Rank CD4+ T-cell Perturb-seq drug targets by whether the readout is dependable, not by how large it is. G-perturb decomposes each perturbation's effect across guides, donors, and activation states with generalizability theory, then re-ranks targets by how reproducibly their signal generalizes.

Built for Built with Claude: Life Sciences (Researcher track, 2026), in partnership with Gladstone Institutes. All work in this repository is done from scratch during the hackathon, per the event rules.

🧬 Preprint: posted on bioRxiv on 2026-07-15, DOI 10.64898/2026.07.13.738312 (read on bioRxiv). Manuscript licensed CC BY-NC 4.0; code Apache-2.0.

Start here: ▶ 3-minute demo video · short summary · full paper (PDF) · demo slides + voiceover.

📄 Full design rationale: docs/design.md (canonical design). 📝 Research write-up (methods + results, paper voice): docs/writeup.md — the bioRxiv-format LaTeX preprint is in manuscript/ (main.pdf). The method is a distribution-light generalizability decomposition — facet variance shares plus an irreducible replication floor. The deliverables are a reliability-weighted target ranking and the dependability findings below, all resting on that decomposition.

On the real 44.6 GB joint pseudobulk (18,129 genes × 4 donors × 3 conditions), decomposed genome-wide:

• Signal was hidden under measurement noise. Raw pseudobulk shows dependable target-signal in only 40 genes; removing the measurement-error floor (estimated from the non-targeting controls) lifts that to 7,674 genes (42%) — the dependable hits were masked by noise, not absent.
• The ranking reshuffles. Weighting effect size by a per-target dependability coefficient reorders the shortlist: the effect-only and dependability-weighted rankings correlate at Spearman ρ = 0.74 overall, yet 50 of the top-100 effect-only targets drop out of the dependability-weighted top-100. Targets you would miss by effect size alone climb sharply — RPS3 #79→#4, QARS1 #53→#5, RPP21 #124→#6, IMP4 #96→#12, NMD3 #158→#14 are dependable but effect-underranked: the "validate these next" candidates.
• Reliability is guide-limited. A D-study shows two guides per gene cap generalizability near 0.53; reaching 0.70 needs ~15 guides, and no number of donors suffices — add guides, not donors (and this holds within every activation state, not just pooled).
• Context-specificity recovers real biology, unprompted. Slicing dependability by activation state surfaces the T-cell-receptor signaling module — CD3D, CD3G, CD247, ZAP70, LAT — as reliably measurable only in activated T cells (resting dependability at the noise floor). With no gene labels, the method reconstructs the TCR module: the sharpest evidence it measures something real.

Figures (submission-grade): analysis/resolution/realdata/figures/. Ranked tables: analysis/resolution/realdata/target_ranking.csv + target_context_specificity.csv. The whole pipeline regenerates every number from one re-run.

Every scientific decision (estimator, thresholds, why a facet is random vs fixed, what counts as identifiable) was filed as a GitHub issue through issue-driven development (IDD) and spec-driven development (Spectra) — open-source Claude Code plugins I built (PsychQuant/issue-driven-development). That auditable decision ledger let a competing frontier model, GPT-5.6 Sol, red-team the design: it returned BLOCKED with 11 methodological holes (3 critical), and each was worked through in the same loop with Claude — gates frozen before results, the method chosen by pre-registered synthetic recovery. The resolutions were then cross-verified across models: a second adversarial pass via codex-pro (Codex) and multi-agent cross-checking through PsychQuant/parallel-ai-agents — another package I wrote that dispatches a task to independent Claude and Codex agents and cross-compares their outputs. The blocked verdict and every resolution are versioned in the repo (docs/reviews/, openspec/changes/, the issue history) — the audit trail is the evidence.

(IDD and Spectra are my pre-existing open-source tooling; the G-perturb analysis they govern here is the new hackathon work.)

Target prioritization in Perturb-seq usually ranks by differential-expression strength — effect size, z-score, FDR, number of downstream DE genes. That answers one question:

Did this perturbation move the transcriptome in this dataset?

It does not answer the question target discovery actually depends on:

Will the effect hold up with another guide, another donor, another cell-state context?

These are different questions. A target can have a large apparent effect but be fragile (two guides disagree, one donor drives it, weak knockdown, off-target flag). A moderate-effect target that is consistent across guides and donors may be the better validation bet. G-perturb treats each perturbation effect as a measured score and decomposes its variance across guide, donor, and condition facets — reporting how much of the effect is dependable versus irreducible idiosyncrasy, distribution-light. The reliability-weighted target ranking (effect magnitude × dependability) is the primary deliverable; the variance decomposition is the engine underneath it.

The experimental design maps cleanly onto a psychometric measurement design:

Classical test theory writes X = T + E. Generalizability theory decomposes the error into named sources:

E = E_guide + E_donor + E_condition + E_interactions + E_residual

A crossed-facet random-effects model over a scalar (or low-dimensional) perturbation score:

y[t,g,d,c] = μ + T_t + G_g + D_d + C_c
             + (TG) + (TD) + (TC) + (GD) + (GC) + (DC) + (TGDC)

with a per-target empirical dependability score

R_dep,t = σ²_signal,t / ( σ²_signal,t + σ²_error,t )

R_dep,t is a target-specific quantity inspired by G-theory. The classical G-theory coefficients — Eρ² (relative decisions) and Φ (absolute decisions) — are design-level: one per measurement design, not one per target. So the per-target score is named R_dep,t, not classical Φ_t.

Two honest constraints shape how this is estimated (not optional caveats — they drive the method):

• Only 2 guides / 4 donors. Per-target variance components are under-identified (2 guides = 1 df). The estimable "complete" decomposition is a hierarchical / partial-pooling REML with target as a random facet, so σ²_guide, σ²_donor, … are estimated by borrowing strength across targets, and R_dep,t is an empirical-Bayes–shrunk quantity. Not per-target ANOVA.
• Condition is biology, not just error. Rest / Stim8hr / Stim48hr are real T-cell states. So the project reports two rankings: a global one (condition treated as a random facet, entering error) and a context-specific one (condition treated as a fixed stratifier, one universe per state).

Priority_t = E_t × R_dep,t × Q_t          (× optional Biology_t)

• E_t — effect magnitude (effect-norm, mean |z|, or number of downstream DE genes)
• R_dep,t — empirical dependability (first-pass from released guide/donor correlation fields; full version from the variance components above). Reported as a scalar for ranking (#2), or as a per-domain profile (R^guide, R^donor, R^condition) for a diagnostic evidence map (#3).
• Q_t — quality penalty: weak on-target knockdown, single-guide estimate, off-target/neighbor flags, low cell support
• Biology_t (optional) — pathway / cell-state / disease-genetics relevance, kept explicit so dependability is never mistaken for importance

Zhu et al. (2025), Genome-scale perturb-seq in primary human CD4+ T cells maps context-specific regulators of T cell programs and human immune traits (Marson & Pritchard labs), bioRxiv doi:10.64898/2025.12.23.696273. Genome-scale CRISPRi, ~22M primary human CD4+ T cells, 4 donors × 3 conditions (Rest, Stim8hr, Stim48hr). Public S3, no credentials:

s3://genome-scale-tcell-perturb-seq/marson2025_data/

Note: the flat DE_stats.suppl_table.csv is a reduced 16-column release — it does not carry the reliability correlation fields (guide_correlation_all, donor_correlation_*). Those live in GWCD4i.DE_stats.h5ad .obs, so even the heuristic MVP reads the h5ad. The per-facet .h5mu files carry the raw guide-/donor-resolved effect vectors needed to fit the crossed variance-component model. Raw data is not committed to this repo (see analysis/data/CODEBOOK.json + fetch_data.sh).

• Reliability-weighted target list — conventional DE rank vs reliability-weighted rank, with rank shift.
• High-effect / low-reliability targets (large DE, guides or donors disagree — fragile).
• Moderate-effect / high-reliability targets (better validation bets than a large fragile hit).
• Context-specific dependable targets (robust within Rest / Stim8hr / Stim48hr).
• Figures: effect-vs-reliability scatter, rank-shift plot, guide/donor agreement, QC waterfall.
• Reproducible notebook/script from released summary statistics — no 22M-cell processing required.

G-perturb/
├── manuscript/   # bioRxiv-format LaTeX preprint (main.tex, main.pdf, figures)
├── SUMMARY.md    # short written summary
├── demo/         # 1920×1080 slide deck (slides.html) + read-aloud voiceover
├── analysis/
│   ├── data/         # released summary statistics + CODEBOOK.json (large files gitignored)
│   └── resolution/   # the pipeline: gates, controls, synthetic recovery, real-data decomposition, ranking
├── docs/         # design rationale, research write-up, the Sol adversarial review
├── openspec/     # spec-driven-development change records (audit trail)
├── archive/      # exploratory work not used in the final pipeline
├── LICENSE       # Apache-2.0
└── NOTICE

Che Cheng — Postdoctoral Research Fellow, Institute of Statistical Science, Academia Sinica · https://che-cheng-website.vercel.app · ORCID https://orcid.org/0000-0003-3376-7833 · GitHub https://github.com/kiki830621

Apache License 2.0 — © 2026 Che Cheng.