Personalized gene and disease prioritization method for rare genetic diseases based on phenotype and genotype data

Abstract

A method for personalized gene / disease prioritization for rare genetic diseases based on phenotype and genotype data is proposed according to the present disclosure. Utilizing graph embedding and graph similarity-based approaches in tandem, disclosed invention presents a marked improvement over the disadvantages of both methods alone, and generating results which are greatly explainable due to its phenotype-to-phenotype comparison centered approach. Disclosed invention also improves the accuracy of the prioritization methods even further with two different techniques Average-N 15 and IC-N, which leverage information content.

Description

[0001] PERSONALIZED GENE AND DISEASE PRIORITIZATION METHOD FOR RARE GENETIC DISEASES BASED ON PHENOTYPE AND GENOTYPE DATA

[0002] Technical Field of the Present Invention

[0003] The present invention relates generally to data driven methods that receive data pertaining to genes and genetic disease information to generate similarities and associations or predictions, and more particularly to methods of processing information for obtaining more in-depth information pertaining to genetic diseases, through which a personalized, patient-specific prioritized gene / disease information is obtained by using the relationship between gene, disease and phenotype from clinical findings.

[0004] Background of the Present Invention

[0005] The problem of disease prioritization is the patient-specific scoring of genes marked with possible diseases that are likely responsible for causing a particular disease from a large set of genetic candidates. This problem has been prominently investigated due to the increasing availability of genomic data and the need for efficient methods to interpret and prioritize diseases based on phenotypic information in the context of human disease.

[0006] Appropriate treatment of a disease depends on the correct diagnosis. There exist many variants (e.g. differences on bases in patient's DNA compared to healthy reference sequence) in the DNA analyses, and therefore the number of possible candidate genes is high. This is why it is generally quite difficult to find the responsible gene and produce a workable diagnosis, as the array of diseases, possible genes as well as genes with variants associated with diseases are immense. Especially in the case of rare diseases, the large number of possible individual diagnoses makes the diagnosis process difficult and unsatisfactory, next to failing to address this problem which is marked by an increased loss of life quality as well as loss of life. The main challenge, therefore, is the fact that there are numerous genes to examine, which is a time-consuming task. Accurate gene / disease prioritization can significantly reduce the time and effort required for experts to examine each gene, thereby expediting the process.

[0007] The Human Phenotype Ontology (HPO) is a popular open database utilized for purposes of scoring for prioritization in the literature, which is described as a directed acyclic graph (DAG). This ontology database defines phenotypes, genes and diseases that have been demonstrated to be correlated in the literature. The relationships between phenotypes, genes and diseases are specified as graph structures where nodes represent phenotype terms, genes and diseases and edges represent their relationships. HPO has become a widely adopted standard for describing phenotypic abnormalities in rare genetic diseases, and been extensively used in various applications including clinical diagnostics and gene prioritization. In HPO database, it is inscribed that a gene is associated with multiple diseases and the phenotypes associated with diseases concern all phenotypes of said associated multiple diseases.

[0008] In recent years, two main approaches have emerged as techniques for gene (disease) prioritization: Models based on classical similarity methods and embedding-based models. Conventional similarity-based models directly use the similarity of data such as their definitions, relationships and information content to calculate the similarity between phenotypes, genes and diseases. Information Content, used in similarity-based methods, indicates how unique a term is and how much additional information it provides beyond the general information. There are different ways of generating information content in the HPO graph. For example, it is possible to extract information content from the HPO graph only in terms of being more unique among HPOs, while it is also possible to extract information content from the HPO-Gene graph in terms of gene relatedness of phenotype terms.

[0009] In 2009, Kohler et al. used the terms phenotype for the information content of phenotype on a disease-by-disease basis in the HPO-Disease sequence (Kohler et al., 2009). In the same publication, the information content based on the use of the highest information content of common ancestors was also defined for the similarity of the two terms. In 2015, Gottlieb et al. (2015) found the information content of phenotype terms in the HPO graph by taking the logarithm of the number of phenotype terms and the logarithm of the number of all phenotype terms following the corresponding phenotype node (Gottlieb et al., 2015). In this method, the information content increases as we move towards the leaves of the HPO directional graph. In 2019, Xue et al. used the information content obtained from the gene relatedness of the phenotype term in the HPO-Gene graph, which we consider as a similarity-based method (Xue et al., 2019). Xue and colleagues addressed the disease prioritization problem.

[0010] Classical similarity-based models have shown promising results in prioritizing disease-causing genes by leveraging the information contained in ontologies and the relationships between phenotypes, genes and diseases. On the other hand, embedding-based models have emerged as a powerful alternative to graph similarity-based approaches. These models aim to learn vector representations (embedding) of genes, diseases and phenotypes in a continuous space while preserving their semantic relationships. Embeddingbased models such as Node2Vec (Grover & Leskovec, 2016) can capture hidden features and relationships between biological ontology nodes by leveraging natural language processing techniques such as Word2Vec (Mikolov et al., 2013). The learned embeddings can then be used to measure similarity between genes and diseases using vector operations such as cosine similarity or dot-product. In Peng et al.'s study, the phenotype-gene graph was embedded using Node2Vec, and then the patient phenotype embeddings and gene embeddings were compared using a dot-product operation (Peng et al., 2021). Embedding-based models have the advantage of efficiently handling large-scale datasets and can potentially discover new relationships not explicitly captured in the original graph.

[0011] European Patent with number EP 3 550 568 Bl concerns graph convolutionbased gene prioritization on heterogeneous networks. The method includes obtaining a set of entities for human rare diseases from one or more sources containing rare diseases, genes, phenotypes for rare diseases and biological pathways and constructing an initial heterogeneous network using each of an entity from the set of entities, the initial heterogeneous network applying Graph Convolution-based Association Scoring (GCAS) to the initial heterogeneous network to derive inferred associations and creating a Heterogeneous Association Network for Rare Diseases (HAN RD) by adding the inferred associations to the initial heterogeneous network and generating a prioritized set of genes for an input query being received in the HANRD.

[0012] EP 4 078 597 Al discloses a computer-implemented method of learning novel relationships among various entities, in particular biological entities such as chemicals, proteins, and diseases, comprises establishing a knowledge graph wherein each of the entities is represented as a node and each relationship between the entities is represented as an edge between the respective nodes, and annotating entities in the knowledge graph with objects of one or more data modalities. A neural network system is trained with the knowledge graph, wherein the neural network system treats the knowledge graph and the objects of a respective one of the data modalities in a unified manner by jointly learning embeddings of the nodes from the knowledge graph and embeddings of the objects of the respective one of the data modalities. The learned embeddings are used for identifying novel relationships among the entities.

[0013] Graph similarity-based models can provide interpretable results, however may suffer at dealing with the more incomplete structure of the ontology. On the other hand, embedding-based models can handle large-scale datasets and capture complex relationships, but their interpretability of the results may be limited and they require careful parameterization and validation.

[0014] Objects of the Present Invention

[0015] Primary object of the present invention is to provide a hybrid method of gene and disease prioritization and scoring.

[0016] Another object of the present invention is to provide a hybrid method of phenotype-based gene / disease ranking which offers greater explanatory strength due to phenotype-phenotype comparisons.

[0017] Another object of the present invention is to provide a hybrid method of phenotype-based gene / disease ranking which offers less runtime memory usage and fast execution.

[0018] Another object of the present invention is to provide a hybrid method of phenotype-based gene / disease ranking which bypasses the need for secondary embedding of both phenotypes and genes / diseases. Another object of the present invention is to provide a hybrid method of phenotype-based gene / disease prioritization which offers enhanced performance based on graph based bespoke features and information content.

[0019] Brief Description of the Present Invention

[0020] Disclosed invention mainly aims to achieve ranking and scoring of diseases and genes linked to phenotypes based on the information available on Human Phenotype Ontology database, which is based on a directed acyclic graph based on a novel approach. Specifically in the context of the invention, rare genetic diseases / genes are prioritized and personalized based on phenotype and genotype data obtainable from open databases incorporating established phenotype hierarchy, phenotype-gene and phenotype-disease information such as HPO. Disclosed invention also uses patient's own phenotype data such that a personalized ranking can be achieved. According to various aspects of the disclosed invention, data sourced from HPO are used to train a model, which is then run on the patient phenotype information to achieve said personalized ranking.

[0021] Disclosed invention offers a hybrid method that combines the two approaches of graph embedding and graph similarity-based methodologies, and thus presents an unprecedented improvement over the disadvantages of both methods alone. Whereas it uses graph embeddings, present disclosure generates results which are greatly explainable due to its phenotype-to- phenotype comparison centered approach. To improve the accuracy even further, two different techniques are also disclosed, namely the Average-N and IC-N, which will be explained in further detail. According to the present disclosure, the method makes use of the information content (IC) computed from the knowledge graph, thereby improving performance with attributes generated in a bespoke fashion which is a marked improvement over the outcomes that can be obtained solely by embedding. Due to the use of embedding, it has a high success rate to deal with ontological deficiencies and informational gaps. According to the invention during the phenotype-phenotype comparison, all gene phenotypes for a patient phenotype are taken into account and a novel approach is proposed on how to incorporate scores obtained for both into the prioritization. When the average of all gene / disease related phenotypes during phenotype-phenotype comparisons are obtained, comparison is also made with unrelated phenotypes and the performance of the algorithm decreases. Present invention remedies this problem by streamlining the scoring process.

[0022] When the most similar of the gene-linked phenotypes are used during phenotype-phenotype comparison, the problem of equal score (i.e. all phenotypes get the same score) and the small width of associated phenotypic environment reduces the performance of the algorithm. For this reason, when averaging, the best K phenotype scores are averaged according to the technique presented herein.

[0023] Brief Description of the Figures of the Present Invention

[0024] Accompanying drawings are given solely for the purpose of exemplifying a method of embedding based disease / gene prioritization model and a system using said model, whose advantages over prior art were outlined above and will be explained in brief hereinafter. The drawings are not meant to delimit the scope of protection as identified in the claims nor should they be referred to alone in an effort to interpret the scope identified in said claims without recourse to the technical disclosure in the description of the present invention.

[0025] Figure 1 demonstrates an example of the relationships between genes and related phenotypes according to an embodiment of the present invention.

[0026] Figure 2 demonstrates an example of the average-N method according to an embodiment of the present invention.

[0027] Figure 3 demonstrates a pseudocode representation of the method of IC-N according to an embodiment of the present invention.

[0028] Detailed Description of the Present Invention

[0029] Disclosed invention mainly aims to achieve ranking and scoring of diseases as well as genes associated with phenotypes based on the information available on Human Phenotype Ontology database, which is based on a directed acyclic graph, with a novel approach. Specifically in the context of the invention, rare genetic diseases and genes associated therewith are prioritized and personalized based on phenotype and genotype data obtainable from open databases incorporating established phenotype hierarchy, phenotype-gene and phenotype-disease information such as HPO.

[0030] In accordance with the present invention, the disclosed system is capable of operating with both genes and diseases as target domains to rank target entities. The detailed descriptions, including the accompanying explanations and formulas, are provided primarily in the context of the gene domain. However, this specification should not be construed as limiting the application of the invention exclusively to the gene domain. The principles and methodologies described herein are equally applicable to the disease domain, thereby ensuring the versatility and adaptability of the invention across different biological target domains.

[0031] The invention according to the present disclosure sets forth a hybrid technique merging classical similarity-based and embedding-based methodologies, marking a significant advancement over the limitations of each method independently. While rooted in embedding, this disclosure produces highly interpretable outcomes through its focus on phenotype-to-phenotype comparison.

[0032] According to the present disclosure, the method leverages information content (IC) derived from the graph, enhancing performance by incorporating attributes manually generated beyond embedding capabilities. By utilizing embedding, it effectively addresses ontological deficiencies and information gaps in the data. Keeping in mind, when averaging all gene / disease-related phenotypes during comparisons, unrelated phenotypes can adversely affect algorithm performance, a mitigating measure of the present invention is the optimization of the scoring process. Specifically, when comparing phenotypes, focusing on the most similar gene-linked phenotypes prevents issues like equal scoring and narrow phenotypic scope, which can diminish algorithm performance. Thus, the technique outlined herein advocates for averaging the top K phenotype scores to address these challenges effectively.

[0033] Disclosed invention has three stages on the high level, which are graph generation, graph embedding, and phenotype comparison and similarity scoring. First, a phenotype-gene or phenotype-disease graph is generated based on the info obtained from HPO database. While constructing the graph, various graph enrichment methods can be used in different embodiments, such as the CADA method applied by Peng et al. The first of these methods is the method of adding clinical cases obtained from various open sources such as ClinVar or affiliated clinics to the graph as HPO-to-Gene links. Next to this method, in addition to adding training cases as links, the frequency of case encounters can also be given as a weight to the link.

[0034] The next step is to create Node2Vec embeddings with the generated graph. Through CADA method (as known in the art) and after the embeddings are created on the HPO-to-gene graph, phenotype nodes and gene nodes are directly compared. Here, diseases can also be directly used instead of genes. As an important aspect of the present disclosure, instead of directly comparing phenotypes and genes, the approach is to compare the set of patient phenotypes with the set of phenotypes to which the gene is linked. This approach forms the basis of the disclosed method. Unlike the study from which CADA method emerges, cosine similarity is used for node comparisons instead of the dot-product method. During gene / disease score generation, the score of the K phenotypes with the highest similarity between a phenotype in the patient phenotype set and the gene phenotype set is considered the score of the gene / disease. The information content which will be subsequently explained in greater detail was also used to find the phenotype score of the gene or disease.

[0035] Information Content, used in the similarity-based method, indicates how unique a term is and how much additional information it provides beyond general knowledge. There are different ways of generating information content within the HPO-Gene graph. For example, in an embodiment, it is possible to extract information content from the HPO graph only if it is more unique among HPO data, while it is also possible to extract information content from the HPO- Gene graph in terms of gene relatedness of phenotype terms.

[0036] In a 2015 study by Gottlieb et al., it was found that the information content of phenotype terms in the HPO graph by taking the logarithm of the number of phenotype terms and the number of all phenotype terms following the relevant phenotype node. This approach posited that, the information content increases as one moves towards the leaves of the HPO directed graph. In 2019, Xue and colleagues used the information content obtained from the gene relatedness of the phenotype term in the HPO-Gene graph, which is considered as a similarity-based method. A similar content was used by Kohler et al. in 2009 for the disease-based information content of phenotype terms in the HPO- Disease sequence. At the same time, the use of the highest information content of common ancestors in finding the similarity of two terms was used in the art by other practitioners. In at least one embodiment, present disclosure utilizes the Node2Vec approach, which is an embedding based technique known in the art. In essence this takes aspects of both a natural language processing approach and graph theory, and generates vector spacing for nodes in a graph,

[0037] The method according to the present disclosure accepts as input the patient HPO terms, a model, a graph, and a top N average, and a listing of ranked genes / diseases. As a first step, a knowledge graph is constructed using HPO dataset, which comprises hierarchical relationships between phenotype terms and the association between phenotypes and genes. Next, a graph enrichment is undertaken. In an embodiment this is by using a Node2Vec model, and generating vector space embeddings for the nodes within aforementioned constructed knowledge graph. Node2Vec commences by initiating random walks from each node in the graph, treating the sequences of nodes encountered during these walks as sentences in which nodes represent words. This approach allows for the application of the SkipGram neural network model, originally developed for the Word2Vec method, to embed these sequences into a Euclidean vector space. By this procedure, the probability of accurately predicting a node's contextual neighbors within a defined window length is maximized, thereby the essence of the graph's topological structure is better captured. Training the SkipGram model involves using one- hot vector encoding of node pairs, with the aim of learning the likelihood of observing a node within another's neighborhood context. By maximizing this likelihood across all nodes in graph, Node2Vec encodes nodes into a vector space where those with similar network neighborhoods are positioned closely together, thereby preserving the graph's intrinsic structural properties in the embedding space.

[0038] Candidate gene(s) can pertain to a list of genes which include a likely list of possible diagnostic leads, either supplied by a prior assessment or at the discretion of a health professional, which can be integrated into the disclosed approach by comparing them to the results obtainable with the method disclosed herein. The output of the method can further be tailored such that the results are confined solely to those candidates that are integrated to the algorithm.

[0039] Subsequent to this, for every candidate gene, HPO terms associated with the gene are determined; for every patient HPO term, a cosine similarity is computed for each gene-adjacent HPO between the patient HPO term and the gene-related HPO term embeddings. After this, the terms which are computed are ranked in descending order of similarity based on the calculations and, the highest N number of similarity scores are averaged, which is then appointed as a candidate of term-gene similarity value. Once this value is obtained for every patient phenotype term, a gene score is obtained by averaging the scores of every single patient phenotype terms. Once more, these averages for every single instance of target genes are listed in ascending order and, in an embodiment the best K number of genes are selected. The method finally retrieves the output of best K number of genes. In an embodiment the number N is selected to be 5.

[0040] Node embeddings, which are obtained from the aforementioned knowledge graph comprising gene and phenotype nodes using node2vec algorithm, are significant in that, in contrast to other embedding-based approaches known in the art, the disclosed invention utilizes phenotype set similarity with these embeddings rather than relying on traditional embedding-based similarity measures. Where the set of patient phenotype is / ’and the set of phenotypes which are related to a gene ^are denoted as G , similarity of / ’to Cis used to calculate gene score. In other words, patient phenotype set is compared with the set of phenotypes related to the gene being prioritized, and a score to be calculated for each gene based on its relevance to the patient's phenotypic profile. When consulting Figure 1, it can be observed that the right-angle arrows connect different instances of HPO phenotype terms, whereas the curvy arrows connect gene instances (on the right-hand side) to associated phenotypes. An example according to the Figure can be that the gene "Entrez:1054" is linked to the hypothetical phenotypes "HP:0003" and "HP: 0004", which together form the gene-related phenotype set Cfor the gene "Entrez:1054".

[0041] Several stages are found according to the disclosed approach, throughout the course of application of which will refer to the cases for prioritizing genes for sake of simplicity, while noting that the cases for diseases can be formulated similarly thereto. The first of these stages is the comparison of each term in the patient phenotype set with the terms in the gene phenotype set. This is specifically utilized in the step of computing a similarity for each gene- associated phenotype between the patient phenotype term and the gene- related phenotype term embeddings, for every patient phenotype term. In this stage, for term similarity, the "information content depending on the number of successive terms" and "common ancestor with the highest information content" methods in the HPO graph were used.

[0042] Z(t)= - log (- - ) (1)

[0043] Above formula, numbered one, is defined as the information content of the phenotype term. Dt denotes the number of successive terms that come after the relevant term (i.e. the descendant number of the term), whereas NHPO denotes the total number of unique terms in HPO database.

[0044] Formula 2 given above expresses the similarity between two terms based on common ancestors, where AnsInt is the intersection of sets of ascendents of the term t and t'. Anstdenotes the set of ascendents of the term t. When two HPO terms are compared according to Formula 2, if term t is the same as f'or if term t is one of the ancestors of term f, the similarity value will be Z as specified according to Formula 1. If there is no direct ancestral relationship between them, the Rvalue of the closest ancestor will be taken instead.

[0045] In the second stage, when averaging the patient phenotypes for the gene score (which can both pertain to genes as well as diseases), the information content of the phenotype on a gene-by-gene basis was found, which was computed using Formula 3 given below. These weights were then used to take the weighted average value.

[0046] These formulae form the basis of the scoring executed as part of the present disclosure's method. One such instance is finding a patient-term / gene-term score computing. Different embodiments exist which utilize this top N averaging approach. (emb(t), emb(t')) (4)

[0047] Above given representation is a cosine similarity metric calculation, where t stands for patient phenotype term, / -'stands for gene-related phenotype term, cos is the cosine similarity operand, emb(t) denotes embedding of phenotype term t.

[0048] According to the Average-N approach in at least one embodiment of the present disclosure, patient and gene similarity is represented as below:

[0049] StepfAvgTop > (SimT(t,t ))}

[0050] Information content, in terms of similarity-based methods and approaches, denotes the uniqueness of a given term in that the amount of useful information it can convey on top of what is already known about the relationships between certain entities, such as genes, phenotypes etc. in a graph. As a corollary, whereas an HPO graph content can only produce information based on the entities in the HPO, additional information can be leveraged based on a HPO-gene graph pertaining to the gene relatedness of phenotype terms. This is achieved, according to various embodiments of the present disclosure, using the formulas 1, 2 and 3 given above.

[0051] In an embodiment, as part of the IC-N approach, a similarity score between patient phenotype term and gene phenotype term entities based on the above given formula 2, which are in turn used as coefficients K.

[0052] In at least an embodiment, as part of the IC-N approach, when finding the term-gene scores, formula 3 is used on the patient terms for finding the information content of the term in the context of a given gene. When computing gene scores in this way, all the term-gene scores of every patient term are averaged, which is a weighted average. The value obtained using formula 3 is used as a weight when calculating the average value.

[0053] What is noteworthy is that in contrast to the established practice known in the art, present disclosure leverages cosine similarity of embedded terms instead of Lin similarity which was first disclosed in Dekang Lin's work titled "An Information-Theoretic Definition of Similarity", as well as the usage of top N number of averages, which is a ranking proposed according to the present invention which improves the legibility and strength of the associations based on the embeddings.

[0054] According to an embodiment, a similarity measure is disclosed wherein embedding similarity and graph-based information content parameters are incorporated, using both embedding-based and information content-based similarities to compute term-wise similarities. IC-N method is subsequently used to compute the gene score. The term information content in the target domain (which can be gene domain or disease domain), i.e. Igene is used to determine the importance of a term in the gene domain as a weight, and a weighted average is obtained to calculate the final patient-gene score. This similarity measure is as follows:

[0055] StepfAvgTop / (simT(t,t' )x SimTJc(t,t' ) x I ene(t)) where SimT and SimT_IC stand for term and information content similarity scores, while still using the Average-N approach as explained hitherto.

[0056] In a nutshell, present disclosure offers a computer-implementable method for personalized gene / disease prioritization for rare genetic diseases based on phenotype and genotype data. Said method comprises steps of, introducing an information graph prepared based on a group of data selectable from phenotype-phenotype, gene-phenotype and disease-phenotype relationship data obtainable from Human Phenotype Ontology database.

[0057] According to an embodiment of the present disclosure, said method comprises training a model for the purpose of embedding graph nodes to vectors based on said information graph.

[0058] According to an embodiment of the present disclosure, said method comprises utilizing vector embeddings obtained according to the previous steps, for every phenotype term associated with any candidate gene / disease, and for every patient phenotype term.

[0059] According to an embodiment of the present disclosure, said method comprises, for every patient phenotype term, computing a similarity metric for each gene / disease-associated phenotype term based on said vector embeddings, wherein an information content in phenotype domain is used as a coefficient.

[0060] According to an embodiment of the present disclosure, said method comprises, for every patient phenotype term, ranking, in descending order of similarity, each of the terms calculated in the preceding step.

[0061] According to an embodiment of the present disclosure, said method comprises, for every patient phenotype term, averaging the highest average N number of similarity scores, N being a predetermined positive integer.

[0062] According to an embodiment of the present disclosure, said method comprises appointing the average calculated in the preceding step as a candidate for patent phenotype similarity value.

[0063] According to an embodiment of the present disclosure, said method comprises obtaining a gene / disease score by averaging the scores of every patient phenotype term, wherein an information content in any target domain as gene or disease domain is used for weight averaging.

[0064] According to an embodiment of the present disclosure, said method comprises listing genes / diseases based on the gene / disease scores in order of ascending likelihood, and returning an output of genes / diseases in list format.

[0065] According to an embodiment of the present disclosure, a system suitable for prioritizing personalized gene / disease for rare genetic diseases based on phenotype and genotype data, comprising at least an input means, a processing means and an output means is proposed. According to an embodiment of the present disclosure, said input means is configured to accept phenotype and / or genotype data.

[0066] According to an embodiment of the present disclosure, said processing means is configured to execute the following actions of introducing an information graph prepared based on a group of data selectable from phenotypephenotype, gene-phenotype and disease-phenotype relationship data obtainable from Human Phenotype Ontology database; training a model for the purpose of embedding graph nodes to vectors based on said information graph; utilizing vector embeddings obtained according to the previous steps, for every phenotype term associated with any candidate gene / disease, and for every patient phenotype term; for every patient phenotype term, computing a similarity metric for each gene / disease-associated phenotype term based on said vector embeddings, wherein an information content in phenotype domain is used as a coefficient; for every patient phenotype term, ranking, in descending order of similarity, each of the terms calculated; for every patient phenotype term, averaging the highest average N number of similarity scores, N being a predetermined positive integer; appointing the average calculated in the preceding step as a candidate for patent phenotype similarity value; obtaining a gene / disease score by averaging the scores of every patient phenotype term, wherein an information content in any target domain as gene or disease domain is used for weight averaging; listing genes / diseases based on the gene / disease scores in order of ascending likelihood, and; returning an output of genes / diseases in list format.

Claims

CLAIMS1) A computer-implementable method for personalized gene / disease prioritization for rare genetic diseases based on phenotype and genotype data, comprising steps of, introducing an information graph prepared based on a group of data selectable from phenotype-phenotype, gene-phenotype and diseasephenotype relationship data obtainable from Human Phenotype Ontology database, training a model for the purpose of embedding graph nodes to vectors based on said information graph, utilizing vector embeddings obtained according to the previous steps, for every phenotype term associated with any candidate gene / disease, and for every patient phenotype term, for every patient phenotype term, computing a similarity metric for each gene / disease-associated phenotype term based on said vector embeddings, wherein an information content in phenotype domain is used as a coefficient, for every patient phenotype term, ranking, in descending order of similarity, each of the terms calculated in the preceding step, for every patient phenotype term, averaging the highest average N number of similarity scores, N being a predetermined positive integer, appointing the average calculated in the preceding step as a candidate for patent phenotype similarity value, obtaining a gene / disease score by averaging the scores of every patient phenotype term, wherein an information content in any target domain as gene or disease domain is used for weight averaging, listing genes / diseases based on the gene / disease scores in order of ascending likelihood, and; returning an output of genes / diseases in list format.2) A computer-implementable method for embedded phenotype ranking and prioritization according to Claim 1, characterized in that said similarity metric is a cosine similarity.3) A computer-implementable method for embedded phenotype ranking and prioritization according to Claim 1 characterized in that said information content in phenotype domain is represented asWhere Dt denotes the number of successive terms that come after the relevant term, being the descendant number of the term, whereas NHPO denotes the total number of unique terms in HPO database.4) A computer-implementable method for embedded phenotype ranking and prioritization according to Claim 1 characterized in that said information content in target domain is represented aswherein G is the set of phenotypes related to a gene, and Gt is the number of genes annotated to specific term t within HPO database.5) A computer-implementable method for embedded phenotype ranking and prioritization according to Claim 1 characterized in that said average of N scores is calculated for every gene based on the representation belowwherein G is the set of phenotypes related to a gene, t and t' being two terms and P is the set of patient phenotypes.6) A computer-implementable method for embedded phenotype ranking and prioritization according to Claim 1 characterized in that said scores for each genes are computed based on information content according to the representation belowStep (AvgTopN eG(simTSimJCNpG=ljtep gene( ')' )' wherein SimT and SimT IC stand for term and information content similarity scores.7) A system suitable for prioritizing personalized gene / disease for rare genetic diseases based on phenotype and genotype data, comprising at least an input means, a processing means and an output means, wherein, said input means is configured to accept phenotype and / or genotype data, said processing means is configured to execute the following actions of introducing an information graph prepared based on a group of data selectable from phenotype-phenotype, gene-phenotype and disease-phenotype relationship data obtainable from Human Phenotype Ontology database; training a model for the purpose of embedding graph nodes to vectors based on said information graph; utilizing vector embeddings obtained according to the previous steps, for every phenotype term associated with any candidate gene / disease, and for every patient phenotype term; for every patient phenotype term, computing a similarity metric for each gene / disease- associated phenotype term based on said vector embeddings, wherein an information content in phenotype domain is used as a coefficient; for every patient phenotype term, ranking, in descending order of similarity, each of the terms calculated; for every patient phenotype term, averaging the highest average N number of similarity scores, N being a predetermined positive integer; appointing the average calculated in the preceding step as a candidate for patent phenotype similarity value; obtaining a gene / disease score by averaging the scores of every patient phenotype term, wherein an information content in any target domain as gene or disease domain is used for weight averaging; listing genes / diseases based on the gene / disease scores in order of ascending likelihood, and; returning an output of genes / diseases in list format.

Images