causalscale v3.4.0 Replication Package

12 engines. d=30 to 17,787. 3 modalities. 33 cancers. 92% test coverage.

This replication package accompanies the KDD 2027 Datasets & Benchmarks submission. It contains all source code, pre-computed results, figures, and one-click reproduction.

Quick Start (3 commands)

pip install -e . dagma statsmodels
python run_all.py --verify      # ~30 sec: validates installation + result integrity
python run_all.py --figures     # ~10 sec: regenerates all 4 paper figures

What's Included

Directory	Contents
causalscale/	Full source code (12 engines, API, CLI, web)
causalscale/pretrained/	Pre-trained models (DepMap, TCGA, Sachs) + benchmark JSONs
causalscale/core/	11 engine implementations (see Engine List below)
results/	17 pre-computed result JSON files (all paper tables/figures)
figures/	4 paper figures (PDF + PNG for fig3)
scripts/gen_figures/	4 figure generation scripts
scripts/	6 experiment scripts (benchmark, ablation, scaling, etc.)
examples/	4 Jupyter notebooks (biology, finance, neuroscience, drug discovery)
tests/	3 test files (92% coverage)
run_all.py	One-click reproduction script
download_data.py	External data download helper
pan_cancer_scan.py	33-cancer TCGA scan script (Table 4)

Reproduction Guide

Verify Installation (30 seconds)

python run_all.py --verify

Checks: import causalscale, quick d=30 test, result file integrity, figure file integrity, pre-trained models.

Regenerate All Figures (10 seconds)

python run_all.py --figures

Runs all 4 scripts/gen_figures/gen_fig*.py scripts. Reads from results/ directory.

Print All Table Data (instant)

python run_all.py --tables

Prints cached results for Tables 1, 4, 5, 6 from results/*.json.

Run Synthetic Benchmark (~15 min on GPU)

python run_all.py --benchmark

Reproduces Table 1 (NOTEARS vs DAGMA vs causalscale, d=30-150, 5 seeds each).

Run Scaling Experiments (~30 min on GPU)

python run_all.py --scaling

Reproduces Table 5 (LowRankGNN rank sensitivity, r=8-128).

Full Reproduction (~1 hour on GPU)

python run_all.py

Runs benchmark + scaling + ablation.

Paper-to-Package Cross-Reference

Tables

Paper Table	Data Source	Reproduce With
Table 1: Synthetic DAG F1	results/exp1_causalscale_er.json + results/dagma_benchmark.json	python run_all.py --benchmark
Table 2: Paired t-test	Computed from Table 1 data	Derivable from results/exp1_causalscale_er.json
Table 3: Feature comparison	Static (no computation)	In paper TeX
Table 4: ARID1A-MTOR	results/pan_cancer_ckpt.json	python pan_cancer_scan.py (needs TCGA data)
Table 5: LowRank scaling	results/exp9_lowrank_scaling.json	python run_all.py --scaling
Table 6: Memory scaling	results/memory_scaling.json	Pre-computed

Figures

Paper Figure	Script	Output
Fig 1: Architecture	scripts/gen_figures/gen_fig1_architecture.py	figures/fig1_architecture.pdf
Fig 2: Benchmark	scripts/gen_figures/gen_fig2_benchmark.py	figures/fig2_benchmark.pdf
Fig 3: ARID1A-MTOR	scripts/gen_figures/gen_fig3_arid1a_mtor.py	figures/fig3_arid1a_mtor.pdf
Fig 4: Timing/Scaling	scripts/gen_figures/gen_fig4_timing.py	figures/fig4_timing.pdf

All 4 figure PDFs are pre-bundled in figures/. To regenerate: python run_all.py --figures.

12 Engines

Core Engines (dimension-based auto-selection)

Engine	Best For	Method	Key Result
dagma	d <= 150	DAGMA (Bello et al., NeurIPS 2022)	F1=0.989 @ d=150
cluster_aware	d <= 200	Verified NOTEARS with exact DAG constraint	Exceeds NOTEARS at all d, exceeds DAGMA at d=30
transformer	d=200-500	Causal Transformer (Gao 2026)	1,028 edges @ d=200, NOTEARS = 0
lowrank	d > 500	LowRankGNN (Gao 2026)	d=17,787, 88.7% STRING/TRRUST precision

Specialized Engines

Engine	Best For	Method	Key Result
multibatch	Multi-dataset	Dataset-specific residual adapters	0.92 shared-edge precision, 93.3% ASCEND
llm_prior	External knowledge	STRING-derived edge prior injection	+12% F1 over vanilla NOTEARS
bayes_lowrank	Uncertainty quantification	Bayesian bootstrap + low-rank NOTEARS	ECE 0.003@d=500, 71.4% BRCA STRING
sc_causal	Single-cell RNA-seq	NB-LR CI test + PC algorithm	14.8% PBMC STRING, 35.8% cell-type
transfer	Cross-dataset	Warm-start NOTEARS/DAGMA/GOLEM	45/45 wins, 68.1% edge retention @ d=889
multiscale	d=500-5,000	Multi-scale low-rank decomposition	16x KM enrichment over concatenation
multimodal	m >= 2	Cross-modal Frobenius consensus	Multi-omics causal discovery
ensemble	Any	3-engine weighted voting	F1 exceeds best single engine by 22-35%

Plus: PCMCI time-series engine (Runge et al., Sci. Adv. 2019).

External Data (Optional)

The pre-computed results in results/ do NOT require any external data downloads. To reproduce experiments that use external biological data:

python download_data.py --string    # STRING PPI + TRRUST (for validate_against_string)
python download_data.py --tcga       # TCGA gene expression (for pan_cancer_scan.py)
python download_data.py --depmap     # DepMap CRISPR (for genome-scale)

Data sources:

• TCGA: UCSC Xena (https://xenabrowser.net/)
• DepMap: Broad Institute (https://depmap.org/, 24Q2 release)
• STRING PPI v12.0: https://string-db.org/
• TRRUST v2: https://www.grnpedia.org/trrust/

Install

pip install causalscale

GPU auto-detected; CPU fallback supported. Python >=3.10, PyTorch >=2.1.

# Full reproduction
git clone https://github.com/sgao-academics/causalscale
cd causalscale
pip install -e .
python run_all.py --verify

Citation

@inproceedings{gao2027causalscale,
  title={causalscale: Scaling Differentiable Causal Discovery to Genome-Wide Resolution},
  author={Gao, Shuaidong},
  booktitle={Proceedings of the 33rd ACM SIGKDD Conference on Knowledge Discovery and Data Mining (KDD)},
  year={2027},
  note={Datasets \& Benchmarks Track}
}

License

MIT. Author: Shuaidong Gao (ORCID: 0009-0004-5641-3581).