Trans-regulatory gene mapping prioritizes disease drivers in asthma

SUMMARY

Deciphering which genes are most important to disease etiology is a central challenge in human genetics. While genome-wide association studies have cataloged thousands of variants, it’s been proposed that most are indirect regulators of a limited, currently unidentified set of central disease-driving genes, defined here as disease-proximal genes (DPGs). Here, we introduce DANDELION, a mediation-inspired statistical framework that prioritizes DPGs by integrating trans-regulatory effects from disease-relevant tissues with gene-level burden from whole-exome sequencing. Applying DANDELION to asthma uncovers novel DPGs that escape detection by conventional methods. CRISPR screens in epithelial and T cells find that most DPGs regulate key asthma-related cellular phenotypes. We also demonstrate that loss of two DPGs, SLC27A3 and SCD, affect inflammation and airway remodeling in a mouse model of allergic asthma. Our study establishes DANDELION as a powerful framework for prioritizing novel, therapeutically actionable genes and pathways underlying disease pathogenesis.

Graphical Abstract

IN BRIEF

DANDELION is a causal mediation-inspired framework, enabling the identification of disease-driving genes missed by existing approaches. Applied to asthma, DANDELION prioritizes two protein palmitoylation enzymes, revealing a therapeutically tractable pathway that is an important and previously unexplored mechanism underlying asthma pathogenesis.

INTRODUCTION

Genetic studies are an invaluable tool in therapeutic development, as drug targets supported by human genetic evidence are nearly three times more likely to achieve clinical success^1,2. However, the highly polygenic nature of complex disease makes it challenging to distinguish central disease drivers from many, sometimes hundreds, of genetic associations^3-8. The omnigenic paradigm of trait regulation proposes that complex diseases are affected by virtually every expressed gene in disease-relevant tissues³, although the overwhelming majority of genes are distal to disease risk. Distal genes are proposed to act indirectly through trans-regulation of a smaller set of central disease-driving genes that directly affects core biological pathways of disease⁴. Although core genes are likely central to disease etiology, their definition has remained abstract. Here, we define disease-proximal genes (DPGs) as those that centrally mediate the effects of other, more distal disease-associated genes within trans-regulatory networks (Figure 1A). Identifying DPGs may enable discovery of new mechanistic pathways and therapeutic targets in polygenic diseases.

Figure 1. DANDELION identifies novel DPGs in blood traits.

A) DANDELION models the path from distal genes to disease-proximal genes to a complex trait as a putative mediation path.

B) DANDELION two-step framework.

C) Enrichment of DPGs in Mendelian blood disorder genes. Paired one-sided Wilcox test.

D) For genes with the same level of burden test significance, DPGs are significantly more enriched in Mendelian disorder genes than non-DPGs.

E) Trans-regulatory network of ETV6. Arrows point from distal genes to candidate DPGs. Gene names in red are significantly associated with platelet crit in burden tests.

F) Venn diagram of the overlap between DANDELION DPGs, WES, and GWAS significant genes/loci.

G) Proportions of DPGs, distal master-regulator, and distal genes in GWAS loci of different significance.

We currently lack methods to systematically identify DPGs. Genome-wide association studies (GWAS) are not suited to DPG identification because they are inherently weighted towards discovery of common variants, which generally have small effect sizes^6,9 and, due to the pressure of selection, are less likely to impact genes involved in critical biological pathways^5,10. While whole-exome sequencing (WES) burden testing can improve identification of genes more directly relevant to a disease by leveraging rare, deleterious variants, its effectiveness is limited by small cohort sample sizes¹¹. For many traits, burden testing has only identified a handful of disease-relevant genes¹². Disease gene discovery with cis-gene expression quantitative loci (cis-eQTLs) has reached a bottleneck, largely due to systematic differences between disease genes and genes with detectable cis-eQTLs; disease genes under stronger selective constraints tend not to have strong cis-eQTL effects¹⁰. In our model, DPGs are trans-regulatory targets of disease-associated genes (such as GWAS loci) and mediate their effects on disease risk⁴. We recently demonstrated that, unlike cis-QTLs, trans-QTL target genes are under similar selective constraints as disease genes and thus useful for mapping disease-relevant genes¹³.

Here we present DANDELION, a mediation-inspired computational framework that prioritizes DPGs by integrating statistical evidence from trans-regulatory associations in disease-relevant tissues with gene-level associations from WES burden tests. DANDELION identifies genes that are not detected by existing approaches or gene prioritization methods, such as GWAS and PoPS¹⁴. To demonstrate that our method provides novel mechanistic insight into polygenic disease, we apply DANDELION to asthma and experimentally validate asthma DPGs using CRISPR screens. We then select two candidate DPGs, SLC27A3 and SCD, for in vivo characterization using a physiologic mouse model of allergic asthma. In doing so, we identify protein palmitoylation as a previously unknown biological mechanism involved in asthma pathogenesis and lung inflammation, further emphasizing DANDELION’s potential to uncover underappreciated aspects of disease biology.

RESULTS

DANDELION nominates disease-proximal genes by leveraging trans-regulatory effects

Complex traits are highly polygenic. In our model, only DPGs have a direct effect on disease risk. The majority of disease-associated genetic variants or genes (e.g. GWAS loci) instead act indirectly by regulating DPGs through trans-regulatory networks⁴. We hypothesized that by modeling the trans-regulatory effects of disease-associated single nucleotide polymorphisms (SNPs) or genes, it might be possible to identify DPGs. To this end, we developed DANDELION, a mediation-inspired computational framework that models the path from disease-associated variants or genes (X) to disease-proximal genes (M) and ultimately to the trait itself (Y; Figure 1A). By identifying the intermediate links in the path from risk variants to phenotypic outcomes, this approach prioritizes genes that have more relevant roles in human traits.

DANDELION identifies significant putative mediation pathways by combining two types of associations: (i) trans-regulatory associations (X → M) between a disease-associated SNP or gene and a putative mediator gene (M), assessed using RNA-sequencing data, and (ii) the gene-level associations (M → Y) between the putative mediator and the trait estimated from WES burden tests. The standard gene-based DANDELION pipeline uses trans associations computed by GBAT, a gene-based association test for trans-gene regulation that requires genotype data¹⁵. When genotype data are not available, SNP-based DANDELION uses the summary statistics of trans-eQTL associations. For a given disease and disease-associated variant/gene, all genes in the genome can serve as potential mediator genes. DANDELION applies Divide-Aggregate Composite-null Test (DACT)¹⁶ to evaluate evidence for putative mediation pathways by combining p-values for X → M and M → Y associations under a composite null of no mediation pathway, yielding a mediation p-value for testing the presence of a non-null mediation pathway (Figure 1B; Methods). The putative mediators on significant paths linking GWAS loci to disease risk are classified as DPGs. These significant paths also provide insight into trans-regulatory networks linking disease-associated variants to candidate genes. While DANDELION is motivated by mediation analysis, it is important to note that it prioritizes candidate DPGs through ranking of DACT p-values, rather than estimating a causal mediation effect.

We performed simulations under realistic settings (Figure S2; Methods) to compare the performance of DANDELION with ARCHIE¹⁷. ARCHIE is not formulated to identify putative mediator genes or DPGs, but it does leverage trans-regulation to identify sets of disease genes that are trans targets of disease-associated SNPs with sparse canonical correlation analysis. DANDELION significantly outperforms ARCHIE at all FDR levels (Figure S2; Supplementary Notes) and has a much lower global FDR across various levels of trans effects (Figure S2). DANDELION can also identify trans regulatory signals of DPGs, though the power to detect trans effects are lower than the power to detect DPGs (Figure S3).

DPGs are highly enriched in Mendelian disorder genes

To demonstrate the ability of DANDELION to prioritize DPGs and key disease pathways, we applied DANDELION to data from the UK Biobank on 27 blood traits with known biology that have been previously studied^18,19. Gene effects on these traits were estimated by burden tests in the UK Biobank WES studies as described in Backman et al. ¹², which aggregated rare putative loss-of-function variants and likely deleterious missense variants with a minor allele frequency (MAF) of up to 1%. We selected this set of variants and MAF cutoff because it identifies the largest number of associations in the original UK Biobank WES study¹². We used estimates of trans-regulatory effects in whole blood expression data from eQTLGen (N=31,684)²⁰, which provides trans-eQTL summary statistics for 10,000 GWAS blood trait-associated SNPs. For each trait and each SNP within 1Mb of the trait-associated locus, DANDELION identified all significant putative mediation paths and DPGs at 5% per-exposure level FDR. We annotated each SNP to the nearest gene with the most significant cis-eQTL p-value in eQTLGen. We identified 14-154 candidate DPGs per trait (Table S1). Using a conventional Bonferroni threshold (0.05/20,000), burden tests identified 7-65 genes significantly associated with each trait, the majority of which (62.9-100%) overlapped with DANDELION-identified DPGs (Figure S4). Burden test-significant genes not captured by DPGs may reflect genes not expressed in the datasets used for trans-QTL mapping or genes with insufficient trans-regulatory signal from distal loci to support detectable putative mediation signal. Importantly, this comparison is not intended to benchmark DANDELION against burden testing, but rather to illustrate that additional genes beyond those identified by burden tests alone can be prioritized by integrating trans-mediated information.

DPGs are likely members of key disease pathways; as such, deleterious mutations in blood trait DPGs are expected to lead to Mendelian blood disorders. To test this prediction, we assessed the enrichment of candidate DPGs in a curated list of 314 Mendelian genes for blood diseases¹⁹. Candidate DPGs were enriched 5.3- to 21.7-fold in Mendelian blood disorder genes compared to randomly selected genes matched by gene length (Figure 1C). Even after removing genes that were significant in burden tests, DANDELION-specific DPGs remained highly enriched in Mendelian genes (1.7- to 21.9-fold enrichment; Figure S5).

The p-values of DANDELION are partially driven by burden test p-values; therefore, many DANDELION-specific DPGs might be statistically suggestive or borderline insignificant in burden testing, akin to the well-recognized phenomenon of sub-threshold SNPs in an initial GWAS that are subsequently genome-wide significant in a later, larger cohort. However, DPGs are significantly more enriched in Mendelian disorder genes than non-DPGs that have similar levels of burden test significance (p-value=0.0011; Figure 1D), suggesting that DPGs are not simply the result of applying a more permissive threshold to burden test p-values. Rather, the mediation-inspired framework prioritizes DPGs that may mediate the effect of GWAS loci (disease-distal genes) on disease risk. As a result, DANDELION helps prioritize genes within a similar range of burden test p-values, rather than uniformly expanding the set of burden test hits.

DPGs and their trans-regulatory network reveal disease biology

Next, we used platelet crit (PLTC) as an example trait to demonstrate how DANDELION-identified DPGs and their associated trans-regulatory networks can provide insight into trait biology. PLTC quantifies the total volume of platelets in the blood and is an indicator of platelet production, destruction, and activation. The genes and pathways involved in this process are well-studied. We performed Gene Ontology (GO)²¹ enrichment analyses on the 73 candidate PLTC DPGs identified by DANDELION (Table S2), identifying the top enriched biological processes as “wound healing” (adjusted p-value=2.5x10⁻⁵) and “hematopoiesis process” (adjusted p-value=1.909x10⁻⁵).

DANDELION prioritizes DPGs with clear functions in platelet activation and blood coagulation. Platelet activation encompasses two major processes—platelet aggregation and platelet adhesion—each mediated by distinct receptor complexes. DPGs ITGA2B and ITBG3 encode the αIIb and β3 subunits of the glycoprotein IIb/IIIa (αIIbβ3) integrin complex, the central receptor that coordinates platelet-platelet aggregation via fibrinogen binding^22,23 (Figure S6A). DPG GP1BA encodes the α-subunit of the glycoprotein Ib complex, which binds von Willebrand factor to initiate platelet adhesion²⁴. All three DPGs are in the same DANDELION-generated trans-regulatory network, putatively mediating the effects of 37 PLTC GWAS loci (Figure S6B). Eighteen of these loci trans-regulate all three DPGs, indicating that platelet aggregation and adhesion are coordinated processes co-regulated by GWAS loci. While GWAS loci for PLTC have previously been reported within 1Mb of ITGA2B, ITBG3 and GP1BA, all GWAS loci are closer to other genes¹⁹. DPGs can therefore provide additional information to fine-map causal genes underlying GWAS loci.

We next focused on DANDELION-specific DPGs—those which are neither significant in burden tests nor within 1Mb of any significant GWAS loci in the UK Biobank—and found that they also have clear biological functions in PLTC. DANDELION identified ETV6 (burden test p-value=7.77x10⁻⁵) as a DPG (Figure 1E). ETV6 mutations cause thrombocytopenia, a condition that occurs when the platelet count is too low²⁵. DANDELION-specific DPG PTGIR (Figure S7; burden test p-value=3.37x10⁻⁵) encodes the prostacyclin receptor, which when bound by prostacyclin inhibits platelet aggregation^26-28. Both ETV6 and PTGIR are regulated in trans by multiple GWAS loci (Figure 1E). DANDELION’s ability to trace the trans-regulatory networks linking GWAS loci to disease-proximal genes improves the interpretability of risk variants.

Most GWAS loci are mapped to disease-distal genes rather than DPGs

DANDELION is premised on a genetic architecture model wherein most loci influencing complex traits are located far from the genes that directly govern relevant biological pathways. Thus, most GWAS loci will not map to DPGs themselves but rather to distal elements that trans-regulate DPGs⁴. To explore this hypothesis, we compared the PLTC DPGs to loci identified by WES and GWAS. WES burden tests identified 39 PLTC-associated genes, 31 of which are identified by DANDELION. GWAS of PLTC in the UK Biobank identified 742 conditionally-independent loci (Table S3). We found 48 DPGs that overlapped with significant GWAS loci, i.e., the DPG’s transcription start site occurs within 1Mb of a GWAS locus. Out of the 30 genes discovered by both WES and GWAS, DANDELION identified 23; an additional 17 DPGs are uniquely identified by DANDELION (Figure 1F).

Next, we mapped the 742 conditionally-independent GWAS loci associated with PLTC to DPG trans-regulatory networks (Table S3). We stratified GWAS loci into eight equal quantiles based on their association p-values. In the top quantile (−log₁₀(p-value)>51), 17.1% of GWAS loci are mapped to DPGs and 40.0% are mapped to disease-distal genes regulating 5 or more DPGs. We refer to these as hyper-connected disease-distal genes (HCGs, Figure 1G). Overall, the proportion of DPGs was 17.1-33.3% in each quantile and independent of GWAS significance, but HCG proportion increased with GWAS significance (correlation = 0.80, p-value = 0.017). These results suggest that the significance of GWAS p-values is more useful in nominating HCGs than DPGs. Moreover, the GWAS effect size and association p-values of DPGs and disease-distal genes are comparable, suggesting it will be difficult to prioritize DPGs based on GWAS results alone (Figure S8). We repeated these analyses for another randomly chosen blood trait and observed the same trends (Figure S9). To further demonstrate that DPGs capture signals distinct from genes prioritized based on GWAS loci and cis-effects, we compared DPGs with genes prioritized by PoPS¹⁴ and TWAS²⁹. Across all traits, the majority of DPGs are uniquely identified by DANDELION and not captured by other methods (Figure S10A). DANDELION-specific DPGs also have higher enrichment in Mendelian blood disorder genes than PoPS- or TWAS-specific genes (Figure S10B).

In summary, many DPGs have clear core functions in complex traits but are often missed by GWAS or WES burden tests. When DPGs are detected by GWAS, they evenly spread across all GWAS significance categories. The large proportions of disease-distal genes in each category, and the comparable GWAS effect sizes and p-values, make it hard to prioritize DPGs from GWAS associations alone. While follow-up studies tend to focus on GWAS loci with the highest significance, our results imply that these loci are more likely to be HCGs than DPGs.

DANDELION identifies novel asthma DPGs

Knowing that GWAS and WES miss DPGs or fail to distinguish them from other genes, we next applied DANDELION to asthma. GWAS have identified nearly 200 asthma risk loci³⁰, but many of the biological mechanisms that mediate heritable risk remain unknown. We used gene effects estimated from WES of asthma from the UK Biobank (N=61,807 European ancestry cases, N=313,658 European ancestry controls). The burden test aggregated rare predicted loss-of-function and likely deleterious missense variants with MAF<1% for each gene¹². We used both standard DANDELION, employing gene-based trans-effects estimated from the Depression Genes and Network study (DGN) RNA-seq dataset of whole blood (N=913)^15,31, and SNP-based DANDELION using SNP-based trans-effects estimated from eQTLGen²⁰. To maximize gene discovery for follow-up experimental validation, we used genome-wide genes/SNPs as the exposure genes/SNPs for DANDELION. We identified 13 disease-proximal genes at 5% per-exposure level FDR using eQTLGen. Since the DGN dataset is much smaller than eQTLGen, we adopted a less stringent threshold (10% per-exposure level FDR) and identified 16 DPGs. Eight DPGs overlapped between the two datasets, yielding a total of 21 DPGs (Figure 2A and Table S4).

Figure 2. DANDELION identifies novel asthma DPGs.

A) Trans-regulatory networks for 21 asthma DPGs. Arrows point from disease distal genes to candidate DPGs. Red circles indicate genes nearest to a locus with genome-wide GWAS significance.

B) Lung cell types with highest expression of asthma DPGs, colored based on cell lineage.

C) DPGs are expressed in multiple lung cell types across several cell lineages. Average expression is the average of the scaled expression (z-scores) among cells in each cell type that express the gene.

Most DPGs were uniquely identified by DANDELION. The burden test only identified two significantly associated genes (FLG and IL33), neither of which were nominated as DPGs by DANDELION. The burden test p-values corresponding to DPGs in our dataset vary widely, from 4x10⁻⁶ to 3x10⁻³. Three DPGs (FAM105A, NRROS, and ARHGAP27) are the nearest genes to significant GWAS loci for asthma in UK Biobank³²; the remaining are all novel asthma genes. GO analysis revealed that DPGs are significantly enriched in "cytosol to ER transport" (adjusted p-value=1.38x10⁻²) and "ER membrane" terms (adjusted p-value=8.12x10⁻⁵; Table S5). ER stress in airway epithelial and inflammatory cells can trigger the unfolded protein response, which promotes inflammation and alters cellular function in asthmatic airways³³. We further validated the DPGs using trans-effects estimated from an independent perturb-seq dataset in primary human CD4+ T cells³⁴ (Methods). Of the 21 DPGs, 16 are expressed in the CD4+ dataset, and all 16 achieve replication significance (p-value < 0.05/16). In addition, DPGs identified from whole-blood trans-eQTL data remain highly ranked in the CD4+ analysis: 15 of 16 fall within the top 0.5% of all 11,096 tested genes, and all are within the top 1% (Figure S11). These results demonstrate strong concordance between whole-blood and CD4+ T cell analyses, supporting that whole blood-derived DANDELION DPGs capture biologically relevant immune contexts.

The largest subnetwork in the DANDELION-generated trans-regulatory network involves three DPGs—FAM105A, SLC27A3, and ATP2A3—which are co-regulated by the transcription factor IKZF1 (Figure 2A). Although IKZF1 itself shows no direct association with asthma in the UK Biobank, large-scale meta-analyses incorporating data from the UK Biobank as well as other studies of diverse populations have identified significant associations with asthma risk³² and childhood asthma age of onset³⁰. Our findings suggest that IKZF1 is involved in asthma pathogenesis through its coordinated regulation of these three DPGs. An additional locus near PKN3 is associated with atopic asthma in the UK Biobank³⁵ and trans-regulatory network tracing revealed that this locus likely affects asthma risk through trans-regulation of two DPGs: SCD and MMP8. To ensure that our DANDELION results are not influenced by artificial trans hubs, we repeated the analysis after removing high-degree regulators and targets from the trans-regulatory effects and then re-identified asthma DPGs (Methods). All previously-identified asthma DPGs remained significant except TAF6L, supporting that DANDELION discoveries are robust to the removal of trans hubs and are therefore unlikely to be primarily driven by degree-related trans regulation artifacts.

Asthma DPGs are active in multiple cell types

DPGs are expected to play direct, mechanistic roles in disease biology; accordingly, we reasoned that DPG expression patterns can assist in the identification of disease-relevant cell types. We used single cell RNA-seq data from the Human Lung Cell Atlas³⁶ to map cell types with high expression of our 21 asthma DPGs. All cell types with high DPG expression (Table S6) were identified using AUCell³⁷ (Figure S12; Methods). Several immune cell types such as macrophages, T cells, and natural killer cells are among the top 15 cell types with highest DPG expression. This agrees with the cell types nominated with asthma GWAS loci³⁸. Interestingly, the cell type with the largest proportion of DPG-expressing cells is alveolar type 2 epithelial cells (Figure 2B), which are known to play a central role in asthma etiology but are often missed by using GWAS loci alone^38,39. Other top cell types include secretory epithelial cells, capillary endothelial cells, and basal epithelial cells (Figure 2B). We reason that DPGs may act in multiple cellular contexts. Consistent with this, we found that several DPGs are active across multiple cell types (Figure 2C; Figure S13). For example, SCD shows high expression in both alveolar type 2 epithelial cells and monocytes, while SLC27A3 is highly expressed in alveolar type 2 epithelial cells, macrophages, and endothelial cells. These observations indicate that DPGs are highly expressed across diverse cell types and therefore may have pleiotropic effects across cell types.

Perturbation of asthma DPGs affects asthma-relevant epithelial cell phenotypes

To functionally characterize asthma DPGs, we developed a set of experimental platforms evaluating the effect of DPG perturbation on molecular and cellular phenotypes in epithelial and T cells, which represent disparate cell types that express most asthma DPGs and share a high fraction of asthma heritability⁴⁰. As we observed the highest expression of DPGs in the epithelium, we first performed an arrayed CRISPR screen in the human bronchial epithelial cell line 16HBE14o- (16HBE; Figure 3A), targeting the 14 DPGs and 29 disease-distal genes expressed in these cells. We screened for genes affecting epithelial barrier integrity, as the epithelium of individuals with asthma is characterized by increased permeability and a reduced ability to heal after injury^41,42. CRISPR/Cas9 knockdown of 71.4% of DPGs (10/14) in 16HBEs led to significant changes in transepithelial electrical resistance (TEER), indicating that the perturbed DPGs are involved in the initial formation (Figure 3B) or stability of the mature epithelial barrier (Figure 3C). In contrast, when we knocked down distal genes, only 14% (4/29) induced changes in TEER (Figure S14A-B). As a negative control, we also tested a set of randomly selected genes expressed at similar levels to those nominated by DANDELION (Figure S14C). We found that only 11.1% (2/18) affected epithelial barrier integrity (Figure S14D-E). To evaluate how DANDELION DPGs performed compared to genes nominated by fine-mapping of asthma GWAS loci, we tested a set of “high-confidence” genes from a recent asthma fine-mapping analysis³⁹. We limited testing to GWAS candidate genes that were either (a) the nearest gene to a fine-mapped causal variant overlapping an open chromatin region in lung or blood cells^43,44, or (b) linked to such a variant by promoter capture Hi-C in epithelial⁴⁵ or T cells⁴⁶. From this subset, we selected a group of thirty genes that were highly expressed in 16HBEs for knockdown studies (Figure S14F-G). Knockdown of four genes (ANP32E, BANF1, CFL1, and CDC6) was lethal. Of the remaining genes, 50% (13/26) significantly affected TEER. In total, 56.7% (17/30) GWAS genes affected epithelial barrier integrity or cell survival. Overall, DANDELION performed better than GWAS fine-mapping as a predictor of direct effects on epithelial barrier integrity (Figure 3D). Perturbation of many fewer disease-distal genes led to changes in epithelial function compared to either DPGs or genes nominated by GWAS functional fine-mapping.

Figure 3. CRISPR screens identify disease-proximal genes that significantly affect asthma cellular phenotypes.

A) Overview of CRISPR screen in epithelial cells.

B) TEER at 500Hz of 16HBE cells treated with CRISPR guides targeting DPGs or non-targeting (NT) control. Grey bar = time of initial cell-cell contact; measurements in right bar chart.

C) TEER normalized to 0 hr measured at 4000Hz. Grey box = mature barrier; measurements in right bar chart.

D) Percent of genes in each nomination category that had a significant effect on TEER.

E) Overview of CRISPR screen in primary Th2 cells.

F) sgRNAs significantly depleted (left) or enriched (right) in cells with highest 15% IL-13 expression compared to all Th2 cells.

G) Percent of genes in each nomination category that were significant hits for IL-13 expression.

B, C: *p<0.05, One-way ANOVA with Dunnett’s multiple comparisons test. Data are represented as mean +/− SD.

Perturbation of asthma DPGs affects asthma-relevant T cell cytokine production

The gene expression patterns of DPGs in lung cells suggest that some DPGs may have pleiotropic effects across cell types. We were particularly interested in testing DPG function in T cells, as allergic asthma is characterized by the presence of antigen-specific CD4+ T helper 2 (Th2) lymphocytes, which express cytokines including IL-13 that lead to defects in airway smooth muscle and goblet cell mucus production⁴⁷. To this end, we designed a pooled CRISPR screen to functionally validate DPGs involved in IL-13 production in Th2 cells. We used a library of single guide RNAs (sgRNAs) targeting 132 genes including disease-proximal, disease-distal, and GWAS candidate genes in a pool of Th2-polarized primary human T cells, then used fluorescence-activated cell sorting to isolate cells with the highest expression (top 15%) of IL-13 (Figure 3E). We then determined which sgRNAs were significantly enriched or depleted in these cells. We included as positive controls sgRNAs targeting IL13 and the Th2 lineage master regulator GATA3; as negative controls we included the two nearest neighbors flanking each DPG. We tested three biological donors and found strong correlation in sgRNA read counts across donors (Figure S14H-I). As expected, positive control guides targeting IL13 and GATA3 were significantly depleted in IL-13-high cells compared to the entire Th2 population (Figure 3F). The most depleted sgRNAs were those targeting DPGs SLC27A3 and TAP1, while the most enriched sgRNAs targeted DPG SMPD4 and disease-distal gene MIER1. Overall, 42.1% (8/19) of DPGs were significantly associated with IL-13 expression in the pooled Th2 cell screen, compared to 23.3% (7/30) of high-confidence fine-mapped GWAS genes, 27.9% (12/43) of disease-distal genes, and 23.6% (9/38) of nearest neighbor controls (Figure 3G, Table S7).

Loss of SLC27A3 leads to changes in lipid transport and protein palmitoylation

Loss of DPG SLC27A3 had the strongest effect on both initial barrier integrity in epithelial cells and on IL-13 expression in Th2 cells. In both cell types, knockdown of SLC27A3 led to asthma-protective phenotypes. To demonstrate that DPGs can illuminate novel disease-relevant biological pathways, we further investigated the role of SLC27A3 in asthma pathogenesis. SLC27A3 is a member of solute carrier family 27, a group of long-chain fatty acid transport proteins⁴⁸. Long-chain fatty acids such as arachidonic acid are important mediators of inflammatory signaling, serving as the dominant substrate for synthesis of eicosanoids such as leukotrienes and prostaglandins. Leukotriene production is increased in the lungs of people with asthma and leukotriene receptor antagonists are prescribed to patients with asthma to reduce inflammation and constriction of the airways⁴⁹. Accordingly, we investigated the effect of SLC27A3 loss on lipid transport by performing lipidomics on 16HBEs and their culture media. After 24 hours, SLC27A3−/− cell culture media contained significantly more long-chain fatty acids such as palmitic (C16:0) and arachidonic acid (C20:4) than the media of WT control cells (Figure 4A, Table S9). Intracellular arachidonic acid storage species were reduced (Figure 4B, Table S9). RNA-seq comparing SLC27A3 knockout cells to those treated with a non-targeting guide identified very few differentially expressed genes (n=53 in 16HBEs, n=9 in Th2 cells; Table S8), suggesting that loss of SLC27A3 is likely mediated not through transcriptional but rather post-translational effects. Overall, our data suggest that SLC27A3−/− cells have reduced uptake and storage of arachidonic acid, potentially reducing the production of lipid mediators of inflammation such as leukotrienes.

Figure 4. Loss of SLC27A3 improves lung function through multiple lipid-related mechanisms.

A) Free fatty acids remaining in cell culture media after 24 hrs.

B) Relative levels of arachidonic acid (C20:4)-containing phospholipid (PL) species normalized to all PL species in 16HBE cells.

C) Relative abundance of palmitate (C16:0)-containing acylcarnitine in 16HBE cells.

D) Overview of protein palmitoylation.

E) Global labeling of proteins in 16HBE cells treated with C16:0 alkyne probe for 6 hrs. Black arrows = protein regions with reduced labeling in cells treated with sgRNA targeting SLC27A3.

F) Western blot and quantification of protein lysate from 16HBE cells that underwent acyl-biotin exchange to label palmitoylated proteins. Streptavidin pull-down and input fractions blotted for STAT3.

G) Western blot and quantification for STAT3 and p-STAT3 in 16HBE cells.

H) Overview of allergic sensitization model.

I) Bronchoalveolar lavage cellularity per mL of fluid from mice treated with PBS (N=2 per genotype) or Alt (N=6 WT, 8 KO).

J) Representative H&E, PAS, and MT images of airways. Black bar = 100 μm.

K-M) Quantification of epithelial thickness from H&E (K), mucin-positive epithelium from PAS (L), and collagen deposition from MT (M).

N) Bronchoalveolar lavage cellular per mL of fluid from mice treated with PBS (N=2 per genotype) or Alt (N=3 per genotype).

O) Representative H&E images of airways. Periairway and perivascular inflammation quantified as number of immune cells per μm basement membrane. Black bar = 50 μm.

P) SLC27A3 gene expression in bronchial brushings. * p<0.05 from edgeR, Benjamini-Hochberg-adjusted.

A-O, *p<0.05, Welch t-test. Data are represented as mean +/− SD.

SLC27A3 has a dual function, serving as both a fatty acid transport protein and an acyl-CoA synthetase⁴⁸ catalyzing the synthesis of acyl-CoA, the substrate for protein palmitoylation (Figure 4D). Palmitoylation is a post-translational modification that regulates protein trafficking and signaling. While palmitoylation has been shown in other disease models⁵⁰ to regulate signaling pathways involved in asthma, such as JAK/STAT⁵¹ and T cell receptor activation⁵², the role of palmitoylation in asthma pathogenesis has yet to be investigated. Intriguingly, a second enzyme in this pathway was also nominated as a DPG by DANDELION. SCD encodes stearoyl CoA desaturase, which can reduce the availability of palmitoyl-CoA as a substrate for palmitoylation by catalyzing its conversion to monounsaturated palmitoleic-CoA. Therefore, SLC27A3 and SCD have opposite directions of effects on the availability of palmitoyl-CoA and are expected to have opposite effects on asthma. Indeed, while SLC27A3 knockdown in CRISPR screens demonstrated protective effects in both epithelial and Th2 cells, we observed opposite and pathogenic effects in cells lacking SCD (Figure 3). Our screen data suggest that loss of SLC27A3 might protect against development of asthma, while loss of SCD might induce or exacerbate asthma.

To assess whether protein palmitoylation is altered in response to loss of SLC27A3, we first investigated levels of palmitate, the substrate for palmitoyl-CoA synthesis. SLC27A3−/− cells had significantly reduced palmitate uptake (Figure 4A) and a reduction in C16:0 acylcarnitine, a source of stored palmitoyl-CoA (Fig 4C) compared to WT cells, suggesting that SLC27A3 loss may impair palmitoyl-CoA levels. We next knocked down SCD or SLC27A3 in 16HBEs and measured global levels of palmitoylation by C16:0-alkyne probe labeling. Cells lacking SCD had a minimal change in levels of total palmitoylation, but loss of SLC27A3 led to decreased global palmitoylation (Figure 4E). We hypothesized that one potential mechanism by which loss of SLC27A3 may protect against asthma-related phenotypes is by altering palmitoylation of proteins involved in asthma pathogenesis. One such protein is STAT3; palmitoylation promotes STAT3 phosphorylation, which is essential for STAT3 activation⁵³, and loss of epithelial Stat3 is protective against airway hyperresponsiveness in mouse models of asthma^54,55. Both STAT3 palmitoylation (Figure 4F) and STAT3 activation by phosphorylation (Figure 4G) were reduced in SLC27A3−/− cells compared to WT, but total levels of STAT3 protein were unaltered. Overall, our data suggest that SLC27A3 plays an important role in mediating asthma-relevant phenotypes in multiple cell types, likely by mediating fatty acid uptake, lipid mediator biosynthesis, and protein palmitoylation.

Loss of Slc27a3 is protective in a mouse model of allergic asthma

Both our screen results and in vitro cell experiments suggested that loss of SLC27A3 might be protective against asthma progression. To test our hypothesis, we generated Slc27a3−/− mice to evaluate the effect on a mouse model of allergic asthma⁵⁶. Mice intranasally sensitized to and then challenged with the fungal allergen Alternaria alternata (Alt) exhibit type 2 airway inflammation and remodeling similar to humans^57-59. Wild type and Slc27a3−/− mice were sensitized with Alt, rested for two weeks, and then challenged with Alt (Figure 4H). We measured several signature phenotypes of asthma physiology. Slc27a3−/− mice had significantly reduced allergen-induced recruitment of eosinophils to the airway compared to wild type controls (p-value=0.0220, Figure 4I). Furthermore, Slc27a3−/− mice exhibited blunted allergen-inducted airway remodeling (Figure 4J) compared to wild type controls, including significantly reduced epithelial thickening by H&E staining (p-value=0.0047; Figure 4J-K), mucus expression by Periodic Acid-Schiff (PAS) staining (p-value=0.0075, Figure 4J, L), and collagen deposition by Masson Trichrome (MT) staining (p-value=0.0440, Figure 4J, M). As we observed in cell models that loss of SLC27A3 and SCD had opposing effects on asthma-related phenotypes, we tested if loss of Scd1, the mouse homolog of SCD, would worsen the response to allergen in vivo. In response to allergic sensitization, Scd1−/− mice had significantly worse inflammation compared to wild type controls, demonstrated by increased airway eosinophilia (p-value=0.0455, Figure 4N) as well as significantly increased lung inflammation proximal to airways (p-value=0.0156) and vessels (p-value=0.0062, Figure 4O).

Overall, our results show that loss of two enzymes with opposing effects on protein palmitoylation also has opposing effects on cellular phenotypes in human epithelial and CD4+ T cells and a mouse model of allergic sensitization. We analyzed publicly available data from the Immune Mechanisms of Severe Asthma (IMSA) cohort⁶⁰ and found significantly increased expression of SLC27A3 in bronchial brushings collected from severe asthmatics (Figure 4P). Our results strongly indicate that the post-translational modification protein palmitoylation is likely a key pathway in asthma pathogenesis. These findings underscore the potential of DANDELION to generate novel insights into the genetics and molecular biology of heritable complex diseases.

DISCUSSION

We introduce DANDELION, a computational framework that prioritizes DPGs and maps their trans-regulatory networks. DANDELION prioritizes candidate genes and pathways that are missed by other methods. When deployed in combination with powerful experimental tools such as CRISPR screens, DANDELION generates actionable insights into the etiology of complex diseases. Applied to asthma, DANDELION identifies protein palmitoylation as a central disease-mediating process. We demonstrate that enzymes regulating substrate availability for palmitoylation are critical mediators of allergic inflammation and candidates for future therapeutic development.

Many DPGs identified by DANDELION are missed by GWAS and other gene prioritization methods using cis effects. DANDELION uses GWAS loci as exposures and prioritizes genes that may mediate the effects of GWAS loci (disease distal genes) on disease risk. Some DPGs are not near any GWAS-significant loci but are instead situated within gene regulatory networks that include GWAS-significant loci at the peripheries (Figures 1E, S6, S7). DANDELION’s ability to prioritize DPGs and link them to GWAS loci in trans-regulatory networks greatly improves the interpretability of GWAS loci. DPGs are evenly distributed across GWAS loci of various p-value strata (Figure 1G), complicating prioritization. Nonetheless, while GWAS p-values cannot nominate DPGs, increasing GWAS significance does scale with increasing proportions of HCGs, disease-distal genes that influence multiple DPGs. Interestingly, GWAS significance of HCGs also correlates with the number of DPGs they regulate (Figure S8 and S9). Genes such as HCGs that have high network connectivity are enriched for disease heritability and likely influence important biological functions^61-63. Our results suggest that GWAS significance can be used to prioritize them. DANDELION-generated trans-regulatory networks connect HCGs to DPGs, increasing functional interpretability of HCGs.

When applying DANDELION, we recommend using trans-regulatory associations measured in disease-relevant cell types, as trans effects are tissue- and cell-type-specific. While we rely on high-quality trans-eQTL resources with extensive correction for technical and biological confounders, trans-eQTLs derived from bulk tissues may still reflect a mixture of intracellular regulatory effects and cell-type composition effects. Although the latter do not represent within-cell gene regulation per se, they remain biologically meaningful, particularly in immune-related traits where variation in cell-type abundance is integral to disease mechanisms. Notably, the DANDELION framework is not limited to bulk trans-QTL data and can be readily applied to trans-regulatory signals derived from perturb-seq⁶⁴ in relatively homogeneous cell populations (e.g., cell lines), which would preferentially capture intracellular trans-regulatory effects. We have demonstrated DANDELION’s application to perturb-seq data from human primary CD4+ cells (Figure S11), consistent with recent studies that have leveraged trans effects from perturb-seq data for gene prioritization^64,65. DANDELION can be applied to many complex traits and diseases to reveal DPGs and their trans-regulatory networks, especially as transcriptome or proteome data become available in more tissues and cell types and WES studies increase in sample size. Our study is limited to the UK Biobank’s European cohort due to the lack of comparably large rare variant burden test results in other populations. Portability across ancestries has not yet been established. Future studies that evaluate cross-population portability of rare variant burden tests and DPGs would be valuable.

DANDELION is implemented using either gene-based or SNP-based trans-eQTLs. The SNP-based implementation leverages summary statistics from trans-eQTL studies, which can be derived from much larger sample sizes than datasets with individual-level genotype data. Our simulations showed that SNP-based DANDELION (trans-eQTL sample size N=30,000) can achieve higher power than gene-based DANDELION based on a much smaller sample (N = 900). Accordingly, our primary analyses utilize trans-eQTL summary statistics from large consortia such as eQTLGen. At the same trans-eQTL sample sizes, standard gene-based DANDELION can be more powerful than the SNP-based DANDELION. However, increased power is often accompanied by elevated false discovery rates, as we have demonstrated in our simulations (Figure S2). Therefore, the choice between implementations should be guided by the study objective, balancing discovery power with statistical significance stringency.

In using DANDELION to map asthma DPGs, we identified a novel pathway in asthma pathogenesis. The DPG whose perturbation led to the strongest asthma-protective phenotype in both epithelial and T cells was a fatty acid transport protein and acyl-CoA synthetase, SLC27A3. Interestingly, a second DPG whose perturbation led to one of the strongest asthma-like phenotypes was SCD, a member of the same pathway (Figure 4D). We showed that SLC27A3 and SCD affected asthma phenotypes through multiple mechanisms. First, we found that SLC27A3 promotes update of arachidonic acid and thus may affect the production of lipid mediators known to cause allergic response, such as leukotrienes and prostaglandins. Second, we found that SLC27A3 also promotes the uptake of palmitic acid and production of palmitoyl-CoA, which in turn affects protein palmitoylation. While over 3,000 proteins are thought to be palmitoylated⁶⁶, this modification has been predominantly investigated in the context of neurological diseases⁶⁷, cancer⁶⁸, and inflammation⁵⁰. Some prior evidence indirectly suggests palmitoylation is important for asthma pathogenesis. Palmitoylation regulates several key cell signaling processes that are important in asthma^53,69,70. Rare, deleterious loss of function mutations in SLC27A3 are associated with decreased risk of childhood-onset asthma and blood eosinophil count in the UK Biobank WES cohort¹² and SLC27A3 is significantly upregulated in bronchial brushings of severe asthma patients (Figure 4P). Conversely, expression of SCD is significantly reduced in bronchial epithelial cells of patients with asthma⁷¹. However, there has been no substantial investigation of what role protein palmitoylation might play in the context of asthma or other lung inflammatory diseases. Here, we demonstrated that loss of Slc27a3 protects against allergen-induced pathogenesis in mice, while loss of the SCD mouse homolog Scd1 exacerbates allergic inflammation. Our in vitro cell model results suggest that the protective effect of SLC27A3 inhibition might arise from a combination of improved epithelial barrier integrity and reduced Th2 cytokine production. Future studies will need to tease apart the distinct contribution of each cell type to the overall phenotype. Our findings represent a new direction for understanding mediators of asthma pathogenesis and a new potential therapeutic intervention in the treatment of asthma, either by targeting SLC27A3 specifically or palmitoylation more broadly.

Overall, our results showcase the ability of DANDELION to generate clinically meaningful insights into disease mechanisms by uncovering DPGs and mapping their trans-regulatory networks. We expect that the deployment of DANDELION to new traits and diseases will reveal a wealth of DPGs with high drug development potential while also facilitating a deeper understanding of the roles these genes play in disease etiology.

Limitations of the Study

Our study may have the following limitations. First, DANDELION may have reduced power to identify highly constrained genes, for which rare loss-of-function variants are extremely scarce and burden tests are inherently underpowered. However, such genes are not fundamentally inaccessible to the framework. By integrating trans-regulatory information from disease-associated loci, DANDELION can prioritize genes with relatively weak burden signals when supported by consistent trans effects. For example, SCD is a highly constrained gene (pLI = 1). DANDELION nominated SCD as a candidate mediator of asthma GWAS locus PKN3, despite its lack of statistically significant burden evidence (burden p-value=0.002), illustrating how trans-regulatory context can rescue otherwise undetectable signals. More broadly, while constrained genes may be missed in the absence of sufficiently powered trans data, their identification within DANDELION depends on the presence of disease-relevant regulatory connections rather than the strength of rare variant burden alone. Second, DANDELION relies on trans regulatory signals in disease-relevant tissues. Using trans effects from the wrong tissue might lead to false discoveries, as the trans regulation from GWAS loci to DPG is highly tissue-specific. Third, we control false discovery rate (FDR) at the exposure (gene/SNP) level rather than globally across all tested exposures, which increases power but implies a more liberal effective trait-level FDR when aggregating results genome-wide. Our simulations demonstrate that applying global FDR control across all exposures leads to a sharp loss of power—particularly for weak trans effects—such that few or no genes are detected under realistic settings (Figure S2E). As many exposure SNPs or genes are correlated, global FDR correction for all exposures is also overly stringent. This challenge is inherent to approaches leveraging trans-regulatory signals, which are individually modest and compounded by limited power in rare variant burden tests. Consistent with this, recent studies using trans effects for gene prioritization^64,65 have similarly adopted per-exposure or per-feature FDR control to balance true and false discoveries. Importantly, although this choice relaxes global FDR control, we observe substantial experimental validation of prioritized genes, supporting their biological relevance. Fourth, because DANDELION operates on GWAS-associated variants as exposures, enrichment of DPGs should not be interpreted as arising independently of GWAS ascertainment, but rather as reflecting the prioritization of putative mediator genes linking GWAS loci to disease risk. Fifth, the current study only used the LOF+missense, MAF<1% variant filter set burden test to nominate DPGs. This variant set has the largest discovery power in blood traits and asthma (Figure S15) but may not generalize to all traits. For certain diseases where loss-of-function variants play a dominant role, a LOF-only mask may be more appropriate than a combined LOF+missense mask.

STAR METHODS

EXPERIMENTAL MODEL AND STUDY PARTICIPANT DETAILS

Cell culture

Human immortalized 16HBE14o- bronchial epithelial cells (16HBEs, Sigma-Aldrich) were grown in DMEM supplemented with 10% FBS, 2 mM L-glutamine, and 1% penicillin/streptomycin on tissue culture-treated plastic coated with 10 μg/mL fibronectin, 100 μg/mL bovine serum albumin fraction V, and 30 μg/mL bovine collagen. Cells were maintained at 30-70% confluency and split as needed. All experiments were performed before passage 14. Human peripheral blood mononuclear cells (PBMCs) were obtained from a commercial vendor (STEMCELL Technologies). For CD4+ T cell assays, naïve CD4+ T cells were isolated from freshly thawed PBMCs using the EasySep Human Naïve CD4+ T Cell Isolation Kit (STEMCELL Technologies) and polarized to Th2 using ImmunoCult-XF T Cell Expansion Medium supplemented with ImmunoCult Th2 Differentiation Supplement (STEMCELL Technologies). Immediately following isolation, naïve CD4+ T cells were diluted to 1x10⁶ cells/mL, then activated with ImmunoCult Human CD3/CD28 T Cell Activator (STEMCELL Technologies). CD4+ T cells were counted every 3-4 days and diluted to 1x10⁶ cells/mL with fresh medium as needed; at day ten, cells were pelleted, resuspended in completely fresh medium, and re-activated.

Mice

All mice (C57BL/6J) were obtained from Charles River Laboratories, Inc. (Wilmington, MA, USA). Slc27a3−/− and Scd1−/− mice were generated using CRISPR/Cas9 genome editing described below. Mice were weaned at four weeks of age and housed up to five per cage. For allergic sensitization and challenge studies, 12-week-old male (Slc27a3−/− and associated WT control) or female (Scd1−/− and associated WT control) mice were anesthetized with isoflurane and intranasally administered 50 μg Alternaria alternata extract (Alt) in 50 μL PBS or PBS alone twice on day 1 and 2. Mice rested for two weeks, then were challenged with 25 μL Alt in 50 μL PBS or PBS alone three times on days 17, 18, and 19. On day 20, animals were terminally anesthetized with ketamine/xylazine and bronchoalveolar cells were collected via tracheostomy lavage in 2mL PBS. Lungs were then collected for further study, described below. Mice were housed on a 12-hour light/dark cycle with ad libitum access to food and water. All procedures were conducted with approval of the Institutional Animal Care and Use Committee (IACUC) of the University of Chicago (ACUP-71656; IBC0934).

METHOD DETAILS

DANDELION method

To identify DPGs, DANDELION combines trans-regulatory effects (γ) and trans target gene association with disease traits (β). DANDELION utilizes both gene-based trans-regulatory effects (gene-to-gene) and the standard SNP-based trans effects (SNP-to-gene). When raw genotypes and gene expression data are available, DANDELION computes gene-based trans-regulatory association p-values using GBAT (gene-based DANDELION). GBAT has two main steps. First, it predicts the expression level of a gene using all cis-SNPs with a leave-one-sample-out cross-validation approach, cvBLUP. Second, GBAT tests the trans-association between the predicted expression level of the gene and all genes that are on different chromosomes. When the raw data are not available, DANDELION can use summary statistics of SNP-based trans-effects (SNP-to-gene; SNP-based DANDELION).

DANDELION combines trans-association p-values and burden test p-values to perform mediation analysis using the Divide-Aggregate Composite-null Test (DACT)¹⁶. For a given disease-associated variant/gene and a disease, genome-wide genes can serve as candidate mediators. DACT decomposes the path from disease-associated variants/genes to genome-wide genes and then to disease into four disjoint cases.

Case 1:	γ = 0, β ≠ 0
Case 2:	γ ≠ 0, β = 0
Case 3:	γ = 0, β = 0,
Case 4:	γ ≠ 0, β ≠ 0

Case 1-3 correspond to the composite null cases of no mediation pathway, whereas Case 4 corresponds to the presence of a non-null mediation pathway.

Specifically, DACT tests for the presence of a mediation pathway using a linear structural equation model for illustration:

$$\begin{gathered} Y={\beta }_0+{\beta }_{x}X+\beta M+{\beta }_Z^{T}Z+{ϵ}_Y \\ M={\gamma }_0+\gamma X+{\gamma }_Z^{T}Z+{ϵ}_Y \end{gathered}$$

where Y is the complex trait, Z is the covariate term, M is the mediator, X is the exposure SNP/gene, and ϵ is the error term with normal distribution of zero mean and constant variance. More generally, the disease trait Y may be continuous or binary, and the framework is not restricted to quantitative traits. The null hypothesis of no mediation pathway is H₀: β_γ = 0, corresponding to Cases 1-3. DACT estimates the proportion of each case from genome-wide data, denoting them ω_case1, ω_case2 and ω_case3 respectively. The DACT statistics of mediator j are defined as: $DAC{T}_j={\widehat{\omega }}_{case1}{p}_{case1}+{\widehat{\omega }}_{case2}{p}_{case2}+{\widehat{\omega }}_{case3}{p}_{case3}$, Where as p_case1 and p_case2 are the p-values of trans-eQTLs and burden test respectively. p_case3 = (P_max)², where P_max is the p-value of the MaxP test.

Next, the DACT statistics are converted into Z-scores using the inverse normal cumulative distribution function:$Z_j^{DACT}={\Phi }^{-1}(1-DAC{T}_j)$, where Φ is the standard normal cumulative distribution function. We further transform the Z-score to the calibrated P-value: $p_j=1-\Phi (\frac{Z_j^{DACT}-\widehat{\delta }}{\widehat{\sigma }})$, where $\widehat{\delta }$ and $\widehat{\sigma }$ are the mean and standard deviation of the empirical null probability density function $f_0(z)$. Significant DACT p-values indicate the presence of a non-null mediation pathway, and the mediators are the DPGs. Details of the causal mediation assumptions, including the composite null’s structure and p-value calibration strategy, are provided in the original DACT paper¹⁶.

To control false positive signals, we convert the DACT p-values to q-values^72,73. In DANDELION, we control for FDR at each exposure gene or SNP, which is consistent with recent work⁶⁴ instead of at a trait (or global) level. We note the global FDR will be higher than the per exposure FDR level. We assessed the global FDR (controlling all genes/SNPs) in the simulations below.

Simulations

We first perform simulations to compare the performance of DANDELION with ARCHIE, which is a method that identifies sets of disease genes that are trans-targets of disease-associated SNPs, using sparse canonical correlation analysis. We simulated mediation paths that include: 1) cis effects of SNPs on disease-distal genes, 2) trans effects of disease-distal genes on target genes, and 3) direct effects of target genes on disease (Figure S1). In our simulations, we designated 10 out of 1000 potential trans-target genes as “disease-proximal genes” and evaluated the performance of both methods by comparing the power, defined as the proportion of disease-relevant genes identified, and global FDR.

Specifically, for our simulations, we generated 90 independent SNPs with minor allele frequencies ranging from 10% to 40% in 230,000 samples, which were divided for simulating complex traits (n= 200,000) and gene expression levels (n=900 for standard DANDELION and n=30,000 for SNP-based DANDELION). This division reflects real-world datasets in which gene expression and GWAS datasets do not have overlapping samples. We simulated a mediation network comprising 18 disease-distal genes regulated in cis by 5 SNPs each and 1,000 target genes regulated by the disease-distal genes in trans. A diagram of our simulation can be found in Figure S1. The cis SNP effects on disease-distal genes follow $N(0,{\sigma }_{cis}^2)$, with the per-SNP cis heritability (${\sigma }_{cis}^2$) of 0.04. The trans effects of disease-distal genes on the trans target genes were simulated as $\gamma \sim N(0,{\sigma }_{\gamma }^2)$, where the trans heritability ${\sigma }_{\gamma }^2$ had values of 0.0025, 0.004, or 0.01. For gene-based DANDELION, we simulated gene expression for 900 samples. We assigned 10 of the 1,000 trans-target genes to have a causal effect β on the trait, thus the 10 genes were disease-proximal genes. The gene effect β followed normal distribution with zero mean and per-gene heritability of 0.004. To run ARCHIE, we picked the set of SNPs significantly associated with the simulated trait, and ARCHIE reported the set of genes associated with the set of SNPs in trans. To correct for multiple testing, DANDELION uses different per-disease-distal -gene false discovery rates (per-gene FDR=0.01 ,0.05 and 0.1). The global FDR was the overall FDR for testing 18 disease-distal genes. The ARCHIE software reports sets of genes, and there is no explicit p-value or FDR controls. Power was computed as the proportion of DANDELION/ARCHIE genes over the 10 putatively causal disease-proximal genes. We performed 300 simulations for each parameter setting.

In both DANDELION and ARCHIE, the power increases with trans effects, but DANDELION significantly outperforms ARCHIE at all FDR levels (Figure S2). We also found that DANDELION has a much lower global FDR across trans effects (Figure S2). Similar to the gene-based DANDELION, the power of SNP-based DANDELION is higher than ARCHIE across most parameter settings. SNP-based DANDELION achieves similar power with a gene expression sample size of 30,000 as the standard DANDELION at a gene expression sample size of 900. Therefore, we recommend gene-based DANDELION whenever possible.

DANDELION can be applied to different numbers of exposure genes/SNPs, but we only control for FDR at each exposure gene level. We evaluated the global FDR (controlling for FDR at all exposure genes) for different numbers of independent exposure genes (N_peripheral=20,100,500,1000). The simulation for this analysis setting was similar to the power simulation. We found that the global FDR of identifying DPGs for 18 exposure genes is slightly higher than the FDR level of each gene, which is expected. Global FDR increased as additional exposure genes/SNPs were included, as expected under per-exposure FDR control. However, this increase was sublinear rather than proportional to the number of tested exposures (Figure S2E). Applying global FDR controls across all exposure genes substantially reduced power, especially when trans-regulatory effects were modest. We did not combine tests of all exposure genes to perform multiple testing correction, as it would be overly stringent, especially when exposure genes are correlated in real data analysis. Instead, DANDELION used per-gene FDR. The global FDR simulations serve as a reference for users when selecting per-exposure gene FDR levels (Figure S2F).

Applying DANDELION to UK Biobank blood traits

We downloaded burden test summary statistics of the 27 different whole blood traits of UK Biobank from Backman et al. (2021). Burden tests of various definitions of variant sets were performed by Backman et al, and we used rare putative loss-of-function variants and likely deleterious missense variants with an MAF of up to 1%, as it generally produced that largest number of significant associations in burden test and largest number of significant DPGs (Figure S15). We also performed sensitivity analyses using different variant filters and found that the loss-of-function and missense with MAF<1% can replicate other variant filters to a large extent (Figure S16). We used trans-eQTLs (n= 31,684)²⁰, which includes 10,317 variants and 15,057 genes. For each trait, SNPs within 1Mb of GWAS significant loci served as exposure SNPs. We applied DANDELION using trans-eQTL p-values and summary statistics from Backman et al. ¹² to identify candidate trans-genes (>5 Mb) associated with the exposure, targeting a per-exposure gene false discovery rate (FDR) of 5%. To map SNPs to their cis-genes, we used cis-eGenes with the smallest p-values from the eQTLGen study²⁰. SNPs without significant cis-eGenes are mapped to the nearest gene.

Mapping blood trait GWAS loci to DANDELION trans-regulatory network

We downloaded independent GWAS significant loci of the UK Biobank for platelet crit (Study ID: GCST90002400) and lymphocyte count (Study ID: GCST90002388) from the GWAS catalog. As there are only 10,361 SNPs from eQTLGen that serves as exposure SNPs, we mapped them to the GWAS loci if they are within the 1Mb range. For DPGs, we mapped GWAS loci to the DPG if they are within 1Mb to the transcription starting sites of the DPG.

Enrichment analyses in Mendelian blood disorder genes

In this section, we estimate the enrichment of genes (such as DPGs) in Mendelian blood disorder genes¹⁹. We first stratify the genes identified by DANDELION and burden tests into several categories: (1) DPGs that are also burden test significant (2) DPGs that are not burden test significant (3) burden test significant genes that are not DPGs. We also create a fourth category of non-DPG genes that match the burden test significance levels of genes in Category 2. This set of genes might barely miss the burden test significance threshold. Category 2 and Category 4 are used to generate results in Figure 1E.

To compute enrichment, we randomly generate 1000 gene sets that match the gene length of genes in each category. We then summarize the number of overlaps between the random gene sets and the Mendelian disorder genes, to obtain baseline distribution of overlapping genes. The fold enrichment is ratio of the actual number of overlaps between genes in each category and Mendelian disorder genes over the mean of the baseline distribution. The p-values are the number of random gene sets with overlaps greater than the actual number of overlaps of each category.

Identify asthma DPGs with DANDELION

We applied DANDELION to identify asthma DPGs. We used burden test summary statistics from Backman et al. (2021), with a variant set containing rare putative loss-of-function variants and likely deleterious missense variants with an MAF of up to 1% (GWAS catalog study ID: GCST90085447).

We used trans-regulatory signals detected from two datasets: eQTLGen (n=31,684)²⁰ and DGN (n=913)³¹. The analysis using eQTLGen is similar to that used for the blood traits and we used per-exposure FDR of 5% to identify significant DPGs. We mapped exposure SNPs to significant cis-eGenes with the smallest p-values in eQTLGen, or the nearest genes if the SNP is not a cis-eQTL in eQTLGen. For DGN, we used gene-based trans effects estimated with GBAT¹⁵. Due to the much smaller sample size, we used a per-exposure FDR cutoff of 10% to identify significant DPGs in DGN.

To directly assess whether DANDELION discoveries were driven by generic trans hubs or degree-related network artifacts, we repeated the analysis after removing high-degree regulators and targets from the trans-regulatory effects and re-identifying asthma DPGs. We performed this analysis using both the GBAT- and eQTLGen-derived trans-effects. For the GBAT analysis, we removed five high-degree regulators that each regulated more than three trans targets; there were no high-degree target genes regulated by multiple regulators. For the eQTLGen network, we removed 1050 high-degree regulator SNPs that each regulated more than five trans targets, as well as one high-degree trans target gene.

Primary CD4+ T Cell Perturb-seq Data

We analyzed a genome-scale CRISPRi perturb-seq dataset generated from primary human CD4+ T cells across four healthy donors 8 hrs after anti-CD2/3/28 restimulation³⁴. Following quality control filtering, which includes retaining cells with ≥ 800 detected genes, ≥ 1,200 UMI counts, and < 5% mitochondrial reads, and excluding cells with multiple guide RNA assignments, we retained approximately 7 million high-quality cells for downstream analysis. All 7 million cells were assigned a single targeting or non-targeting guide RNA. Guide RNAs targeting 12,662 unique genes were detected with an average of approximately 550 cells per perturbed gene. In this replication analysis, we limited to 170 perturbed genes that are around 200kb of exposure genes for asthma DANDELION result.

Identify trans-perturb-QTLs with SCEPTRE

To identify trans-regulatory effects of CRISPRi perturbations on genome-wide gene expression, we applied SCEPTRE⁷⁴ to the CD4+ T cell perturb-seq data. Count matrices from all four donors were merged into a single dataset by exporting CellRanger-format h5ad files with a globally consistent feature set (union of genes and guides across all donors). SCEPTRE objects were constructed using import_data_from_cellranger() with low-MOI settings and ondisc-backed storage for memory efficiency. Donor identity was extracted from cell metadata and included as a covariate alongside log-transformed UMI count and number of non-zero genes in the regression formula. Trans-association pairs were constructed using construct_trans_pairs(), excluding positive control pairs. Association testing was performed using SCEPTRE's default negative binomial regression framework, which accounts for cell-level technical covariates and applies resampling-based calibration. Results were filtered to pairs passing quality control (minimum 7 non-zero cells in treatment and control groups).

Replicating DPGs nominated from whole blood trans-eQTLs by DPGs identified with trans-perturb-QTLs from CD4+ T cells

Trans-association p-values from SCEPTRE for CD4+ T cell perturb-seq data and whole-exome sequencing (WES) burden test results from the UK Biobank were used to nominate DPGs. 16 DPGs are expressed in the perturb-seq data. All 16 DPGs can be replicated by the CD4+ T cell perturb-seq data. Ranking CD4+ T cell DPGs (N = 11,096) by minimum DANDELION p-value, 15/16 DPGs are among the top 0.5% (Figure S11).

Human Lung Cell Atlas data

We analyzed single-cell RNA sequencing data from the Human Lung Cell Atlas (HLCA)³⁶ to assess the cellular contexts of asthma disease-proximal and -distal genes. The HLCA integrates over 2 million cells from the respiratory tracts of 486 donors, compiled from 49 independent studies. We used the core HLCA data, which contains healthy lung tissue from 107 individuals with manually curated cell-type annotations and comprehensive metadata.

The HLCA metadata includes batch-corrected low-dimensional embeddings (UMAP), which we used to visualize both the overall cellular landscape and the expression patterns of disease-proximal genes across cells. To achieve finer resolution in cell-type interpretation, we utilized the hierarchical cell-type annotations provided in the HLCA metadata, combining level 1, which represents broad cell types such as epithelial, endothelial, immune, and stromal cells, and level 3 cell-type annotations, which provide finer subdivisions within these groups (e.g., alveolar type 2 cells, T regulatory cells, etc) for all UMAP visualizations. The dataset is publicly available through CELLxGENE Census (https://cellxgene.cziscience.com/collections/6f6d381a-7701-4781-935c-db10d30de293).

AUCell analysis of disease-proximal genes

To identify the lung cell types with high expressions of asthma DPGs, we applied the AUCell workflow (version 1.16.0) in R using the HLCA core dataset. The single-cell count matrix was loaded from .h5ad format and processed in chunks of 10,000 cells to manage memory efficiently. For each chunk, gene expression values were converted to sparse matrices, and AUCell_buildRankings() was used to rank genes per cell. Chunked rankings were combined into a single matrix and stored in a SummarizedExperiment object for downstream analysis. We used the 21 DPGs as the gene set. The activity of the gene set was computed using AUCell_calcAUC(), which calculates the Area Under the Curve (AUC) for the recovery of the gene set within the ranked gene expression profiles. Active cells were identified using AUCell_exploreThresholds(), which defines threshold values and classifies cells with high AUC scores as actively expressing the DPGs. The distribution of AUC scores was visualized to highlight cells showing the strongest DPG activity. Cell-type identities were annotated using HLCA metadata that combines level 1 and level 3 cell-type annotations, providing detailed resolution of lung cell populations. The number and proportion of nominated cells were calculated to identify cell types most relevant to the DPGs.

Gene nomination for validation assays

The effect of disease-proximal and disease-distal genes on molecular and cellular asthma-relevant phenotypes was evaluated using CRISPR-Cas9 gene knockdown screens in 16HBE and primary T cells. Disease-proximal and disease-distal genes were included in the 16HBE screen if their expression was > 1 TPM by RNA-seq. Negative control genes for the 16HBE screen were randomly selected as a group to generate a similar TPM distribution and their lack of proximity to GWAS loci was confirmed prior to use. Molecular and cellular phenotyping was also performed for genes of interest identified by genome-wide association study (GWAS). Briefly, single nucleotide polymorphisms associated with risk of asthma by fine-mapping³⁹ were linked to potential target genes in two-dimensional (nearest gene) or three-dimensional (promoter capture Hi-C) proximity in bronchial epithelial⁴⁵ or T cells⁴⁶. The 30 GWAS candidate genes with highest expression in 16HBEs were included in the screen, of which six were nearest genes to putative regulatory elements containing SNPs with high PIP, 21 looped to putative regulatory elements in bronchial epithelial cells, and five looped in T cells. The nearest two genes flanking every core gene were also included. The three long non-coding RNAs identified as disease-distal genes by DANDELION were not included in the screen due to technical limitations. All genes of interest and associated guide sequences are listed in Table S10.

Gene knockdown in bronchial epithelial cells

Genes of interest were knocked down in 16HBE cells using Cas9 ribonucleoproteins (RNPs). Briefly, guides targeting expressed exons of interest were designed with GUIDES using GTEx data from lung, skin, and blood⁷⁵. To form RNPs, crRNAs were duplexed with a universal tracrRNA (Integrated DNA Technologies) at equimolar concentrations, then mixed with Alt-R Cas9 enzyme (Integrated DNA Technologies) at a ratio of 1.15:1 with equal volume PBS. 16HBE cells were nucleofected with RNPs using the 4D-Nucleofector X Unit program CM-158 (Lonza). Nucleofection was performed on 2.5x10⁵ cells per knockdown using SG Cell Line 4D-Nucleofector X Kit S (Lonza). Immediately after nucleofection, cells were transferred to a 6-well plate to recover for one week. Knockdown efficiency was determined by qPCR comparing expression of cells that received a targeting crRNA versus a non-targeting crRNA. qPCR primers are listed in Table S11.

Transepithelial electrical resistance

Transepithelial electrical resistance was measured by electric cell-substrate impedance sensing (ECIS) with an ECIS 96-well station (Applied Biophysics). One week after nucleofection, 6x10⁴ cells per well were plated on an extracellular matrix mix-coated 96-well electrode (Applied Biophysics). ECIS measurements were collected at multiple frequencies for 72 hours. To determine barrier strength at initial cell-cell contact, approximated as the time at which capacitance stabilized at 64,000Hz, resistance was measured for four wells per genotype at 500Hz. Statistically significant resistance compared to cells treated with a non-targeting guide was determined by one-way ANOVA with Dunnett’s multiple comparisons test. To determine strength of the mature barrier, occurring 60-72 hours after plating, the average normalized resistance to hour 0 was calculated for four wells per genotype at 4,000Hz. Statistically significant normalized resistance compared to cells treated with a non-targeting guide was determined by one-way ANOVA with Dunnett’s multiple comparisons test.

16HBE RNA-sequencing

RNA was isolated from 16HBEs treated with non-targeting (NT) or SLC27A3-targeting sgRNAs as described above using the RNeasy Mini Kit (Qiagen). Four replicates were treated with NT sgRNA and three replicates were treated with SLC27A3-targeting sgRNA. RNA-seq libraries were prepared using NEBNext Poly(A) mRNA Isolation Module and NEBNext Ultra II Directional RNA Library Prep Kit (New England Biolabs), then sequenced on the Aviti (Element Biosciences) using paired end 50bp reads. Read counts per gene were obtained with Salmon version 1.5.1⁷⁶ on transcripts from human Gencode release 19 (ftp://ftp.ebi.ac.uk/pub/databases/gencode/Gencode_human/release_19/gencode.v19.pc_transcripts.fa.gz and ftp://ftp.ebi.ac.uk/pub/databases/gencode/Gencode_human/release_19/gencode.v19.lncRNA_transcripts.fa.gz). Estimated counts were used in exploratory analysis (transformed with DESeq2’s rlog function) and in DESeq2 version 1.32.0⁷⁷ to identify differentially expressed genes (adjusted p-value ≤ 0.05 and log₂ fold-change of ≥ 0.6).

Pooled CRISPR screen library generation

The CRISPR knockout screens in T cells were performed following Joung et al. ⁷⁵ with modifications. Briefly, four guides per gene of interest were designed with GUIDES using GTEx data from lung, skin, and blood, CRISPOR⁷⁸, or selected from predesigned gRNAs made available by Integrated DNA Technologies. The pool of guide oligos (Twist Biosciences) was amplified for 16 cycles using NEBNext Hi-Fi DNA polymerase (New England Biolabs), cloned into the lentiCRISPRv2GFP backbone (82416, Addgene) via Gibson assembly, and then electroporated into NEB10β electrocompetent cells (New England Biolabs). Input library plasmid DNA was isolated using Nucleobond XF Midi Plus EF kit (Macherey-Nagel). Quality of the input library was assessed by next generation sequencing using the Element Aviti (Element), which found that 94.9% (quality control threshold > 70%) of guides were perfect matches, 0.0% (QC threshold < 0.5%) of guides were undetected in the input library, and the skew ratio between top 10% and bottom 10% was 1.65 (QC threshold < 10%).

Lentivirus production

The guide library was packaged in lentivirus in HEK293FT cells using Lipofectamine PLUS (Invitrogen) and packaging plasmids pMD2.G and psPAX2 (Addgene)⁷⁹. Concentrated lentivirus was resuspended in 100 μL PBS for every 15mL of supernatant. To titer lentivirus and determine the amount required for target MOI, five wells of 5x10⁵ naïve CD4+ were spinfected as described below with lentivirus concentrations ranging from 10%-60%. Seven days post-transduction, transduction efficiency was assessed using a 5-laser Cytek Aurora spectral flow cytometer (Cytek Biosciences) located at the University of Chicago Human Disease and Immune Discovery (HDID) Core Facility (RRID:SCR_022936). Raw spectral data were acquired using SpectroFlo software and unmixed using reference controls comprising a combination of single-color stained compensation beads (UltraComp Plus, Invitrogen) and cells. Post-acquisition analysis was performed using FlowJo software (version 10.10, BD Biosciences). Live, single cells were identified by excluding debris (FSC vs. SSC), doublets (FSC-A vs. FSC-H), and dead cells (LIVE/DEAD Fixable Blue, Thermo Fisher). Transduction efficiency was quantified as the percentage of GFP+ events within the live, single-cell population for each concentration.

Pooled CRISPR screen and analysis

Four samples of CD4+ T cells from three donors were transduced with lentivirus one day after thawing at MOI = 0.2. CD4+ T cells were transduced in a 24-well plate as 6 wells of 9.2x10⁵ cells each. The original media was removed from each well and stored at 37C, then replaced with 1 mL fresh media + 8 μg/mL polybrene + lentivirus. Cells were spinfected at 1000 x g for 90 min at room temperature, then the media was removed in its entirety and replaced with the original media. Fourteen days after spinfection, CD4+ T cells were stimulated with 5ng/mL PMA (Sigma-Aldrich) and 500ng/mL ionomycin (Sigma-Aldrich) for two hours, then 5 μg/mL brefeldin A (Biolegend) and 2 μM monensin (Biolegend) for an additional two hours. Cells were stained with LIVE/DEAD BLUE (Invitrogen), fixed in 1% PFA for 15 min, then stained for markers of interest for 20 min on ice: CD4 (BUV661, BD Biosciences), CD8 (BV785, Biolegend), CD45 (PerCP-Cy5.5, BD Biosciences), and IL-13 (PE/Cy7, Biolegend). Cells were resuspended in PBS + 1% BSA to 1x10⁷ cells/mL. UltraComp eBeads Plus compensation beads (Invitrogen) stained with single antibodies were used as compensation controls. Cell sorting was performed using a BD FACSymphony S6 cell sorter (BD Biosciences) at the University of Chicago Human Disease and Immune Discovery (HDID) Core Facility (RRID:SCR_022936). A 70-μm nozzle was used with purity sort settings. Lymphocytes were identified based on forward (FSC) and side (SSC) scatter properties, followed by singlet discrimination and exclusion of dead cells (LIVE/DEAD Fixable Blue, Thermo Fisher). Live, single cells were sequentially gated for CD45+ GFP+ expression, followed by selection of the CD4+ CD8− T cell subset. From this population, cells were sorted into three distinct fractions based on IL-13 expression intensity: the highest 15%, the lowest 15%, and the middle 55%. Sorted cells were collected into PBS + 1% BSA. After sorting, cells were pelleted by centrifugation and snap-frozen.

Cells were lysed in NK lysis buffer (50 mM Tris, 50 mM EDTA, 1% SDS) with proteinase K overnight at 55C to reverse crosslinking, then DNA was precipitated with isopropanol. sgRNA sequences were amplified for 23 cycles using NEBNext Hi-Fi DNA polymerase (New England Biolabs) with primers containing NGS sequencing barcodes⁷⁵. PCR products were purified using the QIAQuick PCR Purification Kit according to the manufacturer’s instructions. Next generation sequencing libraries were sequenced by the University of Chicago Genomics Facility on an Aviti (Element Biosciences). sgRNA enrichment analysis was performed using MAGeCK⁸⁰, comparing the IL-13-high population counts to counts averaged across IL-13-high, IL-13-low, and middle from the same donor using mageck test --paired. Significantly depleted sgRNAs were identified as having FDR < 0.25 and log₂ fold change less than the non-targeting guide mean (−0.2044). Significantly enriched sgRNAs were identified as having FDR < 0.25 and log₂ fold change greater than 0.2044. For a gene to be reported in Figure 3, it had to be targeted by 2+ significant sgRNAs in 2+ samples.

Primary T cell RNA-sequencing

Primary T cells from three independent donors were transduced with lentivirus containing a GFP reporter, the Cas9 ORF, and sgRNAs targeting SLC27A3 or a non-targeting control as described above. The next day, GFP+ cells were isolated using a FACSymphony B6 (BD Biosciences), then polarized to the Th2 lineage. One sample transduced with lentivirus containing the SLC27A3 sgRNA was contaminated; the non-targeting sample from the same donor was also discarded. Fourteen days after spinfection, cells were stimulated with 5ng/mL PMA (Sigma-Aldrich) and 500ng/mL ionomycin (Sigma-Aldrich) for four hours, then pelleted. RNA was isolated from cells using the RNeasy Micro Kit (Qiagen) and libraries prepared and sequenced as described above. Reads were aligned using STAR⁸¹, read counts generated with HTSeq⁸², and differential gene expression determined using edgeR⁸³ with a model including donor (adjusted p-value ≤ 0.05 and log₂ fold-change of ≥ 0.6).

Generation of knockout cell lines

SLC27A3−/− 16HBEs were generated with CRISPR/Cas9. Briefly, a crRNA targeting exon 5 of SLC27A3 (5’-TGGACTGACAGAGGGCAACG) was duplexed with a universal tracrRNA (Integrated DNA Technologies) at equimolar concentrations, then mixed with Alt-R Cas9 enzyme (Integrated DNA Technologies) at a ratio of 1.15:1 with equal volume PBS. 16HBE cells were nucleofected with RNPs using the 4D-Nucleofector X Unit program CM-158 (Lonza). Nucleofection was performed on 2.5x10⁵ cells per knockdown using SG Cell Line 4D-Nucleofector X Kit S (Lonza). Immediately after nucleofection, cells were transferred to a 6-well plate to recover for 24 hrs. Limited dilution cloning was then used to isolate single clones. Sanger sequencing identified a single base-pair insertion (A) in exon 5 introducing a premature stop codon in exon 6; qRT-PCR confirmed loss of mRNA transcript with primers listed in Table S11.

Lipidomics

For the lipid profiling, 16HBE wild type and SLC27A3−/− cells were grown for 24 hrs in complete culture media, then washed with 0.5 mL PBS at room temperature followed by metabolism quenched with dry ice-cold methanol (300 μL) containing 0.1mg/mL butylated hydroxytoluene (BHT, as an antioxidant). Liquid-liquid extraction was performed. Briefly, 300 μL cell extract was sonicated in water bath for 3 min at room temperature, vortexed for 10 min at 2000rpm and 15°C using a thermomixer, and then 900uL of methyl tert-butyl ether (MTBE) was added shaken for 10 min. To achieve phase separation, 300 μL water was added to the solutions and shaken for 10 min. Finally, samples were centrifuged at 20,000g and 15°C for 20 min. A portion (600 μL) of the upper organic layer containing lipid extract was pipetted out and dried down using the Genevac EZ-2.4 elite evaporator. The dried lipid layer was re-suspended in 60 μL of 3/2/1 isopropanol/acetonitrile/water at room temperature, sonicated for 3 min, vortexed for 15 min at 15 °C, 2,000rpm and centrifuged for 20 min at 18,000g and 15°C; then, 50 μL of supernatant was transferred to a LC–MS vials. A pooled quality control (QC) sample was generated using the remaining samples. A 2uL or 3uL aliquot was injected into the mass spectrometer. A Thermo Scientific Vanquish Horizon UHPLC system (Accucore C30 column, 2.1x150 mm, 2.6 μM) coupled to Orbitrap IQ-X Tribrid mass spectrometer with an H-ESI probe operating in either positive or negative mode was used to separate and detect lipids. The mobile phase A (MPA) was 60/40 acetonitrile/water containing 10 mM ammonium formate + 0.1% formic acid and MPB was 90/10 isopropanol/acetonitrile + 10 mM ammonium formate + 0.1% formic acid. The chromatographic gradient of MPB was, at 0 min: 30%, 2 min:43%, 2.1 min 55% ,12.00 min: 65%, 18.00 min: 85%, 20.00 min:100%, 25 min: 100%, 25.1 min: 30% and 30.00 min: 30%. The column temperature and flow rate were 45°C, and 0.260 mL/min, respectively. MS1 parameters were as follows: spray voltage: 3,500 V for positive ionization and 2,500 V for negative ionization modes, sheath gas: 40, auxiliary gas: 10, sweep gas: 1, ion transfer tube temperature: 300 °C, vaporizer temperature: 350 °C, Orbitrap resolution: 120 K, scan range (m/z): 250–2,000 for positive, 200–2,000 for negative, radio frequency (RF) lens (%): 60, AGC target: 50%, and a maximum injection time (maxIT) of 100 ms. Quadrupole isolation and Internal calibration using Easy IC were enabled. Lipids were identified by performing a comprehensive data dependent HCD MS2 experiment with conditional CID MS2 and MS3 data-acquisition strategy. Initially, MS1 data was acquired in full-scan mass range (200–2,000) followed by the data-dependent (dd) MS2 with HCD collision energy (%) at 25, 30, 35, Orbitrap resolution 30,000 and maxIT 54 ms. For positive mode A total of six injections for positive mode and four injections for negative mode were made to generate fragmentation data using the AcquireX workflow using the pooled QC samples. A list of identified lipids (precursor tolerance ±3 ppm, production tolerance ±5.0 ppm, product threshold 1.0) was generated using Thermo Scientific LipidSearch software version 5.0. The precursor adduct [M + H] was used to identify hexosylceramide (HexCer), sphingomyelin (SM), sphingosine (SPH), methylphosphatidylcholine (MePC), coenzyme Q (CoQ), acylcarnitine, lysophosphatidylcholine (LPC), lysophosphatidylethanolamine (LPE), phosphatidylcholine (PC) and phosphatidylethanolamine (PE) species. To identify triglycerol (TG), diacylglycerol (DG) and cholesteryl ester species, precursor adduct [M + NH4] was used. The identified lipid species were quantified using the Compound Discoverer 3.3 and Skyline45⁸⁴ (v24.1) software. Significant differences between lipid species of interest reported in Figure 4 were determined using Welch’s t-test.

For free fatty acid analysis of culture medium, 30 μL of media was extracted with the 225μL of methanol (with 0.1mg/mL BHT), 750μL of MTBE and 225μL of water. The upper organic layer was collected and dry down using Genevac EZ-2.4 eilite. The samples were re-suspended in 3/2/1 IPA/MeCN/water before injecting on LC-MS. On the day of the LC-MS analysis, samples were re-suspended in 80μL of 3/2/1 isopropanol/acetonitrile/water, sonicated for 2 min, thermomixer for 10 min at 15°C, and supernatant was transferred to LC-MS vial after spinning them down at 18,000g for 20 min at 15°C. Fatty acids were separated on Cortecs T3 (2.1x100 mM, 1.6micron, part # 186008499 Waters Corporation) column connected to a Vanquish Horizon UHPLC system and IQ-X tribrid mass spectrometers. The column temperature, injection volume, and flow rate were 40°C, 5μL and 0.3mL/minute, respectively. The mobile phase A (MPA) was 60/40 acetonitrile/water, 10 mM ammonium formate+0.1% formic acid and MPB was 89.1/9.9/0.99 isopropanol/acetonitrile/water, 10 mM ammonium formate +0.1% formic acid. The chromatographic gradient was 0 min: 5%B, 1.2 min: 5%B, 5 min: 50%B, 9.00 min: 50%B, 13.00 min: 70%B, 15.5 min: 70%B, 17.00 min: 100%B, 20.00 min: 100%B, 20.2 min: 5%B, 28.00 min: 5%B. MS1 parameters were as follows: spray voltage: 2800 V negative, sheath gas: 40, auxiliary gas: 10, sweep gas: 1, ion transfer tube temperature: 300°C, vaporizer temperature: 350°C, orbitrap resolution: 60K, scan range(m/z): 120-500, RF lens(%): 60, automatic gain control (AGC) target: 100%, and a maxIT of 118 milliseconds (ms). The fatty acids were identified by matching their retention time to the respective commercially available reference standards. Significant differences between free fatty acids were determined using Welch’s t-test. P-values were not corrected for multiple comparisons.

C16:0-alkyne probe labeling and click chemistry

Total protein palmitoylation was assessed using C16:0-alkyne probe labeling. Wild type and SLC27A3−/− 16HBEs were treated with 50 μM palmitic acid analogue C16:0-alkyne for 6 hours and then collected and lysed in 1% NP-40 lysis buffer (25mM Tris-HCl pH 8.0, 150 mM NaCl, 10% glycerol, 1% Nonidet P-40) with protease inhibitor cocktail. The supernatant was collected after centrifugation at 17,000xg for 10 min at 4°C. The protein concentration was determined via BCA assay (23225, Thermo Fisher). Click chemistry reagents were added to 50 μL lysates in the following order: 1 μL 4 mM TAMRA-azide (47130, Lumiprobe), 1.2 μL 10 mM tris[(1-benzyl-1H-1,2,3-triazol-4-yl) methyl]amine (TBTA) (T2993, TCI chemicals), 1 μL 40 mM CuSO4, 1 μL 40 mM tris(2-carboxyethyl)phosphine HCl (TCEP-HCl) (580560, Millipore). The reaction mixtures were mixed thoroughly and incubated for 30 min in the dark at room temperature. Total protein was precipitated with chloroform/methanol/water (v/v/v 1:4:3), washed thoroughly with cold methanol, resuspended in resuspension buffer (4% SDS, 50 mM triethanolamine pH 7.4, 150 mM NaCl), and diluted with 6x SDS loading dye (0.33M Tris-HCl (pH 6.8), 50% glycerol, 12% SDS, 0.3M DTT, 0.06% (w/v) bromophenol blue). Samples were heated at 95 °C for 10 min and resolved by SDS–PAGE. The gel was incubated with destaining buffer (50% methanol, 10% acetic acid, 40% water) by shaking for 4 hrs at 4°C, incubated in water to reduce background, then scanned to record the rhodamine fluorescence signal using a ChemiDoc imaging system (Bio-Rad). The gel was then stained with Coomassie Brilliant Blue (B7920, Sigma) to check for protein loading. C16:0-aklyne probe labeling and subsequent click chemistry and gel staining was performed twice.

Acyl-biotin exchange

Palmitoylation of STAT3 was determined using acyl-biotin exchange (ABE)⁸⁵. Briefly, wild type and SLC27A3−/− 16HBEs (n=3 replicates per genotype) were lysed with ABE lysis buffer (100 mM Tris-HCl pH 7.2, 150 mM NaCl, 2.5% SDS, 10 mM N-ethylmaleimide (NEM, 04259, Sigma), protease inhibitor cocktail, nuclease) to block the free cysteines with NEM at room temperature for 1hr. NEM was then quenched with 100 mM 2,3-dimethyl 1,3-butadiene (145491, Sigma) at room temperature for 1r and extracted with one-tenth volume of chloroform. Supernatant was incubated with 0.5M hydroxylamine (467804, Sigma) or Tris-HCl pH 7.5, together with 0.1 mM biotin-HPDP (16459, Cayman) at room temperature for 1hr. Total protein was precipitated with one volume of chloroform, four volumes of methanol and three volumes of ddH₂O, briefly air-dried, and resuspended with ABE resuspension buffer (100 mM Tris-HCl pH 7.2, 150 mM NaCl, 5 mM EDTA, 2.5% SDS, 8 M urea). The dissolved proteins were diluted with PBS and palmitoylated proteins (biotin-HPDP-linked) were enriched through streptavidin beads (20361, Thermo Fisher). Pull-down and input samples were used for immunoblotting to assess STAT3 palmitoylation levels. Densitometry was performed using FIJI. Statistical significance was determined using Welch’s t-test.

Western blots

Relative abundance and phosphorylation of STAT3 was assessed by Western blot. Protein was isolated from wild type and SLC27A3−/− 16HBEs (n=6 replicates per genotype) using RIPA buffer (10 mM Tris-HCl, 1 mM EDTA, 0.5 mM EGTA, 0.1% sodium deoxycholate, 140 mM NaCl, 1% Triton X-100, 0.1% SDS) with protease inhibitor cocktail (Roche) and phos-stop (Roche) for 30 min at 4°C. Samples were centrifuged to pellet insoluble material and then supernatant total protein was quantified using Bradford reagent (Bio-Rad). Protein was separated using 4%–15% gradient gels (Bio-Rad) and then transferred to PVDF membranes (Bio-Rad) for 3 hours at 300 mA, 4°C. Blots were blocked for 90 minutes at room temperature in StartingBlock T20 (Thermo Fisher Scientific) and incubated with primary antibody diluted in T20 overnight at 4°C. Primary antibodies used were against STAT3 (9139, CST), p-STAT3 Tyr705 (9145, CST), and beta-actin (3700, CST). Blots were washed 4 times in 1× TBS with 0.1% Tween 20 (TBST), incubated for 1 hour at room temperature with secondary antibodies conjugated to HRP (Jackson Immuno Research Labs), and washed again in TBST. Blots were imaged with SuperSignal West Pico Chemiluminescent Substrate. Membranes were first blotted for p-STAT3, then stripped for 10 min at RT (Santa Cruz), blocked, and re-blotted for STAT3. Densitometry was performed using FIJI. Statistical significance was determined using Welch’s t-test.

Generation and genotyping of knockout mice

Slc27a3−/− mice were generated by the Transgenic ES Cell Technology Core at the University of Chicago using CRISPR/Cas9 genome editing with sgRNAs targeting exon 2 (sgRNA site 1 5’-TTTCAGGGCCCGTCGCCCAT AGG; sgRNA site 2 5’-GCAGAGAGGTACCCCGGCAC TGG) and intron 9 (sgRNA site 3 5’-GTAGTTTGTCCACTCAGCTC AGG, sgRNA site 4 5’-CAGTAGTGGCCAAAGATTCC TGG). Mutant mice genotyping was performed with PCR amplification of genomic DNA extracted from tails⁸⁶. PCR was performed using Taq 2X Master Mix (New England Biolabs) and primers targeting the deleted gene body (Slc27a3-inner-F 5’-ATGACAGGGGAGCCTATTCG; Slc27a3-inner-R 5’-TCACTTGCCAGAACCCCTAG), bridging the deletion site in exon 2 (Slc27a3-bridge-F 5’-TTTGGGGAGGGATGTGCTAG; Slc27a3-bridge-R 5’-GCACAGTCAGGAAAAGGGTC), and spanning the entire deletion from before exon 2 to after exon 10 (Slc27a3-bridge-F; Slc27a3-span-R 5’-CCCTCTGACTGCCTTTCGTA). Scd1−/− mice were generated as above using sgRNAs targeting exon 2 (sgRNA site 1 5’-TATTCACGACCCCACCTATC AGG; sgRNA site 2 5’-GGTGGTGGTCGTGTAAGAAC TGG) and exon 5 (sgRNA site 3 5’-GCAAGAAGGTGCTAACGAAC AGG, sgRNA site 4 5’-ACGACAAGAACATTCAATCC CGG). Mutant mice genotyping was performed as above using primers targeting the deleted gene body (Scd1-inner-F 5’-CCCTCACCCCAAAACCAAAG; Scd1-inner-R 5’-CGGCTCACTCAGATCATCCT) and spanning the entire deletion (Scd1-span-F 5’-GTCCCAACTTTCGTGCCTTT; Scd1-span-R 5’-TAACAGGTGCCAGAGAAGGG).

Bronchoalveolar lavage cellularity

Bronchoalveolar lavage (BAL) cell concentration was quantified using a Countess II (Thermo Fisher). For Slc27a3−/− animals, cells were attached to glass slides using a Cytospin, then stained for immune cell lineage using HEMA 3 (Fisher Scientific). Approximately 400 cells per animal were counted to determine the percentage of eosinophils, neutrophils, monocytes, and epithelial cells by an experimenter blinded to treatment. For Scd1−/− animals, cells were stained with LIVE/DEAD BLUE (Invitrogen) for 30min on ice, then stained for markers of interest for 20 min on ice: CD11b (BV510, Biolegend), CD4 (BV605, Biolegend), CD19 (BV650, Biolegend), Ly-6G (BV711, Biolegend), MHCII (BV785, Biolegend), Siglec-F (PE, Biolegend), CD8 (PE/Dazzle-594, Biolegend), CD3 (PE/Cy7, Biolegend), CD11c (APC, Biolegend), and CD45 (APC/Cy7, Biolegend). UltraComp eBeads Plus compensation beads (Invitrogen) stained with single antibodies were used as compensation controls. Flow cytometry was performed using a Cytek Aurora Spectral Flow Cytometer (Cytek Biosciences) at the University of Chicago Human Disease and Immune Discovery (HDID) Core Facility (RRID:SCR_022936). Immune cells were identified based on forward (FSC) and side (SSC) scatter properties, followed by singlet discrimination and exclusion of dead cells (LIVE/DEAD Fixable Blue, Thermo Fisher). Live, single cells were sequentially gated for CD45+ expression. Neutrophils were gated as the Ly6G+CD11b+ subset. Alveolar macrophages were gated as the CD11c+Siglec-F+ subset. Eosinophils were gated as the CD11c-Siglec-F+ subset. T cells were gated as the Siglec-F-CD3+CD19− subset. B cells were gated as the Siglec-F-CD3-CD19+ subset. Percentages were multiplied by total cell concentration to determine the number of each cell type per mL of BAL fluid. Statistical significance was determined using Welch’s t test between WT and KO animals, both treated with Alt.

Histology staining and image quantification

Lung remodeling was characterized using histological staining. Following BAL collection, the left lobe from each mouse was fixed in formalin overnight at 4°C, then submitted to the University of Chicago Human Tissue Resource Center for paraffin embedding, sectioning, and staining. Slides containing 5 μm serial sections were stained for hematoxylin & eosin (H&E), Periodic Acid Schiff (PAS), or Alcian Blue (AB). Slides were scanned at the University of Chicago Integrated Light Microscopy Core on a VS200 Slideview (Olympus) at 20X. For the following analyses, two airways of approximately similar size per section were randomly selected for quantification. Epithelial thickness was determined in FIJI by measuring first epithelial cell area (area of inner airway - area of the region encompassed by basement membrane), then dividing by basement membrane length. Mucin expression area was determined in FIJI by first measuring the area of mucin staining using Color Threshold (hue 180-240, saturation 130-255, brightness 0-230), then dividing by epithelial cell area. Collagen deposition was determined in FIJI by measuring the area of collagen staining using Color Threshold (hue 108-190, saturation 0-255, brightness 0-232) normalized to basement membrane length. For perivascular and periairway inflammation, three airways and three vessels of approximately similar size per section stained with H&E were randomly selected for quantification. Immune cells were selected using the wand tool in QuPath⁸⁷ and cell number was determined using the Cell detection tool (requested pixel size 0.5μm, background radius 8μm, median filter radius 0μm, sigma 1.5μm, area 10-400μm², threshold 0.1, max background intensity 2, cell expansion 5μm). The number of immune cells was divided by basement membrane length of the vessel or airway. All airway selection and image quantification was performed by an experimenter blinded to treatment. Statistical significance was determined using Welch’s t test between WT and KO animals, both treated with Alt.

Human bronchial brushing RNA-seq

RNA-seq data from healthy controls, mild/moderate asthmatics, and severe asthmatics were previously generated as part of the Immune Mechanisms in Severe Asthma (IMSA) cohort and available on GEO (GSE158752). Differential gene expression determined using edgeR, with a model including age, sex, steroid use, and batch. P-values were adjusted for multiple comparisons using Benjamini-Hochberg.

QUANTIFICATION AND STATISTICAL ANALYSIS

For all experiments, number of replicates/sample size and age and sex of animals (when relevant) can be found in figure legends, with further details provided in the associated Method Details subsection of the STAR Methods. Statistical details of DANDELION are provided in the Method Details subsection of the STAR Methods. The following statistical tests were performed using GraphPad Prism unless stated otherwise. Statistical significance was identified as p-value<0.05 unless stated otherwise. Significant differences in TEER were determined by one-way ANOVA with Dunnett’s multiple comparisons test. Significant differentially expressed genes by RNA-seq were determined by adjusted p-value≤0.05 and log2 fold-change ≥ 0.6 using DESeq2 or edgeR. Significantly depleted sgRNAs in the Th2 CRISPR screen were identified as having FDR<0.25 and log₂ fold change less than the non-targeting guide mean (−0.2044); significantly enriched sgRNAs were identified as having FDR<0.25 and log₂ fold change greater than 0.2044. Significant differences between free fatty acids or lipid species of interest were determined using Welch’s t-test; p-values were not corrected for multiple comparisons. Western blots were quantified via densitometry using FIJI and statistical significance was determined using Welch’s t-test. Significant differences between populations of bronchoalveolar lavage cell types were determined using Welch’s t test. Lung histology images were quantified using FIJI or QuPath and significant differences were determined using Welch’s t test.

Supplementary Material

2

Table S10. Guide sequences used for CRISPR gene knockdown, related to Figure 3 and STAR Methods.

3

Table S11. qPCR primer sequences, related to Figure 3 and STAR Methods.

4

Table S1. DANDELION blood trait results, related to Figure 1

5

Table S2. Gene ontology enrichment of platelet crit DPGs, related to Figure 1.

6

Table S3. Conditionally independent GWAS loci of PLTC in UKBiobank, related to Figure 1; last few columns are where the GWAS loci map in the network and master regulator status.

7

Table S4. All significant mediations for asthma, using DGN or eQTLGen, related to Figure 2; last column indicates if they are significant if we limit the exposure SNPs/genes to GWAS p value of 10⁻⁴.

8

Table S5. Gene ontology enrichment of asthma DPGs, related to Figure 2.

9

Table S6. Lung cell types with high DPG expression, related to Figure 2.

10

Table S7. sgRNA hits from pooled T cell CRISPR screen, related to Figure 3.

11

Table S8. Differentially expressed genes in SLC27A3−/− epithelial and T cells, related to Figure 4.

12

Table S9. Normalized peak area, lipidomics of SLC27A3−/− cell culture media and cells, related to Figure 4.

KEY RESOURCES TABLE

REAGENT or RESOURCE	SOURCE	IDENTIFIER
Antibodies
STAT3 mouse monoclonal antibody	Cell Signaling Technology	Cat#9139
Phospho-STAT3 (Tyr705) rabbit monoclonal antibody	Cell Signaling Technology	Cat#9145
Beta-actin mouse monoclonal antibody	Cell Signaling Technology	Cat#3700
HRP conjugated goat anti-mouse IgG + IgM (H+L)	Jackson ImmunoResearch	Cat#115036068
HRP conjugated goat anti-rabbit IgG (H+L)	Jackson ImmunoResearch	Cat#111035003
BV 510 anti-mouse/human CD11b	Biolegend	Cat#101245
BV 605 anti-mouse CD4	Biolegend	Cat#100547
BV 650 anti-mouse CD19	Biolegend	Cat#115541
BV 711 anti-mouse Ly-6G	Biolegend	Cat#127643
BV 785 anti-mouse IA-I-E (MHCII)	Biolegend	Cat#107645
PE/Dazzle 594 anti-mouse CD8a	Biolegend	Cat#100761
PE-Cy7 anti-mouse CD3ε	Biolegend	Cat#100319
APC anti-mouse CD11c	Biolegend	Cat#117309
APC-Cy7 anti-mouse CD45	Biolegend	Cat#103115
Siglec-F Rat anti-Mouse, PE	BD Biosciences	Cat#562068
Chemicals, peptides, and recombinant proteins
Alt-R™ S.p. Cas9 Nuclease	IDT	Cat#1081059
Phorbol 12-myristate 13-acetate (PMA)	Millipore-Sigma	Cat#P8139-1MG
Ionomycin	Millipore-Sigma	Cat#I9657-1MG
Monensin Solution (1000X)	Biolegend	Cat#420701
Brefeldin A Solution (1000X)	Biolegend	Cat#420601
TAMRA-azide	Lumiprobe	Cat#47130
Tris[(1-benzyl-1H-1,2,3-triazol-4-yl) methyl]amine (TBTA)	TCI Chemicals	Cat#T2993
Tris(2-carboxyethyl)p hosphine HCl (TCEP-HCl)	Millipore-Sigma	Cat#580560
Coomassie Brilliant Blue	Millipore-Sigma	Cat#B7920
N-ethylmaleimide (NEM)	Millipore-Sigma	Cat#04259
2,3-dimethyl 1,3-butadiene	Millipore-Sigma	Cat#145491
Hydroxylamine	Millipore-Sigma	Cat#467804
Biotin-HPDP	Cayman	Cat#16459
Streptavidin beads	Thermo-Fisher	Cat#20361
cOmplete™ Min i, EDTA-free Protease Inhibitor Cocktail	Millipore-Sigma	Cat#11836170001
PhosSTOP	Millipore-Sigma	Cat#4906845001
Quick Start™ Bradford 1x Dye Reagent	Bio-Rad	Cat#5000205
StartingBlock™ (TBS) Blocking Buffer	Thermo Scientific	Cat#37542
SuperSignal West Pico	Thermo Scientific	Cat#34580
Chemiluminesc ent Substrate
UltraComp eBeads Plus	Invitrogen	Cat#01333341
Live/Dead Fixable Blue	Fisher Scientific	Cat#501121524
Alternaria alternata extract	Stallergenes Greer	Cat#XPM1D3A25
Taq 2X Master Mix	New England Biolabs	Cat#M0270L
NEBNext Q5 Hot Start HiFi PCR Master Mix	New England Biolabs	Cat#M0543S
QIAquick PCR Purification Kit	Qiagen	Cat#28106
Critical commercial assays
Pierce BCA Protein Assay	Thermo Scientific	Cat#23225
Fisher Healthcare PROTOCOL Hema 3 Manual Staining System and Stat Pack	Fisher Scientific	Cat#23123869
Deposited data
trans-eQTLs, eQTLGen	Trans-eQTLs from eQTLGen	https://www.eqtlgen.org/trans-eqtls.html
trans-eQTLs, Depression Genes and Network Study	Trans-eQTLs from Depression Genes and Network (DGN)	https://www.dropbox.com/scl/fo/b1r27fxqlq84ywory8qmo/AKZvSTY5-FfVZAweoiCp8fQ/GBAT/trans_pval?rlkey=n3pzgkdd0fqz9waamidlyfcfy&e=1&dl=0
Burden test summary statistics, asthma	UK Biobank: https://www.ebi.ac.uk/gwas/	Accession ID: GCST90085447
Burden test summary statistics, blood traits	UK Biobank: https://www.ebi.ac.uk/gwas/	Accession IDs in Table S1
Single-cell RNA seq, Human Lung Cell Atlas	CELLxGENE Census	6f6d381a-7701-4781-935c-db10d30de293
Perturb-seq, human CD4+ T cells	Zhu et al. 2025	https://virtualcellmodels.cziscience.com/dataset/genome-scale-tcell-perturb-seq
RNA-seq, bronchial brushings from Immune Mechanisms of Severe Asthma (IMSA) cohort	NCBI GEO	GSE158752
RNA-seq, SLC27A3−/− human CD4+ Th2 cells	NCBI GEO	GSE329629
RNA-seq, SLC27A3−/− 16HBE cells	NCBI GEO	GSE329629
Experimental models: Cell lines
Human: 16HBE14o-bronchial epithelial cell line	Sigma-Aldrich	Cat#SCC150
Human: Peripheral blood mononuclear cells	STEMCELL	Cat#2308401001; Cat#2304806103; Cat#231370501C
Human: HEK293FT embryonic kidney cell line	Xiaochang Zhang	RRID:CVCL_6911
Experimental models: Organisms/strains
Mouse: WT: C57BL/6J-WT	Charles River Laboratory	Cat#027
Mouse: Slc27a3−/−: C57BL/6J-Slc27a3−/−	This study	Slc27a3−/−
Mouse: Scd1−/−: C57BL/6J-Scd1−/−	This study	Scd1−/−
Oligonucleotides
Alt-R™ CRISPR-Cas9 tracrRNA	IDT	Cat#1072533
Alt-R™ CRISPR-Cas9 crRNA	IDT	See Table S10 for sequences
qPCR primers	IDT	See Table S11 for sequences
Recombinant DNA
lentiCRISPRv2 GFP plasmid	Addgene	Cat#82416
pMD2.G plasmid	Addgene	Cat#12259
psPAX2 plasmid	Addgene	Cat#12260
Software and algorithms
DANDELION	This study	https://github.com/mxxptian/DANDELION
FIJI	ImageJ	https://imagej.net/software/fiji/
Prism	GraphPad	Version 10.5.0
ARCHIE	Dutta et al. 2022	https://github.com/diptavo/ARCHIE
DACT	Liu et al. 2022	https://github.com/zhonghualiu/DACT
GBAT	Liu et al. 2020	https://github.com/xuanyao/GBAT
SCEPTRE	Barry et al. 2021	https://github.com/Katsevich-Lab/sceptre
AUCell	Aibar et al. 2016	https://github.com/aertslab/AUCell
GUIDES	Joung et al. 2017	https://github.com/sanjanalab/GUIDES
Salmon	Patro et al. 2017	version 1.5.1
DESeq2	Love et al. 2014	version 1.32.0
MAGECK	Li et al. 2014	https://sourceforge.net/projects/mageck
FlowJo	FlowJo	version 10.10
STAR	Dobin et al. 2013	version 2.2.7a
HTSeq	Anders et al. 2015	version 0.13.5
edgeR	Robinson et al. 2010	version 3.14.0
LipidSearch	Thermo Fisher	version 5.0
Compound Discoverer	Thermo Fisher	version 3.3
Skyline45, version 24.1	Pino et al. 2017	https://skyline.ms/project/home/software/Skyline/begin.view
QuPath	Bankhead et al. 2017	https://qupath.github.io/

Highlights.

• Integrating trans-regulatory signals and burden tests prioritizes disease-driving genes

• Prioritized genes are enriched for Mendelian genes missed by other genomics approaches

• CRISPR screens validate prioritized genes as regulators of key asthma cellular phenotypes

• Loss of palmitoylation enzymes SLC27A3 and SCD affects lung inflammation in vivo