CN-122024820-A - Global optimization method for solving missing observed values in nonlinear regulation network

CN122024820ACN 122024820 ACN122024820 ACN 122024820ACN-122024820-A

Abstract

The invention relates to the technical field of biological information, and discloses a global optimization method for solving a missing observed value in a nonlinear regulation network, which can be used for pharmacological networks, physiological networks, toxicological networks and pathological networks and comprises the steps of acquiring an experimental data set; the method comprises the steps of constructing a nonlinear regulation network model based on a deep learning model according to a target nonlinear regulation network, inputting an experimental data set into the network model, initializing weights and deviations by adopting a default initialization method, carrying out unconstrained training on the initialized network model, taking an average value to obtain a constraint value, inputting the constraint value into a parameter updating method, carrying out constrained training on the initialized network model by adopting a parameter updating method and a parameter constraint method to obtain globally optimized network model parameters, and determining quantitative regulation information of the target nonlinear regulation network according to the globally optimized network model parameters.

Inventors

ZHENG GUANG
LI ZHICHENG
JIA FENG

Assignees

兰州大学

Dates

Publication Date: 20260512
Application Date: 20260130

Claims (10)

1. A global optimization method for solving a missing observed value in a nonlinear regulation network is characterized by comprising the following steps: Obtaining an experimental data set, wherein the experimental data set comprises biological components and expression levels of the biological components, and the biological components comprise pharmaceutical active components, target proteins, biomarkers and genes; Constructing a nonlinear regulation network model based on a deep learning model according to a target nonlinear regulation network; the nonlinear regulation network model comprises an input layer, a plurality of hidden layers and an output layer; Each network node of the input layer, the hidden layer and the output layer represents biological components, the value of each network node represents the expression quantity of the biological components, and the network links among the network nodes represent the interaction among the biological components; Inputting the experimental data set into a nonlinear regulation network model, and initializing weights and deviations by adopting a default initialization method; non-constraint training is carried out on the initialized nonlinear regulation network model, an average value is taken, a constraint value is obtained, and global optimization is achieved; Inputting the constraint value into a parameter updating method, and performing constraint training on the initialized nonlinear regulation network model by adopting the parameter updating method and a parameter constraint method to obtain a globally optimized network model parameter; And determining quantitative regulation information of the target nonlinear regulation network according to the globally optimized network model parameters, wherein the quantitative regulation information is used for predicting a regulation result according to given drug input.
2. The global optimization method for solving missing observations in a nonlinear regulation network according to claim 1, wherein the method for obtaining the experimental data set comprises: Extracting nonlinear regulation network original data from literature experiment data; and expanding the original data of the nonlinear regulation network by adopting a method based on normal distribution or uniform distribution to generate an experimental data set.
3. The global optimization method for solving missing observations in a nonlinear regulation network according to claim 1, wherein a nonlinear regulation network model based on a deep learning model is constructed according to a target nonlinear regulation network, and specifically comprises the following steps: determining an input layer, a plurality of hidden layers and an output layer according to the action sequence of each biological component in the target nonlinear regulation network; The input layer represents the drug input of the target nonlinear regulation network, the hidden layer represents the middle regulation part of the target nonlinear regulation network, and the output layer represents the regulation result of the target nonlinear regulation network; Judging whether direct interaction exists between network nodes of non-adjacent layers according to a target nonlinear regulation network, and filling blank nodes when the direct interaction exists between the network nodes of the non-adjacent layers; Adding network links among all network nodes of an input layer, a hidden layer and an output layer, and setting an activation function, a loss function and an optimizer to obtain a nonlinear regulation network model; The activation function comprises at least one of a Sigmoid function, a Tanh function, a ReLU function, a LeakyReLU function and a Softplus function; the loss function is a mean square error loss function, and the optimizer is an Adam optimizer.
4. The global optimization method for solving the missing observed values in the nonlinear regulation network according to claim 1, wherein the parameter initialization method is adopted to adjust the positive and negative of the weights in the nonlinear regulation network model so that the weights meet the interaction relation among biological components, the interaction relation comprises an activation relation and a suppression relation, the interaction relation among the biological components is represented as the activation relation when the weights are positive values, and the interaction relation among the biological components is represented as the suppression relation when the weights are negative values.
5. The global optimization method for solving the missing observed values in the nonlinear regulation network according to claim 1, wherein the random value is used for carrying out unconstrained training on the initialized nonlinear regulation network model, taking an average value to obtain a constraint value, and realizing global optimization, and the method specifically comprises the following steps: training the initialized nonlinear regulation network model by random value, and respectively calculating the prediction results of all nodes of the output layer; Removing abnormal values in the prediction results of all nodes of the output layer; calculating arithmetic average value of the prediction results of all nodes of the output layer from which abnormal values are removed, obtaining constraint values, and constraining the output results of the nodes by the arithmetic average value to realize global optimization.
6. The global optimization method for solving missing observations in a nonlinear regulation network according to claim 5, wherein the method is characterized by removing abnormal values in the prediction results of all nodes of an output layer, and specifically comprises the steps of calculating a mean value of the prediction values between a first quartile and a third quartile, namely within a quartile range, according to a quartile structure defined by a box diagram so as to represent a central tendency of a distribution main section; And eliminating predicted values outside the distribution main body interval according to the judgment criterion of the polar values in the box diagram.
7. The global optimization method for solving missing observations in a nonlinear regulation network of claim 1, wherein, Adopting a parameter updating method, assigning the weight of the network link of the adjacent layer without interaction to be 0, assigning the weight of the network link of the input blank node to be a constraint value, and assigning the deviation corresponding to the blank node to be 0; Selecting one network node from a hidden layer of a nonlinear regulation network model by adopting a parameter constraint method, determining a target proportion according to a model calculation value and an original data value of the network node, and updating the deviation of the network node according to the target proportion so that the proportion of the model calculation value and the original data value of each network node of the hidden layer is unified as the target proportion; Determining a loss value according to the model calculation value and the original data value of the network node of the output layer, and extracting the weight and the deviation of the nonlinear regulation network model as the overall optimized network model parameters when the loss value is smaller than a set threshold value; Updating the deviation of the network node according to the target proportion, wherein the specific formula is as follows: ; Wherein, the For the updated bias of the network node, For the current deviation of the network node, As the original data value of the network node, Values are calculated for the model of the network node, Is the target ratio.
8. A global optimization system for solving missing observations in a nonlinear regulation network, comprising: The data acquisition module is used for acquiring an experimental data set, wherein the experimental data set comprises biological components and expression levels of the biological components, and the biological components comprise pharmaceutical active components, target proteins, biological markers and genes; The nonlinear regulation network model comprises an input layer, a plurality of hidden layers and an output layer, wherein each network node of the input layer, the hidden layers and the output layer represents the biological component, the value of each network node represents the expression quantity of the biological component, and the network links among the network nodes represent the interaction among the biological components; the model training module is used for inputting the experimental data set into the nonlinear regulation network model, and initializing the weight and the deviation by adopting a default initialization method; randomly taking values to perform non-constraint training on the initialized nonlinear regulation network model, taking an average value to obtain constraint values, and realizing global optimization; Inputting the constraint value into a parameter updating method, and performing constraint training on the initialized nonlinear regulation network model by adopting the parameter updating method and a parameter constraint method to obtain a globally optimized network model parameter; the information determining module is used for determining quantitative regulation information of the target nonlinear regulation network according to the network model parameters, and the quantitative regulation information is used for predicting a regulation result according to given drug input.
9. An electronic device comprising a memory for storing a computer program and a processor that runs the computer program to cause the electronic device to perform the global optimization method of solving missing observations in a nonlinear regulation network according to any one of claims 1 to 7.
10. A computer-readable storage medium, characterized in that it stores a computer program which, when executed by a processor, implements a global optimization method for solving missing observations in a nonlinear regulation network according to any one of claims 1 to 7.

Description

Global optimization method for solving missing observed values in nonlinear regulation network Technical Field The invention relates to the technical field of biological information, in particular to a global optimization method for solving a missing observed value in a nonlinear regulation network. Background In the prior art, the construction of a nonlinear regulation network is generally based on a metabolic pathway diagram and a corresponding structured file, wherein the nonlinear regulation network comprises a pharmacological network, a physiological network, a pathological network and a toxicological network, and the regulation relationship between molecular entities is described through nodes and edges. However, in some regulation structures, although nodes are marked in the path diagram, the side relationship between the nodes and upstream and downstream nodes is not recorded in the structured file, and the nodes are often omitted in the network model or directly connected and replaced by the upstream and downstream nodes, and the nodes comprise parts of the nonlinear regulation network, such as cell types, biological functions, biological anatomy tissues and the like, which cannot acquire the observed values. In the published chinese patent application CN202310852405.7, when the above node is introduced into a nonlinear regulation network model, the node value of the node is usually set empirically or manually, which lacks a unified determination method, and it is difficult to ensure rationality and consistency under different network structures and parameter scales, and to obtain a global optimal solution. Meanwhile, the existing network modeling method based on deep learning generally assumes that nodes have clear supervision information, and for nodes which play an intermediary role in a network but lack direct observation values, effective constraint and stable training are difficult to perform. Therefore, the prior art model parameters are not sufficiently interpretable, and the reliability of the network modeling results is limited. Disclosure of Invention The invention aims to provide a global optimization method for solving a missing observed value in a nonlinear regulation network, which can improve training accuracy by two orders, effectively reduce prediction errors and realize a global optimized numerical solution. The invention is realized in the following way: a global optimization method for solving missing observed values in a nonlinear regulation network comprises the following steps: Obtaining an experimental data set, wherein the experimental data set comprises biological components and expression levels of the biological components, and the biological components comprise pharmaceutical active components, target proteins, biomarkers and genes; Constructing a nonlinear regulation network model based on a deep learning model according to a target nonlinear regulation network; the nonlinear regulation network model comprises an input layer, a plurality of hidden layers and an output layer; Each network node of the input layer, the hidden layer and the output layer represents biological components, the value of each network node represents the expression quantity of the biological components, and the network links among the network nodes represent the interaction among the biological components; Inputting the experimental data set into a nonlinear regulation network model, and initializing weights and deviations by adopting a default initialization method; non-constraint training is carried out on the initialized nonlinear regulation network model, an average value is taken, a constraint value is obtained, and global optimization is achieved; Inputting the constraint value into a parameter updating method, and performing constraint training on the initialized nonlinear regulation network model by adopting the parameter updating method and a parameter constraint method to obtain a globally optimized network model parameter; And determining quantitative regulation information of the target nonlinear regulation network according to the globally optimized network model parameters, wherein the quantitative regulation information is used for predicting a regulation result according to given drug input. Further, the method of acquiring the experimental data set comprises: Extracting nonlinear regulation network original data from literature experiment data; and expanding the original data of the nonlinear regulation network by adopting a method based on normal distribution or uniform distribution to generate an experimental data set. Further, constructing a nonlinear regulation network model based on a deep learning model according to the target nonlinear regulation network, which specifically comprises the following steps: determining an input layer, a plurality of hidden layers and an output layer according to the action sequence of each biological component in the target nonlinear regula