Next-degrader target discovery · atopic dermatitis
The druggable intracellular type-2 network
beyond STAT6
Scoring a genome-wide CD4⁺ T-cell CRISPRi Perturb-seq screen for STAT6/GATA3
collapse phenocopy, then weighting each node by chemical tractability,
stimulation-specificity, AD genetics, and AD patient skin.
3,386 perturbations scored63-node map
off IL4R–STAT6 axispatient-anchored
Companion to the selective-suppressor & effector-validation strands. Zhu & Dann et al. 2025.
The idea
STAT6 was just drugged — what else on the inside can be?
STAT6 was recently orally degraded with biologic-like efficacy in atopic dermatitis, proving an
intracellular master regulator of the type-2 program is druggable. Most Perturb-seq target work stops at
"which knockout shifts the phenotype."
The move here: fuse functional genomics with chemical tractability in one ranking —
score every perturbation for how well it phenocopies the STAT6/GATA3 collapse, then ask which of those nodes
are (i) off the crowded IL4R–STAT6 / JAK axis, (ii) a real small-molecule or degrader handle, (iii)
active only in stimulated cells, and (iv) up in AD patients.
Un-polarized Rest/Stim CD4s — this is a transcriptional-footprint resemblance to the
STAT6/GATA3 collapse, not a functional Th2 differentiation assay. Carried throughout.
Why this calibrates where the atlas did not
Anchor on a positive signal, not a noisy residual
A naive "Th2 down & Th1 flat" residual is dominated by low-power noise — FDR ≈ 1 in the sibling
projects. GATA3 and STAT6, by contrast, have a measurable trans-footprint, so scoring who else reproduces
their collapse is positively anchored. On the informative program genes the two anchors
agree at r = +0.63 (vs ~ −0.1 genome-wide). The classifier calibrates:
- Anchors self-recover — GATA3 #1 in all three conditions (0.95 Rest), STAT6 #2.
- Held-out IL4R recovers out-of-sample — 0.12 → 0.40 → 0.61 across Rest → Stim8hr → Stim48hr.
- Permutation null — top scores 0.42–0.49 vs shuffled 95th ≈ 0.04, p < 0.001.
- Known-regulator recall — median 87th percentile (vs 50th); the naive atlas was recall@250 = 0.
The map
63 nodes, four annotation layers
Top 60 non-anchor collapse phenocopies, each annotated with Open Targets tractability, GWAS Catalog
AD/atopy/allergy support, stimulation class, and AD patient-skin expression.
61 / 63 off-axis32 druggable
57 stim-specific10 AD-GWAS26 patient-up
12 nodes are simultaneously off-axis, druggable, and upregulated in AD lesional skin.
Headline candidates
ITK and INPP5D/SHIP1 lead the off-axis ranking
| gene | collapse | stim class | tractability | AD GWAS | patient logFC | conf |
|---|
| ITK | 0.39 | stim-only | approved/clinical small-molecule | ✓ | +0.14 | 0.69 |
| INPP5D | 0.35 | stim-only | approved/clinical small-molecule | ✓ | -0.22 | 0.66 |
| KDM8 | 0.37 | stim-only | ligandable pocket (SM discovery) | — | +2.30 | 0.64 |
| UBA5 | 0.37 | activation-gated | ligandable pocket (SM discovery) | — | +0.83 | 0.62 |
| PTPRC | 0.39 | activation-gated | ligandable pocket (SM discovery) | ✓ | -0.22 | 0.60 |
| DOHH | 0.33 | activation-gated | ligandable pocket (SM discovery) | — | +0.75 | 0.59 |
| NFKB2 | 0.35 | activation-gated | approved/clinical small-molecule | — | +0.17 | 0.59 |
| MYB | 0.48 | constitutive/rest | ligandable pocket (SM discovery) | ✓ | +0.35 | 0.57 |
ITK & INPP5D/SHIP1 — proximal TCR-signaling nodes with existing
chemical matter and AD/allergy genetics, entirely off the cytokine-receptor axis; both fire only in
stimulated cells. KDM8, UBA5, DOHH — under-explored enzyme handles with the strongest patient-lesional
upregulation.
Patient anchoring
Candidate expression in AD lesional skin (GSE147424)
Per-candidate lesional-vs-control differential expression in the T-cell (CD3⁺) compartment of an AD skin
scRNA-seq cohort. Many candidates are significantly up at the cell level (MWU), including the paper's own
validated hit FBXO32.
Pseudobulk (4 lesional vs 11 control) log-fold-changes are directionally positive but
mostly do not survive multiple-testing correction — small cohort, sparse T-cell sampling. Supporting
evidence of direction, not confirmation.
A coherent under-explored module
UFMylation — a hypothesis, stated honestly
Four UFMylation-pathway members (DDRGK1, UBA5, UFL1, UFM1) phenocopy the collapse and are activation-gated.
Patient anchoring is partial: DDRGK1 (+0.70) and UBA5 (+0.83) are up in AD lesional T cells, but
UFL1 (−0.01) and UFM1 (−0.08) are flat. The ligase/adaptor arm trends up while the modifier does not —
so the module does not co-move coherently in patients.
This is a lead to test, not a validated module —
and saying so is the point of a patient-anchored screen.
Caveats — carried, not buried
What this evidence is and is not
- Un-polarized cells. Transcriptional-footprint resemblance, not a functional type-2 assay.
- Activation-coupled score. Captures the anchors' whole footprint (heavily TCR-linked); correlates
only weakly with cytokine-specific suppression (Spearman −0.14 to −0.19).
- Nominal patient significance. Directionally positive, mostly not adjusted-significant in this cohort.
- Tractability is a database call. A handle existing ≠ a feasible degrader.
- Single-perturbation screen. Mechanism ordering is inferred; the confirming step is a
double-perturbation epistasis test against GATA3.
Bottom line
A druggable type-2 network beyond the Dupixent axis
Anchoring on the STAT6/GATA3 collapse turns a mis-calibrated residual into a validated classifier,
and layering tractability + genetics + AD-skin expression onto it surfaces
ITK and INPP5D/SHIP1 as the most actionable off-axis nodes, with
KDM8, UBA5, DOHH as under-explored enzyme handles up in patients.
Next: functional IL-4/5/13 readout in polarized Th2 + a double-perturbation
epistasis test against GATA3 for the top off-axis leads.
Deliverables: th2_collapse_scorer.py · th2_collapse_scores.csv ·
druggable_type2_network.csv · AD_patient_validation.csv · next_degrader_report.md · figures.
