Synthetic lethal gene prediction method based on random walk algorithm and application

Abstract

The invention discloses a synthetic lethal gene prediction method based on a random walk algorithm and application. The synthetic lethal gene prediction method comprises the following steps: S1, learning a global structure of a single network based on random walk; s2, extracting a prior constraint; s3, using a semi-supervised automatic encoder to integrate constraints, the method is devoted to a prediction task of a synthetic lethal gene function under a cancer background, a cancer-related gene is taken as a research object, a prediction model is established, the task of predicting the synthetic lethal gene function can be carried out, feature embedding is learned by using a feedforward neural network, and a prediction result is obtained. The neural network technology is used as basic development, prediction of the synthetic lethal gene can be completed from more complex gene pairs, and the prediction accuracy of the synthetic lethal gene is effectively improved; the semi-supervised automatic encoder is adopted, dimensionality reduction of high-dimensional input can be completed, overfitting is reduced, and the method has wide application prospects and is beneficial to popularization and application.

Description

Technical Field

[0001] The present invention relates to the technical field of gene prediction, and in particular to a synthetic lethal gene prediction method based on a random walk algorithm and its application. Background Art

[0002] At present, in terms of mining gene topological structure information, traditional synthetic lethality prediction algorithms cannot make good use of the topological structure information in the graph data, which is one of the important features of genes. In addition, in terms of the method of synthetic lethality prediction using multi-perspective data, it is difficult to achieve data aggregation and mining due to the high latitude of multi-perspective data and the use of different perspectives to describe the complexity of genes. Among them, the topological information of the graph can be mined based on deep learning algorithms, but an effective method needs to be sought to aggregate and fuse multi-perspective data. The Mashup framework technology in the prior art mainly includes three steps: First, run a local network diffusion process to obtain the distribution of each node, which captures its correlation with all other nodes in the network. This step describes the topological background of each gene in the network and takes into account the global connection pattern. Second, these distributions are approximated by building a model that is parameterized by low-dimensional feature vectors of each node; these feature vectors are obtained by simultaneously minimizing the difference between the model distribution and the diffusion distribution of all networks. Mashup calculates a low-dimensional vector space representation of all nodes, which can best explain the diffusion or connection pattern in the network. Third, use the learned representations as input features for a wide range of network-based functional reasoning tasks. However, the matrix decomposition-based method used in Mashup is a linear and shallow method that is difficult to capture complex and highly nonlinear structures in all networks, and the accuracy is not high. Therefore, it is urgent to develop a synthetic lethal gene prediction method and application based on a random walk algorithm to solve the above technical problems.

[0003] In view of this, the present invention is proposed. Summary of the invention

[0004] The purpose of the present invention is to provide a synthetic lethal gene prediction method and application based on a random walk algorithm, which is dedicated to the prediction task of synthetic lethal gene function in the context of cancer, takes cancer-related genes as the research object, and can perform the task of synthetic lethal gene function prediction by establishing a prediction model. A feedforward neural network is used to learn feature embedding, and neural network technology is used as the basis for development. The prediction of synthetic lethal genes can be completed from more complex gene pairs, and the accuracy of synthetic lethal gene prediction is effectively improved. A semi-supervised autoencoder is used to reduce the dimensionality of high-dimensional input and reduce the occurrence of overfitting. The method has broad application prospects and is conducive to popularization and application.

[0005] In order to achieve the above object, the present invention provides a method for predicting synthetic lethal genes based on a random walk algorithm, hereinafter referred to as D-FNN, comprising the following steps:

[0006] S1: Learning the global structure of a single network based on random walks;

[0007] S2: Extract prior constraints;

[0008] S3: Integrating constraints using semi-supervised autoencoders.

[0009] Preferably, the specific method of S1 is:

[0010] Formally, the RWR from node i is defined as:

[0011]

[0012] Where P r is the probability of restart, which controls the relative influence of local and global topological information in the diffusion process. The higher the chance of restart, the more emphasis is placed on local structure; e i is an n-dimensional distribution vector, is an n-dimensional distribution vector, where each entry contains the probability of a node being visited after t steps of random walk starting from node i;

[0013] Based on the above method, RWR is run on each network to capture the individual network topology information and convert it into feature representation of nodes;

[0014] Let M k Represents a k-layer network G k =(V, E (K) ), the adjacency matrix of node v i The RWR is described as the following recursive relation:

[0015]

[0016] Where a is the restart probability; e i is an n-dimensional initial eigenvector; is the n-dimensional feature vector of gene i, each item represents the probability of a gene being visited after t steps of random walk; T is the transition probability matrix, each item T ij Save gene j to gene v i The probability of can be calculated as Based on RWR, we get a matrix S, where S ij is node v i Compared with v defined based on the steady-state probability of RWR j The correlation score between .

[0017] Preferably, the specific method of S2 is:

[0018] The pairwise constraints are formatted as a must-connect constraint and a no-connect constraint. The must-connect constraint indicates that these nodes are highly similar or belong to the same cluster, and the no-connect constraint indicates that the two points in the pair of nodes are highly dissimilar or belong to different clusters.

[0019] The Pearson correlation coefficient PCC value is calculated to measure the pairwise similarity between gene nodes. Given two gene feature vectors A and B, the PCC values ​​of these two genes are:

[0020]

[0021] After calculating the PCC values ​​of all gene pairs, two strategies are used to extract constraints. One is to calculate and sort the paired Pearson correlation coefficients PCC based on the feature vectors of all node pairs. The numerator of the formula is considered to be a must-connect constraint, and the denominator is considered to be a no-connect constraint. The other is to set two thresholds for must-connect and no-connect, marked as f1 and f2, respectively. When the PCC value is greater than f1, a pair is considered a must-connect constraint; when the PCC value is less than f2, a pair is considered a no-connect constraint.

[0022] After extracting the constraints of the previous layer, i.e., layer i, the must-connect constraints and the do-not-connect constraints are applied to the next layer, i.e., layer i+1, as prior information.

[0023] Preferably, the specific method of S3 is:

[0024] Using a semi-supervised autoencoder semiAE, starting from the second layer, the input includes the low-dimensional representation and the constraints of the previous layer, merge these constraints, and use the intersection of all constraints as the input of semiAE. Let x i is the i-th input vector or node representation of the network, f and g are the activations of the hidden layer and output layer respectively, and h i =f(Wx i +b) and y i =g(Mh i + d), where Θ = (θ1, θ2) = {W, b, M, d} are the parameters to be learned, f and g are nonlinear operators such as the simoid function (sigmoid (z) = 1 / (1 + exp (-z))), and the optimization goal is to make the original data x i and reconstructed data y i The reconstruction error between them is minimal;

[0025]

[0026] Let M be a set of pairwise must-link constraints, where (x i , x j )∈M means x i and xj There is a strong connection between them. Let C be a pair of cannot-link constraints, where (x i , x j )∈C means x i and x j Irrelevant, the number of constraints is much smaller than the size of the network |M|+|C|≤|S|;

[0027] If x i and x j If there are must-link constraints between the previous layers, they should also be closed based on the low-dimensional space, and the two must-link nodes after encoding are closer, while the cannot-link nodes are farther away; mathematically, let d(h(x i ); h(x j )) is the x in the encoding space i and x j The error score between ; for must-link, d(x i ;x j ) should be greater than d(h(x i ); h(x j )); for cannot-link, d(x i ;x j ) should be less than d(h(x i ); h(x j )); if (x i ;x j ) is a must-link constraint, adding a penalty to the loss function if (x i ;x j ) is a cannot-link constraint, then a reward is added to the loss function, and the loss function of the modeling constraint is defined as follows:

[0028]

[0029] Among them, the matrices M and C are the must-link constraint matrix and cannot-link constraint matrix extracted from the previous layer respectively; h(x i ) and h(x j ) is the input gene representation x j and x j The hidden layer representation of

[0030] The semi-supervised autoencoder integrates equation (4-1) and equation (4-2) and jointly minimizes the following objective function:

[0031]

[0032] The first part of formula (4-3) measures the squared error of the input and output node features, and the second part measures the error score of the hidden layer constraints;

[0033] FNN (Feedforward Neural Network) is a feedforward neural network. Its main framework consists of 3 Dense layers and 2 Dropout layers. The first layer is a fully connected input layer containing 200 neurons, and the activation function is ReLU. This layer reduces the 625-dimensional input to 200 dimensions; the second layer is a Dropout layer, which randomly discards 40% of the neurons to reduce the occurrence of overfitting; the third layer is a fully connected hidden layer, which contains a fully connected layer of 180 neurons, and the activation function is ReLU. This layer extracts features and transforms dimensions from the 200-dimensional input; the fourth layer is a Dropout layer, which randomly discards 35% of the neurons to reduce the occurrence of overfitting; the fifth layer is a fully connected output layer containing 200 neurons, and the activation function is sigmoid. This layer transforms the 180-dimensional input to 200 dimensions and outputs the probability value of the binary classification 0 or 1;

[0034] The core of the D-FNN algorithm is the semi-supervised autoencoder. The goal of this model is to minimize the loss function (4-3). In this section, we will calculate The partial derivative of is used to verify the above model. The integral constraint loss function (4-2) is:

[0035]

[0036] Where L M =D M -M, D M is a diagonal matrix, L C and L M Similarly, H is a simplified representation of the hidden layer, Calculated as:

[0037]

[0038] In formula (4-5), f is the activation function, that is, the sigmoid function or the ReLu function, and L is obtained. MC The partial derivatives of , when the parameters are randomly initialized, the semi-supervised autoencoder uses stochastic gradient descent SGD for optimization.

[0039] Preferably, in the first stage of the D-FNN algorithm, a random walk and restart algorithm is run to learn the global structure of a single biological network, and then the semi-auto encoder is trained to learn low-dimensional features, and prior constraints are extracted based on the network representation of the hidden layer, and finally gene function prediction is performed; in the feature learning stage, the D-FNN algorithm uses an iterative model to train a semi-supervised autoencoder with prior constraints. In each iteration, the D-FNN mainly includes three steps: merging constraints, training semi-auto encoders, and extracting new constraints. As the number of iterations increases, the model tends to converge and the constraints tend to remain unchanged. The D-FNN finally generates several low-dimensional feature representations of the nodes. These features can reflect the importance of the vertices in the graph structure. The D-FNN algorithm is an extensible framework model, and its training complexity is linearly related to the number of vertices N. The constraint extraction part requires the calculation of paired PCC values, which requires O(N 2 ), the training complexity of the D-FNN algorithm is O((N 2 +N)TK), where T is the number of iterations and K is the number of multiple networks.

[0040] The present invention also provides an application of a synthetic lethal gene prediction method based on a random walk algorithm, which includes three stages:

[0041] (1) Feature generation: The main task is to extract the characterization of gene-gene interactions from existing data, especially the characteristics of genes at the molecular level. These characteristics often reveal the potential characteristics of genes, and to perform low-dimensional processing on these characteristics to facilitate subsequent work.

[0042] (2) Model construction: A prediction model needs to be established based on the gene relationship feature matrix of the previous stage;

[0043] (3) Gene prediction: Use the prediction model to make corresponding predictions on the functions of related synthetic lethal genes.

[0044] The present invention provides a synthetic lethal gene prediction method based on a random walk algorithm and its application, which have the following beneficial effects.

[0045] 1. The present invention is dedicated to the task of predicting the function of synthetic lethal genes in the context of cancer. Cancer-related genes are used as research objects. By establishing a prediction model, the task of predicting the function of synthetic lethal genes can be performed, and a feedforward neural network is used to learn feature embedding.

[0046] 2. The present invention is developed based on neural network technology and can complete the prediction of synthetic lethal genes from more complex gene pairs, effectively improving the accuracy of synthetic lethal gene prediction.

[0047] 3. The present invention adopts a semi-supervised automatic encoder, which can reduce the dimensionality of high-dimensional input and reduce the occurrence of overfitting. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 A working principle diagram of a synthetic lethal gene prediction method based on a random walk algorithm provided by the present invention. DETAILED DESCRIPTION

[0049] The present invention is further described below in conjunction with specific embodiments and drawings to help understand the content of the present invention.

[0050] The present invention provides a method for predicting synthetic lethal genes based on a random walk algorithm, comprising the following steps:

[0051] S1: Learning the global structure of a single network based on random walks. The specific method is:

[0052] First, we need to quote RWR (Random Walk with Restart), which is a special random walk algorithm that allows the walker to return to the starting node with a certain probability at each step. This restart mechanism makes the walk process pay more attention to the starting node and its surrounding areas, thereby better capturing the local structural information in the network.

[0053] Formally, the RWR from node i is defined as:

[0054]

[0055] Where P r is the probability of restart, which controls the relative influence of local and global topological information in the diffusion process. The higher the chance of restart, the more emphasis is placed on local structure; e i is an n-dimensional distribution vector, is an n-dimensional distribution vector, where each entry contains the probability of a node being visited after t steps of random walk starting from node i;

[0056] Based on the above method, RWR is run on each network to capture the individual network topology information and convert it into feature representation of nodes. The adjacency matrix only describes the relationship between any directly connected nodes, while ignoring the global structure of the network. RWR can overcome this shortcoming and use these high-dimensional network structure information to represent nodes. In addition, the present invention chooses the RWR method instead of other recently proposed network embedding methods to capture topology information, because these state-of-the-art algorithms are computationally intensive and require additional hyperparameter fitting.

[0057] Let M k Represents a k-layer network G k =(V, E (K) ), the adjacency matrix of node v i The RWR is described as the following recursive relation:

[0058]

[0059] Where a is the restart probability; e i is an n-dimensional initial eigenvector; is the n-dimensional feature vector of gene i, each item represents the probability of a gene being visited after t steps of random walk; T is the transition probability matrix, each item T ij Save gene j to gene v i The probability of can be calculated as Based on RWR, we get a matrix S, where S ij is node v i Compared with v defined based on the steady-state probability of RWR j The correlation score between .

[0060] S2: Extract prior constraints. The specific method is:

[0061] The concept of constraint comes from semi-supervised clustering. Pairwise constraints are formatted as must-connect constraints and no-connect constraints. Pairwise constraints can be described as follows: must-connect constraints indicate that these nodes are highly similar or belong to the same cluster, while no-connect constraints indicate that the two points in this pair of nodes are highly dissimilar or belong to different clusters.

[0062] The Pearson correlation coefficient (PCC) value is calculated to measure the pairwise similarity between gene nodes. Given two gene feature vectors A and B, the PCC values ​​of these two genes are:

[0063]

[0064] After calculating the PCC values ​​of all gene pairs, two strategies are used to extract constraints. One is to calculate and sort the paired Pearson correlation coefficient (PCC) based on the feature vectors of all node pairs. The numerator of the formula is considered to be a must-connect constraint, and the denominator is considered to be a no-connect constraint. The other is to set two thresholds for must-connect and no-connect, marked as f1 and f2, respectively. When the PCC value is greater than f1, a pair can be used as a must-connect constraint; when the PCC value is less than f2, a pair can be used as a no-connect constraint.

[0065] After extracting the constraints of the previous layer (layer i), the must-connect constraints and the do-not-connect constraints are applied to the next layer (layer i+1) as prior information.

[0066] S3: Integrate constraints using a semi-supervised autoencoder by:

[0067] The key problem of D-FNN is how to integrate prior constraints into the network representation through autoencoders. The present invention modifies the original autoencoder and proposes a new autoencoder, called semi-supervised autoencoder (semiAE). Starting from the second layer, the input includes the low-dimensional representation and the constraints of the previous layer. It is worth noting that the constraints from the building blocks of the previous layer are based on different networks. Therefore, the constraints may conflict with each other. To solve this problem, the present invention will merge these constraints and use the intersection of all constraints as the input of semiAE.

[0068] An autoencoder is a classic unsupervised model that aims to learn an effective representation of data and achieve dimensionality reduction during encoding and decoding. It consists of two parts: an encoder and a decoder. The encoder converts the original high-dimensional input data into a lower-dimensional feature representation to retain the important information of the original data and filter out useless information; the decoder recovers the representation of the original data from the low-dimensional feature vector and then uses the low-dimensional code as a compressed representation of the original data. Let x i is the i-th input vector or node representation of the network, f and g are the activations of the hidden layer and output layer respectively. i =f(Wx i +b) and y i =g(Mh i +d), where Θ = (θ1, θ2) = {W, b, M, d} are the parameters to be learned, and f and g are nonlinear operators such as the simoid function (sigmoid (z) = 1 / (1 + exp (-z))). Then the optimization goal is to make the original data x i and reconstructed data y i The reconstruction error is minimal.

[0069]

[0070] The original autoencoder cannot model the constraints obtained from the previous layer. SemiAE takes these limitations into account. Let M be a set of pairwise must-link constraints, where (x i , x j )∈M means x i and x j There is a strong connection between them. Let C be a set of pairwise cannot-link constraints, where (x i , x j )∈C means x i and x j Not relevant. The number of constraints is much smaller than the size of the network |M|+|C|≤|S|.

[0071] If x i and x jIf there are must-link constraints between the previous layers, they should also be closed based on the low-dimensional space. In fact, after encoding, two must-link nodes should be closer, while cannot-link nodes may be farther away. Mathematically, let d(h(x i ); h(x j )) is the x in the encoding space i and x j The error score between must-link, d(x i ;x i ) should be greater than d(h(x i ); h(x j )); for cannot-link, d(x i ;x j ) should be less than d(h(x i ); h(x j )); if (x i ;x j ) is a must-link constraint, adding a penalty to the loss function if (x i ;x j ) is a cannot-link constraint, then a reward is added to the loss function, and the loss function of the modeling constraint is defined as follows:

[0072]

[0073] Among them, the matrices M and C are the must-link constraint matrix and cannot-link constraint matrix extracted from the previous layer respectively; h(x i ) and h(x j ) is the input gene representation x j and x j The hidden layer representation of .

[0074] In order to combine constraints with autoencoders, the present invention proposes a new semi-supervised autoencoder, which integrates equations (4-1) and (4-2) and jointly minimizes the following objective function.

[0075]

[0076] The first part of formula (4-3) measures the squared error of the input and output node features, and the second part measures the error score of the hidden layer constraints.

[0077] like Figure 1, which is a working principle diagram of a synthetic lethal gene prediction method based on a random walk algorithm provided by the present invention. FNN (Feedforward Neural Network) is a feedforward neural network. Its main framework consists of 3 Dense layers and 2 Dropout layers. The first layer is a fully connected input layer containing 200 neurons, and the activation function is ReLU. This layer reduces the 625-dimensional input to 200 dimensions; the second layer is a Dropout layer, which randomly discards 40% of the neurons to reduce the occurrence of overfitting; the third layer is a fully connected hidden layer, which contains a fully connected layer of 180 neurons, and the activation function is ReLU. This layer extracts features and transforms dimensions from the 200-dimensional input; the fourth layer is a Dropout layer, which randomly discards 35% of the neurons to reduce the occurrence of overfitting; the fifth layer is a fully connected output layer containing 200 neurons, and the activation function is sigmoid. This layer transforms the 180-dimensional input to 200 dimensions and outputs the probability value of binary classification (0 or 1).

[0078] The core of the D-FNN algorithm is the semi-supervised autoencoder. The goal of this model is to minimize the loss function (4-3). In this section, we will calculate The partial derivative of is used to verify the above model. Therefore, the integral constraint loss function (4-2) can be rewritten as:

[0079]

[0080] Where L M =D M -M, D M is a diagonal matrix, L C and L M Similarly, H is a simplified representation of the hidden layer. So It can be calculated as:

[0081]

[0082] In formula (4-5), f is the activation function (i.e., sigmoid function or ReLu function), and L is obtained MC Therefore, the proposed semi-supervised autoencoder can be optimized using stochastic gradient descent (SGD) when the parameters are randomly initialized.

[0083] Summary of the method and algorithm of the present invention:

[0084] In the first stage of the D-FNN algorithm, a random walk and restart algorithm is run to learn the global structure of a single biological network. Then, the semi-autoencoder is trained to learn low-dimensional features, and prior constraints are extracted based on the network representation of the hidden layer. Finally, gene function prediction is performed. In the feature learning stage, the D-FNN algorithm uses an iterative model to train a semi-supervised autoencoder with prior constraints. In each iteration, D-FNN mainly includes three steps: merging constraints, training semi-autoencoders, and extracting new constraints. As the number of iterations increases, the model tends to converge and the constraints tend to remain unchanged. D-FNN ultimately generates several low-dimensional feature representations of the nodes, which can reflect the importance of the vertices in the graph structure. The D-FNN algorithm is an extensible framework model whose training complexity is linearly related to the number of vertices N. The constraint extraction part requires the calculation of pairwise PCC values, which requires O(N 2 ). Therefore, the training complexity of the D-FNN algorithm is O((N 2 +N)TK), where T is the number of iterations and K is the number of multiple networks. The D-FNN algorithm is shown in Table 1.

[0085] Table 1 D-FNN algorithm (pseudo code):

[0086]

[0087]

[0088] The present invention provides an application of a synthetic lethal gene prediction method based on a random walk algorithm, which includes three stages:

[0089] (1) Feature generation: The main task is to extract the characterization of gene-gene interactions from existing data, especially the characteristics of genes at the molecular level. These characteristics often reveal the potential characteristics of genes, and to perform low-dimensional processing on these characteristics to facilitate subsequent work.

[0090] (2) Model construction: A prediction model needs to be established based on the gene relationship feature matrix of the previous stage;

[0091] (3) Gene prediction: Use the prediction model to make corresponding predictions on the functions of related synthetic lethal genes.

[0092] The method of the present invention is compared with the Mashup method to evaluate the performance in the task of annotating gene functions. Function prediction can be regarded as a multi-label classification task. Therefore, ACC, F1, and AUPR are used as evaluation indicators. 20% of the entire labeled genes are randomly selected as the test set, and the remaining 80% are used as the training set to annotate the unlabeled gene functions. Each method is repeated 5 times, and the average value is taken as the final experimental result. The experimental results are shown in Table 2.

[0093] Table 2 Experimental results

[0094]

[0095]

[0096] As can be seen from the table, the ACC and F1 values ​​of the new method are higher than those of the Mashup model, but the AUPR value is lower than that of the Mashup model, indicating that the new method has a higher classification accuracy overall, but the performance in the positive category is relatively poor. This may be due to the small number of positive samples in the data set, the imbalance of categories, or the existence of some false positives in the model in the positive category.

[0097] The present invention is dedicated to the prediction task of synthetic lethal gene function in the context of cancer. It takes cancer-related genes as the research object and can perform the task of predicting synthetic lethal gene function by establishing a prediction model. It uses a feedforward neural network to learn feature embedding and is developed based on neural network technology. It can complete the prediction of synthetic lethal genes from more complex gene pairs, effectively improving the accuracy of synthetic lethal gene prediction. It uses a semi-supervised autoencoder to reduce the dimensionality of high-dimensional inputs and reduce the occurrence of overfitting. It has broad application prospects and is conducive to popularization and application.

[0098] The invention concept is described in detail using specific examples herein, and the description of the above embodiments is only used to help understand the core idea of ​​the invention. It should be pointed out that any obvious modification, equivalent substitution or other improvement made by a person of ordinary skill in the art without departing from the invention concept should be included in the protection scope of the present invention.

Claims

1. A method for predicting synthetic lethal genes based on a random walk algorithm, characterized in that: The steps include: S1: Learning the global structure of a single network based on random walks; S2: Extract prior constraints; S3: Integrating constraints using semi-supervised autoencoders.

2. The method for predicting synthetic lethal genes based on a random walk algorithm according to claim 1, characterized in that: The specific method of S1 is: Formally, the RWR from node i is defined as: Where P r is the probability of restart, which controls the relative influence of local and global topological information in the diffusion process. The higher the chance of restart, the more emphasis is placed on local structure; e i is an n-dimensional distribution vector, is an n-dimensional distribution vector, where each entry contains the probability of a node being visited after t steps of random walk starting from node i; Based on the above method, RWR is run on each network to capture the individual network topology information and convert it into feature representation of nodes; Let M k Represents a k-layer network G k =(V, E (K) ), the adjacency matrix of node v i The RWR is described as the following recursive relation: Where a is the restart probability; e i is an n-dimensional initial eigenvector; is the n-dimensional feature vector of gene i, each item represents the probability of a gene being visited after t steps of random walk; T is the transition probability matrix, each item T ij Save gene j to gene v i The probability of can be calculated as Based on RWR, we get a matrix S, where S ij is node v i Compared with v defined based on the steady-state probability of RWR j The correlation score between .

3. The method for predicting synthetic lethal genes based on a random walk algorithm according to claim 2, characterized in that: The specific method of S2 is: The pairwise constraints are formatted as a must-connect constraint and a no-connect constraint. The must-connect constraint indicates that these nodes are highly similar or belong to the same cluster, and the no-connect constraint indicates that the two points in the pair of nodes are highly dissimilar or belong to different clusters. The Pearson correlation coefficient PCC value is calculated to measure the pairwise similarity between gene nodes. Given two gene feature vectors A and B, the PCC values ​​of these two genes are: After calculating the PCC values ​​of all gene pairs, two strategies are used to extract constraints. One is to calculate and sort the paired Pearson correlation coefficients PCC based on the feature vectors of all node pairs. The numerator of the formula is considered as a must-connect constraint, and the denominator is considered as a no-connect constraint. The other is to set two thresholds for must-connect and do not-connect, marked as f1 and f2 respectively. When the PCC value is greater than f1, one pair is used as must-connect constraints; when the PCC value is less than f2, one pair is used as do not-connect constraints. After extracting the constraints of the previous layer, i.e., layer i, the must-connect constraints and the do-not-connect constraints are applied to the next layer, i.e., layer i+1, as prior information.

4. The method for predicting synthetic lethal genes based on a random walk algorithm according to claim 3, characterized in that: The specific method of S3 is: Using a semi-supervised autoencoder semiAE, starting from the second layer, the input includes the low-dimensional representation and the constraints of the previous layer, merge these constraints, and use the intersection of all constraints as the input of semiAE. Let x i is the i-th input vector or node representation of the network, f and g are the activations of the hidden layer and output layer respectively, and i =g(Mh i +d), where Θ = (θ1, θ2) = {W, b, M, d} are the parameters to be learned, f and g are nonlinear operators such as the simoid function (sigmoid (z) = 1 / (1 + exp (-z))), and the optimization goal is to make the original data x i and reconstructed data y i The reconstruction error between them is minimal; Let M be a set of pairwise must-link constraints, where (x i , x j )∈M means x i and x j There is a strong connection between them. Let C be a pair of cannot-link constraints, where (x i ;x j )∈C means x i and x j Irrelevant, the number of constraints is much smaller than the size of the network |M|+|C|≤|S|; If x i and x j If there are must-link constraints between the previous layers, they should also be closed based on the low-dimensional space, and the two must-link nodes after encoding are closer, while the cannot-link nodes are farther away; mathematically, let d(h(x i ); h(x j )) is the x in the encoding space i and x j The error score between ; for must-link, d(x i ;x j ) should be greater than d(h(x i ); h(x j )); for cannot-link, d(x i ;x j ) should be less than d(h(x i ); h(x j )); if (x i , x j ) is a must-link constraint, adding a penalty to the loss function if (x i ;x j ) is a cannot-link constraint, then a reward is added to the loss function, and the loss function of the modeling constraint is defined as follows: Among them, the matrices M and C are the must-link constraint matrix and cannot-link constraint matrix extracted from the previous layer respectively; h(x i ) and h(x j ) is the input gene representation x j and x j The hidden layer representation of The semi-supervised autoencoder integrates equation (4-1) and equation (4-2) and jointly minimizes the following objective function: The first part of formula (4-3) measures the squared error of the input and output node features, and the second part measures the error score of the hidden layer constraints; FNN is a feedforward neural network, and its main framework consists of 3 Dense layers and 2 Dropout layers. The first layer is a fully connected input layer containing 200 neurons, and the activation function is ReLU. This layer reduces the 625-dimensional input to 200 dimensions; the second layer is a Dropout layer, which randomly discards 40% of the neurons to reduce the occurrence of overfitting; the third layer is a fully connected hidden layer, which contains a fully connected layer of 180 neurons, and the activation function is ReLU. This layer extracts features and transforms dimensions from the 200-dimensional input; the fourth layer is a Dropout layer, which randomly discards 35% of the neurons to reduce the occurrence of overfitting; the fifth layer is a fully connected output layer containing 200 neurons, and the activation function is sigmoid. This layer transforms the 180-dimensional input to 200 dimensions and outputs the probability value of the binary classification 0 or 1; The core of the D-FNN algorithm is the semi-supervised autoencoder. The goal of this model is to minimize the loss function (4-3). In this section, we will calculate The partial derivative of is used to verify the above model. The integral constraint loss function (4-2) is: Where L M =D M -M, D M is a diagonal matrix, L C and L M Similarly, H is a simplified representation of the hidden layer, Calculated as: In formula (4-5), f is the activation function, that is, the sigmoid function or the ReLu function, and L is obtained. MC The partial derivatives of , when the parameters are randomly initialized, the semi-supervised autoencoder uses stochastic gradient descent SGD for optimization.

5. The method for predicting synthetic lethal genes based on random walk algorithm according to claim 4, characterized in that: In the first stage of the D-FNN algorithm, a random walk and restart algorithm is run to learn the global structure of a single biological network, and then a semi-auto encoder is trained to learn low-dimensional features and extract prior constraints based on the network representation of the hidden layer, and finally gene function prediction is performed; In the feature learning stage, the D-FNN algorithm uses an iterative model to train a semi-supervised autoencoder with prior constraints. In each iteration, D-FNN mainly includes three steps: merging constraints, training semi-autoencoders, and extracting new constraints. As the number of iterations increases, the model tends to converge and the constraints tend to remain unchanged. D-FNN finally generates several low-dimensional feature representations of the nodes. These features can reflect the importance of the vertices in the graph structure. The D-FNN algorithm is an extensible framework model, and its training complexity is linearly related to the number of vertices N. The constraint extraction part requires calculating the paired PCC values, which requires O(N 2 ), the training complexity of the D-FNN algorithm is O((N 2 +N)TK), where T is the number of iterations and K is the number of multiple networks.

6. An application of the synthetic lethal gene prediction method based on random walk algorithm according to any one of claims 1 to 5, characterized in that: It consists of three stages: (1) Feature generation: The main task is to extract the characterization of gene-gene interactions from existing data, especially the characteristics of genes at the molecular level. These characteristics often reveal the potential characteristics of genes, and to perform low-dimensional processing on these characteristics to facilitate subsequent work. (2) Model construction: A prediction model needs to be established based on the gene relationship feature matrix of the previous stage; (3) Gene prediction: Use the prediction model to make corresponding predictions on the functions of related synthetic lethal genes.

Images