Built with Claude · Life Sciences
About 90% of drugs fail on the wrong target. We built a pipeline that makes a cell screen prove itself against real patient genomes — and it works: it re-finds known biology (including IL23R→IBD, the right gene in the right disease) and returns honest "no" answers for the novel lupus hypotheses. Positive controls fire; the pipeline can say "no".
We take every gene from a huge CD4⁺ T-cell screen and ask one simple question for each: does human genetics agree that changing this gene would help a patient? Each gene goes through five checks — cell biology → can a drug hit it → disease genetics → large biobanks → real patient genomes from All of Us. Only genes that pass every check are trusted. The result: the method re-finds known biology (IL23R → IBD), gives honest "no" answers for the lupus ideas, and flags one gene that it then corrects itself on.
The funnel, with the actual counts: ~4,000 genes narrow to a handful, and only what clears the final patient-genome gate is trusted. The verdicts below are what came out the other end. (How each gate works is detailed further down.)
About 90% of new drugs fail in clinical trials — most often because the target gene was wrong for the disease. Targets that also have human genetic support are about twice as likely to succeed. A screen shows what a gene does inside a cell. It does not tell you whether changing that gene would help a real patient. That is the gap we close.
The screen gives thousands of candidate genes. We link each one to human disease genetics — including real patient genomes from All of Us. A gene is trusted only when the cell biology and the human genetics point the same way. That overlap is where good drug targets are.
Researchers working on autoimmune disease — lupus, IBD, rheumatoid arthritis, MS — and anyone with a gene screen asking "which of these hits is worth turning into a drug?" The answer here rests on five different kinds of evidence, not one experiment.
The method re-finds biology we already know is true (PTPN22→RA, TYK2→RA, IL23R→IBD). It also says a clear "no" to the lupus receptor ideas when the data don't support them. A method that can say "no" is one you can trust when it says "yes". Every number can be reproduced in Claude Science.
Our starting point is one of the largest single-cell experiments ever run in human immune cells — the genome-scale CD4⁺ T-cell Perturb-seq atlas from the Marson Lab at Gladstone Institutes, shared through the CZI Virtual Cells Platform (Dann et al., 2025).
CD4⁺ T cells are the "commander" cells of the immune system. When they behave badly, they drive autoimmune diseases like lupus, IBD, rheumatoid arthritis and MS. So they are exactly the right cells to look in for new autoimmune drug targets.
Perturb-seq means: switch off one gene at a time (here using CRISPRi, which quiets a gene instead of cutting it), then read the full gene-activity profile of each single cell. Do this for almost every gene in the genome, and you learn what each gene actually does to the T cell.
The same gene can matter a lot or not at all depending on whether the T cell is resting or activated. The atlas captures three states — Rest, Stim 8 h, and Stim 48 h (8 and 48 hours after activation) — so a target's effect is measured in the state where it is actually used.
We rank genes within each state separately and never average across them. A gene that only acts in activated cells is not hidden by its "quiet" resting behaviour.
The Marson dataset already links these clusters to autoimmune genetics. Our contribution is the translational layer on top — a per-gene decision framework, not a repeat of the paper's enrichment analysis. The genetics layer, not the bench score, is what re-orders the list.
A translational decision layer, quantified. 7 of our top-20 targets are absent from the paper's autoimmune-intersecting gene set; our composite correlates only ρ=0.55 with the pure-bench rank; and removing the genetics layer collapses top-10 overlap to 0.25 — genetics is the decisive re-ranker.
The scoreboard above shows what passed; here is the method behind each gate. Every layer is a different kind of evidence, drawn from a different data source — a gene earns trust only by clearing checks that ask genuinely different questions. Cell biology and human genetics are strongest when they agree.
We switch each gene off (CRISPRi) and measure how the T-cell activation program changes. Genes are ranked in each condition and must repeat.
→How easy is the gene to hit with a drug? Cell-surface receptors and kinases score higher than harder classes like transporters.
→Is the gene linked to disease in genome-wide studies? We use Open Targets and the GWAS Catalog, and flag known traps (MHC, nearest-gene).
→Does the link repeat in large national biobanks? We check FinnGen and the GWAS Catalog across 14 autoimmune diseases.
→The strongest test: real patient genomes (All of Us). We test both common and rare gene variants — something no public summary table can give you.
Live AlphaFold structures — rotate, zoom, explore. The 3D model is colored by AlphaFold per-residue confidence (blue = high, orange = low). Each target's pipeline verdict is shown by its tab colour and in the card: green = validated, grey = tested null, orange = flagged.
This was never a lupus-only search. Every target is tested against four autoimmune diseases (lupus, IBD, rheumatoid arthritis, MS) in individual-level data — and screened across 14 in public biobanks. Two known controls (PTPN22, TYK2) come out positive — proof the method works on real patient data. IL23R is confirmed in the right disease (IBD). The three novel receptor ideas come back as a clean, well-checked "no" across all four — lupus included, but not lupus alone.
We flipped the usual search to find genes with strong genetics but almost no published papers, then tested them where only patient-level data can — rare variants. The result is honest in both directions.
Instead of collapsing everything into one score, we sort candidates into four honest tiers using a therapeutic-window layer (gnomAD LOEUF), clinical stage, and causal-gene status. This separates clinically supported targets from novel-but-unresolved hypotheses — and flags the ones with strong biology but a poor therapeutic window.
Tiered target cards. IL7R is clinically advancing (Phase 2); IL21R/IL12RB2 are mechanistically convergent; KDM8/BCAT2 are novel-unresolved (pending causal-gene resolution); ZAP70/JAK2 have strong biology but a poor window (constrained LOEUF, pan-T essentiality).
Five common objections, answered with data across the six panels below: arbitrary weighting (weight-sensitivity), nearest-gene (answered twice — a stress test and formal GWAS×eQTL colocalization), single-disease (cross-disease pleiotropy), screen-replication (an orthogonal CRISPR screen), and Europe-only (cross-ancestry BioBank Japan).
The pipeline is a config-driven package — bench2biobank. Point it at any perturbation screen
(CRISPRi/a, ORF, shRNA, drug; any cell type), map your column names, and get a ranked, druggability-weighted,
disease-linked target shortlist plus a genome-scale enrichment test.
# 1. describe YOUR data in a config — just map the column names perturbation_table: path: my_screen/effects.csv columns: perturbation: gene # your header for the perturbed gene effect: logFC # signed effect on the readout stderr: se signature: readout condition: timepoint signature_of_interest: {name: activation, conditions: [8h, 48h], direction: lower} # 2. run it $ bench2biobank --config my_config.yaml [bench2biobank] 3994 perturbations → 2771 candidates → 86 ranked targets [headline] top-9% effect × disease-linked enrichment: OR=1.47, p=2.36e-03 ZAP70 composite=0.631 IL21R composite=0.593 BCAT2 composite=0.546 ...