{"id":"703933ed-cca5-427d-b18c-ff1d7b72fb10","arxiv_id":"2507.02902","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A diffusion model with multi-resolution conditioning and channel-wise attention achieves state-of-the-art accuracy for protein and gene imputation in spatial and single-cell biological data.","lead":"This paper introduces a diffusion model that predicts missing protein or gene channels in multi-channel biological data, such as tissue images and single cells, from whatever subset of channels is observed, and reports state-of-the-art correlations on imaging mass cytometry and CITE-seq benchmarks. If the results hold, researchers could measure fewer markers and impute the rest, cutting cost and enabling merging of datasets with different marker panels.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The arbitrary-subset conditioning claim rests on an untested random-masking assumption; only leave-one-out and full-observation configurations are evaluated.","rationale":"The reader's weakest assumption is the same one I would flag: random masking is assumed to provide informative gradients for all channel subsets, but the experiments only test two extremes. I agree with the reader's conditional verdict. The model architecture is plausible and the reported numbers are consistently favorable, but the central promise of arbitrary conditioning is exactly the part left unmeasured. The missing check is not a formality: for high-dimensional panels, most subsets are astronomically rare under Bernoulli masking even when each subset has nonzero probability, so the amortized estimator's behavior on unseen configurations is genuinely open. The paper deserves credit for proposing a masking-based amortized conditioning scheme and for attempting cross-dataset union/intersection comparisons, but that does not substitute for a direct arbitrary-subset evaluation. The ablation table inconsistency is secondary; I mention it because the instructions require flagging such passages, not because it changes the verdict. Overall, the appropriate disposition remains conditional: the method is promising and the core idea is coherent, but the strongest advertised capability needs direct empirical support before acceptance.","tokens_in":11706,"tokens_out":3983,"duration_ms":42984,"concrete_test":"Run the trained model on the lung IMC dataset with random test-time observed subsets drawn at observed fractions 5%, 25%, 50%, 75%, and 100% (a few draws per fraction), reporting per-protein Pearson r with standard errors, alongside the leave-one-out and full-observation numbers from Table 2. Also report the training masking probability p; if p is not 0.5, retrain under p=0.5 before drawing conclusions. If average correlation drops sharply for low or non-contiguous observed fractions, the arbitrary-subset claim is not supported; if performance degrades gracefully across all fractions, the claim is provisionally validated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is test-time controllability over arbitrary subsets of observed channels (Section 1). This rests on the random-masking assumption in Section 3.2 that partially observed structured data provide informative gradients for every conditional configuration, formalized only as an amortized objective (Eq. 5), with the authors explicitly noting it is not formally addressed. Section 5.2's spatial experiments cover only two configurations: leave-one-out single-channel imputation and full multi-channel output with the complete observed set; the cross-dataset experiment always predicts the intersection set. No experiment samples random subsets of channels at test time, varies the observed fraction, or probes rare combinations (e.g., 1 of 43 channels observed). If masking probability p is not 0.5, some subsets are severely underrepresented during training; p is not reported. Even with p=0.5, the number of channel subsets is exponential, so in 43-channel IMC data most combinations are never seen, and the amortized estimator's performance on them is exactly what the paper promises but does not demonstrate. A secondary inconsistency—Table 3's caption says breast cancer while Section 5.3's text says lung cancer—does not refute the main claim but further undermines confidence in the reported ablations.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a conditional diffusion framework for multi-channel biological data. The method combines hierarchical feature injection from spatially aligned observed channels, channel-wise attention in latent and output spaces, and a random channel-masking training strategy intended to amortize over arbitrary subsets of observed channels. Experiments cover gene-to-protein prediction on four CITE-seq datasets and protein imputation on two IMC datasets, plus a cross-dataset union/intersection study and ablations of the proposed architectural components. The paper reports state-of-the-art Pearson correlations in both the single-cell and spatial settings and describes a one-step SiD-distilled variant for faster inference.","tokens_in":11973,"tokens_out":8442,"duration_ms":90838,"significance":"The modeling idea is attractive and the problem is practically important: biological profiling panels frequently have missing or unmeasured channels, and a single amortized model that can condition on many different subsets would be a useful tool. The paper is clearly written, formulates a general conditional-generation setting that covers both vector and spatial data, and includes a broader set of baselines than many prior works. The random-masking objective is a sensible extension of classifier-free guidance, and the one-step distillation variant is a practical addition. However, the headline capability of arbitrary-subset conditioning is not directly validated, and several reporting gaps prevent me from confirming the quantitative claims. The contribution would be solid if the missing experiments and consistency fixes are supplied, but on the current evidence the central promise remains unverified.","major_comments":[{"comment":"The paper's central claim of test-time controllability over arbitrary subsets of observed channels (Section 1) is not directly tested. Algorithm 1 samples each channel independently with an unreported Bernoulli probability p, and Section 5.2 evaluates only two configurations: full-observation multi-channel output and leave-one-out single-channel imputation. No experiment samples random subsets of channels at test time, varies the size of the observed set, or probes extreme or rare combinations such as 1 of 43 observed channels in the lung IMC data. Because the number of subsets grows exponentially, independent Bernoulli masking does not by itself guarantee coverage of rare configurations, and the paper explicitly states in Section 3.2 that the informativeness assumption is 'not formally addressed.' The arbitrary-subset capability therefore remains an unvalidated promise. Please add random-subset experiments with several observed-set cardinalities, report the masking probability and the induced training distribution over subset sizes, and analyze performance as a function of the observed fraction.","section":"§3.2, §5.2"},{"comment":"The comparative claims are not supported by sufficient statistical reporting. No confidence intervals, standard deviations, or replication counts are given for any table, and the train/validation/test split is not described for either the CITE-seq or IMC experiments. The text lists 'single-protein predictor and kernel ridge regression' as baselines, but Table 2 reports only 'Most correlated protein' and 'Most spatially correlated protein' and has no KRR row, making the statement that all baselines fail to outperform the best linear predictor unverifiable from the displayed results. Please add error bars or per-patient/per-dataset results, describe the splits, and make the baseline table consistent with the methods listed.","section":"§5.2, Table 2"},{"comment":"There are unresolved dataset inconsistencies. The ablation text says 'Using the lung cancer IMC dataset,' while Table 3's caption says 'breast cancer dataset.' In addition, the cross-dataset paragraph reports 23 common proteins, 18 lung-unique proteins, and 21 breast-unique proteins, which sum to 62 and do not reconcile with the stated totals of 43 and 50 channels. These inconsistencies must be corrected and verified before the ablation and generalization conclusions can be attributed to the intended datasets.","section":"§5.3, Table 3, and §5.2 cross-dataset paragraph"}],"minor_comments":[{"comment":"The symbol C is overloaded: it is used both for the number of channels and for the conditional space, and p(c) is written as if C were a set. Please introduce separate notation and define the conditional-configuration distribution explicitly.","section":"§3.2, Eq. (5)"},{"comment":"The column headers for BMNC and HSPC appear misaligned or duplicated, and the numeric rows are not consistently aligned. Please reformat the table so each dataset has a clear rc and rp column.","section":"Table 1"},{"comment":"The sentence 'Fixing the channel known to be missing during training improves performance in the single-protein prediction mode' is unclear, and no displayed result corresponds to this variant. Please explain the procedure or remove the sentence.","section":"§5.2"},{"comment":"The manuscript refers to an appendix with hybrid ControlNet/BrushNet experiments, but no appendix is present in the submitted text. Please include the appendix or remove the reference.","section":"§5.2"},{"comment":"No data availability, code availability, or hyperparameter settings are reported, including the masking probability p, number of diffusion steps, noise schedule, learning rate, batch size, and training budget. These details are needed for reproducibility.","section":"§5, general"},{"comment":"There are copyediting issues, including typos in Algorithm 1 ('conditon', 'encorages'), and the supplied text contains uninterpretable token sequences such as repeated '/gid...' strings. The final manuscript should be carefully proofread and rendered cleanly.","section":"Algorithm 1, full text"},{"comment":"The suggestion that the model may serve as a 'foundation model for spatial biology' is speculative given the scale of the experiments (two IMC datasets and four CITE-seq datasets). Please temper the claim.","section":"§6"}],"recommendation":"major_revision","confidential_remarks":"I am not concerned about the use of the authors' own distillation papers (refs. [30,31]); they are cited for the one-step SiD variant and do not affect the main architectural claim. The more significant editorial concerns are the missing appendix, the Table 2/Table 3 inconsistencies, and the absence of code/data availability statements. The arbitrary-subset claim is the load-bearing issue: it is fixable with targeted random-subset experiments, which is why I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick read on arXiv:2507.02902. The paper is a solid, workmanlike combination of existing diffusion components aimed at a real biological problem: imputing unmeasured protein channels in IMC and CITE-seq data. What is new is not any single block—ControlNet-style hierarchical conditioning, SE and transformer channel attention, random masking—but packaging them for high-dimensional, spatially aligned biological channels, plus the output-space channel attention, which I have not seen applied this way. The benchmark results are consistent: their model tops the baselines on all four CITE-seq datasets and both IMC datasets, and the distilled one-step variant stays close. That is a credible engineering contribution.\n\nThe soft spots are real but addressable. The central claim of test-time controllability over arbitrary subsets of observed channels is not actually tested. Section 5.2 only evaluates leave-one-out single-channel imputation and full multi-channel prediction; the cross-dataset experiment predicts the intersection set. Random masking is supposed to amortize over all subsets, but with 43–50 channels, most combinations are never seen, and the masking probability p is not reported. So the flagship capability is asserted, not demonstrated. I would like to see experiments sampling random subsets, varying the observed fraction, and probing rare configurations (e.g., 1 of 43 channels observed). The absence of error bars matters too: without them, it is hard to tell whether the margin over baselines is meaningful, especially in Table 2, where their own single-channel variant beats their multi-channel model.\n\nThere is also a concrete inconsistency: Table 3's caption says breast cancer while Section 5.3 says lung. Minor, but it chips at confidence in the ablations. No code or data release is mentioned, which for a method paper in computational biology is a limitation.\n\nOn the citation pattern: their use of their own SiD distillation papers is for a speed-up variant and does not affect the main claim. No circularity red flag.\n\nBottom line: this deserves peer review. The method is plausible, the results are uniformly in its favor, and the gaps are fillable with additional experiments. But the arbitrary-subset claim needs direct evidence, and the paper needs error bars and consistent reporting before it is ready. I would send it to review with a request for revisions.","headline":"A plausible diffusion framework for multi-channel biological imputation whose flagship claim—arbitrary subset conditioning—is asserted but not yet directly tested; worth peer review with revision requests.","tokens_in":12428,"tokens_out":1892,"would_cite":true,"duration_ms":19835,"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":"A diffusion model trained with random channel masking reconstructs missing biological channels from any observed subset and outperforms dedicated baselines on spatial and single-cell protein prediction.","keywords":["diffusion models","multi-channel biological data","imaging mass cytometry","single-cell multimodal prediction","conditional generation","random channel masking","channel-wise attention","spatial imputation"],"falsifier":"Train the model on a 43-channel IMC dataset with random masking, then hold out random subsets of observed channels of sizes such as 5, 10, and 20 and measure per-protein Pearson correlation against leave-one-out results; if accuracy collapses on these intermediate subsets, the arbitrary-conditioning claim is not supported.","tokens_in":11539,"feed_emoji":"🧬","tokens_out":8589,"duration_ms":81821,"temperature":0.7,"pith_summary":"Biological profiling technologies such as imaging mass cytometry and single-cell sequencing produce many channels at once, and real experiments often measure only a subset of the channels a researcher cares about. This paper tries to establish that a single diffusion model, trained once, can generate any missing channels from any observed subset while preserving the spatial alignment between channels. The authors build the model around three ideas: injecting observed-channel features into the denoiser at multiple resolutions, applying channel-wise attention in both latent and output space, and training with random channel masking so one amortized model covers all conditioning configurations. If the claim is right, biologists could extend pre-designed antibody panels in silico, rescue corrupted or incomplete acquisitions, and pool datasets with partially overlapping panels without retraining per dataset.","feed_headline":"One diffusion model fills any missing protein channel","feed_subtitle":"One trained model predicts missing proteins in tissue images and single cells, beating dedicated baselines.","key_machinery":"The framework has three load-bearing components. Hierarchical feature injection runs a parallel contextual network over the observed channels and injects its multi-resolution feature maps into the corresponding layers of a UNet denoiser through Squeeze-and-Excitation gating, which preserves spatial alignment while modulating the condition. Channel-wise attention appears in two places: inside the UNet blocks, where global-average-pooled features or transformer-style attention reweight latent channels, and at the output layer, where a residual Squeeze-and-Excitation path models dependencies among the semantic protein channels. Random channel masking samples a random observed subset each training iteration, so the model optimizes a single amortized conditional denoising objective over the space of all subsets. The masking scheme is what carries the arbitrary-conditioning generalization claim; the attention and injection mechanisms carry the fidelity claim.","core_discovery":"The paper's central claim is that multi-channel biological generation should be treated as conditional imputation of the full channel stack from any spatially aligned observed subset, and that a random-masking diffusion model can do this for arbitrary subsets at test time. On single-cell CITE-seq gene-to-protein prediction, the model records the highest protein-level Pearson correlation on all four benchmark datasets. On imaging mass cytometry data, it imputes a held-out protein channel from the remaining channels with higher correlation than kernel ridge regression, ControlNet, and the domain-specific Stem and MULTIPLAI models. The authors also report that training on the union of two partially overlapping protein panels outperforms training only on their intersection, which they interpret as evidence that partially missing channels still provide useful training signal.","pith_inferences":["Not tested in the paper: the experiments use leave-one-out and full-panel configurations, so evaluating random intermediate-size subsets would directly test whether the amortized model really generalizes to arbitrary conditioning.","Not tested in the paper: the same random-masking recipe could scale to spatial transcriptomics panels with thousands of genes, where computing channel-wise attention across all latent channels would be the main cost to watch.","If the union-set result holds beyond these two datasets, each new tissue atlas could improve imputation for all previously collected panels even when channel overlap is small, turning panel design into a collaborative accumulation of channels.","The hierarchical injection mechanism is not specific to biology; aligned multi-channel sensor data such as multispectral imagery could be conditioned the same way, treating each spectral band as an observed channel."],"forward_implications":["A single trained model can output the full protein panel from any measured subset, so a biologist could change the antibody panel after collection and still predict the unmeasured proteins computationally.","The same architecture handles non-spatial single-cell data with many gene channels, so the method is not restricted to image-shaped inputs.","Because all channels are predicted jointly, training across datasets can be done by zero-padding missing channels rather than learning per-dataset models or per-protein heads.","The distilled one-step version keeps most of the prediction accuracy, making the approach usable without running iterative diffusion sampling."],"supporting_citations":[{"why":"Supplies the denoising diffusion objective the model is trained on.","marker":"[9]"},{"why":"Supplies the score-matching view that connects noise prediction to conditional density estimation.","marker":"[5]"},{"why":"Supplies the Squeeze-and-Excitation block used for channel gating and channel attention.","marker":"[13]"},{"why":"Supplies the classifier-free guidance principle that random masking generalizes.","marker":"[17]"},{"why":"Supports the amortized-inference assumption that partially observed data yield informative training gradients.","marker":"[12]"},{"why":"ControlNet is the spatial-conditioning method the paper extends and the main diffusion baseline it compares against.","marker":"[2]"},{"why":"scMM is a single-cell multi-omics baseline the model outperforms in gene-to-protein prediction.","marker":"[20]"},{"why":"Stem is a domain-specific diffusion baseline for spatial data that the model is compared against.","marker":"[23]"},{"why":"Supplies the kernel ridge regression baseline and benchmark datasets used in the single-cell and spatial comparisons.","marker":"[26]"}],"fun_headline_variants":["Diffusion model imputes any missing protein from any channel subset","One model fills all missing protein channels, any observed subset","Masked training lets a diffusion model handle arbitrary channel gaps","Tissue and single cells: one model predicts missing proteins"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that randomly masking channels during training gives informative learning gradients for every combination of observed and missing channels, so the model can generalize to conditioning subsets it never saw during training.","fun_headline_variants_meta":{"raw":{"variants":["Diffusion model imputes any missing protein from any channel subset","One model fills all missing protein channels, any observed subset","Masked training lets a diffusion model handle arbitrary channel gaps","Tissue and single cells: one model predicts missing proteins"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000211,"raw_usage":{"total_tokens":1390,"prompt_tokens":900,"completion_tokens":490,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":516,"completion_tokens_details":{"reasoning_tokens":421}},"tokens_in":516,"tokens_out":490,"duration_ms":6116,"temperature":1.0,"reasoning_tokens":421,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T23:06:44.360306+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the model on a 43-channel IMC dataset with random masking, then hold out random subsets of observed channels of sizes such as 5, 10, and 20 and measure per-protein Pearson correlation against leave-one-out results; if accuracy collapses on these intermediate subsets, the arbitrary-conditioning claim is not supported.","supporting_citations":[{"cited_title":"Generative modeling by estimating gradients of the data distribution","cited_arxiv_id":null,"evidence_quote":"Supplies the score-matching view that connects noise prediction to conditional density estimation."},{"cited_title":"Squeeze-and-excitation networks","cited_arxiv_id":null,"evidence_quote":"Supplies the Squeeze-and-Excitation block used for channel gating and channel attention."},{"cited_title":"A mixture-of-experts deep generative model for integrated analysis of single-cell multiomics data","cited_arxiv_id":null,"evidence_quote":"scMM is a single-cell multi-omics baseline the model outperforms in gene-to-protein prediction."},{"cited_title":"Diffusion generative modeling for spatially resolved gene expression inference from histology images","cited_arxiv_id":null,"evidence_quote":"Stem is a domain-specific diffusion baseline for spatial data that the model is compared against."},{"cited_title":"Luecken, Daniel B","cited_arxiv_id":null,"evidence_quote":"Supplies the kernel ridge regression baseline and benchmark datasets used in the single-cell and spatial comparisons."}],"review_version":1}