Deep learning-based mlncs prediction method

Abstract

The application is suitable for the field of bioinformatics, and provides a mlncs prediction method based on deep learning, which comprises the following steps: step 1, constructing a mlncs positive sample set by combining a calculation method and a deep learning method; step 2, based on a SURE reinforcement learning model, changing input features to construct a mlncs negative sample set; step 3, using a capsule network as a mlncs prediction model, and training the mlncs prediction model by using the constructed mlncs positive sample set and the mlncs negative sample set, wherein the application does not need interaction information between proteins and lncRNAs, and only needs sequence information of lncRNAs to realize prediction, and the calculation method, the deep learning method and the reinforcement learning method are used to construct reliable positive and negative sample sets for mlncs prediction work.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of bioinformatics, and particularly to a mlncs prediction method based on deep learning. BACKGROUND

[0002] Moonlighting proteins (MPs) are proteins with two or more functions, such as enzymes, whose most common function is catalysis, while also participating in physiological and biochemical processes such as transcription regulation, cell transduction and apoptosis. In recent years, research on moonlighting proteins has become one of the hotspots in the field of bioinformatics. However, most of the research on moonlighting in physiological and biochemical processes has focused on proteins and protein-coding genes, with little attention paid to lncRNAs. lncRNAs regulate gene expression and function by interacting with proteins, and current research has shown that mutations in lncRNA sequences are highly related to many diseases. Similarly, moonlighting lncRNAs (mlncs) are lncRNAs with two or more functions. The identification and research of mlncs are of great significance to disease treatment, subcellular localization and physical interactions. However, the existing research on mlncs is very limited.

[0003] The main challenges that mlncs research needs to face are as follows:

[0004] 1. The existing mlncs identification methods are MoonFinder and MoonFinder2.0, both of which are based on the interaction of lncRNAs and proteins to predict mlncs. However, the process of obtaining lncRNA and protein interaction information is very time-consuming and tedious;

[0005] 2. mlncs lacks reliable data sets or databases, which is an important challenge that needs to be addressed in mlncs research;

[0006] 3. After obtaining mlncs positive samples through the existing mlncs identification method, if you want to establish a binary classification machine learning model for mlncs prediction, you also need reliable negative samples for training. The current mainstream method is to directly randomly sample from unlabeled samples, or to construct a single classifier. Both methods have a certain negative impact on prediction accuracy.

[0007] Therefore, in view of the above status quo, there is an urgent need to provide a mlncs prediction method based on deep learning to overcome the deficiencies in current practical applications. SUMMARY

[0008] The present application aims to provide a deep learning-based mlncs prediction method, aiming to solve the problems in the above background.

[0009] The present application is implemented based on a deep learning-based mlncs prediction method, which comprises the following steps:

[0010] Step 1: Constructing a mlncs positive sample set by combining a calculation method and a deep learning method;

[0011] Step 2: Based on the SURE reinforcement learning model, changing the input features to construct a mlncs negative sample set;

[0012] Step 3: Using a capsule network as a mlncs prediction model, and training it with the constructed mlncs positive sample set and mlncs negative sample set.

[0013] As a further scheme of the present application: in step 1, the specific steps for constructing a mlncs positive sample set by combining a calculation method and a deep learning method are as follows:

[0014] Extracting lncRNA-protein interaction pairs with an interaction score greater than 0.5 in the RNAInter database, then performing GO semantic similarity calculation on protein clusters interacting with lncRNA, and predicting MPs of the proteins interacting with lncRNA using MEL-MP, to complete the construction of the mlncs positive sample set.

[0015] As a further scheme of the present application: in step 2, the mlncs negative sample screening model is composed of a negative sample screener and a negative sample supervisor, after the reinforcement learning model is trained, the negative sample screener will have the ability to screen reliable negative samples from unlabeled samples, inputting the unlabeled samples of mlncs into the trained negative sample screener to complete the construction of the mlncs negative sample set.

[0016] As a further scheme of the present application: the negative sample screener contains three reinforcement learning elements: state s i , action a i and policy network π θ (a i |s i );

[0017] Wherein the state s i is the feature representation of sample i, the action a i belongs to {0, 1) is the decision of the negative sample screener on sample i, a i = 0 means that the negative sample screener regards sample i as a negative sample, a i = 1 means that the negative sample screener regards sample i as a positive sample, and the policy network π θ (ai |s i )=p(a i |s i ;θ) is a binary classifier constructed from capsule networks, where θ is a parameter that needs to be updated through backpropagation during training.

[0018] As a further aspect of the present invention: the negative sample supervisor includes two reinforcement learning elements: a sample evaluation network and a reward;

[0019] The negative sample selector selects samples to input into the negative sample supervisor. The negative sample supervisor then calculates the AUC score through five-fold cross-validation. The higher the reliability of the samples selected by the negative sample selector and the lower the noise of the sample set, the better the classification performance of the negative sample supervisor and the greater the reward score returned.

[0020] As a further aspect of this invention: In step 3, sequence features, motif features, physicochemical features, and secondary structure features of lncRNA are extracted; each feature is input into a different feature learning subnetwork composed of fully connected layers for high-level feature extraction and dimensional unification; the high-level feature vectors output by the feature learning subnetwork are combined into capsules, and these capsules are projected into the same high-dimensional prediction space through a transformation matrix; different capsules are added together to obtain new capsules, and the vector length is compressed to between 0 and 1 through a squashing activation function; if the output vector length is greater than 0.5, the input lncRNA is predicted as mlncs; if the output vector length is less than 0.5, the input lncRNA is predicted as non-mlncs.

[0021] As a further aspect of the present invention, the formula for the squash activation function is as follows:

[0022]

[0023] Where s j Let v be the input vector. j This is the output vector.

[0024] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0025] The application proposes a brand-new mlncs prediction model based on deep learning, which does not require the interaction information of proteins and lncRNAs, but only needs the sequence information of lncRNAs to realize prediction; a reliable positive and negative sample set is constructed for mlncs prediction work through calculation method, deep learning method and reinforcement learning method; the mlncs prediction model can also be used for developing an online service platform, which is convenient for clinical and scientific research personnel to use the model end to end; compared with the traditional mlncs prediction method, the application only needs to input the sequence information of lncRNAs, and the cumbersome process of obtaining protein interaction information is saved, in addition, the existing mlncs research method lacks reliable data set support, and the application constructs a reliable positive and negative sample set for mlncs prediction work through calculation method, deep learning method and reinforcement learning method, which provides a new exploration direction for the research of mlncs. BRIEF DESCRIPTION OF DRAWINGS

[0026] Figure 1 The flow chart for constructing the mlncs positive sample set in the embodiment of the application.

[0027] Figure 2 The flow chart for constructing the mlncs negative sample set in the embodiment of the application.

[0028] Figure 3 The work flow chart of the mlncs prediction model in the embodiment of the application.

[0029] Figure 4 The sample distribution chart before and after the negative sample screening in the embodiment of the application. DETAILED DESCRIPTION

[0030] The technical solutions in the embodiments of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, not all the embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor are within the protection scope of the application. The specific implementation of the application will be described in detail below in combination with specific embodiments.

[0031] Please refer to Figures 1-3 The mlncs prediction method based on deep learning provided by the embodiment of the application includes the following steps:

[0032] Step 1: Construct the mlncs positive sample set by combining the calculation method and the deep learning method: the construction of the mlncs positive sample set is completed by performing GO semantic similarity calculation on the protein cluster interacting with lncRNA, and performing MP prediction of the protein interacting with lncRNA by using MEL-MP;

[0033] Specifically, the RNAInter database provides information and correlation scores of lncRNA-protein interactions. In the RNAInter database, 14721 lncRNAs, 14522 proteins and 208913 lncRNA-protein interaction pairs with interaction scores greater than 0.5 were extracted. Among them, one lncRNA interacts with multiple proteins, forming a protein cluster associated with the lncRNA. The construction process of the positive sample set is as shown in Figure 1

[0034] Prediction of mlncs based on GO semantic similarity calculation

[0035] The GO function annotation similarity clustering algorithm is used to predict mlncs. First, the GO annotations of 14522 lncRNA interacting proteins are extracted, and then hypergeometric distribution test is performed to calculate the GO function annotations associated with lncRNAs. The more similar the functions of two GO annotations are, the greater the similarity is. Conversely, the smaller the similarity is. Clustering based on GO annotation similarity is performed to obtain the number of lncRNA functional domains. Clustering of lncRNA GO annotations based on similarity is performed to predict mlncs.

[0036] The GO database provides annotation information describing gene functions from three aspects of cellular component (CC), molecular function (MF) and biological process (BP). The GO function annotations of lncRNAs are obtained from the lncRNA-protein interaction network through hypergeometric distribution test. The formula of hypergeometric distribution test is as follows:

[0037]

[0038] Given an lncRNA x and a GO annotation y, where N is the number of all proteins interacting with all lncRNAs in the dataset, n is the number of proteins interacting with x, M is the number of proteins associated with y, and m is the number of proteins interacting with x and associated with y. When P-value(x, y) < 0.05, it indicates that the given lncRNA x and GO annotation y have significant association.

[0039] After calculating the GO annotations of lncRNAs, the simrel index is used to calculate the similarity between GO terms, and the formula is as follows:

[0040] freq(c) = anno(c) + ∑ h∈children(c) freq(h) (2).​

[0041]

[0042]

[0043] where freq(c) is the frequency of GO term c, anno(c) is the number of products of c, children(c) is the set of first-order neighbor child nodes of c; p(c) is the probability of c, root is the root node in GO ontology; sim Rel(c1，c2) is the similarity of GO terms c1 and c2, S(c1, c2) is the set of ancestor nodes in the ontology graph of c1 and c2;

[0044] Based on the Rel method, the lncRNAs are screened and predicted by the following three conditions:

[0045] 1. Associated with at least eight GO annotations;

[0046] 2. Clustered when the GO semantic similarity threshold is 0.1, and the number of class clusters is at least 2;

[0047] 3. Clustered when the GO semantic similarity threshold is 0.5, and the number of class clusters is at least 4;

[0048] First, according to condition 1, remove the lncRNAs associated with less than eight GO annotations; then, calculate the similarity between each other of the GO annotations of BP and MF ontology respectively to obtain the GO matrix similarity matrix, the formula is as follows:

[0049]

[0050] where n is the number of GO annotations, GO i (i = 1, 2,..., n) is the set of GO annotations under the same ontology;

[0051] The GO semantic similarity clustering algorithm based on the Rel method takes a constant as the threshold, when the similarity of two GO annotations is greater than the threshold, it is determined that they are two connected nodes; the set of connected nodes constitutes a connected graph, and the number of connected graphs in the similarity matrix is the number of class clusters in the graph; the input of the algorithm is the GO similarity matrix Array and the hyperparameter threshold, and the output is the number of connected graphs, that is, the number of class clusters; using the GO semantic similarity clustering algorithm based on the Rel method to analyze the BP ontology and the MF ontology of the lncRNAs, clustering is performed when the threshold is 0.1 and 0.5 respectively, the number of class clusters of the lncRNA GO semantic similarity matrix is obtained, and the conditions 2 and 3 are used for screening, and 5138 mlncs are obtained;

[0052] Prediction of mlncs based on lncRNA-moonlighting protein interaction:

[0053] LncRNAs play roles through interacting with proteins, so when the majority of proteins in the protein cluster interacting with a lncRNA are moonlighting proteins, the lncRNA is most likely a mlnc; the MEL-MP model is an integrated model for predicting moonlighting proteins with superior prediction performance; using the MEL-MP model to predict the proteins in the data, 5651 proteins are predicted to be moonlighting proteins;

[0054] In the protein cluster interacting with each lncRNA, a certain number of proteins are predicted to be moonlighting proteins, and hypergeometric test is performed on them to calculate the significance of MPs in the lncRNA-associated protein cluster; as formula (1), given a lncRNA x, N is the total number of proteins, M is the number of proteins predicted to be moonlighting proteins, n is the number of proteins interacting with x, and m is the number of proteins interacting with x and predicted to be moonlighting proteins; when P value < 0.05, it means that moonlighting proteins exist significantly in the proteins interacting with lncRNA x, and then it is inferred that the lncRNA is a mlnc; after hypergeometric test, 1012 mlncs are obtained;

[0055] Taking the union of the results of the above two prediction methods, a total of 5988 mlncs are obtained, and the construction of the mlncs positive sample set is completed.

[0056] Step 2: Constructing a mlncs negative sample screening model using the SURE model and based on the reinforcement learning model, and extracting a mlncs negative sample set from unlabeled lncRNA samples through the reinforcement learning model: the mlncs negative sample screening model is composed of a negative sample screener and a negative sample supervisor, after the reinforcement learning model is trained, the negative sample screener will have the ability to screen reliable negative samples from unlabeled samples, input the unlabeled samples of mlncs into the trained negative sample screener, and complete the construction of the mlncs negative sample set;

[0057] Specifically, NONCODE is a relatively comprehensive lncRNA database in recent years, which provides sequence information and annotation information; 446867 lncRNA sequence information is obtained from the NONCODE database, and the positive samples that have been predicted to be mlncs and the remaining unlabeled samples are distinguished, and a reinforcement learning model is constructed to screen negative samples;

[0058] SURE is a lncRNA and protein interaction negative sample screening model based on reinforcement learning. The experimental results show that the SURE model has superior performance, and the negative samples screened by the SURE model have high reliability. Based on the SURE model, a negative sample screening model for mlncs is constructed, and the workflow is as shown in Figure 2

[0059] Firstly, the multi-modal feature construction of lncRNA sequence information is carried out, including sequence feature, motif feature, physical and chemical feature and secondary structure feature. After completing the feature extraction, the features of the positive samples and the unlabeled samples are input into the reinforcement learning model for negative sample screening. The model is divided into two components: negative sample selector and negative sample supervisor. The negative sample selector is an agent in reinforcement learning, which interacts with the environment, fully learns the features of the positive samples and reduces the noise of the data set through the feedback of the negative sample supervisor to screen the negative samples. The negative sample supervisor is a classifier, which evaluates the negative samples screened by the negative sample selector, calculates the reward and feeds back to the negative sample supervisor. After the training is completed, the negative sample selector will have the ability to screen reliable negative samples.

[0060] The negative sample selector contains three reinforcement learning elements: state, action and policy network.

[0061] The state s i is the feature representation of the sample i; the action a i belongs to {0, 1} is the decision of the negative sample selector for the sample i, a i = 0 means that the negative sample selector regards the sample i as a negative sample, a i = 1 means that the negative sample selector regards the sample i as a positive sample; the action a i is calculated based on the state s i by the policy network; the policy network π θ (a i | s i ) = p(a i | s i ; θ) is a binary classifier constructed by a capsule network, where θ is the parameter that needs to be updated through back propagation during the training process.

[0062] The negative sample supervisor contains two reinforcement learning elements: sample evaluation network and reward. The sample selected by the negative sample selector is input into the negative sample supervisor, and then the negative sample supervisor calculates the AUC score by five-fold cross-validation. The higher the reliability of the sample selected by the negative sample selector, the smaller the noise of the sample set, and the better the classification performance of the negative sample supervisor, and the larger the reward score returned by the negative sample supervisor;

[0063] The sample evaluation network is a network for binary classification task, where​ For the parameters that need to be updated by back propagation in the training process, s i is the feature vector of the i-th sample, l i is the output of the network; a capsule network with the same structure as the mlncs prediction model is used to build a sample evaluation network, and a gradient descent strategy is used to minimize the cross-entropy loss function;

[0064] When the negative sample selector completes the selection of the current epoch, the positive samples and the selected negative samples are input into the negative sample supervisor to obtain the reward; based on five-fold cross-validation, the sample evaluation network is used to calculate the AUC score of the i-th sample in the k-th epoch, and the reward formula returned by the negative sample supervisor to the negative sample selector is as follows: i is the input, a i is the label, and the AUC score is calculated; the reward formula returned by the negative sample supervisor to the negative sample selector in the k-th epoch is as follows:

[0065] R k = AUC k - MAX_AUC k-1 k≥1 (6);

[0066] wherein, in the k-th epoch, only when the AUC k of the i-th sample exceeds the maximum reward value MAX_AUC k-1 in the first k-1 epochs, the negative sample selector will receive a positive reward; otherwise, the negative sample selector will receive a negative reward; MAX_AUC0 is calculated based on the positive samples and the unlabeled samples, wherein the unlabeled samples are regarded as negative samples;

[0067] After the reinforcement learning training is completed, 446867 lncRNA samples are input into the negative sample selector, and 7292 negative sample information is obtained.

[0068] Step 3: Use a capsule network as the mlncs prediction model, and train it using the constructed mlncs positive sample set and mlncs negative sample set.

[0069] Specifically, after obtaining reliable mlncs positive and negative samples, a multi-modal feature fusion capsule network is constructed as the prediction network of mlncs, and the flow chart of its work is as shown in Figure 3

[0070] ​Firstly, the sequence characteristics, motif characteristics, physical and chemical characteristics and secondary structure characteristics of the lncRNA are extracted; each kind of characteristics is input into different feature learning subnets composed of full connection layers to extract high-level features and unify the dimensions; the high-level feature vectors output by the feature learning subnets form capsules and are projected into the same high-dimensional prediction space through a transformation matrix; the directions of different capsules in the same space may be different; the new capsule is obtained by adding them together, and the vector length is compressed to between 0 and 1 through a squashing activation function; finally, if the output vector length is greater than 0.5, the input lncRNA is predicted as mlncs; if the output vector length is less than 0.5, the input lncRNA is predicted as non-mlncs.

[0071] The formula of the squashing activation function is as follows:

[0072]

[0073] Where s j is the input vector, and v j is the output vector.

[0074] In the embodiment of the application, a brand-new mlncs prediction model based on deep learning is proposed, which does not need the interaction information of proteins and lncRNAs, but only needs the sequence information of lncRNAs to realize prediction; the positive and negative sample sets for mlncs prediction work are constructed by calculation methods, deep learning methods and reinforcement learning methods; the mlncs prediction model can also be used for developing an online service platform, which is convenient for clinical and scientific research personnel to use the model end to end; compared with traditional mlncs prediction methods, the application only needs to input the sequence information of lncRNAs, and the cumbersome process of obtaining protein interaction information is saved; in addition, the existing mlncs research methods lack reliable data set support, and the application constructs the positive and negative sample sets for mlncs prediction work by calculation methods, deep learning methods and reinforcement learning methods, which provides a new exploration direction for the research of mlncs.

[0075] In order to prove the effectiveness and superiority of the above method, comprehensive and systematic comparative experiments are carried out; first, six single classification models are selected to use only the mlncs positive sample set for experiments, then seven binary classification methods including the constructed capsule network are selected, the negative sample set is constructed by random sampling, and experiments are carried out, finally the same seven classification methods as above are selected on the positive and negative sample sets constructed by calculation methods, deep learning methods and machine learning methods, and experiments are carried out, and the experimental results are shown in Table 1:

[0076] Table 1 Experimental results

[0077]

[0078]

[0079]

[0080] In addition, the positive sample and the unmarked sample, and the positive sample and the negative sample screened by reinforcement learning are respectively visualized, Figure 4 The sample distribution before and after sample screening is shown.

[0081] The above only describes the preferred embodiments of the present application and is not used to limit the present application, and any modification, equivalent replacement and improvement made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A deep learning-based mlncs prediction method, characterized by, The method comprises the following steps: Step 1: combine the calculation method with the deep learning method to construct the mlncs positive sample set; Step 2: based on the SURE reinforcement learning model, change the input features to construct the mlncs negative sample set; Step 3: use the capsule network as the mlncs prediction model, and train it with the constructed mlncs positive sample set and mlncs negative sample set; In step 2, the mlncs negative sample screening model is composed of a negative sample selector and a negative sample supervisor. After the reinforcement learning model is trained, the negative sample selector will have the ability to screen reliable negative samples from unlabeled samples. The unlabeled samples of mlncs are input into the trained negative sample selector to complete the construction of the mlncs negative sample set. The negative sample selector comprises three reinforcement learning elements: a state , an action , and a policy network ; where state is the feature representation of sample i, action belongs to {0, 1} is the decision of the negative sample filter for sample i, = 0 means that the negative sample filter considers sample i as a negative sample, strategy network i = 1 means that the negative sample filter considers sample i as a positive sample, policy network is a binary classifier built from a capsule network, where are the parameters that need to be updated by backpropagation during the training process; In step 3, the sequence features, motif features, physical and chemical features, and secondary structure features of lncRNA are extracted; each feature is input into a feature learning subnet composed of a fully connected layer to extract high-level features and unify the dimensions; the high-level feature vectors output by the feature learning subnet form a capsule and are projected into the same high-dimensional prediction space through a transformation matrix; the new capsule is obtained by adding different capsules, and the vector length is compressed to 0-1 through the squashing activation function; if the output vector length is greater than 0.5, the input lncRNA is predicted as mlncs; if the output vector length is less than 0.5, the input lncRNA is predicted as non-mlncs. 2.The deep learning-based mlncs prediction method of claim 1, wherein, In step 1, the specific steps of combining the calculation method with the deep learning method to construct the mlncs positive sample set are as follows: In the RNAInter database, extract lncRNA-protein interaction pairs with an interaction score greater than 0.5, then perform GO semantic similarity calculation on protein clusters interacting with lncRNA, and use MEL-MP to predict MPs for proteins interacting with lncRNA to complete the construction of the mlncs positive sample set. 3.The deep learning-based mlncs prediction method of claim 1, wherein, The negative sample supervisor contains two reinforcement learning elements: a sample evaluation network and a reward; The sample selected by the negative sample selector is input into the negative sample supervisor, and then the negative sample supervisor calculates the AUC score through five-fold cross-validation. The higher the reliability of the sample selected by the negative sample selector, the smaller the noise of the sample set, and the better the classification performance of the negative sample supervisor, and the larger the reward score returned by the negative sample supervisor. 4.The deep learning-based mlncs prediction method of claim 1, wherein, The formula of the squashing activation function is as follows: ； wherein is the input vector, is the output vector.

Images