Genetically supported drug target prioritization for rare diseases

Abstract

RareGPS is a machine-learning framework prioritizing drug targets for rare and uncommon diseases, integrating 11 genetic, clinical, and experimental evidence sources. It uses the full distribution of genetic associations across allele-frequency bins in an allelic-series model. Across 161 phenotypes, RareGPS outperforms existing resources for predicting drug indications and clinical trial progression; top 1% targets show 58-fold higher likelihood of advancing from nonindicated to phase IV and 8-fold from phase I to IV versus the middle 50%. We validated RareGPS using prescriptome analyses in two million patients and an independent literature evaluation tool (AMELIE). We publish predictions for 3,021,965 gene-phenotype pairs.

Supplementary Information

The online version contains supplementary material available at 10.1186/s13073-026-01671-5.

Background

Individuals with rare (< 0.05% prevalence) and uncommon diseases (< 0.2% prevalence) undergo long diagnostic delays due to non-specific and overlapping symptoms, unknown genetic etiologies, and limited physician expertise [1]. Drug discovery for rare and uncommon diseases is regularly characterized by distinct financial and mechanistic challenges compared to that for common diseases [2]. Recouping development costs is difficult due to limited patient populations. Many rare disease targets are not amenable to conventional small-molecule drugs, and instead require costly approaches such as monoclonal antibodies, protein replacement therapies, or cell and gene therapies [3]. The 1983 Orphan Drug Act incentivized drug development for rare diseases, and by 2022, drugs for rare diseases accounted for 49% of novel drugs and biologics approved by the FDA [4]. However, nearly 40% of orphan drug designations and approvals target rare cancers [4, 5], leaving many non-cancer rare diseases with few or no approved therapies. Translating basic research on rare diseases into therapeutic interventions remains slow due to limited biological and clinical knowledge [1, 3].

Most rare diseases are heritable, including 71.9% of diseases in Orphanet [6]. Many result from mutations in a single gene or a small set of genes following Mendelian inheritance. Mendelian genes were historically identified through pedigree studies and case series; however, this approach frequently fails to yield conclusive findings [7]. Recent biobank-scale association studies have identified additional genetic causes [8–10], but only for the few rare diseases with sufficient sample sizes. As a result, the genetic causes of many rare and uncommon diseases remain unknown [11], hampering therapeutic development by impairing target identification and the creation of representative preclinical models. Nevertheless, targeting causal genes represents the most straightforward therapeutic strategy. Drug targets with genetic support are at least twice as likely to progress from phase I to market launch [12], with similar findings for both orphan and non-orphan drugs.

For 399 common drug indications, we previously developed a genetic priority score (GPS) which, by integrating clinical genetics and genetic associations, identified genes with up to an 11-fold higher likelihood of having drug indications and an 8.8-fold higher likelihood of progressing from phase I to phase IV in clinical trials [13]. We subsequently showed that a machine learning-assisted GPS incorporating genetic associations from predicted phenotypes accurately identified genes with an even higher likelihood of drug indications for 112 common chronic diseases [14].

There is an acute need for in silico approaches to therapeutic development for rare and uncommon diseases. However, the feasibility of developing a genetics-based score for this disproportionately understudied and undertreated set of diseases is uncertain. Beyond Mendelian genes, the extent to which genetic evidence supports drug indications for rare diseases is unclear. This is due to the relative paucity of reported Bonferroni-significant common variants and even fewer rare variants for rare diseases [15, 16]. Even with sample sizes ranging from hundreds of thousands to millions, population-based biobanks still lack sufficient statistical power to reliably detect bona fide rare and common variant associations for rare diseases. Hence, novel genetic approaches are required to prioritize drug targets specifically for rare diseases.

Although genetic association studies use strict thresholds to identify associations as statistically significant, prior common variant studies showed that relaxing p-value thresholds can yield additional biologically relevant loci [17, 18]. We similarly propose that drug target identification and prioritization for rare diseases can overcome the statistical power limitations of current genetic studies by incorporating genetic evidence from the full range of genetic associations, including sub-Bonferroni-significant associations and those from ultrarare, rare, and common allele frequency bins within an allelic series framework [19]. An allelic series, composed of variants in a gene that independently produce graded effects on disease, reflects dose-response relationships between target function and phenotype and supports the validity of a target [19, 20]. We further propose enhancing drug target prioritization for rare diseases by integrating other lines of relevant evidence including clinical genetics, mouse models, gene expression data, text mining of experimental literature [21], and a graph neural network capturing gene-disease relationships [22].

Applying this principle, we construct a rare disease genetic priority score (RareGPS) for 64 rare and 97 uncommon disease phenotypes represented using phecodes. We included uncommon phecodes in this study because, like rare phecodes, they are understudied and have fewer available treatments compared to common diseases. Using a gradient boosting framework, RareGPS combines the above sources of evidence and consistently outperforms existing resources in predicting both drug indications and clinical trial advancement (Additional file 1: Fig. S1). We validate RareGPS in multiple ways. Following previous prescriptome analyses [23–25], we analyze electronic health records (EHRs) of two million patients in a major healthcare system and show that RareGPS-prioritized drug mechanisms are differentially prescribed among cases compared to controls, with many mechanisms reflecting off-label use. Furthermore, we validate RareGPS using AMELIE, an automated literature evaluation tool that prioritizes genes for Mendelian diseases [26]. Finally, we generate RareGPS predictions for 3,021,965 gene-phecode (G-P) pairs representing 19,345 protein-coding genes and all 161 phecodes and provide examples of RareGPS applications in drug development and repurposing.

Methods

Definition and selection of phecodes

We represented diseases using phecodeX to avoid redundancy and overlap in phenotyping [27]. Phecodes have four granularity levels: level 0 (e.g., CV_416 Cardiac arrhythmia), level 1 (e.g., CV_416.2 Atrial fibrillation and flutter), level 2 (e.g., CV_416.21 Atrial fibrillation), and level 3 (e.g., CV_416.211 Paroxysmal atrial fibrillation). We considered all phecodes except those in the “Infections,” “Neonatal,” “Neoplasms,” “Pregnancy,” and “Symptoms” categories and any others caused primarily by external factors (diet, iatrogenic factors, medications, and trauma). We excluded neoplasms due to anti-neoplastic drugs being shared between different neoplasms. To avoid overlap, for each top-level phecode and its descendants, we retained the least granular phecode representing a distinct disease process, which we subjectively defined as either a single disease or a set of diseases with shared etiology, presentation, and/or treatment. In most cases, this was the level 1 phecode (Additional file 1: Fig. S2a-b).

Because we performed genetic association testing in the UK Biobank, we initially assessed the proportion of cases for these phecodes among 228,516 of 501,289 UK Biobank participants with both inpatient and outpatient diagnosis data, as well as the total number of cases among all 501,289 participants (Additional file 2: Table S1). We identified cases using diagnostic codes from inpatient diagnoses (category 2000), primary care diagnoses (category 3000), and the death registry (category 100093). Based on observed case proportions, we selected 46 common phecodes (≥ 2%), 97 uncommon phecodes (0.05% to < 2%), and 64 rare phecodes (< 0.05% with ≥ 50 cases) for primary analysis (Additional file 2: Table S2). We selected a minimum threshold of 50 cases following the approach of FinnGen releases 10–12.

Drug indication and mechanism data

We identified drug indications from Citeline Pharmaprojects, using data provided by Minikel et al. [12]; Open Targets Platform (OTP) [21]; and orphan drug designations from the Food and Drug Administration (FDA) and the European Medicines Agency. For Citeline Pharmaprojects data (provided as gene-indication pairs), we mapped indications (provided as MeSH terms) first to Human Phenotype Ontology (HPO) codes using the UMLS Metathesaurus (release 2025AB), and second to phecodes using an HPO to phecodeX map [28]. No individual drug information was available for these data. For OTP data (provided as drug-gene-indication triplets), we mapped indications (provided in several different ontologies) first to either International Classification of Diseases, 10th Revision (ICD-10) codes or HPO codes using mapping files provided by OTP, and second to phecodeX using either an ICD-10 to phecodeX map or an HPO to phecodeX map, respectively. For orphan drug designations (provided as drug-indication pairs), we manually mapped indications to phecodeX (Additional file 2: Table S3). We mapped drugs by name to ChEMBL entries to determine genes targeted by each drug. Clinical trial phases (preclinical to IV) were available for Citeline Pharmaprojects and OTP data. For orphan drug designations, we assigned G-P pairs for approved drugs to phase IV and all other drugs (where clinical trial phase was not available) to preclinical. Finally, we merged all data sources and retained the highest clinical trial phase for each G-P pair (Additional file 1: Fig. S2c-d). We randomly selected 100 OTP indication-phecodeX mappings for manual verification and determined that 90 were valid; of the 10 mismatches, 8 were due to phecodeX or OTP errors, whereas 2 were due to mapping errors (Additional file 2: Table S4).

Target-disease evidence and feature selection

We obtained target-disease evidence from OTP (version 25.09), HGMD Professional (version 2023.3), OMIM (accessed November 14, 2025), Mantis-ML (version 2), and AMELIE (version 3.1.0). OTP includes 22 evidence sources: 9 germline genetics sources [Locus-to-Gene (L2G), ClinVar, Gene Burden, Genomics England PanelApp, Gene2Phenotype, UniProt literature, UniProt variants, ClinGen, Orphanet], 7 systems biology sources (Reactome, CRISPR screens, Project Score, SLAPenrich, Gene signatures, PROGENy, Cancer Biomarkers), 3 somatic variant sources [Cancer Gene Census, IntOGen, ClinVar (somatic)], differential gene expression (from Expression Atlas), text mining (from Europe PMC), and animal models (from IMPC). OTP assigns each evidence source a score from 0 to 1 based on the strength of the association. For HGMD, we converted variants to gene-level scores from 0 to 1 using similar methodology as OTP scoring for ClinVar variants: we mapped variant classifications to scores (DM: 0.9, DM?: 0.7, DFP: 0.5, DP: 0.3, FP: 0.1), calculated a harmonic sum across all variants in each gene, and scaled these scores to a maximum of 1. Both Mantis-ML and AMELIE used HPO codes to define diseases and we mapped these to phecodes using an HPO to phecodeX map.

Except for AMELIE, which we used for independent validation, we considered all these evidence sources as features for RareGPS. Although two features (text mining and Mantis-ML) do not strictly represent genetic evidence, we included them in RareGPS because both features were intended by their respective creators to be used for drug target prioritization and because the data represented in these features (e.g., published experimental evidence and biological pathways) support the strength of each G-P relationship. Due to high inter-feature correlations (Additional file 1: Fig. S3), we collapsed the seven OTP clinical genetics sources (ClinVar, Genomics England PanelApp, Gene2Phenotype, UniProt literature, UniProt variants, ClinGen, Orphanet) into a single feature (“OTP clinical genetics”) by calculating a harmonic sum across sources and scaling to a maximum of 1. To avoid overfitting and poor generalization, we removed features with fewer than 100 non-zero values and/or that were non-zero for fewer than 20 drug indications, including OMIM, all systems biology sources, and all somatic variant sources.

Genetic association testing in the UK Biobank

We performed genetic association testing for all uncommon and rare phecodes in the UK Biobank, which contains deep genetic and phenotypic data for approximately 500,000 volunteers from across the United Kingdom [29]. We retained 485,448 participants (Additional file 2: Table S1), after excluding participants with chromosomal sex discordant with self-reported sex (fields 22001 and 31, respectively), presence of sex chromosome aneuploidy (field 22019), outliers for heterozygosity or missingness (field 22027), and/or ten or more third-degree relatives (field 22021). For three phecodes that were sex-specific (one male-specific and two female-specific), we performed association testing separately among 222,208 male or 263,240 female participants. We performed power calculations for realistic variant minor allele frequencies (MAF) and effect sizes using the genpwr R package (version 1.0.4).

We treated all phecodes as binary phenotypes and included age at enrollment, sex, and 10 principal components of ancestry (field 22009) as covariates. Following previous pooled ancestry approaches for common and rare variant studies [30–32], and because regenie controls for population structure [33], we pooled individuals from all ancestries in each cohort into a single model to maximize power. However, most participants, including cases, had European ancestry (Additional file 2: Table S5). We tested associations using regenie (version 3.3) [33]; step 1 fits a whole genome model using a subset of available genetic markers, while step 2 tests a larger set of markers for association with each trait. For step 1, which is the same for all subsequent analyses, we used unimputed genotype data to generate ridge regression predictions on blocks of 2,000 single-nucleotide variants (SNV). We filtered genotype data for variants with minor allele count (MAC) > 100, MAF ≥ 0.01, genotyping rate > 0.9, and Hardy-Weinberg exact test p-value ≥ 1 × 10^− 15 using PLINK (version 2.0) [34].

We performed step 2 separately for each of four analyses: (1) single-variant testing of common variants, (2) single-variant testing of rare and ultrarare coding variants, (3) gene-level testing of rare deleterious coding variants; and (4) gene-level testing of ultrarare deleterious coding variants. In analyses 1 and 2, we performed single-variant testing of blocks of 500 SNVs using Firth logistic regression. For analysis 1, we tested variants with MAF > 0.01, genotyping rate > 0.9, and Hardy-Weinberg exact test p-value ≥ 1 × 10^− 15 from Haplotype Reference Consortium-imputed genotype data and used p < 5 × 10^− 8 to define genome-wide significant single variant associations. For analysis 2, using exome sequencing data, we tested variants with MAC ≥ 6, MAF < 0.01, genotyping rate > 0.9, and Hardy-Weinberg exact test p-value ≥ 1 × 10^− 15 that were predicted by Ensembl variant effect predictor tool (VEP; version 112) as either protein-truncating variants (PTVs; including “transcript ablation,” “splice acceptor,” “splice donor,” “stop gained,” and “frameshift” consequences) or missense variants (including “missense,” “stop lost,” “start lost,” “transcript amplification,” “inframe insertion,” “inframe deletion,” and “protein altering” consequences) in all MANE Select transcripts for protein-coding genes. Following the approach of Sveinbjornsson et al. [35], we used p < 4.3 × 10^− 7 to define exome-wide significant single variant associations. For analyses 3 and 4, using exome sequencing data, we performed gene-level testing separately for aggregated rare (0.0001 ≤ MAF < 0.01; analysis 3) and ultrarare (MAF < 0.0001; analysis 4) variants that Ensembl VEP predicted as either PTVs or deleterious missense variants, requiring a minimum cumulative MAC ≥ 6 per gene. We defined deleterious missense variants as those predicted to be deleterious or protein intolerant by each of PolyPhen-2 HumVAR, PolyPhen-2 HumDIV, Sorting Intolerant from Tolerant, Likelihood Ratio Test, and MutationTaster. We then performed standard burden tests (BURDEN), sequence kernel association tests (SKAT), optimal unified SKAT (SKATO), and aggregated Cauchy association tests (ACAT) using regenie. We used a Bonferroni-corrected p-value threshold (0.05/number of genes tested, or 0.05/18,520) to define significant gene-level associations.

For analysis 1, to identify linkage disequilibrium (LD)-independent variants, we performed LD clumping using PLINK 2.0 (release 2024-03-02) with a significance threshold of 0.05, r² threshold of 0.1, and a distance of 1 Mb. We mapped these variants to genes using a three-stage approach. For missense and protein-truncating variants, we used the gene assignment from Ensembl VEP. For all other variants, we used the highest scoring gene assignment from the Open Targets Variant-to-Gene pipeline (release 2022-10-06); this pipeline combines expression and protein quantitative trait loci datasets, chromatin interaction and conformation datasets, functional predictions, and distance from the canonical transcript start site [36]. For remaining variants not present in the Variant-to-Gene pipeline, we assigned the closest gene based on distance from the canonical transcript start site, as this is often the causal gene [37–39]. We also calculated gene prioritization scores using MAGMA (version 1.08) and PoPS (release 2022-04-11) but found neither score significantly improved RareGPS performance beyond the above approach [37, 40], and hence did not include them as features.

We included associations from the four analyses as separate features in RareGPS. We encoded all features using -log₁₀(p-values). For analyses 1 and 2, we used the -log₁₀(p-value) corresponding to the most significant association of all variants mapped to each gene. Including the number of Bonferroni-significant associations per gene did not significantly improve RareGPS performance and we did not include them as features.

Drug indication and clinical trial progression outcomes

This study’s primary outcome was drug indication (i.e., whether a G-P pair had an indicated drug in any clinical trial phase). We also examined three secondary outcomes: progression from non-indicated to phase I, from non-indicated to phase IV, and from phase I to phase IV. For progression from non-indicated to phase I, we labeled G-P pairs that reached phase I or higher as successes and non-indicated pairs and those that did not advance beyond the preclinical phase without active development as failures. For progression from non-indicated to phase IV, we labeled G-P pairs that reached phase IV as successes and non-indicated pairs and those that did not advance beyond phase III without active development as failures. For progression from phase I to phase IV, we labeled G-P pairs that reached phase IV as successes and those that reached at least phase I but did not advance beyond phase III without active development as failures.

Machine learning models

We trained all machine learning models to predict drug indications using the XGBoost Python package (version 2.1.1). We included 11 features: 6 representing existing evidence [gene expression, HGMD, L2G, mouse models, OTP clinical genetics, text mining], 4 representing genetic associations [common variants, rare variants, rare variants (gene-level), ultrarare variants (gene-level)], and Mantis-ML. To enable prioritization of G-P pairs in more advanced clinical trial phases, we assigned G-P pairs a score from 0 to 1 based on estimates of clinical trial success rates for non-oncology, non-infectious disease drugs [41]. Specifically, we assigned phase IV G-P pairs a score of 1, phase III a score of 0.732, phase II a score of 0.732 × 0.548 = 0.401, phase I a score of 0.401 × 0.580 = 0.233, preclinical a score of 0.233 × 0.911 = 0.212 (with 0.911 estimated from this paper), and non-indicated a score of 0. We then trained models using a “reg: logistic” scheme where models aim to minimize the root mean squared error (RMSE) between prediction probabilities and these scores.

For main models, we used five-fold nested cross-validation to train and evaluate models where G-P pairs were randomly split into five outer folds. Each outer fold served as a holdout set once for the other four folds, which we combined and further randomly divided into five inner folds. We used four of these inner folds for training and the fifth fold as a validation set for early stopping after 10 rounds of no improvement in RMSE. This process resulted in 5 × 5 = 25 iterations per model. In each iteration, the model generated predictions for the holdout set and for protein-coding genes not included in the training dataset. To avoid overfitting, we used default hyperparameters except for min_child_weight = 10. We calculated feature importances using the SHapley Additive exPlanations (SHAP) Python package (version 0.46.0) when generating predictions for the holdout set.

As sensitivity analyses, we used three alternative cross-validation strategies instead of random splits: gene-based five-fold cross-validation (partitioning genes into buckets), leave-one-phecode-category-out cross-validation (13 categories), and data source holdout validation (training on Citeline versus all other data).

Differential drug prescription patterns in the Mount Sinai Data Warehouse

Mount Sinai Data Warehouse (MSDW) contains more than 87 million deidentified, Epic-derived EHRs from more than 11 million patients seeking care in the Mount Sinai Health System, consisting of six hospitals in the New York City area [42]. MSDW adheres to the Observational Medical Outcomes Partnership (OMOP) standard data model. Because MSDW does not contain EHRs from external hospitals, we analyzed up to 2,027,074 of these patients with records of longitudinal care (Additional file 2: Table S1). Defining “clinical event” as any healthcare encounter, prescription, diagnosis, or procedure, we included only participants who met the following criteria: (1) had at least one outpatient visit; (2) had a clinical event on or after November 1, 2020 (five years prior to the analysis date); (3) had been followed at Mount Sinai (i.e., years between first and last clinical events) for at least one year; (4) were 18 years of age or older as of their last clinical event; (5) had been prescribed at least one medication; and (6) had at least one recorded diagnosis. For each drug mechanism, we further excluded participants whose last clinical event was before or less than one year after the first system-wide prescription date.

MSDW represents diagnoses using SNOMED, which we mapped to ICD-10 codes using a SNOMED CT to ICD-10 map and then to phecodeX. Medications in MSDW have both RxNorm IDs and generic names. We converted RxNorm IDs to DrugBank IDs using the UMLS Metathesaurus and subsequently to ChEMBL IDs using the PubChem Identifier Exchange Service. We simultaneously matched generic names to those in DrugBank and ChEMBL. We then used both DrugBank and ChEMBL to assign indications and mechanisms to medications in MSDW. We excluded GU_617.1 (cervical incompetence) from this analysis as we did not have reliable reproductive history data.

For each phecode and drug mechanism, we performed logistic regression using the statsmodels Python package (version 0.14.4). We defined any prescription of the drug mechanism as the dependent variable and included six independent variables: phecode diagnosis, age at last clinical event, sex, years followed at Mount Sinai, years of potential exposure to drug mechanism, and history of any surgical procedure. We defined potential exposure as any period occurring after both the first recorded system-wide prescription date of the drug mechanism and the patient’s first clinical event and before the patient’s last clinical event. To ensure stable estimates, we analyzed only drug mechanisms with at least 5 prescriptions among cases and 10 total prescriptions among all participants, and discarded estimates if logistic regression did not converge within 100 iterations. For each phecode, we analyzed all approved drug mechanisms, all non-approved prioritized drug mechanisms (i.e., mechanisms with at least one target with RareGPS ≥ 95th percentile), and 50 randomly selected non-approved non-prioritized drug mechanisms (i.e., mechanisms where all targets had RareGPS < 80th percentile).

To calculate the expected proportion of prescriptions occurring after disease diagnosis for each drug mechanism, we first calculated the years of potential exposure to the drug mechanism both before and after disease diagnosis for each patient with disease who had been prescribed the drug mechanism. We averaged these time periods across all eligible patients and divided the average years of possible exposure after diagnosis by the average total years of possible exposure. We then performed a binomial test to assess whether the observed proportion of prescriptions occurring after diagnosis significantly differed from the expected proportion.

Statistical analyses

We defined statistical significance as p < 0.05 unless otherwise specified. We estimated confidence intervals for proportions using the Wilson score interval method as implemented in statsmodels. We calculated machine learning metrics using the scikit-learn Python package (version 1.5.2). We estimated confidence intervals for these metrics through bootstrap resampling with 2,000 iterations. We performed logistic regression to compare G-P predictions in the top X percentiles to those in the 25th -75th percentiles using statsmodels, including phecode categories as covariates.

Results

Rare and uncommon diseases have fewer drug indications and less supporting evidence than common diseases

To examine the extent to which rare diseases are understudied and undertreated, we compared the availability of supporting evidence and number of drug indications for rare and uncommon versus common diseases. Among UK Biobank participants with both inpatient and outpatient diagnosis data (Additional file 2: Table S1), we identified 64 rare phecodes (observed case proportion < 0.05% with at least 50 total cases), 97 uncommon phecodes (0.05% ≤ observed case proportion < 0.2%), and 46 common phecodes (observed case proportion ≥ 2%) representing distinct disease processes (Additional file 2: Table S2). Of the 161 rare and uncommon phecodes, 87 were not included in a prior study of genetics-supported drug development [12], and among demographically diverse participants from an independent hospital system (Additional file 2: Table S1), 150 and 92 of the rare and uncommon phecodes had observed case proportions < 0.2% and < 0.05%, respectively (Additional file 2: Table S6). Rare diseases are also defined using Orphanet nomenclature, and the 161 uncommon and rare phecodes corresponded to 614 distinct Orphanet entries, with 94 phecodes corresponding to at least one entry (Additional file 2: Table S7). Further, 87% of uncommon and 70% of rare phecodes were at the first or second of four phecodeX granularity levels (Additional file 1: Fig. S2a), suggesting they are not infrequent solely due to granularity of definition. Rare, uncommon, and common phecodes varied in their distribution across different categories (Additional file 1: Fig. S2b).

From four data sources (Methods), we identified 1,932, 3,980, and 7,740 G-P pairs with drug indications (i.e., “indicated G-P pairs”) for rare, uncommon, and common phecodes, respectively. There was at least one indicated G-P pair for 42 (66%) of the 64 rare phecodes, 70 (72%) of the 97 uncommon phecodes, and all 46 common phecodes (Fig. 1a). Across all clinical trial phases, there were significantly fewer indicated G-P pairs for uncommon and rare phecodes compared to common phecodes (Fig. 1b).

Fig. 1.

Comparison of drug indications and their supporting evidence for common, uncommon, and rare phecodes. A Number of indicated G-P pairs per phecode. The x-axis is in log₂ scale. B Violin plots showing the distribution of the number of indicated G-P pairs in each clinical trial phase per phecode. Medians are indicated by white horizontal lines and are labeled below each violin plot. Comparing common versus uncommon/rare, p_rank−sum = 0.02 (I), 2 × 10^− 3 (II), 5 × 10^− 4 (III), and 1 × 10^− 9 (IV). C Proportion of phecodes for which at least one gene had supporting evidence from each evidence category. Comparing common versus uncommon/rare, p_z−test = 0.05 (clinical genetics), 4 × 10^− 14 (gene burden), 4 × 10^− 6 (gene expression), 7 × 10^− 8 (L2G), 0.47 (mouse models), 0.01 (somatic variants), and 0.45 (systems biology). D Proportions of indicated G-P pairs with supporting evidence from each evidence category. Comparing common versus uncommon/rare, p_z−test = 2 × 10^− 23 (clinical genetics), 0.05 (gene burden), 5 × 10^− 16 (gene expression), 8 × 10^− 126 (L2G), 3 × 10^− 6 (mouse models), 0.34 (somatic variants), and 1 × 10^− 5 (systems biology). For C-D, all evidence was sourced from OTP except for “Clinical genetics,” which includes both OTP evidence as well as HGMD and OMIM. Error bars represent 95% confidence intervals calculated using the Wilson method

We examined the supporting genetic evidence for rare and uncommon phecodes compared to common phecodes. We obtained drug prioritization resources from the Open Targets Platform (OTP), which aggregates clinical genetics, genetic association, differential gene expression, mouse model, somatic variant, and systems biology data from 22 sources, as well as from the Human Gene Mutation Database (HGMD) and Online Mendelian Inheritance in Man (OMIM) [43, 44]. For all sources of evidence except mouse models and systems biology, a significantly smaller proportion of uncommon and rare phecodes compared to common phecodes had at least one G-P pair with association evidence (Fig. 1c). The greatest differences were for gene burden (representing ultrarare coding variant associations from 11 datasets) and L2G (representing common variant associations from OTP). Additionally, a significantly smaller proportion of indicated G-P pairs for uncommon and rare phecodes received supporting evidence from clinical genetics, gene expression, and L2G compared to indicated G-P pairs for common phecodes (Fig. 1d), and 81.7%, 90.1%, and 90.0% of indicated G-P pairs for common, uncommon, and rare phecodes lacked any supporting evidence, respectively. While a greater proportion of indicated versus non-indicated G-P pairs had supporting evidence for both common and uncommon/rare phecodes (Additional file 1: Fig. S4a-b), there was an inconsistent trend of increasing evidence with more advanced clinical trial phases.

The full distribution of genetic associations is predictive of drug indications for uncommon and rare phecodes

With only 93 of the 5,912 indicated G-P pairs for uncommon and rare phecodes having supporting L2G or gene burden evidence from OTP, which include only Bonferroni-significant associations, we performed separate genetic association testing in the UK Biobank. Power analyses indicated a limited ability to detect Bonferroni-significant single variant associations except at very large effect sizes (Additional file 1: Fig. S5a), although power was sufficient to identify nominally significant associations (Additional file 1: Fig. S5b), and we performed gene-level testing for rare and ultrarare variants to increase power.

We identified 451 G-P pairs with at least one Bonferroni-significant association, implicating 399 distinct genes for 139 distinct rare and uncommon phecodes (Additional file 2: Table S8). We assessed whether these 451 associations were corroborated by clinical genetics, prior genetic associations, and two machine learning methods: Mantis-ML and AMELIE [22, 26]. There was supporting evidence for 26 of the 112 G-P pairs with significant common variants, 97 of the 271 G-P pairs with significant rare variants, 24 of the 52 G-P pairs with significant rare variant gene-level tests, and 43 of the 51 G-P pairs with significant ultrarare variant gene-level tests (Fig. 2a). Notably, both single variant and gene-level testing of rare variants identified Bonferroni-significant associations between IFT140 and GE_976.5 (polycystic kidney disease; n_case = 635) and between FECH and GE_966.2 (disorders of porphyrin metabolism; n_case = 77), supporting our ability to identify even infrequent causes of rare disorders [10, 45].

Fig. 2.

Common, rare, and ultrarare variant genetic associations for 161 rare and uncommon phecodes in the UK Biobank. A G-P pairs with Bonferroni-significant associations, and of those, the number with supporting evidence from different sources. B Drug indication odds ratios for G-P pairs with p-values below specified thresholds compared to G-P pairs with p > 0.05. C Drug indication odds ratios for G-P pairs with p-values below specified thresholds across 1, 2, or 3 allele frequency bins (common, rare, ultrarare). In B-C, numbers below each error bar indicate the number of indicated G-P pairs (top) or the total number of G-P pairs (bottom) meeting the specified thresholds, respectively. Error bars in B-C represent 95% confidence intervals. Results in B-C reflect the training dataset of 356,272 G-P pairs

However, across the 161 rare and uncommon phecodes, only 10 drug indications were supported by these Bonferroni-significant associations, which was too few to train robust models for RareGPS. Addressing this, we postulated that the full distribution of genetic association p-values would be useful for drug target prioritization even if no Bonferroni-significant associations were present. G-P pairs with sub-Bonferroni-significant common and ultrarare variant associations were significantly enriched for drug indications (Fig. 2b), and there was greater enrichment at more stringent significance thresholds. At thresholds of both p < 0.05 and p < 0.01, G-P pairs with support from two or three allele frequency bins were significantly enriched for drug indications (Fig. 2c), supporting the concept that allelic series are valuable for drug target prioritization [19].

RareGPS features

We included 11 features in RareGPS (Fig. 3a). First, we included six features representing existing evidence: differential gene expression, HGMD, L2G, mouse models, OTP clinical genetics, and text mining. Except for HGMD, which we scored separately, OTP assigned these features scores from 0 to 1 based on evidence strength. Second, we included features representing the four genetic association tests we performed in the UK Biobank, encoding each feature as the maximum -log₁₀(p-value) per gene. Third, we included raw Mantis-ML scores, which capture gene-disease relationships on a scale from 0 to 1.

Fig. 3.

Features included in RareGPS. A Features included in RareGPS and their scoring. B Number of non-indicated and indicated G-P pairs for which each feature is non-zero. The y-axis is in log₁₀ scale. C Drug indication odds ratios for G-P pairs with non-zero values for each feature compared to those with values of zero. D Drug indication odds ratios per one unit increase in -log₁₀(p-value) for each genetic association feature. Error bars in C-D represent 95% confidence intervals. All results reflect the training dataset of 356,272 G-P pairs

Our dataset contained 356,272 G-P pairs, of which 5,912 were indicated. There were 3,181 druggable genes for each of 112 phecodes with at least one indicated drug. Within this dataset, none of the 11 features were highly correlated with either other features or with drug indications (Additional file 1: Fig. S6), except for correlations between Mantis-ML and HGMD (r = 0.16), between HGMD and OTP clinical genetics (r = 0.32), and between rare single-variant and rare variant gene-level associations (r = 0.38). L2G and OTP clinical genetics were relatively sparse, with only 1,362 and 991 non-zero values, respectively (Fig. 3b). All features except the single “rare variants” feature were individually significantly associated with drug indication (Fig. 3c-d), with the greatest odds ratios (OR) for HGMD (7.26, 95% CI 6.07–8.69) and OTP clinical genetics (10.41, 95% CI 8.73–12.40).

RareGPS outperforms existing resources in predicting drug indications and clinical trial success for rare and uncommon diseases

We trained and evaluated RareGPS with gradient boosting in a five-fold nested cross-validation procedure (Methods). In holdout evaluation, RareGPS outperformed models trained using only features from each of three categories [Mantis-ML, genetic associations (“Genetic associations”), existing evidence (“Existing”)] or from both Mantis-ML and Existing, achieving an area under the receiver operating characteristic curve (AUROC) of 0.70 (95% CI 0.70–0.70) and area under the precision-recall curve (AUPRC) of 0.06 (0.06–0.06) (Fig. 4a). RareGPS also outperformed other models in predicting whether a G-P pair would progress from being non-indicated to phase I, from being non-indicated to phase IV, and from phase I to phase IV with AUROCs of 0.69 (95% CI 0.69–0.70), 0.81 (0.78–0.83), and 0.67 (0.64–0.69), respectively (Additional file 1: Fig. S7a-b).

Fig. 4.

Performance metrics for RareGPS and comparison models. A Left panel shows areas under the receiver operating characteristic curve (AUROC) of models for predicting drug indications in holdout (n = 356,272 with 5,912 indications) evaluation. Right panel shows areas under the precision-recall curve (AUPRC) of models for predicting drug indications in holdout, external, or combined evaluation. Gray lines represent the proportion of indicated G-P pairs in each dataset. B Drug indication odds ratios in the combined set for G-P pairs in top percentile bins compared to those in the 25th -75th percentiles. C Secondary outcome odds ratios for G-P pairs above the 99th percentile compared to those in the 25th-75th percentiles. In A-C, “Existing” includes the following features: gene expression, HGMD, L2G, mouse models, OTP clinical genetics, and text mining. Error bars represent 95% confidence intervals

RareGPS may be especially useful for target selection at its extremes. Compared to G-P pairs between the 25th and 75th percentiles, G-P pairs at or above the 99th percentile (n = 3,563) were 11.86-fold more likely to have a drug indication (95% CI 10.57–13.30), 13.51-fold more likely to progress from being non-indicated to phase I (95% CI 12.03–15.17), 58.30-fold more likely to progress from being non-indicated to phase IV (95% CI 41.86–81.21), and 8.07-fold more likely to progress from phase I to phase IV (95% CI 5.34–12.21) (Fig. 4b-c). RareGPS predictions, which represent evidence for G-P associations, were also complementary to DrugnomeAI druggability [46]: among G-P pairs above the 99th percentile for RareGPS, there was greater enrichment for drug indications with increasing predicted druggability (Additional file 1: Fig. S8), and G-P pairs with DrugnomeAI scores greater than 0.75 had an OR of 20.29 (95% CI 17.57–23.44).

Feature importance analyses using Shapley Additive Explanations (SHAP) indicated that the most important features were text mining, ultrarare variant gene-level associations, Mantis-ML, and rare variant associations (Additional file 1: Fig. S9a). Although G-P pairs supported by clinical genetics were highly enriched for drug indications (Fig. 3c), these features had lower overall importance due to their sparsity (Fig. 3b). There were also substantial inter-feature interactions between the four genetic association features and Mantis-ML (Additional file 1: Fig. S9b).

RareGPS is robust

We performed subset analyses to assess RareGPS robustness. First, because recent drug programs may be influenced by genetic associations, we compared RareGPS performance between targets first reaching preclinical or phase I before 2005, between 2005 and 2015, and after 2015 and observed AUROCs of 0.72 (95% CI 0.69–0.74), 0.71 (0.69–0.72), and 0.66 (0.64–0.67), respectively (Additional file 1: Fig. S10a-b). Second, RareGPS AUROCs among each of the 13 phecode categories were at or above the overall AUROC except for the blood/immune (0.62, 95% CI 0.58–0.66), gastrointestinal (0.67, 0.63–0.70), genitourinary (0.63, 0.61–0.66), and respiratory (0.56, 0.47–0.67) categories (Additional file 1: Fig. S10c-d). Third, RareGPS had similar performance among 42, 39, and 31 phecodes with UK Biobank observed case proportions of < 0.0005, 0.0005–0.001, and > 0.001, with AUROCs of 0.70 (95% CI 0.69–0.71), 0.71 (0.70–0.72), and 0.69 (0.68–0.70), respectively (Additional file 1: Fig. S10e-f). Notably, models using only genetic associations had similar performance across the three bins, suggesting genetic associations for phecodes with 250 or fewer cases are still valuable for prioritizing drug targets.

We also evaluated RareGPS features with three complementary holdout schemes. First, we used gene-based cross-validation, training models on four-fifths of genes and testing on the remaining fifth. RareGPS achieved an overall AUROC of 0.70 (95% CI 0.69–0.71) across five gene-based folds (Additional file 1: Fig. S11a), similar to the main model’s AUROC (0.70). Second, we used leave-one-phecode-category-out cross-validation, training models on 12 phecode categories and testing on the held-out category. Overall performance across categories was AUROC 0.68 (95% CI 0.67–0.69) (Additional file 1: Fig. S11b). Third, we trained models using drug indications from either Citeline or OTP and orphan indications, and tested them on the opposite source(s). Models trained on OTP and orphan indications achieved AUROC 0.82 (95% CI 0.81–0.83) when tested on Citeline indications (Additional file 1: Fig. S11c), whereas the AUROC was 0.60 (95% CI 0.59–0.61) for the reverse, likely due to the limited coverage of Citeline compared to OTP. These analyses demonstrate that RareGPS features generalize well to unseen genes, disease categories, and data sources.

Drugs targeting RareGPS-prioritized genes have distinct prescription patterns

We examined whether non-approved drugs targeting RareGPS-prioritized genes (≥ 95th percentile), non-approved drugs targeting non-prioritized genes (< 80th percentile), and approved drugs were differentially prescribed among cases compared with controls in two million Mount Sinai patients (Additional file 2: Table S1). To increase power, we collapsed all drugs modulating the same targets in the same direction into drug mechanisms, ultimately assessing 3,983 non-prioritized, 5,685 prioritized, and 108 approved mechanisms across 160 phecodes. Omitting non-biological reasons (e.g., mapping errors and misdiagnoses), we assigned possible reasons for differential prescriptions to six categories based on whether, among cases, there was under- or overprescription and whether a greater or lesser proportion of first prescriptions occurred before or after disease diagnosis (Fig. 5a). We quantified these two metrics using ORs and the ratio of observed to expected (O/E) proportions, respectively (Fig. 5b-c).

Fig. 5.

Prescription patterns of non-prioritized, prioritized, and approved drug mechanisms. A Possible reasons for differential prescriptions stratified by odds ratios (OR) and the ratio of observed to expected (O/E) proportions of first prescriptions occurring after diagnosis. B Scatterplots showing ORs and Benjamini-Hochberg adjusted -log₁₀(p-values) for non-approved non-prioritized, non-approved prioritized, and approved drug mechanisms. Horizontal and vertical solid lines represent nominal significance (p = 0.05) and OR = 1, respectively. Left and right vertical dashed lines represent OR < 0.61 and OR > 4.45, respectively. Both axes are in log₁₀ scale. C Scatterplots showing O/E proportions and binomial test -log₁₀(p-values) for non-approved non-prioritized, non-approved prioritized, and approved drug mechanisms. Horizontal and vertical solid lines represent nominal significance (p = 0.05) and O/E = 1, respectively. The y-axis is in log₁₀ scale. D Percentage of drug mechanisms that are underprescribed (OR < 0.61, p < 0.05) or overprescribed (OR > 4.45, p < 0.05) for each RareGPS bin. Error bars represent 95% confidence intervals calculated using the Wilson method. E Number and percentage of drug mechanisms in each OR and O/E bin. In B-C, numbers in each corner represent the percentage of drug mechanisms within each bounded rectangle

To reduce false positives, we defined underprescription and overprescription using a Benjamini-Hochberg adjusted p-value < 0.05 along with OR thresholds of < 0.61 and > 4.45, respectively, which excluded 90% of non-prioritized drug mechanisms with ORs < 1 and > 1. With these thresholds, a significantly greater percentage of prioritized compared to non-prioritized drug mechanisms were overprescribed among cases, there was a monotonic increase in this percentage with increasing RareGPS percentile bins, and approved drug mechanisms had the highest percentage (74.1%) of overprescription (Fig. 5d). We also observed lower percentages of underprescription among prioritized and approved drug mechanisms, although we only analyzed mechanisms with ≥ 5 prescriptions among cases. Most drug mechanisms, regardless of approval or prioritization, had a greater proportion of prescriptions than expected after disease diagnosis (Fig. 5c and e), perhaps reflecting patients receiving higher levels of care following diagnosis or inaccuracies in the earliest recorded dates of diagnosis and prescriptions. Regardless, the presence of substantial prescriptions after diagnosis supports the safety of a drug mechanism among patients with disease.

Among the non-approved prioritized mechanisms, we examined all 3 underprescribed and the 100 overprescribed mechanisms with the largest ORs to assign possible explanations (Additional file 2: Table S9). Two underprescribed mechanisms involved contraindications (COX-2 inhibitors in diseases with renal dysfunction) and one involved drug-drug interactions with indicated drugs. Of the 100 overprescribed mechanisms, at least 51 involved possible off-label or symptomatic use, such as lumateperone for schizoaffective disorder (RareGPS 0.08, OR 86) [47], riluzole for hereditary ataxias (RareGPS 0.09, OR 70) [48], and denosumab for congenital osteodystrophies (RareGPS 0.09, OR 45) [49]. Other explanations included mechanisms treating comorbidities (n = 18; e.g., cancer drugs in common variable immunodeficiency, which increases malignancy risk), mechanisms causing disease (n = 11; e.g., dopamine agonists causing impulse disorders), mechanisms treating causes of disease (n = 7; e.g., cancer drugs in cancer-associated hypogammaglobulinemia), and possible misdiagnosis (n = 3; e.g., voxelotor, indicated for sickle cell disease, in thalassemia).

RareGPS prioritizes drug targets among all protein-coding genes

We calculated RareGPS for 19,345 protein-coding genes across 161 phecodes (3,021,965 G-P pairs with at least one non-zero feature). Here, a RareGPS threshold of 0.037 (99.6th percentile) maximized F1 = 7% with precision = 5% (proportion of G-P pairs above this threshold that are indicated) and recall = 11% (proportion of indicated G-P pairs that are above this threshold) (Fig. 6a). There were 12,275 G-P pairs above this threshold (4,658 genes; 124 phecodes), of which 10,520 and 5,989 had support from at least two or three feature categories, respectively (Additional file 1: Fig. S12a). Of the 9,934 pairs with supporting genetic associations, 5,686 had at least nominally significant associations from more than one allele frequency bin (Additional file 1: Fig. S12b). We validated these predictions with AMELIE, an automatic literature evaluation tool for Mendelian diseases; compared with all other G-P pairs, a significantly greater percentage of these top G-P pairs had any (69.0%) or strong (39.6%) support (Fig. 6b).

Fig. 6.

RareGPS predictions for all protein-coding genes. A Precision, recall, and F1 score at different RareGPS thresholds among 3,021,965 G-P pairs representing 19,345 protein-coding genes and all 161 phecodes. The vertical dashed line represents the threshold optimizing F1 score (RareGPS = 0.037). B Proportion of G-P pairs in each RareGPS bin with any (> 0) or strong (> 50) support from AMELIE, which outputs a score ranging from 0 to 100 for each publication matching each phenotype query. We retained the highest scoring publication for each query. Error bars represent 95% confidence intervals calculated using the Wilson method. C Precision@k and number needed to screen for different k values. The vertical dashed line represents the threshold optimizing F1 score (k = 12,275 at RareGPS = 0.037)

Because not all protein-coding genes are druggable, we also examined the druggability of top RareGPS predictions. The top 12,275 G-P pairs also had significantly higher DrugnomeAI-predicted druggability compared to all other G-P pairs (median 0.015 versus 0.001, p_rank−sum < 1 × 10^− 325), and 6,411 were either targeted by existing drugs or were predicted as highly druggable (DrugnomeAI > 0.5).

For practical use cases, precision@k and number needed to screen (NNS) indicate strong early precision relative to baseline (0.19%): high-confidence validation (k = 100; precision 30%; NNS 3.3; RareGPS = 0.260), moderate-throughput follow-up (k = 1,000; precision = 13.5%; NNS = 7.4; RareGPS = 0.090), and high-throughput discovery (k = 10,000; precision = 5.49%; NNS = 18.2; RareGPS = 0.039) (Fig. 6c; Additional file 2: Table S10).

Clinical use examples of RareGPS

We publish top RareGPS predictions as an interactive web application, where users can filter predictions by RareGPS, indications, druggability, and evidence from specific sources, as well as view both indicated and non-indicated drugs targeting each gene. These predictions serve several purposes, including (i) identifying opportunities for drug repurposing; (ii) identifying support for drugs in clinical trials; and (iii) identifying evidence-supported targets without current drugs.

For the first use case of identifying opportunities for drug repurposing, we examined 3,492 non-indicated prioritized G-P pairs targeted by existing drugs and highlight examples where an existing drug exists with the correct direction of therapeutic modulation. First, RareGPS supports C3 inhibitors (e.g., AL-78898 A, AMY-101, pegcetacoplan) for focal segmental glomerulosclerosis (C3; score 0.16); increased complement activation and C3 deposition are associated with increased clinical and histological severity [50, 51]. Second, RareGPS supports NF-κB inhibitors (e.g., edasalonexent) for biliary cirrhosis (NFKB1; score 0.15); prior studies showed potential benefit of inhibition in primary biliary cholangitis [52, 53]. Third, RareGPS supports denosumab for juvenile idiopathic arthritis (TNFSF11; score 0.10); denosumab reduced joint destruction in a phase 3 trial for rheumatoid arthritis [54], although increased pediatric safety data are needed [55].

For the second use case of identifying support for drugs in clinical trials, we identified 802 G-P pairs corresponding to drugs in active or recently completed trials, of which 63 have RareGPS > 0.037. The drugs with highest RareGPS included sparsentan for focal segmental glomerulosclerosis (phase 3; NCT03493685; RareGPS 0.11), tofacitinib for systemic sclerosis (phase 2; NCT06044844; RareGPS 0.09), and setrusumab and romosozumab for osteogenesis imperfecta (phase 3; NCT05768854 and NCT05972551; RareGPS 0.08). All these examples were supported by at least two different sources of evidence and at least one source of genetic association evidence.

The third use case of identifying evidence-supported targets without current drugs is causal gene identification, for which existing resources are numerous. However, RareGPS provides additional utility by representing multiple data sources as a single score, such that all genes within each data source are represented (maximizing recall), while genes supported by multiple data sources have higher RareGPS scores (allowing high precision at high thresholds). The top scoring non-indicated G-P pairs are indeed known monogenic causes of disease, including SQSTM1 for Paget’s disease of bone (RareGPS 0.51) and ANO5 for muscular dystrophy (0.50), although many of these genes are challenging to target using conventional small molecule approaches. However, further filtering on DrugnomeAI scores can help identify druggable genes; for example, ENG for hereditary hemorrhagic telangiectasia (0.36) has a DrugnomeAI score of 0.67, and several small molecule approaches to target ENG have been proposed [56,57].

Discussion

RareGPS is a machine learning framework that integrates genetic, clinical, and experimental evidence into a unified score to prioritize drug targets for uncommon and rare phecodes. In holdout and subset evaluation, RareGPS consistently outperformed existing resources in predicting drug indications and clinical trial progression. Among druggable genes, G-P pairs in the top 1% of RareGPS were 58-fold more likely to progress from non-indicated to phase IV and 8-fold more likely to progress from phase I to phase IV compared to G-P pairs in the middle 50%. Further, RareGPS predictions, which represent evidence for target-disease relationships, complement druggability metrics like DrugnomeAI, which assess the feasibility and safety of targeting specific genes.

Consistent with our prior GPS implementation for common chronic phecodes [14], here we found that including the full distribution of genetic associations as -log₁₀(p-values) improved RareGPS performance. Although this approach is inappropriate for genetic discovery due to large type 1 errors, we show that, combined with other evidence and when nominally significant associations span multiple allele frequency bins to form an allelic series, machine learning models can leverage sub-Bonferroni-significant associations for drug target prioritization.

EHR data combine millions of prescriptions, diagnoses, and clinical measurements, representing a valuable but underutilized resource to identify drug repurposing opportunities [23–25]. Here, we used differential prescription patterns among cases versus controls as indicators of the biological relevance of RareGPS-prioritized targets, where both overprescription (potentially due to off-label use or adverse effects) and underprescription (potentially due to contraindications or protective effects) could support functional roles of a target in disease. Many prioritized mechanisms with high ORs likely reflected off-label use, which may be common for rare diseases without FDA-approved treatments [58], and is consistent with estimates that off-label prescriptions constitute 10–20% of all prescriptions [59, 60]. Given that off-label use often lacks rigorous evidence and is associated with high side-effect risks [61], a combination of RareGPS prioritization and significant differential prescription rates could help identify promising drug mechanisms for targeted clinical trials to formally assess efficacy and safety.

The present implementation of RareGPS is limited to 161 phecodes, selected through a semi-subjective process and requiring ≥ 50 UK Biobank cases to enable genetic association testing. This approach excludes phenotypes that are extremely rare, cause early mortality, and/or are otherwise underrepresented in the UK Biobank population. Addressing this, we release trained models that can generate predictions for any phecode, and all RareGPS features besides genetic associations are publicly available for all phenotypes. Although we cannot guarantee accurate predictions, we showed that RareGPS has consistent performance across phecodes, and users can assess performance by comparing predictions against known drug indications from OTP. In addition to RareGPS, we also release trained Existing + Mantis-ML models, which represent RareGPS without UK Biobank genetic associations. This model enables predictions for phecodes too rare for genetic association testing without substantial performance loss. We provide an interactive tutorial in Jupyter notebook format to generate predictions for 120 rare phecodes excluded from this study and observed an AUROC of 0.83 and AUPRC of 0.06 among these phecodes. Finally, RareGPS serves as a general framework for rare disease drug target prioritization, and users can adapt RareGPS to their own datasets with alternative features or phenotype definitions.

This study has several other limitations. First, we performed genetic association testing only in the UK Biobank, which comprises > 80% European-ancestry participants. This may limit the ability of RareGPS to capture genetic variation specific to non-European ancestry populations; however, OTP data in RareGPS includes associations from diverse ancestries (e.g., from the GWAS Catalog [62]), and we performed our prescriptome analysis in a demographically diverse healthcare system. Second, the sources of evidence and drug indications we used employ different phenotype terminologies, and mapping these to phecodes may have introduced errors. However, manual verification suggested that these errors are limited in scope. Third, while RareGPS is designed to facilitate target-based drug discovery, phenotypic drug discovery may sometimes be more effective [63, 64], particularly when disease mechanisms are unknown. Additionally, we treated all targets of a drug equally when training RareGPS, when some targets may be more important for disease modulation than others. Fourth, because none of the features included in RareGPS are comprehensive, a lack of RareGPS prioritization does not necessarily indicate that a target is unsuitable for drug development.

Conclusions

RareGPS represents a machine learning framework that effectively integrates genetic associations, clinical genetics, and multiple other evidence types to prioritize drug targets for rare diseases, demonstrating superior performance in predicting both drug indications and clinical trial progression compared to existing resources. The ability of RareGPS to leverage sub-Bonferroni-significant genetic associations when combined with other evidence types, along with its validation through EHR prescription patterns, suggests it could accelerate the identification of promising therapeutic targets for rare diseases that currently lack effective treatments.

Abbreviations

AUPRC

area under the precision-recall curve

AUROC

area under the receiver operating characteristic curve

CI

confidence interval

EHR

electronic health record

G-P pair

gene-phecode pair

GWAS

genome-wide association study

GPS

Genetic Priority Score

HGMD

Human Gene Mutation Database

HPO

Human Phenotype Ontology

ICD-10

International Classification of Diseases, 10th Revision

L2G

Locus-to-Gene

LD

linkage disequilibrium

MAC

minor allele count

MAF

minor allele frequency

MSDW

Mount Sinai Data Warehouse

NNS

number needed to screen

OMIM

Online Mendelian Inheritance in Man

O/E ratio

observed/expected ratio

OR

odds ratio

OTP

Open Targets Platform

PTV

protein-truncating variant

RMSE

root mean squared error

SHAP

SHapley Additive exPlanations

SNV

single-nucleotide variant

Authors’ contributions

RC, DMJ, and RD conceived and designed the study. RC performed statistical analyses. RC, AD, MM, DNC, GR, DMJ, and RD provided administrative, technical and material support. RC and RD drafted the paper. DMJ and RD supervised the study. RC and RD had access to and verified all of the data in the study. All authors read and approved the final manuscript.

Funding

RC is supported by the National Institute of General Medical Sciences of the NIH (T32-GM007280). DMJ is supported by the National Human Genome Research Institute of the NIH (R01-HG013511). RD is supported by the National Institute of General Medical Sciences of the NIH (R35-GM124836). The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health. The funders had no role in study design, data collection and analysis, decision to publish or preparation of the manuscript.

Data availability

Top RareGPS predictions are interactively viewable at https://rstudio-connect.hpc.mssm.edu/raregps_web. All RareGPS predictions and summary statistics are available at 10.5281/zenodo.17689650 [65]. Other data sources used in this study are publicly available, including ChEMBL (https://chembl.gitbook.io/chembl-interface-documentation/downloads), DrugBank (https://go.drugbank.com/releases/latest), Mantis-ML (https://public.cgr.astrazeneca.com/mantisml/v2/download.html), Open Targets (https://platform.opentargets.org/downloads), phecodeX (https://phewascatalog.org/phewas), and the UMLS Metathesaurus (https://www.nlm.nih.gov/research/umls/licensedcontent/umlsknowledgesources.html). Access to MSDW is restricted to Mount Sinai affiliates as it contains sensitive patient data; the data use agreement and contact information are available at https://labs.icahn.mssm.edu/msdw/. Code to generate RareGPS predictions and train RareGPS models is available at https://github.com/robchiral/RareGPS/ [66]. Other software packages used in this study are publicly available, including Ensembl VEP (https://github.com/Ensembl/ensembl-vep), PLINK (https://www.cog-genomics.org/plink/2.0/), and regenie (https://github.com/rgcgithub/regenie).

Declarations

Ethics approval and consent to participate

We accessed UK Biobank data under application ID 16218. The Institutional Review Board at the Icahn School of Medicine at Mount Sinai approved Mount Sinai Data Warehouse (MSDW) access (GCO no. 07–0529; STUDY-11–01139). This study complied with the Declaration of Helsinki.

Consent for publication

Not applicable.

Competing interests

RD reports being a scientific cofounder, consultant and equity holder for Pensieve Health (pending) and being a consultant for Variant Bio, Character Bio and Cytokinetics. DNC and MM acknowledge receipt of funding from Qiagen Ltd through a License agreement with Cardiff University, which is relevant to the use of HGMD Professional in this work. AD is a current full-time employee of GSK. All other authors have no competing interests to declare.