Claude Hackathon: Life Sciences (Researcher Track), 7–13 July 2026. 📄 Live poster: https://sneakypeat.github.io/HumanCD4CoDEGNet/

Finding. The causal trans-regulatory network of primary human CD4⁺ T cells is hub-dominated and sparse-but-pleiotropic in every activation state — the first causal confirmation of the Barton/Pritchard (Cell Genomics 2026) topology prediction, which was made from twin-study heritability with no perturbation data. The architecture is shape-invariant but identity-labile: hub-dominance is pinned at Gini ≈ 0.92 across Rest → 8 h → 48 h even as the network densifies +32 % and up to 59 % of the top-100 hubs are displaced, with the TCR signalosome switching on as the activation-state broadcaster. The cell keeps the shape of its control while swapping out who is in control.

Data: Marson-lab genome-scale CRISPRi Perturb-seq, ~22 M primary human CD4⁺ T cells (Zhu, Dann et al. 2025, bioRxiv 10.64898/2025.12.23.696273).

Shape (Gini) invariant; identity (which genes are hubs) turns over. Every confound guarded (power ρ ≈ −0.20; validated-KD-only Gini 0.91; not KD-efficiency).

Disease lead: the hubs that switch on specifically with activation are ~2× enriched for monogenic-disease (ClinVar) genes vs stable hubs (OR 2.3, p = 0.01) — a shortlist of druggable state-specific control points (ZAP70, ITK, LCK, PTPRC, IL12RB2). See disease_hubs.py.

cd ~/CoDEG_Tcell
python3.12 -m venv venv && source venv/bin/activate
pip install -r requirements.txt

# either:
jupyter lab notebooks/HumanCD4CoDEGNet_architecture.ipynb   # streams live, 0 errors, figures inline
# or:
python scripts/architecture.py                              # CLI: writes results JSON + both figures

Reproduces in seconds from two released summary fields of the single public S3 file (GWCD4i.DE_stats.h5ad): per-perturbation .obs (n_downstream, ontarget_significant) and per-gene varm/measured_genes_stats_* (n_regulators) — the atlas's authoritative DE statistics, streamed by byte-range. No credentials, no local data.

notebooks/
  HumanCD4CoDEGNet_architecture.ipynb  THE submission notebook (streams live, executed, figures inline)
scripts/
  architecture.py           degree distributions, hub-dominance, rewiring, confound guards, figures
  generality_replogle.py    external replication: Replogle 2022 K562/RPE1 (hub-dominance generalizes)
  further_tests.py          plasticity~complexity (not motifs), locality null, hierarchy scoped out
  disease_hubs.py           disease lead: state-specific hubs enriched for monogenic-disease / druggable genes
  analyze_all.py, run_day1.py         build the committed data inputs (per_gene_full.csv, B_masked.npz)
  build_poster.py           regenerate the poster (docs/index.html)
artifacts/
  figures/fig_arch1..5.png             topology · rewiring · generality · further-tests · disease
  architecture_results.json            architecture + rewiring + confound-guard numbers
  architecture_replogle_results.json   generality-test numbers (K562/RPE1)
  further_tests_results.json           plasticity / locality / hierarchy outcomes
  disease_hubs_results.json            disease-enrichment numbers + candidate shortlist
  arch_perturbation_outdegree.csv      per-perturbation out-degree across states (shipped resource)
  per_gene_full.csv, B_masked.npz      committed data inputs (per-gene stats; on-target-masked edge matrix)
  MECHANISM_NOTE.md                    a mechanism we tested and RETIRED (refuted) — kept for honesty
docs/index.html           the poster, served as a GitHub Pages site
SUMMARY.md                one-page pitch          SUBMISSION.md   full writeup

• Dataset shapes / obs fields / varm groups read from the file, not the docs.
• Edge-total cross-check: sum of out-degree == sum of in-degree per state (592k / 780k / 663k) — in- and out-degree describe the same causal network.
• KD-detectability guard: all cross-state rewiring restricted to regulators whose knockdown was validated (ontarget_significant) in both compared states — so hub emergence is genuine, not "the gene became expressible."
• Hub-dominance is identical on validated-KD-only edges (Gini 0.91 vs 0.92); out-degree is negatively correlated with cells-per-perturbation (not a power artifact).

• The raw n_regulators / n_downstream fields were released by the atlas authors; our contribution is the topology framing, the causal test of Barton/Pritchard (2026), the invariance constant, and the guarded turnover. See SUBMISSION.md for full positioning vs Zhu/Marson 2025 and Mihai 2025 (correlational).
• A secondary essential-gene buffering result is real but not a discovery (Feng et al. 2026 reported the direction, expression-controlled, in iPSCs). A mechanism we explored ("activation routes around essential genes") was refuted and retired — see artifacts/MECHANISM_NOTE.md.