T-cell Evidence Workbench
An evidence-first bioinformatics workbench for auditing and prioritizing primary human CD4+ T-cell CRISPRi Perturb-seq results. A deterministic Python pipeline audits 33,983 perturbation-condition rows across four donors and three culture conditions, applies six registered source-QC gates, and produces 1,652 eligible rows plus 224 cross-condition candidates. The Next.js analyst interface connects every shortlist, comparison, robustness scenario, and finding card to inspectable evidence and provenance. It prioritizes candidates for expert follow-up; it does not claim independent biological validation.
Proof points
- 33,983 Rows Audited
- 224 Cross-Condition Genes
- 6 Registered QC Gates
Technologies
Python, Pandas, Bioinformatics, Perturb-seq, Next.js, TypeScript, React Flow, ELK, Pytest
Links
Problem
A genome-scale CD4+ T-cell Perturb-seq screen contains approximately 22 million cells and 1.58 TiB of cell-level files, while experimental follow-up is expensive. Scientists need a feasible way to audit source results, understand every exclusion, and prioritize candidates without overstating what summary-table analysis can prove.
Approach
Built a deterministic Python pipeline around four public summary tables, six pre-registered source-QC gates, within-condition geometric-mean ranking, weakest-condition cross-condition ranking, and artifact-level provenance. Added a Next.js analyst workspace with candidate search, gene dossiers, evidence graphs, comparisons, robustness analysis, constrained finding cards, and manifest-verified data access.
Results
Audited 33,983 perturbation-condition rows and produced 1,652 eligible rows: 497 in Rest, 593 at 8 hours, and 562 at 48 hours. Identified 224 genes that passed every source-QC gate across all three conditions, with SETDB1 leading the pinned cross-condition run at a weakest-condition score of approximately 0.915. Every result remains framed as prioritization for independent follow-up, not biological validation.
T-cell Evidence Workbench: Bioinformatics Engineering Report
1. Executive Summary
T-cell Evidence Workbench is an evidence-first analysis and review system for a genome-scale CRISPRi Perturb-seq screen of approximately 22 million primary human CD4+ T cells from four donors. The project turns public perturbation-level summary tables into an auditable shortlist of source-QC candidates for Rest, 8-hour stimulation, and 48-hour stimulation conditions.
The workflow deliberately does not reprocess the approximately 1.58 TiB cell-level matrix collection. It audits and prioritizes source-reported results on a normal computer while keeping every claim inside that evidence boundary.
2. Scientific Decision
The intended user is a computational biologist or experimental T-cell scientist deciding which perturbations deserve closer review before expensive independent experiments. Every included or excluded row needs an inspectable reason, and every ranking needs decomposable evidence.
The supported question is: which perturbations pass effective knockdown, coverage, off-target-flag, cross-donor, and cross-guide source-QC checks within each condition, and which pass those checks in all three conditions?
3. Registered Analysis Pipeline
- Audit source schemas, hashes, row counts, and the four-donor by three-condition study design
- Preserve a gate-by-gate decision and exclusion reason for every source row
- Require six registered criteria covering on-target significance and direction, off-target flags, minimum cell coverage, donor concordance, and guide concordance
- Rank eligible rows within condition using the geometric mean of downstream-discovery, donor-concordance, and guide-concordance percentile ranks
- Rank cross-condition candidates by their weakest condition score so one strong condition cannot hide a weaker one
- Generate traceable tables, figures, reports, methods, run summaries, and provenance manifests
4. Analyst Workbench
- Registered-run overview and transparent filter funnel
- Searchable, sortable candidate explorer
- Gene dossiers with React Flow and ELK evidence graphs
- Side-by-side gene and condition comparison
- Exploratory score-component reweighting in a separate robustness lab
- Constrained finding-card generation that preserves the inference boundary
- Provenance ledger for sources, contracts, package versions, and generated artifacts
5. Reproducibility and Product Quality
- Deterministic Python CLI with pinned configuration and dependency lockfile
- Automated tests for statistics, official-data analysis, configuration, and end-to-end workflows
- Manifest fingerprints and SHA-256 verification before the web app reads generated tables
- Privacy-scrubbed client and server error events
- Responsive Next.js interface with Vercel Analytics and Speed Insights
6. Pinned-Run Results
- 33,983 perturbation-condition rows audited
- 497 eligible Rest rows, 593 eligible 8-hour rows, and 562 eligible 48-hour rows
- 1,652 eligible rows across all three conditions
- 224 genes passed every registered gate in all three conditions
- SETDB1 led the cross-condition ranking with a weakest-condition score of approximately 0.915
7. Claim Boundary
These results identify source-QC candidates for independent follow-up. The score is a transparent prioritization heuristic, not a probability, effect size, therapeutic score, or validation claim. Cross-condition comparisons are descriptive because the workflow does not fit a perturbation-by-condition interaction model.