{"id":"a07349f1-e5ff-4c7b-bbd2-2ef68efdb258","arxiv_id":"2412.01124","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"SUICA combines a graph autoencoder with an implicit neural representation to impute and denoise spatial gene expression, reporting improved fidelity and biological signal preservation.","lead":"This paper presents SUICA, a neural network that turns discrete spatial transcriptomics measurements into a continuous function of spatial coordinates, enabling imputation of missing spots, missing genes, and noise in gene expression data. The authors report strong gains over existing methods on several spatial transcriptomics platforms, which could make expensive gene-expression mapping experiments go further.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The spatial-imputation benchmark may be transductive: the GAE is described as pre-trained on the whole ST slice, so held-out spots' expression profiles can enter the latent codes used for evaluation; Table 1 may not test unseen-spot generalization.","rationale":"The strongest claim depends entirely on the 20% held-out spots being unseen, and the text never clearly states that the GAE is trained only on the training subset. The language in Section 3.2.2, which describes GAE pre-training on the given ST slice, and Section 3.2.4, which describes training the decoder on the raw dataset readouts, directly suggests the opposite. That makes the central spatial-imputation result uninterpretable as a test of spatial generalization. I considered other concerns: metrics computed only on non-zero values, ARI above ground truth, duplicate and corrupted manuscript text, lack of multiple seeds, and weaker MERFISH results. These are real issues and support a cautious verdict, but they are secondary. The non-zero-only metric is explicitly justified by zero-inflation; ARI above ground truth could reflect beneficial smoothing; and manuscript artifacts, while damaging readability, do not by themselves refute the method. Appendix H also concedes that SUICA prefers high-quality ST data and degenerates toward vanilla INR on MERFISH, which tempers cross-platform claims but does not displace the leakage concern. The transductive-leakage issue is the one that, if true, invalidates the headline comparison. The proposed check isolates it: restrict GAE pre-training, embedding generation, and decoder finetuning to the 80% training spots and see whether the Table 1 margins survive. Since the reader already flagged exactly this assumption and asked for clarification and a redo, my recommendation is unchanged: CONDITIONAL, pending the protocol check.","tokens_in":28020,"tokens_out":7378,"duration_ms":68502,"concrete_test":"Inspect the public code (github.com/Szym29/SUICA) to determine whether the GAE pre-training and decoder finetuning receive the full expression matrix E and full KNN graph A, including the 20% held-out spots. Then rerun the spatial-imputation benchmark on one MOSTA slice, e.g., E16.5, with the identical 80/20 split, but restrict E and A to the 80 training spots for GAE pre-training, zgt generation, and decoder finetuning. If MAE, cosine similarity, Pearson, Spearman, or ARI on the same 20% test spots changes by more than 10% relative to Table 1, the reported gains are substantively leakage-driven and the spatial-imputation experiment must be redone.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that SUICA predicts gene expression for 20% unseen spots better than baselines (Table 1, Section 4.4). This requires the held-out spots to be unseen during all stages of SUICA training. The paper is ambiguous and likely contradictory here. Section 3.2.1 says 'With the training subset, we train SUICA (GAE+INR), and infer the test subset for evaluation,' but Section 3.2.2 says the GAE is pre-trained using the given ST slice in a self-regressing manner, and that with the pre-trained GAE the authors obtain the encoded latent representation for all spots, denoted zgt. Section 3.2.4 then says the decoder is trained to fit the raw dataset readouts ygt without restricting to the training subset. If the GAE encoder, GAE decoder, and zgt for the 20% held-out spots are computed from the full expression matrix, then the test spots' expression profiles are seen during pre-training and decoder finetuning. The INR only learns x to z on 80% of the coordinates, but the z values it is asked to interpolate already encode the test expressions, and the decoder has been fit to those same expressions. The benchmark then measures transductive autoencoding plus coordinate interpolation, not spatial imputation of unseen spots. The reported MAE/MSE/cosine/Pearson/Spearman/ARI numbers in Table 1 would not establish the claimed generalization. Appendix F's protocol, which varies the training fraction while the test set remains 20%, also suggests the GAE may use all spots, since otherwise the unused spots are unspecified. This is the load-bearing weak point: the method's advantage may come from having seen the target spot's transcriptome, not from continuous spatial modeling.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SUICA, an implicit neural representation (INR) method for spatial transcriptomics (ST). A graph autoencoder (GAE) first maps the high-dimensional, sparse expression matrix into low-dimensional embeddings; an INR then maps spatial coordinates to those embeddings; a finetuned decoder maps the interpolated embeddings back to the raw expression space, using a Dice-loss regularizer to handle zero inflation. The authors evaluate SUICA on spatial imputation, gene imputation, and denoising across Stereo-seq MOSTA, Slide-seqV2, Visium, and MERFISH datasets, reporting improvements in MAE, MSE, cosine similarity, Pearson/Spearman correlations, and ARI over INR baselines and STAGE, with code made available.","tokens_in":28338,"tokens_out":5523,"duration_ms":52873,"significance":"If the reported results are valid, SUICA is a meaningful contribution: it offers a modular INR design that scales to the super-high-dimensional, zero-inflated output space of ST, and the paper provides multi-platform experiments, ablations, and a data-efficiency study in the appendix. The release of code is a concrete strength. However, the central spatial-imputation benchmark rests on an evaluation protocol that is not fully specified and may be circular, because the GAE appears to be pretrained on the full slice including the held-out test spots. Until that issue is resolved, the quantitative claims in Table 1 cannot be taken as evidence of generalization to unseen spatial locations.","major_comments":[{"comment":"The spatial-imputation evaluation may be transductive. Section 3.2.1 says the spots are split into training and test subsets and that SUICA is trained on the training subset, but Section 3.2.2 says the GAE is pretrained \"using the given ST slice in a self-regressing manner\" and that zgt is obtained \"for all spots,\" and Section 3.2.4 says the decoder is trained to fit the \"raw dataset readouts ygt\" without restricting to the training subset. If the GAE encoder, GAE decoder, the latent codes zgt for the 20% test spots, and the decoder finetuning all use the test spots' expression profiles, then Table 1 measures transductive autoencoding plus coordinate interpolation rather than spatial imputation of unseen spots. This would invalidate the central claim in Section 4.4. The authors must state unambiguously how the held-out 20% is excluded from every stage—GAE pretraining, latent-code computation, INR fitting, and decoder finetuning—and rerun the benchmark under that protocol. The same ambiguity affects the gene-imputation results in Table 2, because the GAE is pretrained on the unmuted expression matrix before the 70% masking step.","section":"§3.2.1, §3.2.2, §3.2.4, §4.2, Appendix F"},{"comment":"Numerical fidelity is reported only on non-zero entries. For zero-inflated ST data, a predictor that always outputs zero achieves zero error on the zero entries and is never penalized for false positives under this protocol, so the MAE/MSE/cosine numbers do not establish fidelity for the full expression profile. The paper should report full-entry MSE/MAE and a zero/non-zero classification metric (e.g., IoU or F1) alongside the non-zero-only metrics, or justify with a sensitivity analysis why the non-zero-only choice cannot change the conclusions.","section":"§4.1, Table 1"},{"comment":"No measure of variability is reported. All comparisons are single numbers without standard deviations, number of random splits, or seeds, and several claimed margins are small (e.g., the 4.5–5.6% improvements reported in Section 4.4). Without repeated trials or confidence intervals, the observed differences may not be statistically significant. Report mean and standard deviation over at least three random 80/20 splits or training seeds for the main tables.","section":"§4.2–§4.4, Tables 1–5"}],"minor_comments":[{"comment":"The manuscript contains duplicated blocks of text and CVPR submission boilerplate (e.g., repeated paragraphs around Section 3.2.4 and the phrase \"CVPR#1357 CONFIDENTIAL REVIEW COPY\"). These should be removed before resubmission.","section":"Throughout"},{"comment":"The loss definitions are inconsistent across versions: one formulation of Dice loss uses sgn(ygt) in the numerator, another uses ygt directly, and Eq. (5) reads \"LMAE = LMSE,\" which appears to be a typo. Please harmonize the equations and define all symbols precisely.","section":"§3.2.4, Eqs. (2)–(5)"},{"comment":"The data-efficiency ablation varies the training fraction while keeping the test set at 20%, but the text does not state whether the unused spots are discarded or are still used for GAE pretraining. This directly affects the interpretation of the results and should be clarified.","section":"Appendix F"},{"comment":"The claim that SUICA's ARI exceeds the ground-truth ARI by 3.9% is surprising and should be discussed explicitly: if the predicted data are more clusterable than the ground truth, the paper should explain whether this reflects desirable denoising, an artifact of the metric, or a mismatch with biological cell-type structure.","section":"§4.4"}],"recommendation":"major_revision","confidential_remarks":"The key issue is the unresolved evaluation protocol for spatial imputation; if the authors can clarify and recompute the results with a properly held-out GAE, the paper could become acceptable. The manuscript also appears to be a compilation of earlier CVPR submission versions, and the editorial quality needs attention."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The thing to know about this paper: the method is reasonable but the main evaluation may not test what it claims. SUICA combines a graph autoencoder, an INR over spatial coordinates, and a Dice-loss regularizer to reconstruct gene expression. The combination is new in the spatial transcriptomics space, the code is released, and the authors evaluate across Stereo-seq, Slide-seqV2, Visium, and MERFISH. The ablation study (AE, Dice, Graph) is informative, and the observation that SIREN versus FFN matters depending on spatial density is a useful practical detail.\n\nThe soft spot is load-bearing. The spatial imputation protocol splits spots 80/20, but Section 3.2.1 says the GAE is pre-trained on “the given ST slice” and that zgt is obtained for “all spots.” Section 3.2.4 then trains the decoder to fit ygt without restricting to the training subset. If the GAE encoder, GAE decoder, and decoder fine-tuning all see the 20% test spots’ expression profiles, then Table 1 measures transductive autoencoding plus coordinate interpolation, not spatial generalization to unseen spots. The paper never explicitly rules this out, and Appendix F’s protocol (varying training fraction while test stays 20%) makes it plausible that unused spots are still fed to the GAE. This needs to be stated plainly and, if the GAE does see all spots, the benchmark needs to be redone with the GAE trained only on training spots.\n\nOther issues are real but less severe. Numerical fidelity is reported only on non-zero entries, which flatters zero-inflated data; no standard deviations or repeated seeds are given; and SUICA’s ARI on Stereo-seq exceeds the ground-truth ARI, which is a red flag that the method is doing something beyond faithful reconstruction. The novelty claim is also softened by the fact that the INR maps coordinates to a 32-d embedding, not directly to 20,000 genes. Table 7 on MERFISH shows SUICA’s Spearman is actually the worst of the compared methods, so the “superior across platforms” line should be moderated.\n\nThe arXiv v2 text itself is a corrupted compilation: it contains duplicated sections, leftover “CVPR#1357 CONFIDENTIAL REVIEW COPY” headers, and an author note (“[Qingtian: check Prof. Ding’s version]”). That is not the authors’ science, but it makes the manuscript hard to rely on as-is and must be cleaned up.\n\nWho is this for? Someone working on ST imputation or INR applications to sparse high-dimensional data. The method has a good chance of being useful for gene imputation and denoising, which are evaluated on all spots with synthetic degradation and are not undermined by the transductive concern. The spatial imputation claim urgently needs the protocol clarified. I would send this to review, but with a request to the authors to specify the GAE training split, rerun the benchmark under a clean held-out protocol, report variance across seeds, and soften the cross-platform claim. If the transductive concern is confirmed, the spatial imputation results in Table 1 cannot stand as evidence of spatial generalization.","headline":"SUICA is a plausible GAE+INR recipe for spatial transcriptomics, but the central spatial-imputation benchmark may be transductive because the graph autoencoder appears to see the test spots before evaluation.","tokens_in":28962,"tokens_out":2052,"would_cite":false,"duration_ms":19537,"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":"Sparse gene-expression data can be modeled as a continuous neural map that reconstructs unmeasured tissue spots with better fidelity and biological structure than existing tools.","keywords":["spatial transcriptomics","implicit neural representation","graph autoencoder","spatial imputation","gene imputation","denoising","zero-inflated distribution","regression-by-classification"],"falsifier":"Re-run the spatial-imputation protocol with the GAE pre-trained strictly on the 80% training spots and the KNN graph built without the held-out 20%, then compare MAE, MSE, cosine similarity, Pearson, Spearman, and ARI against the reported MOSTA values of 5.66, 0.85, 0.797, 0.792, 0.447, and 0.343. If the margins over FFN, SIREN, and STAGE shrink substantially, or if the ARI no longer clears the ground-truth reference of 0.312, the spatial-generalization claim would be falsified.","tokens_in":27802,"feed_emoji":"🧬","tokens_out":13152,"duration_ms":102137,"temperature":0.7,"pith_summary":"SUICA aims to model spatial transcriptomics (ST) data — sparse measurements of tens of thousands of genes at discrete, irregularly placed tissue spots — as a continuous, compact function of spatial coordinates. The central claim is that a graph-aware autoencoder coupled with an implicit neural representation (INR), trained under a classification-style loss, lets one pipeline reconstruct unmeasured spots, impute muted genes, and denoise counts across several ST platforms without knowing the degradation type. The stakes are practical: ST data are expensive and dropout-heavy, so a faithful continuous representation could let researchers densify and clean measured slices for downstream analysis. On the main benchmark, the paper reports the lowest errors, cosine similarity 5.6% above the runner-up, Pearson and Spearman correlations of 0.792 and 0.447, and cell-type clustering (ARI 0.343) above the ground-truth slice's reference (0.312).","feed_headline":"Continuous gene map beats rivals by 5.6% on unseen tissue spots","feed_subtitle":"If right, researchers can densify pricey tissue data and recover cell-type structure in one pass.","key_machinery":"The load-bearing machinery is the GAE–INR–decoder cascade. An implicit neural representation (INR) is a neural network that maps spatial coordinates to signal values, giving a continuous model of a discretely sampled signal; SUICA's INR learns $\\mathbf{x} \\mapsto \\hat{z}$ from spot coordinates to the low-dimensional embeddings $z$ produced by the graph autoencoder. The GAE, built on a graph convolutional network in its encoder only, supplies embeddings that are compact, dense, and context-aware; the decoder is detached from the encoder, warmed up separately, then finetuned with the INR frozen to map $\\hat{z}$ back to the raw expression space. The regression-by-classification term — a Dice loss computed on $\\tanh$-mapped outputs — is what stops the reconstruction from collapsing the zero-inflated distribution into a smooth Gaussian one, and the choice of FFN versus SIREN as the INR backbone is made according to spatial density of the slice.","core_discovery":"The paper's discovery is that the real obstacle to applying INRs to ST is not network capacity but the geometry of the target space: raw expression profiles are zero-inflated and can exceed 20,000 channels, so an INR regressing directly onto them washes out the sparsity. SUICA factors the problem: a graph-augmented autoencoder compresses the raw cell-by-gene matrix into low-dimensional, structure-aware embeddings, the INR learns the smooth coordinate-to-embedding map, and a finetuned decoder returns to raw expression space. The skewed zero-versus-nonzero balance is treated as a quasi-classification problem: outputs are pushed through a non-negative tanh and scored with a Dice (intersection-over-union) loss that preserves sparsity in the predictions. The paper presents graph total variation evidence that GAE embeddings retain sharper spatial structure than plain autoencoder embeddings, and reports that the full pipeline beats vanilla INR variants and imputation baselines on numerical fidelity, statistical correlation, and bio-conservation while also amplifying biological signatures for downstream analysis.","pith_inferences":["The decisive check on the spatial-generalization claim is to pre-train the GAE strictly on the 80% training spots, with the KNN graph built without the held-out 20%; the paper's own wording leaves open whether the GAE saw the test spots during pre-training.","The decoupling recipe — let a smooth neural field do spatial interpolation in a learned latent space while a separate head enforces raw-space fidelity — transfers naturally to other sparse, high-dimensional coordinate measurements, such as single-cell atlases or multi-modal tissue imaging, where direct INR regression suffers the same zero-inflation failure.","The ARI above the ground-truth reference implies the representation acts as a denoiser that sharpens cell-type boundaries; a testable prediction is that the amplification concentrates along known anatomical or developmental gradients, and a caveat is that low-signal regions could be over-sharpened."],"forward_implications":["Spatial imputation: expression profiles of held-out spots can be predicted from coordinates alone, with the reported benchmark showing the lowest MAE/MSE, the highest cosine similarity (5.6% above the runner-up), and the best Pearson/Spearman correlations among FFN, SIREN, and STAGE.","Gene imputation and denoising use the identical pipeline with no knowledge of the degradation, making SUICA a reference-free, degradation-agnostic restoration tool rather than a per-task model.","Predicted profiles preserve cell-type structure: clustering SUICA's reconstructions reaches an ARI of 0.343 on the MOSTA slice, above the 0.312 reference ARI of the ground truth, and marker genes such as RPLP0 and AFP localize to the correct anatomical regions.","A single continuous representation spans Stereo-seq, Slide-seqV2, Visium, and MERFISH, so the method is claimed to transfer across sequencing platforms with different spatial densities and dropout rates."],"supporting_citations":[{"why":"Supplies the MOSTA Stereo-seq mouse-embryo dataset on which the main spatial-imputation, gene-imputation, and denoising benchmarks are computed.","marker":"Chen et al., 2022a"},{"why":"Furnishes FFN, the random-Fourier-feature INR used both as a SUICA backbone on dense slices and as a baseline.","marker":"Tancik et al., 2020"},{"why":"Furnishes SIREN, the periodic-activation INR used as SUICA's backbone on sparse slices and as a baseline.","marker":"Sitzmann et al., 2020"},{"why":"Provides STAGE, the learned reference-free spatial-imputation baseline that SUICA must beat for the paper's main comparison.","marker":"Li et al., 2024"},{"why":"Supplies the graph convolutional network that makes the autoencoder's embeddings structure-aware.","marker":"Kipf & Welling, 2017"},{"why":"Supplies the generalized Dice loss that the paper converts into its regression-by-classification training objective.","marker":"Sudre et al., 2017"},{"why":"Provides the Slide-seqV2 mouse-hippocampus dataset used for SUICA's second quantitative benchmark.","marker":"Stickels et al., 2021"}],"fun_headline_variants":["SUICA turns sparse gene data into continuous spatial maps","Sparse genes meet INR: SUICA beats SOTA on fidelity","Graph autoencoder plus classification wins gene mapping","SUICA: compact INR for super-high-dimensional ST","Zero-inflated expression modeled as classification beats rivals"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the held-out 20% of spots were genuinely unseen during graph-autoencoder pre-training; the paper (Sections 3.2.1 and 4.2) never states whether the GAE was fit on only the training subset or on the full slice, so the spatial-imputation benchmark may be autoencoding already-seen data rather than testing spatial generalization.","fun_headline_variants_meta":{"raw":{"variants":["SUICA turns sparse gene data into continuous spatial maps","Sparse genes meet INR: SUICA beats SOTA on fidelity","Graph autoencoder plus classification wins gene mapping","SUICA: compact INR for super-high-dimensional ST","Zero-inflated expression modeled as classification beats rivals"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000293,"raw_usage":{"total_tokens":1724,"prompt_tokens":976,"completion_tokens":748,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":592,"completion_tokens_details":{"reasoning_tokens":671}},"tokens_in":592,"tokens_out":748,"duration_ms":7678,"temperature":1.0,"reasoning_tokens":671,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T04:40:35.106476+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the spatial-imputation protocol with the GAE pre-trained strictly on the 80% training spots and the KNN graph built without the held-out 20%, then compare MAE, MSE, cosine similarity, Pearson, Spearman, and ARI against the reported MOSTA values of 5.66, 0.85, 0.797, 0.792, 0.447, and 0.343. If the margins over FFN, SIREN, and STAGE shrink substantially, or if the ARI no longer clears the ground-truth reference of 0.312, the spatial-generalization claim would be falsified.","supporting_citations":[{"cited_title":"H., Li, W., Vercauteren, T., Ourselin, S., and Jorge Cardoso, M","cited_arxiv_id":null,"evidence_quote":"Supplies the generalized Dice loss that the paper converts into its regression-by-classification training objective."}],"review_version":1}