The steps designed for genomic prioritization and evaluation were briefly described in Figure 1 and detailed below.

The dnet package (46) was used to conduct target set enrichment analysis quantifying the degree to which a predefined gene list is enriched at the leading prioritization. The leading prioritization, visually defined as the left-most region of the peak in running enrichment plot, is the core subset of the prioritized target genes accounting for the enrichment signal. The analysis was applied to clinical proof-of-concept targets in endometriosis, showing that they tended to be highly ranked. This analysis was also applied to: (i) cell-type-specific gene signatures (47) for exploring the cellular basis of all prioritized target genes in endometriosis; (ii) MSigDB hallmark gene sets (48) for revealing molecular hallmarks based on the top 10% prioritized target genes in endometriosis; and (iii) one fibroblast gene expression signature specific to ectopic endometrium (49) and four gene expression signatures involving two endometriosis stages (I/II and III/IV) at two menstrual cycle phases (proliferative and secretory) (50) for examining the highly prioritized target gene expression.

There were three steps to achieve this. First, the XGR package (51) and a collection of KEGG organismal system pathways (52) were used to identify enriched pathways based on the top 1% prioritized genes. Second, pathway crosstalk was identified by searching for a subset of gene interactions (extracted from enriched pathways identified in the previous step); identified so in a manner that the resulting pathway crosstalk contained highly prioritized and interconnected genes (53). Third, attack analysis was conducted to select optimal targeting combinations that maximized the effect of removing nodes on the crosstalk; done so in a manner to maximize the effect of removing either single nodes or specific node combinations. The combinatorial node removal can be 2-node combinations removed in the context of a fixed node ‘ATK1’, and 3-node combinations removed in the context of two fixed nodes ‘AKT1 + ESR1’.

Focusing on the top 5% prioritized target genes (n = 689) in endometriosis, the supraHex package (54) was used to construct cross-disease prioritization map. In brief, a supra-hexagonal map, consisting of 91 hexagons, was trained by the prioritization matrix (containing priority ratings) of 689 target genes versus 7 diseases including endometriosis (this study) and 6 immune-mediated diseases (53). These immune diseases include: (i) inflammatory bowel diseases, subdivided into Crohn’s disease (CRO) and ulcerative colitis (UC); and (ii) inflammatory systemic diseases, including multiple sclerosis (MS), rheumatoid arthritis (RA), Sjögren’s syndrome (SJO), and systemic lupus erythematosus (SLE). The trained map was used to illustrate a gene prioritization profile per disease, and together with a consensus neighbour-joining tree built from the prioritization matrix, to further illustrate inter-disease relationships. The trained map was also divided into target gene clusters in a topology-preserving manner. Enrichment analysis for genes within a cluster was based on one-sided Fisher’s exact test to identify enrichments in terms of: (i) approved drug targets obtained from the ChEMBL database (44); (ii) immune system pathways from the Reactome database (55); (iii) Gene Ontology functional annotations from NCBI (56); and (iv) mouse phenotype annotations using Mammalian Phenotype Ontology (57).

The fpocket software (58) was used to predict druggable pockets of a target gene based on its known protein structure(s). The known protein structures were obtained from the PDB database (59), for example, the access code ‘4N78’ for the WAVE regulatory complex (60). A gene was defined to be tractable if predicted to have drug-like binding sites (that is, druggable pockets). The PDB structure was viewed in 3D as cartoon (secondary structure abstraction), colored by PDB chains and embedded with druggable pockets in blue (61).

We performed genomics-led prioritization, taking GWAS summary statistics in endometriosis as inputs and leveraging the informativeness of regulatory genomics in diverse cell types, activation states and tissues. As outlined in Figure 1 (see Materials and Methods for details), our multi-step prioritization process consisted of: (i) the preparation of genomic predictors; (ii) the evaluation of predictor importance to identify informative predictors; (iii) the assessment of how to combine informative predictors; and (iv) performance evaluation to benchmark competing approaches. We found that using the meta-analysis-like combination strategy, particularly based on order statistic to combine genomic predictors, achieved much better performance than using the direct combination strategy ( Figure 2A ). Benchmarking results showed that our prioritization approach (called ‘END’) was superior to competing approaches ( Figure 2B ). We considered two competing approaches, namely, Open Targets (also using genetics and genomics for target identification and prioritization) (45) and Naïve prioritization (prioritizing a gene by how often it has been targeted by existing drugs), that were similar in performance. Notably, Naïve prioritization, based on the concept of drug repurposing, was limited in being unable to predict new targets. The precision-recall analysis showed that our prioritization achieved the 91% precision at the 56% recall (prioritization coverage; Figure 2C ).

We prioritized a total of >14,000 candidate target genes [with the knowledge of gene interactions sourced from the STRING database (42)], ranked by priority rating ( Table S1 ). The top prioritized targets included genes essential in inflammation and cellular responses, such as JUN (ranked 1^st), NFKB1 (12^th), RUNX1 (17^th), PTEN (21^st), MAPK14 (24^th), FOS (20^th), SMAD3 (22^nd), and RELA (25^th) ( Figure 3A ). Our prioritization recovered existing clinical proof-of-concept target genes in endometriosis ( Figures 3B, C ). These proof-of-concept targets in endometriosis were all at the top 1% prioritized gene list, including: MAPK8 (9^th), MAPK10 (46^th), and MAPK9 (47^th) targeted by bentamapimod, a JNK inhibitor; ESR1 (19^th) targeted by estradiol, an estrogen receptor (ER) alpha agonist; ESR2 (99^th) targeted by prinaberel, an ER beta agonist; AR (26^th) targeted by danazol, an androgen receptor (AR) agonist, and cyproterone acetate, an AR antagonist; NR3C1 (53^th) targeted by cyproterone acetate and mifepristone, two glucocorticoid receptor antagonists; NGF (91^th) targeted by tanezumab, a beta-nerve growth factor inhibitor; PGR (116^th) targeted by selective progesterone receptor modulators (SPRMs); and TNF (136^th) targeted by infliximab, a TNF-alpha inhibitor. According to the latest ESHIR guideline (62), SPRMs such as medroxyprogesterone acetate and levonorgestrel are highly recommended to reduce endometriosis-associated pain, and norethindrone acetate (alongside GnRH) highly recommended to prevent bone loss and hypoestrogenic symptom, but no longer recommended for danazol. Figure 3C also provides a summary of the predictors used and their relative importance (informativeness). We treated the nearby gene predictor as the baseline defining predictive informativeness. We identified informative regulatory genomic predictors that were mostly derived from immune blood cells (with a few from other cell types and tissues). These results necessitated the use of genomic datasets from diverse contexts for target prioritization and identification, which is consistent with endometriosis that is increasingly recognized as a systemic disease (4).

We used rank-based target set enrichment analysis to characterize prioritized target genes using cell-type-specific gene signatures (47). We found the enrichment for immune cells; these included cells in myeloid lineages (such as monocytes, neutrophils, and mast cells) and lymphoid lineages (such as gamma delta T-cells, CD4 + memory T-cells, and CD8+ effector-memory T-cells) but not in B-cell lineages ( Figure 4 and Table S2 ). In addition to immune cells (for example, gamma delta T-cells; Figure S1A ), we also found the enrichment for stromal cells ( Figure S1B ) and epithelial cells ( Figure S1C ). Taken together, our findings are consistent with evidence of altered immunity and inflammation (local and systemic) in endometriosis (3), and also reveal the cellular basis that might involve multiple lineages underlying the pathogenesis of endometriosis.

Next, using MSigDB hallmark gene sets (48), we identified 8 molecular hallmarks enriched in highly prioritized genes ( Figure 5A and Table S3 ). These included: cellular responses to low oxygen levels (hypoxia), ultraviolet radiation and estrogen; signaling pathways (NF-kB signaling in response to TNF, and activation of the PI3K/AKT/mTOR pathway); immune processes (allograft/transplant rejection, and the complement system); and the apical junction complex. For each of these hallmarks, genes found at the leading prioritization are shown in Figure 5B . We identified 21 leading prioritized genes that are responsive to hypoxia, with 4 genes (HSPA5, IL6, PDGFB, and SERPINE1) encoding components of the complement system and 7 genes (CDKN1A, DUSP1, FOS, IL6, IRS2, JUN, and VEGFA) regulated by NF-kB in response to TNF ( Figure 5C ). We identified 26 leading prioritized genes involved in the PI3K/AKT/mTOR pathway ( Figure 5B ), the pathway relevant to the severity of endometriosis stages (63). Among these 26 genes, 7 (AKT1, EGFR, IL2RG, IL4, LCK, MAP3K7, and PRKCB) are also involved in allograft rejection ( Figure 5C ), the consequences of which may cause tissue injury and fibrosis. These results revealed diverse but related molecular hallmarks in endometriosis, indictive of being an inflammatory systemic disease.

Using the KEGG resource (52), we proceeded to perform pathway enrichment analysis for the top 1% prioritized target genes. This identified a wide range of organismal system pathways that can be broadly categorized into immune, endocrine, nervous, and reproductive systems ( Figure 6A ). It is well-recognized that maladaptation of these organismal systems can occur in chronic inflammatory systemic diseases (64). Integrated analysis of these enriched pathways identified a 31-gene network (P = 3.1 × 10⁻²¹³ on permutation test), reflective of crosstalk between pathways, that all contained highly prioritized genes in endometriosis ( Figure 6B and Table S4 ). Using the information on approved therapeutics available in the ChEMBL database (44), we found 5 genes that are already targeted by licensed medications (approved drugs) in other diseases ( Figure 6B ).

Removing/attacking a node critical for the crosstalk would result in large numbers of disconnected nodes. We found that the crosstalk identified above was very robust to node removal. The maximal effect achieved by any individual node removal was as low as 10% network disconnection (that is, removing AKT1 gene; Figure 6C ). The identification of AKT1, or more generally, genes in the PI3K/AKT/mTOR pathway (including MTOR, PIK3CA, and PIK3R1 in addition to ATK1; see also Figure 5 ), is in line with current interests targeting this pathway in endometriosis (65). This motivated us to further perform combinatorial attack analysis in the context of AKT1. The 2-node maximal effect was observed when removing another approved drug target, ESR1, EGFR, JAK1, PIK3CA or SRC ( Figure 6C ). Therapeutic agents targeting ESR1 are now under phase 3 clinical trials in endometriosis (66), and thus, we further performed attack analysis in the context of AKT1 + ESR1, with maximal network disconnection (~16%) achieved by 3-node removal ( Figure 6C ). These findings supported the usefulness of our identified crosstalk genes for drug repurposing ( Figure 7 ). For example, estrogen receptor modulators/agonists targeting ESR1 are now used in the clinic; this includes bazedoxifene acetate and conjugated estrogens for preventing postmenopausal osteoporosis (related to oestrogen deficiency) (67).

A systematic review has confirmed that endometriosis has statistically significant associations with immune-mediated diseases (CRO, MS, RA, SJO, SLE, and UC) (8). Hereinafter, we explored such relationships based on the top 5% prioritized genes in endometriosis. We indeed observed significant correlations of priority ratings between endometriosis and immune diseases ( Figure 8A ). To further capture such correlations, we next employed a supra-hexagonal map (54) for cross-disease comparisons ( Figure 8B , Figure S2 and Table S5 ). Each presentation illustrates a disease-specific prioritization profile (in which hexagons are color-coded to show priority ratings for genes thereof), while consensus neighbor-joining tree captures the (dis)similarity of prioritization profiles between diseases. The observed relationships are consistent with therapeutic/phenotypic similarity. For example, inflammatory bowel diseases (CRO and UC) are grouped together, and inflammatory systemic diseases (SJO, SLE, and RA) stay closer.

To systematically identify target genes that are shared between diseases and are unique to specific diseases, we grouped nearby hexagons and identified 6 target clusters (C1-C6), each containing genes with similar prioritization patterns ( Figure 8C and Table S5 ). Among these, C6 was highly rated in all diseases analyzed, displaying the highest percentage of approved therapeutics ( Figure 8D ). Notably, C5 was highly rated in endometriosis only and contained a relatively low proportion of approved drug genes ( Figure 8D ), indicative of being under-explored. In agreement with this, we found the highest degree of support from clinical evidence (approved drugs) for genes in C6, and no enrichment was observed for genes in C5 ( Figure 8E ).

Next, we characterized C5 and C6 using Reactome pathways (55). We found the enrichment for genes in C6 that are engaged in all major immune system pathways, except for neutrophil degranulation that was specific to C5 ( Figure 8F ). To confirm this, we performed functional enrichment analysis using Gene Ontology (56). We found that genes in C5 are mostly functionally relevant to neutrophil degranulation and are localized in the secretory granule lumen ( Figure S3A ). Using Mammalian Phenotype Ontology (57), we also observed phenotypic enrichments for genes in C5; these genes, when knocked out, tended to cause lethality phenotypes ( Figure S3B ).

Genes in C6 were highly prioritized across diseases ( Figure 8 ). Based on these genes, we explored repurposing opportunities via a heatmap-like illustration ( Figure 9 ). In addition to priority ratings, this illustration collectively showed how these genes are particulate in immune system pathways, including cytokine signaling (interferon signaling and interleukin signaling), innate immunity (toll-like receptor and Fc epsilon receptor 1) and adaptive immunity (T-cell receptor). We identified 20 approved drug targets that are mostly immune-related. These genes provide opportunities for licensed immunomodulatory drugs that could be repurposed for the potential use in endometriosis, particularly disease-modifying anti-rheumatic drugs (DMARDs). They include biological DMARDs (such as inhibitors of TNF, IL6, and IL6R) and targeted synthetic DMARDs (such as kinase inhibitors targeting JAK1/2/3, SRC, LYN, SYK, and LCK). Of particular interest, TNF is a well-established therapeutic target for chronic inflammatory diseases (68). We suggest that drugs targeting this gene should be studied in cohort-scale clinical trials for repurposing in endometriosis, such as infliximab (known as anti-TNF biologics).

As shown in the heatmap visualization in Figure 10A , genes in C5 were specifically prioritized in endometriosis and characterized by neutrophil degranulation. Neutrophils are highly versatile and plastic. Plastic neutrophils are rendered through degranulation, in which cytoplasmic granules are mobilized and fused with the plasma membrane (69). Accumulating evidence supports the tumour metastatic spread model in which neutrophils mediate the migration of cancer cells to distant sites (70). Interestingly, we found that neutrophil degranulation genes in C5 are almost all linked to cancer; for example, those involved in breast cancer are the genes ATG7 (71), CAP1 (72), CCT8 (73), CD14 (74), CYFIP1 (75), and QSOX1 (76). Notably, neutrophil degranulation genes have been reported to be associated with endometriosis, such as A1BG (a diagnostic marker for stage II, III and IV endometriosis) (77) and ATG7 (an autophagy gene in ovarian endometriosis) (78). Finally, we explored the tractable evidence; a tractable gene was defined if its known protein structures were predicted to contain druggable pockets ( Figure 10B ). We suggest that the tractable genes involved in neutrophil degranulation are of particular interest to inform future clinical studies.