Apparatus and method for decision support in standardized phenotypes

Abstract

The present invention relates to a computer-implemented method for decision support of a user to standardize phenotypes in genomic analysis of a subject, wherein the method comprises:-receiving a symptom list comprising at least one symptom observed for the subject; receiving a first graph comprising nodes and weighted links; receiving a second map previously obtained by applying the matrix factorization to the gene-symptom matrix; -outputting at least one gene associated with the symptom list based on the first map and the second map.

Description

Technical Field

[0001] The present invention relates to methods and apparatus for genome analysis. More particularly, the present invention relates to methods and apparatus for calculating symptom-gene and symptom-symptom association maps and their use in standardizing clinical descriptions and in identifying one or more genes that are most likely to be the origin of the disease presented by the subject. Background Art

[0002] Precision medicine relies on patient stratification and identification of clinically relevant groups to improve diagnosis, prognosis, and medical treatment. Phenotyping allows for the composition of homogeneous groups of individuals in which physicians report features that deviate from normal morphology, physiology, and behavior using standardized descriptions in the Human Phenotype Ontology (HPO). Despite a common ontology and abundant clinical data, medical records often lack consistency and comparability between descriptions and practitioners, known as fuzzy matches in phenotypic profiles. Such inconsistent phenotyping is a major barrier to fully utilizing the clinical data contained in medical records.

[0003] Even relying on predefined ontologies, such as the Human Phenotype Ontology (HPO), physicians' descriptions of the same symptoms observed in patients may be different. In particular, the inventors conducted a study (further detailed in this disclosure) on current phenotyping practices that illustrates this heterogeneity. Current algorithms exploit ontology structures to address phenotypic heterogeneity, or extract additional symptom-gene associations from the literature, or assess semantic similarity of symptoms.

[0004] In contrast to these approaches, the inventors of the present invention used HPO as a symptom dictionary and considered the relationships between symptoms only through their co-occurrence in genetic diseases found in the HPO structured and text-mined database.

[0005] In this context, the present disclosure proposes a solution that allows filling the gaps in currently available databases. Summary of the invention

[0006] One aspect of the present disclosure relates to an apparatus for identifying genes with the highest probability of being associated with a symptom observed for a subject, the apparatus comprising:

[0007] - at least one input adapted to receive subject data representative of at least one symptom observed for the subject;

[0008] - at least one processor configured to:

[0009] Compare each subject's data representing a symptom of the subject to a phenotype represented in the genotype-phenotype matrix The gene-phenotype matrix represents the association between at least one phenotype and at least one gene believed to cause the associated at least one phenotype, wherein each coefficient of the gene-phenotype matrix Representative gene and phenotype The probability of association between

[0010] Calculating a subject-specific matrix comprising the phenotypes of the genotype matrix associated with at least one symptom of the subject as columns;

[0011] Calculate the score vector, where each coefficient of the score vector The coefficients of each row of the subject-specific matrix were obtained to represent the probability that each gene of the subject-specific matrix was associated with the symptom observed for the subject;

[0012] - at least one output adapted to output at least one prediction based on the score vector, the prediction providing information about one or more genes that are more likely to be associated with the symptom observed for the subject.

[0013] The present disclosure advantageously provides an apparatus (and a method discussed below) that allows accurate calculation of the probability that different genes are the source of a subject's disease, thereby providing an accurate ranking of the genes. The ranking is the result of a set of simple operations on a numerical matrix (i.e., a gene-phenotype matrix) and can therefore be obtained in a very fast manner.

[0014] In one embodiment, the at least one processor is configured to:

[0015] - Calculate the pathogenicity vector, where each coefficient of the pathogenicity vector represents the The frequency with which variants are reported as pathogenic; and

[0016] - Each coefficient of the score vector Subtract targeted genes Obtain the coefficients of the pathogenicity vector.

[0017] According to one embodiment, at least one processor is configured to:

[0018] - Calculate the correction vector to take into account the associated missing phenotype, where each coefficient of the correction vector is calculated as a subject-specific matrix with genes the sum of the coefficients of the associated row; and

[0019] - Each coefficient of the score vector Subtracting and targeting genes Correlation coefficient of the obtained correction vector.

[0020] According to one embodiment, the at least one processor is further configured to calculate the genotype matrix by:

[0021] -Define the matrix ,in, is the number of genes and is the number of phenotypes present in the list that includes associations between phenotypes and genes; if the list includes the gene and the phenotype The correlation between the gene and phenotype matrix is Set to a predefined maximum non-zero value if the list does not include the gene and the phenotype The correlation between Set to zero;

[0022] - For genes not present in the above list Phenotype Each coefficient of the correlation , according to phenotype and genes The coefficient is calculated by the distance between the closest phenotypes The value of , where the phenotype exists in the list and genes The closest phenotypic association is The distance to the closest phenotype is calculated on a directed acyclic graph representing the standard classification of human abnormal phenotypes and their semantic relations.

[0023] According to one embodiment, the at least one processor is further configured to calculate the genotype matrix by:

[0024] - calculating at least two intermediate gene-phenotype matrices from respective lists comprising associations between at least one phenotype and at least one gene;

[0025] -The linear combination of the calculated intermediate gene-phenotype matrices is used as the gene-phenotype matrix.

[0026] According to one embodiment, in order to calculate the gene-phenotype matrix, at least one processor is further configured to perform a Apply normalized weights, where weights It is phenotype A function of the frequency of occurrence of the genes represented in the gene-phenotype matrix.

[0027] According to one embodiment, the at least one processor is further configured to rank genes associated with symptoms observed in the subject based on a base of the score vectors.

[0028] The present disclosure also relates to a computer-implemented method for identifying genes with the highest probability of being associated with a symptom observed for a subject, the method comprising:

[0029] - receiving subject data representing at least one symptom observed for the subject;

[0030] - Compare each subject's data representing a symptom of the subject to a phenotype represented in the genotype-phenotype matrix The gene-phenotype matrix represents the association between a symptom and at least one gene believed to cause the associated symptom, wherein each coefficient of the gene-phenotype matrix Indicates gene and phenotype The probability of association between

[0031] - calculating a subject-specific matrix comprising phenotypes associated with at least one symptom of the subject as columns;

[0032] -Calculate the score vector, where each coefficient of the score vector is obtained from the coefficients of each row of the subject-specific matrix to represent the probability that each gene represented in the subject-specific matrix is ​​associated with the symptom observed for the subject;

[0033] - outputting at least one prediction based on the score vector, the prediction providing information about one or more genes that are more likely to be associated with the symptom observed for the subject.

[0034] According to one embodiment, the method further comprises:

[0035] - Calculate the pathogenicity vector, where each coefficient of the pathogenicity vector represents the The frequency with which variants are reported as pathogenic; and

[0036] - Each coefficient of the score vector Subtract targeted genes Obtain the coefficients of the pathogenicity vector.

[0037] According to one embodiment, the method further comprises calculating a gene-phenotype matrix by:

[0038] -Define the matrix ,in, is the number of genes and is the number of phenotypes present in the first list of associations between phenotypes and genes; if the first list includes the gene and the phenotype The correlation between the gene and phenotype is calculated by Set to a predefined maximum non-zero value if the list does not include the gene and the phenotype The correlation between Set to zero;

[0039] - for genes not present in the first list Phenotype Each coefficient of the correlation , according to phenotype and genes The distance between the closest phenotypes is calculated for each coefficient where the phenotype exists in the first list and genes The closest phenotype association is The distance to the closest phenotype is calculated on a directed acyclic graph representing the standard classification of human abnormal phenotypes and their semantic relations.

[0040] The present disclosure also relates to a device for calculating a genotype matrix for use as a genotype profile in a method according to the present invention for identifying genes with the highest probability of being associated with a symptom observed for a subject; the device comprising:

[0041] - at least one input adapted to receive a list comprising associations between phenotypes and genes;

[0042] - at least one processor configured to:

[0043] Define the matrix ,in is the number of genes, is the number of phenotypes present in the list including associations between phenotypes and genes; if the list includes the gene and the phenotype The correlation between the gene and phenotype matrix is Set to a predefined maximum non-zero value if the list does not include the gene and the phenotype The correlation between Set to zero;

[0044] For genes not present in the above list Phenotype Each coefficient of the correlation , according to phenotype and genes The coefficient is calculated by the distance between the closest phenotypes The value of , where the phenotype exists in the list and genes The closest phenotypic association is The distance to the closest phenotype is calculated on a directed acyclic graph representing the standard classification of human abnormal phenotypes and their semantic relations.

[0045] - at least one output adapted to provide at least one gene-phenotype matrix representing the association between at least one phenotype and at least one gene believed to be responsible for said associated at least one phenotype.

[0046] According to one embodiment, at least one processor of the apparatus for calculating a gene-phenotype matrix is ​​further configured to calculate the gene-phenotype matrix by:

[0047] - calculating at least two intermediate gene-phenotype matrices from respective lists comprising associations between at least one phenotype and at least one gene;

[0048] - The linear combination of the calculated intermediate gene-phenotype matrices is used as the gene-phenotype matrix (21).

[0049] According to one embodiment, at least one processor of the apparatus for calculating a gene-phenotype matrix is ​​further configured to calculate a value for each column of the gene-phenotype matrix. Apply normalized weights, where weights It is phenotype A function of the frequency of occurrence of genes relative to those represented in the gene-phenotype matrix.

[0050] The present disclosure also relates to a method for generating a gene-phenotype matrix according to the apparatus for calculating the gene-phenotype matrix. The gene-phenotype matrix is ​​an advantageous data structure that is more easily used for rapid calculation of gene priorities while integrating information from many different sources or synonym libraries. In fact, all information is embodied in a numerical matrix.

[0051] Furthermore, the present disclosure relates to a computer program comprising software codes, which, when executed by a processor, is suitable for performing a method for identifying genes or a method for calculating a gene-phenotype matrix in accordance with any of the above-mentioned execution modes.

[0052] The present invention also relates to a computer-readable non-transitory program storage device tangibly embodying a program of instructions executable by a computer to perform a method for identifying genes or a method for calculating a gene-phenotype matrix consistent with the present disclosure.

[0053] Such non-transitory program storage devices may be, but are not limited to, electronic, magnetic, optical, electromagnetic, infrared or semiconductor devices, or any suitable combination of the foregoing. It will be readily appreciated by those skilled in the art that, although more specific examples are provided below, it will be readily appreciated that they are merely illustrative and not an exhaustive list: portable computer floppy disk, hard disk, ROM, EPROM (erasable programmable ROM) or flash memory, portable CD-ROM (compact disk ROM).

[0054] Another aspect of the present invention relates to a computer-implemented method for decision support of a user to standardize phenotypes in genomic analysis of a subject, wherein the method comprises:

[0055] -a) receiving a symptom list comprising at least one symptom observed for the subject, wherein the symptom list comprises an integer Symptoms

[0056] - If integer is above a threshold, the method further comprising a step b1) receiving a first graph comprising nodes and weighted links, wherein:

[0057] oEach node represents a symptom, and each symptom is associated with at least one gene;

[0058] o Each link connecting two nodes represents the similarity between the two symptoms associated with the two connected nodes, and

[0059] oThe weight of each link represents the degree of similarity between the two symptoms associated with the two connected nodes;

[0060] The first graph is previously obtained by calculating symptom pair similarities using information about associations between at least one symptom and at least one gene believed to be responsible for the associated at least one symptom;

[0061] - Otherwise, if the integer is lower than or equal to the threshold, the method further includes b21), b22), b23) and b24), including:

[0062] -b21) receiving a second graph previously obtained by applying matrix decomposition to a gene-symptom matrix, the second graph comprising nodes and weighted links, wherein each node represents a symptom, each symptom is associated with at least one gene; each node is associated with at least one group of nodes of the second graph; and each node is associated with at least one specificity score, each specificity score representing the specificity of the symptom associated with the node relative to the symptoms of the nodes in each group associated with the node; wherein the gene-symptom matrix represents an association between at least one symptom and at least one gene believed to be responsible for the associated at least one symptom, and wherein each coefficient of the gene-symptom matrix Indicates gene With symptoms The probability of association between

[0063] wherein each gene represented in the gene-symptom matrix is ​​associated with each group of nodes of the second graph via a gene-group weighted score;

[0064] -b22) evaluating the specificity of the received symptom list using the specificity scores of the nodes associated with at least one symptom in the second graph, and providing a suggestion of at least one new symptom to the user if the specificity of the symptom list does not satisfy a predefined rule, each of the at least one suggested new symptoms being obtained as a symptom associated with a node in the first graph, the new symptom being connected by a link to a node associated with at least one symptom of the received symptom list;

[0065] -b23) if the user adds at least one suggested new symptom to the received symptom list, evaluating the specificity of the updated symptom list using the specificity scores of the nodes in the second graph associated with the symptoms in the updated symptom list including the at least one received symptom and the at least one newly suggested symptom;

[0066] -b24) whenever the specificity of the symptom list satisfies a predefined rule, using the symptom list and the first graph to obtain at least one gene with the highest probability of being associated with the symptom list;

[0067] The method further comprises the following steps:

[0068] -c) If, after providing a suggestion to the user for at least one new symptom, the specificity of the updated symptom list does not satisfy the predefined rules:

[0069] generating a symptom binary vector representing the symptoms in the symptom list and applying matrix factorization to the binary vector to obtain a representation of the updated symptom list on the groups of the second graph, the representation comprising a patient score for each group of the second graph;

[0070] Calculating the difference between each patient score and each gene-group weighted score for each group of the second graph;

[0071] Normalizing the differences relative to a reference set of distances representing asymptomatic subjects;

[0072] selecting at least one gene with the highest probability of being associated with the updated list of symptoms as the at least one gene with the highest normalized difference;

[0073] -d) Output at least one gene associated with the symptom list or updated symptom list.

[0074] As a result of the methods of this aspect of the invention, heterogeneous phenotypes can be handled by developing a symptom interaction model represented by the first graph and the second graph to standardize clinical descriptions and, based on these standardized clinical descriptions, identify genes with the highest probability of being associated with the symptoms observed for the subject.

[0075] The second graph can indeed link one symptom in the symptom list observed for the subject to a set of symptoms representing the phenotypic pattern of the genetic disease. Since the first graph links symptoms to given similarities, the symptom list may be enriched to help identify the genes most likely associated with the symptoms observed for the subject.

[0076] In one embodiment, the threshold is equal to five.

[0077] Another aspect of the invention relates to a device for decision support of a user to standardize phenotypes in genomic analysis of a subject, the device comprising:

[0078] - at least one input configured to receive:

[0079] o A symptom list comprising at least one symptom observed for the subject, wherein the symptom list comprises integers Symptoms

[0080] o a first graph including nodes and weighted links, wherein each node represents a symptom, each symptom is associated with at least one gene; each link connecting two nodes represents a similarity between the two symptoms associated with the two connected nodes, and the weight of each link represents a degree of similarity between the two symptoms associated with the two connected nodes;

[0081] The first graph is previously obtained by calculating symptom pair similarities using information about associations between at least one symptom and at least one gene believed to be responsible for the associated at least one symptom;

[0082] o a second graph previously obtained by applying matrix decomposition to the gene-symptom matrix, the second graph comprising nodes and weighted links, wherein each node represents a symptom, each symptom is associated with at least one gene; each node is associated with at least one group of nodes of the second graph; and each node is associated with at least one specificity score, each specificity score representing the specificity of the symptom associated with the node relative to the symptoms of nodes in each group associated with the node; wherein the gene-symptom matrix represents an association between at least one symptom and at least one gene believed to be responsible for the associated at least one symptom, and wherein each coefficient of the gene-symptom matrix Indicates gene With symptoms and wherein each gene represented in the gene-symptom matrix is ​​associated with each group of nodes of the second graph by a gene-group weighted score;

[0083] - at least one processor configured to:

[0084] oIf the integer Less than or equal to threshold:

[0085] evaluating the specificity of the symptom list using the specificity scores of the nodes associated with at least one symptom in the second graph, and providing a suggestion of at least one symptom to the user if the specificity of the symptom list does not satisfy a predefined rule, each of the at least one suggested symptom being obtained as a symptom associated with a node in the first graph connected by a link to a node associated with at least one received symptom;

[0086] if the user adds the at least one suggested symptom to the symptom list, evaluating the specificity of the symptom list using specificity scores of nodes in the second graph associated with symptoms in the symptom list including the at least one received symptom and the at least one suggested symptom;

[0087] Whenever the specificity of the symptom list satisfies a predefined rule, using the symptom list and the first graph to obtain at least one gene with the highest probability of being associated with the symptom list;

[0088] o If after providing a suggestion to the user for at least one symptom, the specificity of the symptom list does not satisfy the predefined rules:

[0089] generating a symptom binary vector representing the symptoms in the symptom list and applying matrix factorization to the binary vector to obtain a representation of the symptom list on the groups of the second graph, the representation comprising a patient score for each group of the second graph;

[0090] Calculating the difference between each patient score and each gene-group weighted score for each group of the second graph;

[0091] Normalizing the differences relative to a reference set of distances representing asymptomatic subjects;

[0092] selecting at least one gene having the highest probability of being associated with the list of symptoms as the at least one gene having the highest normalized difference;

[0093] - at least one output configured to provide at least one gene associated with a list of symptoms.

[0094] In one or more embodiments, the matrix factorization is a non-negative matrix factorization.

[0095] Therefore, since the second graph is obtained by decomposing the gene-symptom matrix by non-negative matrix factorization, it is easier to obtain at least one set of nodes of the second graph. This is due to the clustering property inherent in the non-negative matrix factorization method.

[0096] In one or more embodiments, the non-negative matrix factorization applied to the gene-symptom matrix is ​​a non-negative double singular value decomposition (NNDSVD) initialization. Using the NNDSVD initialization technique makes the non-negative matrix factorization process more efficient and faster.

[0097] In one or more embodiments, the information regarding the association between at least one symptom and at least one gene is a gene-symptom list representing the association between at least one symptom and at least one gene believed to cause the associated at least one symptom, or the information is a gene-symptom matrix.

[0098] In one or more embodiments, the weight of each link in the first graph is determined using a node similarity algorithm.

[0099] Since the weight of each link represents the degree of similarity between the two symptoms associated with the two connected nodes, the node similarity algorithm allows calculating pairwise similarities based on the overlap coefficient by comparing a set of nodes based on the nodes they connect. The weight of each link is a calculated pairwise similarity.

[0100] In one or more embodiments, using the symptom list and the first graph to obtain at least one gene with the highest probability of being associated with the symptom list comprises:

[0101] a. determining at least one additional symptom associated with at least one symptom of the symptom list by identifying in the nodes of the first graph at least one node connected to one of the at least one nodes corresponding to the symptom on the symptom list by a weighted link having a weight higher than a predefined threshold;

[0102] b. Calculating a subject-specific matrix comprising the genotype-phenotype matrix for symptoms associated with at least one symptom of the subject included in the symptom list and the at least one additional symptom as columns;

[0103] c. Calculate the score vector, where each coefficient of the score vector is obtained by varying the coefficient of each row of the subject-specific matrix to represent the probability that each gene of the subject-specific matrix is ​​associated with the symptom observed for the subject;

[0104] d. Using the score vector to obtain information about at least one gene that is more likely to be associated with the symptom observed for the subject.

[0105] Thus, as a result of the first graph, at least one additional symptom may be determined to enrich the symptom list based on similarity to the symptoms of the symptom list to aid in the process of determining information about at least one gene that is more likely to be associated with the symptoms observed for the subject.

[0106] In one or more embodiments, the at least one processor is further configured to rank genes associated with the symptoms observed in the subject based on the score vector.

[0107] Furthermore, the present disclosure relates to a computer program comprising software codes, which, when executed by a processor, is adapted to perform a method for decision support for a user to standardize phenotypes in genomic analysis of a subject in accordance with any of the above-mentioned execution modes.

[0108] The present disclosure also relates to a computer-readable, non-transitory program storage device tangibly embodying a program of instructions executable by a computer to perform a method consistent with the present disclosure for decision support for a user to normalize phenotypes in genomic analysis of a subject.

[0109] Such non-transitory program storage devices may be, but are not limited to, electronic, magnetic, optical, electromagnetic, infrared or semiconductor devices, or any suitable combination of the foregoing. It is readily understood that although more specific examples are provided below, those skilled in the art will readily appreciate that they are merely illustrative and not an exhaustive list: portable computer floppy disk, hard disk, ROM, EPROM (erasable programmable ROM) or flash memory, portable CD-ROM (compact disk ROM).

[0110] definition

[0111] In the present invention, the following terms have the following meanings:

[0112] The terms "adapted to" and "configured" in the present disclosure broadly include the initial configuration of the inventive device, subsequent modifications or supplements, or any similar combination thereof, whether implemented by physical or software means (including firmware).

[0113] The term "electronic health record" refers to a digital version of a patient's paper medical history. EHRs are real-time, patient-centric records that make information immediately and securely available to authorized users. Electronic health records contain a patient's medical history, diagnoses, medications, treatment plans, immunization dates, allergies, radiology images, and lab and test results.

[0114] As used herein, the term "subject" refers to a mammal, preferably a human. In one embodiment, the subject may be a "patient," i.e., a warm-blooded animal, more preferably a human, who is awaiting medical care or is receiving medical care, or has been / is / will be the subject of a medical procedure, or is being monitored for the development of a disease (e.g., cancer). The term "mammal" herein refers to any mammal, including humans, livestock and farm animals, and zoo, playground or pet animals (e.g., dogs, cats, cows, horses, sheep, pigs, goats, rabbits, etc.). Preferably, the mammal is a primate, preferably a human.

[0115] The term "processor" is not limited to being understood as hardware capable of executing software, but refers generally to a processing device, which may include, for example, a computer, a microprocessor, an integrated circuit, or a programmable logic device (PLD). A processor may also contain one or more graphics processing units (GPUs), whether for computer graphics and image processing or other functions. In addition, instructions and / or data capable of performing associated and / or resulting functions may be stored on any processor-readable medium, such as an integrated circuit, a hard disk, a CD (compact disk), an optical disk such as a DVD (digital versatile disk), a RAM (random access memory), or a ROM (read-only memory). Instructions may notably be stored in hardware, software, firmware, or any combination thereof.

[0116] The term “symptoms” refers to information derived from observation or measurement of the subject (i.e., blood pressure and other information such as weight, body dimensions), laboratory and test results such as genomic analysis, blood analysis, etc.

[0117] The term "phenotype" refers to any observable characteristic or symptom of a disease, such as morphological, developmental, biochemical, or physiological properties or behavior, without reference to any mechanism. A clinical phenotype refers to the manifestation of a disease in a given individual. Medical ontologies, such as the Human Phenotype Ontology (HPO), provide standardized vocabularies for phenotypic abnormalities encountered in human disease, and medical ontologies can be used to associate observable characteristics or symptoms with corresponding phenotypes.

[0118] The term “gene panel” refers to a test that analyzes multiple genes at once for a specific clinical indication. BRIEF DESCRIPTION OF THE DRAWINGS

[0119] Figure 1 A block diagram schematically showing a specific mode of an apparatus for gene prioritization consistent with the present disclosure.

[0120] Figure 2 FIG. 1 is a block diagram schematically showing a second mode of an apparatus for gene prioritization consistent with the present disclosure.

[0121] Figure 3 To show the use Figure 1 A flowchart of the sequential steps performed by the prediction device.

[0122] Figure 4 To show the use Figure 2 A flowchart of the sequential steps performed by the prediction device.

[0123] Figure 5 Schematically shows the integration of Figure 1 A means for prioritizing functionality of a device.

[0124] FIG. 6A to FIG. 6C is a combination of figures and tables illustrating the practice of phenotyping from a retrospective cohort of 1686 patients and a prospective experiment in which multiple physicians performed phenotyping on clinical reports. Fig. 6A A dendrogram showing the frequency of HPO terms in the retrospective cohort is shown. Figure 6B Violin plots of the HPO term counts for each clinical description for each subgroup in the cohort are shown. Figure 6C Violin plots of the HPO term counts for each clinical description after phenotyping of each clinical report by 12 physicians are shown.

[0125] 7A to 7D Combination of figures quantifying the overlap of symptom-gene associations between a retrospective multicenter cohort of 1686 patients and the medical literature. Fig. 7A Shown is a Venn diagram of symptom-gene associations observed in the cohort overlapping with associations from text mining in the public HPO structured database and free-text databases. Figure 7B The count distribution of symptom-gene associations unique to each database is shown. Figure 7C Taking KMT2D as an example, unique symptom-gene associations found in the Monarch Initiative database (top figure) and the text-mined OMIM database (middle figure) are shown. Fig.7DShown is the distribution of the average lowest distance in the ontology between exclusive terms in the Monarch Initiative database and exclusive terms in our text-mined OMIM database compared to a random selection of HPO terms.

[0126] FIG. 8A to FIG. 8C It is a combination of graphs using a node similarity algorithm to model symptom-symptom interactions in rare diseases using collected symptom-gene associations. Fig. 8A A graphical visualization of symptom relationships based on the HPO's human development architecture is shown. Figure 8B Graph visualization of symptom relationships with node similarity > 80% is shown. Figure 8C Shown are symptom relationships for Kleefstra syndrome clinical reports of EHMT1 variant carriers phenotyped by 12 geneticists.

[0127] 9A to 9C is a combination of graphs for modeling symptom-symptom interactions in genetic diseases using non-negative matrix factorization. Fig.9A Shown is a visualization of symptom relationships based on 390 groups of interacting symptoms in the medical literature. Fig. 9B A graphic representation of the 273 groups with autoimmunity (HP:0002960) as the main symptom. Fig. 9C A UMAP visualization of the clinical description of the cohort projected onto symptom dimension groups, colored and annotated by agglomerative clustering, is shown.

[0128] FIG. 10A to FIG. 10D It is a combination of graphs about modeling symptom interactions as an efficient system for phenotype matching. Fig. 10A An illustration of the phenotype matching principle is shown, searching for genes most associated with the clinical description of two symptoms observed in Kleefstra syndrome including EHMT1 variants: thinning hair (HP:0008070, yellow) and moderate global developmental delay (HP:0011343, purple). Fig. 10B Representation of performance benchmark indicators for diagnostic gene priority ranking (median ranking on the left) and phenotype matching (count of unmatched descriptions on the right) according to the maximum number of symptoms in the clinical descriptions of the cohort is shown. Fig. 10C A benchmark of selected state-of-the-art gene prioritization procedures is shown. Fig. 10D The difference in ranking after removing one symptom according to the number of terms in the clinical description is shown.

[0129] Fig.11 Dendrogram representing the HPO terms in each major category ontology cohort.

[0130] Fig.12 Violin plots representing the average depth of HPO terms from the root entity according to each database.

[0131] Fig.13 Violin plots representing similarity scores of symptom pairs according to ontological distance.

[0132] Fig.14 Represents the distribution of a 1% subsample of similarity pair scores.

[0133] Fig.15 A representation of the subject consistency evaluation for determining the optimal number of groups is shown.

[0134] Fig.16 Violin plot showing symptom-group weight associations.

[0135] Fig.17 is a dendrogram of the hierarchical clustering of clinical observations obtained using agglomerative clustering for a retrospective cohort using 16,600 symptoms from the HPO.

[0136] FIG. 18A to FIG. 18B UMAP visualization representing clinical descriptions of a retrospective cohort of 1686 patients using 16,600 symptoms from the HPO.

[0137] Fig.19 UMAP visualization showing the clinical description of the cohort projected using 16,600 symptoms from the HPO, colored and annotated with agglomerative clusters.

[0138] Fig. 20 Dendrogram of hierarchical clustering of clinical observations obtained using agglomerative clustering on a cohort projection of 390 interacting symptom dimensions.

[0139] Fig.21 UMAP visualization showing the clinical description of the cohort projected using 390 sets of interacting symptoms, colored and annotated by agglomerative clustering.

[0140] Fig. 22 Schematic diagram of the phenotype matching and gene prioritization system.

[0141] Fig.23 Represents the benchmark for prioritization of the most advanced phenotypic driver genes in each subgroup cohort. DETAILED DESCRIPTION

[0142] This specification explains the principles of the present disclosure. It should therefore be appreciated that those skilled in the art will be able to design various arrangements that, although not explicitly described or shown herein, embody the principles of the present disclosure and are included within the scope of the present disclosure.

[0143] All examples and conditional language described herein are used for teaching purposes to help readers understand the principles of the present disclosure and the concepts contributed by the inventors to advance the art, and should be interpreted as not being limited to the specific examples and conditions described.

[0144] In addition, all statements of the principles, aspects and embodiments of the present disclosure described herein and their specific examples are intended to cover their structural and functional equivalents. In addition, such equivalents include both currently known equivalents and equivalents developed in the future, i.e., any elements with the same function developed, regardless of their structure.

[0145] Thus, for example, those skilled in the art will appreciate that the block diagrams presented herein may represent conceptual diagrams of illustrative circuits embodying the principles of the present disclosure. Likewise, it should be appreciated that any flow charts, flowcharts, and the like represent various processes that can be substantially represented by computer-readable media and executed by a computer or processor, whether or not such computer or processor is explicitly shown.

[0146] The functions of the various elements shown in the figures may be provided through the use of dedicated hardware as well as hardware capable of executing software in conjunction with appropriate software. When provided by a processor, the functions may be provided by a single dedicated processor, a single shared processor, or a plurality of individual processors, some of which may be shared.

[0147] It should be understood that the elements shown in the figure can be implemented in various forms of hardware, software or a combination thereof. Preferably, these elements are implemented in a combination of hardware and software on one or more appropriately programmed general-purpose devices (which may include a processor, memory and input / output interface).

[0148] A first aspect of the present disclosure relates to identifying genes with the highest probability of being associated with symptoms observed for a subject (i.e., by creating a gene-phenotype matrix), and will be described with reference to a specific functional embodiment of an apparatus 1 for identifying genes with the highest probability of being associated with symptoms observed for a subject, and for providing decision support to normalize phenotypes in a genomic analysis of a subject, e.g. Figure 1 shown.

[0149] The device 1 is suitable for generating such information about the gene or a small number of genes in a group of genes 24 having the highest probability of causing the subject's symptoms based on subject data 20 representing at least one symptom observed for the subject and information about an association between the at least one symptom and at least one gene believed to be causing the associated at least one symptom.

[0150] In one embodiment, the information about the association between at least one symptom and at least one gene believed to cause the associated at least one symptom is a gene-phenotype matrix 21, which represents a map of genotype-phenotype associations. In more detail, the gene-phenotype matrix 21 is a matrix ,in is the number of genes, is the number of phenotypes present in the list of associations 22 between at least one phenotype and at least one gene believed to be responsible for said associated at least one phenotype.

[0151] Subject data 20, which represents observable symptoms in a subject, may be obtained from an electronic health record associated with the subject.

[0152] Although the presently described device 1 is general and has a number of functions that can be performed alternately or in any cumulative manner, other implementations within the scope of the present disclosure include devices having only a portion of the present functionality.

[0153] Device 1 is advantageously an apparatus or a physical part of an apparatus that is designed, configured and / or adapted to perform the mentioned functions and produce the mentioned effects or results. In alternative implementations, device 1 is embodied as a group of apparatuses or physical parts of apparatuses, whether grouped in the same machine or in different (possibly remote) machines. For example, device 1 may have functionality distributed on a cloud infrastructure and provided to users as a cloud-based service, or have remote functionality accessed through an API.

[0154] In the following, modules are understood to be functional entities, rather than material, physically distinct components. Therefore, they can be embodied as being combined together in the same tangible, concrete component, or distributed in several such components. In addition, each of these modules itself may be shared between at least two physical components. In addition, these modules are also implemented in hardware, software, firmware or a hybrid thereof. These modules are preferably embodied in at least one processor of the device 1.

[0155] The device 1 comprises a receiving module 11 for receiving subject data 20 stored in one or more local or remote databases 10 and information about the association between at least one symptom and at least one gene believed to be responsible for the associated at least one symptom. This information may take the form of a storage resource available from any type of suitable storage means, which may in particular be a RAM or an EEPROM (Electrically Erasable Programmable Read Only Memory), such as a Flash memory, possibly in an SSD (Solid State Disk).

[0156] Module 11 can be configured to receive a first graph 211 including nodes and weighted links. Each node represents a symptom, and each symptom is associated with at least one gene. Each link connecting two nodes represents the similarity between the two symptoms associated with the two connected nodes. The weight of each link represents the degree of similarity between the two symptoms associated with the two connected nodes. The first graph is previously obtained by calculating the symptom pair similarity using information about the association between at least one symptom and at least one gene that is considered to cause the associated at least one symptom. For example, a node similarity algorithm is used to determine the weight of each link in the first graph. The node similarity algorithm calculates pair similarity using information included in the gene-phenotype matrix. Advantageously, the first graph can be used to determine a phenotype similar to a given phenotype based on the links and associated weights that deviate from the given phenotype in the first graph.

[0157] In an alternative embodiment, if Figure 2 As shown in , the receiving module 11 can be configured to receive a second graph 212 to replace the first graph 211. The second graph 212 is previously obtained by applying matrix decomposition to a gene-symptom matrix, which can be the first graph 211 and represents the association between at least one symptom. The second graph 212 includes nodes and weighted links, wherein each node represents a symptom, and each symptom is associated with at least one gene. Each node is associated with at least one group of nodes of the second graph 212; and each node is associated with at least one specificity score, each specificity score representing the specificity of the symptom associated with the node relative to the symptoms of the nodes in each group associated with the node.

[0158] The device 1 also optionally comprises a module 12 for preprocessing the received subject data 20. The module 12 may be particularly suitable for standardizing the language used to report symptoms of the subject data 20. In particular, the module 12 is configured to associate each symptom contained in the subject data 20 with a phenotype based on a predefined ontology (e.g., a human phenotype ontology). It should be noted that the predefined ontology includes information about all phenotypes represented in the gene-phenotype matrix.

[0159] According to various configurations, the module 12 is adapted to perform only some or all of the functions described above, in any possible combination, in any manner suitable for the subsequent processing stages.

[0160] The apparatus 1 may further comprise a module 14 configured to calculate a score vector representing the probability that each gene (ie, a gene included in the first graph 211 or the second graph 212 ) is associated with a symptom observed for the subject.

[0161] Module 14 may also be configured to apply at least one correction factor and / or vector to the score vector. According to one embodiment, the score vector is corrected by a first factor designed to take into account the frequency with which variants of each gene are reported as pathogenic in a predefined medical database, such as the ClinVar database (i.e., a freely accessible public archive of reports on the relationship between human variation and phenotype). The first factor may be a pathogenicity vector, which is constructed such that each of its coefficients Represents each gene in a predefined medical database Advantageously, this first correction factor / pathogenicity vector allows for a higher weighting of genes found more frequently in the diagnosis relative to genes found less frequently in the diagnosis.

[0162] In more detail, genes can be obtained The coefficient of the pathogenicity vector is calculated as a ratio whose numerator is the maximum number of variants expressed by the gene with the largest number of variants. and genes The number of associated variants The denominator of this ratio is the maximum number of variants expressed by the gene with the largest number of variants. and the number of variants expressed by genes with fewer variants The difference (i.e. ).

[0163] Therefore, each coefficient can be Subtract the corresponding gene The score vector is updated by obtaining the coefficient of the pathogenicity vector. The pathogenicity vector may have been previously calculated and stored in the previously cited database 10. Alternatively, the module 14 may be configured to access a predefined medical database and calculate the pathogenicity vector.

[0164] Module 14 may also be configured to calculate and apply a second correction factor configured to account for gene-related factors not observed in the subject. The second correction factor may be a correction vector, wherein its coefficients are calculated from the coefficients of the rows of the subject-specific matrix. More specifically, each coefficient of the correction vector may be calculated as the coefficient of the row of the subject-specific matrix associated with the gene The second correction factor is particularly useful when only a small number of symptoms (i.e., phenotypes) are observed in a subject, because a brief clinical description is often not accurate enough to select the correct gene.

[0165] Therefore, each coefficient of the score vector can be Subtract targeted genes The coefficients of the correction vector obtained are used to update the score vector. Obviously, the first correction factor and the second correction factor can be applied separately (i.e., the score vector is corrected by applying only one correction factor) or jointly (i.e., the two correction factors are applied to the score vector continuously, regardless of the order). Alternatively, the first correction factor, the second correction factor, and the final additional correction factor can be combined into a global correction factor to be directly applied to the score vector in one step.

[0166] It can be observed that the operations of modules 11, 12, 13, and 14 are not necessarily sequential in time, and may overlap, be performed in parallel, or be performed alternately in any suitable manner. For example, new image data may be received and preprocessed incrementally over time while module 13 is processing previously acquired image data. In an alternative example, a batch of image data 21 corresponding to a complete time series range may be completely received and preprocessed before being submitted to module 13.

[0167] The device 1 further comprises a module 15 for extracting a prediction 31 from the score vector calculated by the module 14 and for outputting the prediction 31. According to one embodiment, the module 15 is configured to apply a ranking algorithm to the coefficients of the score vector (or the corrected score vector if one or more correction factors have been applied). Since the (corrected) score vector comprises The probability that each gene in the genes causes the subject's disease can be determined, so the ranking algorithm can provide a list as information, wherein the genes are ordered from the gene with the highest probability to the gene associated with a lower probability. The module 15 can be specifically configured to provide a string as the prediction 31, which specifies the gene with the higher ranking (i.e., with the highest probability of being the origin of the subject's disease).

[0168] The device 1 interacts with a user interface 16, through which a user can input and retrieve information. The user interface 16 comprises any means suitable for inputting or retrieving data, information or instructions, in particular visual, tactile and / or audio capabilities, including any or a combination of the following means familiar to those skilled in the art: screen, keyboard, trackball, touchpad, touchscreen, speaker, voice recognition system.

[0169] In one mode of the present invention, the device 1 also includes a module called a matrix module (not shown), which is configured to calculate a gene-phenotype matrix when the first graph 211 or the second graph 212 is constructed by such a gene-phenotype matrix. In this mode, the receiving module 11 is configured to receive a data set including association information between genotypes and phenotypes. The data set can be a predefined trusted synonym library, that is, a list of genotypes associated with corresponding phenotypes. The human phenotype ontology annotation (HPO) can be referenced in the reference synonym library. The synonym library can also be defined based on information present in the following data sets: Online Mendelian Inheritance in Man (OMIM®), MedGen synonym library, LitVar, etc. Those skilled in the art know synonym libraries including genotype-phenotype associations, or know how to construct synonym libraries that can be used in the present invention. At least one synonym library can be stored in the database 10 or dynamically accessed through a communication network.

[0170] The matrix module can be configured to construct a gene-phenotype matrix based on a given synonym library (i.e., list) by implementing the steps described below. The first step involves defining dimensional zero matrix, where is the number of genes, is the number of phenotypes present in the list of associations between phenotypes and genes. In one example using the HPO thesaurus, the gene-phenotype matrix will have 4531 rows (i.e., the number of genes considered) for every 15785 columns (i.e., the number of phenotype libraries). The second step includes all the coefficients of the gene-phenotype matrix 21 Assign non-zero values, coefficients Indicates gene and the phenotypes included in the list The non-zero value may be set equal to 1.

[0171] All phenotypes are linked to each other, and the interconnections between them can be represented by a hierarchical structure (also called an ontology). In a preferred embodiment, the ontology is a directed acyclic graph representing a standard classification of abnormal human phenotypes and their semantic relationships. In one example, the Human Phenotype Ontology can be used. Advantageously, the gene-phenotype matrix 21 is constructed to correctly reflect the parent-generation relationship between the phenotypes in the gene-phenotype matrix coefficients.

[0172] For this reason, the pairs are still set to 0 (i.e., the genes involved and phenotype Each coefficient of the association of Perform step 3. Each of these coefficients The value of Same gene The distance between the closest phenotypes is calculated, where the phenotype exists in the list Same gene The association between the closest phenotypes. Compute phenotypes on a directed acyclic graph and the distance between the closest phenotype. The distance can be specifically calculated as the phenotype and the same genes as those considered The number of nodes that exist between the closest phenotypes that are related. Since considering completely disconnected relationships between phenotypes would not provide correct information, a maximum distance between phenotypes is preferably set and can be, for example, equal to the distance of two nodes. Based on prior knowledge, directional barriers can also be defined on the directed acyclic graph to avoid associations between certain phenotypes that should not be considered related even if the distance is short (equal to or below the maximum distance) (for example: HP:0040195 head circumference decrease and HP:0040194 head circumference increase). As mentioned before, each of these coefficients The value of is calculated based on a function of the distance, which may be linear, exponential, Gaussian, etc. In a preferred embodiment, a linear decay function is used because linear decay advantageously helps prevent coefficient values ​​from being too low. Since the matrix may be too sparse, it is advantageous to maintain significantly high values, depending on the list or lists used.

[0173] According to one embodiment, the receiving module 11 is configured to receive more than one list and the matrix module is configured to calculate an intermediate gene-phenotype matrix for each corresponding list including the association between at least one phenotype and at least one gene. The intermediate gene-phenotype matrix of each list is calculated according to the above method. Therefore, the (final) gene-phenotype matrix is ​​calculated as a combination of the calculated intermediate gene-phenotype matrices. The combination can be a linear combination and different weights can be associated with the intermediate matrix, for example, according to the accuracy of the information included in the list used to calculate the intermediate matrix. Due to the use of lists obtained from different sources (i.e., synonym libraries), this embodiment advantageously allows a more robust estimate of the coefficients of the gene-phenotype matrix to be obtained. Therefore, the final incorrect association between the phenotype and the genotype that appears in a synonym library will be alleviated by the more accurate information included in one or more other synonym libraries.

[0174] The matrix module can also be configured to perform a normalization operation on the gene-phenotype matrix after the probability propagation step according to the phenotype relationships represented in the ontology. Advantageously, the purpose of this embodiment is to avoid giving too high a weight to phenotypes that may be too frequently associated with genes, and on the other hand to enhance phenotypes that occur less frequently and are said to be more informative. In one embodiment, normalization is performed by multiplying the gene-phenotype matrix by a normalization weight vector To perform, the weight vector The dimension is This operation results in each column of the gene-phenotype matrix being Multiply by weight (i.e., the weight vector coefficient of ).

[0175] Each coefficient weight According to phenotype is calculated relative to the frequency of occurrence of the gene represented in the gene-phenotype matrix 21. Notably, each weight can be calculated as each column of the genotype matrix The average value of the coefficients.

[0176] According to one embodiment, the apparatus 1 further comprises a panel module (not shown) configured to use the score vectors to provide a suggestion of a gene panel for the subject to be further analyzed. The gene panel is defined by the user or the database as the Australian PanelApp (https: / / panelapp.agha.umccr.org / ).

[0177] In more detail, this panel module is configured to:

[0178] - Statistically significant enrichment of symptoms in the gene panel was defined by comparing the counts of each HPO term found in the gene panel group matrix with the counts of the full matrix using Fisher's test, providing odds ratios and p-values ​​for each HPO-panel association;

[0179] - Benjamini Hochberg correction was applied by filtering HPO-panel associations according to p-value thresholds;

[0180] - provides a matrix of statistically significant HPO-panel associations, where rows are associated with panels, columns are associated with HPO terms, and coefficients are odds ratios of HPO-panel associations;

[0181] - Prioritize available gene panels by a score vector of human phenotype ontology enrichment (i.e., the sum of the logarithms of the odds ratios).

[0182] According to one embodiment, the device 1 further comprises an informative phenotype module (not shown), which is configured to suggest at least one informative phenotype based on at least one phenotype observed in the subject. An informative phenotype may be understood as a phenotype that has not yet been observed in the subject, but it may be interesting to know this phenotype in more detail, for example, suggesting a brain exploration when the only symptom currently reported is a kidney cyst.

[0183] The informative phenotype module can be configured to perform the following steps:

[0184] - Define a non-negative matrix factorization model to generalize the gene-phenotype matrix 21 as columns or topics, where The number of phenotypes below the genotype-phenotype matrix ; wherein the model is trained to provide a transformation from a row comprising approximately 15,000 coefficients with values ​​of 0 and 1 if a symptom is present to a row comprising approximately 3,200 values;

[0185] - Information about at least one phenotype is transformed into the model to suggest the most likely topic of the HPO (i.e., the latent space dimension into which all phenotypes are projected).

[0186] In some embodiments, the non-negative matrix factorization applied to the gene-symptom matrix includes a non-negative double singular value decomposition (NNDSVD) initialization step.

[0187] In some embodiments, the informative phenotype module is configured to calculate a second graph 212 based on defined T columns or themes. The second graph includes nodes and weighted links. Each node represents a phenotype in the gene-phenotype matrix 21. T groups of nodes are defined in the second graph, wherein each group of nodes corresponds to one of the T columns or themes. Each node in the second graph 212 is associated with at least one specificity score, and each specificity score represents the specificity of the phenotype associated with the node relative to the phenotype of the node in each group of nodes associated with the node. Advantageously, the second graph 212 helps to associate the phenotype associated with the node with a given node group. In other words, the second graph helps to standardize phenotypic analysis.

[0188] Alternatively, the second map may be pre-computed by another device performing the steps of the informative phenotype module described previously.The device 1 is then able to receive the second map 212 .

[0189] Each gene in the gene-phenotype matrix 21 is associated with each group of nodes in the second graph through a gene-group weighted score. For a given group of nodes, the gene's genome-weighted score represents the prevalence of the gene relative to the group of nodes (i.e., the group of phenotypes corresponding to the group of nodes).

[0190] When the informative phenotype module performs the step of converting information about at least one phenotype into a non-negative matrix factorization model to suggest topics, the informative phenotype module can evaluate the specificity of the received at least one phenotype using the specificity score of the node associated with the at least one phenotype in the second graph. If the specificity of the at least one phenotype does not meet the predefined rules, the informative phenotype module can provide the user with a suggestion of at least one new phenotype, which is calculated by module 17, obtained as a phenotype associated with a node in the first graph and connected by a link to a node associated with at least one symptom in the received symptom list.

[0191] Advantageously, at least one new phenotype may be received by the module 13 configured to calculate a subject-specific matrix, in order to calculate an updated subject-specific matrix.

[0192] According to one embodiment, the device 1 further comprises a module (not shown) configured to suggest metabolic pathways (as defined in gene ontology http: / / geneontology.org / ). Metabolic pathways are groups of genes with the same function in a cell, and thus can be considered as gene panels. Suggestions of metabolic pathways according to individual phenotypes are provided based on a similar approach to gene enrichment in clinical symptoms for panel suggestions.

[0193] The present invention also relates to a device configured to predict genes.

[0194] For example, in the automatic action of device 1, device 1 can perform the following process, such as Figure 3 As shown:

[0195] - receiving the subject data 20, the first map 211 (step 41),

[0196] - preprocessing the subject data 20 to associate each of the subject data 20 with a phenotype Associated (step 42),

[0197] - Calculate a score vector representing the probability that each gene of the subject-specific matrix is ​​associated with the symptoms observed for the subject (step 44)

[0198] - Extracting and providing a prediction 31 based on the score vector (step 45).

[0199] For example, in the automatic action of device 1, device 1 can perform the following process, such as Figure 4 As shown:

[0200] - receiving the subject data 20 and the second graph 212 (step 51 ),

[0201] - Calculate a score vector representing the probability that each gene of the subject-specific matrix is ​​associated with the symptoms observed for the subject (step 55)

[0202] - Extracting and providing a prediction 31 based on the score vector (step 45).

[0203] exist Figure 5 The specific device 9 visible in the figure embodies the apparatus 1 described above. For example, it corresponds to a workstation, a laptop, a tablet, a smartphone or a head mounted display (HMD).

[0204] The device 9 is suitable for gene prediction, related calculation of gene-phenotype matrix and decision support to standardize the phenotype in the genomic analysis of the subject. It includes the following elements, which are interconnected by an address and data bus 95, which also transmits a clock signal:

[0205] - microprocessor 91 (or CPU);

[0206] - Graphics card 92 including multiple graphics processing units (or CPUs) 920 and graphics random access memory (GRAM) 921;

[0207] - a non-volatile memory 96 of the ROM type;

[0208] - RAM 97;

[0209] - one or more I / O (input / output) devices 94, such as a keyboard, a mouse, a trackball, a webcam; other modes of introducing commands, such as voice recognition, are also possible;

[0210] - a power supply 98; and

[0211] - Radio frequency unit 99.

[0212] According to a variant, the power supply 98 is external to the device 9 .

[0213] The device 9 also includes a display screen type display device 93 directly connected to the graphics card 92 to display the composite image calculated and composed in the graphics card. The use of a dedicated bus to connect the display device 93 to the graphics card 92 provides the advantage of having a greater data transmission bit rate, thereby reducing the delay time of displaying the image composed by the graphics card. According to a variant, the display device is located outside the device 9 and is connected to it by a cable or wirelessly to transmit the display signal. The device 9, for example, includes an interface for transmission or connection through the graphics card 92, which is suitable for transmitting the display signal to an external display device, such as an LCD or plasma screen or a video projector. In this regard, the RF unit 99 can be used for wireless transmission.

[0214] It should be noted that the term "register" used hereinafter in the description of the memory 97 and the memory 921 can refer to both a low-capacity memory area (some binary data) and a large-capacity memory area (enabling the storage of an entire program, or enabling the storage of all or part of the data representing the data for calculation or display) in each of the memories mentioned. In addition, the registers of the RAM 97 and the GRAM 921 can be arranged and constituted in any way, each of them does not necessarily correspond to adjacent storage locations, and can be distributed in other ways (this particularly covers the case where one register includes several smaller registers).

[0215] When activated, the microprocessor 91 loads and executes the program included in the RAM 97 .

[0216] As will be appreciated by the skilled person, the presence of the graphics card 92 is not mandatory and may be replaced by full CPU processing and / or a simpler visualization implementation.

[0217] In various modes, the device 9 may include only the functionality of the apparatus 1. Furthermore, the implementation of the apparatus 1 may be different from standalone software, and the device or a group of devices including only part of the device 9 may be utilized through an API call or through a cloud interface.

[0218] According to another advantageous embodiment, the present invention relates to a computer-implemented method for decision support of a user to standardize phenotypes in genomic analysis of a subject, wherein the method comprises:

[0219] - receiving a symptom list comprising at least one symptom observed for the subject;

[0220] - receiving a first graph comprising nodes and weighted links, wherein:

[0221] oEach node represents a symptom, and each symptom is associated with at least one gene;

[0222] o Each link connecting two nodes represents the similarity between the two symptoms associated with the two connected nodes, and

[0223] oThe weight of each link represents the degree of similarity between the two symptoms associated with the two connected nodes;

[0224] The first graph is previously obtained by calculating symptom pair similarities using a gene-symptom matrix; the gene-symptom matrix represents the association between at least one symptom and at least one gene believed to cause the associated at least one symptom, wherein each coefficient of the gene-symptom matrix Indicates gene and symptoms The probability of association between

[0225] - receiving a second graph previously obtained by applying non-negative matrix factorization to the gene-symptom matrix, the second graph comprising nodes and weighted links, wherein:

[0226] o Each node represents a symptom, each symptom is associated with at least one gene; each node is associated with at least one group of nodes of the second graph; and each node is associated with at least one specificity score, each specificity score representing the specificity of the symptom associated with the node relative to the symptoms of nodes in each group associated with the node;

[0227] wherein each gene represented in the gene-symptom matrix is ​​associated with each group of nodes in the second graph through a gene-group weighted score;

[0228] - evaluating the specificity of the symptom list using the specificity scores of the nodes associated with at least one symptom in the second graph, and providing a suggestion of at least one symptom to the user if the specificity of the symptom list does not satisfy a predefined rule, each of the at least one suggested symptom being obtained as a symptom associated with a node in the first graph connected by a link to a node associated with at least one received symptom;

[0229] - if the user adds at least one suggested symptom to the symptom list, evaluating the specificity of the symptom list using the specificity scores of the nodes in the second graph associated with the symptoms in the symptom list including the at least one received symptom and the at least one suggested symptom;

[0230] - whenever the specificity of the symptom list satisfies a predefined rule, using said symptom list and the first graph to obtain at least one gene having the highest probability of being associated with the symptom list;

[0231] - If after providing the user with a suggestion for at least one symptom, the specificity of the symptom list does not satisfy the predefined rules:

[0232] o generating a symptom binary vector representing the symptoms in the symptom list and applying non-negative matrix factorization to the binary vector to obtain a representation of the symptom list on the groups of the second graph, the representation comprising a patient score for each group of the second graph;

[0233] o Calculate the difference between each patient score and each gene-group weighted score for each group in the second graph;

[0234] o normalizing the differences relative to a reference set of distances representing asymptomatic subjects;

[0235] o selecting at least one gene with the highest probability of being associated with the symptom list as the at least one gene with the highest normalized difference;

[0236] - Outputting at least one gene associated with the list of symptoms.

[0237] Another aspect of the invention relates to a device for decision support of a user to standardize phenotypes in genomic analysis of a subject, wherein the device comprises:

[0238] - at least one input configured to receive:

[0239] o a first graph including nodes and weighted links, wherein each node represents a symptom, each symptom is associated with at least one gene; each link connecting two nodes represents the similarity between the two symptoms associated with the two connected nodes, and the weight of each link represents the degree of similarity between the two symptoms associated with the two connected nodes;

[0240] The first graph is previously obtained by calculating the similarity of symptom pairs using a gene-symptom matrix; the gene-symptom matrix represents the association between at least one symptom and at least one gene believed to cause the associated at least one symptom, wherein each coefficient of the gene-symptom matrix Indicates gene and symptoms The probability of association between

[0241] o a second graph obtained by previously applying non-negative matrix factorization to the gene-symptom matrix, the second graph comprising nodes and weighted links, wherein each node represents a symptom, each symptom is associated with at least one gene; each node is associated with at least one group of nodes of the second graph; and each node is associated with at least one specificity score, each specificity score representing the specificity of the symptom associated with the node relative to symptoms of nodes in each group associated with the node; wherein each gene represented in the gene-symptom matrix is ​​associated with each group of nodes of the second graph by a gene-group weighted score;

[0242] - at least one processor configured to:

[0243] o evaluating the specificity of the symptom list using specificity scores of nodes associated with at least one symptom in the second graph, and providing a suggestion of at least one symptom to the user if the specificity of the symptom list does not satisfy a predefined rule, each of the at least one suggested symptom being obtained as a symptom associated with a node in the first graph connected to a node associated with at least one received symptom by a link;

[0244] o if the user adds at least one suggested symptom to the symptom list, evaluating the specificity of the symptom list using specificity scores of nodes in the second graph associated with symptoms in the symptom list including the at least one received symptom and the at least one suggested symptom;

[0245] o whenever the specificity of the symptom list satisfies a predefined rule, using the symptom list and the first graph to obtain at least one gene with the highest probability of being associated with the symptom list;

[0246] oIf after providing the user with a suggestion for at least one symptom, the specificity of the symptom list does not satisfy the predefined rules:

[0247] generating a symptom binary vector representing the symptoms in the symptom list and applying non-negative matrix factorization to the binary vector to obtain a representation of the symptom list on the groups of the second graph, the representation comprising a patient score for each group of the second graph;

[0248] Calculate the difference between each patient score and each gene-group weighted score for each group in the second graph;

[0249] Normalizing the differences relative to a reference set of distances representing asymptomatic subjects;

[0250] Selecting at least one gene with the highest probability of being associated with the symptom list as the at least one gene with the highest normalized difference;

[0251] - at least one output configured to provide at least one gene associated with a list of symptoms.

[0252] In one or more embodiments, the non-negative matrix factorization applied to the gene-symptom matrix is ​​a non-negative double singular value decomposition (NNDSVD) initialization.

[0253] In one or more embodiments, using the symptom list and the first graph to obtain at least one gene with the highest probability of being associated with the symptom list comprises:

[0254] - determining at least one additional symptom associated with at least one symptom of the symptom list by identifying, among the nodes of the first graph, at least one node connected to one of the at least one nodes corresponding to the symptoms on the symptom list by a weighted link having a weight higher than a predefined threshold;

[0255] - calculating a subject-specific matrix comprising the genotype matrix for the symptoms associated with at least one symptom of the subject included in the symptom list and the at least one additional symptom as columns;

[0256] -Calculate the score vector, where each coefficient of the score vector is obtained as a function of the coefficients of each row of said subject-specific matrix to represent the probability that each gene of the subject-specific matrix is ​​associated with the symptom observed for the subject;

[0257] - using the score vector to obtain information about at least one gene that is more likely to be associated with the symptom observed for the subject.

[0258] In one or more embodiments, the at least one processor is further configured to rank genes associated with the symptoms observed in the subject based on a base of the score vectors.

[0259] Example

[0260] The following examples further illustrate the present invention.

[0261] The following example illustrates how the previously described apparatus and methods can be practically used to compute a symptom-symptom association map, and its application in identifying one or more genes with the highest probability of being the origin of a disease manifested by a subject.

[0262] Materials and methods

[0263] Clinical data collection

[0264] Anonymous clinical records from four international cohorts were collected, resulting in a total of 1686 patients with genetic diagnoses and their clinical descriptions in HPO terms. The cohort included 307 patients from the PhenoGenius Consortium of the Grenoble Alpes University Hospital Centre (CHU), CHU de Dijon, CHU de Montpellier, CHU de Rennes, CHU deBrest, and Hospices Civils de Lyon, 140 from Seo et al. (1), 298 from Trujillano et al. (2), and 941 from Peng et al. (3). Clinical descriptions in HPO format were also collected from 12 clinical geneticists from 12 different French hospitals (CHU Lille, CHU Montpellier, CHU Rennes, CHU Rouen, CHU La Réunion, CH Alençon, CHU Poitiers, CHU Limoges, CH Versailles, CHU Toulouse, and CHU Tours). Each physician extracted HPO terms from the same three clinical reports of patients with different diagnostic genes (KMT2D, KMD6A, and C3), one case each from three physicians at French hospitals (CHU Montpellier, CHU Grenoble Alpes, and APHP). Patients or legal guardians provided informed written consent for genetic analysis to be performed at the medical institution. Consent was obtained from clinical geneticists through a survey, which also collected their responses.

[0265] Medical literature database

[0266] Databases were downloaded in May 2022. The Human Phenotype Ontology was downloaded in OBO format from the Monarch Initiative website (https: / / HPO.jax.org / ). Clinical databases in HPO format were downloaded from the Monarch Initiative website in phenotype-to-gene format, EBI Initiative DD2GP (4) CSV files were downloaded from https: / / www.ebi.ac.uk / gene2phenotype / , and Orphanet XML data were downloaded from https: / / www.orpha.net / . Free text databases were downloaded through API requests for OMIM (https: / / www.omim.org / ), NCBI’s MedGen (5) (https: / / www.ncbi.nlm.nih.gov / medgen / ), and NCBI’s PubMed abstracts. For PubMed abstracts, we used a list of all possible pathogenic and pathogenic variants from the ClinVar database (6) to select abstracts of potential interest using the LitVar (7) API.

[0267] Text matching algorithm

[0268] method

[0269] The inventors developed a method to extract symptom-gene associations from free text data in HPO terms and NCBI gene ID format based on Elasticsearch® v5.6. First, these databases were processed to associate the free text data with the corresponding genes in JSON format. Elasticsearch queries were executed to match each HPO available in each gene-free text-associated data and provide a list of HPO-gene associations for each database. The number of gene-HPO associations created was limited to the top 100 associated genes in the HPO ranking.

[0270] Average distance between unique terms in the ontology

[0271] For each exclusive symptom-gene association from the MI database, the ontology distance of each exclusive symptom-gene association in the text-mined OMIM in a common gene was calculated and the minimum distance was kept. For each gene, we processed the average distance between the exclusive terms in the MI database and our text-mined OMIM database in the ontology. We also performed the same experiment in the opposite direction, that is, from the exclusive associations in the text-mined OMIM database to the MI database. To compare the average distance distribution with the random distribution, we calculated the ontology distance of randomly selected HPO terms (not the exclusive terms from other databases).

[0272] Knowledge Data Framework Structure

[0273] The symptom-gene association list for each database is stored in a data frame consisting of 16,600 symptoms (columns) and 5,235 genes (rows). Each cell includes the probability of a symptom-gene association, which depends on the overlap between databases (consensus score based on the mean, e.g., an association found in half of the databases is given a score of 0.5). Databases built in HPO format (MI, DDG2P, and Orphanet) were considered as unique resources, as the associations provided by DDG2P and Orphanet mostly overlap with the MI database and have only 3,492 and 1,849 unique associations, respectively.

[0274] Node Similarity

[0275] The data frame of symptom-gene associations was converted into a data frame of symptom-symptom associations based on symptom associations in the same gene. All existing symptom-to-symptom relationships were injected into a Neo4j® database v4.4.0. The similarity between all pairs of symptoms was processed using the node similarity algorithm (https: / / neo4j.com / docs / graph-data-science / current / algorithms / node-similarity / ). Due to technical reasons (RAM limitations due to the number of combinations), for each symptom, only symptoms with a similarity score > 0.4 were extracted, with a maximum of 1500 related symptoms. If the similarity score was greater than 0.8, the existence of a pair of similar symptoms was reported.

[0276] Collaborative filtering

[0277] method

[0278] The symptom-gene association data frame was converted to a symptom group based symptom association data frame in the same gene using scil-kit learn v.0.24.2 non-negative matrix factorization and non-negative double singular value decomposition initialization. The algorithm provides the requested number of groups, weights of symptom group associations, and weights of genomic associations. For interpretability and illustration, the symptom group associations were filtered to keep only the top 10% of the l2-normalized weights (>0.04) of the symptom group associations. In the phenotype matching assessment, the full symptom group associations were used.

[0279] Determining the optimal number of symptom groups

[0280] A consistency score metric was applied to the processed symptom groups to select their optimal number. The gensim v4.2 implementation of the Consistency Subject Evaluation (https: / / radimrehurek.com / gensim / models / coherencemodel.html) was used to find the range of group numbers with the highest consistency score and lowest coefficient of variance using five random state initializations. The consistency of the consistency between different random states was found to reproduce the same performance with additional data or updates. This is necessary for clinical implementation.

[0281] UMAP and Clustering

[0282] Agglomerative clustering (affinity="euclidean" and linkage="ward" parameters) of Sci-kit learn v0.24.2 was implemented to obtain hierarchical clustering. Data were normalized according to clinical observations and symptom groups. 75 clusters of clinical descriptions were retained to avoid single observation clusters. Dendrograms were obtained using Scipy v1.8 hierarchical module. UMAP visualization was performed using umapv0.5.3 module with the following parameters: number of neighbors = 3 and minimum distance = 0.9 for 390 groups of symptom dimensions and number of neighbors = 2 and minimum distance = 0.9 for 16,600 symptom dimensions. These parameters were determined after observing the dispersion and clustering of clinical descriptions according to the range of number of neighbors (2 to 5) and minimum distance (0.1 to 0.99).

[0283] Graph Visualization

[0284] The exploration graphs were processed using the Python package plotnine v0.9. The graph visualization was obtained using the software Gephi v0.9 (https: / / gephi.org / ) with ForceAtlas2 (8) and Neo4j Bloom v2.3. Retina (https: / / ouestware.gitlab.io / retina / beta / ) provided users with a visual browser for the graphs.

[0285] Phenotypic Matching

[0286] A phenotype match occurs if at least one symptom in the clinical description is associated with a diagnostic gene in the database. The inventors developed a phenotype matching system that matches clinical descriptions with a list of symptom-gene associations available in a knowledge data frame structure. Based on the combination of symptoms from the clinical description, the data frame is filtered to include only selected symptoms or groups of symptom columns. The sum of the consensus scores for each row or gene is processed and the genes are ranked in descending order based on the total score. In the case of equal scores, we applied the worst ranking to all equal genes. When evaluating this phenotype match using an HPO format database and associations of text mining, only the symptoms declared in the clinical description were used. The phenotype matching system based on symptom similarity uses the declared symptoms and adds a virtual symptom that includes all highly similar symptoms and shares their weights. Based on the NMF collaborative filtering method, the symptoms are projected into 390 sets of interacting symptom dimensions using a trained model. Each gene is also projected into the 390 sets of dimensions with different weights. The ranking is calculated based on the standardized Euclidean distance between the 390 sets of projections of each gene and the patient's phenotype. Genes with the largest distances receive the best ranking.

[0287] Comparison of phenotype-driven gene prioritization systems

[0288] Four different phenotype-driven gene prioritization algorithms were implemented: PhenoApt, Phen2Gene, CADA, and LIRICAL (3, 9–11). Because each tool provides a different maximum for the gene ranking list, the diagnostic performance evaluation was based on the top 100 cumulative causal gene rankings that could be interpreted. PhenoApt (https: / / www.phenoapt.org / API) and Phen2Gene (https: / / phen2gene.wglab.org / api) were evaluated using the software's API in May 2022. CADA (https: / / github.com / Chengyao-Peng / CADA, exclusive release) and LIRICAL (https: / / github.com / TheJacksonLaboratory / LIRICAL, v.1.3.4) evaluations were processed using the desktop version through GitHub.

[0289] statistics

[0290] Statistical measures (Kolmogorov-Smirnov test, Spearman correlation coefficient, Fisher exact test) were obtained using Python's SciPy module v1.8. Benjamini-Hochberg correction was obtained using the multipy package v0.16.

[0291] result

[0292] Examples 1 and 2 are intended to illustrate the heterogeneity in current phenotyping practices and the mismatch between symptom-gene associations found in a given clinical study conducted in a predetermined cohort and those found in the medical literature.

[0293] Example 1 : Practice of phenotyping in large cohorts

[0294] In this example, the results of a study conducted by the inventors on current phenotyping practices are described to illustrate the heterogeneity in current phenotyping practices.

[0295] Table 1 shows the results of clinical data collection for 1686 patients in an international cohort. So far, no studies have been conducted on the practice of phenotyping in clinical sequencing. Through 4 international studies, 2501 different symptoms in HPO format and 849 different pathogenic genes were collected. 5-7 .

[0296] Table 1: Clinical data collection of 1686 patients in the international cohort

[0297]

[0298] Table 2 shows the top 10 recurring genes in the four groups in the cohort.

[0299] Table 2: Top 10 recurring genes in the 4 groups in the previous cohort

[0300]

[0301] In the multicenter cohort, nearly half of the patients had symptoms belonging to the categories of neurological disorders (HP: 0000707) and musculoskeletal disorders (HP: 0033127), which illustrates the current attention paid to these rare diseases in clinical practice, e.g. Fig.11 As shown, Fig.11 The following is a dendrogram of the HPO terms for each main ontology in the studied cohort. 538 of the 849 genes were declared only once in the cohort, reflecting the genetic heterogeneity of rare diseases. The most commonly mutated genes occurred in less than 2% of cases (Table 2: ABCC6, n=21).

[0302] Table 3 shows the 10 most common HPO terms in the 4 groups in the cohort.

[0303] Table 3: The 10 most common HPO terms in the 4 groups in the cohort

[0304]

[0305] Heterogeneity in the HPO selection of terms was observed, as 47% of the terms were used only once (in Fig. 6A As shown in Table 3). Figure 6B As shown, this figure is a violin plot of the HPO term counts for each clinical description in each of the four subgroups in the study cohort. The median number of HPO terms per physician clinical description varied across observations, ranging from 3 (Peng et al. 7 ) to 7 (PhenoGenius Alliance, Seo et al., and Trujillano et al. 5,6 ). For patients with the same genetic diagnosis, there was also heterogeneity in the clinical descriptions given by physicians. For genes involved in the diagnosis of more than 10 patients, 67% of the symptoms were mentioned in only one clinical description.

[0306] Table 4 is a description of the clinical geneticist profile for prospective phenotyping trials.

[0307] Table 4: Description of the clinical geneticist profile in prospective phenotyping trials

[0308]

[0309] To exclude the possibility that the observed heterogeneity was caused by variability in clinical examinations, the inventors next investigated whether there was heterogeneity in clinical descriptions if physicians phenotyped the same clinical observations. We performed a prospective experiment in which 12 clinical geneticists with different levels of expertise (shown in Table 4) were asked to phenotype three independent clinical reports associated with genetic test prescriptions, i.e., to convert free text into phenotypes in HPO format. Figure 6C As shown in the violin plots of , we observed heterogeneity in the number and diversity of symptoms stated in each clinical observation. These are violin plots of the HPO term counts for each clinical description after phenotyping each clinical report by 12 clinical geneticists. For example, between 2 and 9 symptoms were listed in the clinical descriptions of the Kleefstra syndrome observations with EHMT1 pathogenic variants. A total of 29 different terms were provided; 17 of these terms were used by 2 or more physicians, but no single term was mentioned by all 12 physicians.

[0310] Example 2 Quantifying the overlap of symptom-gene associations between retrospective cohorts and the medical literature

[0311] In this example, the results of a study aim to quantify how the symptom-gene associations obtained from the retrospective cohort of Example 1 overlap with those obtained from the medical literature.

[0312] To assess whether the clinical description of the cohort used in Example 1 matched existing knowledge in the medical literature, the 11,526 unique symptom-gene associations of the cohort were compared with the HPO structured databases (Orphanet, DDG2P 9 and Monarch Initiative or MI 3 ) were mapped. From these databases, only 4913 associations (43%) were matched, which means that 57% were missing, such as Fig. 7A As shown, Fig. 7A is a Venn diagram of symptom-gene associations observed in the cohort overlapping with text-mined associations in the public HPO structured and free-text databases.

[0313] Since clinical descriptions of genetic diseases in medical literature are mainly provided in free text format, a text mining algorithm was developed based on Elasticsearch to extract symptom-gene associations from free text data in HPO format. 10 、MedGen 11 and PubMed abstracts, found an additional 1,049,522 symptom-gene associations. Figure 7BAs shown, this approach increased the database association of HPO structures by 3.2 times, Figure 7B Represents the count distribution of symptom-gene associations unique to each database.

[0314] Compared to the HPO structured database, the text mining algorithm provided symptom-gene associations where the symptoms were significantly deeper in the ontology (i.e., the text mining database). Fig.12 The violin plots show the average depth of HPO terms from the root entity in each database, with corresponding median depths of 6.7 and 5.2, respectively, and Kolmogorov-Smirnov test p-values ​​< 10 -215 This emphasizes the complementary nature of these approaches, e.g. Figure 7C As shown, KMT2D is associated with abnormal large vessel morphology in the Monarch Initiative database (HP: 0030962) and tetralogy of Fallot in the OMIM database (HP: 0001636). This reflects the variability between individuals in the selection of HPO terms to summarize clinical observations, with 76% of associations being unique to a particular database. The inventors hypothesized that the symptom-gene associations from text mining in the literature are related to the associations available in the HPO structured database. This assumption embodies the concept of fuzzy phenotypes, which provide alternative ways of expressing the same information determined by humans.

[0315] To evaluate this hypothesis, for each gene, the average distance in the ontology between the exclusive symptom-gene associations in the MI database (the largest structured database) and the exclusive symptom-gene associations in the text-mined OMIM database (the largest text-mined database) was compared, as shown in Fig.7D shown. Fig.7D Figure 2 shows the distribution of the average lowest distances in the ontology between proprietary terms in the MI database (left graph) and proprietary terms in the text-mined OMIM database (right graph) compared to randomly selected HPO terms (curve on the right side of each graph). The mean distances for proprietary symptom-gene associations are significantly lower than those for randomly selected HPO terms, indicating that these associations are relevant. Fig.7D In the results, the Kolmogorov-Smirnov test p value is less than 10 -215 .

[0316] Although the number of symptom-gene associations in this trial increased from 734,931 (MI, DDG2P, Orphanet databases) to 1,784,453 (associations discovered by text mining algorithms), only 6226 of 11,526 associations (57%) were matched with the symptom-gene associations of the cohort, meaning that 43% of the matches were still missing, e.g. Fig. 7A The Venn diagram is shown in Figure 2.

[0317] Example 3 : From symptom-gene to symptom-symptom association modeling

[0318] In this example, it will be shown how the symptom-symptom association modeling proposed in this application improves the coverage of symptom-gene associations in a cohort. In particular, it will be shown how modeling associations between symptoms of the same genetic disease improves matching. Symptom-symptom association modeling corresponds to the first figure 211 described previously.

[0319] Since the Human Phenotype Ontology is ordered according to human development, it may not represent the interaction of symptoms in diseases, such as Fig. 8A As shown, different grayscales represent different symptoms. Therefore, we explored an alternative approach to measure symptom-symptom associations in genetic diseases. More specifically, we implemented a node similarity algorithm based on a knowledge graph that stores symptom-gene associations collected from the literature.

[0320] The results of this implementation were that a high correlation was found between the symptom-symptom similarity pair scores and their co-occurrence frequency in clinical observations (Spearman correlation coefficient: 0.99). No correlation was observed between the symptom-symptom similarity pair scores and the distance between symptoms in the HPO (Spearman correlation coefficient: -0.02, Fig.13 ), which reflects that symptom-symptom associations cannot be derived from the ontological framework alone.

[0321] Based on the similarity score distribution, it is inferred that diseases with a similarity of more than 80% may be synonyms or highly similar symptoms. The threshold is Fig.14 As a result, 565,943 pairs of highly similar symptoms were selected, corresponding to the 10% highest symptom-symptom association scores ( Figure 8B Of these pairs, a total of 26% were observed for symptoms in the same ontology category (145,611 out of 565,943), with the majority coming from the musculoskeletal system disorders (HP:0033127) category (51%, 73,817 out of 145,611). Cross-category symptom pairs represented 74% of highly similar symptoms, with the most common pair being metabolic / homeostasis disorders (HP:0001939) and nervous system disorders (HP:0000707) (8%, 35,476 out of 420,332).

[0322] These similarities can be seen in Figure 8C , Figure 8CThe symptom of hypotonia (HP:0001290) reported by 6 of 12 physicians in our study of Kleefstra syndrome with EHMT1 pathogenic variants was used. Dashed arrows connect the closest symptoms in the HPO ontology, and plain arrows connect symptoms with the highest node similarity among the declared symptoms. In the symptom-symptom association graph, the closest term to hypotonia is neurodevelopmental delay (HP:0012758), with a symptom-symptom similarity pair score of 86%. In the HPO, these symptoms are separated by 10 nodes belonging to two different major categories: musculoskeletal system disorders (HP:0033127) and nervous system disorders (HP:0000707). Figure 8C is an example of a portion of the first figure shown above.

[0323] We also investigated the extent to which considering two highly similar symptoms as synonyms improved coverage of symptom-gene associations. Of the 11,526 unique symptom-gene associations in this cohort, only 6,226 associations were found in the HPO structured database and text-mined databases, but this number rose to 8,350 after accounting for similarity. Considering synonyms provided an additional 1,506,469 symptom-gene associations to the previous 1,784,453 associations from the MI, DDG2P, Orphanet, and text-mined databases.

[0324] It can therefore be concluded that modeling the associations between symptoms showed that most of the inter-HPO classes included similar symptoms, highlighting the missing aspects of symptom relationships in the HP ontology. Enhancing symptoms with highly similar pairs improved the coverage of symptom-gene associations in the cohort. However, 27% of associations were still missing.

[0325] It will be shown in the next example (Example 4) that the 27% missing association can be improved by the symptom group modeling described in this disclosure.

[0326] Example 4 :From symptom-gene association to symptom group modeling

[0327] In Example 3, when constructing synonyms based on node similarity, symptom-symptom associations were evaluated independently. To better cover symptom-gene associations, a more sophisticated collaborative filtering approach based on non-negative matrix factorization (NMF) is used in this example. More specifically, Example 4 illustrates the computation of the second graph 212 as described previously.

[0328] Using thematic consistency measures, we determined that the 16,660 HPO terms could be optimally reduced to 390 sets of interacting symptom or phenotypic patterns ( Fig.15, where each cycle corresponds to one of the iterative training experiments initialized from a random state). The position of each symptom in the graph is determined by the algorithmically determined group weight, e.g. Figures 9A-9B shown. Fig.9A Visualization of symptom relationships based on 390 groups of interacting symptoms, where bold boxes indicate specific groups (273 groups). Fig. 9B In more detail, the main symptom of group 273 is autoimmunity (HP:0002960). Fig.9A and 9B In the figure, the line thickness is proportional to the weight of the symptom in the group. Groups are shown in black. For ease of reading, only the top 10% of symptom-group associations are shown. Each gene is associated with an average of 36 groups, and a group contains an average of 501 genes. To compare the recall of NMF and the node similarity model, we only kept the top 10% of the 390 symptom-group weights, as shown in Fig.16 This is a violin plot of the weighted associations of the symptom groups. Fig.16 In the figure, the dashed horizontal line corresponds to the top 10% of the normalized weight thresholds for symptom-group associations (greater than 0.04). Overall, there are 43,308 symptom-group associations leading to 5,971,755 pairs of symptoms in this selection.

[0329] We also investigated to what extent the coverage of symptom-gene associations was improved by taking into account that two symptoms belonging to the same group were synonymous. Using these symptom-association pairs extended the coverage of symptom-gene associations to 11,340 of the 11,526 associations in the cohort, with less than 2% missing matches. This new association detection approach resulted in an additional 2,163,663 NMF-based symptom-gene associations to the 1,784,453 associations previously obtained from MI, DD2P, Orphanet, and text mining databases. The NMF-based symptom-gene associations overlapped 99% (1,497,601 of 1,506,469) with similarity-based associations.

[0330] To assess whether these 390 phenotypic patterns represent the clinical spectrum of genetic diseases, we projected the cohort into the symptom group dimension and performed UMAP visualization (Uniform Manifold Approximation and Projection) 14 We applied agglomerative clustering to the cohort and compared the performance of clustering patients using projections and the 16,600 HPO dimensions. Using the initial list of 16,600 symptoms, 152 patients were found in 14 clusters with significant enrichment of symptoms (see Figures 17 to 19 ) (Fisher's exact test with Benjamini Hochberg correction, p value < 0.05,). Applying the projection of symptom groups, it was found that 1136 patients were distributed in 51 clusters that were significantly enriched in symptom groups, such as Fig. 20 and Fig.21 The projection was applied to the three clinical reports of phenotype analysis by 12 physicians in our experiment listed in Table 4 to evaluate whether the projection can standardize clinical descriptions. When a sufficient number of HPO terms are given (KMT2D reports), the projection method shows a high degree of consistency even in the presence of symptom heterogeneity, see Fig. 9C , Fig. 9C UMAP visualization representing the projection of the clinical descriptions of the cohort onto groups of symptom dimensions, colored and annotated according to agglomerative clustering. When fewer than 5 terms were provided, the clinical description predictions still grouped patients, but with less homogeneity (EHMT1, C3).

[0331] These results therefore suggest that by following a more sophisticated collaborative filtering approach based on non-negative matrix factorization (NMF), it is possible to describe 390 sets of interacting symptoms, thereby expanding the coverage of available knowledge on genetic diseases and providing a way to standardize clinical descriptions based on the HP ontology.

[0332] Example 5 Symptom interaction model as an efficient and robust system for phenotype matching

[0333] In this example, we describe the use of the previously described symptom interaction modeling, as well as the methods illustrated in Examples 3 and 4, to develop a phenotype matching system.

[0334] To evaluate the clinical relevance of the symptom interaction model, the inventors designed and implemented a phenotype matching experiment and a diagnostic gene ranking experiment. When at least one symptom in the clinical description is associated with a diagnostic gene, it is defined as a phenotype match, such as Fig. 10A shown. Fig. 10A The principle of phenotypic matching is illustrated to find the genes most associated with the clinical description (including two symptoms observed in Kleefstra syndrome with EHMT1 variants: thinning hair (HP:0008070, yellow) and moderate global developmental delay (HP:0011343, purple). Therefore, Fig. 10A Two examples of node groups for a typical second graph introduced earlier are shown. The line thickness is proportional to the probability score of the symptom-gene associations available in the joint HPO structured and text-mined database. Based on the match count for each gene, a personalized gene ranking list is provided, such as Fig. 22 These experiments conducted clinical observations on 1,686 patients.

[0335] Using HPO structured databases (MI, DDG2P, Orphanet), 1566 phenotypic matches were obtained for clinical observations, with a median ranking of diagnostic genes of 251. After applying text mining associations, 1628 matches were obtained for clinical observations, with a median ranking of 40. Fig. 10B Two performance benchmarks are illustrated: diagnostic gene prioritization ranking based on the maximum number of symptoms in the cohort clinical descriptions (median ranking on the left) and phenotype matching (count of mismatched descriptions on the right). The best performing diagnostic median ranking was node similarity symptom association ( Fig. 10B ) (median rank 37 vs. 58 for non-negative matrix factorization - NMF), but NMF (curve #4) was able to more fully cover the clinical observations (1682 terms vs. 1663 for node similarity). This coverage gap was only observed when the clinical descriptions included 5 or fewer terms (4 unmatched descriptions vs. 25 for node similarity). Since each symptom interaction model provides a different degree of inductive reasoning, one of the two models was applied conditionally based on the number of symptoms in the clinical descriptions. The Applicant's PhenoGenius combined system performed best (median rank 41) and achieved almost complete phenotypic matching of diagnostic genes for all clinical descriptions (99.8%, 1682 / 1686).

[0336] To illustrate this phenotype matching system, we considered clinical descriptions that include two symptoms observed in Kleefstra syndrome with EHMT1 pathogenic variants: thinning hair (HP:0008070) and moderate global developmental delay (HP:0011343). These terms were not matched to EHMT1 in the HPO structured database. No matches were found from text-mined symptom-gene associations. Phenotype matching was achieved according to the symptom interaction model of the present disclosure, ranking 1244 of 5235 genes (top 25% of genes) using the similarity model and 851 of 5235 genes (top 17% of genes) using the NMF model and the PhenoGenius combined system.

[0337] The combined system (PhenoGenius) was compared with four recently published phenotype-driver gene prioritization algorithms: PhenoApt, Phen2Gene, CADA, and LIRICAL 7,15-17 Despite using different prioritization methods, the four programs showed similar performance in phenotype matching, e.g. Fig. 10CAs shown, the top curve corresponds to the combined system, while the similarly positioned curves below represent the four algorithms being compared. Using symptom interaction modeling, the phengenius combined system (median ranking of 41) improved the median ranking of diagnostic genes by 42% compared to the best competitor Phen2Gene (median ranking of 71, while the median rankings of other methods ranged from 73-80). This improvement was replicated in every studied subgroup in the cohort, highlighting the clinical relevance of symptom interactions in genetic disease models. Fig.23 The results of this benchmark are shown, and on each graph the top curve corresponds to the results obtained using the PhenoGenius combined system.

[0338] A further study was performed aimed at evaluating the robustness of gene prioritization. For each symptom from the clinical descriptions, we randomly removed two or more terms and measured the impact on the ranking of causal genes for the top half of descriptions (ranked 41 or lower) in the cohort. Overall, 701 clinical descriptions resulted in 6331 symptom removal experiments. Fig. 10D As shown, in most cases, phenotype matching remained robust after symptoms were removed. The causative genes were ranked identically in 35% of cases (2274 of 6331 cases), and the median absolute difference between the rankings was only 1. However, nine extreme drops in ranking (>1000) were observed in clinical descriptions with three or fewer terms, including two complete losses of phenotype matching for descriptions with two symptoms. No extreme drops in gene ranking were found for clinical descriptions with four or more terms.

[0339] References in the Materials and Methods section

[0340] 1. GH Seo, T. Kim, IH Choi, J.-Y. Park, J. Lee, S. Kim, D.-G.Won, A. Oh, Y. Lee, J. Choi, H. Lee, HG Kang, HY Cho, MH Cho, YJKim, YH Yoon, B.-L. Eun, RJ Desnick, C. Keum, BH Lee, Diagnostic yield and clinical utility of whole exome sequencing using an automatedvariant prioritization system, EVIDENCE. Clin. Genet. 98, 562–570 (2020).

[0341] 2. D. Trujillano, AM Bertoli-Avella, K. Kumar Kandaswamy, MEWeiss, J. Köster, A. Marais, O. Paknia, R. Schröder, JM Garcia-Aznar, M.Werber, O. Brandau, M. Calvo Del Castillo, C. Baldi, K. Wessel, S. N. Kishore. Nahavandi, W. Eyaid, MT Al Rifai, A. Al-Rumayyan, W. Al-Twaijri, A.Alothaim, A. Alhashem, N. Al-Sannaa, M. Al-Balwi, M. Alfadhel, A. Rolfs, R.Abou Jamra, Clinical exome sequencing: results from 2019 samples reflecting families. Time. J. Hum. Genet. 25, 176–182 (2017).

[0342] 3. J. Peng, S. Dieck, A. Schmid, A. Ahmad, A. Knaus, M. Wenzel, L.Mehnert, B. Zirn, T. Haack, S. Ossowski, M. Wagner, T. Brunet, N. Ehmke, M.Daniel, S. Rosnev, T. Namphans, N. F. Kamphischer, N.H. Fröhlich, P.Krawitz, CADA: phenotype-driven gene prioritization based on a case-enrichedknowledge graph. NAR Genome Bioinform. 3, lqab078 (2021).

[0343] 4. A. Thormann, M. Halachev, W. McLaren, D. J. Moore, V. Svinti, A.Campbell, S. M. Kerr, M. Tischkowitz, S. E. Hunt, M. G. Dunlop, M. E. Hurles,C. F. Wright, H. V. Firth, F. Cunningham, D. R. FitzPatrick, Flexible andscalable diagnostic filtering of genomic variants using G2P with Ensembl VEP.Nat. Commun. 10, 2373 (2019).

[0344] 5. D. N. Louden, MedGen: NCBI’s Portal to Information on MedicalConditions with a Genetic Component. Med. Ref. Serv. Q. 39, 183–191 (2020).

[0345] 6. M. J. Landrum, S. Chitipiralla, G. R. Brown, C. Chen, B. Gu, J.Hart, D. Hoffman, W. Jang, K. Kaur, C. Liu, V. Lyoshin, Z. Maddipatla, R.Maiti, J. Mitchell, N. O’Leary, G. R. Riley, W. Shi, G. Zhou, V. Schneider,D. Maglott, J. B. Holmes, B. L. Kattman, ClinVar: improvements to accessingdata. Nucleic Acids Res. 48, D835–D844 (2020).

[0346] 7. A. Allot, Y. Peng, C.-H. Wei, K. Lee, L. Phan, Z. Lu, LitVar: asemantic search engine for linking genomic variant data in PubMed and PMC.Nucleic Acids Res. 46, W530–W536 (2018).

[0347] 8. M. Jacomy, T. Venturini, S. Heymann, M. Bastian, ForceAtlas2, acontinuous graph layout algorithm for handy network visualization designedfor the Gephi software. PLoS One. 9, e98679 (2014).

[0348] 9. Z. Chen, Y. Zheng, Y. Yang, Y. Huang, S. Zhao, H. Zhao, C. Yu, X.Dong, Y. Zhang, L. Wang, Z. Zhao, S. Wang, Y. Yang, Y. Ming, J. Su, G. Qiu,Z. Wu, T. J. Zhang, N. Wu, PhenoApt leverages clinical expertise toprioritize candidate genes via machine learning. Am. J. Hum. Genet. 109, 270–281 (2022).

[0349] 10. M. Zhao, J. M. Havrilla, L. Fang, Y. Chen, J. Peng, C. Liu, C.Wu, M. Sarmady, P. Botas, J. Isla, G. J. Lyon, C. Weng, K. Wang, Phen2Gene:rapid phenotype-driven gene prioritization for rare diseases. NAR GenomBioinform. 2, lqaa032 (2020).

[0350] 11. P. N. Robinson, V. Ravanmehr, J. O. B. Jacobsen, D. Danis, X. A.Zhang, L. C. Carmody, M. A. Gargano, C. L. Thaxton, UNC Biocuration Core, G.Karlebach, J. Reese, M. Holtgrewe, S. Köhler, J. A. McMurry, M. A. Haendel,D. Smedley, Interpretable Clinical Genomics with a Likelihood Ratio Paradigm.Am. J. Hum. Genet. 107, 403–417 (2020).

[0351] General References

[0352] 1. Dewey, F. E. et al. Distribution and clinical impact of functionalvariants in 50,726 whole-exome sequences from the DiscovEHR study. Science354, (2016).

[0353] 2. Robinson, P. N. Deep phenotyping for precision medicine. HumanMutation vol. 33 777–780 Preprint at https: / / doi.org / 10.1002 / humu.22080(2012).

[0354] 3. Köhler, S., Kindle, G. & Robinson, P. N. The Human PhenotypeOntology in 2021. (2021).

[0355] 4. Haendel, M. A., Chute, C. G. & Robinson, P. N. Classification,Ontology, and Precision Medicine. N. Engl. J. Med. 379, 1452–1462 (2018).

[0356] 5. Seo, G. H. et al. Diagnostic yield and clinical utility of wholeexome sequencing using an automated variant prioritization system, EVIDENCE.Clin. Genet. 98, 562–570 (2020).

[0357] 6. Trujillano, D. et al. Clinical exome sequencing: results from 2819samples reflecting 1000 families. Eur. J. Hum. Genet. 25, 176–182 (2017).

[0358] 7. Peng, C. et al. CADA: phenotype-driven gene prioritization basedon a case-enriched knowledge graph. NAR Genom Bioinform 3, lqab078 (2021).

[0359] 8. Osmond, M. et al. PhenomeCentral: 7 years of rare diseasematchmaking. Hum. Mutat. 43, 674–681 (2022).

[0360] 9. Thormann, A. et al. Flexible and scalable diagnostic filtering ofgenomic variants using G2P with Ensembl VEP. Nat. Commun. 10, 2373 (2019).

[0361] 10. Amberger, J. S., Bocchini, C. A., Schiettecatte, F., Scott, A. F.& Hamosh, A.OMIM.org: Online Mendelian Inheritance in Man (OMIM®), an onlinecatalog of human genes and genetic disorders. Nucleic Acids Res. 43, D789–98(2015).

[0362] 11. Louden, D. N. MedGen: NCBI’s Portal to Information on MedicalConditions with a Genetic Component. Med. Ref. Serv. Q. 39, 183–191 (2020).

[0363] 12. Lee, D. D. & Seung, H. S. Learning the parts of objects by non-negative matrix factorization. Nature 401, 788–791 (1999).

[0364] 13. Röder, M., Both, A. & Hinneburg, A. Exploring the space of topiccoherence measures. in Proceedings of the Eighth ACM International Conferenceon Web Search and Data Mining - WSDM ’15 (ACM Press, 2015). doi:10.1145 / 2684822.2685324.

[0365] 14. McInnes, L., Healy, J., Saul, N. & Großberger, L. UMAP: UniformManifold Approximation and Projection. Journal of Open Source Software vol. 3861 Preprint at https: / / doi.org / 10.21105 / joss.00861 (2018).

[0366] 15. Chen, Z. et al. PhenoApt leverages clinical expertise toprioritize candidate genes via machine learning. Am. J. Hum. Genet. 109, 270–281 (2022).

[0367] 16. Zhao, M. et al. Phen2Gene: rapid phenotype-driven geneprioritization for rare diseases. NAR Genom Bioinform 2, lqaa032 (2020).

[0368] 17. Robinson, P. N. et al. Interpretable Clinical Genomics with aLikelihood Ratio Paradigm. Am. J. Hum. Genet. 107, 403–417 (2020).

[0369] 18. 100,000 Genomes Project Pilot Investigators et al. 100,000Genomes Pilot on Rare-Disease Diagnosis in Health Care - Preliminary Report.N. Engl. J. Med. 385, 1868–1880 (2021).

[0370] 19. Larkin, J., McDermott, J., Simon, D. P. & Simon, H. A. Expert andnovice performance in solving physics problems. Science 208, 1335–1342(1980).

[0371] 20. Coderre, S., Mandin, H., Harasym, P. H. & Fick, G. H. Diagnosticreasoning strategies and diagnostic success. Med. Educ. 37, 695–703 (2003).

[0372] 21. Shen, F. et al. HPO2Vec+: Leveraging heterogeneous knowledgeresources to enrich node embeddings for the Human Phenotype Ontology. J.Biomed. Inform. 96, 103246 (2019).

[0373] 22. Tiddi, I. & Schlobach, S. Knowledge graphs as tools forexplainable machine learning: A survey. Artif. Intell. 302, 103627 (2022).

[0374] 23. Zhang, Y. et al. High-throughput phenotyping with electronicmedical record data using a common semi-supervised approach (PheCAP). Nat.Protoc. 14, 3426–3444 (2019).

[0375] 24. Jacobsen, JOB et al. The GA4GH Phenopacket schema defines acomputable representation of clinical data. Nat. Biotechnol. 40, 817–820(2022).

[0376] 25. Yurkovich, JT, Tian, ​​Q., Price, ND & Hood, L. A systems approach to clinical oncology uses deep phenotyping to deliver personalized care. Nat. Rev. Clin. Oncol. 17, 183–194 (2020).

[0377] 26. Chute, CG Clinical Data Retrieval and Analysis. I've Seen aCase Like That Before. Annals of the New York Academy of Sciences vol. 670133–140 Preprint at https: / / doi.org / 10.1111 / j.1749-6632.1992.tb26083.x(1992).

[0378] Another aspect of the present disclosure relates to:

[0379] Item A. A device (1) for identifying a gene with the highest probability of being associated with a symptom observed for a subject, the device comprising:

[0380] - at least one input adapted to receive subject data (20) representative of at least one symptom observed for said subject;

[0381] - at least one processor configured to:

[0382] Compare each subject's data (20) representing a symptom of the subject with a phenotype represented in the genotype-phenotype matrix (21) The gene-phenotype matrix (21) represents an association (22) between at least one phenotype and at least one gene believed to cause the at least one phenotype associated with the gene-phenotype matrix (21), wherein each coefficient Indicates gene and phenotype The probability of association between

[0383] Calculating a subject-specific matrix comprising the phenotypes of the genotype matrix (21) associated with at least one symptom of the subject as columns;

[0384] Calculate the score vector, where each coefficient of the score vector is obtained from the coefficients of each row of the subject-specific matrix to represent the probability that each gene of the subject-specific matrix is ​​associated with the symptom observed for the subject;

[0385] - at least one output adapted to output at least one prediction (31) based on the score vector, the prediction (31) providing information about one or more genes that are more likely to be associated with the symptom observed for the subject.

[0386] Item B. The apparatus of Item A, wherein the at least one processor is configured to:

[0387] - Calculate the pathogenicity vector, where each coefficient of the pathogenicity vector represents the The frequency with which variants are reported as pathogenic; and

[0388] - Each coefficient of the score vector Subtract targeted genes Obtain the coefficients of the pathogenicity vector.

[0389] Item C. The apparatus of item A or item B, wherein the at least one processor is configured to:

[0390] - Calculate the correction vector to take into account the associated missing phenotype, where each coefficient of the correction vector is calculated as a subject-specific matrix with genes the sum of the coefficients of the associated row; and

[0391] - Each coefficient of the score vector Subtracting and targeting genes Correlation coefficient of the obtained correction vector.

[0392] Item D. The apparatus according to any one of Items A to C, wherein the at least one processor is further configured to calculate the gene-phenotype matrix (21) by:

[0393] -Define the matrix ,in, is the number of genes and is the number of phenotypes present in the list that includes associations between phenotypes and genes; if the list includes the gene and the phenotype The correlation between the gene and phenotype is calculated by Set to a predefined maximum non-zero value if the list does not include the gene and the phenotype The correlation between Set to zero;

[0394] - For genes not present in the above list Phenotype Each coefficient of the correlation , according to phenotype and genes The coefficient is calculated by the distance between the closest phenotypes The value of , where the phenotype exists in the list and genes The closest phenotypic association is The distance to the closest phenotype is calculated on a directed acyclic graph representing the standard classification of human abnormal phenotypes and their semantic relations.

[0395] Item E. The apparatus according to any one of Items A to D, wherein the at least one processor is further configured to calculate the gene-phenotype matrix (21) by:

[0396] - calculating at least two intermediate gene-phenotype matrices from respective lists comprising associations between at least one phenotype and at least one gene;

[0397] - The linear combination of the calculated intermediate gene-phenotype matrices is used as the gene-phenotype matrix (21).

[0398] Item F. The apparatus according to item D or item E, wherein, in order to calculate the gene-phenotype matrix (21), at least one processor is further configured to calculate each column of the gene-phenotype matrix (21). Apply normalized weights, where weights It is phenotype A function of the frequency of occurrence of the genes represented in the gene-phenotype matrix ( 21 ).

[0399] Item G. The apparatus of Item D or Item E, wherein the at least one processor is further configured to rank genes associated with symptoms observed in the subject based on a basis of the score vector.

[0400] Item H. A computer-implemented method for identifying genes with the highest probability of being associated with a symptom observed for a subject, the method comprising:

[0401] - receiving (41) subject data (20) representing at least one symptom observed for the subject;

[0402] - Compare each subject data (20) representing a symptom of the subject with a phenotype represented in the gene-phenotype matrix (21) association (42); the gene-phenotype matrix (21) represents the association between a symptom and at least one gene believed to cause the associated symptom, wherein each coefficient of the gene-phenotype matrix (21) Indicates gene and phenotype The probability of association between

[0403] - calculating (43) a subject-specific matrix (21) comprising phenotypes associated with at least one symptom of the subject as columns;

[0404] - Calculate (44) the score vector, where each coefficient of the score vector is obtained from the coefficients of each row of the subject-specific matrix to represent the probability that each gene represented in the subject-specific matrix is ​​associated with the symptom observed for the subject;

[0405] - outputting at least one prediction (31) based on the score vector, the prediction (31) providing information about one or more genes that are more likely to be associated with the symptom observed for the subject.

[0406] Item I. The method according to Item H, further comprising:

[0407] - Calculate the pathogenicity vector, where each coefficient of the pathogenicity vector represents the The frequency with which variants are reported as pathogenic; and

[0408] - Each coefficient of the score vector Subtract targeted genes Obtain the coefficients of the pathogenicity vector.

[0409] Item J. The method according to item H or item I, further comprising:

[0410] - Calculate the correction vector to take into account the associated missing phenotype, where each coefficient of the correction vector is calculated as the sum of the gene-phenotype matrix (21) with the gene the sum of the coefficients of the associated row; and

[0411] - Each coefficient of the score vector Subtracting and targeting genes Correlation coefficient of the obtained correction vector.

[0412] Item K. The method according to any one of Items H to J, further comprising calculating (52) the gene-phenotype matrix (21) by:

[0413] -Define the matrix ,in, is the number of genes and is the number of phenotypes present in the first list of associations between phenotypes and genes; if the first list includes the gene and the phenotype The correlation between the gene and phenotype is calculated by Set to a predefined maximum non-zero value if the list does not include the gene and the phenotype The correlation between Set to zero;

[0414] - for genes not present in the first list Phenotype Each coefficient of the correlation , according to phenotype and genes The distance between the closest phenotypes is calculated for each coefficient where the phenotype exists in the first list and genes The closest phenotype association is The distance to the closest phenotype is calculated on a directed acyclic graph representing the standard classification of human abnormal phenotypes and their semantic relations.

[0415] Item L. A computer program comprising software code adapted to perform at least one of the methods for identifying a gene with the highest probability of being associated with a symptom observed for a subject according to any one of Items H to K.

[0416] Item N. A computer-readable storage medium comprising instructions that, when executed by a computer, cause the computer to perform a method for identifying a gene with the highest probability of being associated with a symptom observed for a subject according to any one of Items H to K.

Claims

1. A computer-implemented method for decision support of a user to standardize phenotypes in genomic analysis of a subject, wherein: The method comprises: -a) receiving a symptom list comprising at least one symptom observed for the subject, wherein the symptom list comprises integers Symptoms - If integer is above a threshold, the method further comprising a step b1) receiving a first graph comprising nodes and weighted links, wherein: oEach node represents a symptom, and each symptom is associated with at least one gene; o Each link connecting two nodes represents the similarity between the two symptoms associated with the two connected nodes, and oThe weight of each link represents the degree of similarity between the two symptoms associated with the two connected nodes; The first graph is previously obtained by calculating symptom pair similarities using information about associations between at least one symptom and at least one gene believed to be responsible for the associated at least one symptom; - Otherwise, if the integer is lower than or equal to the threshold, the method further comprises b21), b22), b23) and b24), which comprise: -b21) receiving a second graph previously obtained by applying matrix decomposition to a gene-symptom matrix, the second graph comprising nodes and weighted links, wherein each node represents a symptom, each symptom is associated with at least one gene; each node is associated with at least one group of nodes of the second graph; and each node is associated with at least one specificity score, each specificity score representing the specificity of the symptom associated with the node relative to the symptoms of the nodes in each group associated with the node; wherein the gene-symptom matrix represents an association between at least one symptom and at least one gene believed to be responsible for the associated at least one symptom, and wherein each coefficient of the gene-symptom matrix Indicates gene With symptoms The probability of association between wherein each gene represented in the gene-symptom matrix is ​​associated with each group of nodes of the second graph via a gene-group weighted score; -b22) evaluating the specificity of the received symptom list using the specificity scores of the nodes associated with at least one symptom in the second graph, and providing a suggestion of at least one new symptom to the user if the specificity of the symptom list does not satisfy a predefined rule, each of the at least one suggested new symptoms being obtained as a symptom associated with a node in the first graph, the new symptom being connected to a node associated with at least one symptom of the received symptom list by a link; -b23) if the user adds at least one suggested new symptom to the received symptom list, evaluating the specificity of the updated symptom list using the specificity scores of the nodes in the second graph associated with the symptoms in the updated symptom list including at least one received symptom and at least one newly suggested symptom; -b24) whenever the specificity of the symptom list satisfies a predefined rule, using the symptom list and the first graph to obtain at least one gene with the highest probability of being associated with the symptom list; The method further comprises the following steps: -c) If, after providing a suggestion to the user for at least one new symptom, the specificity of the updated symptom list does not satisfy the predefined rules: generating a symptom binary vector representing the symptoms in the symptom list and applying matrix factorization to the binary vector to obtain a representation of the updated symptom list on the groups of the second graph, the representation comprising a patient score for each group of the second graph; Calculating the difference between each patient score and each gene-group weighted score for each group of the second graph; Normalizing the differences relative to a reference set of distances representing asymptomatic subjects; selecting at least one gene with the highest probability of being associated with the updated list of symptoms as the at least one gene with the highest normalized difference; -d) outputting at least one gene associated with the symptom list or the updated symptom list.

2. A device for decision support of a user to standardize phenotypes in genomic analysis of a subject, the device comprising: - at least one input configured to receive: o A symptom list comprising at least one symptom observed for the subject, wherein the symptom list comprises integers Symptoms o a first graph including nodes and weighted links, wherein each node represents a symptom, each symptom is associated with at least one gene; each link connecting two nodes represents a similarity between the two symptoms associated with the two connected nodes, and the weight of each link represents a degree of similarity between the two symptoms associated with the two connected nodes; The first graph is previously obtained by calculating symptom pair similarities using information about associations between at least one symptom and at least one gene believed to be responsible for the associated at least one symptom; o a second graph previously obtained by applying matrix decomposition to the gene-symptom matrix, the second graph comprising nodes and weighted links, wherein each node represents a symptom, each symptom is associated with at least one gene; each node is associated with at least one group of nodes of the second graph; and each node is associated with at least one specificity score, each specificity score representing the specificity of the symptom associated with the node relative to the symptoms of nodes in each group associated with the node; wherein the gene-symptom matrix represents an association between at least one symptom and at least one gene believed to be responsible for the associated at least one symptom, and wherein each coefficient of the gene-symptom matrix Indicates gene With symptoms and wherein each gene represented in the gene-symptom matrix is ​​associated with each group of nodes of the second graph by a gene-group weighted score; - at least one processor configured to: oIf the integer Less than or equal to threshold: evaluating the specificity of the symptom list using the specificity scores of the nodes associated with at least one symptom in the second graph, and providing a suggestion of at least one symptom to the user if the specificity of the symptom list does not satisfy a predefined rule, each of the at least one suggested symptom being obtained as a symptom associated with a node in the first graph connected by a link to a node associated with at least one received symptom; if the user adds the at least one suggested symptom to the symptom list, evaluating the specificity of the symptom list using specificity scores of nodes in the second graph associated with symptoms in the symptom list including the at least one received symptom and the at least one suggested symptom; Whenever the specificity of the symptom list satisfies a predefined rule, using the symptom list and the first graph to obtain at least one gene with the highest probability of being associated with the symptom list; o If after providing a suggestion to the user for at least one symptom, the specificity of the symptom list does not satisfy the predefined rules: generating a symptom binary vector representing the symptoms in the symptom list and applying matrix factorization to the binary vector to obtain a representation of the symptom list on the groups of the second graph, the representation comprising a patient score for each group of the second graph; Calculating the difference between each patient score and each gene-group weighted score for each group of the second graph; Normalizing the differences relative to a reference set of distances representing asymptomatic subjects; selecting at least one gene having the highest probability of being associated with the list of symptoms as the at least one gene having the highest normalized difference; - at least one output configured to provide said at least one gene associated with said list of symptoms.

3. The device according to claim 2, wherein: The matrix decomposition is a non-negative matrix decomposition.

4. The device according to claim 3, wherein: The non-negative matrix factorization applied to the gene-symptom matrix is ​​a non-negative double singular value decomposition (NNDSVD) initialization.

5. The device according to any one of claims 2 to 4, wherein: The information about the association between at least one symptom and at least one gene is a gene-symptom list representing the association between at least one symptom and at least one gene believed to cause the associated at least one symptom, or the information is a gene-symptom matrix.

6. The device according to any one of claims 2 to 5, wherein: The weight of each link in the first graph is determined using a node similarity algorithm.

7. The apparatus according to any one of claims 2 to 6, wherein: Using the symptom list and the first graph to obtain at least one gene with the highest probability of being associated with the symptom list comprises: - determining at least one additional symptom associated with at least one symptom of the symptom list by identifying, among the nodes of the first graph, at least one node connected to one of the at least one nodes corresponding to a symptom on the symptom list by a weighted link having a weight higher than a predefined threshold; - calculating a subject-specific matrix comprising the genotype matrix for the symptoms associated with at least one symptom of the subject included in the symptom list and the at least one additional symptom as columns; - Calculate a score vector, where each coefficient of the score vector is obtained based on the coefficients of each row of the subject-specific matrix to represent the probability that each gene of the subject-specific matrix is ​​associated with the symptom observed for the subject; - using the score vector to obtain information about at least one gene that is more likely to be associated with the symptom observed for the subject.

8. The apparatus according to any one of claims 2 to 7, wherein: The at least one processor is further configured to rank genes associated with symptoms observed in the subject based on a basis of the score vector.

9. A computer program product for decision support for users to standardize phenotypes in genomic analysis of subjects, comprising instructions that, when executed by a computer, cause the computer to execute the method for decision support for users to standardize phenotypes in genomic analysis of subjects according to claim 1. 10 . A computer-readable storage medium comprising instructions, which, when executed by a computer, causes the computer to perform the method for decision support for a user to standardize phenotypes in genomic analysis of a subject according to claim 1 .

Images