influence of every other cell on a given cell's motion is learned from data and can be read off as an attention score. On held-out time-point reconstruction tasks across three scRNA-seq datasets, the model reports lower Wasserstein distances than four baselines on most tasks, with the largest gains on extrapolation tasks, and the learned attention matrices are asymmetric, which the authors interpret as non-reciprocal cell-cell communication.","pith_inferences":["If the attention scores truly reflect biological communication, then comparing them with spatial transcriptomics or lineage-tracing constraints would be a direct validation; the paper does not include such a test.","The mean-field integral is approximated from a minibatch of 512 cells, so the learned interaction strengths may depend on sampling; quantifying this sampling variance would clarify how stable the inferred interactions are.","The extrapolation gains could come from the interaction term acting as a regularizer on trajectories rather than from capturing true communication; an ablation study would distinguish these explanations."],"forward_implications":["Held-out time points, including extrapolated late stages, can be reconstructed from snapshot scRNA-seq data with lower Wasserstein error than current single-cell-level baselines.","The estimated attention scores provide time-resolved, directional maps of which cell types influence which, without requiring predefined receptor-ligand pairs.","Non-reciprocal interaction patterns can emerge naturally from the asymmetry of learned attention, consistent with out-of-equilibrium multicellular systems.","In near-equilibrium in vitro systems, simpler single-cell SDE models may remain sufficient, while interaction-aware modeling matters most for branching in vivo development."],"supporting_citations":[{"why":"Supplies the McKean-Vlasov SDE formalism that makes the drift distribution-dependent.","marker":"16–18"},{"why":"Provides the self-attention architecture used to encode cell-cell interactions.","marker":"19"},{"why":"Supplies the NeuralSDE framework used to solve and optimize the stochastic differential equation.","marker":"20, 21"},{"why":"Sets the held-out-task benchmark protocol and provides scNODE, a key baseline.","marker":"12"},{"why":"Provide the PRESCIENT and PI-SDE baselines and the least-action objective.","marker":"10, 11"},{"why":"Provides the zebrafish embryogenesis dataset used for the in vivo benchmarks.","marker":"22"},{"why":"Provides the MEF reprogramming dataset used for the in vivo benchmarks.","marker":"23"},{"why":"Provides the pancreatic beta-cell differentiation dataset used for the in vitro benchmark.","marker":"24"},{"why":"Justifies decomposing the drift into intrinsic and mean-field interaction terms.","marker":"27"},{"why":"Provides the Sinkhorn algorithm used to compute Wasserstein losses.","marker":"30"}],"fun_headline_variants":["Cell interactions improve single-cell dynamics forecasts","Transformer neural SDE learns collective cell dynamics from scRNA-seq","Modeling cell communication enhances single-cell trajectory reconstruction","scIMF uses transformer attention to predict multicellular dynamics"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a cell's fate is influenced by the full population through pairwise, mean-field interactions that the attention mechanism can represent; if real communication is sparse, local, or higher-order, the learned dynamics and the interaction interpretation could be wrong.","fun_headline_variants_meta":{"raw":{"variants":["Cell interactions improve single-cell dynamics forecasts","Transformer neural SDE learns collective cell dynamics from scRNA-seq","Modeling cell communication enhances single-cell trajectory reconstruction","scIMF uses transformer attention to predict multicellular dynamics"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000166,"raw_usage":{"total_tokens":1227,"prompt_tokens":889,"completion_tokens":338,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":505,"completion_tokens_details":{"reasoning_tokens":277}},"tokens_in":505,"tokens_out":338,"duration_ms":3365,"temperature":1.0,"reasoning_tokens":277,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T14:59:47.313007+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same held-out tasks with the interaction term ablated, modeling only the cell-intrinsic drift while keeping the architecture and training budget identical; if the full model does not show clearly lower Wasserstein distances than the ablation, the claim that interaction modeling drives the improvement is falsified.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the self-attention architecture used to encode cell-cell interactions."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Sets the held-out-task benchmark protocol and provides scNODE, a key baseline."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the zebrafish embryogenesis dataset used for the in vivo benchmarks."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the MEF reprogramming dataset used for the in vivo benchmarks."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the pancreatic beta-cell differentiation dataset used for the in vitro benchmark."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Justifies decomposing the drift into intrinsic and mean-field interaction terms."},{"cited_title":"true McKean-Vlasov case","cited_arxiv_id":null,"evidence_quote":"Provides the Sinkhorn algorithm used to compute Wasserstein losses."}],"review_version":1}