{"id":"2ff1f63c-c148-460b-9e82-087fbd47fc89","arxiv_id":"2605.28659","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":5.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":0,"one_line_summary":"Temporal graph models on pseudotime-resolved gene regulatory networks from single-cell transcriptomics outperform static foundation models like scGPT on gene-expression forecasting, link prediction, and centrality tasks in two mouse developmental datasets.","lead":"The paper tests temporal graph neural networks on evolving gene regulatory networks built from single-cell data snapshots to forecast gene expression, regulatory links, and gene importance over developmental time. If the approach holds, it offers a way to track how cellular programs change dynamically rather than treating cell states as fixed.","discovery_kind":"new_application","skeptic_critique":{"model":"grok-4.3","headline":"Pseudotime ordering and per-snapshot GRN inference may introduce artifacts that favor temporal graph models over static baselines","rationale":"The reader's weakest_assumption directly identifies the same upstream pipeline vulnerability that would have to hold for the central claim to be credible. Because the review was performed on the abstract alone, the same concern remains load-bearing even after the placeholder full-text note; no other internal inconsistency is visible from the supplied material.","tokens_in":1746,"tokens_out":316,"duration_ms":13299,"concrete_test":"Recompute all three tasks on both datasets after replacing the reported pseudotime with (a) a random ordering of the same cells and (b) trajectories from an independent tool (e.g., Slingshot instead of the original method); if the graph-model advantage disappears or reverses under either substitution, the original gains are likely driven by inference artifacts.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The headline claim requires that the reported gains on gene-expression forecasting, link prediction, and centrality tasks arise from genuine modeling of evolving regulatory structure rather than from biases in the upstream pipeline (pseudotime discretization into snapshots followed by GRN reconstruction). If the chosen trajectory inference or network inference method systematically produces temporally coherent but biologically spurious edges, temporal GNNs could exploit those artifacts while static foundation models (scGPT, scFoundation) cannot. The abstract provides no quantitative validation that the inferred networks recover known regulatory interactions or that performance is robust to alternative pseudotime or GRN methods.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.3","summary":"The manuscript proposes a temporal graph-based framework for modeling dynamics in biological systems from single-cell transcriptomic data. Starting from public mouse developmental datasets (erythroid gastrulation and pancreatic endocrinogenesis), the approach infers pseudotime trajectories, discretizes cells into developmental snapshots, reconstructs one gene regulatory network per snapshot, and applies temporal graph neural networks to three tasks: gene-expression forecasting, link prediction, and out-degree centrality prediction. The central claim is that these graph-based models outperform static foundation models such as scGPT and scFoundation, indicating that explicit modeling of evolving regulatory structure supplies information beyond static pretrained representations.","tokens_in":1868,"tokens_out":570,"duration_ms":21006,"significance":"If the empirical claims are substantiated with quantitative results and controls, the work would establish temporal graph learning as a viable complementary paradigm to transformer-based foundation models in single-cell biology. It would demonstrate that explicitly tracking the evolution of gene regulatory networks across pseudotime can improve forecasting of expression, links, and centrality, potentially aiding interpretation of developmental programs and disease progression.","major_comments":[{"comment":"Abstract: The claim that 'graph-based models outperform well-known foundation-model such as scGPT and scFoundation' on gene-expression forecasting, link prediction, and centrality tasks is presented without any quantitative results, error bars, statistical tests, baseline implementation details, or ablation studies. This absence is load-bearing for the central empirical claim and prevents evaluation of whether the reported gains arise from temporal structure modeling.","section":"Abstract"},{"comment":"Abstract (framework description): The pipeline depends on pseudotime trajectory inference followed by per-snapshot GRN reconstruction, yet no validation is supplied that the inferred networks recover known regulatory interactions or that performance remains stable under alternative pseudotime or GRN inference methods. This is load-bearing because systematic artifacts in the upstream discretization or network inference steps could preferentially benefit temporal GNNs while leaving static baselines unaffected.","section":"Abstract"}],"minor_comments":[{"comment":"Abstract: Grammatical error in 'well-known foundation-model such as scGPT' should read 'well-known foundation models such as scGPT'.","section":"Abstract"},{"comment":"Abstract: The phrase 'For link prediction and centrality forecasting, temporal graph learning captures non-trivial regulatory dynamics' is stated without reference to any supporting figure, table, or quantitative metric.","section":"Abstract"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is explicitly labeled work-in-progress and the abstract contains no results section. This may indicate the submission is premature for standard journal review; the editor may wish to confirm whether a full methods/results version exists before proceeding."},"author_rebuttal":{"model":"grok-4.3","summary":"We thank the referee for the constructive comments on our work-in-progress manuscript. The feedback correctly identifies areas where additional evidence is needed to support the central claims. We respond to each major comment below and will incorporate revisions accordingly.","responses":[{"response":"We agree that the abstract currently states the outperformance claim without accompanying quantitative details, which limits evaluation. In the revised manuscript we will augment the abstract with key metrics (including means, standard deviations or error bars, and p-values where appropriate) and ensure the main text supplies full baseline implementation details, ablation studies, and statistical comparisons. These additions will allow direct assessment of whether gains derive from temporal graph modeling.","revision_made":"yes","referee_comment":"[Abstract] Abstract: The claim that 'graph-based models outperform well-known foundation-model such as scGPT and scFoundation' on gene-expression forecasting, link prediction, and centrality tasks is presented without any quantitative results, error bars, statistical tests, baseline implementation details, or ablation studies. This absence is load-bearing for the central empirical claim and prevents evaluation of whether the reported gains arise from temporal structure modeling."},{"response":"We concur that validation of the pseudotime and GRN steps is essential to rule out upstream artifacts. In the revision we will add (i) overlap analyses between inferred edges and curated regulatory databases or ChIP-seq resources and (ii) robustness experiments that repeat the full pipeline with alternative trajectory inference algorithms and GRN methods, reporting performance variance across these choices. This will confirm that the reported advantages are not driven by particular discretization or inference choices.","revision_made":"yes","referee_comment":"[Abstract] Abstract (framework description): The pipeline depends on pseudotime trajectory inference followed by per-snapshot GRN reconstruction, yet no validation is supplied that the inferred networks recover known regulatory interactions or that performance remains stable under alternative pseudotime or GRN inference methods. This is load-bearing because systematic artifacts in the upstream discretization or network inference steps could preferentially benefit temporal GNNs while leaving static baselines unaffected."}],"tokens_in":1463,"tokens_out":447,"duration_ms":22878,"standing_objections":[]},"desk_editor":{"model":"grok-4.3","letter":"The main thing to know is that this work-in-progress paper describes a pipeline that turns single-cell data into pseudotime snapshots, builds a gene regulatory network for each snapshot, and feeds the sequence of graphs into temporal GNNs to forecast expression, links, and centrality. It claims this beats scGPT and scFoundation on two mouse developmental datasets.\n\nWhat is new is the joint use of per-snapshot GRNs and temporal GNNs for those three specific forecasting tasks. The paper does a reasonable job of spelling out why static foundation models might miss the progressive reorganization of regulatory structure during differentiation.\n\nThe soft spots are straightforward. The abstract reports outperformance but gives no quantitative results, baselines, error bars, or ablation details, so there is nothing to evaluate. The stress-test concern about pseudotime discretization and GRN inference creating artifacts that favor the temporal model is valid on current evidence; nothing in the text shows robustness to alternative trajectory or network methods or validates the inferred edges against known biology. No equations or parameter counts are described, which is consistent with an empirical comparison but leaves the work thin.\n\nThis is for readers who track graph-based alternatives to transformer foundation models in single-cell biology. It does not yet have the substance or evidence to justify sending it to peer review.","headline":"The temporal graph pipeline for single-cell dynamics is a sensible framing but the abstract contains no numbers or details, so the outperformance claim cannot be assessed.","tokens_in":2356,"tokens_out":337,"would_cite":false,"duration_ms":22565,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.3","headline":"Temporal graph models of evolving gene regulatory networks outperform static foundation models in forecasting single-cell dynamics.","keywords":["temporal graph neural networks","gene regulatory networks","pseudotime","single-cell transcriptomics","developmental dynamics","link prediction","biological network forecasting"],"falsifier":"Re-running the experiments after randomly permuting the pseudotime order or after replacing the inferred networks with networks drawn from a null model that preserves degree distribution but removes temporal structure; if the temporal-graph advantage disappears under these controls, the central claim is falsified.","tokens_in":2640,"feed_emoji":"🧬","tokens_out":661,"duration_ms":27656,"temperature":0.7,"pith_summary":"The paper tests whether representing cellular development as a sequence of gene regulatory networks, each inferred from a pseudotime-discretized snapshot of single-cell data, allows temporal graph neural networks to predict future biological states more accurately than static pretrained models. It evaluates the approach on two mouse developmental datasets across gene-expression forecasting, regulatory-link prediction, and gene-centrality forecasting. The results indicate that the explicit temporal structure supplies information that static foundation models such as scGPT and scFoundation do not capture. This framing treats development as an evolving graph process rather than a static embedding problem.","feed_headline":"Temporal graphs beat static models at cell development forecasts","feed_subtitle":"Modeling sequences of pseudotime-inferred gene networks improves expression, link, and centrality predictions over foundation models on mous","key_machinery":"Temporal graph neural networks operating on a sequence of pseudotime-resolved gene regulatory networks, where each network encodes regulatory interactions at one developmental stage and the model learns how those interactions change across stages.","core_discovery":"By inferring pseudotime trajectories from single-cell transcriptomes, dividing cells into ordered developmental snapshots, reconstructing one gene regulatory network per snapshot, and feeding the resulting sequence of graphs into temporal graph neural networks, the method produces forecasts of gene expression, link presence, and out-degree centrality that exceed those of static foundation models on erythroid gastrulation and pancreatic endocrinogenesis data. The performance advantage suggests that modeling the temporal evolution of regulatory interactions supplies useful dynamic information beyond what is available from static representations.","pith_inferences":["The same pipeline could be applied to disease trajectories by substituting pathological pseudotime orderings for developmental ones.","Hybrid models that combine large-scale static pretraining with subsequent temporal-graph fine-tuning on developmental snapshots may improve both accuracy and data efficiency.","Because the method separates network inference from temporal modeling, it can incorporate alternative network-reconstruction algorithms without retraining the entire foundation model."],"forward_implications":["Temporal graph learning captures non-trivial regulatory dynamics that static models miss.","The same framework identifies temporally important gene hubs via centrality forecasting.","Temporal graph models constitute a complementary paradigm to current foundation-model approaches in single-cell biology."],"fun_headline_variants":["Temporal graphs forecast cell development through gene networks","Sequences of regulatory networks enable temporal predictions in cells","Forecasting gene expression via evolving regulatory graphs","Temporal graph neural networks for single-cell developmental dynamics"],"cache_read_input_tokens":2112,"weakest_assumption_plain":"Pseudotime trajectories correctly order the true developmental sequence and gene regulatory networks inferred from each snapshot contain biologically meaningful edges rather than artifacts that drive the observed performance differences.","fun_headline_variants_meta":{"raw":{"variants":["Temporal graphs forecast cell development through gene networks","Sequences of regulatory networks enable temporal predictions in cells","Forecasting gene expression via evolving regulatory graphs","Temporal graph neural networks for single-cell developmental dynamics"]},"model":"grok-4.3","cost_usd":0.011017,"raw_usage":{"total_tokens":4798,"prompt_tokens":729,"num_sources_used":0,"completion_tokens":55,"cost_in_usd_ticks":110165500,"prompt_tokens_details":{"text_tokens":729,"audio_tokens":0,"image_tokens":0,"cached_tokens":64},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":4014,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":729,"tokens_out":55,"duration_ms":43770,"temperature":1.0,"reasoning_tokens":4014,"cache_read_input_tokens":64,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-06-29T13:54:04.352760+00:00","model_set":{"reader":"grok-4.3"},"falsifier":"Re-running the experiments after randomly permuting the pseudotime order or after replacing the inferred networks with networks drawn from a null model that preserves degree distribution but removes temporal structure; if the temporal-graph advantage disappears under these controls, the central claim is falsified.","supporting_citations":[],"review_version":1}