Method for generating negative sample set of predicting intermolecular interaction, method for predicting intermolecular interaction, and method for training model

Abstract

A method of generating a negative sample set for predicting inter-macromolecular interactions is provided. The method includes receiving a positive sample set comprising pairs of a first type of macromolecule and a second type of macromolecule having an inter-macromolecular interaction; generating a first similarity graph of the first type of macromolecule; generating a second similarity graph of the second type of macromolecule; generating a vectorized representation of nodes in the first similarity graph and a vectorized representation of nodes in the second similarity graph; and generating the negative sample set using the vectorized representation of nodes in the first similarity graph and the vectorized representation of nodes in the second similarity graph.

Description

TECHNICAL FIELD

[0001] The present application relates to machine learning techniques, and more particularly, to a method for generating a negative sample set for predicting inter-macromolecular interactions, a method for predicting inter-macromolecular interactions, a training method for generating a model for generating a negative sample set for predicting inter-macromolecular interactions, and a neural network model for predicting inter-macromolecular interactions. BACKGROUND

[0002] Protein-RNA interactions play an important role in various processes of cells, including post-transcriptional regulation of gene expression, protein translation, RNA post-transcriptional modification, and cellular regulation. In recent years, much effort has been made on the prediction of protein-RNA interactions. Generally, the prediction methods are based on the structure, chemical properties, and biological functions of RNA molecules and protein molecules. SUMMARY

[0003] In one aspect, the present disclosure provides a method for generating a negative sample set for predicting inter-macromolecular interactions, comprising: receiving a positive sample set comprising pairs of first type macromolecules and second type macromolecules having inter-macromolecular interactions; generating a first similarity graph of the first type macromolecules; generating a second similarity graph of the second type macromolecules; generating a vectorized representation of nodes in the first similarity graph and a vectorized representation of nodes in the second similarity graph; and generating the negative sample set using the vectorized representation of nodes in the first similarity graph and the vectorized representation of nodes in the second similarity graph.

[0004] Optionally, the first similarity graph or the second similarity graph comprises nodes and edges connecting adjacent nodes, wherein each node represents a respective macromolecule, each edge represents a respective distance between a respective pair of macromolecules, and a respective weight of each edge represents a respective similarity between the respective pair of macromolecules.

[0005] Optionally, the method further comprises generating a plurality of intermediate sets; wherein the positive sample set is represented by {(m1 i , m2 i ), i = 1,..., K}, where m1 i represents the i-th first type macromolecule, and m2 i represents the i-th second type macromolecule; wherein generating a respective intermediate set in the plurality of intermediate sets comprises: sorting m1 i (j = 1,..., K, and j ≠ i) based on the similarity between m1 j and m1 j to obtain a subset of m1 j (j = 1,..., K, and j ≠ i); determining m2 ia probability of interaction between m2

[0006] Optionally, the method further comprises calculating m1 i a probability of interaction between m1 j and each sample in the subset:

[0007]

[0008] wherein dr j represents a vectorized representation of nodes in the subset; and i represents a vectorized representation of m1 i .

[0009] Optionally, m2 i is determined by:

[0010]

[0011] wherein dr j represents a vectorized representation of nodes in the subset; and i represents a vectorized representation of m2 i ; P(1|dr j , dp i ) represents a probability of interaction between m2 i and each sample in the subset; and represents a concatenation between elements; represents a product of two vectors; and

[0012] Optionally, the method further comprises placing (m1 j , 1-P(1|dr i , dpi)) into the respective intermediate set when P(1|dr j , dp j ) is less than a threshold.

[0013] Optionally, based on probabilities {p k , k = 1,.., |T|}, L negative samples are sampled from the respective intermediate set;

[0014] wherein

[0015] |T| represents a number of elements in the respective intermediate set, and (m1 j , 1-P(1|dr j , dp i )) represents a k-th element in the respective intermediate set.

[0016] Optionally, generating the negative sample set comprises: sampling L negative samples from a respective one of the plurality of intermediate sets; wherein the negative sample set comprises negative samples sampled from the plurality of intermediate sets; and L is an integer greater than or equal to 1.

[0017] Optionally, the first type of macromolecule comprises RNA molecules and the second type of macromolecule comprises protein molecules.

[0018] In another aspect, the present disclosure provides a method of predicting an inter-macromolecular interaction using the positive sample set and the negative sample set generated by the method for generating a negative sample set described above.

[0019] In another aspect, the present disclosure provides a method of training a model for generating a negative sample set for predicting an inter-macromolecular interaction, comprising: receiving a positive sample set comprising pairs of a first type of macromolecule and a second type of macromolecule having an inter-macromolecular interaction; generating a first similarity graph of the first type of macromolecule; generating a second similarity graph of the second type of macromolecule; generating a vectorized representation of nodes in the first similarity graph and a vectorized representation of nodes in the second similarity graph; determining a probability of an interaction between a first respective vectorized representation of a node in the first similarity graph and a second respective vectorized representation of a node in the second similarity graph; and training the model based at least in part on the probability of the interaction.

[0020] Optionally, the probability of the interaction is determined by the following formula:

[0021]

[0022] wherein dmi i represents the vectorized representation of a node in the first similarity graph; dm2 j represents the vectorized representation of a node in the second similarity graph; p(1|dmi i , dm2 i represents the probability of an interaction between a first respective vectorized representation of a node in the first similarity graph and a second respective vectorized representation of a node in the second similarity graph; [.] represents concatenation between elements; represents the product of two vectors; 0 represents a tunable parameter.

[0023] Optionally, the first similarity graph or the second similarity graph comprises nodes and edges connecting adjacent nodes, wherein each node represents a respective macromolecule, each edge represents a respective distance between a respective pair of macromolecules, and a respective weight of each edge represents a respective similarity between a respective pair of macromolecules.

[0024] Optionally, a respective similarity between a respective pair of first-type macromolecules is represented as:

[0025] sim1(m 1-1 , m 1-2 ) = 1 - d1(m 1-1 , m 1-2 ); wherein (m 1-1 , m 1-2 ) represents the respective pair of first-type macromolecules, sim1 represents the respective similarity between the respective pair of first-type macromolecules, and d1 represents a distance between the respective pair of first-type macromolecules.

[0026] Optionally, d1 is expressed as:

[0027]

[0028] wherein lev(m 1-1 , m 1-2 ) represents an edit distance between the respective pair of first-type macromolecules, len(m 1-1 ) represents a length of a first first-type macromolecule in the respective pair, and len(m 1-2 ) represents a length of a second first-type macromolecule in the respective pair.

[0029] Optionally, a respective similarity between a respective pair of second-type macromolecules is represented as:

[0030] sim2(m 2-1 , m 2-2 ) = 1 - d2(m 2-1 , m 2-2 );

[0031] wherein (m 2-1 , m 2-2 ) represents the respective pair of second-type macromolecules, sim2 represents the respective similarity between the respective pair of second-type macromolecules, and d2 represents a distance between the respective pair of second-type macromolecules.

[0032] Optionally, d2 is expressed as:

[0033]

[0034] wherein lev(m 2-1 , m 2-2 ) represents an edit distance between the respective pair of second-type macromolecules, len(m 2-1 ) represents a length of a first second-type macromolecule in the respective pair, and len(m 2-2 ) represents a length of a second second-type macromolecule in the respective pair.

[0035] Optionally, the first similarity graph comprises N1nodes {e i ,i = 1,..., N1}; and M1edges {r j ,j = 1,..., M};

[0036] The respective vectorized representation of a respective node in the first similarity graph is expressed as:

[0037]

[0038] where e i represents the respective node in the first similarity graph; h t1 (e i ) represents the respective vectorized representation of the respective node e i before the t1th iteration; h t1+1 (e i ) represents the updated respective vectorized representation of the respective node e i after the t1th iteration; σ represents a leaky rectified linear unit activation function; N(e i ) represents a set of nodes adjacent to the respective node e i ; W p , W ph represent parameters of a graph neural network used to generate the vectorized representation.

[0039] Optionally, the second similarity graph comprises N2nodes {e’ i ,i = 1,..., N2}; and M2edges {r’ j ,j = 1,..., M2};

[0040] The respective vectorized representation of a respective node in the second similarity graph is expressed as:

[0041]

[0042]

[0043] where e’ i represents the respective node in the second similarity graph; h t2 (e′ i ) represents the respective vectorized representation of the respective node e’ i before the t2th iteration; h t2+1 (e′ i ) represents the updated respective vectorized representation of the respective node e’ i after the t2th iteration; σ represents a leaky rectified linear unit activation function; N(e’ i ) represents a set of nodes adjacent to the respective node e’ i ;t2 (e′ i ), h t2 (e′ k ) represents the inner product of h t2 (e′ i ) and h t2 (e′ k ); represent parameters of a graph neural network used to generate the vectorized representations; represent attention weights representing the link strength between nodes e’ i and e’ k .

[0044] Optionally, the positive sample set is represented by {(m1 i , m2 i ), i = 1,..., K}, where m1 i represents the i-th first type macromolecule, and m2 i represents the i-th second type macromolecule; wherein training the model comprises minimizing a loss function:

[0045]

[0046] wherein dm1 i represents the vectorized representation of a node in the first similar graph; dm2 j represents the vectorized representation of a node in the second similar graph; p(1|dm1 i , dm2 i ) represents the probability of the interaction between the first respective vectorized representation of a node in the first similar graph and the second respective vectorized representation of a node in the second similar graph.

[0047] Optionally, the first type macromolecule comprises an RNA molecule, and the second type macromolecule comprises a protein molecule.

[0048] In another aspect, the present disclosure provides a neural network model for predicting inter-macromolecular interactions, which is trained by the method of training the model described herein. BRIEF DESCRIPTION OF DRAWINGS

[0049] According to various disclosed embodiments, the following drawings are merely examples for illustrative purposes only, and are not intended to limit the scope of the present disclosure.

[0050] Figure 1 A process of generating negative samples for predicting inter-macromolecular interactions is shown.

[0051] Figure 2A process of generating negative samples for predicting inter-macromolecular interactions is shown.

[0052] Figure 3 A process of generating negative samples for predicting inter-macromolecular interactions is shown.

[0053] Figure 4 A process of training a model for generating a set of negative samples for predicting inter-macromolecular interactions in accordance with some embodiments of the present disclosure is shown.

[0054] Figure 5 A process of generating a set of negative samples for predicting inter-macromolecular interactions in accordance with some embodiments of the present disclosure is shown.

[0055] Figure 6 A specific example of generating a set of negative samples for predicting inter-macromolecular interactions in accordance with some embodiments of the present disclosure is shown.

[0056] Figure 7 is a schematic diagram of the structure of a device in accordance with some embodiments of the present disclosure. DETAILED DESCRIPTION

[0057] The present disclosure will now be described in greater detail in connection with the following embodiments. It should be noted that the following description of some embodiments presented herein is merely intended to illustrate and describe, and is not exhaustive or limited to the precise forms disclosed.

[0058] The present disclosure provides, among other things, a method of generating a set of negative samples for predicting inter-macromolecular interactions, a method of predicting inter-macromolecular interactions, a training method of a model for generating a set of negative samples for predicting inter-macromolecular interactions, and a neural network model for predicting inter-macromolecular interactions, which substantially obviate one or more problems due to limitations and disadvantages of the related art. In one aspect, the present disclosure provides a method of generating a set of negative samples for predicting inter-macromolecular interactions. In some embodiments, the method includes receiving a set of positive samples, the set of positive samples including pairs of a first type of macromolecule and a second type of macromolecule having an inter-macromolecular interaction; generating a first similarity graph of the first type of macromolecule; generating a second similarity graph of the second type of macromolecule; generating a vectorized representation of nodes in the first similarity graph and a vectorized representation of nodes in the second similarity graph; and generating the set of negative samples using the vectorized representation of nodes in the first similarity graph and the vectorized representation of nodes in the second similarity graph. As used herein, the term “set of samples” can include one or more samples. In one example, a set of samples (e.g., a set of negative samples or a set of positive samples) includes a single sample. In another example, a set of samples (e.g., a set of negative samples or a set of positive samples) includes multiple samples.

[0059] As used herein, the term "macromolecule" refers to any protein, ribonucleic acid (RNA), deoxyribonucleic acid (DNA), carbohydrate, polypeptide, polynucleotide, and other large biological molecules.

[0060] Using a computational model to predict inter-macromolecular interactions (e.g., protein-RNA interactions) often requires positive and negative samples. Positive samples include known pairs of macromolecules (e.g., protein-RNA pairs) based on experiments. However, it is difficult to find experimentally verified negative samples. Although negative samples can be randomly generated, the quality of these randomly generated negative samples cannot be guaranteed because they inevitably include false negative samples. Because the quality of negative samples is not satisfactory, the prediction performance of a computational model using these randomly generated negative samples is unreliable.

[0061] Figure 1 A process of generating negative samples for predicting inter-macromolecular interactions is shown. Referring to Figure 1 , triangles represent positive samples, and circles represent negative samples. In Figure 1 , shaded circles represent negative samples that are similar to positive samples. For example, a positive sample is represented as (r, p), and a similar negative sample can be represented as (r1, p), where r1 is similar to r. Ideally, a positive sample dataset for predicting inter-macromolecular interactions (e.g., protein-RNA interactions) should not include shaded circles, and a negative sample dataset should include these shaded circles. The line between a triangle and a shaded circle indicates classifying a sample into a positive sample dataset or a negative sample dataset.

[0062] Figure 2 A process of generating negative samples for predicting inter-macromolecular interactions is shown. Figure 2 A typical result of a method of using randomly generated negative samples is shown. Referring to Figure 2 , in a method of randomly generating negative samples, it is relatively easy to identify negative samples that are not very similar to positive samples (represented by circles with solid lines). However, it is very difficult to identify negative samples that are similar to positive samples (circles with dashed lines, e.g., (r1, p)). Machine learning classifiers in inter-macromolecular interaction (e.g., protein-RNA interaction) prediction can mistakenly classify them into a positive sample dataset, as indicated by the line between a circle with a dashed line and a circle with a solid line.

[0063] The inventors of the present disclosure found that determining the similarity of negative samples is another factor that can affect classification accuracy. Figure 3 A process of generating negative samples for predicting inter-macromolecular interactions is shown. Referring to Figure 3, a triangle with a dashed line represents a sample point that is highly similar to a positive sample (triangle with a solid line). These samples are actually positive samples, but are incorrectly identified as negative samples because of the use of high similarity in generating negative samples, resulting in misclassification of these sample points into the negative sample dataset, as indicated by the line between the triangle with a dashed line and the triangle with a solid line.

[0064] Figure 4 A process of training a model for generating a negative sample set for predicting inter-macromolecular interactions is shown, in accordance with some embodiments of the present disclosure. Reference is made to Figure 4 In some embodiments, the present method includes generating a first similarity graph of a first type of macromolecule (e.g., an RNA molecule); and generating a second similarity graph of a second type of macromolecule (e.g., a protein molecule). In some embodiments, the first similarity graph includes similarities of multiple pairs (e.g., all pairs) of the first type of macromolecule; and the second similarity graph includes similarities of multiple pairs (e.g., all pairs) of the second type of macromolecule.

[0065] Various suitable methods can be used to determine the similarity between pairs of macromolecules. Examples of suitable methods for determining similarity include edit distance comparison methods, token-based comparison methods, and sequence-based comparison methods. An edit distance comparison method determines the number of operations required to transform a first macromolecule into a second macromolecule. The more operations required, the less similar the macromolecules are. Specific examples of edit distance comparison methods include Hamming distance methods, Levenshtein distance methods, and Jaro-Winkler methods. In some embodiments, similarity is calculated by a distance between respective pairs of macromolecules. In one example, respective similarities between respective pairs of the first type of macromolecule are denoted as:

[0066] sim1(m 1-1 , m 1-2 ) = 1 - d1(m 1-1 , m 1-2 )

[0067] where (m 1-1 , m 1-2 ) represents a respective pair of the first type of macromolecule, sim1 represents a respective similarity between the respective pair of the first type of macromolecule, and d1 represents a distance between the respective pair of the first type of macromolecule.

[0068] Optionally, the distance d1 is denoted as:

[0069]

[0070] where lev(m 1-1 , m 1-2represents the length of the first first-type macromolecule in the respective pair, len(m 1-1 represents the length of the second first-type macromolecule in the respective pair. 1-2 represents the length of the second first-type macromolecule in the respective pair.

[0071] In another example, the respective similarity between the respective pair of second-type macromolecules is represented as:

[0072] sim2(m 2-1 , m 2-2 ) = 1 - d2(m 2-1 , m 2-2 );

[0073] where (m 2-1 , m 2-2 ) represents the respective pair of second-type macromolecules, sim2 represents the respective similarity between the respective pair of second-type macromolecules, and d2 represents the distance between the respective pair of second-type macromolecules.

[0074] Optionally, the distance d2 is represented as:

[0075]

[0076] where len(m 2-1 , m 2-2 ) represents the edit distance between the respective pair of second-type macromolecules, len(m 2-1 ) represents the length of the first second-type macromolecule in the respective pair, and len(m 2-2 ) represents the length of the second second-type macromolecule in the respective pair.

[0077] In some embodiments, the nodes in the first similarity graph represent respective sequences of respective first-type macromolecules, the edges connecting two adjacent nodes in the first similarity graph represent distances between respective pairs of first-type macromolecules, and the weights of the edges represent respective similarities between the two adjacent nodes. The nodes in the second similarity graph represent respective sequences of respective second-type macromolecules, the edges connecting two adjacent nodes in the second similarity graph represent distances between respective pairs of second-type macromolecules, and the weights of the edges represent respective similarities between the two adjacent nodes.

[0078] With reference to Figure 4 In some embodiments, the method further includes generating a vectorized representation of the nodes in the first similarity graph; and generating a vectorized representation of the nodes in the second similarity graph. In one specific example, a graph neural network (GNN) can be used to generate the vectorized representation of the nodes.

[0079] In some embodiments, the first similarity graph includes N1nodes {e i{e j ,j = 1,..., M}. In one example, the respective vectorized representation of a respective node in the first similarity graph is expressed as:

[0080]

[0081] where e i represents a respective node in the first similarity graph; h y1 (e i ) represents the respective vectorized representation of the respective node e i before the t1th iteration; h t1+1 (e i ) represents the updated respective vectorized representation of the respective node e i after the t1th iteration; σ represents a leaky relu activation function; N(e i ) represents a set of nodes adjacent to the respective node e i ; W p , W ph represent parameters of a graph neural network used to generate the vectorized representation. Optionally, the method randomly initializes the parameters of the graph neural network and the initial respective vectorized representation h i (e 0 ) of the respective node e i . Optionally, a maximum value of t1 can be used. In one example, the maximum value of t1 is a positive integer, such as 10.

[0082] In one example, the vectorized representation of a node in the first similarity graph is represented by dr i .

[0083] In one example, the first type of macromolecule is a protein, the first similarity graph comprises similarities between pairs of proteins; and the respective vectorized representation is computed for the first similarity graph comprising similarities between pairs of proteins. In another example, the first type of macromolecule is an RNA molecule, the first similarity graph comprises similarities between pairs of RNA molecules; and the respective vectorized representation is computed for the first similarity graph comprising similarities between pairs of RNA molecules. Optionally, the first similarity graph is a similarity graph of RNA molecules.

[0084] In some embodiments, the second similarity graph comprises N2nodes {e’ i ,i = 1,..., N2}; and M2edges {r’ j ,j = 1,..., M2}. In one example, the respective vectorized representation of a respective node in the second similarity graph is expressed as:

[0085]

[0086]

[0087] Among them, e' i Represents the corresponding node in the second similar graph; h t2 (e′ i ) represents the corresponding node e' before the t2th iteration. i The corresponding vectorized representation; h t2+1 (e′ i ) represents the corresponding node e' after the t2th iteration. i The updated corresponding vectorized representation; σ represents the activation function of the leakage rectifier linear unit; N(e′) i ) represents the corresponding node e' i A pair of adjacent nodes; <h t2 (e′ i ), h t2 (e′ k )> represents h t2 (e′ i ) and h t2 (e′ k The inner product of ) These represent the parameters used by the graph neural network to generate the vectorized representation; The representative represents node e' i and e' k Attention weights are assigned to the strength of the connections between nodes. Optionally, this method randomly initializes the parameters of the graph neural network and the corresponding nodes e'. i The initial corresponding vectorized representation h 0 (e′ i Alternatively, the maximum value of t2 can be used. In one example, the maximum value of t2 is a positive integer, such as 6.

[0088] In one example, the vectorized representation of nodes in the second similarity graph is given by dp. j express.

[0089] In one example, the second type of macromolecule is an RNA molecule, and the second similarity map includes similarities between multiple pairs of RNA molecules; and for the second similarity map including similarities between multiple pairs of RNA molecules, a corresponding vectorized representation is calculated. In another example, the second type of macromolecule is a protein, and the second similarity map includes similarities between multiple pairs of proteins; and for the second similarity map including similarities between multiple pairs of proteins, a corresponding vectorized representation is calculated. Optionally, the second similarity map is a similarity map of protein molecules.

[0090] refer to Figure 4In some embodiments, the method further comprises determining a probability of an interaction between a first respective vectorized representation of a node in the first similarity graph and a second respective vectorized representation of a node in the second similarity graph by:

[0091]

[0092] where dm1 i represents a vectorized representation of a node in the first similarity graph; dm2 j represents a vectorized representation of a node in the second similarity graph; p(l | dm1 i , dm2 i ) represents a probability of an interaction between a first respective vectorized representation of a node in the first similarity graph and a second respective vectorized representation of a node in the second similarity graph; [ ] represents concatenation between elements separated by “,” (e.g., represents concatenation between dm1 i , dm2 j , dm1 i |-dm2 j and ; represents the product of two vectors; and θ represents a tunable parameter (e.g., a parameter that can be tuned by a training method).

[0093] In some embodiments, the method further comprises training the model using a set of positive samples {(m1 i , m2 i ), i = 1,..., K}, where m1represents a first type of macromolecule and m2represents a second type of macromolecule. Various suitable methods can be used to train the model. In one example, the method comprises training the model using stochastic gradient descent to minimize a loss function L:

[0094]

[0095] By minimizing the loss function L, the model can be fine-tuned and the model parameters can be optimized to better detect interactions between the first type of macromolecule and the second type of macromolecule. Because the inputs to the model include the first similarity graph and the first similarity graph, machine learning can be performed using these similarity graphs. Examples of parameters that are optimized by minimizing the loss function L include W p , W ph , and θ.

[0096] In another aspect, the disclosure provides a neural network model for predicting interactions between macromolecules trained by the method described herein for generating a model for predicting a set of negative samples of interactions between macromolecules.

[0097] In another aspect, the present disclosure provides methods of generating a negative sample set for predicting inter-macromolecular interactions. Figure 5 A process of generating a negative sample set for predicting inter-macromolecular interactions according to some embodiments of the present disclosure is shown. Reference is made to Figure 5 In some embodiments, the method includes receiving a positive sample set including pairs of first-type macromolecules and second-type macromolecules having inter-macromolecular interactions; generating a first similarity graph of the first-type macromolecules; generating a second similarity graph of the second-type macromolecules; generating vectorized representations of nodes in the first similarity graph and vectorized representations of nodes in the second similarity graph; and generating the negative sample set using the vectorized representations of nodes in the first similarity graph and the vectorized representations of nodes in the second similarity graph. The vectorized representations of nodes, the first similarity graph, and the second similarity graph can be generated and stored in any suitable manner. In one example, the vectorized representations of nodes, the first similarity graph, and the second similarity graph are generated, e.g., from scratch, each time the method is executed to generate a negative sample set. In another example, the vectorized representations of nodes, the first similarity graph, and the second similarity graph are generated during a process of training a model, and stored in a memory for later use in generating one or more negative sample sets.

[0098] Figure 6 A specific example of generating a negative sample set for predicting inter-macromolecular interactions according to some embodiments of the present disclosure is shown. Reference is made to Figure 6 In the specific example, the nodes in the first similarity graph represent respective sequences of the respective first-type macromolecules, the edges connecting two adjacent nodes in the first similarity graph represent distances between the respective pairs of first-type macromolecules, and the weights of the edges represent respective similarities between the two adjacent nodes. The nodes in the second similarity graph represent respective sequences of the respective second-type macromolecules, the edges connecting two adjacent nodes in the second similarity graph represent distances between the respective pairs of second-type macromolecules, and the weights of the edges represent respective similarities between the two adjacent nodes. As Figure 6 In one specific example, the vectorized representations of nodes can be generated using a graph neural network (GNN), as shown.

[0099] Reference is made to Figure 5 and Figure 6 In some embodiments, the method of generating a negative sample set further includes receiving a positive sample set {(m1 i , m2 i ), i = 1,..., K} and generating vectorized representations dr i of nodes in the first similarity graph and vectorized representations dp j of nodes in the second similarity graph.

[0100] In some embodiments, generating the negative sample set further includes computing, with respect to m2 i , m1i similarity between m1 j (j = 1, …, K, and j ≠ i). Various suitable algorithms can be used to compute the similarity. Examples of suitable algorithms include matching, Shingling, SimHash, random projection, and SpotSig. In one example, the similarity between m1 i similarity between m1 j (j = 1, …, K, and j ≠ i):

[0101]

[0102] In some embodiments, generating the negative sample set further comprises sorting m1 i similarity between m1 j (j = 1, …, K, and j ≠ i) to obtain a subset; determining a probability of interaction between m2 j (j = 1, …, K, and j ≠ i) to obtain a subset; determining a probability of interaction between m2 i (j = 1, …, K, and j ≠ i) to obtain a subset; determining a probability of interaction between m2 i (j = 1, …, K, and j ≠ i) to obtain a subset; determining a probability of interaction between m2

[0103]

[0104] wherein dr j represents the vectorized representation of a node in the subset; dp i represents the vectorized representation of m2 i ; P(l|dr j , dp i ) represents the probability of interaction between m2 i and each sample in the subset; [, ] represents concatenation between elements separated by “,” (e.g., represents concatenation between elements dr j , dp i , dr j -dp i and ; represents the product of two vectors; and 0 represents a tunable parameter (e.g., a parameter that can be tuned by a training method).

[0105] In some embodiments, generating the negative sample set further comprises generating a plurality of intermediate sets; and sampling L negative samples from a respective intermediate set of the plurality of intermediate sets. Optionally, the negative sample set comprises negative samples sampled from the plurality of intermediate sets; L is an integer greater than or equal to 1, such as 1, 2, 3, 4, 5, or 6. Optionally, the negative sample set consists of a single negative sample, and thus generating the negative sample set comprises generating a single intermediate set; and sampling a single negative sample from the single intermediate sample set.

[0106] In some embodiments, the positive sample set is represented by {(m1 i ,m2 i ), i = 1,..., K}, where m1 i represents the i-th first type macromolecule, and m2 i represents the i-th second type macromolecule. Optionally, generating a respective intermediate set of the plurality of intermediate sets comprises determining a probability of interaction between m2 i and each sample in m1 i ; and generating the respective intermediate set based on the probability of interaction.

[0107] In one example, generating the negative sample set further comprises placing (m1 j , 1 - P(1 | dr j , dp i )) into an intermediate set when P(1 | dr j , dp i ) is less than a threshold value (such as 0.4, 0.45, 0.5, 0.55, or 0.6). In one example, the threshold value is 0.5. If the intermediate set is empty, it indicates that the given positive sample cannot be used to generate a negative sample.

[0108] In some embodiments, when the intermediate set is not empty, generating the negative sample set further comprises sampling L negative samples from the intermediate set based on the probabilities {p k , k = 1,..., |T|}, where

[0109] and

[0110] |T| represents the number of elements in the intermediate set, and (m1 j , 1 - P(1 | dr j , dp i )) represents the k-th element in the intermediate set.

[0111] Optionally, the first type macromolecule comprises an RNA molecule, and the second type macromolecule comprises a protein molecule.

[0112] Optionally, the first type macromolecule comprises a protein molecule, and the second type macromolecule comprises an RNA molecule.

[0113] Referring again to Figure 5 In some embodiments, the method includes generating a first similarity graph of the first type of macromolecules. As described above, in some embodiments, the respective similarity between each pair of the first type of macromolecules is represented as:

[0114] sim1(m 1-1 , m 1-2 ) = 1 - d1(m 1-1 , m 1-2 )

[0115] where (m 1-1 , m 1-2 ) represents a respective pair of the first type of macromolecules, sim1 represents the respective similarity between the respective pair of the first type of macromolecules, and d1 represents the distance between the respective pair of the first type of macromolecules.

[0116] Optionally, the distance d1 is represented as:

[0117]

[0118] where lev(m 1-1 , m 1-2 ) represents the Levenshtein distance between the respective pair of the first type of macromolecules, len(m 1-1 ) represents the length of the first of the first type of macromolecules in the respective pair, and len(m 1-2 ) represents the length of the second of the first type of macromolecules in the respective pair.

[0119] Referring again to Figure 5 In some embodiments, the method includes generating a second similarity graph of the second type of macromolecules; as described above, in some embodiments, the respective similarity between each pair of the second type of macromolecules is represented as:

[0120] sim2(m 2-1 , m 2-2 ) = 1 - d2(m 2-1 , m 2-2 )

[0121] where (m 2-1 , m 2-2 ) represents a respective pair of the second type of macromolecules, sim2 represents the respective similarity between the respective pair of the second type of macromolecules, and d2 represents the distance between the respective pair of the second type of macromolecules.

[0122] Optionally, the distance d2 is represented as:

[0123]

[0124] where lev(m 2-1 , m 2-2) represents the length of the first second-type macromolecule in the corresponding pair, len(m 2-1 ) represents the length of the first second-type macromolecule in the corresponding pair, len(m 2-2 ) represents the length of the second second-type macromolecule in the corresponding pair.

[0125] Referring again to Figure 5 , in some embodiments, the method includes generating vectorized representations of nodes in the first similarity graph and vectorized representations of nodes in the second similarity graph. As described above, in some embodiments, the first similarity graph includes N1nodes {e i ,i = 1, …, N1}; and M1edges {r j ,j = 1, …, M}. In one example, the respective vectorized representation of a respective node in the first similarity graph is expressed as:

[0126]

[0127] where e i represents the respective node in the first similarity graph; h t1 (e i ) represents the respective vectorized representation of the respective node e i before the t1th iteration; h t1+1 (e i ) represents the updated respective vectorized representation of the respective node e i after the t1th iteration; σ represents a leaky rectified linear unit activation function; N(e i ) represents a set of nodes adjacent to the respective node e i ; W p , W ph represent parameters of a graph neural network used to generate the vectorized representation. Optionally, the method randomly initializes the parameters of the graph neural network and the initial respective vectorized representation h i (e 0 ) of the respective node e i . Optionally, a maximum value of t1 can be used. In one example, the maximum value of t1 is a positive integer, such as 10.

[0128] In one example, the vectorized representation of a node in the first similarity graph is represented by dr i .

[0129] In some embodiments, the second similarity graph includes N2nodes {e’ i ,i = 1, …, N2}; and M2edges {r’ j ,j = 1, …, M2}. In one example, the respective vectorized representation of a respective node in the second similarity graph is expressed as:

[0130]

[0131]

[0132] Among them, e' i Represents the corresponding node in the second similar graph; h t2 (e′ i ) represents the corresponding node e' before the t2th iteration. i The corresponding vectorized representation; h t2+1 (e′ i ) represents the corresponding node e' after the t2th iteration. i The updated corresponding vectorized representation; σ represents the activation function of the leakage rectifier linear unit; N(e′) i ) represents the corresponding node e' i A pair of adjacent nodes; <h t2 (e′ i ), h t2 (e′ k )> represents h t2 (e′ i ) and h t2 (e′ k The inner product of ) These represent the parameters used by the graph neural network to generate the vectorized representation; The representative represents node e' i and e' k Attention weights are assigned to the strength of the connections between nodes. Optionally, this method randomly initializes the parameters of the graph neural network and the corresponding nodes e'. i The initial corresponding vectorized representation h 0 (e′ i Alternatively, the maximum value of t2 can be used. In one example, the maximum value of t2 is a positive integer, such as 6.

[0133] In one example, the vectorized representation of nodes in the second similarity graph is given by dp. j express.

[0134] In another aspect, this disclosure provides a method for predicting macromolecular interactions using a positive sample set and a negative sample set generated by the methods described in this disclosure.

[0135] On the other hand, this disclosure provides an apparatus. Figure 7 This is a schematic diagram illustrating a device according to some embodiments of the present disclosure. Reference Figure 7The device 1000 can include any appropriate type of TV, such as a plasma TV, a liquid crystal display (LCD) TV, a touch screen TV, a projection TV, a non-smart TV, a smart TV, etc. The device 1000 can also include other computing systems, such as a personal computer (PC), a tablet or a notebook computer, a smartphone, etc. In addition, the device 1000 can be any appropriate content presentation apparatus that is capable of presenting appropriate content. A user can interact with the device 1000 to perform other operations of interest.

[0136] As shown in FIG. 1, the device 1000 can include a processor 1002, a storage medium 1004, a display 1006, a communication module 1008, a database 1010, and a peripheral device 1012. Certain apparatuses can be omitted and other apparatuses can be included for better description of the related embodiments. Figure 7

[0137] The processor 1002 can include any appropriate one or more processors. The processor 1002 can include multiple cores for multi-threading or parallel processing. The processor 1002 can execute sequences of computer program instructions to perform various processes. The storage medium 1004 can include memory modules (e.g., ROM, RAM, flash memory modules) and mass storage devices (e.g., CD-ROM and hard disks), etc. The storage medium 1004 can store computer programs to implement various processes when the computer programs are executed by the processor 1002. For example, the storage medium 1004 can store computer programs to implement various algorithms when the computer programs are executed by the processor 1002.

[0138] Further, the communication module 1008 can include certain network interface apparatuses for establishing connections through communication networks (e.g., TV cable networks, wireless networks, the Internet). The database 1010 can include one or more databases for storing certain data and performing certain operations on the stored certain data, such as database searches.

[0139] The display 1006 can provide information to a user. The display 1006 can include any appropriate type of computer display apparatus or electronic device display, such as LCD or OLED based apparatuses. The peripheral device 1012 can include various sensors or other I / O apparatuses, such as a keyboard or a mouse.

[0140] ​All or some of the steps, functions, modules / units in the above disclosed methods, systems, apparatuses can be implemented as software, firmware, hardware, or any suitable combination thereof. In hardware implementation, the division of functionality between the above described functional modules / units is not necessarily a division in physical components. For example, one physical component can serve multiple functions, or one function or step can be performed by several physical components working in cooperation. Certain physical components or all physical components can be implemented as software executed by a processor, such as a central processing unit, a digital signal processor, or a microprocessor, or as hardware, or as an integrated circuit, such as an application specific integrated circuit. Such software can be distributed on computer readable storage media, which can include any medium or means that embodies or stores the computer readable instructions for execution by a computer. Computer storage media includes volatile and non-volatile, removable and non-removable media implemented in any method or technology for storage of information such as computer readable instructions, data structures, program modules or other data. Computer storage media includes, but is not limited to, RAM, ROM, EEPROM, flash memory or other memory technology, CD-ROM, digital versatile disks (DVD) or other optical disk storage, magnetic cassettes, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other medium that can be used to store the desired information and that can be accessed by a computer. Further, it should be appreciated that a computer program product can be tangibly embodied as computer-readable data on a computer-readable storage medium or computer-readable storage device. The computer-readable storage media or computer-readable storage device does not include carrier waves and is not a transitory propagating signal.

[0141] The flow and block diagrams in the drawings show architectural, functional, and operational architectures of possible implementations of apparatuses, methods, and computer program products according to various embodiments of the present disclosure. In this regard, each block in the flow and block diagrams can represent a module, a segment, or a portion of code that comprises one or more executable instructions for implementing the specified logical function(s). It should also be noted that in some alternative implementations, the functions noted in the blocks can occur out of the order noted in the figures. For example, two blocks shown in succession may, in fact, be executed substantially concurrently or the blocks can sometimes be executed in the reverse order, depending upon the functionality involved. It will also be noted that each block of the block diagrams and / or flowchart illustrations, and combinations thereof, can be implemented by special purpose hardware-based systems that perform the specified functions or operations, or combinations of special purpose hardware and computer instructions.

[0142] In some embodiments, the device includes one or more memories and one or more processors, wherein the one or more memories and the one or more processors are connected to each other. In some embodiments, the one or more memories store computer-executable instructions for controlling the one or more processors to: receive a set of positive samples comprising pairs of first type macromolecules and second type macromolecules having inter-macromolecular interactions; generate a first similarity graph of the first type macromolecules; generate a second similarity graph of the second type macromolecules; generate vectorized representations of nodes in the first similarity graph and vectorized representations of nodes in the second similarity graph; and generate the set of negative samples using the vectorized representations of nodes in the first similarity graph and the vectorized representations of nodes in the second similarity graph.

[0143] In some embodiments, the one or more memories store computer-executable instructions for controlling the one or more processors to: receive a set of positive samples comprising pairs of first type macromolecules and second type macromolecules having inter-macromolecular interactions; generate a first similarity graph of the first type macromolecules; generate a second similarity graph of the second type macromolecules; generate vectorized representations of nodes in the first similarity graph and vectorized representations of nodes in the second similarity graph; determine probabilities of interactions between first respective vectorized representations of nodes in the first similarity graph and second respective vectorized representations of nodes in the second similarity graph; and train a model using a loss function.

[0144] In another aspect, the present disclosure provides a computer program product comprising a non-transitory, tangible computer-readable medium having computer-readable instructions thereon. In some embodiments, the computer-readable instructions are executable by a processor to cause the processor to perform: receiving a set of positive samples comprising pairs of first type macromolecules and second type macromolecules having inter-macromolecular interactions; generating a first similarity graph of the first type macromolecules; generating a second similarity graph of the second type macromolecules; generating vectorized representations of nodes in the first similarity graph and vectorized representations of nodes in the second similarity graph; and generating the set of negative samples using the vectorized representations of nodes in the first similarity graph and the vectorized representations of nodes in the second similarity graph.

[0145] In some embodiments, the computer-readable instructions are executable by the processor to cause the processor to perform: receiving a set of positive samples comprising pairs of first type macromolecules and second type macromolecules having inter-macromolecular interactions; generating a first similarity graph of the first type macromolecules; generating a second similarity graph of the second type macromolecules; generating vectorized representations of nodes in the first similarity graph and vectorized representations of nodes in the second similarity graph; determining probabilities of interactions between first respective vectorized representations of nodes in the first similarity graph and second respective vectorized representations of nodes in the second similarity graph; and training a model using a loss function.

[0146] The various illustrative neural networks, layers, units, channels, blocks, and other operations described in connection with the configurations disclosed herein can be implemented to operate with electronic hardware, computer software, or combinations of both. These neural networks, layers, units, channels, blocks, and other operations can be implemented or performed with a general purpose processor, a digital signal processor (DSP), an ASIC or ASSP, an FPGA or other programmable logic device, discrete gate or transistor logic, discrete hardware components, or any combination thereof designed to generate the configurations disclosed herein. For example, such configurations can be implemented at least in part as hard-wired circuitry, as a configuration of circuitry manufactured into an application-specific integrated circuit, or as a firmware program loaded into non-volatile storage, or software program loaded from or into a data storage medium as machine-readable code, such code being instructions executable by an array of logic elements such as a general purpose processor or other digital signal processing unit. The general purpose processor can be a microprocessor, but in the alternative, the processor can be any conventional processor, controller, microcontroller, or state machine. A processor can also be implemented as a combination of computing devices, e.g., a combination of a DSP and a microprocessor, a plurality of microprocessors, one or more microprocessors in conjunction with a DSP core, or any other such configuration. The software modules can reside in a non-transitory storage medium, such as RAM (random-access memory), ROM (read-only memory), non-volatile RAM (NVRAM) such as flash memory, erasable programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), registers, hard disk, a removable disk, or a CD-ROM; or in any other form of storage medium known in the art. An illustrative storage medium is coupled to the processor such that the processor can read information from, and write information to, the storage medium. In the alternative, the storage medium can be integral to the processor. The processor and the storage medium can reside in an ASIC. The ASIC can reside in a user terminal. In the alternative, the processor and the storage medium can reside as discrete components in a user terminal.

[0147] The foregoing description of embodiments of the application has been presented for the purposes of illustration and description. It is not intended to be exhaustive or to limit the application to the precise form disclosed. The description was presented as illustrative of the broadest aspects of the application that are and can be claimed as it is and only limited by the appended claims. Obviously, many modifications and variations will be apparent to those skilled in the art. The embodiments were chosen and described in order to explain the principles of the application and its best mode of practical application to thereby enable others skilled in the art to understand the application for various embodiments and with various modifications that are suited to the particular use or implementation. The scope of the application is to be defined by the claims and their equivalents, where all terms are meant to be construed in their broadest reasonable sense, unless otherwise indicated. Thus, the terms "the invention," "the present invention," or similar referents used in the context of the detailed description are not intended to be limiting of the claimed subject matter, and will be included in the scope of the claims along with the equivalents thereof. The present invention is only limited by the claims appended hereto and encompasses all variations falling within the scope of the claims, which are to be interpreted in accordance with the principles of patent law. Furthermore, these claims can refer to "a," "an" or "the" article, which is intended to be interpreted to mean "at least one" or "one or more." Also, any application recitations of "first," "second," or "third" or similar language are not intended to be construed as limiting the number or order of elements. Any advantages and benefits provided by the described embodiments can not apply in all instances. It should be understood that various changes can be made by those skilled in the art which would be apparent to one skilled in the art. Such changes are not to be considered limiting of the scope of the application as defined by the appended claims and their equivalents. Furthermore, elements and components of the disclosure can be arranged and designed in a wide variety of different configurations, all of which are intended to fall within the scope of the present disclosure. Also, the elements and components of the disclosure can be interchanged with other elements and components that serve the same function or are otherwise suitable for the purposes contemplated by the disclosure.

Claims

1. A method for generating a negative sample set for predicting macromolecular interactions, comprising: Receive a positive sample set, which includes pairs of first-type macromolecules and second-type macromolecules with intermolecular interactions; Generate the first similarity map of the first type of macromolecule; Generate a second similarity graph of the second type of macromolecule; the first similarity graph or the second similarity graph includes nodes and edges connecting adjacent nodes, wherein each node represents a macromolecule, each edge represents the corresponding distance between corresponding pairs of macromolecules, and the corresponding weight of each edge represents the corresponding similarity between corresponding pairs of macromolecules; Generate vectorized representations of nodes in the first similarity graph and vectorized representations of nodes in the second similarity graph; wherein, The first similar graph includes N1 nodes {e i {i = 1, ..., N1}; and M1 edges {r j The corresponding vectorized representation of the corresponding node in the first similarity graph is expressed as: Among them, e i Represents the corresponding node in the first similarity graph; h t1 (e i ) represents the corresponding node e before the t1th iteration. i The corresponding vectorized representation; h t1+1 (e i ) represents the corresponding node e after the t1th iteration. i The updated corresponding vectorized representation; σ represents the leaky relu (rectified linear unit) activation function; N(e i ) represents the corresponding node e i A pair of adjacent nodes; W p W ph These represent the parameters used by the graph neural network to generate the vectorized representation; The second similar graph includes N2 nodes {e' i {i = 1, ..., N2}; and M2 edges {r' j The corresponding vectorized representation of the corresponding node in the second similar graph is expressed as: Among them, e' i Represents the corresponding node in the second similar graph; h t2 (e′ i ) represents the corresponding node e' before the t2th iteration. i The corresponding vectorized representation; h t2+1 (e′ i ) represents the corresponding node e' after the t2th iteration. i The updated corresponding vectorized representation; σ represents the activation function of the leakage rectifier linear unit; N(e′) i ) represents the corresponding node e' i A pair of adjacent nodes; <h t2 (e′ i ), h t2 (e′ k )> represents h t2 (e′ i ) and h t2 (e′ k The inner product of ); W p t+1 These represent the parameters used by the graph neural network to generate the vectorized representation; The attention weight represents the strength of the connection between nodes e'i and e'k; as well as Generating the negative sample set using the vectorized representations of nodes in the first similarity graph and the vectorized representations of nodes in the second similarity graph includes: generating multiple intermediate sets; sampling L negative samples from corresponding intermediate sets in the multiple intermediate sets; wherein the negative sample set includes negative samples sampled from the multiple intermediate sets; and L is an integer greater than or equal to 1. Wherein, the positive sample set consists of {(m1)} i m2 i Let m1, i = 1, ..., K represent the expression, where m1 = ... i Represents the i-th type I macromolecule, m2 i Represents the i-th second-type macromolecule; wherein, generating the corresponding intermediate sets in the plurality of intermediate sets includes: based on m1 i With m1 j The similarity between them, for m1 j Sort the data (j = 1, ..., K, and j ≠ i) to obtain m1. j A subset of (j = 1, ..., K, and j ≠ i); determine m2 i The probability of interaction with each sample in the subset; and, based on the probability of the interaction, generating a corresponding intermediate set; Among them, m2 i The probability of interaction with each sample in the subset is determined by the following formula: Among them, dr j The vectorized representation of the nodes in the subset; dp i Represents m2 i Vectorized representation of P(1|dr j dp i ) represents m2 i The probability of interaction with each sample in the subset; [,] represents the concatenation between elements; θ represents the product of two vectors; θ represents a tunable parameter. When P(1|dr) j dp i When (m1) is less than the threshold, (m1) j ,1-P(1|dr j dp i )) Insert into the corresponding intermediate set, or, based on probability {p k L negative samples are sampled from the corresponding intermediate set, k = 1, ..., |T|. in, |T| represents the number of elements in the corresponding intermediate set, and (m1 j ,1-P(1|dr j dp i )) represents the k-th element in the corresponding intermediate set.

2. The method according to claim 1, further comprising calculating m1 using the following formula i With m1 j Similarity among (j = 1, ..., K, and j ≠ i): in, dr j Vectorized representation of nodes in a subset; dr i Represents m1 i The vectorized representation of .

3. The method according to claim 1 or 2, wherein, The first type of macromolecule includes RNA molecules, and the second type of macromolecule includes protein molecules.

4. A method for predicting macromolecular interactions, the method using the positive sample set and the negative sample set generated by the method according to any one of claims 1 to 3.

5. A method for training a model to generate a negative sample set for predicting macromolecular interactions, comprising: Receive a positive sample set, which includes pairs of first-type macromolecules and second-type macromolecules with intermolecular interactions; Generate the first similarity map of the first type of macromolecule; Generate a second similarity graph of the second type of macromolecule; the first similarity graph or the second similarity graph includes nodes and edges connecting adjacent nodes, wherein each node represents a macromolecule, each edge represents the corresponding distance between corresponding pairs of macromolecules, and the corresponding weight of each edge represents the corresponding similarity between corresponding pairs of macromolecules; Generate vectorized representations of nodes in the first similarity graph and vectorized representations of nodes in the second similarity graph; wherein, The first similarity graph includes N1 nodes {e i {i = 1, ..., N1}; and M1 edges {r j ,j=1,…,M; The corresponding vectorized representation of the corresponding node in the first similarity graph is expressed as: Among them, e i Represents the corresponding node in the first similarity graph; h t1 (e i ) represents the corresponding node e mentioned before the t1th iteration. i The corresponding vectorized representation; h t1+1 (e i ) represents the corresponding node e after the t1-th iteration. i The updated corresponding vectorized representation; σ represents the activation function of the leakage rectifier linear unit; N(e i ) represents the corresponding node e i A pair of adjacent nodes; W p W ph These represent the parameters used by the graph neural network to generate the vectorized representation; The second similarity graph includes N2 nodes {e' i {i = 1, ..., N2}; and M2 edges {r' j ,j=1,…,M2}; The corresponding vectorized representation of the corresponding node in the second similarity graph is expressed as follows: Among them, e' i Represents the corresponding node in the second similarity graph; h t2 (e′ i ) represents the corresponding node e' mentioned before the t2th iteration. i The corresponding vectorized representation; h t2+1 (e′ i ) represents the vectorized representation of the update of the corresponding node e'i after the t2th iteration; σ represents the activation function of the leakage rectifier linear unit; N(e′ i ) represents a group of nodes adjacent to the corresponding node e'i; <h t2 (e′ i ),h t2 (e′ k )> represents h t2 (e′ i ) and h t2 (e′ k The inner product of ); W p t+1 These represent the parameters used by the graph neural network to generate the vectorized representation; The attention weight represents the strength of the connection between nodes e'i and e'k; Determining the probability of an interaction between a first corresponding vectorized representation of a node in the first similarity graph and a second corresponding vectorized representation of a node in the second similarity graph includes: determining the probability of the interaction using the following formula: Among them, dm1 i The vectorized representation of the nodes in the first similarity graph; dm2 j The vectorized representation of the nodes in the second similarity graph; p(1|dm1) i dm2 i ) represents the probability of interaction between the first corresponding vectorized representation of a node in the first similarity graph and the second corresponding vectorized representation of a node in the second similarity graph; [,] represents the concatenation between elements; θ represents the product of two vectors; θ represents a tunable parameter. as well as The model is trained based at least in part on the probabilities of the said interactions.

6. The method according to claim 5, wherein, The corresponding similarity between the first type of macromolecules is expressed as follows: sim1(m 1-1 ,m 1-2 )=1-d1(m 1-1 ,m 1-2 ); Among them, (m 1-1 m 1-2 ) represents the corresponding pair of first-type macromolecules, sim1 represents the corresponding similarity between the corresponding pair of first-type macromolecules, and d1 represents the distance between the corresponding pair of first-type macromolecules.

7. The method according to claim 6, wherein, d1 is expressed as: Where, lev(m) 1-1 ,m 1-2 ) represents the edit distance between the corresponding pairs of first-type macromolecules, len(m 1-1 ) represents the length of the first type I macromolecule in the corresponding pair, len(m) 1-2 ) represents the length of the second first-type macromolecule in the corresponding pair.

8. The method according to claim 5, wherein, The corresponding similarities between the second type of macromolecules are expressed as follows: sim2(m 2-1 ,m 2-2 )=1-d2(m 2-1 ,m 2-2 ); Among them, (m 2-1 m 2-2 ) represents the corresponding pair of second-type macromolecules, sim2 represents the corresponding similarity between the corresponding pair of second-type macromolecules, and d2 represents the distance between the corresponding pair of second-type macromolecules.

9. The method according to claim 8, wherein, d2 is expressed as: Where, lev(m) 2-1 m 2-2 ) represents the edit distance between the corresponding pairs of second-type macromolecules, len(m 2-1 ) represents the length of the first second-type macromolecule in the corresponding pair, len(m 2-2 ) represents the length of the second type II macromolecule in the corresponding pair.

10. The method according to claim 5, wherein, The positive sample set consists of {(m1)} i m2 i Let m1, i = 1, ..., K represent the expression, where m1 = ... i Let m2 represent the i-th type I macromolecule, and m2 i Represents the i-th type II macromolecule; Training the model includes minimizing the loss function: Among them, dm1 i The vectorized representation of the nodes in the first similarity graph; dm2 j The vectorized representation of the nodes in the second similarity graph; p(1|dm1) i dm2 i ) represents the probability of interaction between the first corresponding vectorized representation of a node in the first similarity graph and the second corresponding vectorized representation of a node in the second similarity graph.

11. The method according to any one of claims 5 to 10, wherein, The first type of macromolecule includes RNA molecules, and the second type of macromolecule includes protein molecules.

12. A neural network model for predicting macromolecular interactions, said neural network model being trained by any one of claims 5 to 11.

Images