{"id":"225468f1-9c14-47d8-8e77-a9eaa92bd3f4","arxiv_id":"2607.04527","paper_version":2,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.5,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"ASCEND recovers ancestral gene-regulatory relationships at genome scale by replacing full-background conditioning with dynamically updated nearest-ancestor sets under a known two-tier hierarchy.","lead":"ASCEND finds directed ancestral relationships among genes by conditioning only on small, dynamically updated nearest-ancestor sets drawn from a known upstream layer such as SNPs. It makes genome-scale multi-omic causal discovery practical and recovers coherent regulatory hubs that match known biology.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"No significant objection identified beyond the two-tier assumption already flagged by the reader.","rationale":"The central claim is that nearest-ancestor conditioning yields higher-precision ancestral recovery at far lower cost than full-background or association-based baselines under a known two-tier ordering. The reader correctly isolates Assumption 1 as the condition that must hold for the initial non-descendant sets and the guarded union (Eq. 2) to remain valid; the supplement’s inductive argument for soundness rests on that invariant. All other design choices (IAMB blankets, R1–R3 rules, closure) are standard or explicitly proved relative to the lazy-oracle class. Empirical superiority is demonstrated across twelve simulation cells, DGRP hubs, yeast eQTL, and head-to-head causal baselines, with matched-K F1 and wall-clock numbers that are large enough to be robust to modest implementation variation. Because no stronger or independent load-bearing flaw was located, the CONDITIONAL verdict and HIGH confidence stand; the concrete reverse-edge injection test simply quantifies the already-flagged sensitivity.","tokens_in":25769,"tokens_out":528,"duration_ms":6121,"concrete_test":"Re-run the primary simulation cell (n=2000, sp=0.9, R^{2}=0.7) after injecting a controlled fraction of reverse X\to Z edges (e.g. 5 % of possible cross-tier edges) while keeping all other generative parameters fixed; if ASCEND’s F1 and direction accuracy remain statistically indistinguishable from the no-reverse baseline, the practical robustness claim is strengthened; a sharp drop confirms the assumption is load-bearing exactly as stated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The reader's weakest assumption (Assumption 1: no X\to Z edges) is correctly identified as the load-bearing condition for soundness of the guarded conditioning sets S_ij and the initial T_X = Z. Under that assumption the supplement proofs (Lemmas 1–3, Theorems 1–5) establish soundness, lazy-oracle completeness relative to CBL, and polynomial complexity; the empirical claims (F1 0.589 vs 0.358, 27–5000\times speed-ups) are supported by the reported tables and figures. No additional internal inconsistency, hidden circularity, or unstated assumption that would independently undermine the central algorithmic claim was found. The linear-Gaussian simulation regime and post-hoc gene filtering are acknowledged limitations but do not break the argument under the stated conditions.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"ASCEND is a constraint-based causal discovery method for two-tier multi-omics data (background Z known to precede foreground X). It maintains dynamically updated nearest-ancestor conditioning sets for each foreground variable, tests pairwise CI among X variables only on the guarded union of those sets, and orients ancestral relations via deactivation/activation/independence rules plus transitive closure. Under the two-tier ordering, Markov and faithfulness assumptions, the supplement proves soundness, lazy-oracle completeness relative to CBL, and polynomial complexity. Empirical claims: on linear-Gaussian simulations ASCEND beats GENIE3/ARACNe/WGCNA on matched-K F1 (primary cell 0.589 vs 0.358) and CBL/GES/LiNGAM/PC on precision and runtime (27–5000× vs CBL); on DGRP it recovers immune hubs; on yeast eQTL it outperforms TRIGGER and BFCS on ancestral F1.","tokens_in":25974,"tokens_out":896,"duration_ms":11744,"significance":"If the two-tier assumption holds, the paper supplies a practical, theoretically grounded route to genome-scale ancestral discovery that existing full-background or association-based methods cannot match. The nearest-ancestor localisation is a clean algorithmic contribution with formal guarantees (soundness, lazy-oracle completeness, O(d_X^{2}(d_X+d_Z)s̄) complexity) and large, reproducible speed-ups. Code and preprocessed matrices are released. The DGRP immune-module recovery and yeast ancestral-benchmark results give concrete biological utility. The work is a genuine advance over CBL for the high-dimensional multi-omics setting the field actually faces.","major_comments":[{"comment":"Assumption 1 (no X→Z edges) is load-bearing for Lemmas 1–3 and Theorems 1–5 and for the initial T_X=Z. The Discussion correctly flags this, but the manuscript never quantifies robustness when the assumption is mildly violated (e.g., reverse causation or unmeasured feedback). A short simulation or analytic bound under controlled reverse-edge rates would make the scope of the guarantees clearer for multi-omic applications where the ordering is only approximately known.","section":null},{"comment":"All synthetic grids (Tables 1–3, Figures 2, 5–6) use linear-Gaussian SEMs and Fisher-z tests. The Discussion notes that non-linear/non-Gaussian regimes will need alternative CI tests, yet no such experiment is reported. Because the primary claim is superiority for multi-omics, at least one non-linear or non-Gaussian cell (or a real-data ablation with a non-parametric CI test) is needed to show that the F1 and runtime advantages are not artefacts of the linear-Gaussian match.","section":null}],"minor_comments":[{"comment":"Figure 1 caption and Step-2 text refer to Anc*(X,T_Xi) while the formal definition uses Pa(X;T); unify notation.","section":null},{"comment":"Table 1 reports Dir. acc. only for ASCEND; a short note that competitors cannot orient would avoid the impression of missing columns.","section":null},{"comment":"DGRP variance-selection cutoff (top 250 genes) is a free parameter; a brief sensitivity check or justification would strengthen the biological claims.","section":null},{"comment":"Typo: “scales properly and much faster” in the abstract; “recepter” → “receptor” in the DGRP section.","section":null},{"comment":"Supplementary Algorithm 1 uses bracket notation for (de)activation that is defined only later; a forward pointer would help.","section":null}],"recommendation":"minor_revision","confidential_remarks":"The paper is a solid incremental advance over the authors’ own CBL work; the heavy self-citation is justified by the direct comparison but the novelty claim should be framed as “localisation of CBL’s conditioning sets” rather than a wholly new paradigm. Fit for a methods-oriented bioinformatics or causal-ML venue is good; the linear-Gaussian limitation is the only real risk of over-claim."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The real contribution here is not the two-tier idea itself (that is CBL) but the guarded nearest-ancestor construction: keep only Pa(X; T) for each foreground node, update T from the orientation rules, and condition on the both-sides-safe union. That change is what turns an exponential or full-background method into something that finishes in sub-second time while still matching CBL’s lazy-oracle power. The supplement actually proves soundness, the localisation lemma, and the O(d_X^{2} (d_X + d_Z) s̄) bound; the empirical tables back the claim (primary-cell F1 0.589 vs GENIE3 0.358, 27–5000× vs CBL, direction accuracy ~77 % in the sparse regime).\n\nWhat they do well: clean experimental design (matched-K F1, 50-rep cells, 81-cell causal grid, explicit timeouts), honest coverage of NA verdicts, and a DGRP application that recovers a coherent Toll/immune module rather than a random gene list. Code and data are promised on GitHub. Self-citation of CBL is heavy but legitimate; they are measuring against the natural baseline.\n\nSoft spots, in proportion. The load-bearing assumption is exactly the one the reader flagged: no X → Z edges. If that fails, the initial T_X = Z and the guarded sets lose soundness. Simulations are linear-Gaussian; real multi-omics will need non-linear tests. DGRP filtering (top-250 variance genes, MAF window) is post-hoc and shapes the hubs they report. None of these break the algorithmic claim under the stated conditions; they just bound the domain of use.\n\nThis is for people who already have matched multi-omic layers and want directed ancestral structure without conditioning on thousands of SNPs. It is not a general causal-discovery breakthrough, but it is a practical, formally grounded tool that removes a real computational wall. I would send it to referees; the proofs and the scaling numbers deserve a careful look, not a desk reject. Worth citing if you work on multi-omic GRNs or two-tier methods.","headline":"Solid engineering of nearest-ancestor conditioning that makes two-tier ancestral discovery actually usable at omics scale, with proofs and large empirical gains over CBL and GRN baselines.","tokens_in":26576,"tokens_out":551,"would_cite":true,"duration_ms":6249,"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":"ASCEND recovers ancestral gene-regulatory relationships at genome scale by conditioning only on dynamically updated nearest ancestors, not the full upstream layer.","keywords":["causal discovery","multi-omics","gene regulatory networks","two-tier structure","nearest ancestors","conditional independence","high dimensionality"],"falsifier":"On a multi-omic dataset with independently verified reverse edges from expression into genotype or methylation, ASCEND should produce systematically wrong ancestral orientations or collapse in coverage relative to a method that does not assume the two-tier order.","tokens_in":26651,"feed_emoji":"🧬","tokens_out":566,"duration_ms":8498,"temperature":0.7,"pith_summary":"Biological systems have a natural hierarchy: upstream variables such as SNPs or methylation sites precede downstream variables such as gene expression. Most causal-discovery methods either ignore that hierarchy or condition on every upstream variable at once, which is impossible at omics scale. ASCEND turns the known two-tier ordering into a computational advantage. For each foreground variable it maintains a small, growing set of nearest ancestors and uses only those as the conditioning set for conditional-independence tests. The resulting algorithm recovers directed ancestral edges with higher precision and F1 than standard gene-regulatory and causal baselines, runs orders of magnitude faster, and has polynomial rather than exponential complexity. On Drosophila multi-omic data it recovers a coherent immune-module network whose hubs match known Toll-pathway genes, showing that the same machinery works on real population data.","feed_headline":"Nearest-ancestor sets make causal discovery scale to genomes","feed_subtitle":"ASCEND recovers directed ancestral edges faster and more precisely than full-background methods","key_machinery":"Dynamically updated nearest-ancestor conditioning sets (the \"inherited descent\" of each foreground variable), guarded so that every member remains a non-descendant of both endpoints of the tested pair; these sets replace the full background layer in the three orientation rules (deactivation, activation, independence).","core_discovery":"Under a known two-tier causal ordering, the guarded union of nearest-ancestor sets is a minimally valid conditioning set for ancestral orientation. Conditioning on those sets alone lets ASCEND recover the same ancestral relations as full-background methods, but with far fewer and cheaper conditional-independence tests, polynomial runtime, and higher precision on sparse regulatory graphs.","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["Nearest-ancestor sets cut CI tests for genome-scale discovery","ASCEND recovers ancestral edges via guarded two-tier sets","Two-tier order yields polynomial ancestral causal discovery","Minimal nearest-ancestor conditioning matches full-background precision","ASCEND orients multi-omics edges faster than full GRN methods"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"The method assumes that no downstream (foreground) variable can ever cause an upstream (background) variable; if reverse causation or feedback exists, the initial and updated conditioning sets lose their soundness guarantee.","fun_headline_variants_meta":{"raw":{"variants":["Nearest-ancestor sets cut CI tests for genome-scale discovery","ASCEND recovers ancestral edges via guarded two-tier sets","Two-tier order yields polynomial ancestral causal discovery","Minimal nearest-ancestor conditioning matches full-background precision","ASCEND orients multi-omics edges faster than full GRN methods"]},"model":"grok-4.5","effort":"low","cost_usd":0.007016,"raw_usage":{"total_tokens":1704,"prompt_tokens":754,"num_sources_used":0,"completion_tokens":84,"cost_in_usd_ticks":70160000,"prompt_tokens_details":{"text_tokens":754,"audio_tokens":0,"image_tokens":0,"cached_tokens":128},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":866,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":754,"tokens_out":84,"duration_ms":6837,"temperature":1.0,"reasoning_tokens":866,"cache_read_input_tokens":128,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-13T06:57:42.333365+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"On a multi-omic dataset with independently verified reverse edges from expression into genotype or methylation, ASCEND should produce systematically wrong ancestral orientations or collapse in coverage relative to a method that does not assume the two-tier order.","supporting_citations":[],"review_version":2}