{"id":"d7cf57f0-80b1-4d56-9ef2-5fb2370fff89","arxiv_id":"2506.13344","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A conditional graph diffusion model with Laplacian positional encodings and spectral adversarial edge-weight perturbations reports improved MMD and Wasserstein metrics on four scRNA-seq datasets.","lead":"LapDDPM is a graph-based diffusion model that generates synthetic single-cell RNA-sequencing data, conditioned on cell type, using a graph encoder with a new spectral perturbation of edge weights. The authors report lower distribution-matching errors than scVI, scDiffusion, CFGen, and scGAN on four scRNA-seq datasets, but the promised robustness experiments are missing from the text.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper's distinctive robustness claim is never tested: Section 2.4 promises adversarial-graph experiments in Section 4, but Section 4 only evaluates clean graphs, so the 'robust' benchmark is unsupported.","rationale":"The reader's verdict is CONDITIONAL with moderate confidence, and the rationale notes the missing robustness evaluation as one reason. However, the reader's weakest_assumption field frames the primary risk as the fidelity of the k-NN graph and the sufficiency of MMD/WD metrics. My stress-test identifies a more specific, more load-bearing gap: the unique selling point of the method—spectral adversarial perturbation for robustness—is never tested at all. This is not a question of metric validity; it is a missing experiment that the paper itself promises. Without it, the robustness claim is not merely weaker, it is unevidenced. The clean-graph performance in Section 4 cannot substitute, because robustness to structural variation requires perturbing the graph at evaluation time. The ablation in Table 3 is useful for showing the perturbation helps clean-graph fidelity, but it does not address the stated robustness motivation. I therefore agree with the CONDITIONAL verdict: the generative performance claim may hold, but the robustness claim needs a dedicated test. I would not escalate to REJECT because the core generative idea is coherent and the clean-graph results are plausible; the missing evidence is fixable. I mark agreement as 'partial' because the reader mentioned the missing robustness evaluation in the rationale but placed the weakest-assumption emphasis elsewhere.","tokens_in":13704,"tokens_out":2791,"duration_ms":30416,"concrete_test":"Run the adversarial robustness experiment promised in Section 2.4 on PBMC3K: after training LapDDPM and the no-spectral-perturbation ablation, poison the test-time k-NN graph by flipping edges (random, DICE, GF-Attack) at perturbation rates from 5% to 20%, then recompute conditional MMD and 2-Wasserstein distances. If the full model does not degrade significantly less than the ablation (preferably also less than baselines), the central robustness claim fails; if it does, the claim is provisionally supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central novel contribution is the spectral adversarial perturbation of graph edge weights, which the Abstract and Section 3.3 state 'boosts robustness against structural variations.' Section 2.4 explicitly promises an evaluation in Section 4 against adversarially poisoned input graphs using random attacks, DICE, GF-Attack, and Mettack. However, Section 4 (Experiments) contains only MMD and 2-Wasserstein distances computed on the original, unperturbed k-NN graphs. No attack, no graph perturbation, no robustness metric appears anywhere in the results. The ablation in Table 3 shows the full model has lower MMD/WD on clean graphs than the variant without spectral perturbation, but that is a quality improvement, not evidence of robustness: it does not measure behavior under structural noise. Because the spectral perturbation is the load-bearing element that differentiates LapDDPM from prior graph diffusion models, the headline claim that LapDDPM is a 'robust tool' and 'sets a new benchmark' is not supported by the presented evidence. A model that performs well only on clean graphs may be fragile to realistic cell-cell graph noise; the paper's own promise to demonstrate robustness is unfulfilled.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"LapDDPM is a conditional graph diffusion model for generating synthetic scRNA-seq data. The method builds a k-NN graph on PCA-reduced, log-normalized gene expression, augments cell features with Laplacian positional encodings, and trains a latent score-based diffusion model conditioned on cell-type or tissue labels. A spectral adversarial perturbation module modifies graph edge weights based on the principal eigenvector of the adjacency matrix during encoder training, ostensibly to improve robustness. The paper evaluates conditional and unconditional generation on PBMC3K, Dentate Gyrus, Tabula Muris, and HLCA using RBF-kernel MMD and 2-Wasserstein distances in a 30-dimensional PCA projection, and reports ablations of the main components.","tokens_in":13936,"tokens_out":7158,"duration_ms":63947,"significance":"If the proposed robustness mechanism worked as intended, the work would be a useful contribution to conditional scRNA-seq generation, combining graph structure with diffusion models. The paper offers a broad set of baselines and ablations, and the clean-graph results are competitive. However, the central novelty—spectral adversarial perturbation for robustness—is never directly evaluated, and the biological plausibility claim rests on distribution distances alone. As it stands, the evidence does not support the abstract's claims of a 'robust tool' or a 'new benchmark'; the paper would become significantly stronger if the authors supplied the promised adversarial experiments and biological validation.","major_comments":[{"comment":"Section 2.4 promises an evaluation in Section 4 of robustness against structural attacks (random, DICE, GF-Attack, Mettack), but Section 4 contains no such experiments; Table 1 and Table 3 report only clean-graph MMD and 2-Wasserstein distances. Since the spectral perturbation is the paper's distinguishing contribution, the robustness claim is unsupported. Moreover, because the graph is used only during encoder training and not at generation time, the mechanism by which this training-time perturbation would confer test-time robustness is not explained; the authors should add the promised evaluation or revise the claims.","section":"Section 2.4, Section 4, Table 3"},{"comment":"The quantitative evaluation is limited to RBF-kernel MMD and 2-Wasserstein distance on 30-dimensional PCA projections. No biological validation (e.g., marker-gene expression, cluster structure, differential expression, or downstream task performance) is provided, so the claims of generating 'biologically-plausible, cell-type-specific samples' are not substantiated.","section":"Section 4.1"},{"comment":"Several entries show overlapping error bars with the strongest baseline, such as Tabula Muris conditional MMD (LapDDPM 0.19±0.02 vs CFGen 0.19±0.02) and Dentate Gyrus conditional MMD (LapDDPM 1.04±0.08 vs CFGen 1.12±0.04). No significance tests are reported over the 10 runs, so the statement that LapDDPM 'consistently' outperforms baselines is not fully supported.","section":"Table 1"}],"minor_comments":[{"comment":"The text calls wdiff, wKL, and wrec 'learnable weights,' but Algorithm 1 lists them as fixed input hyperparameters; please clarify whether they are optimized during training or hand-tuned.","section":"Section 3.4"},{"comment":"Mettack is cited as (Gosch et al., 2024b), but Mettack originates from Zügner et al. (2018), which is already cited in the same paragraph; please correct this reference.","section":"Section 2.4"},{"comment":"The abbreviations 'C-CFG EN' and 'U-CFG EN' are unclear; please rename them to 'CFGen (conditional)' and 'CFGen (unconditional)' for readability.","section":"Table 1"},{"comment":"The text states that the ablation 'assesses ... robustness to structural variations,' but Table 3 only reports clean-graph MMD/WD; the wording should match the actual experiment.","section":"Section 5.2, Table 3"},{"comment":"The paper does not state whether code is publicly available; providing a repository would improve reproducibility.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The missing robustness evaluation is the paper's main weakness, and it is strongly tied to the claimed contribution. The authors themselves promise these experiments in Section 2.4, so adding them is feasible and should be required. If the authors cannot provide the adversarial experiments, the robustness claim would need to be removed or substantially weakened, which might change the novelty assessment."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is a plausible but incremental combination of known pieces for conditional scRNA-seq generation, and its headline robustness claim is not actually tested anywhere in the results. That is the thing to know.\n\nWhat's new and good: The architecture itself — k-NN graph with Laplacian positional encodings, a ChebConv encoder, a latent VP-SDE diffusion model, and a Poisson decoder — is sensible, and each component gets an ablation. The ablations show that removing LPEs, input masking, or the spectral perturbation degrades MMD/WD on clean graphs. On four datasets, the model often has lower MMD and 2-Wasserstein distances than scVI, scDiffusion, and CFGen for conditional generation, though the wins are not universal (e.g., on Tabula Muris conditional MMD, LapDDPM ties CFGen; on unconditional MMD, CFGen is actually better). The complexity analysis is a nice touch.\n\nWhere the paper falls down: The distinctive contribution is the spectral adversarial perturbation on edge weights, and the abstract and Section 3.3 claim it \"boosts robustness against structural variations.\" Section 2.4 promises an evaluation against random attacks, DICE, GF-Attack, and Mettack. That evaluation never appears. Section 4 only reports metrics on the original, unperturbed k-NN graphs. Table 3 shows the full model has lower MMD/WD than the variant without perturbation, but that is a clean-quality improvement, not evidence of robustness under structural noise. So the central technical novelty is effectively untested. This is a load-bearing gap, not a cosmetic one.\n\nThe evaluation is also thin on the biological side: only distribution distances in a 30-dimensional PCA projection, no marker gene checks, no downstream task validation. Some error bars overlap across baselines, and the \"consistently superior\" phrasing in the abstract overstates the table. There's also a citation slip: Mettack is attributed to Gosch et al. 2024b, but Mettack is from Zügner et al. 2018, and the two Gosch entries appear to be the same paper. No code or hyperparameters are provided, which makes reproduction harder.\n\nWho this is for: Someone working on generative models for single-cell data might find this a useful baseline if the robustness gap gets fixed. As it stands, it's a reasonable workshop-level paper with a clear, fixable hole. I'd support sending it to peer review with major revision: add the promised robustness experiments, strengthen the biological validation, and release the code. Without those, the \"robust\" claim shouldn't stand.","headline":"A sensible but incremental scRNA-seq generator whose central robustness claim is promised but never tested; the clean-graph results are decent, but the paper needs major revision.","tokens_in":14452,"tokens_out":3233,"would_cite":false,"duration_ms":27848,"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":"LapDDPM, a conditional graph diffusion model for scRNA-seq generation, claims a new benchmark in matching real single-cell distributions.","keywords":["single-cell RNA-seq generation","conditional diffusion models","graph neural networks","Laplacian positional encodings","spectral adversarial perturbations","maximum mean discrepancy","Wasserstein distance","synthetic single-cell data"],"falsifier":"Compute marker-gene recall and cell-type separation on generated profiles: if synthetic cells for a known type do not express its canonical markers, or if their clusters overlap real cells of other types, the paper's claim of biologically plausible cell-type-specific generation is falsified. Similarly, a graph attack that randomly rewires edges or removes hub nodes would falsify the robustness claim if generation quality collapses.","tokens_in":13488,"feed_emoji":"🧬","tokens_out":7625,"duration_ms":65537,"temperature":0.7,"pith_summary":"LapDDPM is a generative model for single-cell RNA sequencing data that aims to produce synthetic count matrices matching real cells, with generation controlled by cell-type labels. The paper's central claim is that combining a graph representation of cell–cell similarity with a latent score-based diffusion model, plus a spectral adversarial perturbation of graph edge weights during training, yields higher-fidelity conditional generation than existing scRNA-seq generators. On four datasets spanning roughly 2,600 to 585,000 cells, LapDDPM reports lower RBF-kernel MMD and 2-Wasserstein distances to held-out real cells than scVI, scDiffusion, and CFGen in conditional mode, and than scGAN in unconditional mode. Ablations on PBMC3K attribute part of the gain to each component: spectral adversarial perturbations, input gene masking, and Laplacian positional encodings. If these results hold, the method offers a scalable way to synthesize cell-type-specific scRNA-seq data for data augmentation, benchmarking, and in-silico experiments.","feed_headline":"Conditional graph diffusion wins on single-cell RNA benchmarks","feed_subtitle":"Generated cells match real distributions on four datasets; each graph and diffusion component earns its place.","key_machinery":"The load-bearing object is the graph representation of the cell population: a k-NN graph built on PCA-reduced log-normalized expression, enriched by Laplacian positional encodings derived from the normalized graph Laplacian. The spectral encoder uses Chebyshev convolutions to embed nodes, and a conditional variance-preserving score-based diffusion model generates latent states that a Poisson decoder turns into count data. The distinctive mechanism is the spectral adversarial perturbation: a perturbation to edge weights is set proportional to products of principal-eigenvector components $v_i v_j$ of the adjacency matrix, so training sees graphs whose dominant spectral modes have been altered; this is what the paper claims supplies robustness to structural variation.","core_discovery":"LapDDPM is a conditional generative model that maps gene expression counts into a latent space through a Chebyshev graph convolutional encoder operating on a k-NN cell graph augmented with Laplacian positional encodings, then learns a score-based diffusion process in that latent space conditioned on cell-type labels, and decodes samples back into gene expression rates via a Poisson likelihood. The paper reports that this system achieves lower RBF-kernel MMD and 2-Wasserstein distances to held-out real cells than scVI, scDiffusion, and CFGen for conditional generation, and lower unconditional distances than scGAN, across PBMC3K, Dentate Gyrus, Tabula Muris, and HLCA. Ablations indicate that removing the spectral adversarial perturbation, the input gene masking, or the Laplacian positional encodings each degrades these metrics, which the paper reads as evidence that the components contribute to fidelity and robustness.","pith_inferences":["Because the paper evaluates only distribution distances in PCA space, the more consequential claim of biological plausibility remains untested; marker-gene, trajectory, and differential-expression checks would be the natural next experiments.","The spectral adversarial perturbation has the flavor of a spectral data augmentation and could plausibly transfer to other graph-conditioned generative tasks beyond single cells, such as spatial transcriptomics or graph-based drug response modeling, whenever the graph is noisy.","If the k-NN graph is itself a noisy estimate of cell state, the robustness conferred by training on perturbed graphs may also make the encoder less sensitive to the choice of k and PCA dimension, which would simplify practical deployment."],"forward_implications":["Conditional generation can supply synthetic cells for rare cell types whose real counts are too low for training or testing.","The latent-space diffusion setup with a Poisson decoder gives a recipe for generating count matrices rather than normalized expression only, which is what many downstream single-cell tools expect.","The ablations imply each architectural choice (LPEs, input masking, spectral perturbation) is doing measurable work, so future generators in this line should keep or replace them deliberately.","Because preprocessing, training, and generation scale approximately linearly in the number of cells, the method is positioned for atlas-scale datasets such as HLCA.","Robustness to perturbed edge weights means the generator could tolerate noisy cell graphs from fresh experiments or different preprocessing pipelines."],"supporting_citations":[{"why":"Supplies the denoising diffusion probabilistic model framework that the latent diffusion component builds on.","marker":"Ho et al. 2020"},{"why":"Supplies the score-based SDE formulation used for the forward and reverse diffusion processes.","marker":"Song & Ermon 2020"},{"why":"Provides scVI, the main VAE baseline for conditional scRNA-seq generation that LapDDPM compares against.","marker":"Gayoso et al. 2021"},{"why":"Provides scDiffusion, the continuous latent-diffusion baseline for scRNA-seq generation.","marker":"Luo et al. 2024"},{"why":"Provides CFGen, the flow-based conditional baseline used in both conditional and unconditional comparisons.","marker":"Palma et al. 2025"},{"why":"Provides scGAN, the unconditional GAN baseline that LapDDPM reports beating in unconditional mode.","marker":"Marouf et al. 2020"},{"why":"Supplies the RBF-kernel MMD metric used to measure distributional fidelity between generated and real cells.","marker":"Borgwardt et al. 2006"},{"why":"Supplies the PBMC3K dataset used for the main quantitative comparison and ablation study.","marker":"10x Genomics 2017"},{"why":"Supplies the Human Lung Cell Atlas, the largest and most heterogeneous evaluation dataset.","marker":"Sikkema et al. 2023"}],"fun_headline_variants":["LapDDPM: spectral adversarial diffusion for scRNA-seq","Graph diffusion with spectral noise wins on cell benchmarks","Robust conditional graph diffusion for single-cell RNA","Adversarial spectral perturbations sharpen cell-type generation"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central claim stands on the assumption that the nearest-neighbour graph built from the leading principal components of log-normalized expression faithfully represents cell–cell relationships, and that matching real data on two distribution-distance metrics in a 30-dimensional projection is enough to certify biological plausibility and robustness.","fun_headline_variants_meta":{"raw":{"variants":["LapDDPM: spectral adversarial diffusion for scRNA-seq","Graph diffusion with spectral noise wins on cell benchmarks","Robust conditional graph diffusion for single-cell RNA","Adversarial spectral perturbations sharpen cell-type generation"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000259,"raw_usage":{"total_tokens":1586,"prompt_tokens":948,"completion_tokens":638,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":564,"completion_tokens_details":{"reasoning_tokens":576}},"tokens_in":564,"tokens_out":638,"duration_ms":6230,"temperature":1.0,"reasoning_tokens":576,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:04:01.851273+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute marker-gene recall and cell-type separation on generated profiles: if synthetic cells for a known type do not express its canonical markers, or if their clusters overlap real cells of other types, the paper's claim of biologically plausible cell-type-specific generation is falsified. Similarly, a graph attack that randomly rewires edges or removes hub nodes would falsify the robustness claim if generation quality collapses.","supporting_citations":[{"cited_title":"and Ermon, S","cited_arxiv_id":null,"evidence_quote":"Supplies the score-based SDE formulation used for the forward and reverse diffusion processes."},{"cited_title":"I., Stegle, O., Theis, F","cited_arxiv_id":null,"evidence_quote":"Provides scVI, the main VAE baseline for conditional scRNA-seq generation that LapDDPM compares against."},{"cited_title":"Learning Dynamics from Multicellular Graphs with Deep Neural Networks","cited_arxiv_id":"2401.12196","evidence_quote":"Provides scDiffusion, the continuous latent-diffusion baseline for scRNA-seq generation."},{"cited_title":"o rnemann, J., Gribov, E., He, B., K \\","cited_arxiv_id":null,"evidence_quote":"Provides scGAN, the unconditional GAN baseline that LapDDPM reports beating in unconditional mode."},{"cited_title":"M., Gretton, A., Sch \\\"o lkopf, B., Smola, A","cited_arxiv_id":null,"evidence_quote":"Supplies the RBF-kernel MMD metric used to measure distributional fidelity between generated and real cells."},{"cited_title":"PBMC3K dataset, 2017","cited_arxiv_id":null,"evidence_quote":"Supplies the PBMC3K dataset used for the main quantitative comparison and ablation study."},{"cited_title":"C., Gillett, T","cited_arxiv_id":null,"evidence_quote":"Supplies the Human Lung Cell Atlas, the largest and most heterogeneous evaluation dataset."}],"review_version":1}