Recently, to diagnose rare genetic diseases, Virtual Gene Panels (VGP) comprising sets of disease-related causal genes have been used to interpret candidate genes detected through whole-genome sequencing and whole-exome sequencing. In a pilot study of the UK 100,000 Genome Project for diagnosing rare diseases, VGPs from PanelApp software were used to filter candidate genes instead of manual interpretation to improve diagnostic efficiency. However, in about 50% of cases, the PanelApp VGPs also filtered out disease-causing genes. Here, we propose several methods using the hierarchical structure of Mondo disease ontology to design VGPs that avoid filtering disease-causing genes with high probability. To determine the best VGP design method among those proposed, we evaluated our methods with various parameters via computational experiments with an evaluation data set composed of 74 patients. Our results show that our proposed method can contribute significantly to automatically filtering candidate genes as well as shortening the interpretation time for diagnosing rare diseases.
Rare diseases are those with a very low prevalence rate, of which there are estimated to be about
10,000 in the world [
Recently, to reduce the time needed for manual search, a Virtual Gene Panel (VGP) comprising a
set of disease-related causal genes has been used to filter candidates. By taking the intersection of
candidate genes and the set of genes of the VGP corresponding to an initial diagnosis, the number of
potential candidates can be reduced significantly. As part of the UK 100,000 Genomes Project, a pilot
study for rare disease diagnosis, the PanelApp software, including 332 VGPs, was developed and
applied to the manual interpretation of whole-genome sequencing results [
There are two reasons why disease-causing genes were not effectively filtered in half of the pilot study cases. First, there were only 332 VGPs included in PanelApp. VGPs in PanelApp are designated through manual curation, which takes time to design. Therefore, there were not always appropriate VGPs corresponding to initial diagnoses. Furthermore, if the initial diagnosis is even slightly wrong, the disease-causing gene may not be included in the VGP.
In this study, we propose a method to design VGPs automatically using a knowledge graph.
Additionally, we propose several methods to expand VGPs using the hierarchical structure of Mondo
disease ontology (Mondo) [
We constructed a knowledge graph based on the Resource Description Framework to design VGPs
automatically (https://integbio.jp/rdf/dataset/pubcasefinder), using existing classes for interoperability
with other knowledge graphs. All diseases and genes were defined as instances of "med2rdf:Disease"
and "med2rdf:Gene" classes, respectively, by Med2RDF-ontology (http://med2rdf.org/) [
We collected gene-disease associations from the following three data sources.
MIM2Gene (https://ftp.ncbi.nlm.nih.gov/gene/DATA/mim2gene_medgen).
Associations between MIM
Numbers with type "phenotype" and NCBI Gene IDs were extracted.
OrphaData (http://www.orphadata.org/data/xml/en_product6.xml).
Associations between OrphaCodes and Gene symbols were extracted.
GenCC (https://search.thegencc.org/download) [
Associations between disease IDs including MIM Numbers, OrphaCodes, and Mondo IDs and HGNC IDs were extracted.
∪ ∪ 2.2.
We connected to MIM numbers in Online Mendelian Inheritance in Man [
Initially, when a disease, d, is diagnosed, G(d) can be regarded as a VGP for d. Original set (OR): For a given disease d, output a set G(d) of genes. Using OR, we can obtain as many VGPs as the number of Mondo IDs. However, like PanelApp, OR also has the problem of filtering out the disease-causing gene if the initial diagnosis is slightly different. To overcome this problem, we propose an expansion of the VGP designed by OR.
Mondo forms a directed acyclic graph, in which nodes represent diseases and edges represent "rdfs:subClassOf" relationships between diseases. Using this structure, we expanded the VGPs designed by OR by adding genes of superclasses in Mondo.
Here we describe four simple rolling-up methods to expand VGP using rdfs:subClassOf. One class up for all paths (1UAP):
For disease d, compute a set
of
Two classes up for all paths (2UAP): For disease d, compute a set of genes ′, rdfs:subClassOf ′.
| ∈ set of genes ′, rdfs:subClassOf ′. minimum paths (1UMP): For disease d, compute a set of genes , rdfs:subClassOf ′
′ rdfs:subClassOf ′′.
Two classes up for minimum paths (2UMP): For disease d, compute a |G(d'')| | ∈ ′′, rdfs:subClassOf ′ ′ rdfs:subClassOf ′′.
genes | ∈ One class up for |G(d')| | ∈ if|′| else
∪ ′ ∪ ∈ return ∅ while ′′ ′′ return if|′| else ∪ ′ ′.push′′ for ′′ such that ′ is a subclass of ′′
Input: a Mondo ID d, the Mondo Ontology Output: a set G of genes .push //′ is a stack is not empty .pop
Additionally, we describe the following two methods to expand VGP by rolling-up until the number of genes exceeds the threshold k.
All paths up with threshold k (TH k AP)
Minimum path up with threshold k (TH k MP)
For an initially diagnosed disease d, TH k AP and TH k MP compute a set D of Mondo IDs, such size of ′
. The algorithms of TH k AP and TH k MP are shown in Algorithms 1 and 2. that each
in D is d or ancestor of d and the size of
for all ′ in D, and TH k MP outputs ′ is equal to or greater than the threshold k. for ′ with minimum
Input: a Mondo ID d, the Mondo Ontology Output: a set G of genes ∅ .push //′ is a stack while is not empty .pop ′.push′′ for ′′ such that ′ is a subclass of ′′ |G(d)| | ∈
To evaluate the diagnostic efficiency of VGPs developed according to the proposed methods, a data set from 74 patients obtained via whole-exome sequencing (WES) was used. The data set includes candidate genes from WES, initially diagnosed diseases, and disease-causing genes. The median of the number of candidate genes in the 74 patients was 384. For each patient, the number of disease-causing genes, which would be included in G, was one. The number of initially diagnosed diseases, which is the input of the methods, takes from one to three, with a median of one.
Using the knowledge graph, VGPs for about 25,000 diseases contained in Mondo can be automatically designed by the OR method. To evaluate the performance of the seven expansion methods, we computed Coverage, Median, and Expectation. Table 1 shows the experimental results of OR, 1UAP, 1UMP, 2UAP, and 2UMP. Table 2 and Table 3 show the results for TH k AP and TH k MP when increasing the threshold from 50 to 300 at increments of 50. Coverage refers to the ratio of VGPs that include disease-causing genes. Median is the median of the sizes of intersections of the patient's candidate genes and VGPs. Expectation is the expected number of genes to analyze for determining the final disease-causing gene. If VGPs include disease-causing genes, only genes in intersections of the patient's candidate genes and VGPs should be analyzed. However, if VGPs do not include diseasecausing genes, all candidate genes should be analyzed. Therefore, the expectation can be computed using the following formula.
Expectation ∈ VGP | ∩ VGP | ∉ VGP | | / 74 where, for patient , a set of genes in VGP, the disease-causing gene, and a set of candidate genes are denoted by VGP( ), , and , respectively.
The results of the experiments using OR, 1UAP, 1UMP, 2UAP and 2UMP Coverage Median Expectation 2UAP performed best in terms of Coverage, although this had the worst Median. Conversely, OR had the best Median but worst Coverage. Regarding Expectation, TH 200 MP performed best.
Among the seven methods, the best performance observed was for TH k MP, especially when k = 200. As k increases, both the coverage of TH k MP and the median of the number of genes to be analyzed become larger. Therefore, there is a trade-off between coverage and the median. By selecting suitable values of k, users can design a VGP with preferred coverage and median.
Using VGPs, the number of candidate genes to be analyzed can be reduced. If a VGP was not used,
all candidate genes had to be analyzed (median of 388). However, by using VGPs designed by TH 200
MP, the median of the number of candidate genes was reduced to 14.5 if the VGPs included
diseasecausing genes. Therefore, VGPs designed by our proposed methods may be useful for gene ranking
systems, such as PubCaseFinder [
In this study, seven methods were presented to develop expanded VGPs for initial clinical diagnosis using knowledge graphs including Mondo. As a result, the TH 200 MP method, i.e., the minimum path navigation using a threshold of 200, achieved the best performance regarding Expectation. We found that this could contribute significantly to automatically reducing candidate genes as well as shortening the interpretation time for diagnosing rare diseases. We expect our methods to be widely used by clinicians to diagnose rare diseases with NGS analysis technology.
We are very grateful to Prof. Hirotomo Saitsu for providing the evaluation dataset. This work was supported by the National Bioscience Database Center of the Japan Science and Technology Agency, by "Challenging Exploratory Research Projects for the Future" grant from Research Organization of Information and Systems and by JSPS KAKENHI grant number 21K12148.