T cell Perturb-seq Target Discovery

Ranking CRISPR perturbations in primary human CD4+ T cells to find candidate therapeutic gene targets. Built for the "Built with Claude: Life Sciences" hackathon (Research track).

▶ Live report · 3-min demo video

Primary analysis — disease-signature reversal (CMap-style). For each of 11,281 CRISPRi knockdowns and each of three autoimmune diseases (rheumatoid arthritis, psoriasis, multiple sclerosis), a weighted GSEA-style enrichment tests whether the knockdown's genome-wide effect runs opposite the disease's treatment-reversal signature — i.e. the knockdown mimics effective therapy. Cross-disease consensus, a STRING/GWAS network-support gate, and Open Targets tractability leave a druggable shortlist led by ITK, INPP5D, TET2, MAPK14, CD52.

Part two — cross-validation across seven diseases. The top candidates are then re-scored against four more autoimmune diseases (T1D, IBD, lupus, celiac) and classified by breadth. Exactly two are universal regulators (reverse ≥4 of 7 diseases): TET2 and KAT6B — the latter a literature-blind hit on the clinical-stage KAT6A/B pathway.

A second, orthogonal Th1/Th2 cell-state method (below) converges on overlapping targets and supplies the validation / genetics / robustness scaffold. The whole pipeline is deterministic and reproduces exactly across machines (a separate rigor note: docs/methods/R0_reproducibility.md).

• Plan: docs/plan.md
• Tasks: docs/tasks.md
• Original brief: docs/requirements.md
• Dataset: https://virtualcellmodels.cziscience.com/dataset/genome-scale-tcell-perturb-seq

Setup

Fresh Mac (no conda yet)? Run the bootstrap — it installs Miniforge and creates the tcell-perturb env:

bash src/setup/setup_env.sh

Then, in a new terminal, download the processed data (public S3 + GitHub, no login) into data/raw/, and follow docs/tasks.md:

conda activate tcell-perturb
bash src/download/2026_07_08_step_1_download_raw.sh

Notes:

• If you already have conda: conda env create -f envs/environment.yml, then run the download script.
• On Linux, src/setup/linux_setup_and_download.sh does env + download in one step (edit path_folder_output at the top first).
• The 1.6 TiB cell-level *.assigned_guide.h5ad shards are intentionally skipped; local work runs off the summary tables and lazy S3 reads.

Run the analysis

After setup + download:

bash src/run_pipeline.sh      # A → R1–R4 (reversal) → CV1–CV3 (cross-validation) → B–J (convergent Th1/Th2)
python src/reversal/2026_07_11_step_R4_figures.py   # reversal report figures
python docs/build_report.py   # rebuild docs/report.html

To reproduce from scratch or change any analysis decision, see docs/REPRODUCE.md — it maps every step to its script/module/config key and names each change-point (target state, signature, QC, scoring/FDR, disease panel), so the biology can be re-pointed without touching the machinery.

Results

Primary — disease-signature reversal (R1–R3)

1. A stream & cache the 15.6 GB DE matrix from S3 (row-blocks; column slicing is pathological).
2. R1 weighted-GSEA reversal score per knockdown × disease vs published GEO treatment-reversal signatures; permutation null + BH-FDR → RA 1,713 · MS 1,181 · Psoriasis 7 significant knockdowns.
3. R2 cross-disease consensus → 238 genes significant at FDR<0.1 in ≥2 diseases.
4. R3 six-filter plausibility cascade (expression, effect size, STRING/GWAS network support, novelty, threshold stability) + Open Targets tractability → 40-gene master shortlist, 7 directly targetable (led by ITK).

Top reversal candidates (consensus, network-supported, druggable): ITK (→ ritlecitinib/pazopanib), INPP5D (SHIP1), TET2, MAPK14 (p38α), CD52 (→ alemtuzumab), CD33, KAT6A/B. See results/tables/reversal_master_candidates.csv and the Claude-generated, citation-verified mechanistic review docs/reversal_candidate_interpretation.md. The pipeline's determinism / cross-machine reproducibility is documented separately in reversal_reproduction_consistency.md.

Cross-validation — universal multi-disease targets (CV1–CV3)

Extend the top-40 candidates to four more autoimmune diseases (T1D, IBD, lupus, celiac) — seven total — and classify each gene by breadth: 9. CV1 re-score the 40 candidates against each new disease (same weighted-GSEA + pooled null; candidate-panel BH-FDR) — reproduces the reference at Pearson r = 1.0000. 10. CV2 build the 40×7 hit matrix and classify: UNIVERSAL (≥4 diseases) > MULTI-DISEASE (2-3) > DISEASE-SPECIFIC. Exactly two universal regulators: TET2 and KAT6B (both reverse RA, MS, IBD, celiac). 11. KAT6B — the "wait-really" lead: a literature-blind screen nominated the KAT6A/B histone-acetyltransferase pathway in exactly the diseases (MS, IBD) where independent mouse genetics already showed the paralog is causal — and a clinical-stage dual KAT6A/B inhibitor exists. Broad-essentiality safety risk flagged honestly (as is TET2's clonal-hematopoiesis risk). See results/tables/crossval_breadth_classification.csv and docs/crossval_candidate_interpretation.md.

Convergent method + rigor — Th1/Th2 cell state (B–J)

5. B signed Th1/Th2 signature (513 genes) from two published contrasts; C QC → 2,850 / 33,983 reproducible perturbations.
6. D ULM score vs signature (+=KD→Th2, −=KD→Th1) + empirical-null BH-FDR → 843 genes (q<0.10).
7. E recovers known regulators — AUROC 0.92/0.86 (95% CI), PR-AUC 30–47× chance; G independent human genetics (Open Targets, genetics-only) enriches known regulators p=4.2e-11; J cross-study robustness (Spearman ρ=0.66).
8. Convergence: the two orthogonal methods agree on overlapping targets (INPP5D, TET2, VAV1), and the cell-state method honestly flags its own weak hits (PTPRC, HDAC7, RIPK1). See results/tables/final_priority_candidates.csv and docs/candidate_interpretation.md.

Live report: https://hackathon-reversing-autoimmunity.netlify.app/ — or locally docs/report.html (Demo · Voiceover · Manuscript · Methods · Philosophy · Team tabs; the Voiceover tab embeds the AI-narrated demo audio). Demo video (3 min): https://youtu.be/1ov9PXek3gM — a short live intro, then the Demo slides auto-scroll in sync with the AI narration. Manuscript: docs/manuscript.pdf (source docs/manuscript.md; rebuild with python docs/build_manuscript.py). Figures in results/figures/. Submission summary: docs/summary.md. Demo script: docs/demo/transcript.md.