Pig feed efficiency prediction model and system based on multi-omics data

Abstract

The invention relates to the crossing field of artificial intelligence technology and bioinformatics, and discloses a pig feed efficiency prediction model and system based on multi-omics data. Modulating a neural differential equation which runs on a priori knowledge graph and is realized by a graph neural network by using the matrix so as to solve and generate a continuous evolution trajectory of an individual physiological state; and finally, aggregating the tracks, combining the constraint matrix, and outputting a feed efficiency prediction value through a second preset model. The invention further provides a corresponding prediction system which comprises a static constraint module, a dynamic core module and a prediction module. According to the method, static genetic constraint and dynamic physiological process simulation are combined, genetic differences among different individuals can be reflected, and the biological consistency and individualization precision of a prediction model are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of artificial intelligence technology and bioinformatics, in particular to a pig feed efficiency prediction model and system based on multi-omics data. BACKGROUND

[0002] The feed efficiency of pigs, usually referring to the amount of feed consumed by pigs per unit of weight gain, is one of the key indicators for measuring the economic benefits of pig breeding. Accurate prediction of feed efficiency helps to select individuals with excellent genetic potential at an early stage of breeding, thereby accelerating the breeding process and reducing breeding costs, which has important economic value for the animal husbandry industry.

[0003] Currently, the prediction of complex traits such as pig feed efficiency mainly relies on genetic evaluation and statistical models. With the development of technology, researchers have begun to use machine learning or deep learning models combined with genomic, transcriptomic, metabolomic and other multi-omics data for prediction in order to improve the accuracy of prediction. However, the existing technical solutions still have certain limitations in dealing with such problems.

[0004] On the one hand, the growth and development of an organism is a continuous dynamic process, and its physiological state changes continuously over time throughout its life cycle. Existing prediction models are usually based on modeling of omics "snapshot" data at discrete time points, which may ignore the continuity information in the state evolution process and fail to fully capture the dynamic biological processes behind the formation of traits, thereby affecting the model's ability to represent the cumulative effects throughout the life cycle.

[0005] On the other hand, an individual's genetic background is an intrinsic factor that determines its life activity patterns. In the existing technical solutions, genomic data is used as input features, but it is usually on an equal footing with other omics data, failing to reflect its role as a fundamental constraint condition in the dynamic evolution of individual physiological processes. This separation or simple combination of static genetic information and dynamic physiological processes limits the depth of the model in simulating the "genotype to phenotype" complex process, and the biological consistency and individualized prediction accuracy of the model need to be further improved. SUMMARY

[0006] In view of the deficiencies of the prior art, the present application provides a pig feed efficiency prediction model and system based on multi-omics data, which solves the problem that the existing pig feed efficiency prediction technical solutions do not couple the individual's static genetic information with the continuous dynamic physiological process throughout its life cycle, resulting in room for improvement in the accuracy of the prediction results.

[0007] To achieve the above purpose, the present application realizes the following technical solutions: a pig feed efficiency prediction model and system based on multi-omics data, comprising:

[0008] Firstly, the model obtains genomic data of the individual to be tested, and processes the genomic data through a first preset model to generate a personalized system constraint matrix.

[0009] Subsequently, the model generates a continuous hidden state evolution trajectory of the individual within a preset growth time period on a preset prior knowledge graph containing biological entity nodes and inter-entity relationship edges by solving a neural differential equation modulated by the personalized system constraint matrix. The physiological state of the individual is represented by a hidden state vector h(t) corresponding to the nodes of the prior knowledge graph. The dynamic function of the neural differential equation is implemented by a graph neural network. The neural differential equation can be represented as:

[0010]

[0011] wherein, represents the rate of change of the hidden state vector with respect to time t; f dyn represents the dynamic function implemented by the graph neural network; h(t) is the hidden state vector at the current time; is the topological structure of the prior knowledge graph; Θ G is the personalized system constraint matrix; ψ is a preset parameter of the dynamic function.

[0012] The generation step of the continuous hidden state evolution trajectory includes: taking the value of the hidden state vector at the current time as the input of the graph neural network; calculating the rate of change of the hidden state vector with respect to time by running the graph neural network on the topological structure of the prior knowledge graph; and integrating the rate of change within the preset growth time period to obtain the continuous hidden state evolution trajectory.

[0013] The modulation of the personalized system constraint matrix to the graph neural network is implemented by at least one of the following ways:

[0014] Node modulation: adjusting the calculation parameters associated with a specific biological entity node in the graph neural network according to the numerical value of the personalized system constraint matrix.

[0015] Edge modulation: adjusting the message passing weight associated with a specific inter-entity relationship edge in the graph neural network according to the numerical value of the personalized system constraint matrix.

[0016] Finally, the model performs aggregated network processing on the continuous hidden state evolution trajectory, and inputs the processing result together with the personalized system constraint matrix into a second preset model, and outputs the feed efficiency prediction value of the individual by the second preset model.

[0017] The second aspect of the present application provides a pig feed efficiency prediction system based on multi-omics data.

[0018] The system is configured to implement the model described in any of the preceding embodiments, and in one embodiment, the system comprises:

[0019] A static constraint module is configured to obtain genomic data of a subject to be tested, and process the genomic data through a first preset model to generate a personalized system constraint matrix, and output the personalized system constraint matrix.

[0020] A dynamic core module is configured to receive the personalized system constraint matrix output from the static constraint module, and based on the matrix, solve a neural differential equation modulated by the personalized system constraint matrix on a preset prior knowledge graph comprising biological entity nodes and inter-entity relationship edges, to generate a continuous hidden state evolution trajectory of the subject within a preset growth period, and output the continuous hidden state evolution trajectory.

[0021] A prediction module is configured to receive the continuous hidden state evolution trajectory output from the dynamic core module and the personalized system constraint matrix output from the static constraint module, and perform aggregated network processing on the continuous hidden state evolution trajectory, and input the processing result and the personalized system constraint matrix into a second preset model to output a feed efficiency prediction value of the subject.

[0022] The present application provides a pig feed efficiency prediction model and system based on multi-omics data. The present application has the following beneficial effects:

[0023] 1. The present application generates a personalized system constraint matrix from the genomic data of the subject through a first preset model, and modulates the neural differential equation describing the continuous change of physiological state using the matrix. This technical solution fundamentally constrains the simulation of physiological evolution trajectory by the genetic characteristics of the subject, ensuring that the prediction model can reflect the genetic differences between different individuals, thereby making the prediction results more consistent with the biological nature of the individual and improving the individualized accuracy of the prediction.

[0024] 2. The present application uses a neural differential equation to solve on a preset prior knowledge graph, generating a hidden state evolution trajectory that can reflect the continuous change of physiological state over time. This processing method not only better reflects the continuity of the physical process of biological growth and development, but also ensures that the internal state change of the model has a clear biological reference, enhancing the interpretability of the technical solution.

[0025] 3、The complete continuous hidden state evolution trajectory is processed through the polymerization network to capture the cumulative effect of the trait in the entire life cycle, and the personalized system constraint matrix is combined again to take into account the static genetic effect. By fusing dynamic process information and static genetic information as a prediction basis, the input information of the final prediction is more comprehensive, thereby improving the accuracy of the prediction result of the complex trait of feed efficiency. BRIEF DESCRIPTION OF DRAWINGS

[0026] Figure 1 The system overall structure block diagram of the present application is shown in the figure.

[0027] Figure 2 The function flow chart of the static constraint module of the present application is shown in the figure.

[0028] Figure 3 The function flow chart of the dynamic core module of the present application is shown in the figure.

[0029] Figure 4 The function flow chart of the prediction module of the present application is shown in the figure.

[0030] Figure 5 The modulation mechanism schematic diagram of the personalized system constraint matrix of the present application is shown in the figure. DETAILED DESCRIPTION

[0031] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the specification of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0032] Please refer to the drawings in the specification of the present application Figure 1 The pig feed efficiency prediction model and system based on multi-omics data provided by the embodiments of the present application include:

[0033] The system is configured to perform a calculation process defined by a pig feed efficiency prediction model based on multi-omics data.

[0034] In one embodiment, the system can include a static constraint module 10, a dynamic core module 20, and a prediction module 30. The three modules are connected in sequence and interact with data through a pre-set data interface.

[0035] The static constraint module 10 is configured to obtain the genomic data of the individual to be tested, and process the genomic data through a first pre-set model to generate a personalized system constraint matrix. The calculation result of the module, i.e. the personalized system constraint matrix, is output to the dynamic core module 20 and the prediction module 30.

[0036] The dynamic core module 20 has its input connected to the output of the static constraint module 10. This module is configured to receive a personalized system constraint matrix and, based on this matrix, generate a continuous hidden state evolution trajectory of an individual over a preset growth period by solving a neural differential equation on a preset prior knowledge graph containing biological entity nodes and inter-entity relationship edges.

[0037] Neural differential equations are used to characterize the continuous changes in an individual's physiological state over time. An individual's physiological state is represented by a hidden state vector h(t) corresponding to a node in a prior knowledge graph. The specific form of the neural differential equation is as follows:

[0038]

[0039] in:

[0040] f represents the rate of change of the hidden state vector with respect to time t; dyn Let h(t) represent a dynamic function implemented by a graph neural network; h(t) represents the value of the hidden state vector at time t. Represents the topological structure of the prior knowledge graph; Θ G ψ represents the personalized system constraint matrix generated by static constraint module 10; ψ represents the dynamic function f. dyn The preset parameter set. The calculation results of the dynamic core module 20, namely the continuous hidden state evolution trajectory, are output to the prediction module 30.

[0041] The calculation results of the dynamic core module 20, namely the continuous hidden state evolution trajectory, are output to the prediction module 30.

[0042] The prediction module 30 has its inputs connected to the outputs of the static constraint module 10 and the dynamic core module 20, respectively. This module is configured to receive continuous hidden state evolution trajectories and personalized system constraint matrices.

[0043] The prediction module 30 performs aggregation network processing on the continuous hidden state evolution trajectory and inputs the processing results along with the personalized system constraint matrix into a second preset model. The second preset model calculates and outputs the final predicted value of the individual's feed efficiency.

[0044] See attached document Figure 2 , Figure 2 This is a functional flowchart of a static constraint module 10 according to an embodiment of the present invention. The specific functional implementation flow of the static constraint module 10 is as follows.

[0045] First, the static constraint module 10 acquires the genomic data of the individual to be tested. In one specific embodiment, the genomic data is single nucleotide polymorphism (SNP) data. The SNP data undergoes preprocessing, including quality control, missing value imputation, and numerical encoding (e.g., encoding genotypes AA, Aa, and aa as 0, 1, and 2, respectively), ultimately forming a numerical matrix as input to this module. Next, the static constraint module 10 constructs an SNP graph structure based on the SNP data. In this graph structure, each SNP site is defined as a node. Whether an edge exists between two nodes depends on the physical distance between the corresponding two SNP sites on the chromosome. If the physical distance between two SNP sites is less than a preset threshold, an edge is established between the two nodes.

[0046] Subsequently, the static constraint module 10 inputs the SNP graph structure into the first preset model. In one embodiment, the first preset model is a graph attention network (GAT). The graph attention network performs operations on the topology of the SNP graph structure. For each target node in the graph, the network calculates the attention coefficients between the target node and all its neighboring nodes.

[0047] The graph attention network weights the feature vector of each neighbor node based on the calculated attention coefficients. Then, the network aggregates all the weighted neighbor feature vectors and generates a new feature vector for the target node in the next layer through a non-linear transformation. This process is repeated across multiple layers of the network to obtain the final feature representation of each SNP node.

[0048] Finally, the static constraint module 10 generates a personalized system constraint matrix Θ based on the final feature representations of each SNP node output by the graph attention network. G In one specific embodiment, this generation step includes performing a pre-defined pooling operation on the final feature representations of all nodes, and then linearly transforming the pooled result through a fully connected layer to obtain a personalized system constraint matrix Θ with predetermined dimensions and structure. G .

[0049] After completing the calculation, the static constraint module 10 will generate the personalized system constraint matrix Θ. G Transmitted to the dynamic core module 20 and the prediction module 30.

[0050] See attached document Figure 3 , Figure 3 This is a functional flowchart of a dynamic core module 20 according to an embodiment of the present invention. The input terminal of the dynamic core module 20 is connected to the output terminal of the static constraint module 10, and is used to receive the personalized system constraint matrix Θ. G .

[0051] Before performing the computation, the dynamic core module 20 loads a pre-defined prior knowledge graph. In one embodiment, the map is constructed based on a publicly available biological database (such as the KEGG or Reactome database). Nodes in the map represent biological entities, such as genes, proteins, or metabolites; edges represent interactions between biological entities, such as activation, inhibition, or metabolic transformation.

[0052] The core function of the dynamic core module 20 is to solve the neural differential equations mentioned above to generate continuous hidden state evolution trajectories. This solution process takes place over a preset growth time period [t0, t...]. final The process takes place within a specified time frame. The inputs to this process include: an initial hidden state vector h(t0) generated based on prior knowledge or randomly, and a time interval [t0, t...]. final ], and the dynamic function f implemented by a graph neural network. dyn The dynamic core module 20 employs a numerical integrator (e.g., a Runge-Kutta solver) to solve the neural differential equations. At any time point t, the numerical integrator calls the dynamic function f. d yn is used to calculate the rate of change dh(t) / dt of the current hidden state h(t). The integrator uses this rate of change to calculate the hidden state after the next small time step. This is achieved over the entire time interval [t0, t...]. final This process is repeated within the range, eventually generating a continuous hidden state evolution trajectory consisting of a series of hidden state vectors arranged in chronological order.

[0053] dynamic function f dym It is implemented using a graph neural network. In each computational step, the graph neural network receives the hidden state vector h(t) at the current time step as its input node features, and then compares it with the prior knowledge graph. Message passing and node update operations are performed on the topology. After the operation is completed, the graph neural network outputs a vector with the same dimension as the input, which is the rate of change of the hidden state vector at the current time step: dh(t) / dt.

[0054] The computation process of a graph neural network is subject to the personalized system constraint matrix Θ. G Modulation. In one embodiment, modulation is achieved through at least one of the following methods:

[0055] Node modulation: In the computation process of graph neural networks, from the personalized system constraint matrix Θ G A subset of values ​​is extracted and transformed using a predefined transformation (e.g., linear transformation) to generate a node modulation vector. This node modulation vector is then applied to the feature vectors of specific biological entity nodes in the graph neural network through element-wise addition or multiplication to adjust their underlying activity parameters.

[0056] Side modulation: In the message passing process of a graph neural network, from the personalized system constraint matrix Θ G Another portion of the values ​​is extracted, and a modulation weight is generated for a specific relation edge in the prior knowledge graph through a preset transformation. This modulation weight is used to scale the size of the message transmitted along the edge, thereby adjusting the strength of the interaction between the entities represented by the edge.

[0057] After completing the calculation, the dynamic core module 20 transmits the generated continuous hidden state evolution trajectory to the prediction module 30.

[0058] See attached document Figure 4 , Figure 4 This is a functional flowchart of the prediction module 30 according to an embodiment of the present invention. The input terminals of the prediction module 30 are connected to the output terminals of the static constraint module 10 and the dynamic core module 20, respectively.

[0059] The prediction module 30 is configured to receive the continuous hidden state evolution trajectory generated by the dynamic core module 20 and the personalized system constraint matrix Θ generated by the static constraint module 10. G .

[0060] First, the prediction module 30 processes the continuous hidden state evolution trajectory through an aggregation network. The continuous hidden state evolution trajectory is a sequence of multiple hidden state vectors arranged in chronological order. The aggregation network transforms this sequence data into a fixed-dimensional trajectory feature vector.

[0061] In one embodiment, the aggregation network is a temporal pooling layer based on a self-attention mechanism. This network calculates the correlation weights between all hidden state vectors in the trajectory and performs a weighted summation of all hidden state vectors based on these weights to generate a trajectory feature vector that can characterize the cumulative features of the dynamic process throughout the entire growth period.

[0062] Next, the prediction module 30 combines the trajectory feature vector generated in the previous step with the received personalized system constraint matrix Θ. G The combination operation involves concatenating two vectors or matrices to form a combined feature vector.

[0063] Finally, the prediction module 30 inputs the combined feature vector into a second preset model to calculate and output the predicted feed efficiency value for the individual. In one embodiment, the second preset model is a multilayer perceptron (MLP). The MLP receives the combined feature vector and, after calculation by one or more fully connected layers and nonlinear activation functions contained within it, finally outputs a scalar value. This scalar value is the final predicted result of the individual's feed efficiency.

[0064] See attached document Figure 5 , Figure 5 This is a schematic diagram of the structure of a computing device that can be used to implement the system of the present invention, according to an embodiment of the present invention.

[0065] like Figure 5 As shown, the computing device 500 may include one or more processors 510, a memory 520, and a bus 530 for connecting the processors 510 and the memory 520. Those skilled in the art will understand that this structure is merely an example and does not constitute a limitation on the computing device 500.

[0066] Memory 520 is configured to store computer program instructions. Memory 520 may include high-speed random access memory (RAM) or non-volatile memory, such as at least one disk storage device or flash memory.

[0067] The processor 510 is configured to execute computer program instructions stored in the memory 520 to implement the pig feed efficiency prediction method based on multi-omics data in any of the foregoing embodiments.

[0068] In one embodiment, the computing device 500 may further include an input / output interface for connecting to an external data input / output device, and a network interface for communicating with one or more other computing devices.

[0069] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, can perform the steps described in the foregoing method embodiments. The computer-readable storage medium can be any device, apparatus, tangible medium, or article capable of storing program code.

[0070] Those skilled in the art will understand that the static constraint module 10, the dynamic core module 20, and the prediction module 30 described in the foregoing embodiments, as well as the more specific processing units within the modules, can all implement their functions by the processor 510 executing the corresponding computer program instructions stored in the memory 520.

[0071] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A pig feed efficiency prediction model based on multi-omics data, characterized in that, include: The genomic data of the individual to be tested is acquired and processed through a first preset model to generate a personalized system constraint matrix; On a pre-defined prior knowledge graph containing biological entity nodes and inter-entity relationship edges, a neural differential equation modulated by the personalized system constraint matrix is ​​solved to generate a continuous hidden state evolution trajectory of the individual within a pre-defined growth period. The dynamic function of the neural differential equation is implemented by a graph neural network, and the physiological state of the individual is represented by a hidden state vector corresponding to the prior knowledge graph. The continuous hidden state evolution trajectory is processed by an aggregation network, and combined with the personalized system constraint matrix, the feed efficiency of the individual is predicted and output through a second preset model.

2. The pig feed efficiency prediction model based on multi-omics data according to claim 1, characterized in that, The steps for generating the personalized system constraint matrix are as follows: The genomic data is input into the first preset model, which is a graph attention network. The graph attention network captures the epistatic effects and linkage relationships between single nucleotide polymorphism sites to generate the personalized system constraint matrix.

3. The pig feed efficiency prediction model based on multi-omics data according to claim 1, characterized in that, The steps for generating the continuous hidden state evolution trajectory are as follows: The value of the hidden state vector at the current time is used as the input of the graph neural network; The rate of change of the hidden state vector with respect to time is calculated using the graph neural network. The rate of change is integrated to solve the neural differential equation, thereby obtaining the evolution trajectory of the continuous hidden state.

4. The pig feed efficiency prediction model based on multi-omics data according to claim 3, characterized in that, During the computation of the graph neural network, the graph neural network is modulated using the personalized system constraint matrix. The modulation step specifically includes at least one of the following: Node modulation: Using the personalized system constraint matrix, the basic activity or sensitivity to input associated with a node in the graph neural network that is related to a specific biological entity; Edge modulation: Using the personalized system constraint matrix, adjust the message passing efficiency or weights related to the relationship edges between specific entities in the graph neural network.

5. The pig feed efficiency prediction model based on multi-omics data according to claim 1, characterized in that, The steps for generating the feed efficiency of the individual are as follows: The continuous hidden state evolution trajectory is processed by an aggregation network to generate a trajectory feature vector that summarizes the entire life cycle; The trajectory feature vector and the personalized system constraint matrix are used together as input to the second preset model to predict the feed efficiency of the individual.

6. A pig feed efficiency prediction system based on multi-omics data, characterized in that, The pig feed efficiency prediction model based on multi-omics data, applied to any one of claims 1-6, includes: The static constraint module is used to acquire the genomic data of the individual to be tested and process the genomic data through a first preset model to generate a personalized system constraint matrix; The dynamic core module is used to generate a continuous hidden state evolution trajectory of the individual within a preset growth period by solving a neural differential equation modulated by the constraint matrix of the personalized system on a preset prior knowledge graph containing biological entity nodes and relationship edges between entities. The prediction module is used to perform aggregation network processing on the continuous hidden state evolution trajectory, and combine the prediction of the individual with the personalized system constraint matrix and the second preset model to output the feed efficiency of the individual.

7. The pig feed efficiency prediction system based on multi-omics data according to claim 6, characterized in that, The static constraint module includes: A matrix generation unit is used to input the genomic data into the first preset model, which is a graph attention network, and to capture the epistatic effects and linkage relationships between single nucleotide polymorphism sites through the graph attention network to generate the personalized system constraint matrix.

8. The pig feed efficiency prediction system based on multi-omics data according to claim 6, characterized in that, The dynamic core module includes: A trajectory generation unit is used to take the value of the individual's physiological state at the current moment as the input of the graph neural network, calculate the rate of change of the hidden state vector with respect to time through the graph neural network, and integrate the rate of change to solve the neural differential equation, thereby obtaining the continuous hidden state evolution trajectory.

9. The pig feed efficiency prediction system based on multi-omics data according to claim 6, characterized in that, The prediction module includes: The feature aggregation unit is used to process the continuous hidden state evolution trajectory through an aggregation network to generate a trajectory feature vector that summarizes the entire life cycle. An efficiency prediction unit is configured to receive the trajectory feature vector and use the trajectory feature vector and the personalized system constraint matrix together as input to the second preset model to predict the feed efficiency of the individual.

Images