Methods and systems for learning novel relationships among various biological entities

Abstract

A computer-implemented method of learning novel relationships between various entities, in particular biological entities such as chemical substances, proteins and diseases, comprises establishing a knowledge graph (100), wherein each of the entities is represented as a node (110) and each relationship between the entities is represented as an edge (120) between respective nodes (110), and annotating the entities in the knowledge graph (100) with objects (130) of one or more data modalities. A neural network system (200) is trained with the knowledge graph (100), wherein the neural network system (200) processes the knowledge graph (100) and the objects (130) of the respective one of the data modalities in a unified manner by jointly learning embeddings of the nodes (110) from the knowledge graph (100) and embeddings of the objects (130) of the respective one of the data modalities. The learned embeddings are used to identify novel relationships between the entities.

Description

[0001] This invention relates to a computer-implemented method and a computer system for learning novel relationships between various entities, particularly biological entities such as chemical substances, proteins, and diseases.

[0002] Biological processes within the human body result from complex interactions between various molecules, such as DNA, RNA, proteins, metabolites, and enzymes. Disease phenotypes are caused by disruptions in these processes; similarly, effective treatment aims to restore these processes to their normal, healthy state. Therefore, it is essential to accurately and thoroughly characterize the relationships between the molecules, diseases, and other entities involved in these processes.

[0003] Traditional methods for characterizing these relationships require expensive and time-consuming biological wet laboratory experiments. The cost of these experiments results in highly incomplete datasets. Therefore, while this experimental approach remains the gold standard for determining their validity, it is important to design more cost-effective methods to characterize these relationships.

[0004] Mounting evidence demonstrates that artificial intelligence (AI) methods are also highly effective in identifying novel relationships. Of course, high-confidence relationships predicted by AI methods can still (and should) be validated using traditional wet laboratory methods.

[0005] To address this issue, network-based methods have been proposed in the prior art, where each entity is represented as a node in a graph, and each relationship is represented as an edge between the corresponding entities. Different relationship types are represented by different edge types. These network-based methods can be categorized into three types: neighborhood-based, diffusion-based, and representation-based. These methods rely on different topological properties of the network (such as motifs, communities, hubs, clusters, node centrality, shortest paths, etc.) to predict new edges. Furthermore, some methods also consider the biological characteristics of nodes.

[0006] For example, consider representation-based methods most similar to this invention, which use neural networks to learn embeddings or real-valued vectors such that entities that are close in the graph and have similar relationships also have similar embeddings. A machine learning model is then trained to predict novel relationships based on these embeddings.

[0007] However, the known solutions mentioned above have proven to be disadvantageous because all existing methods are limited by the network structure.

[0008] Therefore, the object of the present invention is to improve and further develop methods and systems of the type originally described for learning novel relationships between various entities in order to improve the view of the knowledge graph.

[0009] In an embodiment, the present invention provides a computer-implemented method for learning novel relationships between various entities, particularly biological entities such as chemical substances, proteins, and diseases, the method comprising:

[0010] Construct a knowledge graph where each entity is represented as a node, and each relationship between entities is represented as an edge between the corresponding nodes;

[0011] Annotate entities in the knowledge graph with objects of one or more data modalities;

[0012] The knowledge graph is used to train a neural network system, which processes the knowledge graph and the objects of the corresponding data modalities in a unified manner by jointly learning the embeddings of nodes from the knowledge graph and the embeddings of objects of the corresponding data modalities; and

[0013] Use the learned embeddings to identify novel relationships between these entities.

[0014] According to the present invention, it is first recognized that the above-mentioned objectives can be achieved through a method that learns consistent representations of multiple data modalities, particularly biological entities such as proteins and diseases, based on a knowledge graph (KG) and any additional data modalities such as structured annotations describing entities and free text. These representations can then be used to identify novel relationships, for example, proteins associated with diseases. Novel relationships can be used to prioritize protein targets for new drugs. Embodiments of the present invention explicitly incorporate the consistency of representations into the training process, compared to prior art methods. In particular, embodiments of the present invention consider the KG structure and documents about entities in a uniform manner. More specifically, the joint embedding and learning to predict true relationships is performed in a single step, i.e., the embedding and prediction model are learned jointly. This ensures that the embedding is optimized for prediction.

[0015] Embodiments of the present invention relate to a method and system for learning novel relationships between various biological entities such as proteins, drugs, and diseases using a network-based approach. In this network-based approach, each entity is represented as a node in a graph, and each relationship is represented as an edge between the corresponding entities, wherein different relationship types are represented by different edge types. In the context of this disclosure, such a network is referred to as a knowledge graph (KG). Learning novel relationships then equates to predicting “missing” edges based on the KG.

[0016] An obvious approach to combining the KG structure and the document (which exists in the literature) is to train separate ML models on the KG and the document, and then combine them in some post-analysis manner. In the context of neural network embeddings, this corresponds to learning an embedding from the KG structure and an embedding from the document for each entity; however, even with co-training and end-to-end training (as done in previous literature), the two embeddings can still be completely different for each entity. That is, the embeddings are not uniform, but rather entirely independent.

[0017] To address this problem, embodiments of the present invention relate to a link prediction system that implements a neural network learning strategy that combines KG structure and documents within a unified framework. The two embeddings are forced to be similar, unlike "obvious" methods. A key advantage over obvious methods is that this "pulls together" embeddings of entities that are geographically distant in the KG but have similar documents. Similarly, it "pulls apart" embeddings of entities that are closely related in the KG but have very different documents. Thus, a unified view of the KG and the documents is used to predict novel relationships. Therefore, these embodiments offer the advantage that the neural network system learns entity embeddings that combine the KG structure with additional information from the documents. In particular, the neural network system can learn document similarity based on the context of associated entities within the documents, regardless of syntactic or semantic overlap in the documents' vocabulary.

[0018] In an embodiment, the present invention provides a method for identifying novel candidate drug targets by predicting genes associated with a disease.

[0019] In another embodiment, the present invention provides a method for drug reuse in which existing drugs are used to treat new diseases by predicting the 'drug-treats-disease' edge. This is particularly relevant for immune checkpoint inhibitors (ICIs), which have proven highly effective for some diseases but have serious side effects for others. ICIs are often used in conjunction with personalized cancer vaccines based on neoantigens.

[0020] There are several ways to advantageously design and further develop the teachings of the invention. For this purpose, reference is made, on the one hand, to the dependent claims, and on the other hand, to the following explanation of preferred embodiments of the invention illustrated by way of example and in conjunction with the accompanying drawings. The general preferred embodiments and further developments of the teachings will be explained in conjunction with the explanation of preferred embodiments of the invention with the aid of the drawings. In the drawings...

[0021] Figure 1This is a conceptual overview of a framework for learning novel relationships between various biological entities according to embodiments of the present invention, and

[0022] Figure 2 This is a schematic overview of a neural network structure used in a method for learning novel relationships between various biological entities according to embodiments of the present invention.

[0023] In the following description of preferred embodiments of the present invention, the following notes will be used:

[0024]

[0025] Figure 1 A general conceptual framework for a system for learning novel relationships between various biological entities, according to embodiments of the present invention, is described. Figure 1 The illustrated embodiment of the framework includes various aspects and components, including a biomedical document mining component 150 (shown in Part 1), a biomedical knowledge graph KG 100 (shown in Part 2), an embedding space 160 (shown in Part 3), and a neural network system 200 (shown in Part 4). The specific embodiment is situated within the context of drug development. However, as those skilled in the art will understand, in other embodiments, the framework can be equally relevant to other applications.

[0026] Biomedical Knowledge Graph

[0027] According to an embodiment of the present invention, a biomedical knowledge graph 100 (denoted herein as G = (V, E)) may consist of a node set 110 (denoted as V) and an edge set 120 (denoted as E). Each node V in the biomedical knowledge graph 100 corresponds to an entity, such as a chemical substance, protein, or disease. Edges E represent known relationships between these entities.

[0028] In this embodiment, the construction of the knowledge graph G can be performed manually by considering various structured data sources. Specifically, according to the embodiment, entities and their relationships can be extracted from various sources, including:

[0029] - From DisGeNET: "disease--associated_with--protein" relationship

[0030] -From SIDER: , chemical--treats--disease (chemical substances-treatments-diseases) "relationship"

[0031] - From STITCH: "chemical--reacts_with--chemical (chemical substance-with chemical substance-reaction)"

[0032] - From STITCH: "chemical--interacts_with--protein" relationship

[0033] -From the STRING "protein--interacts_with--protein" relationship

[0034] According to the embodiments, the disclosed methods can be followed to create additional "disease--similar_to--disease" relationships.

[0035] Furthermore, since some data sources provide confidence values ​​for relationships, these confidence values ​​can be considered to filter out relationships with low reliability, since those relationships are already predictions.

[0036] In summary, based on the described embodiments, knowledge graph 100 considers three types of entities: diseases, genes (proteins), and chemical substances (drugs). Furthermore, the knowledge graph includes six relation (edge) types. According to the annotations used herein, e t It refers to edge type t.

[0037] As those skilled in the art will understand, other entity, data source, and relation extraction strategies besides those explicitly mentioned above can also be used to construct KG 100.

[0038] In the following text, facts encoded by graphs in G are represented as a set of triples of the form (h,r,t), where h∈V and t∈V are the head entity and tail entity, and r∈E is the relation type.

[0039] For example, Figure 1 Part 2 presents an excerpt of KG 100 according to an embodiment of the present invention. For example, this part illustrates the association of the disease "asthma" (D1) with "gene 108" (P2) and "gene 142" (P1). Therefore, written as a triple of the above form, the relationship present in KG 100 can be represented as (D1 is 'associated' with P2). From a practical perspective, identifying novel relationships is costly, and most of them remain unknown.

[0040] Mining biomedical documents

[0041] According to embodiments of the invention, entities in KG 100 (which can be considered objects of a first data modality) are annotated with objects 130 of one or more other data modalities. In the described embodiment, entities in KG 100 are annotated with objects 130 of another data modality (i.e., biomedical documents). The biomedical documents describe the relevant entities in KG 100 in more detail. For example, a protein may be annotated with its function or text from academic papers describing the protein. Hereinafter, this additional information is simply referred to as the document 140 of the corresponding entity.

[0042] According to the embodiments, document 140 can be retrieved manually or automatically by crawling various data sources, including unstructured sources such as academic papers from PubMed or structured data sources such as ontology such as gene ontology. Possible details described by document 140 may include location (e.g., at which tissue has specific protein activity), biochemical properties (e.g., the hydrophobicity or molecular weight of a drug), functional behavior (e.g., a description of the phenotype of an individual with a disease), or higher-level descriptions (e.g., the results of clinical trials involving specific drugs and diseases).

[0043] According to the embodiments, all documents 140 can be prepared based on a public vocabulary for further use. For example, word segmentation-based methods or frequency-based methods can be used. Furthermore, standard preprocessing steps such as tokenization, lemmatization, and stemming can be used. Figure 1 Part 1 schematically illustrates an embodiment of a mining biomedical document component 150 used to create document 140 from an original data source.

[0044] Furthermore, each document 140 is associated with at least one entity from KG 100. A document 140 may be associated with more than one entity, and each entity may be associated with more than one document 140. Numerous methods can be used to assign documents 140 to entities. For structured data sources, entity identifiers are typically provided; for unstructured sources, natural language processing methods such as named entity recognition can be used to extract entity identifiers from the text. This invention is not limited to the exact method used to create biomedical documents 140 from the original data source and to match biomedical documents with entities.

[0045] According to an embodiment, the output of the biomedical document mining component 150 can be a biomedical document set D, where each d∈D is a set of symbols from a vocabulary. Further, think Here, (V,D) is a pair. This is denoted as G. B =((V,E),(V,D)). GB It will be used as the input to the neural network component 200.

[0046] According to embodiments of the present invention, in order to create document 140, the following exemplary information can be extracted from various public sources:

[0047] - Gene Ontology: Protein Function and Location in the Cell (Structured Annotation)

[0048] - Human phenotypic ontology: Phenotypes associated with disease (structured annotation)

[0049] -SIDER, OFFSIDES: Chemical (Drug) Side Effects (Structured Annotation)

[0050] -MyGene.info: Protein description (free text)

[0051] -DrugBank: Drug Descriptions (Free Text)

[0052] -DisGeNET: Disease Mechanisms and Associated Mutations (Free Text)

[0053] Figure 1 Part 2 presents document 140 and protein P. i An example of the relationship between entities. Typically, all entities can have associated documents 140, and documents 140 can be associated with multiple entities.

[0054] All current methods known in the prior art are limited by the network structure. In particular, these current methods do not consider the structure of the knowledge graph 100 and the documents 140 about entities in a uniform manner. Embodiments of the present invention utilize a machine learning-based approach to address this shortcoming, in which the knowledge graph 100 and the documents 140 are combined in a uniform manner, as will be described in detail below.

[0055] Neural Network System

[0056] Figure 1 The embodiment of the neural network system 200 schematically shown in Part 4 is... Figure 2 A more detailed description follows. The neural network system 200 maintains a machine learning model responsible for using the biomedical knowledge graph G provided above. B And the relevant biomedical document set D is used to predict novel relationships. Specifically, Figure 2 The neural network structure described in detail below is illustrated graphically.

[0057] train

[0058] For illustration, the following description exemplarily refers to "protein" and "disease" as possible head and tail entities, respectively, where the correspondence is "associated with" (e.g., protein and disease). d_aw_p It should be noted that the remaining combinations of head and tail entities with corresponding relationships are trained in a corresponding manner.

[0059] To train a machine learning model, a specific disease (e.g., "cardiomyopathy") and a specific relationship r are used. k (e.g., "associated_with") remain unchanged as the head h and relation type r, and proteins are grouped as and in It is believed that cardiomyopathy exists in KG, associated with... Triples in the form of ) are considered to be... and

[0060] Similarly, and Described and The corresponding biomedical documents for the elements in the document. In other words, Each protein in the triplet is the correct tail, and Each protein in the sequence has an incorrect tail.

[0061] like Figure 2 As shown, conceptually, the machine learning model is trained using a single sample each time. That is, selecting... and corresponding documents And as a first step, the sample is embedded, where the embedding of node V and the embedding of document D occur independently, as shown at 202 for node V and at 204 for document D.

[0062] According to the embodiment, as shown at 206 for node V and at 208 for document D, the embedding can be performed by the following embedding function:

[0063]

[0064] Where U∈{V,D}, It is the embedding of element i, x j It is the corresponding element (that is, v) j or d j ), i = |x j |。 d j The cardinality is defined by the word count in the relevant documentation. j The cardinality is always 1; the cardinality is the index of the entity in the graph.

[0065] Secondly, as shown at 210 for node V and at 212 for document D: negative sampling can be applied to randomly select and D p - ,in, and In other words, these sets contain incorrect tails for the selected heads and relations. Negative sampling can be performed according to the method described in the following literature: Garcia-Duran, Alberto, and Mathias Niepert: “KBlrn: End-to-end learning of knowledge base representations with latent, relational, and numerical features”, Proceedings of the 34 th The paper, *Conference on Uncertainty in Artificial Intelligence* (2018), is incorporated herein by reference.

[0066] Similar to the embedding of positive elements, the set of negative elements can be embedded as follows, as shown at 214 for node V and at 216 for document D:

[0067]

[0068] in, It is the negative sample set (that is, V) s - or V s - and The cardinality of both is equal to the number of negative samples. The number of negative samples is a hyperparameter of this method.

[0069] The embedding of relations (as shown at 218) is defined as follows:

[0070]

[0071] Where k is an index for a specific relation type. (Based on the exemplary embodiments described herein, there are six relation types.)

[0072] More specifically, Emb V It is a neural network that provides the embeddings of each entity in the KG, while Emb... RIt provides a neural network for embedding each type of relation. Additionally, Emb... D It is a neural network that gives the embedding of each symbol in the vocabulary, and the embedding of a document is simply the average embedding of all its symbols. It is a simple function that gives the average embedding of all entities passed to it.

[0073] Learning Unified Embedding

[0074] According to an embodiment of the present invention, the neural network system 200 is configured to compute positive samples x. j Embedding and negative samples Euclidean distance between the centers:

[0075]

[0076] As shown at 220 for node V and at 222 for document D.

[0077] Furthermore, as shown at position 224, distance V With distance D The distance between them is calculated as L 偏移 :

[0078]

[0079] In other words, distance V (v j V s - The expression gives the distance between the embeddings of positive samples from KG and the centers of negative samples, while... The distance between the embedding of the positive sample from the document and the center of the negative sample is given.

[0080] It is worth noting that these distances can vary considerably in standard methods; in this case, L 偏移 It will be very large. In contrast, according to an embodiment of the invention, the neural network system 200 is designed to make L... 偏移 Minimize. Therefore, this will make v j and d j With the corresponding negative embedding sample V s - and The embedding distance between them is minimized. As a result, when learning document embeddings, the model explicitly considers similarities in the KG and vice versa. That is, the model explicitly unifies the embeddings of the KG structure with the embeddings of associated documents, which distinguishes the solution according to the invention from other known prior art methods.

[0081] It is worth noting that, according to the above implementation method, the model adjusts the learned embeddings of entities to find a balance between evidence in the KG and the document. Figure 1 Part 3 illustrates examples of how this can affect the embedding space. The left side of the figure shows an embedding learned using standard methods: similar documents indicate that gene 1017 is moderately close to other asthma-asthmatic proteins in the embedding space. However, as from... Figure 1 As seen in Part 2, gene 1017 is not closely related to other genes in KG, so it is unclear whether it should be associated with asthma. However, on the right side of the attached figure in Part 3, we can see the use of L... 偏移 Impact: Due to the unified learning as explained above, document embeddings provide information for KG embeddings. Therefore, it is now clear that gene 1017 should also be associated with asthma.

[0082] Learning accurate relationship prediction

[0083] According to an embodiment of the present invention, except L 偏移 In addition, the neural network component 200 can also track the prediction accuracy of the model; that is, the neural network component measures whether the learned weights of the embedding function are reliable for accurately predicting relationships. Specifically, the neural network component 200 can use embedding layers (Emb... V Emb D and Emb R The learned weights are used to predict relationships known in KG 100.

[0084] More specifically, as mentioned above, embodiments of the present invention aim to predict the true tail of a triple when the head and relation type are fixed. That is, assuming, for example, a specific disease as a given tail entity h (e.g., "cardiomyopathy" according to the example given above) and a specific relation type r. k ("associated_with") aims to predict the correct tail entity t of the triple (h,r,t). In the following text, the node associated with the head and the document are referred to as v, respectively. 头部 and d 头部 The prediction result is denoted as L. 预测 The function is calculated.

[0085] With L 偏移 In comparison, L 预测 No, not (from V) s - and Instead of averaging the embeddings of the negative samples, all negative samples are coupled with the correct tail (v). j and d jThe embeddings of fixed headers and fixed relation types can be compared to measure predictive performance. Specifically, according to an embodiment, embeddings of fixed headers and fixed relation types can be combined (element-wise multiplication, "·") to group nodes (as shown at 226) and documents (as shown at 228) in the embedding space by relation type:

[0086] Group U (x 头部 ,r k ) = Emb U (x 头部 Emb R (r k )

[0087] Finally, according to the embodiment, the score of each candidate tail (that is, j and each negative sample) is calculated as its embedding with the Group. U (x 头部 ,r k The dot product of ) is shown below (as shown at 230 for the node and at 232 for the document):

[0088] Predict U (x 头部 ,r k ,x j ) = Group U (x 头部 ,r k Emb U (x j )

[0089] Each score is a measure of the probability that the corresponding entity correctly completes the triple (h, r, ?). In the best case, the score for a negative sample should be 0, while the score for the true tail should be 1.

[0090] like Figure 2 As shown, the above steps are performed for both modalities (i.e., nodes and documents) respectively, and the resulting vectors are combined (elementally added) to form a score, as shown below (at position 234):

[0091] Predict(header, r) k ,j)=Predict V (v 头部 ,r k ,v j )+Predict D (d 头部 ,r k ,d j )

[0092] This score is calculated for all negative samples and the true tail. Because the exemplary embodiment discussed in this paper considers two modalities, the combined score for the true tail should be 2, while the combined score for the negative samples should still be 0. More generally, considering n different modalities, Predict(head, r) k The above equation for (j) will consist of n addends, and the combined score of the true tail will be n. 偏移 The calculation is an exponent of the number of modes included.

[0093] exist Figure 2 In the exemplary embodiment with two modes, the obtained scores are considered to calculate the binary cross-entropy loss (as shown at 236):

[0094]

[0095] in, It is the i-th negative sample.

[0096] Finally, according to embodiments of the present invention, these two loss values ​​can be combined to measure the performance of the corresponding machine learning model, for example as follows:

[0097] L=∝L 预测 +(1-∝)L 偏移 ,

[0098] Here, ∝∈[0,1] is a hyperparameter that controls the trade-off between accuracy and uniform embedding. Then, the standard backpropagation algorithm can be used to update Emb. V Emb R and Emb D The parameters are adjusted to make the predictions more accurate while still adhering to the consistency between KG embeddings and document embeddings; that is, the parameters are updated to minimize L.

[0099] For example, refer to Figure 1 Part 2. By considering graph structures and biomedical documentation and unifying their embeddings, the above mechanism recognizes that "gene 1017" (P3) is also associated with disease D1 (i.e., asthma). As a result, this invention identifies a novel relationship (e d_aw_p This novel relationship can be investigated and considered in more detail when developing novel drugs. As those skilled in the art will understand, this is merely an example of a particular type of edge, and similar situations exist for the remaining relationships that will also be disclosed in this invention.

[0100] Ranking Novel Relationships

[0101] According to an embodiment of the present invention, after training the model, the embedding function Emb... V Emb Rand Emb D The parameter is set to produce a high Predict(·) value for true triples and a low value for incorrect triples.

[0102] In other words, the trained model can answer queries of the following forms: (h,r,?), (h,?,t), and (?,r,t). Specifically, given two of these elements, Predict(·) is computed to obtain all possible values ​​for the missing element in the triple. The corresponding score is used to rank the triple from most likely (highest value) completion to least likely (lowest value) completion. For cases where correct query completion is known, this ranking can be evaluated using standard information retrieval metrics such as the area under the receiver operation feature curve, precision@k, and the area under the precision-recall curve.

[0103] As mentioned above, Figure 2 The pipeline outlined herein is not limited to embedding common learning nodes and document embedding. In fact, according to a further embodiment of the invention, the system can be easily enhanced by adding additional objects of other data modalities (such as images). In this case, the loss function L will be calculated for each pair. 偏移 Therefore, considering nodes, documents, and images will result in three pairs: (node, document), (node, image), and (document, image). Thus, L will be computed separately for each pair. 偏移 As described above and as per the description for that pair (node, document) Figure 2 As shown at position 224. However, it should be noted that this method still ensures that all modes are learned collectively. In summary, the total loss can be calculated as follows:

[0104] L = xL 预测 +y1L 偏移 …+y n L 偏移

[0105] in, m represents the number of data modes, and

[0106] In this embodiment, the present invention relates to a method for identifying disease-gene associations, which can be applied, for example, in conjunction with automated CRISPR-Cas9 knockdown experiments (as described, for example, in Zhou, Y., Zhu, S., Cai, C. et al. High-throughput screening of a CRISPR / Cas9 library for functional genomics in human cells. Nature 509, 487-491 (2014)) to understand the effect of a specific gene on a disease. In short, the CRISPR-Cas9 knockdown experiment uses a guide RNA that targets a specific gene; after the application of a CRISPR-Cas9 compound with the specific guide RNA, the target gene no longer functions. (The target gene is “knocked down.”) The goal of this system is to identify the genes most relevant to the disease by prioritizing those used in the knockdown experiment. (Presumably, these genes are then good candidates for drug discovery, but we do not explore this idea further in this disclosure.)

[0107] In this context, the nodes and edges in the knowledge graph correspond to genes, chemicals, and diseases, as already described and as... Figure 1 As shown in Part 2. Further, a set of guide RNAs for a specific gene is known to be physically available in a storage container. Such sequence libraries are publicly available (for reference, see, for example, Doench, J., Fusi, N., Sullender, M. et al., Optimized sgRNA design to maximize activity and minimize off-target effects of CRISPR-Cas9. Nat Biotechnol 34, 184-191 (2016). A robot is capable of retrieving the specific guide RNA from the storage device and placing it in a designated set of wells in a microtiter plate. For example, Agilent's BenchBot robot is capable of doing this.

[0108] First, a disease of interest is selected. Importantly, a cell line similar to that disease must be available; for example, KG-1 and HL-60 are leukemia cell lines (see, for reference, Koeffler HP, Golde DW. Human myeloid leukemia cell lines: a review. Blood, Sep 1980; 56(3):344-50). Second, the robot dispenses Cas9 into all the wells of the plate. Then, the method for learning novel relationships between various biological entities according to embodiments of the invention can be used to identify genes most likely associated with the disease. That is, the method predicts the “disease-associated gene” relationship for the selected disease and takes the top-ranked genes. The number of genes selected can be determined by the number of wells on the plate. Each gene is assigned to one or more wells on the plate. The robot automatically selects the corresponding guide RNA from the storage device and places them in the known wells of the microtiter plate for co-incubation with Cas9. After sufficient time, cells from the cell line are combined into the system using appropriate techniques such as electroporation.

[0109] Finally, another assay can be used to assess the effect of each gene knockout in the cell line. For example, Konstantinos Tzelepis et al. A CRISPR Dropout Screen Identifies Genetic Vulnerabilities and Therapeutic Targets in Acute Myeloid Leukemia. Cell Reports. Vol. 17, No. 4, 18 October 2016, pp. 1193-1205, used RNA sequencing to identify genetic pathways that function differently in the knockout cells compared to the unaltered cell line.

[0110] In an alternative embodiment, the present invention relates to a method for identifying disease-gene associations, which can be applied, for example, in conjunction with a personalized medicine system for developing personalized medicines. The goal of such a system is to select the best chemical substance to treat a disease for a specific patient. In this case, the nodes in the knowledge graph (KG) again correspond to genes, chemical substances, and diseases, as... Figure 1 As shown in Part 2. Furthermore, a subset of the chemicals included in the KG can be physically obtained in the storage container. The robot is able to extract specific chemicals from the storage device and place them in a designated set of wells in a microtiter plate. For example, Agilent's BenchBot robot is capable of doing this.

[0111] First, the patient is diagnosed with a specific disease. Second, a biopsy is performed on the patient; depending on the type of disease, the biopsy can be liquid (e.g., for leukemia) or solid (e.g., for ovarian cancer). Relevant cells are extracted from the biopsy using standard methods. For example, a standard protocol (see, for reference, Panda, SK and Ravindran, B. (2013). Isolation of Human PBMCs. Bio-protocol 3(3):e323) can be used to extract immune cells from blood. A robot dispenses the cells into wells on one or more microtiter plates. Then, a method for learning novel relationships between various biological entities according to embodiments of the invention can be used to identify the most likely chemical substances to treat the disease. That is, the method predicts a “chemical substance treats disease” relationship for the patient’s disease and selects the top-ranked chemical substances. The number of selected chemical substances can be determined by the number of wells on the plate. Each chemical substance is assigned to one or more wells on the plate. The robot automatically selects those chemical substances from a storage device and places them into the known wells of the microtiter plate.

[0112] Another assay can be used to determine the effect of each chemical on the patient's cells. For example, a high-throughput ELISpot assay, such as those available from R&D Systems, can be used to determine which chemicals elicit a response from immune system cells. The number of spots indicates how effective the chemical is for the patient. Finally, the most effective chemical for a specific patient can be selected for treatment.

[0113] As those skilled in the art will understand, the ability to begin with the methods described herein depends on having sufficient information to construct the KG and documentation. In some domains, this may be problematic. However, in the biomedical field, and specifically in the aforementioned biomedical application scenarios, ample data is available.

[0114] Many modifications and other embodiments of the invention set forth herein will occur to those skilled in the art upon benefiting from the teachings presented in the foregoing description and associated drawings. Therefore, it should be understood that the invention is not intended to be limited to the specific embodiments disclosed, and that modifications and other embodiments are intended to be included within the scope of the appended claims. Although specific terminology is used herein, it is used only in a general and descriptive sense and not for purposes of limitation.

Claims

1. A computer-implemented method for learning novel relationships between various entities, said various entities including biological entities, said biological entities including chemical substances, proteins, and diseases, the method comprising: Construct a knowledge graph (100), where each entity among these entities is represented as a node (110), and each relationship between these entities is represented as an edge (120) between the corresponding nodes (110). The entities in the knowledge graph (100) are annotated with objects (130) of one or more data modalities; The knowledge graph (100) is used to train a neural network system (200), wherein the neural network system (200) processes the knowledge graph (100) and the objects (130) of the corresponding data modal in a uniform manner by jointly learning the embeddings of nodes (110) from the knowledge graph (100) and the embeddings of objects (130) of the corresponding data modal in these data modal. The neural network system (200) is configured to minimize the loss function of a particular positive sample, which calculates the Euclidean distance between the following two items: the distance between the embedding of the node (110) of the positive sample from the knowledge graph (100) and the center of the corresponding negative sample and the distance between the embedding of the positive sample of the particular object (130) from these data modal and the center of the negative sample. Using learned embeddings to identify novel relationships between these entities, including: selecting a specific disease; using the neural network system (200) to predict the selected disease relationship in the form of 'genes associated with the disease' or 'chemical treatment of the disease'; The predicted genes or chemicals are ranked according to the probability that a corresponding one of the predicted genes or chemicals is associated with the selected disease; and Select a predefined number of top-ranked genes as candidates for knockdown experiments, or select a predefined number of top-ranked chemical substances as candidates for personalized medicine development.

2. The method according to claim 1, wherein, The one or more data modalities include biomedical documents (140) in the form of details describing these entities and / or objects (130) in the form of images.

3. The method according to claim 1 or 2, wherein, Each time, a single sample is used to train the machine learning model maintained by the neural network system (200). The sample is provided as a triple consisting of a head entity, a tail entity, and the correspondence type between the head entity and the tail entity. For specific head entities and specific relation types, these tail entities are grouped into positive subsets and negative subsets in such a way that for each tail entity in the positive subset, the corresponding triple is considered to exist in the knowledge graph (100), while for each tail entity in the negative subset, the corresponding triple is considered not to exist in the knowledge graph (100).

4. The method according to claim 1, wherein, The loss function is calculated separately for each pair of data modes used.

5. The method according to any one of claims 1 to 4, wherein, The neural network system (200) measures the prediction accuracy of a corresponding machine learning model in a trained machine learning model based on these embedded learned weights.

6. The method according to any one of claims 1 to 5, further comprising introducing a hyperparameter for controlling the trade-off between prediction accuracy and uniform embedding.

7. A computer system for learning novel relationships between various entities, including biological entities, which include chemical substances, proteins, and diseases, the computer system comprising a memory and one or more processors, the one or more processors being configured individually or in combination to provide execution of methods comprising: Construct a knowledge graph (100), where, Each of these entities is represented as a node (110), and each relationship between these entities is represented as an edge (120) between the corresponding nodes (110). The entities in the knowledge graph (100) are annotated with objects (130) of one or more data modalities; The knowledge graph (100) is used to train a neural network system (200), wherein the neural network system (200) processes the knowledge graph (100) and the objects (130) of the corresponding data modal in a uniform manner by jointly learning the embeddings of nodes (110) from the knowledge graph (100) and the embeddings of objects (130) of the corresponding data modal in these data modalities, wherein the neural network system (200) is configured to minimize a loss function for a specific positive sample, the loss function calculating the Euclidean distance between the following two items: the distance between the embedding of the node (110) of the positive sample from the knowledge graph (100) and the center of the corresponding negative sample and the distance between the embedding of the specific object (130) of the positive sample from these data modalities and the center of the negative sample; and Using learned embeddings to identify novel relationships between these entities, including: selecting a specific disease; using the neural network system (200) to predict the selected disease relationship in the form of 'genes associated with the disease' or 'chemical treatment of the disease'; The predicted genes or chemicals are ranked according to the probability that a corresponding one of the predicted genes or chemicals is associated with the selected disease; and Select a predefined number of top-ranked genes as candidates for knockdown experiments, or select a predefined number of top-ranked chemical substances as candidates for personalized medicine development.

8. The computer system according to claim 7, wherein, The neural network system (200) is configured to minimize a loss function for a specific positive sample, which calculates the Euclidean distance between the following two items: the distance between the embedding of the positive sample from the node (110) of the knowledge graph (100) and the center of the corresponding negative sample, and the distance between the embedding of the positive sample from the specific object (130) in these data modalities and the center of these negative samples.

9. The computer system according to claim 8, wherein, The loss function is calculated separately for each pair of data modes used.

10. The computer system according to any one of claims 7 to 9, wherein, The neural network system (200) measures the prediction accuracy of a corresponding machine learning model in a trained machine learning model based on these embedded learned weights.

11. The computer system according to any one of claims 7 to 10, further comprising a biomedical document mining component (150), the biomedical document mining component being configured to prepare biomedical documents based on a public vocabulary and generate a biomedical document set, wherein, Each element in this set is a set of symbols drawn from the vocabulary.

12. A non-transitory computer-readable medium having instructions thereon that, when executed by one or more processors, individually or in combination and using memory, provide for the execution of a method comprising: Construct a knowledge graph (100), where, Each of these entities is represented as a node (110), and each relationship between these entities is represented as an edge (120) between the corresponding nodes (110). The entities in the knowledge graph (100) are annotated with objects (130) of one or more data modalities; The knowledge graph (100) is used to train a neural network system (200), wherein the neural network system (200) processes the knowledge graph (100) and the objects (130) of the corresponding data modal in a uniform manner by jointly learning the embeddings of nodes (110) from the knowledge graph (100) and the embeddings of objects (130) of the corresponding data modal in these data modalities, wherein the neural network system (200) is configured to minimize a loss function for a specific positive sample, the loss function calculating the Euclidean distance between the following two items: the distance between the embedding of the node (110) of the positive sample from the knowledge graph (100) and the center of the corresponding negative sample and the distance between the embedding of the specific object (130) of the positive sample from these data modalities and the center of the negative sample; and Using learned embeddings to identify novel relationships between these entities, including: selecting a specific disease; using the neural network system (200) to predict the selected disease relationship in the form of 'genes associated with the disease' or 'chemical treatment of the disease'; The predicted genes or chemicals are ranked according to the probability that a corresponding one of the predicted genes or chemicals is associated with the selected disease; and Select a predefined number of top-ranked genes as candidates for knockdown experiments, or select a predefined number of top-ranked chemical substances as candidates for personalized medicine development.

Images