Disease-Specific T-Cell Autoimmune Target Discovery and Cross-Disease Validation
CMap-style reversal signature matching against genome-scale CD4+ T cell Perturb-seq knockdown data, across 7 autoimmune diseases
Dinesh Ravindra Raju, Ze Yu
Built with Claude: Life Sciences hackathon (Research track), July 2026.
Framing note. Every gene named in this report is a computationally nominated candidate for experimental validation in disease-specific T-cell models — not a confirmed drug target. Each survived a genome-wide statistical screen plus a set of plausibility filters, but none has been validated in an independent dataset, cell system, or animal model within this analysis. See Section 6 (Validation & Caveats) before acting on any candidate.
Executive Summary
This project asks a single computational question: which CD4+ T-cell gene knockdowns push the transcriptome in the direction of successful disease treatment, across multiple autoimmune diseases at once? It answers that question in two phases.
Phase 1 — Discovery (3 diseases). Rheumatoid arthritis (RA), psoriasis, and multiple sclerosis (MS) were each represented by a public disease-state signature and a treatment-reversal signature (before vs. after effective therapy). All 11,281 genome-scale CRISPRi knockdowns from the Marson/Pritchard Perturb-seq CD4+ T-cell atlas were scored, Connectivity-Map style, against each disease's reversal signature. A cross-disease consensus ranking, a six-filter plausibility cascade, and Open Targets druggability annotation narrowed this to a 40-gene priority shortlist.
Phase 2 — Cross-disease validation (4 more diseases). The same 40 candidates were re-scored against four additional autoimmune diseases with T-cell-driven pathology — type 1 diabetes (T1D), inflammatory bowel disease (IBD), lupus (SLE), and celiac disease — expanding the screen to 7 diseases total. This surfaced two "universal regulators" (TET2, KAT6B — each significant in 4 of 7 diseases) and, on closer inspection, one standout finding: KAT6B has no prior T-cell literature of its own, but its paralog KAT6A has independently published causal evidence in MS and IBD mouse models, and a dual KAT6A/B inhibitor already has Phase 1 human safety data from oncology trials — a purely computational, literature-blind screen converging on a pathway independent wet-lab genetics had already validated.
Bottom line for a reader in a hurry: 40 candidates, fully ranked and annotated, are in MASTER_ranked_candidate_table.csv and crossvalidation_hit_matrix.csv. TET2 meets the "universal" statistical bar but is not recommended due to a known safety liability (clonal hematopoiesis risk). KAT6B is the most defensible repurposing lead this analysis surfaced. Every number here needs experimental follow-up before it means anything clinically.
Figure 0. End-to-end pipeline architecture: disease signature curation through the discovery phase (top) and the cross-disease validation extension (bottom).
1. Disease Signatures Curated
1.1 Original 3 diseases (Phase 1)
Three autoimmune diseases with strong T-cell pathology and available public transcriptomic data were selected: RA, psoriasis, and MS. For each, a disease-state signature (active disease vs. healthy) and a treatment-reversal signature (before vs. after effective treatment) were curated from GEO.
| Disease | Signature type | GEO accession | N | Platform | Tissue/cell type |
|---|---|---|---|---|---|
| RA | Disease-state | GSE56649 | 13 active RA vs. 9 healthy | Affymetrix HG-U133 Plus 2.0 (GPL570) | Purified peripheral blood CD4+ T cells |
| RA | Treatment-reversal | GSE97165 | 19 paired pre/post triple-DMARD | RNA-seq (GPL11154) | Synovial biopsy (not blood/CD4+ T) |
| Psoriasis | Disease-state | GSE30999 | 89 subjects total (170 samples: 85 lesional + 85 non-lesional); 81 subjects had both a lesional AND non-lesional sample (8 had only one biopsy type) | Affymetrix HG-U133 Plus 2.0 (GPL570) | Skin biopsy |
| Psoriasis | Treatment-reversal | GSE137218 | 14 paired day-0/day-84 secukinumab (anti-IL-17A) | Affymetrix Clariom D (GPL23126) | Lesional skin biopsy |
| MS | Disease-state | GSE21942 | 14 MS vs. 15 healthy | Affymetrix HG-U133 Plus 2.0 (GPL570) | PBMC |
| MS | Treatment-reversal | GSE138064 | 12 complete-responder patients, 0h vs. 24h post IFN-β injection | Affymetrix HTA 2.0 (GPL17586) | PBMC |
Cross-platform considerations: All disease-state signatures share the GPL570 microarray platform, giving consistent probe-to-gene mapping. GSE97165's series-matrix file carries metadata only — its expression data was pulled from a separate supplementary counts file, GSE97165_counts.txt.gz. The two newer-generation reversal-signature platforms (Clariom D, HTA 2.0) required Bioconductor annotation packages (clariomdhumantranscriptcluster.db, hta20transcriptcluster.db) for probe-to-gene-symbol mapping.
The RA reversal signature is the weakest link in the original panel: no genuinely paired pre/post anti-TNF blood or CD4+ T-cell dataset could be located in GEO after an extensive search (GSE58795, GSE12051, GSE20690 were all confirmed cross-sectional responder/non-responder designs, not paired timepoints). GSE97165, a paired pre/post-DMARD synovial-biopsy RNA-seq dataset, was used as a documented compromise — it captures joint-tissue treatment response rather than circulating or CD4+ T-cell-intrinsic response.
1.2 Cross-validation: 4 new diseases (Phase 2)
| Disease | Disease-state GEO | N (disease vs. control) | Tissue | Quality | Reversal GEO | Reversal design | N pairs | Quality |
|---|---|---|---|---|---|---|---|---|
| Type 1 Diabetes | GSE55098 | 12 T1D vs 10 healthy | PBMC | Moderate | GSE84934 | Teplizumab responder vs non-responder, single timepoint — not paired pre/post | 16 vs 16 | Weak |
| Inflammatory Bowel Disease | GSE16879 | 43 disease (CD+UC) vs 6 healthy | Colonic mucosal biopsy | Strong | GSE16879 (same series) | Paired before/after first infliximab, responders only | 27 matched pairs | Strong |
| Lupus (SLE) | GSE49454 | 157 SLE vs 20 healthy | PBMC/whole blood | Strong | GSE49454 (same series) | Longitudinal within-patient active (max SLEDAI) vs quiescent (min SLEDAI) visit | 35 patients | Moderate |
| Celiac Disease | GSE145358 | 30 celiac vs 6 healthy | Duodenal biopsy | Moderate | GSE145358 (same series) | Paired gluten-free-diet vs post-gluten-challenge biopsies | 15 matched pairs | Moderate |
IBD (GSE16879) is the strongest new addition — the same series provides both a large disease-state contrast and a genuinely paired responder pre/post-infliximab reversal design, matching the rigor of the best datasets in the original panel. T1D has no genuinely paired pre/post immunotherapy dataset in GEO; the responder-vs-non-responder comparison is a documented compromise, following the same pattern as the original RA-reversal gap. Lupus's disease-state signature is dominated by a strong interferon-stimulated-gene signature (IFI35, CMPK2, IFIT5, USP18, RSAD2), consistent with known SLE biology, but its longitudinal reversal signature is statistically weak (reflecting natural disease-activity fluctuation, not a specific drug). Celiac is a small RNA-seq dataset but its paired diet-remission vs. gluten-challenge-flare design is mechanistically clean and tissue-matched.
Reversal gene sets used for scoring were the top-150 most significant up- and down-regulated genes (by p-value) in each treatment-reversal signature, restricted to genes also measured in the Perturb-seq transcriptome panel.
2. Methods
2.1 Perturb-seq knockdown data
Source: Marson/Pritchard genome-scale CD4+ T-cell Perturb-seq dataset (s3://genome-scale-tcell-perturb-seq/marson2025_data/), specifically GWCD4i.DE_stats.h5ad (16.8 GB) — genome-wide pseudobulk differential expression for CRISPRi knockdowns across the transcriptome. All scoring used the Stim48hr culture condition (11,281 of ~12,748 total gene knockdowns; the remainder excluded per-condition due to insufficient cell counts or QC filters upstream in the source data), against a background of 10,282 reliably detected measured genes. Each knockdown carries per-gene log2FC, raw p-value, and adjusted (10% FDR) p-value relative to non-targeting controls. Two separate extractions were made from the raw file: a compact checkpoint (logfc/padj/pval matrices, 11,281 × 10,282, used for reversal scoring and the on-target knockdown-efficacy check) and a smaller, separate pull of the baseMean layer (median baseline expression per gene in CD4+ T cells) used only for the Filter 1 expression-plausibility check in Section 2.4 — these are not the same file.
Why Stim48hr and not Rest or Stim8hr. The dataset provides three culture conditions per knockdown; Stim48hr was chosen deliberately rather than by default. Biologically, it best represents an activated CD4+ T-cell state relevant to autoimmune pathology — autoreactive T cells drive disease while actively signaling, not at rest, and 48 hours captures the sustained-activation transcriptome after the acute early-response programs seen at 8 hours have settled. The two alternatives were considered and rejected on the same grounds: Rest would test target relevance in quiescent cells, which is less relevant to an actively-signaling autoimmune process; Stim8hr captures early-activation kinetics that may not reflect the more stable regulatory states a knockdown's downstream disease-relevant effect depends on. Restricting to one condition also kept the statistical burden manageable — scoring all three conditions would have tripled the number of tests (11,281 knockdowns × 3 conditions × 7 diseases) without a clear way to combine three per-condition reversal scores into one ranking. This is a documented design choice, not an oversight: whether the TET2/KAT6B/TCR-complex signal is activation-state-dependent remains untested (see Section 7, Next Step 4).
2.2 Disease-signature differential expression
Disease-state signatures (active disease vs. healthy): Welch's t-test per gene (unequal variance), Benjamini-Hochberg FDR correction. Treatment-reversal signatures (before vs. after effective treatment): paired t-test per gene where matched pre/post samples exist (RA, Psoriasis, MS, IBD, Celiac); for T1D, a responder-vs-non-responder Welch's t-test proxy was used in place of a genuine paired design (no such dataset exists in GEO); for Lupus, a within-patient active-vs-quiescent longitudinal proxy built from repeated clinical visits. Microarray probes were collapsed to gene symbols via platform-specific annotation (GPL570 .annot.gz; GPL10558 .annot.gz; Clariom D and HTA 2.0 via Bioconductor *transcriptcluster.db packages). RNA-seq counts (GSE97165 for RA reversal; GSE145358 for celiac disease-state/reversal) used pre-normalized counts with gene symbols already provided (or Ensembl IDs mapped to symbols for GSE97165), log2-transformed with a floor of 1 count. GSE16879 (IBD) and GSE49454 (Lupus) are GPL570/GPL10558 microarray, not RNA-seq, and were processed identically to the other microarray signatures above.
2.3 CMap-style reversal scoring
For each of the 11,281 knockdowns (Phase 1) or 40 shortlisted candidates (Phase 2) and each disease, a weighted Kolmogorov-Smirnov-style enrichment score (Subramanian et al. GSEA formulation) tested whether genes up in the knockdown were enriched among genes down in the disease's treatment-reversal signature (and vice versa). The final reversal score is (ES_up − ES_down) / 2; a positive score means the knockdown's expression signature runs opposite the disease-reversal direction (i.e., the knockdown mimics treatment-induced reversal).
A pooled empirical null (100 permutations of matched-size random gene labels × a 2,000-knockdown subsample = 200,000 null scores) was used to compute two-sided empirical p-values. In Phase 1, Benjamini-Hochberg FDR correction was applied genome-wide (11,281 tests) per disease. In Phase 2, because the same 150-up/150-down gene-set size was used, the identical pooled null was reused directly, with FDR correction applied across only the 40 genes tested per disease (a much smaller multiple-testing burden, appropriate since these genes were pre-selected, not screened fresh).
2.4 Cross-disease consensus and filter cascade (Phase 1 only)
Genes significant (FDR<0.1) in the 3-disease screen were pooled and ranked by how many diseases each passed in, then run through six sequential filters:
| Filter | Criterion | Result |
|---|---|---|
| 1. Expression plausibility | Median Perturb-seq baseMean above 10th percentile across all measured genes | 2,009 / 2,663 pass |
| 2. Effect-size consistency | |log2FC| ≥ 0.3 on shared reversal-signature genes | Flagged (not excluded): 105/2,009 robust |
| 3. Cross-disease consensus | Significant (FDR<0.1) in ≥2 diseases | 177 candidates carried forward |
| 4. Network/pathway support | Direct STRING interaction (score≥400) with ~160 curated autoimmune/GWAS genes | 34/177 direct hits |
| 5. Novelty | Not itself a known GWAS gene; knockdown profile not highly correlated (r>0.5) with one | 171/177 novel |
| 6. Threshold sensitivity | Sign of reversal score stable across top-100/150/200 gene-set sizes | 177/177 stable |
The 40 highest-priority candidates (weighted by effect-size robustness, network support, novelty, and consensus FDR) were then queried against the Open Targets Platform for protein class, tractability, and known/investigational drugs, yielding DIRECTLY TARGETABLE / INDIRECTLY TARGETABLE / CHALLENGING tiers.
2.5 Cross-disease validation re-scoring (Phase 2)
The same 40 candidates were re-scored against each of the 4 new diseases using the identical weighted-KS engine and reused permutation null (Section 2.3). Results were merged into a single 40-gene × 7-disease hit matrix (score, FDR, significance flag per cell), then genes were classified by breadth:
- UNIVERSAL REGULATOR: significant (FDR<0.1) in 4+ of 7 diseases
- MULTI-DISEASE REGULATOR: significant in 2-3 diseases
- DISEASE-SPECIFIC: significant in only 1 disease
A STRING protein-interaction analysis (score≥400) among the 40 candidates themselves identified multi-gene clusters; a cluster was flagged pathway-level signal only if a gene had 2+ other cluster members each significant in different diseases from one another (the brief's explicit criterion) — a bar that only a 6-member cluster among the 40 could satisfy, since 2-member clusters have just one possible "other" gene and cannot mathematically qualify. Finally, the two UNIVERSAL REGULATORS were assessed for single-molecule multi-disease targeting feasibility: mechanism plausibility per disease, existing single-drug multi-indication precedent, and a safety/toxicity cross-check using CD4+ T-cell expression breadth as a (partial) proxy.
3. Results — Phase 1: 3-Disease Discovery
3.1 Reversal scoring
| Disease | Genes tested | FDR < 0.1 |
|---|---|---|
| RA | 11,281 | 1,713 |
| Psoriasis | 11,281 | 7 |
| MS | 11,281 | 1,181 |
The wide disparity between diseases (RA/MS yielding hundreds to >1,000 hits, psoriasis only 7) most likely reflects the smaller, noisier reversal signature available for psoriasis (n=14 paired biopsies from a single time-course) relative to RA/MS.
Figure 1. Genome-wide reversal scores per disease (11,281 CD4+ T-cell knockdowns, Stim48hr). Red points mark the 40 priority candidates that survived the full filter cascade.
3.2 Cross-disease consensus
Genes were pooled across the three disease-specific FDR<0.1 lists (2,663 union). 238 genes reached FDR<0.1 in ≥2 diseases; none reached significance in all 3 (driven by psoriasis's sparse hit list).
Figure 2. Reversal score by disease for the top-30 priority candidates.
Figure 3. Sequential reduction from the genome-wide screen (11,281 knockdowns) to the final 40-candidate priority shortlist.
3.3 Druggability
| Tier | N | Examples |
|---|---|---|
| DIRECTLY TARGETABLE | 7 | ITK, KAT6A, INPP5D, PRKAG1, MAPK14, CD52, CD33 |
| INDIRECTLY TARGETABLE | 8 | SETD7, USP18, TET2, PPP3R1, PTPN18, KAT6B, HELZ, PRDX2 |
| CHALLENGING | 25 | VAV1, TBX21, HIF1A, and 22 largely uncharacterized/unclassified proteins |
3.4 Summary statistics
- Genes screened genome-wide: 11,281
- Passing FDR<0.1 in ≥1 disease: 2,663
- Passing FDR<0.1 in ≥2 diseases (consensus): 238
- Passing FDR<0.1 in all 3 diseases: 0
- Final annotated priority shortlist: 40
- Novel vs. known-gene-correlated: 36 novel / 4 correlated with an established GWAS gene
3.5 Top 10 candidates (Phase 1 ranking)
- ITK (RA, score=−0.48, FDR=0.002) — TEC-family kinase central to TCR signaling; dense STRING neighborhood with CD2/CD247/CD28/LAT/LCK/VAV1. Directly targetable (ritlecitinib, pazopanib precedent).
- KAT6A (MS, score=0.40, FDR=0.030) — MYST-family lysine acetyltransferase with a documented high-quality small-molecule binding pocket; network-linked to RUNX1.
- INPP5D (SHIP1) (RA, score=−0.47, FDR=0.002) — Inositol phosphatase dampening PI3K/AKT signaling; network-linked to GAB2/PTPN11/STAT3.
- PRKAG1 (MS, score=−0.46, FDR=0.005) — AMPK regulatory gamma-1 subunit, network-linked to MAP3K7 (TAK1).
- MAPK14 (p38α) (Psoriasis, score=0.46, FDR=0.040) — Extensively drugged kinase (28 known compounds), STRING-linked to TAB1.
- CD52 (RA, score=−0.39, FDR=0.031) — Surface antigen with an approved therapeutic antibody (alemtuzumab, used clinically in MS).
- CD33 (MS, score=0.43, FDR=0.015) — Myeloid/lymphoid inhibitory receptor with oncology antibody-drug conjugates (gemtuzumab ozogamicin).
- SETD7 (RA, score=0.39, FDR=0.033) — SET-domain lysine methyltransferase, network-linked to TBX21 (T-bet).
- USP18 (RA, score=0.37, FDR=0.057) — Deubiquitinase / ISG15-specific protease regulating type-I interferon signaling.
- TET2 (RA, score=0.50, FDR=0.002) — DNA demethylase with an established role in T-cell differentiation; network-linked to JAK2/RUNX1/SH2B3.
Full master table (all 40 candidates, 18 annotated columns): MASTER_ranked_candidate_table.csv.
4. Results — Phase 2: Cross-Disease Validation (7 Diseases Total)
4.1 Re-scoring the top 40 against 4 new diseases
| Disease | N significant (of 40) | Top hit |
|---|---|---|
| IBD | 13 | TET2 (FDR=0.0006) |
| Celiac | 6 | TET2, MAPK14 (FDR=0.016 each) |
| T1D | 4 | EMC3, PRDX2, PTPN18 (FDR=0.037 each) |
| Lupus | 0 | GNAI2 (FDR=0.47, not significant) |
IBD gave by far the strongest cross-validation signal, consistent with its high-quality paired-reversal data. Lupus produced no significant hits at the 40-gene scale — expected given its reversal signature's inherent weakness.
4.2 Full cross-disease hit matrix
Figure 4. Reversal score (color) and significance (asterisk, FDR<0.1) for all 40 candidates across all 7 diseases. Full data: crossvalidation_hit_matrix.csv.
4.3 Breadth classification
Figure 5. (a) Breadth classification counts. (b) Distribution of the number of diseases each gene reached significance in.
| Classification | N genes | Criterion |
|---|---|---|
| UNIVERSAL REGULATOR | 2 | Significant (FDR<0.1) in 4+ of 7 diseases |
| MULTI-DISEASE REGULATOR | 38 | Significant in 2-3 diseases |
| DISEASE-SPECIFIC | 0 | Significant in only 1 disease |
No gene fell to DISEASE-SPECIFIC — every gene in the top-40 shortlist was already an RA+MS consensus hit by construction, so all 40 tested at 2+ diseases going in. The two UNIVERSAL REGULATORS: TET2 and KAT6B (each significant in RA, MS, IBD, Celiac).
Pathway-level signal. STRING analysis among the 40 candidates found 5 multi-gene clusters; only the largest satisfies the brief's "2+ other genes, each significant in a different disease" bar: the TCR-proximal signaling complex — ITK, VAV1, CD52, CD33, ARHGAP30, TBC1D10C (6 members, each individually hitting only 2-3 diseases, but the cluster's union covers 4: RA, MS, IBD, Celiac). The two-member clusters (SETD7/TBX21, KAT6A/KAT6B, LEO1/ELOF1) cannot mathematically satisfy this criterion, since each has only one possible "other" gene — noted descriptively but not formally flagged.
4.4 Single-molecule multi-disease targeting assessment
TET2 — NOT RECOMMENDED despite meeting the universal threshold. A DNA-demethylating enzyme with an established, disease-general role in T-cell cytokine regulation and Treg/Tfh lineage stability — a plausible shared epigenetic mechanism. However, TET2 loss-of-function is strongly and independently linked to clonal hematopoiesis (CHIP) and myeloid malignancy risk in humans — a serious safety liability invisible to the CD4+-T-cell-expression-only safety proxy used in Filter 1 (which read this gene as LOW toxicity risk). Flagged as a hard safety stop.
KAT6B — the standout finding of this cross-validation. KAT6B has no dedicated prior autoimmune or T-cell literature of its own. But its functionally redundant MYST-family paralog, KAT6A, was independently shown in 2024 (Fu et al., Cell Metabolism) to drive CD4+ T-cell glycolytic reprogramming and proinflammatory differentiation — and KAT6A-deficient mice are protected from experimental autoimmune encephalomyelitis (the MS model) and colitis (the IBD model), exactly two of KAT6B's four computationally-nominated disease hits. A purely computational, literature-blind screen independently converged on the same pathway that independent wet-lab genetics had already validated, in the same disease areas. A first-in-class dual KAT6A/KAT6B inhibitor, PF-07248144 (Pfizer), already has published Phase 1 human safety/tolerability data and is in Phase 3 oncology trials for breast cancer — meaning a repurposing path would not need to start from a de novo compound. Caveat: KAT6A/B are broadly essential for hematopoietic and neural stem cell maintenance; chronic autoimmune dosing would face substantial toxicity hurdles distinct from short-course oncology use.
Comparator precedent: TNF inhibitors (etanercept, infliximab, adalimumab) and IL-17A inhibitors (secukinumab) are approved across multiple autoimmune indications, establishing that "one drug, many diseases" is achievable when the mechanism is genuinely shared — the benchmark against which TET2 and KAT6B are judged.
CD33 — an instructive near-miss, not a candidate. CD33 has an FDA-approved targeted drug (gemtuzumab ozogamicin), but it is a cytotoxic antibody-drug conjugate mechanistically incompatible with chronic immunomodulatory use — illustrating that "has an approved drug" is necessary but not sufficient for a repurposing claim.
Bottom line: KAT6B is the top single-molecule multi-disease candidate surfaced by this analysis — not risk-free, but the most defensible, evidence-convergent lead, given independent causal genetics in two of its four hit diseases plus an existing clinical-stage inhibitor with human safety data. TET2 formally meets the universal threshold but is explicitly not recommended. Neither is validated; both require full experimental follow-up.
4.5 Broadly-protective pathways/genes — explicit list
- TET2 — implicated across RA, MS, IBD, celiac (4 diseases); epigenetic/Treg-stability mechanism, but carries a CHIP safety liability.
- KAT6B / KAT6A-KAT6B pathway — implicated across RA, MS, IBD, celiac (4 diseases); independently validated causal role for paralog KAT6A in MS and IBD mouse models.
- TCR-proximal signaling complex (ITK, VAV1, CD52, CD33, ARHGAP30, TBC1D10C) — collectively spans RA, MS, IBD, celiac (4 diseases) even though individual members hit only 2-3 each.
- TBX21/SETD7 pair (descriptive note; does not meet the formal pathway-flag bar) — union spans RA, MS, celiac, IBD; TBX21 (T-bet) is the master Th1 transcription factor.
5. The "Wait, Really?" Finding
A purely computational, statistically rigorous screen — one that never consulted the T-cell literature — independently rediscovered a druggable node (the KAT6A/B histone-acetyltransferase pair) that wet-lab biology had already validated, in a different paralog (KAT6A vs. our KAT6B hit) and in different disease models (EAE/MS and colitis/IBD mouse models), with an existing Phase 1-human-dosed, Phase 3 clinical-stage inhibitor (PF-07248144) available for potential repurposing. This is the single most defensible "surprising result" this analysis surfaced: not the top-ranked score, but the point where an independent computational and an independent experimental line of evidence converged on the same target.
6. Validation & Caveats
6.1 Statistical validation performed
- Threshold sensitivity (Phase 1, Filter 6): reversal-score sign was stable across top-100/150/200 reversal gene-set sizes for 177/177 consensus candidates (100%) — the ranking is not an artifact of the arbitrary top-150 cutoff choice.
- Permutation-based empirical null: 100 permutations × a 2,000-knockdown subsample (200,000 null scores) used for empirical p-values in both phases, rather than relying on a parametric approximation.
- Genome-wide multiple-testing correction (Phase 1): Benjamini-Hochberg FDR applied across all 11,281 knockdowns per disease; shortlist-scale correction (Phase 2): BH-FDR applied across the 40 pre-selected candidates per new disease (appropriately smaller correction, since these genes were not screened fresh).
- Cross-disease reuse of the null: because gene-set size (150 up/150 down) was identical across all 7 diseases, the Phase 1 pooled null was reused directly for Phase 2 scoring rather than re-derived — internally consistent, but see the caveat below.
6.2 Caveats — cross-platform, tissue, and design limitations
- Cross-platform/cross-tissue mismatch. Disease-state signatures are CD4+ T-cell- or PBMC-derived (RA, MS, T1D, Lupus) or biopsy-derived (Psoriasis: skin; IBD: colonic mucosa; Celiac: duodenum). The RA, IBD, and Celiac reversal signatures are all tissue-derived (synovial, colonic, duodenal) rather than blood/CD4+ T cell — a substantive tissue mismatch relative to the pure CD4+ T-cell Perturb-seq screen used for scoring.
- CRISPRi partial knockdown, not knockout. All Perturb-seq perturbations are CRISPRi-mediated partial loss-of-function, not complete knockouts — effects may be under- or over-estimated relative to a true null allele, especially for dosage-sensitive genes.
- Permutation null does not model gene-gene correlation. The null is built from random gene-label shuffles and does not capture real co-expression structure; observed reversal-score variance somewhat exceeds null variance (e.g., RA reversal-score SD of 0.22 vs. null SD of 0.17), meaning reported FDR values are likely anti-conservative — a ranking signal, not a calibrated significance threshold.
- Cell-state heterogeneity. The Perturb-seq screen profiles a mixed CD4+ T-cell population at a single activation timepoint (Stim48hr, 8-48h post-activation), not FACS-sorted Th1/Th2/Th17/Treg subsets or long-term chronic-activation states.
- Zero genes reached significance in all 3 (Phase 1) or all 7 (Phase 2) diseases. In Phase 1 this reflects psoriasis's sparse reversal signal (n=14 paired biopsies); "consensus" in Phase 1 is effectively "significant in RA and/or MS." In Phase 2, the "universal regulator" bar (4+ of 7) is a substantially lower threshold than full 7-disease agreement and should be read accordingly.
- T1D and Lupus reversal signatures are compromises, not genuine drug-response designs — results for these two diseases should be weighted less heavily than IBD or the original RA/Psoriasis/MS reversal signals.
- Small sample sizes for T1D (12 vs 10) and celiac (30 vs 6, only 15 paired reversal samples) reduce statistical power relative to IBD (43 vs 6) and lupus (157 vs 20).
- Curated GWAS reference list is not exhaustive. The novelty filter (Filter 5, Phase 1) used a manually curated list of ~160 established autoimmune GWAS genes rather than a live, comprehensive GWAS Catalog query — some "novel" candidates may have unlisted GWAS associations.
- Druggability annotation reflects current Open Targets records only and does not include newer literature, patents, or proprietary pipeline programs not yet indexed there.
- The CD4+-T-cell-expression-based safety proxy has known blind spots. Both TET2 and KAT6B were flagged LOW toxicity risk by this proxy, yet both carry serious, well-documented safety liabilities (CHIP/myeloid malignancy for TET2; broad hematopoietic/developmental essentiality for KAT6A/B) that a T-cell-expression-only filter cannot detect. This is itself an important validation lesson: expression breadth in the tissue of interest is necessary but not sufficient evidence of safety.
7. Suggested Next Steps
- Independent validation of the top candidates (ITK, MAPK14, CD52, CD33, INPP5D, PRKAG1, KAT6A, and especially KAT6B) in an orthogonal CD4+ T-cell perturbation system (arrayed CRISPR knockout or siRNA), with full gene knockout rather than CRISPRi knockdown.
- Subset-resolved re-analysis — if single-cell-resolution data is available, repeat scoring within FACS/clustering-defined Th1/Th2/Th17/Treg subsets to test whether reversal effects are subset-specific.
- Obtain a genuinely paired pre/post-treatment dataset for RA (blood/CD4+ T cell, not synovial) and for T1D (teplizumab or abatacept trial with matched pre/post transcriptomics) to replace the current documented compromises.
- Extend to Rest and Stim8hr Perturb-seq culture conditions (only Stim48hr was scored here) to test whether candidate effects, especially the TET2/KAT6B/TCR-complex signal, are activation-state-dependent.
- Directly test the KAT6A/B–autoimmunity hypothesis: assess whether PF-07248144 (or a research-grade KAT6A/B inhibitor) modulates CD4+ T-cell activation/differentiation in vitro, building on the existing KAT6A EAE/colitis mouse data.
- Directly assess TET2 inhibition's effect on clonal hematopoiesis risk markers before any further autoimmune-indication development, given the safety flag raised in Section 6.
- Formal GWAS Catalog cross-reference using a live, comprehensive query to sharpen the novelty classification before prioritizing candidates for wet-lab follow-up.
- Extend the STRING-based pathway clustering with a formal enrichment test (e.g., hypergeometric or permutation-based pathway score) rather than the qualitative flagging used here.
Appendix: Full Deliverables
| File | Description |
|---|---|
MASTER_ranked_candidate_table.csv |
Phase 1: 40 candidates, 18 annotated columns |
crossvalidation_hit_matrix.csv |
Phase 2: 40 genes × 7 diseases, score/FDR/significance |
gene_classification_breadth.csv |
Phase 2: universal/multi-disease classification, pathway flags, safety notes |
single_molecule_targeting_assessment.csv |
Phase 2: TET2/KAT6B/CD33 targeting analysis |
stage3_reversal_scores_{RA,Psoriasis,MS}.csv |
Phase 1: genome-wide reversal scores per disease |
stage_crossval_reversal_scores_{T1D,IBD,Lupus,Celiac}.csv |
Phase 2: 40-gene reversal scores per new disease |
crossval_disease_data_quality.csv |
Phase 2: data-quality ratings per new disease |
pipeline_architecture_diagram.png, fig1-3_*.png, crossval_*.png |
All figures |