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

Technical Summary: Causal ASCEND

Problem Statement
The paper addresses the challenge of performing causal discovery on high-dimensional multi-omics data. Biological systems possess an inherent hierarchical structure where upstream regulators (background variables, e.g., SNPs, methylation) causally precede downstream effects (foreground variables, e.g., gene expression). While this "two-tiered" ordering offers a natural scaffold for causal inference, existing methods struggle to leverage it effectively at scale.

• Limitations of Current Approaches: Traditional Gene Regulatory Network (GRN) inference methods (e.g., GENIE3, ARACNE, WGCNA) often treat variables as homogeneous, failing to resolve directionality or ignoring the tiered structure. Methods that do attempt to use background variables (e.g., TRIGGER, Confounder Blanket Learner [CBL]) often require conditioning on the entire upstream layer. In high-dimensional settings, conditioning on all upstream variables leads to an exponential blow-up in the number of conditional independence (CI) tests required, creating a "computational wall" that renders genome-scale discovery infeasible.

Methodology: ASCEND
The authors propose ASCEND (Ancestral Scalable Causal discovEry via iNherited Descent), a constraint-based framework designed specifically for high-dimensional, two-tiered biological systems.

• Core Assumption: The method assumes a known causal ordering where background variables ($Z$) precede foreground variables ($X$).
• Divide-and-Conquer Strategy: Instead of conditioning on the full set of background variables, ASCEND employs a dynamic strategy to identify and maintain a set of "nearest ancestors" ($Anc^*$) for each foreground variable.
• Algorithmic Workflow:
  1. Estimation: For each foreground variable, the algorithm estimates the nearest ancestors from the background layer.
  2. Conditioning: Pairwise CI tests between foreground variables are performed by conditioning only on the union of their nearest ancestor sets, rather than the entire background layer.
  3. Inference Rules: The method applies specific inference rules (Minimal Deactivation, Minimal Activation, Conditional Independence) to orient edges and update the set of non-descendants.
  4. Transitive Closure: The process iterates to resolve ancestral relationships, producing a directed acyclic graph (DAG) representing the ancestral structure of the foreground variables.
• Complexity: By restricting conditioning sets to dynamically updated nearest ancestors, ASCEND reduces the complexity of CI testing from exponential to polynomial time relative to the dimensionality of the background layer.

Key Contributions

1. Scalable Framework: ASCEND is the first general framework to leverage known two-tiered biological hierarchies to make genome-wide causal discovery computationally tractable.
2. Dynamic Conditioning: It replaces the static, full-background conditioning sets used in methods like CBL with dynamically updated "nearest ancestor" sets, drastically reducing the number of required CI tests.
3. Directionality Resolution: Unlike standard association-based GRN methods, ASCEND recovers the directionality of causal relationships (ancestral ordering).

Results
The authors evaluated ASCEND using extensive simulations and real biological data:

• Synthetic Data Benchmarks:
  • Performance: Across 12 simulation conditions (varying sample size, sparsity, and signal strength), ASCEND consistently outperformed standard GRN tools (GENIE3, ARACNE, WGCNA) in F1 score, precision, and AUPR. In the primary benchmark ($n=2000$, sparsity=0.9), ASCEND achieved an F1 of 0.589 compared to 0.358 for the next-best method (GENIE3).
  • Speed: ASCEND demonstrated a speedup of two to three orders of magnitude over GENIE3 (e.g., ~0.4s vs. ~149s per replicate).
  • Comparison with Causal Baselines: When compared against other causal discovery algorithms (CBL, GES, LiNGAM, PC), ASCEND achieved the highest F1 scores in sparse, signal-rich regimes (most relevant to real GRNs). It showed superior precision compared to GES and significantly higher decisiveness (coverage) and speed compared to CBL.
• Real Data Application (DGRP):
  • Applied to the Drosophila Genetic Reference Panel (DGRP), ASCEND analyzed 200 inbred lines with genomic variants and transcriptomic profiles.
  • It identified 880 directed causal edges and 61 undirected edges among 250 transcripts.
  • The method successfully recovered a coherent, literature-validated immune module, identifying known upstream regulators (e.g., GNBP-like3, Metchnikowin) and their downstream targets, demonstrating its ability to track real axes of genetic variation.
• Comparison with TRIGGER and BFCS: On a yeast eQTL dataset, ASCEND outperformed TRIGGER and BFCS in recovering both direct and ancestral regulatory relationships, achieving significantly higher F1 scores and Jaccard similarity with curated ground truth.

Significance and Claims
The paper claims that ASCEND bridges the gap between the theoretical appeal of causal discovery and the practical realities of high-dimensional omics data.

• Feasibility: It makes genome-scale causal inference feasible by reducing computational complexity from exponential to polynomial time through the exploitation of biological priors.
• Utility: The method is presented as both a standalone tool for ancestral discovery and a scalable preprocessor that can define constraints for other causal discovery algorithms (e.g., PC).
• Biological Insight: By resolving directionality and identifying upstream master regulators, ASCEND moves analysis beyond static "parts lists" toward functional, directed models of gene regulation, providing high-leverage targets for experimental follow-up.

Limitations Noted
The authors acknowledge that ASCEND relies on the assumption of a known causal ordering between tiers; if this ordering is uncertain or absent, the method is not applicable. Additionally, the current implementation uses linear-Gaussian assumptions and specific CI tests (Fisher's Z), which may require adaptation for non-linear or non-Gaussian relationships. The method also notes that in dense, weak-signal regimes, score-based methods like GES may remain competitive.