A rare disease in the United States is defined by the 1983 Orphan Drug Act as a condition that affects fewer than 200,000 people [1], whereas the analogous legislation introduced in the European Union in 2000 considers a disease to be rare when it affects fewer than 1 in 2,000 people [2]. In comparison to common diseases, health care providers are challenged by a lack of familiarity with diagnosing and treating rare diseases, which can lead to missed, delayed, or inaccurate diagnoses even when an approved, effective therapy is available [3]. Improved understanding and recognition of rare diseases are key for accurate and timely diagnosis, and this relies on broad dissemination of and access to knowledge about rare diseases [4]. A huge amount of effort has been made to develop rare disease resources for patients, families, and clinicians, such as the Genetic and Rare Diseases Information Center (GARD) [5], Orphanet [6], and Monarch Disease Ontology (MONDO) [7]; however, disparate data and incomplete data harmonization are still major barriers to improved coordination across specialists, leading to inefficiencies and delays in diagnosis, care, and treatment. This is exemplified by the difficulty faced in accurately answering the question, how many total rare diseases are there? A recent report by Haendel et al [8], after an examination of multiple rare resources, concluded that “there are total of 10,393 rare diseases in MONDO…the majority, 6370 rare diseases, are presented in three or more resources, whereas 4023 are unique to one source.” The fact that more than one-third of rare diseases are unique to 1 source highlights a reality that those resources continue to use their own disease definitions or harmonization rules to develop their rare disease vocabularies. Insufficient effort put toward data harmonization ultimately leads to redundancy in categorization efforts and a resulting inconsistency of rare disease representation globally.

The goal of data harmonization is to improve the compatibility of data collected from independent sources (horizontally) in order to better understand disease etiology from different angles, which may forward the discovery of therapeutic approaches for rare diseases. For each individual source, data harmonization is crucial to better represent and organize data for supporting data harmonization horizontally. Current data harmonization efforts are primarily aligning standard nomenclatures or human efforts to translate specific medical and clinical features into a standardized and sharable format. For instance, Pontikos et al [9] introduced Phenooplis, an open platform for the harmonization and analysis of genetic and phenotypic data that harmonize phenotypes with the help of Human Phenotype Ontology (HPO). The International Cancer Genome Consortium (ICGC) and The Cancer Genome Atlas (TCGA) invited the cancer-genomics and bioinformatics communities to work together to identify the best pipelines for the detection of mutations in DNA-sequencing reads for cancer genomes in order to facilitate the harmonization of mutation-calling procedures among institutions [10,11]. Orphanet and OMIM (Online Mendelian Inheritance of Man) heavily relied on human efforts for their data curation and harmonization [12,13]. To avoid cumbersome human efforts and a lack of rare disease standards in this study, we proposed to systematically identify phenotypically similar rare diseases from GARD and determine their similarity types, including duplicate diseases, sibling diseases, and subtypes for supporting rare disease data harmonization.

Rare disease designations are often in conflict across different datasets due to the differing statutory requirements used in defining a rare disease in different countries, and as such, useful methods to improve interoperability across these broad terminologies and standards are required. With the aim of eliminating data redundancy and inconsistency across different resources, improving data interoperability, and facilitating data harmonization, the implementation of a knowledge graph is capable of semantically organizing and integrating complex networks of data into one collection. Knowledge graphs have been widely applied in the medical domain and in the rare disease field. For instance, Reumann et al [14] reported their solution for cognitive differential diagnosis (DDx) in rare diseases based on knowledge graph technology that incorporates data from ICD-10, DOID, medDRA, PubMed, Wikipedia, Orphanet, the CDC, and anonymized patient data. Li et al [15] presented their work to develop a rare disease classification algorithm established on a knowledge graph. Sosa et al [16] applied a knowledge graph–embedding method that explicitly models the uncertainty associated with literature-derived relationships and uses link prediction to generate drug repurposing hypotheses for rare diseases. In this study, we accessed data from an integrative knowledge graph that we developed from our previous study [17] with a variety of rare disease-related resources for phenotypical similarity identification among GARD diseases.

In this study, we report our preliminary work to identify phenotypically similar GARD diseases from an integrative knowledge graph using 2 approaches: (1) disease mapping comparison, and (2) phenotypical similarity derivation from disease classification systems. This effort will not only direct GARD data harmonization but will also support data harmonization across different resources, and eventually support clinical decision making. Phenotypically similar GARD diseases applied in this study specifically refer to disease subtypes and sibling diseases that share similar clinical manifestations. For example, 2 GARD diseases of “lactate dehydrogenase deficiency” and “lactate dehydrogenase A deficiency” are subtypes, and they have similar phenotypical profiles.

In this study, we aimed to identify phenotypical similarity among rare diseases to support data harmonization and data interoperability with existing standardized terminologies and ontologies. We designed two complementary approaches: (1) analysis of disease mappings to Orphanet, OMIM, and the UMLS to measure phenotypical similarity among GARD diseases; (2) prioritizing phenotypical similarity derived from MONDO and Orphanet disease classification systems with shared phenotypes from the HPO and genes from OMIM. The architecture of this study is shown in Figure 1.

In this study, we identified and prioritized phenotypical similarity among GARD diseases by comparing disease similarity and deriving phenotypical similarity from disease classification systems. As a proof-of-concept, we demonstrated the usefulness of the identified phenotypically similar disease pairs to support data harmonization for GARD. By incorporating these identified similar diseases, GARD will have the capability of supporting education and clinical decision making; for instance, GARD can provide more complementary information not only for the inquiry disease but also for phenotypically similar diseases.

There are many different rare disease resources available, and each of them has their own strength and focus. OMIM classifies diseases based on their genetic cause, Orphanet defines rare diseases based on phenotypical characteristics, and UMLS incorporates biomedical vocabulary and standards to define their disease concepts. Given the complementary definition of disease concepts from these 3 resources, we employed their mappings to the GARD diseases for disease similarity comparison. Of the 392 disease pairs, 271 disease pairs (80%) with similarity scores greater than 0.5 were successfully validated as clinically relevant by our genetic specialists. Besides these true positives, feedback from our subject matter experts on the false positives [ie, 16 disease pairs (~4%) with similarity scores greater than 0.5 were manually determined as irrelevant] and false negatives [ie, 70 disease pairs (~18%) with similarity scores less than 0.5 were manually determined as relevant] illustrates that it is important to accurately capture the latest information in regard to disease mappings across different resources, and to incorporate human interpretations. For example, “Spondylothoracic dysostosis (GARD:0006798)” and “Spondylocostal dysostosis 1 (GARD:0010726)” share 3 of the same mappings, “ORPHA:2311,” “UMLS:C0265343,” and “OMIM:277300,” so their similarity score equals 1.0, indicating that they should be highly similar. However, our experts marked them as “Unrelated” due to the fact that these 2 conditions were grouped together in the past (both were previously referred to as Jarcho-Levin syndrome); they are considered as distinct conditions now, according to references from GHR (Genetic Home Reference) [29,30]. Berdon et al [31] also discussed the clinical and radiological distinction between these 2 diseases. Another example is “Hunter Carpenter Macdonald syndrome (GARD:0002751)” and “Infantile neuroaxonal dystrophy (GARD:0003957),” which have a similarity score of 0.35, indicated they should be less relevant. However, it was marked as relevant by our experts given that PLA2G6-associated neurodegeneration (PLAN) comprises a continuum of 3 phenotypes with overlapping clinical and radiologic features for these 2 diseases, and similar evidence can be found at Orphanet [32] that reveals that Hunter-Carpenter-McDonald syndrome has been moved to “Infantile neuroaxonal dystrophy.” In comparison of the total 13,705,230 GARD disease pairs, there are only 392 disease pairs with similarity scores greater than 0, which might direct the extension in 2 ways. First, 3 selected resources might not be comprehensive enough to cover all GARD diseases for disease similarity comparison based on their disease mappings. Therefore, we plan to extend our work with additional rare disease resources, such as MONDO, Disease Ontology, NCI Thesaurus, etc. Second, external disease mappings curated by GARD are accurate but might be incomplete due to cumbersome human effort. Thus, we will extend the disease mappings by inferring new associations via network analysis from the Neo4j knowledge graph.

Phenotypical similarity derivation from disease classifications resulted in 360 disease pairs shared with at least 1 phenotype and gene, and they are grouped as sibling diseases. Among 241,604 disease pairs retrieved from the disease classification trees, there are 84,054 disease pairs that share at least 1 phenotype and 396 disease pairs that share at least 1 gene. Compared to the number of disease pairs with shared phenotypes, a relatively small number of disease pairs shared at least 1 gene; we are planning to obtain more genes for GARD diseases from other resources, including DisGeNet [33] and ClinVar [34]. Given the success we gained from this study in identifying phenotypical similarity derived from sibling diseases from disease classifications, we propose to extend this work with subtype diseases (ie, parent diseases and child diseases) by mining disease classifications. Once we have GARD diseases that we are able to assign to those relevant categories, we will develop our own disease classification system, which will not only define more accurate disease definitions and relationships among those diseases but will also serve as a unique, rare disease resource in the United States.

By combining 2 sets generated by our 2 approaches, we identified 662 phenotypically similar disease pairs and mapped them to 4 phenotypical similarity types, namely, “Duplicates,” “Subtypes,” “Siblings,” and “Unrelated,” which will be applied to direct GARD data harmonization. To be specific, for “Duplicate” disease pairs, we will select and keep primary diseases in the GARD database; “Siblings” and “Subtypes” will direct GARD disease classification regeneration; for “Unrelated” diseases, we will keep these 2 diseases separately in the GARD database.

By comparing these 2 sets, there are 39 overlapped disease pairs. These 39 disease pairs were grouped as “Siblings” by the second approach of disease classification derivation. However, based on the evaluation result (Table 4) from the first approach of disease mapping analysis, of these 39 disease pairs, 25 disease pairs were grouped as “Siblings,” 7 pairs were grouped as “Subtypes,” 2 pairs were grouped as “Duplicated,” and 5 pairs were grouped as “Unrelated.” For instance, “Malignant hyperthermia” and “King Denborough syndrome” are classified as sibling diseases by the second approach, since they are siblings in Orphanet, which groups them under the same disease parent class of “Rare Disease With Malignant Hyperthermia (ORPHA:466658).” However, they are determined as different diseases by our subject-matter experts, and the same statement has been made in the GARD page for “King-Denborough syndrome (GARD:0008433),” claiming that “King-Denborough syndrome is a congenital myopathy associated with susceptibility to malignant hyperthermia (GARD:0006964)” [35]. Such discrepancies occurring across different resources unveiled from this study illustrate that there is an urgent need to propose a standard protocol for guiding data harmonization in the rare disease field globally. Regardless of phenotypical similarity types, the process our subject-matter experts took in the evaluation step is crucial to re-evaluate rare diseases with the latest prevalence data, which is one critical step to determine their eligibility of RARE. For instance, there are more than 200,000 individuals in the United States who are affected with familial Alzheimer disease (GARD:0000632) [36,37]; thus, the prevalence rate of this disease does not meet the criteria of the United States’ rare disease definition, so it will be retired from the GARD database.