A cross-disease, pleiotropy-driven approach for therapeutic target prioritization and evaluation.

In this study, we aimed to determine whether the principle and two recipes outlined earlier are also applicable to cross-disease pleiotropic association summary data. Specifically, we sought to address the “ W-H-W ” challenge: (1) whether pleiotropic associations are useful for therapeutic target discovery, (2) how to implement and evaluate target identification based on pleiotropic evidence, and (3) in which specific diseases pleiotropy can inform clinical therapeutics. To tackle this challenge, we developed a pleiotropy-driven approach, “PIPE,” with further details provided in the subsequent subsection. As indicated by its name, PIPE features two components — prioritization and evaluation. It excels in ab initio resolving pleiotropic associations derived from cross-disease genetic findings, facilitating therapeutic target prioritization and evaluation, and generating accessible online resources ( Figure 1 A). To achieve this, we devised three prioritization steps ( Figure 1 B) and three evaluation analyses ( Figure 1 C). Using pleiotropic associations as inputs, PIPE systematically runs through three sequential steps to output therapeutic targets at both the gene and pathway crosstalk levels ( Figure 1 B; STAR Methods ). A distinctive feature of PIPE is its capacity to establish the link from pleiotropic loci, including those located within the non-coding genome, to candidate target genes. This is made possible by harnessing the value of multimodal regulatory genomic datasets (genomic evidence) and leveraging knowledge of protein interactions (network evidence). Accordingly, at Step 1, right from pleiotropic associations, we identified pleiotropic genes supported by genomic evidence, termed “pleiotropic (genomic) genes.” These encompassed three types (or predictors): nearby genes ( nGenes ) with evidence of genomic proximity, conformation genes ( cGenes ) identified using promoter-capture Hi-C datasets, 32 , 33 and expression-associated genes ( eGenes ) identified using eQTL datasets. 34 For each type (or per dataset) of pleiotropic genes, we further identified peripheral genes supported by network evidence, termed “peripheral (networked) genes,” identified so through utilizing high-quality protein interaction data. 35 It is noteworthy that only one nGene predictor was prepared, while the number of cGene predictors (or eGene predictors) equaled the number of datasets used (i.e., one predictor per dataset). We then assessed the relative importance of predictors, retaining subsets of cGene and eGene predictors that were at least as important as the nGene predictor (conventionally used as the baseline predictor). This process enabled the extraction of informative predictors, subsequently forming a gene-predictor matrix. Similar to Fisher’s meta-analysis combining predictors in the gene-predictor matrix, at Step 2, we generated a comprehensive list of prioritized target genes, each assigned a credit score on a 10 point scale. Moving to Step 3, from this compilation of all prioritized target genes, we identified a concise and manageable list of genes likely mediating the crosstalk between molecular pathways, representing targets at the pathway crosstalk level. To evaluate the performance of PIPE in target gene prioritization, we designed three instrumental analyses ( Figure 1 C; STAR Methods ). Firstly, we outlined the supportive procedures for defining clinical proof-of-concept targets (phase 2 or higher) for a specific disease and simulating negative controls that are unlikely targets for this disease. Secondly, we benchmarked PIPE against the state-of-the-art prioritization approach, namely Open Targets 36 , 37 —the most relevant approach that also utilizes human genetics and genomics for target identification and validation. This benchmarking was measured by the area under the ROC curve (AUC), assessing the ability to distinguish clinical proof-of-concept targets from simulated negative targets. Thirdly, we introduced a new metric, called “pleiotropy informing clinical therapeutics (PICT),” to quantify the tendency of prioritized genes to be clinical proof-of-concept targets via leading prioritization analysis (LPA). First, we evaluated the effectiveness of PIPE in predicting therapeutic targets from cross-disease pleiotropic associations, focusing on neuropsychiatric disorders that encompassed eight specific diseases. 18 As illustrated in Figures 2 A and S1 A, our benchmarking strongly supported the validity of PIPE in predicting therapeutic targets, especially when compared to Open Targets, 37 particularly its Genetics Portal. 36 Regarding Open Targets, prioritizations based on individual evidence were used for evaluation, namely, genetic association and text mining, both obtained from the Open Targets Platform. 37 Despite being purely pleiotropy-driven, PIPE outperformed Open Targets (genetic association) in all diseases analyzed and demonstrated competitiveness with Open Targets (text mining) for most diseases. Subsequently, we performed LPA to quantify the PICT potential for each specific disease ( Figure 2 B; Table S1 ). LPA assesses the extent to which clinical proof-of-concept targets are enriched at the “leading prioritization” of the entire prioritized gene list and reports three pieces of information on the enrichment results: normalized enrichment score, enrichment significance (false discovery rate), and the fraction of clinical proof-of-concept targets recovered at the leading prioritization (coverage). All three results were factored into the calculation of PICT. This enabled the identification of specific diseases in which pleiotropy can inform clinical therapeutics. The highest PICT was observed for attention-deficit hyperactivity disorder, with its leading prioritization plot presented in Figure 2 B (top right), where clinical proof-of-concept targets are explicitly labeled. Figure 2 Performance evaluation of PIPE in recovering therapeutic targets for neuropsychiatric disorders and its utility in determining target-level disease relationships (A) Heatmap-like performance comparisons between PIPE and Open Targets. The performance is measured by the area under the ROC curve (AUC), which separates clinical proof-of-concept targets from simulated negative targets. (B) LPA for eight neuropsychiatric disorders. Individual diseases are arranged in clockwise order using the pie chart sized by PICT. Each leading prioritization plot presents the normalized enrichment score (NES), enrichment significance (false discovery rate [FDR]), and the fraction of clinical proof-of-concept targets found at the leading prioritization (coverage). All three variables contribute to the calculation of disease-specific PICT. The leading prioritization is defined as the leftmost region ahead of the peak (the dark bar). Inserted at the top right is the leading prioritization plot for attention-deficit hyperactivity disorder (ADHD), where recovered clinical proof-of-concept targets are indicated by vertical lines, labeled by gene symbols, and color coded by the credit score. (C) Network-like representation of diseases and member genes (i.e., clinical proof-of-concept targets recovered at the leading prioritization). Member genes shared between any two diseases are also labeled, including target ranks (in the parentheses). The network layout was rendered using the stress majorization algorithm. (D) Inter-disease relationships inferred by the sharing of member genes. Nodes are sized by the number of member genes, and edges represent relationships (with the thickness proportional to the number of genes shared between two endpoint diseases). See also Figure S1 and Table S1 . Performance evaluation of PIPE in recovering therapeutic targets for neuropsychiatric disorders and its utility in determining target-level disease relationships (A) Heatmap-like performance comparisons between PIPE and Open Targets. The performance is measured by the area under the ROC curve (AUC), which separates clinical proof-of-concept targets from simulated negative targets. (B) LPA for eight neuropsychiatric disorders. Individual diseases are arranged in clockwise order using the pie chart sized by PICT. Each leading prioritization plot presents the normalized enrichment score (NES), enrichment significance (false discovery rate [FDR]), and the fraction of clinical proof-of-concept targets found at the leading prioritization (coverage). All three variables contribute to the calculation of disease-specific PICT. The leading prioritization is defined as the leftmost region ahead of the peak (the dark bar). Inserted at the top right is the leading prioritization plot for attention-deficit hyperactivity disorder (ADHD), where recovered clinical proof-of-concept targets are indicated by vertical lines, labeled by gene symbols, and color coded by the credit score. (C) Network-like representation of diseases and member genes (i.e., clinical proof-of-concept targets recovered at the leading prioritization). Member genes shared between any two diseases are also labeled, including target ranks (in the parentheses). The network layout was rendered using the stress majorization algorithm. (D) Inter-disease relationships inferred by the sharing of member genes. Nodes are sized by the number of member genes, and edges represent relationships (with the thickness proportional to the number of genes shared between two endpoint diseases). See also Figure S1 and Table S1 . Next, we demonstrated the utility of PIPE in elucidating inter-disease relationships within neuropsychiatric disorders ( Figures 2 C and 2D) based on the shared prioritization of therapeutic targets. For each disease, we defined its member genes as clinical proof-of-concept targets at the leading prioritization. The relationship between any two diseases was inferred by the extent of sharing their member genes ( Figure 2 C). As shown in Figure 2 D, based on the degree of shared prioritized target genes, the strongest relationships were observed among three specific neuropsychiatric diseases that impact the mind—bipolar disorder, schizophrenia, and unipolar depression. To visually represent the inferred relationships, we provided a side-by-side summary of how PIPE recovered clinical proof-of-concept targets for each disease within the neuropsychiatric disorder system ( Figure 3 ; Table S2 ). Notably, approved drug targets are highlighted in bold in Figures 3 B and 3C, suggesting potential opportunities for drug repurposing within neuropsychiatric disorders. Figure 3 Illustration of clinical proof-of-concept therapeutic targets recovered at the leading prioritization for each individual disease within neuropsychiatric disorders (A) Heatmap for clinical proof-of-concept therapeutic targets. The credit score is color coded, with the number in the parentheses indicating the target rank. (B) Annotations to clinical proof-of-concept targets per disease. (C) Annotations for approved drug targets, disease indications, and approved drugs. See also Figure S1 and Table S2 . Illustration of clinical proof-of-concept therapeutic targets recovered at the leading prioritization for each individual disease within neuropsychiatric disorders (A) Heatmap for clinical proof-of-concept therapeutic targets. The credit score is color coded, with the number in the parentheses indicating the target rank. (B) Annotations to clinical proof-of-concept targets per disease. (C) Annotations for approved drug targets, disease indications, and approved drugs. See also Figure S1 and Table S2 . Moreover, we substantiated the validity and utility of PIPE by analyzing cross-disease pleiotropic associations for inflammatory disorders encompassing five specific diseases. Our results demonstrated the superior performance of PIPE across all specific diseases analyzed ( Figures 4 A and S1 B). Notably, the highest PICT was seen for psoriasis and inflammatory bowel disease ( Figure 4 B; Table S3 ). Inflammatory bowel disease can be subcategorized into Crohn’s disease (CRO) and ulcerative colitis (UC; Figure 4 C). Consistent with clinical distinctions, we observed a close relationship between CRO and UC, which exhibited the strongest inferred connection based on the number of shared member genes ( Figures 4 D, 4E, and S2 ). Comparing PIPE to our previous Pi approach, 38 , 39 which concentrated on single-disease GWASs, highlighted the superior performance of PIPE ( Figure 4 F). This underscores the added value of pleiotropy studies, particularly employing pleiotropy approaches like PIPE, in enhancing target prioritization. To elucidate the impact of different types of PIPE predictors on prioritization, we also assessed their relevant importance, with cGene emerging as the most important predictor, followed by eGene and nGene ( Figure 4 G). This finding aligns with our previous findings from the Pi approach. 38 , 39 Figure 4 Performance evaluation, PICT, and inter-disease relationships for inflammatory disorders (A) Performance comparison between PIPE and Open Targets. AUC plots compare the performances that separate clinical proof-of-concept targets from simulated negative controls. (B) Illustration of PICT for each individual disease within inflammatory disorders arranged in a clockwise order using the pie chart. Leading prioritization plot for each disease presents the NES, enrichment significance (FDR), and the fraction of clinical proof-of-concept targets found at the leading prioritization (coverage). (C) Illustration of the leading prioritization for ulcerative colitis (UC). Clinical proof-of-concept targets are indicated by vertical lines, labeled with gene symbols, and color coded by credit scores. (D) Network-like representation of diseases and member genes. Member genes shared between any two diseases are labeled, including target ranks (in the parentheses). The network layout was rendered using the stress majorization algorithm. (E) Inferred inter-disease relationships based on shared member genes. Nodes are sized by the number of member genes, and the edge thickness represents the number of genes shared between two endpoint diseases. The layout is rendered using the Fruchterman-Reingold algorithm. (F) Bar plot comparing the performance of PIPE with a single-disease prioritization approach (our previous Pi approach). (G) Boxplot illustrating the relevant importance of predictors. Predictor importance is quantified by the decrease in accuracy when disabling/removing a predictor. Notably, removing an informative/important predictor would result in a significant decrease in accuracy. Different types of predictors are color coded. See also Figure S2 and Table S3 . Performance evaluation, PICT, and inter-disease relationships for inflammatory disorders (A) Performance comparison between PIPE and Open Targets. AUC plots compare the performances that separate clinical proof-of-concept targets from simulated negative controls. (B) Illustration of PICT for each individual disease within inflammatory disorders arranged in a clockwise order using the pie chart. Leading prioritization plot for each disease presents the NES, enrichment significance (FDR), and the fraction of clinical proof-of-concept targets found at the leading prioritization (coverage). (C) Illustration of the leading prioritization for ulcerative colitis (UC). Clinical proof-of-concept targets are indicated by vertical lines, labeled with gene symbols, and color coded by credit scores. (D) Network-like representation of diseases and member genes. Member genes shared between any two diseases are labeled, including target ranks (in the parentheses). The network layout was rendered using the stress majorization algorithm. (E) Inferred inter-disease relationships based on shared member genes. Nodes are sized by the number of member genes, and the edge thickness represents the number of genes shared between two endpoint diseases. The layout is rendered using the Fruchterman-Reingold algorithm. (F) Bar plot comparing the performance of PIPE with a single-disease prioritization approach (our previous Pi approach). (G) Boxplot illustrating the relevant importance of predictors. Predictor importance is quantified by the decrease in accuracy when disabling/removing a predictor. Notably, removing an informative/important predictor would result in a significant decrease in accuracy. Different types of predictors are color coded. See also Figure S2 and Table S3 . Taken together, our results strongly validate the utility of PIPE in addressing the " W-H-W" challenge, providing a robust framework for therapeutic target prioritization and evaluation for both neuropsychiatric and inflammatory disorders. The validity of PIPE also motivates further exploration into its advanced utilities in therapeutic discovery. We used Manhattan plots to illustrate a comprehensive list of target genes prioritized across the genome ( Figure S3 A). Additionally, we identified the most significant pathways wherein highly prioritized genes act coordinately. They included circadian entrainment and GABAergic/serotonergic synapses for neuropsychiatric disorders 40 , 41 and tumor necrosis factor (TNF), NOD-like receptor, and nuclear factor κB (NF-κB) signaling for inflammatory disorders 42 , 43 ( Figure S3 B). These significant pathways were consistently identified using molecular hallmarks as well ( Figure S3 C). Beyond individual pathways, PIPE is also unique in identifying a network of highly prioritized genes that mediate crosstalk between molecular pathways, generating a concise and manageable list of target genes at the pathway crosstalk level. To achieve this, we integrated the prioritization information with gene interactions (merged from KEGG pathways; illustrated in step 3 of Figure 1 B). As shown in Figure 5 (also see Table S4 ), we identified a 57-gene pathway crosstalk in neuropsychiatric disorders ( p  = 1.77 × 10 −60 on the permutation test) and a 59-gene pathway crosstalk in inflammatory disorders ( p  = 1.39 × 10 −181 on the permutation test). Figure 5 Identification of targets at the pathway crosstalk level for neuropsychiatric and inflammatory disorders Visualizations of pathway crosstalks are illustrated separately for neuropsychiatric disorders (top) and inflammatory disorders (bottom). The crosstalk was identified by integrating target credit score information with pathway-derived gene interactions (in orange; merged from KEGG pathways). Nodes are labeled by gene symbols (ranks), and if identified as genomic/pleiotropic genes, they are highlighted as pie segments (with genomic predictors inside). See also Figure S3 and Table S4 . Identification of targets at the pathway crosstalk level for neuropsychiatric and inflammatory disorders Visualizations of pathway crosstalks are illustrated separately for neuropsychiatric disorders (top) and inflammatory disorders (bottom). The crosstalk was identified by integrating target credit score information with pathway-derived gene interactions (in orange; merged from KEGG pathways). Nodes are labeled by gene symbols (ranks), and if identified as genomic/pleiotropic genes, they are highlighted as pie segments (with genomic predictors inside). See also Figure S3 and Table S4 . Next, we delved into unraveling the modular structure of the network by merging the identified instances of two pathway crosstalks. This exploration revealed five distinct modules (M1–M5) inherent in the network, identified so without presumptive prior knowledge. These modules were intricately associated with discernible molecular pathways ( Figures 6 A and S4 A; Table S5 ), signifying a functional modular design. Among them, two modules (M3 and M5) exhibited links with inflammatory pathways, including JAK-STAT signaling with M3, TNF, and NF-κB signaling with M5. Additionally, three modules were characterized by pathways relevant to neuropathology, including cAMP and cGMP-PKG signaling (M1), cAMP and calcium signaling (M2), and apelin and PI3K-ATK signaling (M4). Figure 6 Modular analysis of pathway crosstalks in neuropsychiatric and inflammatory disorders This analysis involved examining a network of gene interactions constructed by merging two instances of pathway crosstalks (as presented in Figure 5 ). Five modules (M1–M5) were identified through network modular analysis. Enrichment analysis was performed using a one-sided Fisher’s exact test to calculate the significance level (FDR), odds ratio, and 95% confidence interval. (A) Modular visualization of the gene network. Modules and member genes are color coded as indicated. (B) Disorder-specific illustration of the gene network. The same network layout as in (A) is presented, with nodes color coded according to credit scores in neuropsychiatric disorders (left) and in inflammatory disorders (right). (C) Forest plot of approved or non-approved phased drug target enrichments. Approved and non-approved phased drug targets were sourced from ChEMBL. (D) Forest plot of phylostrata enrichments. A phylostratum contains a group of genes that were first created at that phylostratum, defined by gene evolutionary age as dated via genomic phylostratigraphy (PSG). Plylostrata are ordered according to evolutionary history from the earliest (PSG01) to the most recent (PSG16). Genes first created at each enriched phylostratum are also listed. See also Figure S4 and Table S5 . Modular analysis of pathway crosstalks in neuropsychiatric and inflammatory disorders This analysis involved examining a network of gene interactions constructed by merging two instances of pathway crosstalks (as presented in Figure 5 ). Five modules (M1–M5) were identified through network modular analysis. Enrichment analysis was performed using a one-sided Fisher’s exact test to calculate the significance level (FDR), odds ratio, and 95% confidence interval. (A) Modular visualization of the gene network. Modules and member genes are color coded as indicated. (B) Disorder-specific illustration of the gene network. The same network layout as in (A) is presented, with nodes color coded according to credit scores in neuropsychiatric disorders (left) and in inflammatory disorders (right). (C) Forest plot of approved or non-approved phased drug target enrichments. Approved and non-approved phased drug targets were sourced from ChEMBL. (D) Forest plot of phylostrata enrichments. A phylostratum contains a group of genes that were first created at that phylostratum, defined by gene evolutionary age as dated via genomic phylostratigraphy (PSG). Plylostrata are ordered according to evolutionary history from the earliest (PSG01) to the most recent (PSG16). Genes first created at each enriched phylostratum are also listed. See also Figure S4 and Table S5 . When modules were color-coded by credit scores ( Figure 6 B), an intriguing pattern emerged. Genes in M3 and M5 generally exhibited higher credit scores in inflammatory disorders, indicating inflammatory-disorder-specific modules. In contrast, genes in M1, M2, and M4 showcased higher credit scores for neuropsychiatric disorders, signifying neuropsychiatric-disorder-specific modules ( Figure 6 B). Moreover, our analysis revealed a high degree of support from both preclinical evidence (non-approved phased drugs) and clinical evidence (approved drugs) for inflammatory-disorder-specific modules ( Figure 6 C). However, support from clinical evidence, albeit not from preclinical evidence, was notable for neuropsychiatric-disorder-specific modules ( Figure 6 C). Collectively, our investigation of pathways and modules for targets prioritized from pleiotropy GWASs unveiled a notable separation between neuroscience and immunological diseases. While acknowledging that GWAS loci tend to concentrate in specific genomic locations irrespective of the trait, we underscore that such concentration does not necessarily imply shared mechanisms. Our findings underscore the significance of further exploring the divergence or convergence of pathways/modules across different traits, adding a layer of complexity and relevance to the understanding of the molecular basis of complex diseases. We hypothesized that disorder-specific functional modules might bear an evolutionary origin. To test this hypothesis, we traced the evolutionary history of gene creation, denoted as “gene evolutionary age.” For a human gene, its evolutionary age is defined as the phylostratum in which it first emerged, and gene creation events were dated through genomic phylostratigraphy. 44 Our investigation revealed that genes within inflammatory disorder-specific modules tended to originate in Bilateria , especially in its descendant Euteleostomi (or “bony fish”), coinciding with the emergence of genes involved in NF-κB and JAK-STAT signalings ( Figures 6 D and S4 B). This co-emergence aligns with the findings from evolutionary analysis for immune system genes in teleost/bony fishes. 45 In contrast, we did not find evidence supporting a preferential gene creation pattern for neuropsychiatric disorder-specific modules. Finally, focusing on the genes implicated in two instances of pathway crosstalks, we employed a supra-hexagonal map 46 , 47 , 48 , 49 , 50 to construct a prioritization map ( Figures 7 A and S5 A). The trained map was color-coded separately for neuropsychiatric and inflammatory disorders ( Figures 7 B and S5 B). We grouped adjacent hexagons in a topology-preserving manner, leading to the identification of three target clusters (C1–C3) that exhibited both shared and distinct characteristics between the two disorder systems ( Figure 7 C; Table S6 ). Genes in cluster C1 demonstrated significantly higher credit scores in inflammatory disorders. Although genes in cluster C2 were slightly shared, they leaned toward stronger associations with inflammatory disorders. In contrast, genes in cluster C3 displayed markedly higher credit scores in neuropsychiatric disorders. Relative to the clusters specific to inflammatory disorders, the neuropsychiatric disorder-specific cluster C3 contained a notably lower proportion of structurally targetable genes ( Figure 7 C), indicating that these genes are under-studied. A gene was classified as targetable if its known protein structures 51 , 52 , 53 , 54 were predicted to contain druggable pockets. Despite genes in C3 being under-studied (i.e., lacking genes already targeted by non-approved or approved drugs), a closer inspection of the structural evidence enabled the identification of the targetable genes involved in calcium signaling ( CAMK2A , CAMK2D , and CALML3 ), which were also predicted to contain druggable pockets ( Figures 7 D and 7E). Considering the significance of calcium/calmodulin-dependent protein kinases in synaptic plasticity and neuropsychiatric disease, 55 , 56 we propose that these structurally targetable calcium signaling genes hold particular interest in informing disease-specific targeting potential in neuropsychiatric disorders. Figure 7 Crosstalk-based prioritization map for neuropsychiatric and inflammatory disorders (A) Illustration of the supra-hexagonal map. The map was used to analyze a matrix (containing credit scores) of 112 pathway crosstalk target genes identified in either neuropsychiatric disorders or inflammatory disorders. Top: the 2D map architecture consists of 37 hexagons indexed circling outward (H1–H37). Bottom: the number of genes mapped onto each hexagon. (B) Prioritization map showing credit score profiles for neuropsychiatric disorders (top) and inflammatory disorders (bottom). (C) Target gene clusters. The prioritization map was divided into 3 clusters (C1–C3), each covering continuous hexagons separated by white borders. Each hexagon is color coded by the probability of containing structurally targetable genes. The percentage of targetable genes per cluster is also shown below (polar bar). A gene was defined as targetable if it was predicted to have drug-like binding sites (i.e., druggable pockets) based on its known protein structure(s). (D) Druggable pockets for genes in C3 specific for neuropsychiatric disorders. The dot plot presents 15 targetable genes (y axis) and their PDB known protein structures (x axis). The number of druggable pockets predicted based on PDB structures is color coded. Highlighted in bold are structurally targetable genes, which are also marked if involved in calcium signaling. (E) Illustration of druggable pockets viewed in 3D within PDB structures for calcium signaling genes in C3. For each structure, a druggable pocket is illustrated using 3D blue balls. See also Figure S5 and Table S6 . Crosstalk-based prioritization map for neuropsychiatric and inflammatory disorders (A) Illustration of the supra-hexagonal map. The map was used to analyze a matrix (containing credit scores) of 112 pathway crosstalk target genes identified in either neuropsychiatric disorders or inflammatory disorders. Top: the 2D map architecture consists of 37 hexagons indexed circling outward (H1–H37). Bottom: the number of genes mapped onto each hexagon. (B) Prioritization map showing credit score profiles for neuropsychiatric disorders (top) and inflammatory disorders (bottom). (C) Target gene clusters. The prioritization map was divided into 3 clusters (C1–C3), each covering continuous hexagons separated by white borders. Each hexagon is color coded by the probability of containing structurally targetable genes. The percentage of targetable genes per cluster is also shown below (polar bar). A gene was defined as targetable if it was predicted to have drug-like binding sites (i.e., druggable pockets) based on its known protein structure(s). (D) Druggable pockets for genes in C3 specific for neuropsychiatric disorders. The dot plot presents 15 targetable genes (y axis) and their PDB known protein structures (x axis). The number of druggable pockets predicted based on PDB structures is color coded. Highlighted in bold are structurally targetable genes, which are also marked if involved in calcium signaling. (E) Illustration of druggable pockets viewed in 3D within PDB structures for calcium signaling genes in C3. For each structure, a druggable pocket is illustrated using 3D blue balls. See also Figure S5 and Table S6 .

This study convincingly tackles the " W-H-W" challenge through the introduction of a pleiotropy-driven approach designed to resolve pleiotropic association data for translational use in therapeutic target identification and validation. By improving therapeutic target discovery, PIPE represents a substantial leap toward narrowing the gap between cross-disease pleiotropy discovery and translational therapeutic applications. Our work adds a new translational strategy for early-stage therapeutics development with a specific focus on pleiotropic evidence. This approach proves particularly valuable for diseases that, despite manifesting distinct phenotypes, share therapeutic relevance. Consequently, our efforts deepen the ongoing quest for accurate therapeutic target prediction and validation. In a broader context, PIPE advances computational medicine that leverages human disease genetics and genomics for therapeutic target discovery. We have demonstrated the efficacy of PIPE across two paradigmatic systems—neuropsychiatric and inflammatory disorders. These systems, often employed as models for studying cross-disease genetic pleiotropy due to their high genetic heritability and therapeutic overlap, serve as robust testing grounds for the capabilities of PIPE. It is to identify target genes that mediate crosstalk between molecular pathways. Within the crosstalk network, functional modules are enriched for approved drug targets and thus offer insights into therapeutic potential. Additionally, the crosstalk-based prioritization map can reveal structurally targetable calcium signaling genes specific to neuropsychiatric disorders. The scientific community is encouraged to leverage PIPE for their own datasets, extrapolating our insights to diverse disorder systems. Both PIPE and Open Targets rely on similar features—distance, promoter-capture Hi-C, and eQTLs—in their target prioritization processes. Consequently, a target gene supported by multiple lines of evidence is assigned with a higher priority in both approaches. Nevertheless, PIPE distinguishes itself from Open Targets in two crucial aspects. Firstly, Open Targets employs a weighted harmonic sum strategy to aggregate scores specific to individual data sources, including the locus-to-gene score from the Open Targets Genetics Portal. In contrast, PIPE calculates the credit score using Fisher’s combined method applied to the gene-predictor matrix that contains affinity scores. Secondly, PIPE incorporates the ability to identify functionally linked targets with no direct genetic evidence (i.e., peripheral genes) within the context of gene interaction networks — the capability absent in Open Targets (and its Genetics Portal). These distinctions, particularly the additional use of gene interaction networks (potentially eQTL networks 57 ), likely contribute to the observed improved performance over genetic association scores from Open Targets. It should be noted, however, that PIPE is limited by the availability of regulatory genomic datasets. The use of such datasets is pivotal for establishing links between non-coding loci and genomic/pleiotropic genes responsible for pleiotropic associations. Presently, the available datasets predominantly focus on immune- and brain-related contexts, thus requiring caution when extrapolating PIPE to other systems of disorders. With the increasing accessibility of promoter-capture Hi-C and eQTL datasets spanning a broader spectrum of cell types and tissues, we anticipate that pleiotropy-centered target identification and validation, via PIPE, will gain enhanced efficacy. Another potential avenue for improving target prioritization involves weighting the predictors according to their importance instead of the equal weighting implemented in this study. This alternative is particularly valuable for singling out targets, for example, supported by tissue- or cell-type-specific eGenes , thereby placing an emphasis on specific tissues or cell types for experimental validation. Moreover, although beyond the scope of this study, it would be potentially beneficial for future improvements to explore the inclusion of fine-mapping results in PIPE. This might involve borrowing from multidisease fine-mapping methods such as CAFEH, 58 flashfm, 59 msCAVIAR, 60 and PAINTOR, 61 which leverage summary statistics and linkage disequilibrium matrices.

Computational medicine, though lacking a precise definition, is increasingly recognized for its rapid facilitation of bridging the gap between the discovery of cross-disease pleiotropy and the application of translational therapeutics. This trend aligns with the evolving focus of human genetics research and drug development (R&D), which is progressively shifting toward translational use (of genetic findings), integrative prioritization (together with performance evaluation), and cross-disease analyses (based on population-scale cohorts with rich phenotypes). It has been over 15 years since the 2007 landmark genome-wide association studies (GWASs) in common chronic diseases, 1 , 2 and we find ourselves in an era of “trying and testing” within human genetics R&D, as highlighted in Alzheimer’s disease 3 and type 2 diabetes. 4 Despite notable advancements, the drug attrition rate remains high, with estimates suggesting that more than 90% of phase 1 clinical trial drug candidates fail to progress through the drug discovery pipeline. 5 This underscores the urgent need to increase the success rate along the pipeline. Human genetics R&D, particularly in the post-GWAS translational medicine era, is poised to play a pivotal role in addressing this challenge, marked by the three aforementioned shifts, which will be explained in greater detail below. The first and foremost shift involves moving the emphasis from merely understanding the genetics of disease to the translational use of genetic findings. 6 , 7 , 8 This shift is motivated by retrospective analyses revealing that human genetic evidence supporting drug-target pairs significantly enhances the success rate of drug development, 9 , 10 , 11 especially when drugs are supported by targets with a causal relation to the disease. 12 The second shift involves the development of drug-target integrative prioritization approaches, encompassing two key recipes: (1) the utilization of multimodal genomic datasets on gene regulation and (2) the incorporation of network knowledge on protein interactions. In Recipe 1, the use of regulatory genomic datasets addresses the intrinsic difficulty in linking genetic loci, mostly non-coding in common diseases, to effector genes and pathways. The regulatory effects of non-coding loci on distant genes are often exerted over considerable distances and in a cell-type-specific manner. These effects can be measured quantitively. For example, promoter-capture Hi-C studies map long-range physical interactions with gene promoters, while expression quantitative trait locus (eQTL) studies map genetic associations with gene expression. Recipe 2 emphasizes the integration of knowledge about protein interactions. This integration is motivated by the recognition that combining human genetic findings with protein interaction data significantly enhances the recovery of known therapeutic drug targets. 13 , 14 The most fundamental is the shift from single-disease associations to cross-disease analyses, which was made possible by population-scale genomic and phenomic data, 15 particularly the introduction of phenomics 16 and translational phenomics. 17 Appreciation is growing for the simultaneous analysis of a broad spectrum of diseases, recognizing its potential to boost the discovery in genetics. Cross-disease analyses have proven instrumental in identifying shared genetic associations, each exhibiting pleiotropic effects across multiple diseases. This cross-disease analysis has found success in the investigation of neuropsychiatric disorders 18 and inflammatory disorders. 19 The term pleiotropic originates from the Greek words “pleio” (many) and “tropic” (ways). In the context of disease genetics, pleiotropy refers to a phenomenon where a genetic locus or single gene influences multiple traits. 20 Specifically, the term “pleiotropic loci” denotes shared genetic loci each influencing two or more diseases. This concept differs from “polygenic loci,” which describes multiple genetic loci converging to impact a single disease. Remarkably, pleiotropic effects manifest in diseases with often divergent phenotypes or distinct clinical manifestations. This aligns with Mendel’s study of inheritance in pea plants a century ago observing that the “seed coat color” gene affected apparently unrelated traits, such as flower and axil pigmentation. Diseases with diverse phenotypes, such as neuropsychiatric disorders, may share underlying pathogenic mechanisms, presenting significant opportunities for therapeutic sharing. Identifying shared pleiotropic loci and elucidating their effector genes and pathways hold profound implications for therapeutic target discovery. The central question arising is how to effectively identify therapeutic targets from cross-disease pleiotropic associations. We can learn from what is already known, particularly insights from the two aforementioned recipes when developing integrative approaches primarily for individual diseases. Alongside contributions from others, 21 our work positions us at the forefront of developing integrative prioritization approaches and tools tailored for individual diseases, such as Alzheimer’s disease, 22 ankylosing spondylitis, 23 asthma, 24 critical COVID-19, 25 endometriosis, 26 immunoglobulin A nephropathy, 27 kidney stone disease, 28 myasthenia gravis, 29 , 30 and type 1 diabetes. 31 These tools harness human disease genetics in conjunction with a wealth of functional genomic datasets aimed at facilitating early-stage therapeutic target discovery. The principle of genetics-led target discovery is gaining increasing recognition and is being employed to enhance late-stage drug approval processes. In this study, we establish that this principle is also applicable to cross-disease pleiotropic associations. This is demonstrated by the development of a pleiotropy-driven approach, pleiotropy informing prioritization and evaluation (or PIPE), designed to enable pleiotropy-driven therapeutic target prioritization and evaluation ( Figure 1 ). We substantiate the validity of PIPE by showcasing its improved performance in generating therapeutic target resources for neuropsychiatric and inflammatory disorders. We also show the potential of PIPE in: (1) identifying target genes involved in pathway crosstalk; (2) revealing functional modules with therapeutic potential and evolutionary origin, particularly co-arising of inflammatory disorder-specific modules with the speciation of bony fish; and (3) constructing a prioritization map in which structurally targetable calcium signaling genes are specific to neuropsychiatric disorders. Alongside freely accessible resources and step-by-step reproducible showcases, PIPE represents a valuable tool for in silico early-stage therapeutic target discovery, with a specific focus on pleiotropic evidence. Figure 1 Overview of pleiotropy informing prioritization and evaluation (PIPE) (A) Conceptual illustration. The artwork PIPE is designed to resemble a computational framework for pleiotropy-driven therapeutic target prioritization and evaluation. (B) Three steps designed for therapeutic target prioritization at both the gene and pathway crosstalk levels. Notably, nGene (in blue) refers to nearby genes with evidence of genomic proximity, cGene (in green) refers to conformation genes with evidence of gene promoters physically interacting with SNP-harboring genomic regions, and eGene (in pink) refers to expression-associated genes with evidence of genetic regulation of gene expression. (C) Three instrumental analyses enabling performance evaluation. They include the supportive procedure of how to define clinical proof-of-concept targets and how to simulate negative controls, performance benchmarking, and pleiotropy informing clinical therapeutics (PICT) via leading prioritization analysis (LPA). Overview of pleiotropy informing prioritization and evaluation (PIPE) (A) Conceptual illustration. The artwork PIPE is designed to resemble a computational framework for pleiotropy-driven therapeutic target prioritization and evaluation. (B) Three steps designed for therapeutic target prioritization at both the gene and pathway crosstalk levels. Notably, nGene (in blue) refers to nearby genes with evidence of genomic proximity, cGene (in green) refers to conformation genes with evidence of gene promoters physically interacting with SNP-harboring genomic regions, and eGene (in pink) refers to expression-associated genes with evidence of genetic regulation of gene expression. (C) Three instrumental analyses enabling performance evaluation. They include the supportive procedure of how to define clinical proof-of-concept targets and how to simulate negative controls, performance benchmarking, and pleiotropy informing clinical therapeutics (PICT) via leading prioritization analysis (LPA).

REAGENT or RESOURCE SOURCE IDENTIFIER Deposited data Cross-disease GWAS summary data for neuropsychiatric disorders Lee et al. 18 N/A Cross-disease GWAS summary data for inflammatory disorders Ellinghaus et al. 19 N/A GWAS Catalog Sollis et al. 62 https://www.ebi.ac.uk/gwas Promoter capture Hi-C for 27 human cell and tissue types Jung et al. 32 N/A Promoter capture Hi-C for 4 human neural cells Song et al. 33 N/A eQTL Catalog Kerimov et al. 34 https://www.ebi.ac.uk/eqtl Open Targets Ochoa et al. 37 https://www.opentargets.org ChEMBL Mendez et al. 63 https://www.ebi.ac.uk/chembl RCSB Protein Data Bank Burley et al. 51 https://www.rcsb.org Pleiotropy-based target resources for neuropsychiatric disorders and inflammatory disorders This study http://www.genetictargets.com/PIPE Software and algorithms fpocket Schmidtke et al. 64 https://github.com/Discngine/fpocket supraHex Fang & Gough. 46 https://bioconductor.org/packages/supraHex PIPE This study https://doi.org/10.6084/m9.figshare.25331527.v1 and https://hfang-bristol.github.io/PIPE Further information and request for resources should be directed to and will be fulfilled by the lead contact, Prof. Hai Fang ( [email protected] ). This study did not generate new unique reagents. • All data used and generated in the current study are accessible through the dedicated web portal at http://www.genetictargets.com/PIPE . This portal provides an interactive platform for exploring pleiotropy-based target resources for neuropsychiatric disorders and inflammatory disorders, as well as functional modules and the prioritisation map. In support of reproducible research, we have provided comprehensive showcases that empower users to reproduce all findings from this study. These showcases include input data, line-by-line codes, and both tabular and graphical outputs, all encapsulated within a single self-contained HTML file. This file also contains detailed, step-by-step instructions on how to use the showcases and what results to expect. • The code for PIPE is also available at https://hfang-bristol.github.io/PIPE and https://doi.org/10.6084/m9.figshare.25331527.v1 . • Any additional information required to reanalyse the data reported in this paper is available from the lead contact upon request. All data used and generated in the current study are accessible through the dedicated web portal at http://www.genetictargets.com/PIPE . This portal provides an interactive platform for exploring pleiotropy-based target resources for neuropsychiatric disorders and inflammatory disorders, as well as functional modules and the prioritisation map. In support of reproducible research, we have provided comprehensive showcases that empower users to reproduce all findings from this study. These showcases include input data, line-by-line codes, and both tabular and graphical outputs, all encapsulated within a single self-contained HTML file. This file also contains detailed, step-by-step instructions on how to use the showcases and what results to expect. The code for PIPE is also available at https://hfang-bristol.github.io/PIPE and https://doi.org/10.6084/m9.figshare.25331527.v1 . Any additional information required to reanalyse the data reported in this paper is available from the lead contact upon request. PIPE was implemented as an R package, specially designed for ab initio resolving cross-disease pleiotropic association data, as described below in the subsection ‘ processing input cross-disease pleiotropic association data ’. Its primary objective is to accelerate translational use in therapeutic target identification and validation. Notably, PIPE is not devised for the upstream generation of pleiotropic association data but is rather tailored to support downstream applications in therapeutic target discovery. PIPE operates as a pleiotropy-driven unsupervised tool, featuring two integral components ( prioritisation and evaluation ). The resulting resources are made accessible via a web-based portal, and all showcases are made reproducible as well ( Figure 1 A). The prioritisation component of PIPE consists of three sequential steps to accomplish therapeutic target identification at both the gene and pathway crosstalk levels, as graphically illustrated in Figure 1 B. • Step 1. Identify genomic/pleiotropic genes with genomic evidence by leveraging multimodal regulatory genomic datasets (see ‘ identification of pleiotropic genes with genomic evidence ’ for greater details) and networked/peripheral genes with network evidence by leveraging knowledge of protein interactions (see ‘ identification of peripheral genes with network evidence ’), to prepare informative predictors (see ‘ extraction of informative predictors and preparation of gene-predictor matrix ’). • Step 2. Combine informative predictors to prioritise target genes across the genome (that is, targets at the gene level; see ‘ prioritisation of targets at the gene level ’). • Step 3. Identify a network of genes that mediate crosstalk between molecular pathways (that is, targets at the pathway crosstalk level; see ‘ identification of targets at the pathway crosstalk level ’), followed by crosstalk-based advanced analyses, including network modular analysis (see ‘ crosstalk-based network modular analysis ’) and prioritisation map analysis (see ‘ crosstalk-based prioritisation map analysis ’). Step 1. Identify genomic/pleiotropic genes with genomic evidence by leveraging multimodal regulatory genomic datasets (see ‘ identification of pleiotropic genes with genomic evidence ’ for greater details) and networked/peripheral genes with network evidence by leveraging knowledge of protein interactions (see ‘ identification of peripheral genes with network evidence ’), to prepare informative predictors (see ‘ extraction of informative predictors and preparation of gene-predictor matrix ’). Step 2. Combine informative predictors to prioritise target genes across the genome (that is, targets at the gene level; see ‘ prioritisation of targets at the gene level ’). Step 3. Identify a network of genes that mediate crosstalk between molecular pathways (that is, targets at the pathway crosstalk level; see ‘ identification of targets at the pathway crosstalk level ’), followed by crosstalk-based advanced analyses, including network modular analysis (see ‘ crosstalk-based network modular analysis ’) and prioritisation map analysis (see ‘ crosstalk-based prioritisation map analysis ’). The evaluation component of PIPE facilitates three instrumental analyses to enable performance evaluations, as graphically illustrated in Figure 1 C. • Analysis 1. Define clinical proof-of-concept targets (see ‘ defining clinical proof-of-concept targets ’) and simulate negative controls (see ‘ simulating negative targets as controls ’). • Analysis 2. Do benchmarking to compare prioritisation performance in terms of separating clinical proof-of-concept targets from simulated controls (see ‘ benchmarking for performance comparisons ’). • Analysis 3. Evaluate the potential of pleiotropy in informing clinical proof-of-concept therapeutics (see ‘ leading prioritisation analysis (LPA) ’ and ‘ pleiotropy informing clinical therapeutics (PICT) ’) and determine target-level disease relationships (see ‘ determining target-level disease relationships ’). Analysis 1. Define clinical proof-of-concept targets (see ‘ defining clinical proof-of-concept targets ’) and simulate negative controls (see ‘ simulating negative targets as controls ’). Analysis 2. Do benchmarking to compare prioritisation performance in terms of separating clinical proof-of-concept targets from simulated controls (see ‘ benchmarking for performance comparisons ’). Analysis 3. Evaluate the potential of pleiotropy in informing clinical proof-of-concept therapeutics (see ‘ leading prioritisation analysis (LPA) ’ and ‘ pleiotropy informing clinical therapeutics (PICT) ’) and determine target-level disease relationships (see ‘ determining target-level disease relationships ’). Input pleiotropic association data were a list of SNPs, each exerting pleiotropic effects on multiple diseases, along with summary statistics indicating the significance level (i.e., pleiotropic p-values combined across multiple diseases). These pleiotropic loci were identified from cross-disease GWASs (illustrated in Figure 1 A). In this study, the input data encompassed a total of 142 lead SNPs (significant at p-value <5 × 10 −8 ) associated with two or more neuropsychiatric disorders, 18 and a total of 224 lead SNPs (p-value <5 × 10 −8 ) associated with two or more inflammatory disorders. 19 These data, manually curated in the NHGRI-EBI GWAS Catalog, 62 included pleiotropic lead SNPs and their p-values based on the European population for subsequent analysis. Additional SNPs in linkage disequilibrium (LD; R 2 ≥ 0.8) with lead SNPs were calculated according to the European population. 65 The rationale for employing LD expansion on lead SNPs is the ease of its widespread applicability. While an alternative approach via fine-mapping could be explored in the future, it is currently constrained by the availability of summary statistics, which may not always be readily accessible. Lead and LD SNPs were scored by considering pleiotropic association p-values and LD strength R 2 ( Equation 1 ). Equation 1 S S N P = R 2 × ( log 10 1 − P S N P P S N P − log 10 1 − 5 × 10 − 8 5 × 10 − 8 ) Where S S N P and P S N P denote the SNP score and pleiotropic p-value combined across multiple diseases, respectively, and R 2 is the LD strength. Lead and LD SNPs from the aforementioned pleiotropic association data were used to identify pleiotropic genes. These pleiotropic genes were categorised into three types according to the genomic evidence (i.e., which regulatory genomic datasets were used): (i) nearby genes ( nGene ), identified with evidence of genomic proximity; (ii) conformation genes ( cGene ) identified using promoter capture Hi-C datasets, reflecting evidence of gene promoters physically interacting with SNP-harbouring genomic regions; and (iii) expression-associated genes ( eGene ) identified based on eQTL datasets, reflecting evidence of genetic regulation of gene expression. Notably, immune- and brain-related datasets were used in this study, including 24 promoter capture Hi-C datasets 32 , 33 and 69 eQTL datasets 34 ; accordingly, there were 24 sets of cGenes and 69 sets of eGenes . The scoring for nGene considered the influential range of distance ( Equation 2 ). The scoring for cGene (or eGene ) was tissue- or cell type-specific: for each tissue cGene (or eGene ), it considered the empirical cumulative distribution function (eCDF) of the strength or significance level linking an SNP to a gene ( Equations 3 and 4 ). The eCDF was estimated based on all SNP-gene pairs, ensuring the scaling into the 0–1 range. It is also worth noting that these three types of genes could overlap; a gene might simultaneously be categorised as both nGene and eGene . Equation 2 S n G e n e = max S N P ∈ Ω S S N P Where S n G e n e is the score for nGene , S S N P is the SNP score, Ω stands for collections of lead and LD SNPs located within 20kb of nGene , and max denotes the maximum scoring scheme, keeping the most informative one when multiple SNPs are located within the same pleiotropic region. In simpler terms, any genes within ±20kb around the lead and LD SNPs will be considered as nGenes . Notably, the findings were robust, with the trivial effect of changing the distance windows on the performance ( Figure S6 ). Equation 3 x = − log 10 ( S S N P ↔ G e n e ) Equation 4 S c G e n e o r S e G e n e = max S N P ∈ Ω [ S S N P × e C D F ( x ) ] Where S c G e n e is the score for cGene , S e G e n e is the score for eGene , S S N P is the SNP score, S S N P ↔ G e n e denotes the strength or significance level linking an SNP to a gene (that is, the significance level of genetic association with gene expression for an eQTL dataset, and the physical interaction strength with a gene promoter for a promoter capture Hi-C dataset), Ω stands for collections of lead and LD SNPs, and max denotes the maximum scoring scheme. Notably, the scoring for cGene (or eGene ) was handled for each tissue or cell type. We employed a random walk with restart algorithm to identify peripheral genes under network influence, that is, peripheral genes connected or networked to pleiotropic genes. Each type of pleiotropic genes (or dataset of the same evidence type) was considered as core seeds, and we used them to further identify peripheral genes by leveraging the network connectivity or affinity of protein interaction information. The protein interaction data were sourced from the STRING database, 35 focusing on interactions with high-confidence scores (≥700) and high-quality sources (limited to evidence codes ‘experiments’ or ‘databases’). This corresponded to approximately 15,000 nodes or genes that underwent prioritisation. Starting from a seed node, the random walker faces two choices at each iterative step: either moving to a randomly chosen neighbor or jumping back to the seed node ( Equation 5 ). Upon reaching stability after a finite number of steps, the resultant steady probability vector measures the “influential impact” over the network imposed by the starting seed nodes ( Equation 6 ). Affinity scores were computed for each gene based on its connectivity to pleiotropic genes, and the vector comprising these scores constituted a predictor for each evidence type or dataset of the same evidence type. Without a loss of generality, a gene with a higher connectivity to pleiotropic genes received a higher affinity score, with highly networked pleiotropic genes securing markedly higher affinity scores. We generated one predictor for nGene , 24 predictors for promoter capture Hi-C datasets ( cGenes ), and 69 predictors for eQTL datasets ( eGenes ). Equation 5 P t ⇀ = ( 1 − γ ) × A × P t − 1 ⇀ + γ × P 0 ⇀ Equation 6 P ∞ ⇀ = lim t → ∞ P t ⇀ Where P ∞ ⇀ represents the steady probability vector (attained upon reaching stability) that stores the affinity score of all nodes in the network to the starting seed nodes (i.e., pleiotropic genes), P t ⇀ is the probability vector dictating the walker’s visits to network nodes at step t , P 0 ⇀ is the starting probability vector that contains gene scores for seed nodes (i.e., scores in Equations 2 and 4 ) and 0 otherwise, γ is the restarting probability (probability of jumping back to a seed node, and 1 − γ for the probability of moving to a neighbor), and A is the normalised Laplacian adjacency matrix of the network being traversed. As illustrated at Step 1 of Figure 1 B, we used random forest to assess the relative importance of predictors. We retained only informative predictors, identified as those whose removal resulted in a substantial decrease in accuracy; see detailed methodology. 38 Since the nGene predictor was considered the baseline, we retained subsets of cGene and eGene predictors that were deemed at least as important as nGene . In essence, utilising pleiotropic association data, we prepared a gene-predictor matrix consisting of affinity scores, where informative predictors as columns and pleiotropic and peripheral genes as rows (totaling approximately 15,000 genes). We performed the gene-level target prioritisation by combining informative predictors using a method akin to Fisher’s combined meta-analysis. For each predictor in the gene-predictor matrix, we converted affinity scores into P -like values through eCDF based on the affinity scores for all genes in that predictor ( Equation 7 ). Then, for each gene in the gene-predictor matrix, we combined these converted p-values across predictors using Fisher’s combined method ( Equations 8 , 9 , and 10 ). This method is suitable when the evidence is concentrated or moderately strong. 66 Finally, we rescaled the combined p-value into a credit score, ranging from 0 to 10 ( Equation 11 ). Equation 7 P i j = e C D F ( A F i j ) Where A F i j denotes the affinity score for the i th gene in terms of the j th predictor, P i j is the corresponding converted p-value, and eCDF is estimated based on all genes. Equation 8 x i = − 2 ∑ j J log P i j Equation 9 x i ∼ χ 2 ( 2 J ) Equation 10 C P i = C D F ( x i ) Equation 11 C S i = 10 × − log C P i − MIN k K ( − log C P k ) MAX k K ( − log C P k ) − MIN k K ( − log C P k ) Where P i j denotes the converted p-value for the i th gene in terms of the j th predictor, J is the number of the informative predictors, χ 2 ( 2 J ) denotes the Chi-squared distribution with 2J degrees of freedom, CP i represents the combined p-value for the i th gene (i.e., eCDF of the Chi-squared distribution valued at x i ), and CS i is the credit score for the i th gene (among K genes). First, we constructed a gene network by merging KEGG pathways. 67 Next, we extended our previously established algorithm 68 , 69 , 70 to identify a subset of the pathway-merged gene network that contains highly prioritised/credited and interconnected genes. This identified gene subnetwork, referred to as ‘pathway crosstalk’, underwent a degree-preserving node permutation test to estimate its significance (p-value) relative to chance occurrences (i.e., by counting how frequently it would occur by chance; see detailed information 70 ). Information on current therapeutics, including drugs, development phases, target genes, and disease indications, was retrieved from the ChEMBL database. 63 Specifically, for a particular disease, we defined ‘ clinical proof-of-concept targets ’ as therapeutic genes targeted by drugs that have advanced to development phase 2 or above, signifying substantial evidence of their efficacy in treating the respective disease. In our performance evaluation, negative targets were simulated for a particular disease following the strategy outlined in Analysis 1 of Figure 1 C, with two key steps. The first step is to define druggable landscape. We delineated the druggable landscape, encompassing all targetable genes across various disease indications and at all drug development phases. The second step is to simulate negative targets. These negative targets were simulated by selecting genes from the druggable landscape, but after systematically excluding genes known to be specific to the disease under consideration, along with their interacting neighbors. Interacting neighbors were determined based on interaction data sourced from databases including STRING 35 and Pathway Commons. 71 Taken together, this simulation strategy aimed to ensure that the negative targets, under the constraint of the existing druggable landscape (i.e., evidence already explored) and protein interaction information (due to the tendency of known targets to interact together), were unlikely to be target genes for the disease under consideration. A benchmarking analysis was conducted to assess and compare the performance of PIPE with the state-of-the-art Open Targets. 37 The performance focused on the ability to distinguish clinical proof-of-concept drug targets from simulated negative targets. In the Open Targets model, the prioritisations were evaluated based on individual pieces of evidence, specifically genetic association (for details see the open targets genetics portal 36 ) and text mining . Information on genetic association scores and text mining scores was extracted from the Open Targets Platform, accessible at https://platform.opentargets.org . 37 LPA was performed to quantify the extent of enrichment in a target set, such as clinical proof-of-concept targets, at the ‘leading prioritisation’ within the prioritised gene list ranked by credit scores. Analogous to gene set enrichment analysis, the leading prioritisation (or the leading edge) encompasses the leftmost subset of the highly prioritised/credited gene list. The normalised enrichment score (NES), calculated as the observed running enrichment score divided by the expected score, was determined, with the significance level (p-value) estimated through a permutation test conducted 50,000 times. We introduced a new metric, PICT, to quantify the Pleiotropy Informing Clinical Therapeutics potential for a particular disease. PICT measures the inclination of the prioritised gene list to contain clinical proof-of-concept targets. This tendency was assessed through LPA, evaluating the degree to which clinical proof-of-concept targets are enriched at the top/leftmost of the prioritised gene list. PICT was computed using the formula specified in Equation 12 , where diseases with higher PICT exhibit more substantial enrichment (FDR; p-value adjusted accounting for multiple tests), a higher NES, and a higher coverage (Coverage, defined as the fraction of clinical proof-of-concept targets recovered at the ‘leading prioritisation’ of the prioritised list). Equation 12 P I C T = log 10 ( N E S × C o v e r a g e / F D R ) We presented the relationships among diseases as a network, with diseases represented as nodes and their estimated relations as edges. For each disease, member genes were defined as clinical proof-of-concept targets recovered at the leading prioritisation. We initially inferred edges if any member genes were shared between diseases and subsequently refined by identifying a minimum spanning tree. Only edges present in the resulting tree were retained. The thickness of an edge was proportional to the number of member genes shared between the two endpoint diseases. We employed a spin-glass model and simulated annealing to identify network modules within crosstalks. For each module, we performed enrichment analysis to unveil its functional relevance, therapeutic potential, and evolutionary origin. Functional relevance was determined by identifying enriched KEGG pathways. We identified enriched approved drug targets and non-approved phased drug targets using the ChEMBL database 63 to assess therapeutic potential. Therapeutic targets were categorised into two groups: one for approved drug targets (i.e., genes targeted by any approved drugs), and the other for non-approved phased drug targets (i.e., genes targeted by non-approved phased drugs). Additionally, genes created in each of 16 phylostrata, arranged from the earliest (PSG01) to the most recent (PSG16), were obtained from this study 44 for enrichment analysis to reveal evolutionary origin. We identified two instances of pathway crosstalks, one associated with neuropsychiatric disorders and the other with inflammatory disorders, yielding a total of 116 crosstalk target genes. We utilised the supraHex package 46 to cluster and visualise these genes on a supra-hexagonal map consisting of 91 hexagons. The map was trained using a matrix containing credit scores for the 116 target genes across two systems of disorders (i.e., neuropsychiatric and inflammatory). This trained map was used to illustrate a gene prioritisation profile for each disorder system. The trained map was also topologically divided into target gene clusters. We overlaid the trained map with druggable pocket data, estimating the probability of each hexagon containing structurally targetable genes. A gene was considered targetable if predicted to harbor drug-like binding sites (i.e., druggable pockets) based on its known protein structure(s). All known protein structures were sourced from the PDB database, 51 with druggable pockets predicted using the fpocket software. 64