Pith. sign in

REVIEW 2 major objections 5 minor 17 references

Causal ASCEND: Scalable Two-tier Causal Discovery on High Dimensional Multi-omics Data

T0 review · 2 major / 5 minor · reviewed 2026-07-13 · grok-4.5

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read ASCEND recovers ancestral gene-regulatory relationships at genome scale by conditioning only on dynamically updated nearest ancestors, not the full upstream layer.

desk verdict 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. read the letter →

arxiv 2607.04527 v2 pith:3GI2UHIQ submitted 2026-07-05 stat.ML cs.LGq-bio.GN

Stephen Asiedu , David Watson This is my paper · ORCID
classification stat.MLcs.LGq-bio.GN
keywords causaldiscoverymulti-omicsgeneregulatorynetworkstwo-tierstructurenearestancestorsconditionalindependencehighdimensionality
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

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.

What carries the argument

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).

What would settle it

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.

Watch

Extended reading notes

Core claim

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.

Load-bearing premise

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.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

2 major / 5 minor

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.

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 (2)
  1. 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.
  2. 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.
minor comments (5)
  1. Figure 1 caption and Step-2 text refer to Anc*(X,T_Xi) while the formal definition uses Pa(X;T); unify notation.
  2. Table 1 reports Dir. acc. only for ASCEND; a short note that competitors cannot orient would avoid the impression of missing columns.
  3. DGRP variance-selection cutoff (top 250 genes) is a free parameter; a brief sensitivity check or justification would strengthen the biological claims.
  4. Typo: “scales properly and much faster” in the abstract; “recepter” → “receptor” in the DGRP section.
  5. Supplementary Algorithm 1 uses bracket notation for (de)activation that is defined only later; a forward pointer would help.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: ASCEND's soundness and complexity claims are proved from d-separation and Markov-blanket arguments under an explicit two-tier assumption, and empirical claims are scored against external ground truth.

full rationale

The derivation chain is self-contained. ASCEND initialises non-descendant sets from the stated two-tier ordering (Assumption 1), learns nearest-ancestor blankets via IAMB, conditions on the guarded union S_ij, and orients with standard (de)activation rules (R1–R3) drawn from Entner et al., Magliacane et al., and Claassen & Heskes. Soundness (Theorem 1), localisation equivalence to full-background conditioning (Lemma 4), lazy-oracle completeness relative to CBL (Theorem 4), and polynomial complexity (Theorem 5) are proved in the supplement from those assumptions and an exact CI oracle; they do not reduce to a fitted quantity or to a renamed empirical pattern. Self-citation of Watson & Silva (CBL, 2022) is present and frames the completeness comparison, but CBL is an independent published baseline whose full-background conditioning is strictly more expensive; ASCEND's nearest-ancestor construction is a new, proved localisation rather than an imported uniqueness claim or ansatz. Empirical F1/precision/runtime claims are evaluated against simulator ground-truth ancestral skeletons and external curated networks (YEASTRACT+), not against quantities fitted from the same targets. No self-definitional loop, fitted-input-as-prediction, or load-bearing self-citation chain was found.

Assumptions & free parameters 4 free parameters · 3 assumptions · 1 invented entities

The method rests on two standard causal assumptions plus a domain-specific biological ordering; free parameters are ordinary significance thresholds; no new physical entities are postulated.

free parameters (4)
  • alpha (CI test threshold) = 0.05
    Fixed at 0.05 for pairwise tests and Markov-blanket discovery; controls Type-I error and therefore the operating point of precision/recall.
  • alpha_mb (Markov-blanket FDR) = 0.05
    IAMB false-discovery rate, also set to 0.05; affects size of nearest-ancestor sets.
  • maxiter = 10
    Maximum number of outer iterations; set to 10 in experiments.
  • variance-selection cutoff (DGRP) = 250
    Top 250 genes by variance retained after expression filtering; post-hoc choice that shapes the real-data network.
assumptions (3)
  • domain assumption Two-tier ordering: no foreground variable is an ancestor of any background variable (Assumption 1).
    Initialises every non-descendant set T_X = Z and is required for the guarded conditioning set to satisfy the valid-conditioning property (Lemma 3).
  • standard math Markov and faithfulness: conditional independence in the data coincides with d-separation in the DAG (Assumption 2).
    Standard causal-discovery assumption that lets oracle CI statements be treated as graph statements; invoked throughout the soundness proofs.
  • domain assumption Linear-Gaussian SEM for synthetic data generation and partial-correlation tests.
    Used both to generate ground-truth graphs and as the working model for Fisher-z tests; non-linear or non-Gaussian regimes are left for future work.
invented entities (1)
  • Nearest-ancestor (relativised parent) set Pa(X; T)
    purpose: Minimal valid conditioning set that replaces the full background layer while preserving d-separation for ancestral queries.
    Defined in Definition 1 as the Markov blanket of X restricted to known non-descendants; the algorithmic novelty of ASCEND rests on maintaining and guarding these sets.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Causal ASCEND: Scalable Two-tier Causal Discovery on High Dimensional Multi-omics Data." pith.science (2026). https://pith.science/paper/3GI2UHIQ

@misc{pith2026260704527,
  author       = {Pith},
  title        = {Pith review of: Causal ASCEND: Scalable Two-tier Causal Discovery on High Dimensional Multi-omics Data},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3GI2UHIQ}},
  note         = {Machine review of arXiv:2607.04527}
}
read the original abstract

Biological systems exhibit a hierarchical structure, characterised by directed flow from upstream regulators to downstream effects. Although this ordering provides a natural scaffold for causal inference, most causal discovery and GRN methods either ignore the tiered organisation or condition on all upstream variables, which becomes infeasible for high-dimensional omics data. We present ASCEND (Ancestral Scalable Causal discovEry via iNherited Descent), a constraint-based framework that leverages known two-tiered structure to enable genome-scale causal discovery. ASCEND introduces a divide-and-conquer strategy that maintains dynamically updated ancestral conditioning sets for each downstream variable, dramatically reducing the number of conditional independence tests required, and achieves polynomial-time complexity where traditional approaches face exponential blow-up. Through extensive simulations and real biological data, we demonstrate that ASCEND accurately recovers ancestral relationships, scales properly and much faster, and outperforms existing gene regulatory network inference methods in both causal precision and computational efficiency. The algorithm's ability to resolve directionality makes it particularly suited for integrating multi-omic data where upstream regulators (e.g., SNPs, methylation sites) and downstream responses (e.g., gene expression) are measured jointly.

Figures

Figures reproduced from arXiv: 2607.04527 by the authors.

Figure 1
Figure 1. ASCEND workflow. (I) The two-tiered data structure where some background Z causally precede foreground X. (II) Structure learning: Test pairwise conditional independencies among X variables while conditioning only on Anc∗ (X, TXi ) ∪ Anc∗ (X, TXj ). (III) Final Output after iterations and closure depicting ancestral relationship. Edges may represent direct or indirect causes. met in observational human multi-omic da… view at source ↗
Figure 2
Figure 2. F1 at matched edge count across the full sweep. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. The genomic distribution of causal influence scores. The y￾axis represents the causal out-degree (the number of downstream genes controlled by a specific locus). Peaks on Chromosomes 2R and 3R identify localised regulatory hotspots. Identification of Genomic Causal Hubs To transition from statistical associations to a structural understanding of the DGRP transcriptome, we employed ASCEND, which oriented 72 causal ed… view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: Comparison of ASCEND, TRIGGER and BFCS. Left: edge￾level precision, recall and F1 scores for direct and ancestral regulatory relationships. Right: per-regulator Jaccard similarity between predicted and curated regulons. directed regulatory interactions using instrument…
Figure 5
Figure 5. Figure 5: ASCEND vs. CBL runtime, computational cost, and discovery quality. [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: ASCEND vs. four causal discovery baselines on synthetic benchmarks. [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: Minimal DAGs detected by the orientation rules, with witness W [PITH_FULL_IMAGE:figures/full_fig_p011_7.png]
Figure 7
Figure 7. Figure 7: Minimal DAGs detected by the orientation rules, with witness W [PITH_FULL_IMAGE:figures/full_fig_p012_7.png]
Figure 8
Figure 8. Figure 8: Example graphs illustrating how we exploits transitivity and antisymmetry to infer causal structure. Theoretical Guarantees Throughout this section I is the exact oracle of Assumption 2, so all independence statements are d-separation statements. In this section we pro…
Figure 8
Figure 8. Figure 8: Example graphs illustrating how we exploits transitivity and antisymmetry to infer causal structure. Theoretical Guarantees Throughout this section I is the exact oracle of Assumption 2, so all independence statements are d-separation statements. In this section we pro…

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

17 extracted references · 3 linked inside Pith

  1. [1]

    doi: 10.3389/fcell.2022.1016400. S. Asiedu and D. Watson. Multi-omic causal discovery using genotypes and gene expression.arXiv preprint arXiv:2505.15866,

    doi:10.3389/fcell.2022.1016400 2022

  2. [2]

    doi: 10.1073/pnas.2007743117. I. G. Bucur, T. Bussel, T. Claassen, and T. Heskes. A bayesian approach for inferring local causal structure in gene regulatory networks. InInternational Conference on Probabilistic Graphical Models, pages 37–48. PMLR,

    doi:10.1073/pnas.2007743117

  3. [3]

    Chevalley, Y

    M. Chevalley, Y. Roohani, A. Mehrjou, J. Leskovec, and P. Schwab. Causalbench: A large-scale benchmark for network inference from single-cell perturbation data.arXiv 9 preprint arXiv:2210.17283,

    arXiv

  4. [4]

    Claassen and T

    T. Claassen and T. Heskes. A logical characterization of constraint-based causal discovery.arXiv preprint arXiv:1202.3711,

    arXiv

  5. [5]

    H. Dai, I. Ng, G. Luo, P. Spirtes, P. Stojanov, and K. Zhang. Gene regulatory network inference in the presence of dropouts: a causal view.arXiv preprint arXiv:2403.15500,

    arXiv

  6. [6]

    doi: 10.1002/PMIC.200600442. D. Entner, P. Hoyer, and P. Spirtes. Data-driven covariate selection for nonparametric estimation of causal effects. In Artificial intelligence and statistics, pages 256–264. PMLR,

    doi:10.1002/pmic.200600442

  7. [7]

    doi: 10.3390/ijms18020412. P. Hu, R. Jiao, L. Jin, and M. Xiong. Application of causal inference to genomic analysis: Advances in methodology. Frontiers in Genetics, 9,

    doi:10.3390/ijms18020412

  8. [8]

    doi: 10.1007/978-3-319-46807-5

    doi:10.1007/978-3-319-46807-5

Show all 17 references
  1. [9]

    doi: 10.1093/nar/ gkac1041. I. Tsamardinos, C. F. Aliferis, A. R. Statnikov, and E. Statnikov. Algorithms for large scale markov blanket discovery. InFLAIRS, volume 2, pages 376–81,

    doi

  2. [10]

    doi: 10.1186/gm339. D. S. Watson and R. Silva. Causal discovery under a confounder blanket. InUncertainty in Artificial Intelligence, pages 2096–2106. PMLR,

    doi

  3. [11]

    Xis a non-descendant ofY

    10 Asiedu et al. Supplementary Material ASCEND: Method Details ASCEND is a constraint-based causal-discovery framework for high-dimensional, two-tier biological systems in which a set of background variablesZis known to causally precede a set of foreground variablesX. ASCEND e...

    2012

  4. [12]

    Each iteration performs four steps

    and updated by T (t+1) X ={Y∈Z∪X:Y⪯ t X},(1) where⪯ t is the partial order recorded after iterationt. Each iteration performs four steps. Step 1 (Nearest ancestors). For eachXcompute Pa(X;T (t) X ) by Markov-blanket discovery restricted toT (t) X (Definition 1); in practice we...

    2003

  5. [13]

    Remark 1 (Why the both-sides condition is required)The naive unionPa(X i;T i)∪Pa(X j;T j)is unsafe: a mediator MonX i →M→X j is a non-descendant ofX j yet a descendant ofX i

    We then testX i ⊥ ⊥Xj |S ij by a likelihood-ratio comparison of linear models, equivalent to the partial-correlation test and uniformly most powerful under our assumptions [Lehmann and Romano, 2005]. Remark 1 (Why the both-sides condition is required)The naive unionPa(X i;T i)...

    2005

  6. [14]

    (R2)Activation.If∃W:W̸ ⊥ ⊥X i |S \W ∪[X j], thenX i ⪯X j

    and Watson and Silva [2022]: (R1)Deactivation.If∃W:W⊥ ⊥X j |S \W ∪[Xi], thenX i ≺X j. (R2)Activation.If∃W:W̸ ⊥ ⊥X i |S \W ∪[X j], thenX i ⪯X j. (R3)Independence.IfX i ⊥ ⊥Xj |S ij, thenX i ∼X j. Minimal examples of structures identifiable by these rules are shown in Fig

    2022

  7. [15]

    [2016], entailing Xi ∈An(X j), i.e.X i ≺X j

    and Magliacane et al. [2016], entailing Xi ∈An(X j), i.e.X i ≺X j. R2.The activationW̸ ⊥ ⊥X i |S \W ∪[X j] means addingX j opens aW–X i path; a node opens a path only as (a descendant of) a collider, soX j is such a collider, givingX i /∈De(Xj), i.e. Xi ⪯X j. R3.SupposeX i ⊥d ...

    2016

  8. [16]

    These are exactly the (de)activation and separation queries of R1–R3, conditioned on the common known non-descendants

    Alazy oracle algorithmstarts from an uninformativeMand updates each round using only oracle answers to queries of two types: (i)W⊥ ⊥X i |S ij \ {W} ∪ϕ(X j), withW∈S ij andϕ(X j)∈ {∅,{X j }}; and (ii)X i ⊥ ⊥Xj |S ij. These are exactly the (de)activation and separation queries o...

    2022

  9. [17]

    Convergence is reached only when no further rule can fire.□ Theorem 4(Lazy-oracle completeness)ascend-oracleis lazy-oracle complete: no lazy oracle algorithm dominates it

    and contains every nearest- ancestor witness, so every rule application available to the full- conditioning oracle is available to ASCEND and fires the same relation. Convergence is reached only when no further rule can fire.□ Theorem 4(Lazy-oracle completeness)ascend-oracleis...

    2022

Pith tools

Reviewed July 13, 2026 · model on record in the stance chip above.