{"id":"64d5f8c8-c986-4b3d-9ce5-347629104e36","arxiv_id":"2502.02629","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"scGSL tunes cell-cell graphs with GNN graph structure learning and pairwise alignment transfer, improving cell type annotation over 10 baselines and yielding graphs proposed to reflect tumor microenvironment interactions.","lead":"A new graph neural network called scGSL combines graph structure learning with graph domain adaptation to label cell types in non-spatial tumor single-cell RNA-seq data, reporting about 85% accuracy across three cancers. The same learned cell-cell graphs are offered as a way to read off tumor microenvironment interactions without spatial data.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The biological-interaction claim rests on a circular validation: edge-connected vs non-connected cells differ for any variable gene, so the Mann-Whitney test cannot establish gene-pair interactions.","rationale":"The reader's weakest assumption targets the GDA pairwise alignment assumptions (Section 2.4). That is a legitimate concern, but it is an inherited methodological assumption that can be checked with the provided code and may hold approximately for these datasets. The interpretability validation, by contrast, is internally flawed: the Mann-Whitney test in Section 3.4 compares expression of single genes between edge-connected and non-connected cells, but edge connectivity is generated from expression similarity, so significance is expected for most variable genes. This makes the abstract's central claim about 'biologically meaningful gene interactions' vacuous as presented. The concern is not about external consensus but about the logical validity of the evidence. A concrete null-control experiment would settle it. Since both the reader's conditional verdict and mine require additional evidence, the verdict remains CONDITIONAL; no change to the verdict is needed, but the condition should explicitly include a non-circular validation of the interaction claim.","tokens_in":16730,"tokens_out":3797,"duration_ms":38207,"concrete_test":"Re-run the Section 3.4 protocol with two null controls: (1) a graph of the same cells with edges randomly permuted within each cell type while preserving node degree, and (2) a KNN graph built from the same HVGs but without GSL refinement. Additionally, test 100 random gene pairs matched for expression variance as negative controls. If known pairs show comparable significance rates in the random-edge graph, or if random pairs show significance rates comparable to Table 4–6 in the scGSL graph, the interaction-identification claim is unsupported. Report false discovery rates across random pairs for the scGSL graph.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract's second claim, that scGSL 'proves to robustly identify biologically meaningful gene interactions in an unsupervised manner,' is supported in Section 3.4 only by testing whether individual genes in known pairs show expression differences between cells connected by an edge and cells not connected (Tables 4–6). This is not a test of gene-pair interactions. The graph edges are themselves constructed from expression similarity via KNN and the GSL feature-smoothness loss (Eq. 4), so connected cells are transcriptionally similar by construction. Any gene that varies across cell states will differ between the connected and background groups, independent of whether the gene pair mediates a real cell-cell interaction. The validation is therefore tautological and would also pass for random gene pairs. The 'unsupervised' framing is also misleading: GSL uses cell-type labels (Eq. 1) during graph learning, so the graph is supervised by cell type even though the specific validated pairs were not used. Because the interpretable graph is a core claimed novelty, this invalid validation is at least as load-bearing as the GDA assumptions identified by the reader; if the interaction-identification claim fails, the paper reduces to a modest accuracy gain on top of an existing graph domain adaptation method.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes scGSL, a GNN-based method that combines graph structure learning (GSL) with graph domain adaptation (pairwise alignment, PA) to annotate cell types in non-spatial scRNA-seq data across patients of the same cancer type, and to construct cell-cell interaction graphs for tumor microenvironment analysis. The model is evaluated on 49,024 cells from 19 patients across three cancers (leukemia, breast cancer, colorectal cancer) in pairwise patient-transfer experiments, reporting an average accuracy of 84.83% against ten baseline methods. The authors further claim that the learned graphs identify biologically meaningful gene interactions in an unsupervised manner, validated by Mann-Whitney tests on six literature-known gene pairs.","tokens_in":16982,"tokens_out":4924,"duration_ms":44855,"significance":"If the reported accuracy gains and the biological interpretability claims were robust, scGSL would be a useful addition to single-cell analysis methods that extract cell-type annotations and cell communication hypotheses from non-spatial data. The paper provides a public code repository and uses a patient-pair transfer benchmark that is relevant to clinical applications. However, the current evidence is not fully convincing: the breast-cancer average accuracy is internally inconsistent, hyperparameters are selected on the benchmark data without a held-out protocol, no uncertainty estimates accompany the accuracy comparisons, and the biological validation is circular because graph edges are built from expression similarity. The interpretability claim, which is a core claimed novelty, is not supported by the presented tests. The idea of adapting graph-domain alignment to cross-patient cell graphs is plausible, but the paper needs stronger validation before its central claims can be accepted.","major_comments":[{"comment":"There is an internal inconsistency in the reported breast cancer average accuracy. Section 3.1 states the average accuracy for the Breast Invasive Carcinoma dataset as 83.56%, whereas the 'Average' row of Table 2 and the text in Section 3.2 both report 82.14%. Recomputing the average of the 'Ours' column in Table 2 gives 83.56%, which is also the value consistent with the overall unweighted average of 84.83% across all 40 patient-pair experiments. The manuscript should be corrected to use one consistent number, and the aggregate averaging scheme should be stated explicitly.","section":"Section 3.1, Table 2, Section 3.2"},{"comment":"The accuracy comparisons are reported as single point estimates with no variance or significance testing. Many of the claimed advantages over the second-best method are 2-4 percentage points, and with one run per patient pair it is unclear whether these differences are above noise. Provide standard deviations, confidence intervals, or paired significance tests (e.g., Wilcoxon signed-rank over the patient-pair experiments) for the comparisons.","section":"Section 3.2, Tables 1-3"},{"comment":"Hyperparameters, including KNN neighbor fraction K=0.2% and embedding size ES=128, are selected by comparing performance on the same datasets that are later used in the benchmark (Fig. 5a). As no held-out validation or nested cross-validation is described, the reported gains over baselines may be inflated by selection on the test data. Please either use a separate validation set for hyperparameter tuning or demonstrate that all baselines were tuned under the same protocol.","section":"Section 3.3, Fig. 5"},{"comment":"The Mann-Whitney validation is circular and does not support the claim that the model identifies biologically meaningful gene interactions. Cell-cell graphs are constructed from expression similarity via KNN and the feature smoothness loss (Eq. 4), so cells connected by an edge are transcriptionally similar by construction; any gene whose expression varies across cell states will differ between the connected and background groups, independent of whether the gene pair actually mediates cell-cell communication. A control with random gene pairs, or a permutation test that shuffles edge assignments while preserving expression, is needed. Additionally, the statement that the graphs are generated 'in an unsupervised manner' is misleading: the GSL objective in Eq. (1) explicitly uses cell-type labels, so the graph is supervised by cell type even if the specific validated gene pairs were not used in training.","section":"Section 3.4, Tables 4-6"},{"comment":"The graph domain adaptation procedure relies on the pairwise alignment assumptions from ref. [25]: the conditional edge distribution P(A|Y) of the target is assumed to be a density-ratio reweighting of the source, and pseudo-labels from the classifier are used to estimate the quantities Sigma, nu, and C. On cross-patient tumor scRNA-seq data these assumptions may be violated by batch effects and biological shifts, but no diagnostics are provided. Please report experiments that assess whether the estimated gamma and beta actually improve target accuracy relative to an ablated version without reweighting, and evaluate pseudo-label reliability on the source domain.","section":"Section 2.4, Eqs. (6)-(16)"}],"minor_comments":[{"comment":"The caption reads 'Flowchat of scGCN model' but should read 'Flowchart of scGSL model'.","section":"Fig. 2 caption"},{"comment":"Equation (3) defines the weights as ws and wl, but the following sentence states that 'the values of α and β were respectively set at 5e-4 and 1.5', introducing undefined symbols α and β. Please align the notation.","section":"Eq. (3), Section 2.3"},{"comment":"The abstract reports a dataset of 49,020 cells, while Section 2.1 states 49,024 cells. These numbers should be reconciled.","section":"Abstract vs Section 2.1"},{"comment":"The paper states 'we introduce for the first time the concept of graph domain adaptation' in this context, but the pairwise alignment method is taken from ref. [25]. Please clarify that the contribution is the application to single-cell cell-cell graphs, not the introduction of the GDA concept itself.","section":"Section 2.4, introduction"},{"comment":"The tables report performance for each patient pair, but no information is given about cell type class distributions or which cell types are present. Adding the number of cells per class and the label distributions would help interpret the per-pair variability.","section":"Section 3.2, Tables 1-3"}],"recommendation":"major_revision","confidential_remarks":"The paper is on a topic suitable for the journal and the code/data availability is a positive feature. The main concern for me is that the biological interaction validation is not a valid test of the claimed interpretability, and the accuracy claims currently rest on an internally inconsistent table and a hyperparameter-selection protocol that may overstate the gains. Both concerns are addressable through additional experiments and corrections. I would also encourage the editor to verify that the GitHub repository contains the exact code used for the reported experiments, as the manuscript does not specify a versioned release or commit."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Worth a look if you work on single-cell annotation, but keep your expectations in check. The paper's real contribution is empirical: it grafts a published pairwise-alignment graph domain adaptation method onto a published graph structure learning objective, and shows that the combination beats ten baselines by 2–4 accuracy points in patient-pair transfer across three cancer types. That benchmark is the strongest part of the paper. The experiments are combinatorial over patient pairs, the code and data are public, and the ablation shows the GSL component actually helps. Fine.\n\nThe soft spots are real, though. The reporting is sloppy: Table 2's average for the Ours column is wrong (I get about 83.9, not 82.14), Fig. 2 is captioned \"Flowchart of scGCN,\" and the paper mixes up alpha/beta and ws/wl in Section 2.3. The hyperparameters are tuned on the same data with no uncertainty estimates, so the reported gains are likely optimistic. These are fixable.\n\nThe bigger problem is Section 3.4, the \"biologically meaningful gene interactions\" claim. The validation compares expression of individual genes between edge-connected and non-connected cells. But the edges are built from expression similarity via KNN and a feature-smoothness loss, so connected cells are transcriptionally similar by construction. Any gene that varies across cell states will show a significant Mann-Whitney U statistic, whether or not the gene pair participates in a real interaction. This would pass for random gene pairs. It is not a test of gene-pair interactions, and the \"unsupervised\" framing is misleading because the GSL objective uses cell-type labels. The stress-test note is right: if this interaction claim fails, the paper reduces to a modest accuracy gain on top of an existing method.\n\nThe GDA transfer assumption (same conditional edge distribution up to per-class-pair density ratio) is also unexamined, but because the benchmark is empirical, that is a lesser concern than the circular validation. The annotation claim can be checked from the repo; the interaction claim needs a redesign.\n\nVerdict: the annotation benchmark deserves a serious referee, and the paper should go to peer review. But the reviewers should demand negative controls for the graph interpretability analysis (random gene pairs, permutation tests) and a validation that does not rely on expression similarity by construction, plus fixes to the reporting problems. The paper is honest in its limitations, but the central second claim is currently unsupported.","headline":"A competent integration of existing GSL and graph domain adaptation for scRNA-seq cell type annotation, with solid benchmarks and a circular interpretability validation that needs to be redone.","tokens_in":17502,"tokens_out":1880,"would_cite":false,"duration_ms":19262,"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":"A self-learned cell graph can annotate tumor cell types from ordinary scRNA-seq data and simultaneously reveal cell-cell interactions in the tumor microenvironment.","keywords":["single-cell RNA sequencing","tumor microenvironment","cell type annotation","graph neural network","graph structure learning","graph domain adaptation","cell-cell communication","ligand-receptor interactions"],"falsifier":"Take any reference-query patient pair, estimate the density-ratio matrix $\\gamma$ with the paper's equations, and compare the reweighted source edge-type distribution against the observed target edge-type distribution; if the two diverge and the reweighting lowers accuracy below the best baseline on that pair, the core transfer assumption is falsified. A second decisive check would be to run scGSL on a dataset with matched spatial transcriptomics and ask whether the learned cell-cell edges coincide with spatial adjacency or experimentally validated ligand-receptor contacts.","tokens_in":16550,"feed_emoji":"🧬","tokens_out":7181,"duration_ms":66940,"temperature":0.7,"pith_summary":"The paper claims that a graph neural network that learns its own cell-cell graph, rather than relying on fixed similarity graphs or precompiled ligand-receptor lists, can both annotate cell types in non-spatial tumor scRNA-seq data and produce an interpretable picture of cell-cell communication in the tumor microenvironment. Tested by transferring labels between patient samples of the same cancer, the model attains an average accuracy of 84.83%, precision of 86.23%, recall of 81.51%, and F1 of 80.92% across leukemia, breast cancer, and colorectal cancer datasets, consistently two to four points above ten established baselines. The same learned graph, built without supervision from gene interaction knowledge, flags literature-validated gene pairs as significantly differentially expressed in relevant cell types. If correct, the result would make patient-level cell annotation and communication analysis feasible from ordinary scRNA-seq, without spatial assays or complete ligand-receptor databases.","feed_headline":"Self-learned cell graphs beat ten established cell-type annotators","feed_subtitle":"scGSL reaches 84.83 percent accuracy across 19 tumor patients and recovers known gene interactions on its own.","key_machinery":"The load-bearing object is the learned cell-cell adjacency matrix $S$ produced by graph structure learning. It is initialized as a KNN adjacency and refined by alternating optimization of a classification loss, a graph fidelity loss with $\\ell^1$ and nuclear-norm penalties for sparsity and low rank, and a feature-smoothness Laplacian term. A second mechanism, pairwise alignment from graph domain adaptation, computes per-class-pair density ratios $\\gamma$ that reweight source edges so the source conditional edge distribution matches the target, and a label-shift ratio $\\beta$ reweights the classifier loss. Together these mechanisms make the model's predictions transferable across patients and give the final query graph a biological reading.","core_discovery":"On the paper's own terms, scGSL combines graph structure learning with pairwise-alignment graph domain adaptation so that a classifier trained on one patient's labeled cells can annotate another patient's unlabeled cells of the same cancer type. For each source graph, the graph structure learning module refines a KNN adjacency matrix into a symmetric matrix $S$ supervised by cell-type labels, regularized toward low rank and sparsity, and smoothed over gene-expression features; the query graph is initially built by KNN alone. The pairwise alignment step estimates density ratios $\\gamma$ between source and target edge-type distributions, reweights the source graph, and iterates with pseudo-labels to reduce conditional structure shift and label shift. After prediction, the query graph is re-fed through the same graph structure learning module, and the resulting edges are read as putative cell-cell interactions. Across all patient-pair experiments, the method reports average accuracy 84.83%, and for literature-validated gene pairs such as SPRY1-LAT, BSG-ATP2B4, ATP2A3-SP1, and BIRC5-CASP6, Mann-Whitney U tests find significant expression differences between edge-connected and non-edge cells in most experiments, supporting the claim that the learned graph is biologically meaningful rather than merely predictive.","pith_inferences":["If the learned graph edges really track communication, the method could be used to propose novel ligand-receptor or gene-pair hypotheses from non-spatial data alone; a prospective test would require validating newly discovered pairs with spatial or experimental evidence rather than checking already-known ones.","The pairwise alignment reweighting is the most fragile link: batch effects that change the conditional edge distribution beyond a density-ratio rescaling, or unreliable pseudo-labels for rare cell types, would silently distort the target graph; diagnostic plots of estimated versus observed edge distributions would make this visible.","The interpretability results are retrospective, testing genes already known for each cancer; a stronger claim would need the pipeline to surface reproducible new interactions across independent patient cohorts."],"forward_implications":["Cell type annotation can be carried out from one patient's labeled cells to another's unlabeled cells of the same cancer, at accuracies roughly 2 to 4 points above ten existing methods.","Adding graph structure learning to the GNN raises average accuracy from 87.55% to 90.89% on the GSE132509 leukemia dataset, indicating that graph learning itself, not just the GNN, drives the improvement.","The final cell-cell graph can be inspected for cell-type-specific gene pairs, and known interactions show significant expression differences, implying the graph carries communication-relevant signal.","Because the query graph is learned after prediction, the same pipeline yields both labels and a candidate interaction map from a single non-spatial scRNA-seq run."],"supporting_citations":[{"why":"Supplies the pairwise alignment graph domain adaptation procedure (density-ratio reweighting and label-shift correction) that carries the cross-patient transfer.","marker":"[25]"},{"why":"Supplies the matrix-completion result that justifies using l1 and nuclear-norm penalties to enforce sparsity and low rank in the learned graph.","marker":"[26]"},{"why":"Establishes the prior GNN-based cell-cell graph annotation approach that scGSL extends and with which it shares the graph-based premise.","marker":"[15]"},{"why":"Serves as a baseline cell type annotation method in all three disease-specific comparisons.","marker":"[11]"},{"why":"Serves as the strongest or second-strongest baseline across most patient-pair experiments.","marker":"[12]"},{"why":"Supplies leukemia scRNA-seq patient samples used in the patient-pair transfer experiments.","marker":"[20]"},{"why":"Supplies additional leukemia scRNA-seq patient samples used in the transfer and ablation experiments.","marker":"[21]"},{"why":"Supplies breast and colorectal cancer patient samples, including the E-MTAB-8107 cohort used in the gene-selection ablation.","marker":"[23]"},{"why":"Supplies breast cancer patient samples used in the transfer experiments and gene-pair validation.","marker":"[24]"}],"fun_headline_variants":["scGSL graph learner hits 84.83% accuracy on tumor cell typing","scGSL: 84.83% accuracy, auto-discovers gene interactions","Graph structure learning boosts tumor cell annotation to 84.83%","Unsupervised edge discovery in scGSL hits 84.83% accuracy","scGSL auto-learns graphs, annotates cells at 84.83% accuracy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The transfer step assumes that two patients of the same cancer type have the same conditional distribution of cell-cell edges given cell type, differing only by per-class density ratios, and that the model's pseudo-labels are accurate enough to estimate those ratios.","fun_headline_variants_meta":{"raw":{"variants":["scGSL graph learner hits 84.83% accuracy on tumor cell typing","scGSL: 84.83% accuracy, auto-discovers gene interactions","Graph structure learning boosts tumor cell annotation to 84.83%","Unsupervised edge discovery in scGSL hits 84.83% accuracy","scGSL auto-learns graphs, annotates cells at 84.83% accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001492,"raw_usage":{"total_tokens":6048,"prompt_tokens":1061,"completion_tokens":4987,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":677,"completion_tokens_details":{"reasoning_tokens":4896}},"tokens_in":677,"tokens_out":4987,"duration_ms":32212,"temperature":1.0,"reasoning_tokens":4896,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T11:45:50.728255+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take any reference-query patient pair, estimate the density-ratio matrix $\\gamma$ with the paper's equations, and compare the reweighted source edge-type distribution against the observed target edge-type distribution; if the two diverge and the reweighting lowers accuracy below the best baseline on that pair, the core transfer assumption is falsified. A second decisive check would be to run scGSL on a dataset with matched spatial transcriptomics and ask whether the learned cell-cell edges coincide with spatial adjacency or experimentally validated ligand-receptor contacts.","supporting_citations":[{"cited_title":"Single-cell analysis of childhood leukemia reveals a link between developmental states and ribosomal protein expression as a source of intra -individual heterogeneity,","cited_arxiv_id":null,"evidence_quote":"Supplies the pairwise alignment graph domain adaptation procedure (density-ratio reweighting and label-shift correction) that carries the cross-patient transfer."},{"cited_title":"Risk-associated alterations in marrow T cells in pediatric leukemia,","cited_arxiv_id":null,"evidence_quote":"Supplies the matrix-completion result that justifies using l1 and nuclear-norm penalties to enforce sparsity and low rank in the learned graph."},{"cited_title":"scGCN is a graph convolutional networks algorithm for knowledge transfer in single cell omics,","cited_arxiv_id":null,"evidence_quote":"Establishes the prior GNN-based cell-cell graph annotation approach that scGSL extends and with which it shares the graph-based premise."},{"cited_title":"Comprehensive integration of single -cell data,","cited_arxiv_id":null,"evidence_quote":"Serves as a baseline cell type annotation method in all three disease-specific comparisons."},{"cited_title":"Inferring spatial and signaling relationships between cells from single cell transcriptomic data,","cited_arxiv_id":null,"evidence_quote":"Serves as the strongest or second-strongest baseline across most patient-pair experiments."},{"cited_title":"Inference and analysis of cell -cell communication using CellChat,","cited_arxiv_id":null,"evidence_quote":"Supplies leukemia scRNA-seq patient samples used in the patient-pair transfer experiments."},{"cited_title":"scPML: pathway-based multi-view learning for cell type annotation from single-cell RNA-seq data,","cited_arxiv_id":null,"evidence_quote":"Supplies additional leukemia scRNA-seq patient samples used in the transfer and ablation experiments."},{"cited_title":"KEGG: kyoto encyclopedia of genes and genomes,","cited_arxiv_id":null,"evidence_quote":"Supplies breast and colorectal cancer patient samples, including the E-MTAB-8107 cohort used in the gene-selection ablation."},{"cited_title":"C/EBPα and GATA-2 mutations induce bilineage acute erythroid leukemia through transformation of a neomorphic neutrophil -erythroid progenitor,","cited_arxiv_id":null,"evidence_quote":"Supplies breast cancer patient samples used in the transfer experiments and gene-pair validation."}],"review_version":1}