Method, system and storage medium for miRNA target gene prediction and model training

Abstract

The application provides a method, system and storage medium for miRNA target gene prediction and model training. The training method of the miRNA target gene prediction model comprises: projecting input data spliced from the sequence of miRNA and the sequence of mRNA into A base space, U base space, G base space and C base space, thereby obtaining A base vector, U base vector, G base vector and C base vector; extracting feature tensor from the A base vector, the U base vector, the G base vector and the C base vector by using the convolution layer, the activation layer and the pooling layer of the miRNA target gene prediction model connected in four levels in sequence; inputting the extracted feature tensor into a full connection layer to obtain a prediction result; comparing the obtained prediction result with a reference result; and optimizing the miRNA target gene prediction model based on the comparison result.

Description

TECHNICAL FIELD

[0001] The present application relates to bioinformatics, and more particularly, to a method and system for miRNA target gene prediction and model training, and a storage medium. BACKGROUND

[0002] microRNA (hereinafter referred to as "miRNA") is a class of non-coding single-stranded RNA molecules encoded by endogenous genes, generally about 18-25 nt in length. In the process of biological development, miRNA plays an important regulatory role. miRNA mainly realizes the negative regulation of target gene expression by participating in post-transcriptional regulation of genes. The specific action mode is mainly translation inhibition (common in animals) and degradation of target genes (common in plants). miRNA plays an extremely important regulatory role in tumor occurrence and development, biological development, organ formation, virus defense, epigenetic regulation, and metabolism. Understanding the target genes regulated by miRNA is of great significance for tumor prevention and treatment as well as diagnosis and treatment of other diseases.

[0003] However, miRNA has a very complex regulatory network. One miRNA can often regulate multiple target genes, and the same target gene can also have multiple miRNAs for regulation. At present, miRanda, RNAhybrid, PITA, TargetScan and other software are widely used in the industry to predict miRNA target genes. The idea of such software is to calculate the complementary pairing of miRNA and target genes, and further judge whether the target gene is a gene interacting with miRNA according to the thermodynamic stability of the miRNA and target gene combination. Although such software considers the base complementarity of miRNA specific sequences and target genes, the cross-species conservation of target gene untranslated regions, and the thermodynamic stability of miRNA and target gene dimers in the prediction process, and such methods can be used for any species, the calculation is relatively small, however, the algorithm of such software does not truly reflect the biological mechanism of the interaction between miRNA and target genes. Therefore, the accuracy of such software for miRNA target gene prediction is often less than 40%. This greatly increases the workload of later verification, so the time and economic costs are not low. In addition, existing miRNA target gene prediction software (such as miranda, PITA, RNAhybrid, etc.) usually requires users to input information such as thermodynamic energy threshold, score threshold, and target upstream and downstream positions to be considered for folding UTR when performing target gene prediction. On the one hand, these information has a certain bias interference on the prediction results (especially in the current situation where the mechanism of miRNA and target gene interaction is not truly understood), on the other hand, it also brings some inconvenience in use. Therefore, there is a need in the market for a relatively efficient, accurate and convenient method for miRNA target gene prediction. SUMMARY

[0004] The application provides a training method of a miRNA target gene prediction model, comprising: projecting input data spliced from a sequence of miRNA and a sequence of mRNA into A base space, U base space, G base space and C base space, thereby obtaining an A base vector, a U base vector, a G base vector and a C base vector; extracting a feature tensor from the A base vector, the U base vector, the G base vector and the C base vector by using a four-level convolution layer, an activation layer and a pooling layer of the miRNA target gene prediction model connected in sequence; inputting the extracted feature tensor into a full connection layer to obtain a prediction result; comparing the obtained prediction result with a reference result; and optimizing the miRNA target gene prediction model based on a comparison result.

[0005] According to the embodiment of the application, the sequence of miRNA and the sequence of mRNA include the sequence of miRNA and the sequence of mRNA that interact with each other in a positive training set and the sequence of miRNA and the sequence of mRNA randomly generated in a negative training set, wherein: the sequence of miRNA and the sequence of mRNA that interact with each other in the positive training set are the sequence of miRNA and the sequence of mRNA that interact with each other and are verified by low-throughput experiments and are extracted from at least one database of ENCORI, miRDB, miRTarBase, miRNet, miRWalk; the sequence of miRNA and the sequence of mRNA randomly generated in the negative training set are randomly generated sequences, and the randomly generated sequences exclude the sequence of miRNA and the sequence of mRNA that interact with each other in the positive training set and the sequence of miRNA and the sequence of mRNA that interact with each other predicted by miRanda, RNAhybrid and PITA.

[0006] According to the embodiment of the application, the number of pairs of the sequence of miRNA and the sequence of mRNA that interact with each other in the positive training set is the same as the number of pairs of the sequence of miRNA and the sequence of mRNA randomly generated in the negative training set.

[0007] According to the embodiment of the application, the four-level convolution layer, the activation layer and the pooling layer connected in sequence include: a first level with 16 4-channel convolution kernels, a relu activation function and a maximum pooling layer with a window of 2; a second level with 32 16-channel convolution kernels, a relu activation function and a maximum pooling layer with a window of 2; a third level with 64 32-channel convolution kernels, a relu activation function and a maximum pooling layer with a window of 2; and a fourth level with 128 64-channel convolution kernels, a relu activation function and a maximum pooling layer with a window of 2.

[0008] According to an embodiment of the present application, inputting the extracted feature tensor into the fully connected layer to obtain a prediction result comprises: reducing the extracted feature tensor into a first vector of one dimension; randomly pruning half of the elements of the first vector to form a second vector; inputting the second vector into a first fully connected layer with relu as an activation function to obtain a third vector; randomly pruning half of the elements of the third vector to form a fourth vector; and inputting the fourth vector into a second fully connected layer with sigmoid as an activation function to obtain the prediction result.

[0009] According to an embodiment of the present application, optimizing the miRNA target gene prediction model based on the comparison result comprises: using an Adam optimization function to correct parameters of each layer of the miRNA target gene prediction model with binary classification cross-entropy as a loss function.

[0010] The present application also provides a training system of a miRNA target gene prediction model, comprising: a memory storing executable instructions; and one or more processors in communication with the memory to execute the executable instructions to complete the following operations: projecting input data spliced from a sequence of miRNA and a sequence of mRNA into A base space, U base space, G base space and C base space to obtain an A base vector, a U base vector, a G base vector and a C base vector; extracting a feature tensor from the A base vector, the U base vector, the G base vector and the C base vector by using a four-level convolution layer, an activation layer and a pooling layer connected in sequence of a miRNA target gene prediction model; inputting the extracted feature tensor into a fully connected layer to obtain a prediction result; comparing the obtained prediction result with a reference result; and optimizing the miRNA target gene prediction model based on the comparison result.

[0011] The present application also provides a computer readable storage medium for training of a miRNA target gene prediction model, characterized in that the computer readable storage medium stores executable instructions capable of being executed by one or more processors to complete the following operations: projecting input data spliced from a sequence of miRNA and a sequence of mRNA into A base space, U base space, G base space and C base space to obtain an A base vector, a U base vector, a G base vector and a C base vector; extracting a feature tensor from the A base vector, the U base vector, the G base vector and the C base vector by using a four-level convolution layer, an activation layer and a pooling layer connected in sequence of a miRNA target gene prediction model; inputting the extracted feature tensor into a fully connected layer to obtain a prediction result; comparing the obtained prediction result with a reference result; and optimizing the miRNA target gene prediction model based on the comparison result.

[0012] The application also provides a prediction method of a miRNA target gene, comprising: inputting a target miRNA sequence and a target mRNA sequence into a miRNA target gene database integrated based on an existing database to query whether the target miRNA sequence and the target mRNA sequence interact; in response to that the target miRNA sequence and the target mRNA sequence are not queried to interact in the miRNA target gene database, inputting the target miRNA sequence and the target mRNA sequence into a miRNA target gene prediction model trained by the training method provided by the application to predict whether the target miRNA sequence and the target mRNA sequence interact.

[0013] According to the embodiment of the application, the miRNA target gene database integrated based on the existing database comprises: sequences of miRNAs and corresponding mRNA sequences of the miRNAs that interact with each other and are verified by low-throughput experiments, extracted from at least one of ENCORI, miRDB, miRTarBase, miRNet and miRWalk; and sequences of miRNAs and corresponding mRNA sequences of the miRNAs that are verified by high-throughput sequencing after being screened by a predetermined threshold.

[0014] The training method of the miRNA target gene prediction model provided by the application innovatively converts the interaction judgment of the miRNA sequence and the mRNA sequence into a calculation task of image processing, so that a convolutional neural network (CNN) can be used to perform the task of miRNA target gene prediction, thereby providing a more efficient, accurate and convenient prediction scheme of the miRNA target gene. BRIEF DESCRIPTION OF DRAWINGS

[0015] Other features, objects and advantages of the application will become more apparent from the following detailed description of non-limiting embodiments, made with reference to the accompanying drawings:

[0016] Figure 1 is a flowchart of the training method of the miRNA target gene prediction model according to the embodiment of the application;

[0017] Figure 2 is a schematic diagram of projecting sequence data into a base space according to the embodiment of the application;

[0018] Figure 3 is a flowchart of the prediction method of the miRNA target gene according to the embodiment of the application; and

[0019] Figure 4 is a schematic block diagram of the miRNA target gene prediction model training system and the miRNA target gene prediction system according to the embodiment of the application;

[0020] Figure 5is a schematic diagram of a miRNA target gene prediction result according to an embodiment of the present application. DETAILED DESCRIPTION

[0021] For a better understanding of the present application, reference will be made to the technical solutions of the present application in more detail with reference to the accompanying drawings. It should be understood that these detailed descriptions are only a description of exemplary embodiments of the present application, and are not intended to limit the scope of the present application in any way. Throughout the specification, like reference numerals refer to like elements. The expression "and / or" includes any combination or all combinations of one or more of the associated listed items.

[0022] It should be noted that in the present specification, the expressions "first", "second", "third" and the like are only used to distinguish one feature from another feature, and do not represent any limitation on the features. Therefore, the first vector discussed below can also be referred to as the second vector without departing from the teachings of the present application. Conversely, the same also applies.

[0023] In the drawings, the size, proportions and shapes of the legends have been slightly adjusted for ease of illustration. The drawings are merely examples and are not drawn strictly to scale. As used in this document, the words "substantially", "approximately", and similar expressions are used as terms of approximation and not as terms of degree, and are intended to account for the inherent deviations in measured or calculated values that would be recognized by those of ordinary skill in the art.

[0024] It should also be understood that expressions such as "include", "including", "have", "has", "contain" and / or "containing" and the like are open-ended terms that are used in this specification to indicate the presence of the stated feature, element, component, and / or the like, but do not preclude the presence or addition of one or more other features, elements, components and / or combinations thereof. In addition, when expressions such as "at least one of" appear after a list of items, they modify the entire list of items and do not merely modify the individual items in the list. Furthermore, when describing embodiments of the present application, the use of "may" indicates that "one or more embodiments of the present application". Also, the word "exemplary" is intended to mean example or illustrative.

[0025] Unless otherwise defined, all words and phrases used in this document, including engineering and scientific terms, have the same meaning as commonly understood by those of ordinary skill in the art to which the present application belongs. It should also be understood that words defined in commonly used dictionaries should be interpreted as having a meaning consistent with their meaning in the context of the relevant art and should not be interpreted in an idealized or overly formal sense unless specifically so stated in the present application.

[0026] It should be noted that, unless otherwise specified, the embodiments and features in the examples of this application can be combined with each other. Furthermore, unless explicitly limited or contradicted by the context, the specific steps included in the methods described in this application are not limited to the order in which they are described, but can be performed in any order or in parallel. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0027] As mentioned above, since the mechanism of interaction between miRNAs and target genes has not yet been fully understood, existing miRNA target gene prediction software does not truly reflect the biological mechanism of miRNA-target gene interaction in its algorithms. This results in the low prediction accuracy of existing miRNA target gene prediction software.

[0028] Deep learning is a rapidly developing field in recent years. Its greatest feature is its ability to autonomously learn data classification features from data, thus it is widely used in complex feature recognition and classification tasks. The breakthroughs in deep learning in recent years have provided a new possibility for miRNA target gene prediction.

[0029] Recurrent Neural Networks (RNNs) in deep learning are widely used in tasks such as natural language recognition. miRNA and mRNA sequences share some characteristics with natural language in their sequence structure; for example, like natural language, they are sequences composed of characters arranged in sequence. Therefore, when using deep learning to predict miRNA target genes, it might be relatively easy to think of using RNNs to build a model for miRNA target gene prediction.

[0030] However, the applicant recognized many differences between gene sequences and natural language, including but not limited to the fact that gene sequences do not possess the strong contextual semantic relationships of natural language. In fact, gene sequences exhibit strong disorder. Therefore, this application innovatively proposes a method for deconstructing miRNA and mRNA sequences based on image concepts.

[0031] Specifically, both miRNA and mRNA sequences are composed of adenine (A), uracil (U), guanine (G), and cytosine (C). Therefore, this application proposes that the different base information in miRNA and mRNA sequences can be processed by referring to color channels such as R, G, and B in an image. In this case, a convolutional neural network (CNN) can be used to build a model for predicting miRNA target genes, extracting key features in the input layer using convolution.

[0032] Figure 1 This is a flowchart of a training method for a miRNA target gene prediction model according to an embodiment of this application.

[0033] At step S1010, the input data spliced from the sequence of miRNA and the sequence of mRNA is projected to the A base space, the U base space, the G base space and the C base space, thereby obtaining an A base vector, a U base vector, a G base vector and a C base vector.

[0034] When the input data is projected to the A base space, the U base space, the G base space and the C base space, only the elements of the base positions with the same name as the base space are valued as 1, and the elements of the remaining base positions are valued as 0. Specifically, referring to Figure 2 , it is assumed that the input data 2000 spliced from the sequence of miRNA and the sequence of mRNA is

ACUGUACG

[0035] At step S1020, the feature tensor is extracted from the A base vector, the U base vector, the G base vector and the C base vector by using the convolution layer, the activation layer and the pooling layer connected in four levels of the miRNA target gene prediction model. In step S1020, the feature extraction can be performed on different base vectors by referring to the process of feature extraction on input vectors of different color channels in image processing tasks.

[0036] At step S1030, the extracted feature tensor is input into a fully connected layer to obtain a prediction result. The fully connected layer can be connected with all nodes of the previous layer, thereby integrating the attributes of all features.

[0037] At step S1040, the obtained prediction result is compared with the reference result (Ground Truth). The prediction result can be the confidence of the interaction between the sequence of miRNA and the sequence of mRNA.

[0038] Finally, at step S1050, the miRNA target gene prediction model is optimized based on the result of comparison. For example, the parameters of each layer of the miRNA target gene prediction model can be corrected by back propagation based on the result of comparison.

[0039] The training method of the miRNA target gene prediction model provided in the application innovatively converts the interaction judgment of miRNA sequences and mRNA sequences into a calculation task of image processing, so that a convolutional neural network (CNN) can be used to perform the task of miRNA target gene prediction, thereby providing a more efficient and accurate miRNA target gene prediction scheme.

[0040] According to the application, the sequences of the miRNAs and the sequences of the mRNAs include the sequences of the miRNAs and the sequences of the mRNAs that interact with each other in the positive training set and the sequences of the miRNAs and the sequences of the mRNAs that are randomly generated in the negative training set, wherein: the sequences of the miRNAs and the sequences of the mRNAs that interact with each other in the positive training set are the sequences of the miRNAs and the sequences of the mRNAs that are verified by low-throughput experiments and interact with each other, which are extracted from at least one of the databases of ENCORI, miRDB, miRTarBase, miRNet and miRWalk; the sequences of the miRNAs and the sequences of the mRNAs that are randomly generated in the negative training set are randomly generated sequences, and the randomly generated sequences exclude the sequences of the miRNAs and the sequences of the mRNAs that interact with each other in the positive training set and the sequences of the miRNAs and the sequences of the mRNAs that are predicted to interact by miRanda, RNAhybrid and PITA.

[0041] In order to better train the miRNA target gene prediction model, a labeled training set needs to be prepared.

[0042] According to the application, first, gene data of authoritative databases (hereinafter referred to as “existing databases”) such as ENCORI, miRDB, miRNET, miRTarBase and miRWalk can be collected, and then the data is screened. For the data verified by low-throughput experiments such as fluorescence quantitative PCR, northern blot, luciferase reporter, western blot and CLIP in these databases, the data can be selected as positive data. In order to better train the miRNA target gene prediction model, the data from different databases can be cleaned and integrated in advance. For example, entries with mRNA sequence length less than 10 nt in the above databases can be screened out, and redundant data from different databases can be removed. For another example, the format of the data can be unified into the format of

name of miRNA; sequence of miRNA; name of mRNA; sequence of mRNA

[0043] In the process of integrating data of existing databases, for the database with active sites, the sequence of active sites is directly extracted, and the sequence of miRNA target gene is randomly completed to about 90bp; for the database without active sites, the active sites are found by RNAhybrid, and then the corresponding sequence is completed.

[0044] The reason for completing the sequence of miRNA is to simulate the situation that may occur in the actual use of the model. In the process of actual use of the model, the sequence length of mRNA input into the miRNA target gene prediction model is usually greater than 90bp, for example, more than 500bp. If the sequence length of mRNA input into the miRNA target gene prediction model in the process of actual use of the model is less than the sequence length of mRNA input into the model in the process of training the model, the prediction of miRNA target gene may not be performed due to the lack of long enough mRNA sequence information in the process of actual use of the model. However, if the sequence length of input mRNA is greater than the sequence length of mRNA input into the model in the process of training the model, the input mRNA sequence can be "data grouped" by "sliding window method", the grouped data is combined with miRNA (the sequence length is usually about 20bp) one by one, and finally input into the miRNA target gene prediction model to predict the target gene.

[0045] According to the present application, the length of the sliding window is about 90bp, and the step length of the sliding window is 20bp. For example, assuming that the sequence length of mRNA input into the miRNA target gene prediction model is 130bp, the sequence data obtained by the sliding window is the data from 1stbp to 90thbp, the data from 21stbp to 110thbp, and the data from 41stbp to 130thbp.

[0046] In the process of preparing negative data, miRNA and mRNA sequences can be randomly generated and paired. However, in this process, the pairs of miRNA and mRNA sequences that have been verified to interact in existing databases should be excluded, and the pairs of miRNA and mRNA sequences that are predicted to interact by miRanda, RNAhybrid and PITA should also be excluded (even if the confidence of these predicted interactions is low).

[0047] After obtaining positive data and negative data, the positive data and negative data can be prepared into data packets in a ratio of 1:1. Most of the data in the data packet will be divided into a training set (Training Set) for training the miRNA target gene prediction model; and the remaining data will be divided into a testing set (Testing Set) for testing the training results of the miRNA target gene prediction model.

[0048] For example, according to one embodiment of the present application, 96.8% of the data in the data set is selected into the training set, and the remaining 3.2% of the data is selected into the test set.

[0049] According to the present application, the four sequentially connected convolutional layers, activation layers and pooling layers described above include: a first stage having 16 4-channel convolutional kernels, a relu activation function and a maximum pooling layer with a window of 2; a second stage having 32 16-channel convolutional kernels, a relu activation function and a maximum pooling layer with a window of 2; a third stage having 64 32-channel convolutional kernels, a relu activation function and a maximum pooling layer with a window of 2; and a fourth stage having 128 64-channel convolutional kernels, a relu activation function and a maximum pooling layer with a window of 2.

[0050] The data in the training set, i.e., the input data spliced from the sequence of miRNA and the sequence of mRNA, is converted into A base vector, U base vector, G base vector and C base vector after projection. Since the length of miRNA is generally about 20 bp, and the length of mRNA in the training set is generally completed to about 90 bp, the length of A base vector, U base vector, G base vector and C base vector in the present application is set to 110.

[0051] The four-channel base vector is input to the convolutional layer of the first stage. The convolutional layer of the first stage includes 16 4-channel convolutional kernels, thereby convolving the four-channel base vector described above. According to the present application, zero padding can be performed on the data edges before convolution, so that the size of the data obtained after convolution remains consistent with the size of the base vector before convolution. Then, the relu activation function can be used to provide nonlinear characteristics for the data after convolution. The activated data is input to the maximum pooling layer with a window of 2 to reduce the data size and prevent overfitting. In the application scenario of the present application, compared with the average value pooling and other pooling schemes, the maximum pooling scheme is more effective in preventing overfitting. Therefore, after the first stage, an intermediate feature tensor of 16 channels and 55 lengths can be obtained.

[0052] Similarly, the 16-channel intermediate feature tensor is input to the convolutional layer of the second stage. The convolutional layer of the second stage includes 32 16-channel convolutional kernels, thereby convolving the 16-channel intermediate feature tensor described above. According to the present application, zero padding can be performed on the data edges before convolution, so that the size of the data obtained after convolution remains consistent with the size of the intermediate feature tensor before convolution. Then, the relu activation function can be used to provide nonlinear characteristics for the data after convolution. The activated data is input to the maximum pooling layer with a window of 2 to reduce the data size and prevent overfitting. Therefore, after the second stage, an intermediate feature tensor of 32 channels and 28 lengths can be obtained.

[0053] The 32-channel intermediate feature tensor is input to the third-level convolution layer. The third-level convolution layer includes 64 32-channel convolution kernels, so as to perform convolution on the 32-channel intermediate feature tensor. According to the present application, zero padding can be performed on the data edges before the convolution, so that the size of the data obtained after the convolution is consistent with the size of the intermediate feature tensor before the convolution. Then, the relu activation function can be used to provide nonlinear characteristics for the data after the convolution. The activated data is input to the max pooling layer with a window size of 2 to reduce the data size and prevent overfitting. Therefore, after the third level, a 64-channel intermediate feature tensor with a length of 14 can be obtained.

[0054] The 64-channel intermediate feature tensor is input to the fourth-level convolution layer. The fourth-level convolution layer includes 128 64-channel convolution kernels, so as to perform convolution on the 64-channel intermediate feature tensor. According to the present application, zero padding can be performed on the data edges before the convolution, so that the size of the data obtained after the convolution is consistent with the size of the intermediate feature tensor before the convolution. Then, the relu activation function can be used to provide nonlinear characteristics for the data after the convolution. The activated data is input to the max pooling layer with a window size of 2 to reduce the data size and prevent overfitting. Therefore, after the fourth level, a 128-channel feature tensor with a length of 7 can be obtained.

[0055] The 128-channel feature tensor can be reduced in dimension by the Flatten method, so as to obtain a one-dimensional vector with a length of 896, which is referred to as a first vector. Then, half of the elements can be randomly pruned to form a second vector. Randomly pruning half of the elements can effectively prevent overfitting of the data. The second vector is input to the first fully connected layer composed of 128 units with the relu activation function. In the present application, the data processed by the first fully connected layer is referred to as a third vector. Half of the elements of the third vector can be randomly pruned again to form a fourth vector. Then, the fourth vector can be input to the second fully connected layer composed of 1 unit with the sigmoid activation function, so as to obtain the miRNA target gene prediction result.

[0056] After the predicted result is compared with the reference result, the difference between the two is used as a feedback value to optimize the model parameters. Specifically, the binary cross-entropy can be used as the loss function, and the Adam optimization function can be used to correct the parameters of each layer of the miRNA target gene prediction model.

[0057] Figure 3 is a flowchart of the miRNA target gene prediction method according to the embodiments of the present application.

[0058] Referring toFigure 3 The prediction method 3000 of the miRNA target gene comprises two steps. In step S3010, the target miRNA sequence and the target mRNA sequence are input into the miRNA target gene database based on the existing database integration to query whether the target miRNA sequence interacts with the target mRNA sequence. The existing database can be ENCORI, miRDB, miRTarBase, miRNet, miRWalk, etc. The integrated database not only includes the sequences of the miRNAs and the corresponding mRNA sequences that interact with each other and are verified by low-throughput experiments extracted from the existing database, but also includes the sequences of the miRNAs and the corresponding mRNA sequences that are verified by high-throughput sequencing after being screened by a predetermined threshold. The above threshold can be set to be higher to ensure the quality of the database.

[0059] If it is queried in the integrated miRNA target gene database that the target miRNA sequence and the target mRNA sequence interact with each other, the query result is directly returned. If it is not queried in the miRNA target gene database that the target miRNA sequence interacts with the target mRNA sequence, in step S3020, the target miRNA sequence and the target mRNA sequence are input into the miRNA target gene prediction model trained by the training method described above to predict whether the target miRNA sequence interacts with the target mRNA sequence.

[0060] The prediction method of the miRNA target gene proposed in the present application can not only make full use of the existing gene database, but also predict the interaction relationship between the miRNA and the mRNA that is not included in the existing gene data through the strong learning ability of the CNN, and is a relatively efficient and accurate miRNA target gene prediction method.

[0061] The following refers to Table 1 to illustrate a verification example of the miRNA target gene prediction model proposed in the present application.

[0062] First, the species “human” is selected from the miRBase official website, and a miRNA derived from human is randomly selected, for example, hsa-miR-125b-5p.

[0063] Table 1-List of hsa-miR-125b-5p related information in miRBase official website

[0064]

[0065]

[0066] Then, the mRNA (STARD13) with the highest matching degree is selected from the database and its sequence is obtained.

[0067] The sequence of the miRNA and the sequence of the mRNA are input into the miRNA target gene prediction model proposed in the present application. In order to verify the quality of the model, the data in the miRBase are not used as the training set. The final output of the model is: True; 0.76700366. Wherein, True indicates that the miRNA interacts with the mRNA, and the number indicates the probability of interaction.

[0068] As a control, the sequence of the miRNA and a randomly selected sequence that does not interact with the miRNA are input into the miRNA target gene prediction model proposed in the present application, and the final output of the model is: False (no interaction).

[0069] The present application also proposes a miRNA target gene prediction system and a model training system thereof, which can be realized in the form of a mobile terminal, a personal computer (PC), a tablet computer, a server, etc. Reference is made below to Figure 4 which shows a structural schematic diagram of a system suitable for realizing the embodiments of the present application.

[0070] As shown in Figure 4 , the computer system includes one or more processors, a communication unit, etc., such as one or more central processing units (CPUs) 401 and / or one or more graphic processing units (GPUs) 413, etc. The processor can perform various appropriate actions and processes according to executable instructions stored in a read-only memory (ROM) 402 or loaded into a random access memory (RAM) 403 from a storage unit 408. The communication unit 412 can include but is not limited to a network card, which can include but is not limited to an IB (Infiniband) network card.

[0071] The processor can communicate with the read-only memory 402 and / or the random access memory 403 to execute the executable instructions, be connected to the communication unit 412 through the bus 404, and communicate with other target devices through the communication unit 412, so as to complete the operations corresponding to any one of the methods proposed in the embodiments of the present application, such as: projecting the input data spliced by the sequence of the miRNA and the sequence of the mRNA into the A base space, the U base space, the G base space and the C base space, so as to obtain the A base vector, the U base vector, the G base vector and the C base vector; extracting the feature tensor from the A base vector, the U base vector, the G base vector and the C base vector by using the convolution layer, the activation layer and the pooling layer connected in four levels of the miRNA target gene prediction model in sequence; inputting the extracted feature tensor into the full connection layer to obtain the prediction result; comparing the obtained prediction result with the reference result; and optimizing the miRNA target gene prediction model based on the comparison result.

[0072] Further, in the RAM 403, various programs and data required for device operation can also be stored. The CPU 401, the ROM 402, and the RAM 403 are connected to each other via the bus 404. In the case of the RAM 403, the ROM 402 is an optional module. The RAM 403 stores executable instructions, or writes executable instructions to the ROM 402 at runtime, which causes the processor 401 to perform operations corresponding to the above-described communication method. An input / output interface (I / O interface) 405 is also connected to the bus 404. The communication unit 412 can be integrally provided or can be provided as a plurality of sub-modules (for example, a plurality of IB network cards) and linked on the bus.

[0073] The following components are connected to the I / O interface 405: an input unit 406 including a keyboard, a mouse, and the like; an output unit 407 including a cathode ray tube (CRT), a liquid crystal display (LCD), and the like, and a speaker, and the like; a storage unit 408 including a hard disk, and the like; and a communication unit 409 including a network interface card such as a LAN card, a modem, and the like. The communication unit 409 performs communication processing via a network such as the Internet. A drive 410 is also connected to the I / O interface 405 as needed. A removable medium 411 such as a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory, and the like is mounted on the drive 410 as needed.

[0074] Need to be explained, as Figure 4 the architecture shown is only one optional implementation, in the specific practice process, can be selected according to actual needs, the number and type of the above Figure 4 Components are selected, reduced, increased or replaced; in different functional component settings, separate settings or integrated settings can also be used, for example, GPU and CPU can be set separately or GPU can be integrated on CPU, communication unit 412 can be set separately, or integrated on CPU or GPU, and the like. These alternative embodiments all fall within the scope of the present application.

[0075] In particular, according to the present application, with reference to the flow Figure 1The described process can be implemented as a computer program product. For example, the present application proposes a computer program product comprising computer readable instructions which, when executed by a processor, implement the following operations: projecting input data spliced from a sequence of miRNA and a sequence of mRNA into A base space, U base space, G base space and C base space, thereby obtaining an A base vector, a U base vector, a G base vector and a C base vector; extracting a feature tensor from the A base vector, the U base vector, the G base vector and the C base vector using a four-level sequentially connected convolution layer, an activation layer and a pooling layer of a miRNA target gene prediction model; inputting the extracted feature tensor into a fully connected layer to obtain a prediction result; comparing the obtained prediction result with a reference result; and optimizing the miRNA target gene prediction model based on a result of the comparison.

[0076] In such an embodiment, the computer program product can be downloaded and installed from a network by the communication unit 409, and / or read and installed from the detachable medium 411. When the computer program product is executed by the central processing unit (CPU) 401, the above-described functions defined in the method of the present application are executed.

[0077] The technical solution of the present application can be implemented in many ways. For example, the technical solution of the present application can be implemented by software, hardware, firmware, or any combination of software, hardware, and firmware. The order of steps for describing the method is only provided for the purpose of more clearly illustrating the technical solution. Unless specifically limited, the steps of the method of the present application are not limited to the above specifically described order. In addition, in some embodiments, the present application can also be implemented as a storage medium storing a computer program product.

[0078] Figure 5 A comparison chart showing the results of target gene prediction using a miRNA target gene prediction model trained according to an embodiment of the present application alone, the results of target gene prediction using a miRNA target gene prediction model trained according to an embodiment of the present application and a miRNA target gene database together, and the results of target gene prediction using existing target gene prediction software is shown.

[0079] As Figure 5As shown, the miRNA target gene prediction is performed on ENCORI database, miRNet database and miRTarbase database respectively. In the prediction results made on each database, from left to right are the results predicted by PITA, the results predicted by miranda, the results predicted only by the miRNA target gene prediction model according to the embodiments of the present application, and the results predicted by both the miRNA target gene prediction model according to the embodiments of the present application and the miRNA target gene database. It should be noted that the miRNA target gene prediction model according to the embodiments of the present application is trained based on the screened gene data in the existing database, while the miRNA target gene database is trained based on the gene data in the existing database without screening. Figure 5 The database total number in the above formula includes the data in the existing database which is not screened into positive data. Therefore, Figure 5 The results predicted by both the miRNA target gene prediction model according to the embodiments of the present application and the miRNA target gene database shown above are not 100%, but have a smaller deviation compared with 100%. By Figure 5 It can be seen that the prediction accuracy of the miRNA target gene prediction model trained according to the embodiments of the present application has exceeded that of the conventional target gene prediction software, and the prediction effect obtained by combining the miRNA target gene database and the miRNA target gene prediction model has far exceeded that of the conventional target gene prediction software.

[0080] The above description is only for the embodiments of the present application and the explanation of the technical principles applied. It should be understood by those skilled in the art that the protection scope of the present application is not limited to the technical solutions formed by the specific combination of the technical features described above, and should also cover other technical solutions formed by any combination of the above technical features or equivalent features without departing from the technical concept. For example, the technical solutions formed by mutually replacing the above features with the technical features disclosed in the present application (but not limited to) having similar functions.

Claims

1. A training method for a miRNA target gene prediction model, characterized in that, The training method comprises: projecting input data spliced from sequences of miRNAs and sequences of mRNAs into A base space, U base space, G base space and C base space, thereby obtaining A base vector, U base vector, G base vector and C base vector; extracting feature tensor from the A base vector, the U base vector, the G base vector and the C base vector by using convolution layers, activation layers and pooling layers connected in four levels of the miRNA target gene prediction model in sequence; inputting the extracted feature tensor into a fully connected layer to obtain a prediction result; comparing the obtained prediction result with a reference result; and optimizing the miRNA target gene prediction model based on the comparison result. The sequences of the miRNAs and the sequences of the mRNAs comprise sequences of miRNAs and sequences of mRNAs that interact with each other in a positive training set and sequences of miRNAs and sequences of mRNAs randomly generated in a negative training set, wherein: The sequences of miRNAs and the sequences of mRNAs that interact with each other in the positive training set are sequences of miRNAs and corresponding sequences of mRNAs that interact with each other and are verified by low-throughput experiments and are extracted from at least one of ENCORI, miRDB, miRTarBase, miRNet and miRWalk; the sequences of miRNAs and the sequences of mRNAs randomly generated in the negative training set are randomly generated sequences, and the randomly generated sequences exclude the sequences of miRNAs and the sequences of mRNAs that interact with each other in the positive training set and sequences of miRNAs and sequences of mRNAs that are predicted to interact by miRanda, RNAhybrid and PITA; The convolution layers, activation layers and pooling layers connected in four levels in sequence comprise: a first level with 16 4-channel convolution kernels, a relu activation function and a maximum pooling layer with a window of 2; a second level with 32 16-channel convolution kernels, a relu activation function and a maximum pooling layer with a window of 2; a third level with 64 32-channel convolution kernels, a relu activation function and a maximum pooling layer with a window of 2; and a fourth level with 128 64-channel convolution kernels, a relu activation function and a maximum pooling layer with a window of 2; Inputting the extracted feature tensor into a fully connected layer to obtain a prediction result comprises: reducing the extracted feature tensor to a first vector of one dimension; randomly pruning half of the elements of the first vector to form a second vector; inputting the second vector into a first fully connected layer with relu as the activation function to obtain a third vector; randomly pruning half of the elements of the third vector to form a fourth vector; inputting the fourth vector into a second fully connected layer with sigmoid as the activation function to obtain the prediction result; The optimization of the miRNA target gene prediction model based on the comparison result comprises: using the Adam optimization function to correct the parameters of each layer of the miRNA target gene prediction model with the binary cross-entropy as the loss function.

2. The training method of claim 1, wherein, The pairs of sequences of miRNAs and sequences of mRNAs that interact with each other in the positive training set are the same as pairs of sequences of miRNAs and sequences of mRNAs that are randomly generated in the negative training set. 3.A training system of a miRNA target gene prediction model, characterized in that, The training system comprises: a memory storing executable instructions; and one or more processors in communication with the memory to execute the executable instructions to accomplish the following operations: projecting input data spliced from sequences of miRNAs and sequences of mRNAs into A base space, U base space, G base space and C base space to obtain A base vectors, U base vectors, G base vectors and C base vectors; extracting feature tensors from the A base vectors, the U base vectors, the G base vectors and the C base vectors by using four levels of convolution layers, activation layers and pooling layers connected in sequence of a miRNA target gene prediction model; inputting the extracted feature tensors into a fully connected layer to obtain a prediction result; comparing the obtained prediction result with a reference result; and optimizing the miRNA target gene prediction model based on a result of the comparison; The sequences of miRNAs and the sequences of mRNAs comprise sequences of miRNAs and sequences of mRNAs that interact with each other in a positive training set and sequences of miRNAs and sequences of mRNAs that are randomly generated in a negative training set, wherein: The sequences of miRNAs and the sequences of mRNAs that interact with each other in the positive training set are sequences of miRNAs and corresponding sequences of mRNAs that are verified by low-throughput experiments and that interact with each other, which are extracted from at least one of ENCORI, miRDB, miRTarBase, miRNet and miRWalk; The sequences of miRNAs and the sequences of mRNAs that are randomly generated in the negative training set are randomly generated sequences, and the randomly generated sequences exclude the sequences of miRNAs and the sequences of mRNAs that interact with each other in the positive training set and sequences of miRNAs and sequences of mRNAs that are predicted to interact by miRanda, RNAhybrid and PITA; The four levels of convolution layers, activation layers and pooling layers connected in sequence comprise: a first level having 16 4-channel convolution kernels, a relu activation function and a maximum pooling layer with a window of 2; a second level having 32 16-channel convolution kernels, a relu activation function and a maximum pooling layer with a window of 2; a third level having 64 32-channel convolution kernels, a relu activation function and a maximum pooling layer with a window of 2; and a fourth level having 128 64-channel convolution kernels, a relu activation function and a maximum pooling layer with a window of 2; Inputting the extracted feature tensors into a fully connected layer to obtain a prediction result comprises: dimensional first vector; randomly pruning half of elements of the first vector to form a second vector; inputting the second vector into a first fully connected layer with a relu as an activation function to obtain a third vector; randomly pruning half of elements of the third vector to form a fourth vector; inputting the fourth vector into a second fully connected layer with a sigmoid as an activation function to obtain the prediction result; The optimizing the miRNA target gene prediction model based on the comparison result comprises: using an Adam optimization function to correct parameters of each layer of the miRNA target gene prediction model with a binary classification cross-entropy as a loss function. 4.A computer readable storage medium for training of a miRNA target gene prediction model, characterized in that, The computer readable storage medium stores executable instructions that can be executed by one or more processors to accomplish the following operations: projecting input data spliced from a sequence of miRNA and a sequence of mRNA into A base space, U base space, G base space and C base space, thereby obtaining an A base vector, a U base vector, a G base vector and a C base vector; extracting a feature tensor from the A base vector, the U base vector, the G base vector and the C base vector by using a four-level convolution layer, an activation layer and a pooling layer connected in sequence of the miRNA target gene prediction model; inputting the extracted feature tensor into a fully connected layer to obtain a prediction result; comparing the obtained prediction result with a reference result; and optimizing the miRNA target gene prediction model based on the comparison result; The sequence of miRNA and the sequence of mRNA comprise sequences of miRNA and mRNA that interact with each other in a positive training set and sequences of miRNA and mRNA randomly generated in a negative training set, wherein: The sequences of miRNA and mRNA that interact with each other in the positive training set are sequences of miRNA and corresponding mRNA that interact with each other and are verified by low-throughput experiments and are extracted from at least one of ENCORI, miRDB, miRTarBase, miRNet and miRWalk; The sequences of miRNA and mRNA randomly generated in the negative training set are randomly generated sequences, and the randomly generated sequences exclude the sequences of miRNA and mRNA that interact with each other in the positive training set and sequences of miRNA and mRNA that are predicted to interact by miRanda, RNAhybrid and PITA; The four-level convolution layer, the activation layer and the pooling layer connected in sequence comprise: a first level with 16 4-channel convolution kernels, a relu activation function and a maximum pooling layer with a window of 2; a second level with 32 16-channel convolution kernels, a relu activation function and a maximum pooling layer with a window of 2; a third level with 64 32-channel convolution kernels, a relu activation function and a maximum pooling layer with a window of 2; and a fourth level with 128 64-channel convolution kernels, a relu activation function and a maximum pooling layer with a window of 2; The inputting the extracted feature tensor into a full connection layer to obtain a prediction result comprises: reducing the extracted feature tensor into a first vector of one dimension; randomly cutting half of elements of the first vector to form a second vector; inputting the second vector into a first full connection layer with a relu as an activation function to obtain a third vector; randomly cutting half of elements of the third vector to form a fourth vector; inputting the fourth vector into a second full connection layer with a sigmoid as an activation function to obtain the prediction result; The optimizing the miRNA target gene prediction model based on the comparison result comprises: using an Adam optimization function to correct parameters of each layer of the miRNA target gene prediction model with a binary classification cross entropy as a loss function.

5. A method of predicting a target gene of a miRNA, characterized by, The prediction method comprises: inputting a target miRNA sequence and a target mRNA sequence into a miRNA target gene database integrated based on an existing database to query whether the target miRNA sequence and the target mRNA sequence interact; in response to that the target miRNA sequence and the target mRNA sequence are not queried to interact in the miRNA target gene database, inputting the target miRNA sequence and the target mRNA sequence into the miRNA target gene prediction model trained by the training method according to claim 1 to predict whether the target miRNA sequence and the target mRNA sequence interact.

6. The method for predicting miRNA target genes according to claim 5, characterized in that, The miRNA target gene database integrated based on the existing database comprises: sequences of miRNAs and corresponding mRNA sequences of the miRNAs that interact with each other and are verified by low-throughput experiments extracted from at least one of ENCORI, miRDB, miRTarBase, miRNet and miRWalk; and sequences of miRNAs and corresponding mRNA sequences of the miRNAs that are verified by high-throughput sequencing after being filtered by a predetermined threshold.

Images