{"id":"bd9838cf-4115-4ad8-b14d-84aa3474033c","arxiv_id":"2607.28090","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Reliability-weighted ridge transport of measured source perturbation responses improves recipient-context mean-effect MSE by 4.1% over a recipient-local low-rank base on Frangieh Perturb-CITE-seq.","lead":"PerturbMap predicts a missing cell-context gene-perturbation effect by transporting the same perturbation’s measured responses from other contexts through ridge maps whose reliability is fixed on held-out training anchors. It offers a practical way to fill structured holes in single-cell atlases without pooling all data centrally.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"No significant objection identified beyond the reader's already-stated transfer and harm caveats.","rationale":"The strongest claim is a carefully scoped empirical result under an explicit information boundary, not a universal virtual-cell guarantee. The paper's own diagnostics (shuffled affine worse by 5.60e-5; equal-token ridge/MLP still behind; route confidences varying 6.2×; Jiang reliability check independent and sealed) support that the gain is from identity-aligned source-response transport plus train-only route weighting rather than from leakage or underpowered baselines. The reader's weakest assumption is exactly the place where the argument is least secure, and the reported 19.5% harm rate already quantifies residual negative transfer. No separate internal contradiction rises to the same load-bearing level. Therefore the CONDITIONAL / MODERATE verdict stands without adjustment: accept the Frangieh claim as stated once artifacts allow re-execution, without elevating it to a settled multi-cohort operating point.","tokens_in":12094,"tokens_out":555,"duration_ms":11861,"concrete_test":"Re-run the frozen Frangieh five-fold protocol with an additional leave-one-context-out stress: fit all route maps and (α, ρ) using only two contexts as sources/recipients in training, then score held queries into the held-out third context; if the paired ΔMSE vs LowRank loses significance or harm rate rises materially above 19.5%, the train-anchor reliability transfer assumption is weaker than the main tables imply.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The reader's weakest assumption already captures the central soft spot: that train-only affine ridge maps and validation-anchor (α, ρ) scores (Sec. 4.4–4.5, Eqs. 5–8) transfer to held query identities without reading sealed recipient responses, while ~19.5% of identities are still harmed vs LowRank. The Frangieh protocol is internally consistent with that claim—identity-held splits, copy/shuffle/token-matched controls, paired bootstrap CI excluding zero on the 4.1% MSE gain, and near-parity to TokPool. I do not find a deeper load-bearing inconsistency (e.g., leakage across the sealed boundary, mis-specified primary metric, or controls that fail to isolate identity-aligned transport). Residual limits are scope and magnitude (condition means; small absolute Δ; single primary cohort), which the reader already flags for a CONDITIONAL verdict.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"PerturbMap addresses source-observed, recipient-unmeasured perturbation response prediction: for a query perturbation measured in source contexts but missing in a recipient context, it combines a recipient-local low-rank base with source-to-recipient affine ridge experts fit on paired training identities, then aggregates accepted routes by train-only reliability (α, ρ) estimated on disjoint validation anchors. On the Frangieh Perturb-CITE-seq melanoma cohort (200 identity-held queries, three contexts), PerturbMap-GR reduces full-effect MSE by 4.1% versus the registered LowRank base (paired 95% CI excluding zero), beats no-token, copy, calibrated-copy, and identity-shuffled affine controls, and stays within 2.82×10−6 MSE of a stronger centralized TokPool reference. Secondary shape metrics and a sealed Jiang multi-source reliability-fusion check move in the same direction. The paper emphasizes a sealed identity-held information boundary and reports residual harm (~19.5% of identities vs LowRank).","tokens_in":12294,"tokens_out":1556,"duration_ms":40085,"significance":"The problem formulation is practically relevant for incomplete multi-context perturbation atlases and is cleaner than standard unseen-perturbation prediction because the query carries measured source-effect tokens. Methodologically the work is careful: identity-held outer folds, manifest-authenticated sealed scoring, information-matched controls (including identity-shuffled affine and equal-token ridge/MLP), paired bootstrap uncertainty, and near-parity to a stronger pooled interface under a weaker client-local training regime. Those design choices are real strengths and make the Frangieh claim falsifiable within its contract. If the result holds under broader cohorts, reliability-weighted response transport would be a useful atlas-completion primitive. The absolute gain is modest and the evaluation is condition-mean rather than full single-cell distributional, so significance is incremental rather than transformative, but the contribution is concrete and well-scoped.","major_comments":[{"comment":"§6.1 / Table 1: the headline 4.1% full-effect MSE gain (Δ = 6.80×10−5 on a LowRank baseline of 1.649×10−3) is statistically supported by a paired CI that excludes zero, but the manuscript does not establish that this absolute improvement is biologically or operationally meaningful for downstream atlas use (e.g., pathway ranking, experimental prioritization). Please add a calibrated discussion of effect size—ideally tied to top-20 gene recovery, sign agreement, or a simple decision-utility proxy—so readers can judge whether the gain justifies deployment versus the simpler LowRank fallback.","section":"§6.1, Table 1"},{"comment":"§4.4–4.5, Eqs. (5)–(8) and §6.1: the central selective-trust claim rests on train-only affine maps and validation-anchor (α, ρ) transferring to held identities, yet PerturbMap-GR still harms 39/200 identities (19.5%) relative to LowRank (Table 2). The paper reports the rate but does not characterize when harm occurs (e.g., by recipient context, route confidence in Table 7, source-response norm/sparsity, or program class). Without that failure-mode analysis, it is hard to know whether ρ is a sufficient proxy or whether the residual harm is structured and avoidable. A short stratified harm audit—and clearer guidance on when to fall back to the base—should be load-bearing for the reliability-weighting narrative.","section":"§4.4–4.5, Eqs. (5)–(8); §6.1; Table 2; Table 7"},{"comment":"§5.1 and §6.4: the primary claim is carried almost entirely by one three-context Frangieh cohort; Jiang is explicitly a secondary sealed reliability-fusion check under a different protocol and is not used to set the Frangieh operating point. That is honest, but it leaves open whether identity-aligned ridge transport plus train-only ρ generalizes beyond this melanoma immune-evasion setting. At minimum, the Discussion should state more sharply what would falsify the method on a new atlas (minimum source support, required paired-anchor count, expected harm ceiling), and if any additional public multi-context resource can be scored under the same sealed contract without retuning, that would substantially strengthen the paper.","section":"§5.1; §6.4; §7"}],"minor_comments":[{"comment":"Fig. 1 and Fig. 2 are referenced as motivation and protocol overview but are not self-contained in the text; captions should state the information boundary (what is sealed vs query-visible) explicitly so the figures stand alone.","section":"Fig. 1; Fig. 2"},{"comment":"§4.3: the perturbation descriptor (correlations with 64 control-anchor genes plus control mean/sd/detection) is important for interpreting LowRank as a no-token baseline; a one-sentence justification for this choice versus simpler gene-identity or embedding features would help non-specialist readers.","section":"§4.3"},{"comment":"Table 1 reports TokPool without a Δ-vs-LowRank column while other methods have one; either add the contrast or state in the caption that TokPool is a ceiling reference only, to avoid visual miscomparison.","section":"Table 1"},{"comment":"§6.2 / Table 4: rank-8/16/32 and MAX variants are statistically tied with GR; the text correctly keeps rank-16 as the frozen headline, but a single sentence on why communication cost (MB) does not change the scientific recommendation would tighten the ablation narrative.","section":"§6.2, Table 4"},{"comment":"Minor prose inconsistencies: “Severalvariantsarestatisticallytied” (missing spaces) in §6.2; “Tab.5testswhether…” similarly; and abstract vs §8 report slightly different MSE decimals (1.581 vs 1.5809×10−3)—align rounding.","section":"§6.2; §8; Abstract"},{"comment":"Related Work cites strong virtual-cell / OT baselines; a clearer one-line contrast that PerturbMap predicts condition-mean effects under sealed recipient outcomes (not cell-level counterfactual generation) would reduce scope confusion.","section":"§2"}],"recommendation":"minor_revision","confidential_remarks":"I agree with the reader’s CONDITIONAL/moderate stance and with the skeptic that there is no sealed-boundary leakage or control-design failure. The work is above the bar for a methods contribution if the authors add effect-size interpretation and a harm-mode audit; I would not demand a new primary cohort as a condition of acceptance, but I would weight generalization language carefully in the Decision letter. Fit is better for a methods / computational-biology venue than for a broad AI venue expecting large absolute gains."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The thing worth knowing is that this is a clean methods paper on a real atlas problem: you have the same perturbation measured in source contexts and need the missing recipient mean effect, without opening the sealed recipient rows. They define that contract tightly and stick to it.\n\nWhat is new is not ridge maps or residual weighting in isolation. It is the sealed source-observed / recipient-unmeasured setup plus train-only reliability (alpha, rho on disjoint validation anchors) with the right controls: no-token LowRank and FedAvg, raw and calibrated copy, identity-shuffled affine, equal-token ridge/MLP, and a stronger centralized TokPool that they nearly match (gap 2.82e-6). On Frangieh, PerturbMap-GR cuts full-effect MSE by 4.1% vs the registered low-rank base, paired CI excludes zero, wins 161/200 identities, and lifts cosine top-10 retrieval 74.5% → 80.5%. Split discipline, manifest authentication, and route tables look serious. Math is standard affine ridge in a shared train-only basis; citations cover GEARS/CPA, OT/virtual-cell, and federated fusion without obvious padding.\n\nSoft spots are real but already in proportion. Absolute delta is small. About 19.5% of identities are still harmed vs LowRank; the adaptive gate helps directionally but is not headline-ready. Target is condition means on one primary melanoma cohort (Jiang is a secondary reliability check only). Free parameters (K, ridge grid, base net) are frozen before test, which is good practice, not magic. No shipped code in the text, so reproducibility is mid-scale until artifacts land.\n\nThis is for people building multi-context perturbation atlases or federated-style response reuse, not for full virtual-cell distribution modeling. I would bring it to reading group as a protocol example. It deserves a serious referee; I would cite the task definition and control battery if I work in this regime. Send to peer review.","headline":"Careful identity-held protocol for reusing measured source perturbation effects across contexts; 4.1% MSE gain is real on Frangieh but small, with residual harm and condition-mean scope.","tokens_in":12944,"tokens_out":507,"would_cite":true,"duration_ms":10191,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"Measured source-context perturbation responses can fill missing recipient-context effects when transported through train-only reliability-weighted maps rather than copied or ignored.","keywords":["single-cell perturbation","cross-context transfer","response transport","reliability-weighted fusion","Perturb-CITE-seq","low-rank response basis","identity-held evaluation"],"falsifier":"On the same identity-held Frangieh protocol, if sealing recipient answers and scoring full-effect MSE showed PerturbMap no better than the recipient-local base (paired CI covering zero) or no better than identity-shuffled affine transport, the central reuse claim would fail.","tokens_in":12893,"feed_emoji":"🧬","tokens_out":972,"duration_ms":23228,"temperature":0.7,"pith_summary":"Single-cell perturbation atlases almost never measure every genetic intervention in every cellular context. When a perturbation has already been measured elsewhere, throwing that experiment away wastes query-specific evidence, but naively copying it across contexts can transfer the wrong signal. This paper defines that setting as source-observed, recipient-unmeasured prediction and introduces PerturbMap: a recipient-local low-rank base plus source-to-recipient ridge transports whose acceptance and weights are fixed only on training validation anchors, before any held recipient answer is opened. On a multi-context melanoma Perturb-CITE-seq cohort under identity-held splits, the method cuts full-effect error by about four percent versus the local base, beats copy, shuffle, and no-token controls, and nearly matches a stronger centralized pooled reference. A sympathetic reader cares because atlas missingness is structural: the same intervention is often already measured in other contexts, and a disciplined reuse rule turns those measurements into usable recipient predictions without pooling every client’s training interface.","feed_headline":"Source experiments cut missing-context error by 4.1%","feed_subtitle":"Reliability-weighted transport reuses measured perturbation responses without copying the wrong signal across cell contexts.","key_machinery":"Train-only reliability-weighted response transport: shared low-rank coordinates, a recipient-local base, per-route affine ridge experts fit on paired training perturbations, and route interpolation strength α plus reliability ρ estimated solely on disjoint validation anchors, then used at inference to weight accepted source proposals for the sealed query.","core_discovery":"On identity-held Frangieh Perturb-CITE-seq queries, PerturbMap’s train-only reliability-weighted transport of measured source responses reduces full-effect MSE by 4.1% relative to a recipient-local low-rank base that never sees the query source tokens, outperforms FedAvg, zero-response, raw-copy, calibrated-copy, and identity-shuffled affine controls, and stays within 2.82×10−6 MSE of a stronger centralized token-matched pooled reference, with same-recipient top-10 cosine retrieval rising from 74.5% to 80.5%.","pith_inferences":["The same sealed source-token contract could be stress-tested on combination perturbations or on contexts farther apart than the three melanoma conditions studied here.","If route reliability were allowed to depend on pathway or gene-program labels, harm on the worst identities might fall further without opening recipient answers.","Atlas imputation pipelines could expose per-route α and ρ as audit trails so experimentalists see which source experiments actually drove a filled-in effect."],"forward_implications":["Atlas builders can treat a perturbation measured in other contexts as query-time evidence for a missing recipient mean effect, under an identity-held seal.","Client-local training with validation-fixed route weights can approach a stronger centralized token-matched pooled trainer without sharing a pooled training interface.","Route-specific trust is warranted: source-to-recipient paths are not exchangeable, and reliability weighting can beat uniform fusion when several sources are available.","Negative transfer remains real for a minority of identities, so harm-rate reporting and optional query-adaptive gates are part of responsible deployment, not optional extras.","The method completes condition-mean effects; uses that need full single-cell distributions still require generators that keep the same sealed source-token boundary."],"fun_headline_variants":["PerturbMap cuts missing-context MSE 4.1% via reliability-weighted transport","Source-to-recipient ridge experts lift full-effect accuracy over low-rank base","Train-only route weights reuse measured perturbations without wrong-signal copy","Same-recipient top-10 cosine retrieval rises 74.5% to 80.5% with PerturbMap","PerturbMap stays within 2.82e-6 MSE of stronger centralized pooled reference"],"cache_read_input_tokens":128,"weakest_assumption_plain":"Maps and route trust scores learned only from training and validation perturbation pairs still describe how held query interventions should move from source contexts into the recipient.","fun_headline_variants_meta":{"raw":{"variants":["PerturbMap cuts missing-context MSE 4.1% via reliability-weighted transport","Source-to-recipient ridge experts lift full-effect accuracy over low-rank base","Train-only route weights reuse measured perturbations without wrong-signal copy","Same-recipient top-10 cosine retrieval rises 74.5% to 80.5% with PerturbMap","PerturbMap stays within 2.82e-6 MSE of stronger centralized pooled reference"]},"model":"grok-4.5","effort":"low","cost_usd":0.004868,"raw_usage":{"total_tokens":1409,"prompt_tokens":842,"num_sources_used":0,"completion_tokens":99,"cost_in_usd_ticks":48684000,"prompt_tokens_details":{"text_tokens":842,"audio_tokens":0,"image_tokens":0,"cached_tokens":128},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":468,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":842,"tokens_out":99,"duration_ms":8527,"temperature":1.0,"reasoning_tokens":468,"cache_read_input_tokens":128,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-31T18:05:39.633529+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"On the same identity-held Frangieh protocol, if sealing recipient answers and scoring full-effect MSE showed PerturbMap no better than the recipient-local base (paired CI covering zero) or no better than identity-shuffled affine transport, the central reuse claim would fail.","supporting_citations":[],"review_version":1}