{"id":"a3699823-1f00-42e7-bcd4-6f06794b0bf7","arxiv_id":"2608.00985","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":9,"one_line_summary":"CoCoS adds gated, complementary-view contrastive learning to masked expression reconstruction and reports the best mean point estimates among compared models on ten cell-type annotation datasets and six GRN networks.","lead":"This paper introduces CoCoS, a pretraining method that learns single-cell representations by contrasting two complementary gene-expression views of each cell rather than only reconstructing masked gene values. It reports higher average accuracy on ten cell-type annotation benchmarks and six gene-regulatory-network evaluations than several existing single-cell foundation models.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Transfer claim depends on downstream evaluation cells being absent from the 1.8M-row CELLxGENE pretraining corpus; no overlap audit is reported.","rationale":"The paper's own framing makes transfer the central claim. The dataset section says pretraining follows Qi et al. data construction from CELLxGENE and downstream evaluation uses Qi et al.'s ten datasets. Since Qi et al. assembled those datasets from public single-cell collections, overlap is plausible. The paper's 'labels do not enter pretraining' sentence addresses label leakage only, not cell identity leakage. A model trained on the same cells could achieve inflated k-NN accuracy through memorized cell-level identity, independent of learning transferable gene-value correspondence. The GRN experiment similarly uses BEELINE networks built from scRNA-seq data that may appear in CELLxGENE. I considered alternatives: single-seed training is a statistical concern but affects all compared methods equally and is explicitly disclosed; the close control variants (balanced-random panels at 74.34 vs. 74.82 k-NN) are concerning, but the overlap confound is logically prior because it threatens even the best-case comparison. Thus the overlap audit is the single most load-bearing test. The reader identified the same assumption, so I agree. The conditional verdict already reflects the need for this audit; no change in verdict is required beyond what the reader recommended.","tokens_in":11725,"tokens_out":3524,"duration_ms":35576,"concrete_test":"Compute the overlap between the 1,799,684 training rows (after excluding held-out cohorts) and each of the ten downstream cell-type evaluation matrices plus the six BEELINE networks: first by exact source-row/barcode identity, then by near-duplicate expression profile (e.g., Pearson correlation >0.99 over shared variable genes). Report the fraction of downstream cells with a training match. If nonzero, re-run the frozen-embedding k-NN and linear-probe evaluation after removing all matching training rows from pretraining (or retraining if needed) and compare CoCoS-GPC against the split-view reconstruction-only control. If gains persist after removal, the transfer claim survives; if they shrink materially, the reported advantage is partly memorization of pretraining cells rather than transfer.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central transfer claim — that properly bounded whole-cell contrastive supervision yields more discriminative and transferable cell representations than pure reconstruction-based methods — requires that the downstream evaluation cells are not in the pretraining corpus. The paper establishes that downstream labels do not enter pretraining, but it never establishes that the cells themselves do not. The pretraining corpus is 1,813,780 rows following the Qi et al. data construction from CELLxGENE; the ten cell-type datasets are 'assembled by Qi et al. (2025)' from the same ecosystem, and the six BEELINE networks are built from scRNA-seq atlases that plausibly overlap the same CELLxGENE collection. Exact cell overlap would let contrastive pretraining memorize cell identities and inflate k-NN and linear-transfer scores. The reported margins are small enough that overlap artifacts could account for them: +3.83 k-NN over scGPT and +0.88 over the uniform-sampling control. The sentence 'Annotation and GRN labels are used only for downstream evaluation; neither enters pretraining' addresses label leakage, not sample overlap. Because the transfer conclusion is the stated purpose of the paper, this unverified disjointness assumption is the most load-bearing point.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes CoCoS, a contrastive pretraining framework for single-cell transcriptomics that augments masked expression reconstruction with three components: co-expression-guided gene partitioning into two complementary views, expression-aware contrast-set construction with fixed-identity value derangements, and competence-gated contrastive onset. The authors evaluate frozen embeddings on ten cell-type annotation datasets and six BEELINE GRN networks, reporting mean k-NN and linear-probe accuracies as well as AUROC/AUPRC point estimates. The central claim is that properly bounded whole-cell contrastive supervision yields more discriminative and transferable cell representations than reconstruction alone. The paper is candid about limitations, including one training seed, point-estimate GRN results, and a short-fork diagnostic that does not measure downstream performance.","tokens_in":12055,"tokens_out":7421,"duration_ms":72594,"significance":"If the central claim holds, CoCoS offers a practical recipe for adding cell-level contrastive supervision to masked single-cell pretraining, addressing two real transcriptome-specific challenges: constructing biologically valid positive pairs and avoiding gene-identity shortcuts. The method is principled, and the paper includes useful controlled diagnostics: the fixed-identity shortcut experiment in Figure 3 is well designed, and the same-architecture scBenchmark baseline provides a more trustworthy comparison than the external scGPT checkpoint. The manuscript is also unusually transparent about its limitations. However, the transfer claim currently rests on an unverified data-disjointness assumption and on single-seed point estimates, and the balanced-random panel control nearly matches the full method, weakening the specific case for co-expression-guided partitioning. These issues are fixable and do not invalidate the framework, but they must be addressed before the broad transferability conclusion is established.","major_comments":[{"comment":"The transfer conclusion presupposes that the downstream evaluation cells are not in the 1,813,780-row pretraining corpus. The paper says only that annotation and GRN labels do not enter pretraining; it does not audit cell overlap. Both the pretraining corpus and the ten Qi et al. cell-type datasets are drawn from the CELLxGENE ecosystem, and the BEELINE GRN atlases plausibly overlap it. Exact or near overlap would let contrastive pretraining memorize cell identities and inflate the reported gains; the margins (+3.83 k-NN over scGPT, +0.88 over uniform sampling) are small enough for such artifacts to matter. Please add an overlap audit (e.g., barcode or expression-profile deduplication) for all downstream sets, or restrict the claim to non-transfer performance.","section":"Datasets / Downstream evaluation"},{"comment":"All main results use one pretraining seed (42) and one fixed downstream split. The paper's central comparative claim is therefore supported only by point estimates without uncertainty or significance. For example, Table 3 differences between balanced-random and proposed (0.48 k-NN, 0.25 linear) and between uniform and expression-aware sampling (0.88 k-NN, 0.25 linear) are within plausible training noise, and the +3.83 k-NN margin over scGPT could be seed-specific. The split-seed 42-44 reevaluation varies only the downstream split for a fixed checkpoint and does not address pretraining variance. Please provide multiple pretraining seeds (or paired bootstrap) for the main tables.","section":"Experimental Setup; Tables 1-3"},{"comment":"The partition objective in Eq. (2) maximizes the total co-occurrence weight of cross-view gene pairs, but the algorithm that follows assigns each gene opposite the panel containing more of its already-assigned neighbors, counting each neighbor once 'irrespective of w_gh.' Edge weights therefore enter only through the visit order, not through the placement decision, so the method does not actually optimize Eq. (2). If the count-based heuristic is intentional, it should be stated as an approximation, and the claim that views are 'co-expression-guided' should be tied to the heuristic actually used. As written, the formal objective and the implemented procedure are inconsistent.","section":"Method, Eq. (2)"},{"comment":"The balanced-random panel control is nearly as strong as the full method (74.34 vs. 74.82 k-NN; 78.44 vs. 78.85 linear). Since this control differs only in how the two panels are chosen, Table 3 does not demonstrate that co-expression-guided partitioning contributes materially to the proposed system. Given that the first contribution is 'co-expression-guided gene partitioning,' the paper should report an ablation that isolates the partition choice while holding all other components (including the flip probability and GPC triggering) fixed, and state the size of the partition effect.","section":"Table 3"}],"minor_comments":[{"comment":"The description of Geneformer contains an incomplete cross-reference: 'as detailed in Section.' with no section number.","section":"Experimental Setup"},{"comment":"The repertoire score 'q_p = ||std(Z_p)||_2 / p_erank_p' is ambiguous: define p and 'erank' explicitly, and clarify how Z_p is normalized before computing the standard deviation.","section":"Method, Eq. (4)"},{"comment":"The GPC row is described as an unmatched reference, but the surrounding text tends to read as a comparison with the fixed-onset controls. Please state explicitly that the GPC configuration differs in multiple factors and therefore does not by itself establish superiority of the data-dependent onset.","section":"Table 4"},{"comment":"The text reports 74.87 ± 0.06 mean k-NN over downstream split seeds 42-44, while Table 1 reports 74.82 for seed 42. Clarify whether the checkpoint, evaluator, or averaging procedure differs, or correct the discrepancy.","section":"Reevaluation paragraph"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is honest about its limitations, and the same-architecture scBenchmark control is a strength. However, the transfer claim cannot be evaluated without an overlap audit; I would be willing to accept after that is addressed and after at least a small multi-seed robustness check. The balanced-random panel closeness also needs an explicit interpretation, as it bears on the novelty of the co-expression-guided partition component."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a real empirical contribution, honestly reported, with one partially load-bearing assumption that should be checked before the transfer claim is marketed.\n\nWhat's new: the specific combination of co-expression-guided gene partitioning, fixed-identity value derangements for hard negatives, and competence-gated contrastive onset. They explicitly say cell-level contrast is not new and credit scConcept, Tabula, LangCell. The diagnostic for the gene-identity shortcut is the best part: from a shared checkpoint, adding K=4 identity-matched negatives collapses the confidence gap from 0.358 to 0.028, which directly demonstrates the mechanism works. They also compare against fixed-onset controls, a split-view reconstruction-only control, and a balanced-random panel, and they report the balanced-random panel is only 0.48 k-NN points behind the full method. That's honest reporting, even though it thins the claim that co-expression partitioning is essential.\n\nSoft spots, in proportion. First, the main transfer claim assumes the downstream evaluation cells are not in the pretraining corpus. The paper says labels do not enter pretraining, but that is not the same as saying the cells themselves are disjoint. Both the pretraining corpus and the ten downstream datasets are assembled from CELLxGENE via Qi et al. If there is measurable overlap, the +3.83 k-NN gain over scGPT and +0.88 over uniform sampling could be partly memorization. This is a testable question: a cell-barcode overlap audit or a nearest-neighbor leakage check would settle it. Given the margins are single-digit points, this is not a fatal flaw but it should be addressed.\n\nSecond, all main tables use one training seed and one downstream split. The authors report downstream split-seed variation for the selected checkpoint only, not training-seed variation. The GRN winners also vary across the six networks; the mean AUROC/AUPRC is highest but there is no uncertainty estimate. These are standard tightening requests, not reasons to reject.\n\nOverall: the point-estimate claims are internally consistent, the method is clearly specified, and the paper is candid about its comparability limitations. The value for the field is a well-defined pretraining recipe with a plausible mechanism. I'd send it to peer review, and the main requests would be a data-overlap audit, multi-seed training, and code release.\n\nFor a reading group working on single-cell representation learning, it's worth an hour. If I were writing a paper on contrastive single-cell pretraining, I'd cite this. Recommend engaging with it and sending to serious peer review.","headline":"A solid, honest empirical recipe for contrastive single-cell pretraining; the transfer claim needs a data-overlap audit and multi-seed runs before it is fully convincing.","tokens_in":12543,"tokens_out":3348,"would_cite":true,"duration_ms":30121,"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":"Cell-level contrastive pretraining on complementary gene views produces more discriminative, transferable single-cell representations than masked reconstruction alone.","keywords":["single-cell transcriptomics","contrastive learning","cell representation learning","masked expression reconstruction","single-cell foundation models","cell-type annotation","gene regulatory network inference","pretraining objective"],"falsifier":"An overlap audit: compare the 1,799,684 training rows against the cell rows used in the ten cell-type annotation datasets and the six regulatory-network evaluation sets. If a substantial fraction of downstream cells or their exact transcriptomic profiles appear verbatim in the training pool, the reported k-NN and GRN gains would be partly memorization rather than transfer, weakening the central claim; if overlap is negligible, the transfer conclusion stands. A secondary check is to run the K=0 versus K=4 shortcut diagnostic for the full 140,600-update training rather than 1,000 updates, testin","tokens_in":11596,"feed_emoji":"🧬","tokens_out":13742,"duration_ms":121251,"temperature":0.7,"pith_summary":"Single-cell foundation models are pretrained mainly by reconstructing masked expression values, an objective that supervises individual genes but says nothing about the geometry of the whole-cell embedding. This paper argues that a cell-level contrastive objective, applied the right way, fills that gap. The method, CoCoS, gives each cell two complementary views by splitting its genes along co-expression structure, builds hard negatives that permute expression values while keeping gene identities fixed, and switches the contrastive loss on only after a label-free readiness check on held-out cells. With these pieces, the CoCoS-GPC embeddings beat reconstruction-only references on cell-type annotation (74.82% vs 70.99% mean k-NN against scGPT, 78.85% vs 77.02% linear probe) and give the highest mean AUROC/AUPRC point estimates on six gene-regulatory networks. If the paper is right, a bounded contrastive term on complementary views is a viable complement to masked reconstruction in single-cell pretraining.","feed_headline":"Contrasting two gene views beats reconstruction-only pretraining","feed_subtitle":"The embeddings transfer better to cell-type annotation and gene-network inference than masked-reconstruction models.","key_machinery":"The load-bearing mechanism is the complementary view pair. A weighted gene co-occurrence graph is bipartitioned so strongly co-occurring genes land in opposite panels; each cell is masked twice, once per panel, so every gene value is observed in one view and masked in the other. This yields a biologically anchored positive pair while keeping the reconstruction objective intact. Around that pair, CoCoS builds a contrast set of same-view and cross-view in-batch cells plus $K=4$ fixed-identity value derangements, values permuted but gene identities and masking preserved, forcing the encoder to use gene–value correspondence rather than gene-set composition. A competence gate keeps the contrastiv","core_discovery":"The paper claims that masked expression-value reconstruction alone leaves whole-cell representation geometry unsupervised, and that a properly bounded contrastive objective over complementary transcriptomic views closes that gap. CoCoS keeps every measured gene–value pair intact and instead partitions each cell's genes by co-expression into two complementary views; it builds contrast sets whose hardest negatives are fixed-identity value derangements, so gene-set composition cannot separate positives from negatives; and it leaves the contrastive weight at zero until label-free probes on held-out cells show stable cross-view concordance, low reconstruction error, and preserved embedding divers","pith_inferences":["My inference: the label-free competence gate is a general pattern. Any pretraining run that couples a stabilizing objective with a potentially corrupting one could trigger the second objective using the same held-out readiness probes, not just single-cell contrast.","My inference: because each positive pair is two views of the same measured cell rather than a perturbed copy, the design sidesteps the open question of which expression perturbations preserve cell identity; the same view-construction idea could transfer to other contrastive methods on sparse, high-dimensional molecular data.","My inference: the paper reports one training seed and point estimates, so the reported margins are not yet distributional claims; a multi-seed rerun of the ten-dataset protocol would show whether the k-NN gain is stable or within noise, especially where per-network GRN winners vary.","My inference: a natural scale-up test follows from the paper's own numbers. Expression-aware sampling shifted the onset from 96,000 to 93,000 steps and improved accuracy by 0.88 k-NN points, suggesting the sampler and the gate interact; ablating them on larger corpora would show whether that interaction grows or saturates."],"forward_implications":["If CoCoS-GPC is right, single-cell foundation-model pretraining should include a bounded cell-level contrastive objective alongside masked reconstruction: frozen whole-cell embeddings transfer better to cell-type annotation than the reconstruction-only reference and scGPT (74.82% vs 70.99% mean k-NN).","The fixed-identity derangement diagnostic shows a measurable shortcut: without them, confidence against identity-matched negatives sits at 0.639 with an ordinary–identity gap of 0.358; with $K=4$ these move to 0.969 and 0.028, so the contrast objective learns gene–value correspondence, not gene-set composition.","Gating contrastive onset on label-free competence beats fixed schedules: the GPC trigger at step 93,000 improves over immediate contrast by 1.50 k-NN points and over both one- and two-epoch warmups, making data-dependent onset a load-bearing design choice.","The same frozen encoder supports gene-regulatory edge prediction, with the highest mean AUROC (0.8733) and AUPRC (0.7120) point estimates among compared variants on six networks, although per-network winners differ.","Because both views share one encoder and one optimizer update per minibatch, the contrastive addition is compatible with a modest training budget (10 epochs, one RTX 4090), making the recipe practical to reproduce."],"supporting_citations":[{"why":"Supplies the pretraining data construction, the Transformer backbone, and the unified ten-dataset cell-forward evaluation protocol used in the main comparisons.","marker":"Qi et al. 2025"},{"why":"scGPT, the external foundation-model baseline whose cell-type and GRN scores CoCoS-GPC is compared against.","marker":"Cui et al. 2024"},{"why":"Geneformer V2-104M, the second external reference checkpoint evaluated under the same downstream cell rows and splits.","marker":"Theodoris et al. 2023"},{"why":"Source of the six 1000-gene regulatory networks used for the GRN edge-prediction evaluation.","marker":"Pratapa et al. 2020"},{"why":"The contrastive learning paradigm whose naive transfer to transcriptomes motivates the three CoCoS adaptations.","marker":"Chen et al. 2020"},{"why":"Grounds the argument that expression-value perturbations can alter cellular identity, motivating complementary views over augmented values.","marker":"Kotliar et al. 2019"},{"why":"LangCell, prior masked-gene-plus-contrastive pretraining that the paper positions CoCoS beyond.","marker":"Zhao et al. 2024"},{"why":"Tabula, prior joint column-wise reconstruction and row-wise contrastive pretraining that CoCoS distinguishes its design from.","marker":"Ding et al. 2025"},{"why":"scConcept, the closest prior contrastive scheme over disjoint gene panels that CoCoS builds on and differs from.","marker":"Bahrami et al. 2025"}],"fun_headline_variants":["Contrastive gene-view pretraining outdoes masked reconstruction","CoCoS learns cell representations from complementary gene views","Two transcriptomic views beat one reconstruction objective for cells","Contrastive cell pretraining beats reconstruction on transfer tasks","Gene co-expression splits sharpen cell embeddings beyond reconstruction"],"cache_read_input_tokens":2816,"weakest_assumption_plain":"The transfer claim presupposes that the ten downstream cell-type datasets and six gene-regulatory networks are not materially present in the 1.8-million-row pretraining pool built from the same public single-cell atlas; the paper guarantees only that downstream labels never enter pretraining, not that the same cells never appear in both.","fun_headline_variants_meta":{"raw":{"variants":["Contrastive gene-view pretraining outdoes masked reconstruction","CoCoS learns cell representations from complementary gene views","Two transcriptomic views beat one reconstruction objective for cells","Contrastive cell pretraining beats reconstruction on transfer tasks","Gene co-expression splits sharpen cell embeddings beyond reconstruction"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00054,"raw_usage":{"total_tokens":2419,"prompt_tokens":733,"completion_tokens":1686,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":477,"completion_tokens_details":{"reasoning_tokens":1609}},"tokens_in":477,"tokens_out":1686,"duration_ms":13524,"temperature":1.0,"reasoning_tokens":1609,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T00:34:52.506156+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"An overlap audit: compare the 1,799,684 training rows against the cell rows used in the ten cell-type annotation datasets and the six regulatory-network evaluation sets. If a substantial fraction of downstream cells or their exact transcriptomic profiles appear verbatim in the training pool, the reported k-NN and GRN gains would be partly memorization rather than transfer, weakening the central claim; if overlap is negligible, the transfer conclusion stands. A secondary check is to run the K=0 versus K=4 shortcut diagnostic for the full 140,600-update training rather than 1,000 updates, testin","supporting_citations":[{"cited_title":"Hinton , title=","cited_arxiv_id":null,"evidence_quote":"The contrastive learning paradigm whose naive transfer to transcriptomes motivates the three CoCoS adaptations."},{"cited_title":"Aurel and Tabrizi, Shervin and Hodis, Eran and Melton, Douglas A","cited_arxiv_id":null,"evidence_quote":"Grounds the argument that expression-value perturbations can alter cellular identity, motivating complementary views over augmented values."},{"cited_title":"Nature Methods , volume=","cited_arxiv_id":null,"evidence_quote":"Tabula, prior joint column-wise reconstruction and row-wise contrastive pretraining that CoCoS distinguishes its design from."}],"review_version":1}