DRIVE-KG: Enhancing variant-phenotype association discovery in understudied complex diseases using heterogeneous knowledge graphs

To construct a heterogeneous knowledge graph (KG), we integrated multiple biological data sources (node and edge types summarized in Table 1 ). The graph is comprised of four primary node types: single nucleotide polymorphisms (SNPs), genes, proteins, and phenotypes. We derived SNP nodes from the Single Nucleotide Polymorphism Database (dbSNP). 16 To avoid node type imbalance caused by including tens of millions of variants, we included only SNPs with minor allele frequency (MAF) ≥ 0.01 in the Penn Medicine BioBank (PMBB). We obtained phenotype nodes from the Human Phenotype Ontology (HPO), 17 which contains standardized disease and trait terminology. We sourced gene information from the National Center for Biotechnology Information (NCBI) gene database, 18 and derived protein nodes from the Universal Protein Resource (UniProt). 19 We incorporated five distinct edge types connecting the four biological entities (SNP, gene, phenotype, and protein) into the KG. We established SNP-to-gene associations using expression quantitative trait loci (eQTL) data from version 8 of the Genotype-Tissue Expression (GTEx) project. 20 We derived SNP-to-phenotype relationships from Open Targets, 21 which aggregates evidence from genome-wide association studies (GWAS) and other sources. We obtained gene-to-phenotype associations from PheWAS (phenome-wide association study) summary statistics, published from OmicsPred. 22 We established SNP-to-protein connections through OmicsPred, and sourced gene-to-protein relationships from UniProt annotations. 19 When designing the DRIVE-KG schema ( Figure 1 ), we prioritized having dense node properties and fewer edge types to avoid sparsity issues when performing machine learning analyses on the graph. We maintained cross references of node type identifiers across databases (‘dbXrefs’) as properties of the gene and phenotype node for possible future querying by other researchers. The ‘tissue expression’ property of the ‘hasEqtl’ edge is an array of length 53 (number of tissues) that contains all tissue-specific beta values for eGenes in this dataset. Since eGenes (a gene whose expression is influenced by eQTLs) are tissue-specific in the Omics Pred dataset, we aggregated all tissue-specific effects for a particular SNP–Gene pair into this list. If a particular SNP–Gene pair did not have a tissue-specific association, we set the beta value for that entry to 0. Omics Pred pQTl data pertain to the same 53 tissue types as published by the GTEx consortium, with the exception of Bladder, Fallopian tube, and Cervix (Endocervix & Ectocervix), resulting in a total of 49 tissue types. We mapped the tissue expression array in the same alphabetical order as used by the GTEx consortium, 20 so the index of a given beta value in this array can be used to infer from which tissue it came from. The DRIVE-KG construction process involved several key steps: first, we standardized identifiers across all data sources (e.g., harmonized to the same Entrez identifier for all gene-related data and ensured that hg38 was used to represent all SNP-related information). Finally, we ingested these files into Memgraph for subsequent visualization and analysis. For the purposes of model training, we removed all unconnected nodes (node with degree = 0) in the graph (removing 94.2% of the original 13,806,079 nodes), yielding a total of 801,112 nodes and 1,390,440 edges in the graph. The removed nodes consisted mainly of highly specific genetic variants and phenotypes without documented associations in our data sources. This pre-processing step preserved the core connectivity structure necessary for effective graph-based learning and link prediction analyses. To identify previously unreported, biologically relevant relationships within DRIVE-KG, we employed link prediction methods that leverage its multi-modal data structure. The core premise of our link prediction framework is that missing edges can be inferred from graph connectivity patterns, because biologically meaningful associations often exhibit characteristic topologies. For instance, if a SNP is associated with multiple genes that are collectively linked to a specific phenotype, this suggests a potential direct SNP-phenotype relationship that may not be explicitly captured in current databases. Similarly, proteins sharing common genetic variants and phenotypic associations may indicate previously uncharacterized protein-protein interactions or shared functional pathways. Given that DRIVE-KG is created with disease-agnostic biological data, it can be queried for links to any phenotype (using its HPO identifier). We were interested in predicting whether there should be a link between a given SNP node and two pilot phenotypes: endometriosis and obesity, both complex traits with contrasting levels of genetic characterization (67 versus 112 existing genome-wide significant variants, respectively). We generated 64-dimension node embeddings using Memgraph’s node2vec 23 module and then used these embeddings as input for a pairwise multilayer perceptron (MLP) model developed using the PyTorch Lightning 24 module. The three-layer feed-forward MLP used rectified linear unit (ReLU) activations and dropout, projecting 64-dim embeddings to 128, then sequentially reducing to 64 and finally 1 dimension. The output scalar represents the probability of a particular SNP-phenotype edge. We trained the link prediction module on all snpAssociatedWithDisease (SNP-Phenotype) edges with Memgraph’s native link prediction module, applying a 70–30 training-validation split. To evaluate specific phenotypes, we queried the trained model with their HPO identifiers (HP 0030127 for endometriosis and HP 0001513 for obesity). The model outputs a score in the range [0,1] (from the final sigmoid activation), representing its confidence in a potential association. To assess the sensibility of these predictions, we focused on high-confidence edges (model score ≥ 0.95) and cross-referenced them with published GWAS results. 12 We also explored the neighborhood of these candidate associations within DRIVE-KG. In order to assess the utility of DRIVE-KG for disease risk stratification, we trained a graph convolutional network (separate from the link prediction procedure). In particular, we explored the potential of this graph to accurately stratify disease risk at the patient level when applied to endometriosis/adenomyosis. We obtained chart reviews for a cohort of 1,441 patients from the PMBB 7 with their status of endometriosis or adenomyosis (case or control) labeled, as these conditions are often grouped together during phenotyping despite likely differences in their underlying pathogenesis. 25 For endometriosis cases, the chart review also documented surgical disease stage according to established clinical classification systems, enabling both binary case-control analyses and stage-specific investigations. To calculate genetically inferred ancestry (GIA) for each participant, principal components were calculated from PMBB genotype data and then projected onto the reference HapMap3 dataset. 26 GIA for each participant was assigned by comparing their projected principal component scores to the HapMap3 reference populations and classifying them using a kernel density estimation approach. 27 Table 2 contains additional cohort demographics and GIA; a supplemental version with relevant co-morbidities is available in the project’s Github repository. To curate the graph classification dataset, we generated per-patient copies of DRIVE-KG from the chart-reviewed cohort. We did this by first exporting the base KG (with existing network edges) and graph connectivity information from Memgraph into Python (PyTorch Geometric library 28 ). Then, we assigned the value of each SNP node, for each patient’s graph, as the imputed genotype dosage (where the value indicates the dosage of the alternate allele at a given locus) from imputed genotyping data in the 2024 PMBB data release 3.0, which was processed using PLINK2.0. 29 Here, the prediction task is whether the entire graph represents endometriosis/adenomyosis (case) or not (control). For all non-SNP node types, we represented them with 8-dimension embeddings learned using the node2vec module. We also added per-patient covariates (age and the first 5 of overall ancestry principal components (PCs)) to the classifier. We leveraged PyTorch Lightning to load the patient graph objects and train the graph classifier. The graph classifier architecture contains two convolutional layers for all edge types, with SAGEConv layers wrapping each of the edges. We chose SAGEConv due to its inductive capabilities, and ability to efficiently scale to large graphs. We used 64-dimension projections of the 8-dimension embeddings learned on all nodes as model input. We used ReLU activation functions for the two convolutional layers, and the ‘sum’ aggregation function (which generates 64-dimension tensors for each node type). We then generated pooled GNN embeddings over the SNP node type, which creates 64-dimension tensors for the entire graph. We pooled over the SNP node type alone since only the SNP node values differ across patient graphs; therefore, it is likely to contain biological signal that can be used to distinguish patients. We concatenated the pooled, 64-dimension tensors with the per-patient covariates data (6-dimension: age and 5 PCs) to get 70-dimension tensors. We fed these tensors into the first linear layer, which transformed them into 64-dimension tensors with the ReLU activation function. The second linear layer contained a sigmoid activation function that generated a binary classification value (single probability per graph/patient) from the 64-dimension tensors. For each epoch, we specified a batch size of 16, a learning rate of 0.001, and used the Adam optimization function. We compared the performance of disease risk prediction using DRIVE-KG to a genetic risk score (GRS) 12 developed and evaluated on the same patient cohort (with the same training-validation split of patients between DRIVE-KG and the GRS). We calculate the GRS by linearly combining the patient’s genetic information, age and 5 PCs (as covariates) to ensure a fair comparison.

We investigated link predictions for two phenotypes of interest: endometriosis (HP 0030127) and obesity (HP_0001513). For each phenotype, we obtained recommendations from the model of SNP-phenotype links, ranked by their binary classification score and annotated with whether they were existing or candidate associations. Our link prediction analysis identified 66 high confidence, candidate SNP-endometriosis associations. Notably, many of these predicted associations involved SNPs with well-established links to other complex traits and conditions yet had not been previously associated with endometriosis in the literature (except for rs10828249 located on MLLT10 30 ). Specifically, 24.2% of the top-ranked predicted SNP-endometriosis links involved variants previously associated with body mass index (BMI)/obesity-related traits, 6% with lipid metabolism (triglycerides and HDL), and 4.5% with depressive disorders. Additionally, these top-ranked predicted SNPs are enriched in previous endometriosis GWAS for p-values with an average magnitude of 0.473. The majority of these associations (60/66) are also not in LD with each other ( R 2 ≤ 0.1), with the exception of 3 variant pairs with R 2 values ranging from 0.4–0.65. We also estimated narrow sense heritability within the PMBB cohort, using previous GWAS associations versus these associations in addition to the 66 SNP-endometriosis candidates. The variant set from previous GWAS (136 SNPs) 12 explained an average variance of 16.73% (pseudo− R 2 90% CI: 13.03–20.39%). Adding the 66 DRIVE-KG variants increased this to 24.64% (pseudo− R 2 90% CI: 20.30–28.82%), an improvement of 7.91%. We observed a higher confidence for candidate associations compared to existing association scores. This could be because the model is undertrained on endometriosis associations compared to more well-studied diseases (i.e., those with a greater number of genome-wide significant associations) and biased towards positive predictions as a result. Candidate SNP-endometriosis associations had a mean confidence of 0.757 ± 0.179, whereas the 67 known SNP-endometriosis edges had a mean confidence of 0.634 ± 0.163. Figure 2 compares known versus candidate SNP−endometriosis associations in DRIVE-KG with − log 2 link prediction score on the y−axis. We also visualize the neighborhood of candidate signals, which are not explicitly connected to endometriosis as a phenotype, but rather relevant pathways and associated conditions. In contrast to the endometriosis results, link prediction for obesity yielded markedly different patterns, reflecting extensive existing knowledge for these well-characterized phenotypes. The link prediction model ranked the 112 existing SNP-obesity associations with a mean confidence of 0.937 ± 0.111 (much higher than that of endometriosis), whereas the candidate SNP-obesity associations were ranked substantially lower with a mean confidence of 0.255 ± 0.178. Of the 1,575 high-confidence, candidate SNP–obesity associations, 348 could be evaluated using LDlink, as many variants lacked corresponding entries in the GWAS Catalog and thus could not be assessed. Among these, 38.22% were in high LD ( R 2 ≥ 0.8) with previously established loci ( NEGR1 31 and DNAJC27-AS1 32 ) for obesity or related comorbidities, rather than unreported associations. This suggests that the predicted associations largely recapitulated existing knowledge of metabolic pathways and obesity-related biological mechanisms already well-documented in the literature. The network visualizations for the high-confidence, candidate SNP-obesity associations in Figure 3 (rs74432706 and rs329642) demonstrate that these predictions are primarily driven by neighboring connections to obesity phenotypes, as opposed to distinct biological signals. Our cohort consisted of 1,411 PMBB participants assigned female at birth with imputed genotyping data and a chart reviewed label (0 or 1) for having endometriosis/adenomyosis or not. With a 70–30 training-validation split, the graph classifier had an F1 score of 0.752. We compared this to a GRS developed and evaluated on the same patient cohort, which had an F1 score of 0.698 (additional metrics displayed in Table 3 ). Since this is a binary classification task, the model provides a score (value between 0 and 1, inclusive) for its predicted likelihood of a given patient having either an endometriosis or adenomyosis diagnosis. Figure 4 provides model score distributions stratified both by disease subtypes from chart reviews ( Figure 4a ) and by surgically confirmed endometriosis stages (1–4) ( Figure 4b ).

This study demonstrates the potential of heterogeneous KG approaches to advance our understanding of understudied diseases where traditional genomic methods have plateaued. Our findings provide several key insights into the utility of integrating multi-omics data through graph-based representations for both discovery and clinical prediction applications. Link prediction revealed a striking contrast between well-studied and understudied phenotypes, validating our central hypothesis. For endometriosis, we identified high-confidence candidate SNPs largely independent of one another ( R 2 < 0.1) and enriched for variants linked to BMI, lipid metabolism, and depressive disorders—traits frequently co-morbid with endometriosis. 33 – 35 Network visualization further connected these variants to underlying inflammatory and cell proliferation pathways, providing additional insight into shared biological mechanisms. This pattern suggests shared biological pathways between endometriosis and these co-morbid conditions, which is consistent with emerging mechanistic hypotheses 12 , 30 , 36 and our PMBB chart reviewed cohort, which shows elevated risks of hyperlipidemia (RR 1.85), obesity (2.53), CVD (2.03), hypertension (1.79), and depression (1.95). These findings help address the “missing heritability” gap between traditional polygenic risk scores (explaining 11%) and the broad sense heritability estimate from a twin study (47%). 12 Conversely, link predictions for obesity—a trait with extensive genomic characterization—yielded lower confidence scores and predominantly identified variants in LD with known associations. This validates that DRIVE-KG appropriately finds fewer novel discoveries for well-mapped biological landscapes while successfully uncovering genuine candidates for understudied conditions. Our patient classification for endometriosis demonstrates modest but meaningful improvements over traditional genetic approaches. The DRIVE-KG-based model achieved an F1 score of 0.752 compared to 0.698 for the GRS. These moderate effect sizes show potential clinical relevance for a condition where existing genomic tools perform poorly: the model showed enhanced prediction for adenomyosis and demonstrated particular strength in identifying patients with severe endometriosis (stage 4). The improved performance likely stems from DRIVE-KG’s ability to capture complex multi-omics relationships that single-modality approaches miss, even when the only patient-specific input is genetic data. Rather than relying solely on individual genetic variants, our framework leverages broader biological knowledge encoded in protein associations, gene expression patterns, and phenotypic associations. This integrated approach may be particularly valuable for conditions like endometriosis where the genetic architecture is complex and poorly understood. These findings have broader implications for addressing healthcare disparities in women’s health research. Endometriosis exemplifies the challenges facing many gynecological conditions: high prevalence (affecting around 10% of women of reproductive-age 9 ), significant clinical impact, yet persistent research neglect leading to limited therapeutic options. Our results suggest that graph-based approaches can help bridge these knowledge gaps by extracting additional insights from existing multi-omics data without requiring massive new cohort studies that may be infeasible for understudied populations. The identification of shared pathways with well-studied traits (obesity, 33 depression 35 ) also suggests opportunities for re-purposing existing therapeutic targets and biomarkers. If endometriosis shares mechanistic pathways with metabolic or psychiatric conditions, this could accelerate drug development and improve patient care through a better understanding of comorbidity patterns. Our study has several limitations. First, the construction of DRIVE-KG relied on existing databases that may contain biases toward well-studied biological pathways, potentially limiting discovery of entirely new mechanisms. Second, while we harmonized variants across alleles when creating DRIVE-KG, we do not fully account for differences in the effect allele when processing the PMBB genotype data to make patient-specific graphs and simply drop variants that are in PMBB data but not in DRIVE-KG. Changing the SNP node representation to be per alternate allele would enable us to more completely utilize genotype data during patient classification. Third, we filtered the SNP nodes at a MAF ≥ 0.01 to accommodate memory constraints from Memgraph, but this restricted our analysis to common variants. Our future work will incorporate rare variants to explore associations with higher effect sizes. For patient classification, our modest AUC improvements will require further validation in independent cohorts, and generalizability across diverse, multi-institution populations remains to be established. Lastly, we cannot yet automatically incorporate data from source databases of DRIVE-KG. Ensuring the continual refreshment of its underlying data is critical for scalable biological discovery, and we plan to address this in subsequent development. Our future work will focus on incorporating additional data modalities (e.g., transcriptomics, metabolomics, protein-protein interaction and existing genetic pathways) to further enhance the comprehensiveness of DRIVE-KG. Additionally, experimental validation of our predicted SNP-endometriosis associations will be crucial for translating these computational findings into biological insights. In conclusion, our study demonstrates that DRIVE-KG offers a promising approach to advance precision medicine in understudied diseases. By integrating multi-omics data in a unified framework, we identified previously unreported biological candidates for endometriosis (that underscore emerging hypotheses about disease mechanisms) and improved patient classification beyond traditional genomic approaches. These findings highlight the value of graph-based methods for conditions where conventional genomic tools have reached their current limits, offering a path forward to address knowledge gaps in understudied populations.

Multi-omics data integration involves combining diverse biological data types (including genomic, proteomic, epigenomic, and metabolomic data) to provide comprehensive insights into the mechanisms of diseases. 1 This approach is particularly useful for uncovering complex molecular interactions and biological pathways that remain hidden when analyzing individual data types in isolation, and leads to a more complete understanding of disease etiology and progression. 2 Despite this potential, significant computational challenges exist with multi-omics data integration. The high-dimensional, heterogeneous, and sparse nature of multi-omics data creates substantial barriers in developing analysis methods that are both biologically interpretable and computationally efficient. 3 These integration challenges highlight the need for analytical frameworks that can effectively represent the inherently complex relationships in multi-omics data. Knowledge graphs have emerged as a promising solution to address these multi-omics integration challenges. By representing biological entities as nodes and the known relationships that exist between them as edges, knowledge graphs provide a framework that can capture the complex, interconnected nature of biological systems across data modalities. 3 Graph-based approaches are particularly well-suited for large-scale, multi-omics data because they can naturally accommodate heterogeneous data types while preserving their semantic relationships and biological context. For multi-omics integration, a heterogeneous knowledge graph representation offers particular advantages. 4 Unlike homogeneous graphs where all nodes and edges represent the same type of entity or relationship, heterogeneous knowledge graphs can simultaneously represent multiple entity types (e.g., genes, proteins, metabolites) and diverse relationship types (e.g., regulatory interactions, protein-protein interactions, metabolic pathways) within a single structure. 3 Another advantage of knowledge graph-based approaches for multi-omics integration is that they do not require data from the same individuals across different -omics layers. Instead, these methods can leverage population-level patterns and biological knowledge to create comprehensive representations that integrate information from disparate studies and cohorts. Existing biological knowledge encoded in the graph structure can then be leveraged via network analyses for a variety of purposes: making use of existing relationships in the graph to predict new connections, or for downstream tasks such as patient-level classification. For many well-characterized complex diseases like obesity, extensive genetic studies have successfully explained a substantial proportion of disease heritability. Furthermore, polygenic risk scores (PRS) demonstrate strong predictive performance; for obesity, typical AUC values for PRS are around 0.8. 5 These genetic studies are enabled by large-scale biobanks (e.g., UK Biobank, 6 Penn Medicine BioBank, 7 and All of Us 8 ) which are collections of de-identified genetic, lifestyle, clinical and biological data. In contrast, endometriosis—one of many complex gynecological conditions—affects approximately 10% of reproductive age women yet remains severely understudied and underfunded. 9 Endometriosis is characterized by the presence of endometrial-like tissue outside the uterus and can cause chronic pain symptoms along with comorbidities including infertility. 10 Due to its underdiagnosis, traditional genomic approaches fall short as there is a much lower prevalence observed in biobanks. Despite a broad sense heritability estimation from a twin study of 47%, 11 variants identified through GWAS studies of endometriosis to-date ( N = 928, 413; n cases = 44, 125) 12 have only been able to explain roughly 11% of this heritability, and the predictive accuracy of genetic risk scores have been inconsistent as a result. 13 The ability of knowledge graph methods to integrate information across disparate studies and populations, without requiring matched individual-level data, makes them especially suited to address these gaps. While previous studies have applied graph representations to integrate multi-omics data for disease discovery 14 and prediction, 15 we are not aware of effective applications of these methods to complex and understudied conditions such as those in women’s health. By leveraging existing multi-omics data from independent studies and incorporating established biological relationships contributing to disease, these approaches can potentially uncover disease mechanisms and improve predictive models for conditions where traditional large-scale genomic studies have been insufficient. This study presents an innovative framework for integrating multi-omics data about disease associations through a heterogeneous knowledge graph representation. We hypothesize that by systematically incorporating knowledge from diverse -omics data sources within a unified graph structure, we can identify previously hidden candidate relationships that are both biologically meaningful and mechanistically relevant to disease pathogenesis. Combining large-scale biological data in this way will improve our ability to discover meaningful and interpretable insights about disease pathogenesis. This is particularly impactful for historically understudied women’s health conditions where traditional genomic methods have not fully captured the complexity of disease etiology.