Biological network method for cancer prediction and oncogene attribution analysis based on deep learning

Abstract

The invention provides a biological network method for cancer prediction and oncogene attribution analysis based on deep learning. The method comprises the following steps: firstly, collecting transcriptome data, a potential oncogene set and a pathway database, and obtaining a sample expression matrix, a pathway network and a data set by using the biological data; constructing an integrated biological model by using a sample expression matrix neutralization pathway network; training the model by adopting a knowledge trainable mode; patient transcriptome data is input into the model to obtain the cancer incidence probability of the patient, and oncogenes affecting cancer can be attributed and analyzed. The method relates to the field of medical and industrial combination crossing, and is mainly applied to the fields of early screening and early warning of cancers and discovery of potential new targets and treatment strategies of cancers by attribution analysis. The method has the following advantages that firstly, biological knowledge is fused, the sparse model calculation amount is small, and the training iteration speed is high; secondly, knowledge can be trained, the model prediction performance is improved, and downstream explanation tasks are assisted; and 3, the decision of the explaining algorithm explaining model.

Description

Technical Field

[0001] The present invention belongs to the cross-field of medicine and engineering that combines computers and medicine, and mainly involves artificial intelligence deep learning, cancer prediction and gene attribution analysis. Background Art

[0002] Traditional fields: In medicine, high-throughput sequencing technology is often used to obtain a large amount of transcriptome data. These data can be processed through differential expression analysis, cluster analysis and other methods to find genes that are significantly different between tumor tissues and normal tissues. These genes are considered to be potential factors involved in tumorigenesis. Next, through functional annotation and pathway analysis of these differentially expressed genes, we can further understand the biological functions of potential genes and related signal pathways. Through methods such as gene network analysis and gene interaction network analysis, we can explore the interaction relationship between these genes, thereby discovering deeper biological mechanisms and potential therapeutic targets, identifying specific tumor subtypes, predicting patient survival and treatment response, and discovering new targets and treatment strategies.

[0003] Emerging technologies: With the rise of artificial intelligence, especially deep learning, in recent years, the medical field has also begun to gradually apply these technologies to process massive amounts of medical data. These technologies can utilize the potential feature relationships between massive amounts of data, making a huge contribution to the application and research in the medical field. "Visible" neural networks have been developed to simulate the effects of gene interactions on yeast cell growth and to use path-related sparse neural networks to solve the prognosis of patients with glioblastoma multiforme; the AlphaFold model predicts the structure of unknown proteins by learning a large amount of known protein structure data. Its prediction accuracy is very high, and it can even predict some protein structures that have not been solved so far. The emergence of this model has set off a revolution in the field of protein structure prediction, bringing new opportunities and challenges to fields such as biology and drug development. Its highly accurate prediction results can not only help scientists better understand the structure and function of proteins, but also provide new ideas and methods for drug development; there is also Pnet, which is based on biological data modeling and training to predict the primary and metastatic nature of prostate tumors. Surprisingly, the results of its attribution analysis not only conform to reality, but also discover a new target gene MDM4, which has been confirmed in clinical trials. The method proposed in the present invention not only increases the trainable knowledge on the basis of the former, obtains more accurate prediction performance and is more conducive to downstream interpretation tasks, but also proposes the concept of gene embedding, which contains gene prior knowledge, which can be sent into the model together with gene expression information to enhance performance. In addition, mathematical methods can be used to explore the similarities between genes based on gene embedding information alone, and there is still a lot of room for exploration.

[0004] Although current deep learning techniques can establish strong links between biological signals and measured phenotypes, these methods are limited in identifying the causal structure of neural networks. With the application of "AI+medical care", the requirements for the interpretability and visualization of related learning models are becoming increasingly higher. Therefore, developing new methods to improve the interpretability and visualization of deep learning models in order to better understand how these models make predictions and decisions and better explain these results is an urgent problem to be solved. Summary of the invention

[0005] The goal of this invention is to achieve cancer prediction and early warning at the molecular level, promote early screening and treatment of tumors and precise diagnosis and treatment, and discover potential oncogenes through attribution analysis to obtain potential new targets and new treatment strategies.

[0006] The present invention proposes a biological network method based on deep learning that can be used for cancer prediction and oncogene attribution analysis. By integrating biological knowledge into modeling and adopting a knowledge trainable method for training, cancer incidence prediction results and attribution analysis results are obtained. The overall process steps are as shown in Figure 1, and the method specifically includes the following steps.

[0007] 1. Collect data:

[0008] 1.1) Collect the transcriptome sequencing data of cancer patients and the transcriptome sequencing data of patients to be analyzed, and convert the transcriptome sequencing data of the corresponding cancer patients into an expression matrix through preprocessing Each element Indicates gene g j In samples i The expression of , reflects the real biochemical reaction and molecular level in the patient;

[0009] 1.2) Pathway datasets are divided into pathway datasets and the Pathways and Genes dataset G includes the set of path nodes and the edge set ε, Indicates that there is a path from To the access road Each gene in the gene set g i The channel That means g i Participated in the pathway biochemical reactions;

[0010] 1.3) The candidate oncogene set is a gene set that is most likely to cause cancer and is selected using prior knowledge;

[0011] 1.4) Gene embedding information: using prior biological knowledge to obtain the representation of genes in vector space. Prior biological knowledge can be factual knowledge such as gene interaction networks, which contains rich gene information.

[0012] 2. Establish biological model:

[0013] 2.1) Using the expression matrix and candidate gene expression set X candidate Take the intersection to obtain the expression matrix containing only the candidate genes, and delete the values ​​of unnecessary genes in the expression matrix;

[0014] 2.2) Represent the pathway and its database as a directed graph, and find all nodes with in-degree 0 as representations of higher-order biological processes. Create a new node root and connect it to all nodes with in-degree 0. Traverse the search tree starting from the root node using a breadth-first search algorithm, retaining only nodes with a height of the first n layers, and deleting nodes with a height greater than n. At the same time, add new nodes in such a way that all leaf nodes are located at the nth layer.

[0015] 2.3) Using the search tree with a height of n obtained in the previous step and the gene data set, connect the genes in the candidate gene expression matrix to the leaf nodes (pathway nodes) of the search tree. Specifically, connect the gene nodes and the pathway nodes according to the display relationship, which is equivalent to adding a new layer of nodes (gene nodes) under the leaf nodes (pathway nodes) of the original search tree with a height of n, forming a search tree with a height of n+1. The n+1th layer is the gene node, from the nth layer to the second layer are the pathway nodes, and the first layer is the virtual root node.

[0016] 2.4) The architecture of the model includes:

[0017] a. A search tree with a height of n+1 is used as the basic structure of the model;

[0018] b. Add a mask matrix to the data flow from each layer to the next layer to ensure that the direction of data flow follows the rules of action of genes and pathways;

[0019] c. From the n+1th layer to the 3rd layer, each layer adds a fully connected layer and activation function y=

[0020] sigmoid(W T x+b), used to represent the prediction of each layer separately;

[0021] d. Take the weighted sum of the outputs of the fully connected layer and the root node to get the final prediction result;

[0022] 2.5) First, the gene expression matrix is ​​classified by samples, each sample contains the expression information of thousands to tens of thousands of genes, and then the gene embedding information containing prior knowledge is combined with the embedding containing gene expression information to obtain new samples, and stratified sampling is performed to divide the candidate gene expression matrix into training set, test set and validation set with a quantity ratio of 8:1:1 according to the samples.

[0023] 2.6) Use Adam optimizer, binary cross entropy loss function, and linear decreasing learning rate to train the model for downstream cancer prediction and oncogene attribution analysis.

[0024] 3. Attribution analysis of oncogenes and cancer incidence:

[0025] 3.1) The patient's transcriptome sequencing data x is fed into the model, and the model's prediction f can be calculated. w (), which is the predicted value of cancer incidence.

[0026] 3.2) Oncogenes attributable to analysis include:

[0027] a. Input the patient’s transcriptome sequencing data x into the model and get the output f w ();

[0028] b. Input the standard reference into x 0 Input the model separately and get the standard output f w (x 0 );

[0029] c. Calculate the difference between the two outputs, i.e. f w ()-f w ( 0 );

[0030] d. Use the back propagation algorithm to calculate the contribution of each neuron (gene node) in the n+1th layer of the model to the output difference value;

[0031] e. The contribution of each gene to the predicted value, that is, the contribution of the gene to cancer;

[0032] The characteristics and beneficial effects of the present invention are:

[0033] Integrating biological knowledge into deep learning neural networks can, firstly, reduce the number of network structure parameters by using biological prior knowledge; secondly, use knowledge training to improve the prediction performance of the model; and thirdly, use interpretable algorithms to explain the contribution of genes to prediction and infer oncogenes, which has practical significance for discovering new targets and new treatment strategies. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 is the overall flow chart.

[0035] Figure 2 is a diagram of a neural network model. DETAILED DESCRIPTION

[0036] The present invention proposes a biological network method based on deep learning that can be used for cancer prediction and oncogene attribution analysis, which is further described in detail below in conjunction with the accompanying drawings and specific embodiments.

[0037] 1. Collect data, including the following steps:

[0038] 1.1) Transcriptome sequencing data: Collect the transcriptome sequencing data of the corresponding cancer patients and the transcriptome sequencing data of the patients to be analyzed. The transcriptome sequencing data contains various gene expression conditions, reflecting the results and molecular levels of the real biochemical reactions in the patient's body. First, pre-process the transcriptome sequencing data of the corresponding cancer patients into an expression matrix Where s represents the number of samples, g represents the number of genes measured in one sample, and the values ​​in the matrix Indicates gene g j In samples i The transcriptional sequencing data of the patient to be analyzed is x = [v g1 ,v g2 ,...,v gn ], including the patient's own gene expression profile.

[0039] 1.2) Pathway Dataset: The pathway dataset can be represented as two parts of data. The first part is the pathway and the pathway dataset. represents the set of path nodes, ε represents the set of edges, Indicates that there is a path from To the access road The directed edge of .

[0040] The second data set is the pathway and gene dataset Indicates the pathway to which a gene belongs, This indicates that gene g i Participated in this pathway 1.3) Candidate oncogene set: The candidate oncogene set is a gene set X candidate =

[0041] {g 1 ,g 2 ,...,g n}, the gene set most likely to cause cancer selected using prior knowledge.

[0042] 1.4) Gene embedding information: Gene embedding information X gene =[v 1 ,v 2 ,...,v n] That is, the representation of genes in vector space, which is similar to the word embedding in natural language processing that contains word information, and gene embedding that contains gene information.

[0043] 2. Establish a biological model, including the following steps:

[0044] 2.1) Preprocessing the transcriptome data to obtain an expression matrix containing only candidate genes: Using the expression matrix and candidate gene expression set X candidate Construct an expression matrix containing only candidate genes That is, the gene set in the expression matrix is ​​intersected with the candidate gene set, and only the values ​​of these genes in the matrix are retained. The values ​​of other unnecessary genes are deleted to obtain an expression matrix containing only candidate genes.

[0045] 2.2) Pathway database builds pathway network: The pathway and pathway database itself is a directed graph G. Find all nodes with in-degree 0 in the graph. They are high-order biological processes and contain rich biological information. Create a new node root and connect the root node to all nodes with in-degree 0. Now you can traverse all pathways from the root node. Use the breadth-first search algorithm to traverse from the root node to get the search tree. Only the first n layers of the search tree are retained, and nodes with a height exceeding n are deleted. At the same time, some leaf nodes may have a height less than n. Add nodes to make all leaf nodes located at the nth layer.

[0046] 2.3) Genes in pathway networks and gene expression matrices Obtain a network structure containing genes and pathways: Based on the search tree with a layer height of n obtained in the previous step, use the pathway and gene dataset p = {g 1 ,g 2 ,...,g n}, connect the genes in the candidate gene expression matrix to the leaf nodes (pathway nodes) of the search tree, and the connection relationship is expressed as g i →p j ,g i ∈X candidate , g i ∈p j , which is equivalent to adding a new layer of nodes (gene nodes) under the leaf nodes (pathway nodes) of the original n-layer search tree to form a search tree with a height of n+1. The n+1th layer is all gene nodes, from the nth layer to the second layer are path nodes, and the first layer is a virtual root node.

[0047] 2.4) Construct a sparse neural network model based on the network structure: Based on the n+1 layer search tree in the previous step, construct a neural network model with the same structure, and add a mask matrix to the data flow from each layer to the next layer of the network model. The specific formula can be expressed as y=f((Mask*W) T x+b), so that the direction of data flow in the model follows the rules of action of genes and pathways in the real world. From the n+1th layer to the third layer, each layer adds a fully connected layer and an activation function y=sigmoid(W T x+b), which represents the prediction of each layer separately. Finally, the output of the fully connected layer and the root node is weighted and summed to get the final value, which is the predicted value of cancer incidence. This is the architecture of the model. The final structure is shown in Figure 2. n is a 5-layer search tree. H1 to H5 are the pathway layers composed of pathway nodes. H0 is the gene layer composed of gene nodes. According to the actual relationship, the connection between H0-H5 obeys the actual relationship, so they are all sparsely connected and controlled by mask matrices. L0 to L5 are the fully connected layers of each layer from H0 to H5, representing the individual prediction output of each layer. Finally, the weighted sum of the prediction outputs from L0 to L5 is the prediction output of the entire model.

[0048] 2.5) Expression matrix to obtain training set: In the candidate gene expression matrix, for each gene, its gene embedding information is attached, so that any sample in the gene expression matrix will contain gene expression information and gene embedding information. The gene expression matrix is ​​divided into training set, test set and validation set according to the ratio of 8:1:1 in sample number. At the same time, the stratified sampling mechanism is used to make the proportion of different cancers in the training set, test set and validation set the same to avoid imbalance. The training set and validation set are used in the model training stage to train the model and evaluate whether the model hyperparameters are appropriate, while the test set is a data set for evaluating model performance.

[0049] 2.6) Use the gradient descent method to train the model in a knowledge-trainable way: In the process of training the model, a batch size of 50 is used, the Adam optimizer is used, the L2 norm is used to prevent overfitting, and the loss function uses binary cross entropy. The formula is as follows:

[0050]

[0051] Where W y Represents the weight coefficient. For the current prediction class y, the weight is used to amplify or reduce the loss in order to avoid the impact of sample imbalance. At the same time, the knowledge is trained in a way that can be trained, that is, during the training process, the labeling mask matrix can also be trained to capture more biological information. After the loss value is obtained, the weights of the parameters of the neural network can be updated through the gradient descent algorithm. The formula is as follows:

[0052]

[0053] where w i is the parameter that needs to be trained in the network, and α is the learning rate, which is used to control the speed of model learning. It is not good to be too large or too small. Here, the initial learning rate is set to 0.001. The model is trained for a total of 300 rounds, and the learning rate is multiplied by 0.25 every 50 rounds to obtain a well-trained model for downstream cancer prediction and oncogene attribution analysis.

[0054] 3. Attribution analysis of oncogenes and cancer incidence, including the following steps:

[0055] 3.1) Cancer incidence: The patient's transcriptome sequencing data x is fed into the model, and the model's prediction f can be calculated. w (), which is the predicted value of cancer incidence.

[0056] 3.2) Attribution analysis of oncogenes: Put the patient's transcriptome sequencing data x into the model and get the output f w () Then input the standard reference into x 0 Put them into the model respectively and get the standard output f w (x 0 ), then calculate the difference between the two outputs, and then use back propagation to calculate the contribution of each neuron (gene node) in the n+1th layer of the model. This method can be used to obtain the contribution of each gene to the predicted value, that is, the contribution of genes to cancer. This is also a new means of discovering potential targets and treatment strategies.

[0057] The method of the present invention can be implemented by programming by ordinary technicians in this field. Ordinary technicians in this field can understand that the above model method can be completed by a program, and the program can be stored in a computer-readable storage medium.

[0058] The above is only a specific embodiment of the present invention, which obviously cannot be used to limit the scope of the present invention. Therefore, equivalent changes made according to the claims of the present invention are still within the scope of the present invention.

Claims

1. A biological network method for cancer prediction and oncogene attribution analysis, Its characteristics include the following steps: 1.1) Collect biological data; 1.2) Use biological data to build universal biological models; 1.3) Analyze the incidence of cancer and oncogenes in patients.

2. The biological network method for cancer prediction and oncogene attribution analysis according to claim 1, It is characterized in that The collection of biological data specifically includes: 2.1) Biological data such as transcriptome sequencing of cancer patients; 2.2) Pathway databases, protein interaction networks and other data; 2.3) Candidate oncogene set; 2.4) Gene embedding information.

3. The biological network method for cancer prediction and oncogene attribution analysis according to claim 1, It is characterized in that The method of establishing a universal biological model using biological data specifically includes: 3.1) The collected biological data such as transcriptomes are combined with the candidate oncogene set to obtain an expression matrix containing candidate genes; 3.2) Use pathway database to construct graph-like pathway network; 3.3) Using pathway networks and protein interaction networks, we can obtain a network structure that contains both gene nodes and pathway nodes; 3.4) Use the above gene pathway network structure to build a neural network model with similar structure; 3.5) Obtaining a data set for model training based on samples in the gene expression matrix; 3.6) Put the training set into the neural network model and use the gradient descent algorithm and knowledge trainable method to obtain a well-trained model that can be applied to biological data.

4. The biological network method for cancer prediction and oncogene attribution analysis according to claim 1, It is characterized in that The attribution analysis of the cancer incidence and oncogenes of patients specifically includes the following steps: 4.1) Put the patient's transcriptome data into the model, and let the model predict the patient's cancer incidence; 4.2) The transcriptome data is fed into the model, and the explainable attribution method is used to obtain the contribution of each gene to cancer, that is, the impact of oncogenes on cancer, and obtain potential new targets.

5. The step of collecting biological data according to claim 2, It is characterized in that The gene embedding information specifically comprises the following steps: 5.1) Factual biological knowledge can be obtained through deep learning methods to obtain gene embedding information, such as using protein interaction networks to obtain gene maps, and then extracting gene embedding information of gene nodes through map information; 5.2) The embedding information obtained by different methods can be fused to obtain the final gene embedding information.

6. The step of establishing a universal biological model using biological data according to claim 3, It is characterized in that The construction of the neural network model specifically includes the following steps: 6.1) Remove all nodes with in-degree 0 in the path network; 6.2) Abstract the root node root, and use the root node to point to all the nodes with in-degree 0 obtained in the previous step; 6.3) With the root node as the center, use the breadth-first search algorithm to obtain the search tree, retaining the nodes that can be reached within n steps, that is, only retaining the search tree with a layer height of n or less; 6.4) For leaf nodes with a layer height less than n layers, add nodes to extend them so that all leaf nodes are at the nth layer; 6.5) A pathway is a series of biochemical reactions, which also includes genes. The genes in the gene expression matrix are connected to the pathways at the nth level, that is, at the n+1th level of the search tree, based on the actual relationship, and the connection relationship between each node in the search tree and the next node is preserved in the form of an adjacency matrix; 6.6) Based on the above n+1-layer search tree structure, a neural network with a similar structure is constructed. A mask matrix is ​​also added between each layer of the neural network to participate in the operations between layers. Its value completely corresponds to the adjacency matrix, so that the flow of data in the network obeys the interaction between genes and pathways in the real world. In the actual model training, a knowledge-trainable method is adopted. The value of the mask matrix is ​​updated with training, and a fully connected layer is added from the n+1th layer to the 2nd layer (the root node is the 1st layer). The weighted sum of all fully connected layers is normalized and is the output of the model.

7. The step of attributing analysis of cancer incidence and oncogenes in patients according to claim 4, It is characterized in that The explainable attribution analysis specifically includes the following steps: 7.1) Collect patient transcriptome sequencing data and standard input for attribution analysis; 7.2) The model forward propagates the standard input and patient transcriptome sequencing data to obtain the standard output and cancer prediction value respectively; 7.3) The difference between the patient's cancer prediction value and the standard reference output is the difference value; 7.4) By backpropagating the difference value to the input layer of the model, which is the gene layer, we can find the contribution of each gene to the difference value, that is, the contribution of the gene to the incidence of cancer.

Images