A large model training method and device for gene regulatory network prediction

Abstract

This specification discloses a large-model training method and device for gene regulatory network prediction. In the constructed gene regulatory network, a one-hop subgraph centered on each node is determined, and the graph embedding features of the subgraph, the positive sample of the subgraph, and the negative sample of the subgraph are respectively obtained through a preset graph encoding model, and the contrast loss is calculated based on the graph embedding features. Then, through the first prompt information, the large model learns the correspondence between the graph embedding features of the subgraph and the text embedding features of each gene in the subgraph, and calculates the cross-entropy loss. The parameters of the large model are jointly adjusted by the contrast loss and the cross-entropy loss. In this method, the graph structure information represented by the graph embedding features obtained by the graph encoding model is migrated to the large model, so that the large model performs joint reasoning based on the correspondence between the learned graph structure information of the gene regulatory network and the text description information of the gene, and has the ability to predict gene regulatory relationships.

Description

Technical Field

[0001] This specification relates to the field of computer technology, and in particular to a large-scale model training method and device for gene regulatory network prediction. Background Art

[0002] Gene regulatory networks are complex systems that describe the interactions between genes. Their topological structure directly influences key biological processes such as cell differentiation and disease development. Traditional gene regulatory network (GRN) discovery methods rely on experimental techniques such as gene knockout (KO) and chromatin immunoprecipitation (ChIP-seq). These methods suffer from the following drawbacks: high cost: a single experiment can cost thousands of dollars, and the cost of constructing large-scale networks increases exponentially; low throughput: experimental cycles typically take weeks to months, making them difficult to meet the needs of dynamic network analysis; and limited coverage: they only capture the regulatory relationships between known transcription factors and target genes, ignoring emerging regulatory elements such as non-coding RNAs. Existing computational models (such as GENIE3 and SCENIC) can partially replace experimental methods, but they are based on statistical correlation rather than causal inference, resulting in a high false-positive rate.

[0003] Based on this, this specification provides a large model training method for gene regulatory network prediction. Summary of the Invention

[0004] This specification provides a large-model training method, device, storage medium and electronic device for gene regulatory network prediction to at least partially solve the above-mentioned problems existing in the prior art.

[0005] This manual adopts the following technical solutions:

[0006] This specification provides a large model training method for gene regulatory network prediction, including:

[0007] Constructing a gene regulatory network based on gene pairs with regulatory relationships, wherein nodes in the gene regulatory network represent genes and edges represent regulatory relationships between genes;

[0008] In the gene regulatory network, a one-hop subgraph centered on a node of each gene is determined respectively;

[0009] For each subgraph, perform structural perturbation on the subgraph to generate positive and negative samples of the subgraph;

[0010] Obtaining, by means of a preset graph encoding model, graph embedding features of the subgraph, a positive sample of the subgraph, and a negative sample of the subgraph, respectively, and obtaining a contrast loss based on the graph embedding feature similarity between the subgraph and the positive sample and the graph embedding feature similarity between the subgraph and the negative sample;

[0011] For each subgraph, using the first prompt information, the large model learns the correspondence between the graph embedding features of the subgraph and the text embedding features of each gene in the subgraph, and calculates the cross entropy loss based on the first result output by the large model, wherein the text embedding features are obtained by feature extraction of pre-acquired text description information of the genes;

[0012] Parameters of the large model are adjusted according to the contrast loss and the cross entropy loss.

[0013] Optionally, for each subgraph, structural perturbation is performed on the subgraph to generate positive samples and negative samples of the subgraph, specifically including:

[0014] For each subgraph, delete the edges in the subgraph at a ratio less than a first set value to obtain a positive sample of the subgraph;

[0015] The edge connection relationship in the subgraph is changed at a ratio greater than a second set value to obtain a negative sample of the subgraph, wherein the first set value is less than the second set value.

[0016] Optionally, the first prompt information includes a graph feature placeholder, and the first prompt information is used to instruct the large model to sort the gene description information in a disrupted order;

[0017] The first prompt information enables the large model to learn the correspondence between the graph embedding feature of the subgraph and the text embedding feature of each gene in the subgraph, and calculates the cross entropy loss according to the first result output by the large model, specifically including:

[0018] Obtaining text embedding features extracted by the large model from the first prompt information, wherein the text embedding features include a graph feature placeholder, and the graph feature placeholder is not encoded;

[0019] Replacing the graph feature placeholder with the graph embedding feature to obtain a graph-text joint feature;

[0020] The large model performs reasoning based on the joint features of the graph and text, and outputs a ranking prediction result of the gene description information corresponding to each node contained in the subgraph;

[0021] A cross entropy loss is determined based on the difference between the ranking prediction result and the correct ranking result.

[0022] Optionally, the first prompt information includes description information of each node in the subgraph in order, and gene description information of each node in a disrupted order.

[0023] Optionally, after adjusting the parameters of the large model, the method further includes:

[0024] Saving the adjusted parameters of the large model;

[0025] Obtaining a sample gene pair, inputting the sample gene pair and second prompt information into the large model, and obtaining a second result output by the large model, wherein the second prompt information is used to indicate whether the large model predicts whether there is a regulatory relationship between the input gene pair;

[0026] determining an inference loss based on a difference between the second result and the true regulatory relationship of the sample gene pair;

[0027] Based on the inference loss, parameters of the large model are adjusted.

[0028] Optionally, the second result is a discrete confidence level, and the true regulatory relationship is represented by an existence binary label and an existence confidence level label;

[0029] Determining an inference loss based on a difference between the second result and the true regulatory relationship of the sample gene pair, specifically comprising:

[0030] Obtaining the predicted probability of the regulatory relationship between the sample gene pairs obtained by the large model;

[0031] Determining a first loss based on a difference between the predicted probability and the binary label of the presence of the sample gene pair; mapping the presence confidence level label of the sample gene pair to a continuous value, and determining a second loss based on a difference between the predicted probability and the mapped continuous value;

[0032] An inference loss is determined based on the first loss and the second loss.

[0033] Optionally, a gene pair with an unknown regulatory relationship is input into the large model, and the large model outputs a prediction result representing the regulatory relationship between the gene pair with the unknown regulatory relationship.

[0034] This specification provides a large model training device for gene regulatory network prediction, the device comprising:

[0035] A graph construction module constructs a gene regulatory network based on gene pairs with regulatory relationships, wherein nodes in the gene regulatory network represent genes and edges represent regulatory relationships between genes;

[0036] a subgraph extraction module, which determines, in the gene regulatory network, a one-hop subgraph centered on the node of each gene;

[0037] The sample generation module performs structural perturbations on each subgraph to generate positive and negative samples of the subgraph.

[0038] A contrastive learning module, which obtains graph embedding features of the subgraph, a positive sample of the subgraph, and a negative sample of the subgraph through a preset graph encoding model, and obtains a contrast loss based on the graph embedding feature similarity between the subgraph and the positive sample and the graph embedding feature similarity between the subgraph and the negative sample;

[0039] The graph-text combination module, for each subgraph, uses the first prompt information to enable the large model to learn the correspondence between the graph embedding features of the subgraph and the text embedding features of each gene in the subgraph, and calculates the cross entropy loss based on the first result output by the large model, wherein the text embedding features are obtained by feature extraction of the pre-acquired text description information of the genes;

[0040] A parameter adjustment module adjusts the parameters of the large model according to the contrast loss and the cross entropy loss.

[0041] This specification provides a computer-readable storage medium, which stores a computer program. When the computer program is executed by a processor, it implements the above-mentioned large model training method for gene regulatory network prediction.

[0042] This specification provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, the large model training method for gene regulatory network prediction is implemented.

[0043] At least one of the above technical solutions adopted in this specification can achieve the following beneficial effects:

[0044] In the large-model training method for gene regulatory network prediction provided in this specification, in the constructed gene regulatory network, a one-hop subgraph centered on each node is determined, and the graph embedding features of the subgraph, the positive sample of the subgraph, and the negative sample of the subgraph are respectively obtained through a preset graph encoding model, and the contrast loss is calculated based on the graph embedding features. Then, through the first prompt information, the large model learns the correspondence between the graph embedding features of the subgraph and the text embedding features of each gene in the subgraph, and calculates the cross-entropy loss. The large model parameters are jointly adjusted by contrast loss and cross-entropy loss. In this method, the graph structure information represented by the graph embedding features obtained by the graph encoding model is migrated to the large model, so that the large model performs joint reasoning based on the correspondence between the learned graph structure information of the gene regulatory network and the text description information of the gene, and has the ability to predict gene regulatory relationships. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] The drawings described herein are used to provide a further understanding of this specification and constitute a part of this specification. The exemplary embodiments and descriptions of this specification are used to explain this specification and do not constitute an improper limitation of this specification. In the drawings:

[0046] Figure 1 This is a flowchart of a large model training method for gene regulatory network prediction in this specification;

[0047] Figure 2 A model training architecture diagram provided in an embodiment of this specification;

[0048] Figure 3 A data construction process diagram provided in an embodiment of this specification;

[0049] Figure 4 A schematic diagram of a large model training device for gene regulatory network prediction provided in this specification;

[0050] Figure 5 The corresponding Figure 1 Schematic diagram of electronic equipment. DETAILED DESCRIPTION

[0051] To make the purpose, technical solutions, and advantages of this specification more clear, the technical solutions of this specification will be clearly and completely described below in conjunction with the specific embodiments of this specification and the corresponding drawings. Obviously, the embodiments described are only part of the embodiments of this specification, not all of the embodiments. Based on the embodiments in this specification, all other embodiments obtained by ordinary technicians in this field without making any creative efforts are within the scope of protection of this application.

[0052] The technical solutions provided by the various embodiments of this specification are described in detail below with reference to the accompanying drawings.

[0053] Figure 1 The following is a flow chart of a large model training method for gene regulatory network prediction in this specification, which specifically includes the following steps:

[0054] S100: Constructing a gene regulatory network based on gene pairs with regulatory relationships, wherein nodes in the gene regulatory network represent genes, and edges represent regulatory relationships between genes.

[0055] All steps in the large-scale model training method for gene regulatory network prediction provided in this specification can be implemented by any electronic device with computing capabilities, such as a terminal, server, or other device. For ease of description, the large-scale model training method for gene regulatory network prediction provided in this specification will be described below using only the server as the execution entity.

[0056] In this specification, gene regulatory networks can be constructed based on known gene pairs with regulatory relationships for training large models. Gene regulatory networks are graph data network structures, where nodes represent genes and edges represent regulatory relationships. Gene pairs can be obtained from public datasets on the Internet. The gene regulatory network can be represented as G=(V,E), where V represents the set of gene nodes and E represents the set of edges.

[0057] S102: In the gene regulatory network, determine a one-hop subgraph centered on each node.

[0058] In a gene regulatory network, an edge connecting two nodes indicates a regulatory relationship between the genes represented by these two nodes. For each node in the gene regulatory graph, a one-hop subgraph centered on that node—that is, a subgraph consisting of nodes directly connected to that node—represents all genes that have a regulatory relationship with the gene represented by that node.

[0059] Therefore, in order to enable the large model to learn the regulatory relationship between genes, the server extracts a one-hop subgraph of each node in the gene regulatory network with each node as the center.

[0060] S104: For each subgraph, perform structural perturbation on the subgraph to generate positive samples and negative samples of the subgraph.

[0061] In this method, the large model is optimized using contrastive loss. This step first constructs the positive and negative samples required for contrastive loss calculation from each subgraph. This is done by perturbing each subgraph's structure to determine the positive and negative samples for each subgraph.

[0062] In one embodiment of the present specification, the server may perform a relatively minor structural perturbation on the subgraph to generate a positive sample of the subgraph, which is highly similar to but not identical to the subgraph itself. The server may also perform a relatively major structural perturbation on the subgraph to generate a negative sample of the subgraph, which is significantly different from the subgraph itself.

[0063] Specifically, for each subgraph, the server may delete edges in the subgraph at a ratio less than a first set value to obtain a positive sample of the subgraph, and change the link relationship of the edges in the subgraph at a ratio greater than a second set value to obtain a negative sample of the subgraph.

[0064] The first set value is smaller than the second set value to ensure that the positive sample has a high similarity to the sub-image itself and the negative sample has a low similarity to the sub-image itself. For example, the first set value can be set to a value less than 10%, and the second set value can be set to 50%.

[0065] In this embodiment, structural perturbations are performed when determining both positive and negative samples, rather than only when generating negative samples. This can reduce the impact of data errors in the gene pairs obtained in step S100 on model reasoning during training, thereby enhancing the robustness and generalization of large-model gene regulatory relationship predictions.

[0066] S106: Obtain graph embedding features of the subgraph, the positive sample of the subgraph, and the negative sample of the subgraph respectively through a preset graph encoding model, and obtain contrast loss according to the graph embedding feature similarity between the subgraph and the positive sample and the graph embedding feature similarity between the subgraph and the negative sample.

[0067] The large model is designed for text encoding and cannot directly process graph data. This specification uses an additional graph encoding model to encode each subgraph, its positive samples, and its negative samples, obtaining graph embedding features corresponding to each subgraph, its positive samples, and its negative samples. This specification does not restrict the specific type of graph encoding model; you can choose one as needed.

[0068] Then, the contrastive loss is calculated based on the graph embedding feature similarity between each subgraph and the positive sample of the subgraph, as well as the graph embedding feature similarity between each subgraph and the negative sample of the subgraph.

[0069] Among them, if Graph Attention Networks (GAT) is used as the graph encoding model, Represents the graph embedding feature of the subgraph G, Represents the positive sample of this subgraph The graph embedding features of Represents the negative sample of this sub-image The graph embedding features, then the contrast loss It can be determined according to the following formula:

[0070]

[0071]

[0072]

[0073] In the above formula Represents the temperature parameter, which is used to adjust the sharpness of the similarity distribution. The larger the value, the flatter the distribution. The smaller the value, the sharper the distribution and the stronger the contrast. It can usually be set to 0.05~0.2. K represents the number of negative samples.

[0074] By calculating the contrast loss, the similarity between the sub-graph and the positive sample can be shortened, and the similarity between the sub-graph and the negative sample can be shortened during the large model inference process.

[0075] It's important to note that the graph encoding model used in this specification is an existing, pre-trained model for book corner encoding. The encoding results of this graph encoding model are used to train the large model. Therefore, while the contrastive loss is calculated based on the encoding results of the graph encoding model, it is used to adjust the parameters of the large model to transfer the subgraph structure knowledge encoded by the graph encoding model to the large model.

[0076] S108: For each subgraph, the first prompt information is used to enable the large model to learn the correspondence between the graph embedding features of the subgraph and the text embedding features of each gene in the subgraph, and calculate the cross entropy loss according to the first result output by the large model, wherein the text embedding features are obtained by feature extraction of the text description information of the genes obtained in advance.

[0077] The method in this specification can be trained based on any type of large model, which can be a language model with a large parameter scale, such as the Generative Pre-trained Transformer (GPT), or a lightweight language model with a small parameter scale, such as phi3-mini.

[0078] In this specification, the training data includes not only gene pairs with regulatory relationships but also textual descriptions of each gene. This textual description can also be obtained from public datasets to describe the gene function.

[0079] The gene regulatory network constructed based on gene pairs with known regulatory relationships is graph data, and the gene description information of each gene is text data. In this step, through the first prompt information, the large model learns the correspondence between the graph embedding features of the subgraph and the text embedding features of each gene in the subgraph, and based on the graph data and text data, two types of data related to gene regulatory relationships, reasoning is performed to learn the regulatory relationship between gene pairs, so that the large model has the ability to predict the regulatory relationship of gene pairs with unknown gene regulatory relationships.

[0080] When training the large model to learn the correspondence between the graph embedding features of the subgraph and the text embedding features of each gene in the subgraph, various methods can be used. Different training methods will result in different first prompt information content and the first result output by the large model.

[0081] In one embodiment of this specification, the large model learns the correspondence between the graph embedding features of the subgraph and the text embedding features of each gene in the subgraph by having the large model sort the shuffled gene description information. The first prompt information is used to instruct the large model to sort the shuffled gene description information.

[0082] In this embodiment, the first prompt information includes a node list and a gene description list. The node list is the nodes of the subgraph in a given order, and the gene description list is the gene description information of each node contained in the subgraph in a given order. The order of the gene descriptions in the gene description list does not completely correspond to the order of the nodes in the node list.

[0083] The large model outputs reordered gene descriptions according to the first prompt information so that they correspond to the order of the nodes in the node list. The gene description order output by the large model is used as the first result. The cross-entropy loss is calculated based on the difference between the first result and the correct order of the gene descriptions that corresponds to the order of the nodes in the node list.

[0084] If the correct sorting result is expressed as , and express the first result as , then the cross entropy loss It can be expressed as . Among them, CE represents the cross entropy loss calculation function.

[0085] S110: Adjusting parameters of the large model according to the contrast loss and the cross entropy loss.

[0086] Contrastive loss can enhance the robustness of large models to graph embedding features, optimizing their ability to understand graph structural information. Cross-entropy loss enables large models to learn the correspondence between gene regulatory network structures and textual descriptions of genes, enabling them to reason jointly based on graph and text data, thus optimizing language generation tasks.

[0087] In this manual, the total loss is determined by comparing the loss and the cross-entropy loss, and the large model is jointly optimized. At the same time, the graph structure understanding task and the language generation task are optimized, so that the optimized large model can achieve more accurate prediction of gene regulatory relationships.

[0088] The total loss can be the weighted sum of the contrast loss and the cross entropy loss, and the weighted coefficient is not limited in this specification. The weight coefficient is 0.5, and the cross entropy loss When the weight coefficient is 1, the total loss It can be expressed as: .

[0089] When the total loss converges, the model parameters are saved and the trained large model is obtained. The large model can be used for gene regulatory relationship prediction tasks. Specifically, the nodes corresponding to the gene pairs of unknown gene regulatory relationships can be The second prompt information is input into the large model, and the second prompt information is used to instruct the large model to predict the gene regulatory relationship between the input gene pairs and output the prediction result.

[0090] Based on the above Figure 1 The large model training method for gene regulatory network prediction shown in the figure determines a one-hop subgraph centered on each node in the constructed gene regulatory network, obtains the graph embedding features of the subgraph, the positive sample of the subgraph and the negative sample of the subgraph respectively through a preset graph encoding model, and calculates the contrast loss based on the graph embedding features. Then, through the first prompt information, the large model learns the correspondence between the graph embedding features of the subgraph and the text embedding features of each gene in the subgraph, and calculates the cross entropy loss. The parameters of the large model are jointly adjusted by the contrast loss and the cross entropy loss. In this method, the graph structure information represented by the graph embedding features obtained by the graph encoding model is migrated to the large model, so that the large model performs joint reasoning based on the correspondence between the learned graph structure information of the gene regulatory network and the text description information of the gene, and has the ability to predict gene regulatory relationships.

[0091] In step S102, when generating positive and negative samples, a dynamic structural perturbation strategy can be used to adjust the perturbation rate based on the node degree. High-degree nodes often represent more core regulatory relationships. When generating positive samples, the connection relationships corresponding to nodes with lower degrees are preferentially perturbed to ensure that the positive samples have a greater topological similarity with the original subgraph. When generating negative samples, the connection relationships corresponding to nodes with higher degrees are preferentially perturbed to ensure that the positive samples have a smaller topological similarity with the original subgraph, thereby enhancing the training effect of contrastive learning.

[0092] Specifically, when generating positive samples, nodes with degrees lower than a specified value are identified in the subgraph as first target nodes, and edges connected to the first target nodes are deleted at a ratio less than the first set value. When generating negative samples, nodes with degrees higher than a second specified value are identified in the subgraph as second target nodes, and the connection relationship of edges connected to the second target nodes is changed at a ratio greater than the second set value.

[0093] Alternatively, in another embodiment, when determining whether to perturb an edge, the retention probability corresponding to the edge may be calculated based on the degrees of the two nodes connected to the edge. The calculation formula for the retention probability P is as follows:

[0094] For the i-th edge contained in the subgraph , according to the two nodes connected by this edge ( and ) corresponds to the node degree, and calculates the geometric mean of the two node degrees as the importance of the edge The specific process is as follows: .

[0095] Then, among the importance levels corresponding to the edges of the subgraph, determine the maximum importance level. , used to normalize the retention probability and reduce the impact of the factor graph scale on the retention probability calculation.

[0096] Then the i-th edge The probability of retention The calculation formula is: .

[0097] For each subgraph, the retention probability of all edges contained in the subgraph is calculated, and the edges are arranged according to the size of the retention probability.

[0098] When generating positive samples, edges in a first preset ratio are selected from smallest to largest as candidate perturbation edges. Only these candidate perturbation edges are deleted, leaving the topology of other edges unchanged. When generating negative samples, edges in a second preset ratio are selected from largest to smallest as candidate perturbation edges, and their connections are replaced.

[0099] Because a larger degree indicates a more complex node connection, it represents an important regulatory relationship within the subgraph. The larger the degree product of the two nodes corresponding to an edge, the greater its retention probability. The retention probability indicates the importance of the edge in the subgraph topology. A node with a higher retention probability will have a greater impact on the subgraph topology after being deleted.

[0100] Therefore, when generating positive samples, we prioritize deleting edges with a lower probability of retention, so that the topological structure of the positive sample is more similar to that of the original subgraph. When generating negative samples, we prioritize changing the link relationship of edges with a higher probability of retention, so that the topological structure of the negative sample is less similar to that of the original subgraph.

[0101] When determining the cross entropy loss in the above step S108, the graph embedding features of the subgraph and the text embedding features of the text description information can be fused, so that the large model can perform reasoning based on the fused graph-text joint features, better learn the relationship between the graph structure information of the subgraph and the text semantics of the gene description information, and generate more accurate results.

[0102] When the first prompt information is used to instruct the large model to sort the shuffled gene description information, the first prompt information may include a graph feature placeholder. This graph feature placeholder is a special identification symbol pre-specified for large model encoding. If the large model does not encode the graph feature placeholder during text encoding, the graph feature placeholder will still be stored in the encoded result.

[0103] In this embodiment, the specific process of determining the cross entropy loss is as follows.

[0104] First, after the first prompt information is input into the large model, the large model performs text encoding on the first prompt information to obtain the text embedding feature extracted by the large model for the first prompt information. The text embedding feature contains an unencoded graph feature placeholder.

[0105] Then, for each subgraph, a preset graph encoding model is used to extract features of the subgraph to obtain graph embedding features, and the graph feature placeholder is replaced by the graph embedding features to obtain a joint graph-text feature.

[0106] If the graph feature placeholder is represented by <graph>,by The text embedding feature representing the first prompt information is represents the graph embedding feature of the subgraph, then the text embedding feature before replacement can be expressed as: , replaced text embedding features, (also known as image-text joint features ) can be expressed as: Here The symbol represents feature splicing, Indicates the text content before the feature placeholder in the first prompt information. Indicates the text content after the feature placeholder in the first prompt information.

[0107] After obtaining the joint features of the image and text, the joint features of the image and text are input into the large model again to obtain the first output of the large model. When phi3-mini is selected as the large model, the first output can be expressed as: .

[0108] In this embodiment, the second prompt information also includes a corresponding graph feature placeholder. During inference in the large model, the input gene pair with an unknown gene regulatory relationship is used as the gene pair to be predicted. The large model determines a one-hop subgraph in the gene regulatory network centered on the two gene nodes in the gene pair to be predicted. The graph encoding model then determines the graph embedding features of the two subgraphs. These subgraphs are used to replace the graph feature placeholders in the text embedding features encoded from the second prompt information, resulting in a joint graph-text feature. The large model then infers based on this joint graph-text feature and outputs a prediction result.

[0109] The large model trained in step S110 is now capable of predicting gene regulatory relationships based on the prompt words representing the edge regulatory relationship prediction task.

[0110] However, in order to enhance the inference accuracy of the large model for edge prediction tasks, in one embodiment of the present specification, after obtaining the large model trained in step S110, the large model can be further optimized for the edge regulation relationship prediction task.

[0111] The sample data for this optimization task consists of several sample gene pairs with known regulatory relationships. For each sample gene pair, the sample gene pair and the second hint information are input into the large model, and the large model outputs a second result. The second hint information instructs the large model to predict whether a regulatory relationship exists between the input gene pairs.

[0112] The server then determines the inference loss based on the difference between the second result and the true regulatory relationship of the sample gene pair, and continues to optimize the large model based on the inference loss.

[0113] The specific form of the second result obtained by the large model for predicting the regulatory relationship can be a binary result representing whether it exists, or a discrete confidence level representing the strength of the regulatory relationship.

[0114] If the second result is a binary label, then the true regulatory relationship is correspondingly a binary data. If the second result is a discrete confidence level, then the true regulatory relationship is also correspondingly a pre-classified confidence level label.

[0115] Furthermore, in another embodiment, two types of annotations can be used, namely, the true regulatory relationship includes a binary label representing whether it exists and a discrete confidence level label representing the strength of the regulatory relationship. Two types of losses are constructed through these two types of annotations, and the large model is optimized at the same time.

[0116] First, the predicted probability of a regulatory relationship between a sample gene pair is obtained from the large model. A first loss is then determined based on the difference between this predicted probability and the binary presence label of the sample gene pair. The presence confidence level label of the sample gene pair is mapped to a continuous value, and a second loss is determined based on the difference between this predicted probability and the mapped continuous value.

[0117] Specifically, the first loss is the binary cross entropy loss, which can be determined by the following formula:

[0118]

[0119] Among them, p represents the predicted probability output by the large model, y represents the binary label, .

[0120] When calculating the second loss, the discrete confidence levels must first be mapped to continuous values. This specification does not limit the specific mapping method. The mapping can be uniform or uneven according to the number of pre-divided confidence levels.

[0121] For example, if the confidence level is divided into 6 levels, including levels AE representing the confidence of the existence of a regulatory relationship from weak to strong, and confidence level N representing the confidence of the existence of a regulatory relationship is zero, that is, there is no regulatory relationship. Then a continuous value mapping method in this example is as follows: N→0, A→0.2, B→0.4, C→0.6, D→0.8, D→1.0.

[0122] The second loss can be calculated according to the following formula :

[0123]

[0124] in, Indicates the confidence level prediction result output by the model, To control the error threshold, it is usually set to 0.2~0.5.

[0125] In the calculation of the second loss, when , that is, when the difference between the confidence level prediction result output by the model and the confidence level label is small, the mean squared error (MSE) is used. The gradient decreases as the error decreases, and the adjustment of the model parameters is more refined. When the error exceeds the threshold When , it indicates that the error is large, and the Mean Absolute Error (MAE) is used to prevent the model loss from diverging due to extreme errors, avoid gradient explosion, and enhance the stability of training gradients.

[0126] Then, after obtaining the first loss and the second loss, the inference loss is determined based on the first loss and the second loss. Specifically, the inference loss It can be determined according to the following formula: .

[0127] After the optimization of the large model in this embodiment is completed, the gene pairs to be predicted can be Input the big model, and through the pre-set second prompt information, prompt the big model to output the confidence level of the regulatory relationship between the gene pair .

[0128] If the large model is not optimized for the confidence level prediction task of this embodiment, based on the second prompt information instructing the large model to output a confidence level, the large model can still output a confidence level prediction result based on its own semantic understanding capabilities according to the second prompt information. However, after optimizing the confidence level prediction task of this embodiment, the large model's accuracy in predicting the confidence level of gene regulatory relationships can be significantly increased.

[0129] In the above step S100, when constructing a gene regulatory network based on gene pairs with known regulatory relationships in a public dataset, the construction can be specifically carried out as follows.

[0130] First, the features of each gene are extracted through a preset feature extraction model to obtain the node embedding features of each gene.

[0131] This specification does not impose any specific restrictions on the feature extraction model herein. It can be a feature extraction model in a general field or a dedicated model in the genetic field, such as a single-cell generative pre-trained transformer (scGPT) based model.

[0132] Then, the node embedding feature is used as the feature identifier of the node to generate each gene node, and the gene nodes corresponding to the genes with regulatory relationships are connected with edges to obtain the gene regulatory network.

[0133] Figure 2 A model training architecture diagram provided in the embodiment of this specification, such as Figure 2 As shown in the figure, the data used for learning gene regulatory relationships are gene pairs with known regulatory relationships, gene description information, and gene node embedding features. Figure 2 The gene pairs in the diagram are: TF(1)→TF(2), TF(3)→gene(1), TF(4)→miRNA(1), miRNA(2)→TF(1), and miRNA(2)→gene(1), where TF stands for transcription factor, gene stands for gene, miRNA stands for microRNA, and → stands for regulatory relationship. The numbers in the brackets are used to identify different types of TF, gene, and miRNA. Figure 2 The gene regulatory information in the article was obtained by crawling public online data. The content is: USF1 encodes a member of the basic helixoop-helix leucine zipper family, and can function as a celular transcription factor. The encoded protein can activate transcription through pyrimidine-nich initiator (inr) elements and E-box motfs. This gene has been linked to famlal combined hyperipidemia (FCHL). Alterative splicing of this gene results in mufiple transcript vaiants. A related pseudogene has been defined on chromosome 21. The Chinese interpretation is: USF1 encodes a member of the basic helixoop-helix leucine zipper family, and can function as a celular transcription factor. The encoded protein can activate transcription through pyrimidine-nich initiator (inr) elements and E-box motfs. This gene has been linked to famlal combined hyperipidemia (FCHL). Alterative splicing of this gene results in mufiple transcript vaiants. A related pseudogene has been defined on chromosome 21. Alternative splicing of this gene can produce multiple transcript variants, and its related pseudogene is located on chromosome 21.

[0134] Based on the gene pairs with known regulatory relationships and the node embedding features of genes, a gene regulatory network is constructed, and the one-hop subgraph of each node in the gene regulatory network is encoded through the graph encoding model, the gene description information is encoded through the large model, and the graph embedding features of the subgraph are used. , replace the unencoded graph feature placeholders in the text embedding features to obtain the graph-text joint features, so that the large model can perform reasoning based on the graph-text joint features.

[0135] Figure 2 As shown in the figure, the joint feature of image and text , the text content before the feature placeholder in the first prompt information , graph embedding features and the text after the feature placeholder in the first prompt message Spliced ​​together.

[0136] In one embodiment of the present specification, the data input to the large model is organized according to a specific data format, and the training process of the above steps S100 to S110 is performed. Figure 3 A data construction process diagram provided in an embodiment of this specification is provided. Figure 5 The process shown specifically includes the following steps:

[0137] (1.1) Obtain gene pairs with regulatory relationships based on public online data, and use Python code to construct a known gene regulatory network G = (V, E) based on the gene pairs, where V is the set of gene nodes and E is the set of regulatory edges;

[0138] (1.2) Obtain the feature embedding of the gene in the scGPT model as the node embedding feature of the gene;

[0139] (1.3) Using Python crawlers, obtain text descriptions of genes from publicly available online data;

[0140] (1.4) Use Python to generate JSON data in the following format:

[0141] {

[0142] "id": dataid,

[0143] "graph": {

[0144] "node_idx": node_idx,

[0145] "edge_index": edge_index,

[0146] "node_list": node_list

[0147] },

[0148] "conversations": [

[0149] {"from": "human ", "value": conversation_human},

[0150] {"from": "gpt", "value": conversation_gpt} ]

[0152] }

[0153] Among them, "id" is the unique ID number corresponding to each data; "graph" represents the one-hop subgraph of the gene related to this data extracted from G, including the central node ID "node_idx", the edge index "edge_index" and the node list "node_list" in the one-hop subgraph excluding the central node. "conversations" is the training data in the form of conversations, which is used to guide model learning. "conversation_human" is the model training prompt word, and " <graph>” is the information placeholder for graph embedding features, and “conversation_gpt” is the target for model training.

[0154] (1.5) For the regulatory relationship learning task, generate task-specific data. The data content guides the model to learn the structure of a gene regulatory network with a single-hop subgraph centered on a gene. Specifically, the data format is the same as that described in step 1.4. The "conversation_human" data structure requires the model to sort the shuffled gene descriptions so that the gene descriptions correspond to the order of the gene points. The "conversation_gpt" data structure contains the correct order of the gene descriptions.

[0155] (1.6) For the large-model edge prediction inference task, generate data with a specific structure for this task. After model training is complete, the data format for the inference and prediction process follows this data structure. The data content guides the model to predict the degree of possible regulatory relationship between two genes. Specifically, the data format is the same as that described in step 1.4, where "conversation_human" describes two gene loci, and "conversation_gpt" contains the confidence level of the possible regulatory relationship between the two gene loci. The confidence level is categorized as N, A, or E, with E indicating the highest confidence level for possible edge relationship, A indicating the lowest confidence level, and N indicating no regulatory relationship.

[0156] In the above step S108, when the large model learns the correspondence between the graph embedding features of the subgraph and the text embedding features of each gene in the subgraph, the method is not limited to sorting the gene description information, and other methods can also be used.

[0157] For example, given a one-hop subgraph centered on a node and textual descriptions of several genes, the first prompt prompts the large model to select the textual description corresponding to the subgraph's central node from among the multiple textual descriptions. Alternatively, the gene descriptions of each gene pair describing the subgraph's topological regulatory relationships can be randomly masked, with the first prompt instructing the large model to restore the complete, unmasked gene regulatory information based on the subgraph's structural information.

[0158] The above is the large model training method X for gene regulatory network prediction provided in this specification. Based on the same idea, this specification also provides a corresponding large model training device for gene regulatory network prediction, such as Figure 4 shown.

[0159] Figure 4 A schematic diagram of a large model training device for gene regulatory network prediction provided in this specification, specifically including:

[0160] A graph construction module 200 is used to construct a gene regulatory network based on gene pairs with regulatory relationships, wherein nodes in the gene regulatory network represent genes and edges represent regulatory relationships between genes;

[0161] A subgraph extraction module 202 is used to determine a one-hop subgraph centered on a node of each gene in the gene regulatory network;

[0162] A sample generation module 204 is configured to perform structural perturbation on each subgraph to generate positive and negative samples of the subgraph;

[0163] a contrastive learning module 206 for respectively obtaining graph embedding features of the subgraph, a positive sample of the subgraph, and a negative sample of the subgraph using a preset graph coding model, and obtaining a contrastive loss based on the graph embedding feature similarity between the subgraph and the positive sample and the graph embedding feature similarity between the subgraph and the negative sample;

[0164] A graph-text combination module 208 is configured to, for each subgraph, enable the large model to learn the correspondence between the graph embedding features of the subgraph and the text embedding features of each gene in the subgraph through the first prompt information, and calculate the cross entropy loss based on the first result output by the large model, wherein the text embedding features are obtained by feature extraction of pre-acquired text description information of the genes;

[0165] The parameter adjustment module 210 is used to adjust the parameters of the large model according to the contrast loss and the cross entropy loss.

[0166] Optionally, the sample generation module 204 is specifically configured to, for each subgraph, delete edges in the subgraph at a ratio less than a first set value to obtain a positive sample of the subgraph, and change edge connectivity in the subgraph at a ratio greater than a second set value to obtain a negative sample of the subgraph, wherein the first set value is less than the second set value.

[0167] Optionally, the first prompt information includes a graph feature placeholder, and the first prompt information is used to instruct the large model to sort the gene description information in a scrambled order. The graph-text joint module 208 is specifically used to obtain the text embedding feature extracted by the large model from the first prompt information, wherein the text embedding feature includes a graph feature placeholder, and the graph feature placeholder is not encoded. The graph feature placeholder is replaced by the graph embedding feature to obtain the graph-text joint feature. The large model performs reasoning based on the graph-text joint feature, outputs the ranking prediction result of the gene description information corresponding to each node included in the subgraph, and determines the cross entropy loss based on the difference between the ranking prediction result and the correct ranking result.

[0168] Optionally, the first prompt information includes description information of each node in the subgraph in order, and gene description information of each node in a disrupted order.

[0169] Optionally, the device further includes an inference optimization module 212 .

[0170] The inference optimization module 212 is used to save the adjusted parameters of the large model;

[0171] Obtain a sample gene pair, input the sample gene pair and second prompt information into the large model, obtain a second result output by the large model, wherein the second prompt information is used to indicate that the large model makes a prediction on whether a regulatory relationship exists for the input gene pair, determine the inference loss based on the difference between the second result and the actual regulatory relationship of the sample gene pair, and adjust the parameters of the large model based on the inference loss.

[0172] Optionally, the second result is a discrete confidence level, and the true regulatory relationship is represented by an existence binary label and an existence confidence level label. The reasoning optimization module 212 is specifically configured to obtain the predicted probability of the existence of a regulatory relationship between the sample gene pair obtained by the large model, determine a first loss based on the difference between the predicted probability and the existence binary label of the sample gene pair; map the existence confidence level label of the sample gene pair to a continuous value, determine a second loss based on the difference between the predicted probability and the mapped continuous value, and determine an inference loss based on the first loss and the second loss.

[0173] Optionally, the apparatus further includes a prediction module 214 .

[0174] The prediction module 214 is specifically configured to input the gene pairs with unknown regulatory relationships into the large model, and the large model outputs a prediction result indicating the regulatory relationship between the gene pairs with unknown regulatory relationships.

[0175] This specification also provides a computer-readable storage medium, which stores a computer program that can be used to execute the above Figure 1 Provides a large model training method for gene regulatory network prediction.

[0176] This manual also provides Figure 5 The schematic structure diagram of the electronic device shown in FIG. Figure 5 As mentioned above, at the hardware level, the electronic device includes a processor, an internal bus, a network interface, a memory and a non-volatile memory, and may also include other hardware required for the business. The processor reads the corresponding computer program from the non-volatile memory into the memory and then runs it to achieve the above Figure 1 The large-scale model training method for gene regulatory network prediction described above. Of course, in addition to software implementation, this specification does not exclude other implementation methods, such as logic devices or a combination of software and hardware, etc. In other words, the execution body of the following processing flow is not limited to each logic unit, but can also be hardware or logic devices.

[0177] Technological improvements can be clearly categorized as either hardware improvements (for example, improvements to circuit structures like diodes, transistors, and switches) or software improvements (improvements to process flows). However, with technological advancements, many process flow improvements can now be considered direct hardware circuit structure improvements. Designers almost always create the corresponding hardware circuit structure by programming the improved process flow into the hardware circuit. Therefore, it cannot be said that a process flow improvement cannot be implemented using physical hardware modules. For example, a programmable logic device (PLD), such as a field programmable gate array (FPGA), is an integrated circuit whose logical function is determined by user programming. Designers can "integrate" a digital system onto a PLD by programming it themselves, without having to hire a chip manufacturer to design and manufacture a dedicated integrated circuit chip. Moreover, nowadays, instead of manually fabricating integrated circuit chips, this programming is mostly performed using "logic compiler" software. This is similar to the software compiler used during program development. Before compilation, the original code must be written in a specific programming language, called a hardware description language (HDL). There is not just one HDL, but many, such as ABEL (Advanced Boolean Expression Language), AHDL (Altera Hardware Description Language), Confluence, CUPL (Cornell University Programming Language), HDCal, JHDL (Java Hardware Description Language), Lava, Lola, MyHDL, PALASM, and RHDL (Ruby Hardware Description Language). Currently, the most commonly used are VHDL (Very-High-Speed ​​Integrated Circuit Hardware Description Language) and Verilog. Those skilled in the art will also understand that by simply programming the method flow in one of these hardware description languages ​​and then programming it into an integrated circuit, a hardware circuit that implements the logic method flow can be easily obtained.

[0178] The controller can be implemented in any suitable manner. For example, the controller can take the form of a microprocessor or processor and a computer-readable medium storing computer-readable program code (e.g., software or firmware) executable by the (micro)processor, logic gates, switches, an application-specific integrated circuit (ASIC), a programmable logic controller, and an embedded microcontroller. Examples of controllers include, but are not limited to, the following microcontrollers: ARC 625D, Atmel AT91SAM, Microchip PIC18F26K20, and Silicone Labs C8051F320. The memory controller can also be implemented as part of the memory control logic. Those skilled in the art will also appreciate that, in addition to implementing the controller purely in computer-readable program code, the controller can also be implemented in the form of logic gates, switches, an application-specific integrated circuit, a programmable logic controller, an embedded microcontroller, etc. by logically programming the method steps. Therefore, such a controller can be considered a hardware component, and the means for implementing the various functions included therein can also be considered as structures within the hardware component. Alternatively, the means for implementing the various functions can be considered both a software module implementing the method and a structure within the hardware component.

[0179] The systems, devices, modules, or units described in the above embodiments may be implemented by computer chips or entities, or by products having certain functions. A typical implementation device is a computer. Specifically, the computer may be, for example, a personal computer, a laptop computer, a cellular phone, a camera phone, a smartphone, a personal digital assistant, a media player, a navigation device, an email device, a game console, a tablet computer, a wearable device, or a combination of any of these devices.

[0180] For the convenience of description, the above devices are described as being divided into various units according to their functions. Of course, when implementing this specification, the functions of each unit can be implemented in the same or multiple software and / or hardware.

[0181] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware. Furthermore, the present application may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0182] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the embodiments of the present application. It should be understood that each process and / or block in the flowchart and / or block diagram, as well as the combination of processes and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 a process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0183] These computer program instructions may also be stored in a computer readable memory that can direct a computer or other programmable data processing device to work in a specific manner, so that the instructions stored in the computer readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 a process or multiple processes and / or boxes Figure 1 The function specified in one or more boxes.

[0184] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operational steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing the instructions executed on the computer or other programmable device for implementing the process. Figure 1 a process or multiple processes and / or boxes Figure 1 A step that specifies a function in one or more boxes.

[0185] In a typical configuration, a computing device includes one or more processors (CPUs), input / output interfaces, network interfaces, and memory.

[0186] Memory may include non-permanent storage in a computer-readable medium, random access memory (RAM) and / or non-volatile memory in the form of read-only memory (ROM) or flash RAM. Memory is an example of a computer-readable medium.

[0187] Computer-readable media includes permanent and non-permanent, removable and non-removable media that can be implemented by any method or technology to store information. The information can be computer-readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassettes, magnetic tape, magnetic disk storage or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined herein, computer-readable media does not include transitory computer-readable media (transitory media), such as modulated data signals and carrier waves.

[0188] It should also be noted that the terms "comprises," "includes," or any other variations thereof are intended to encompass non-exclusive inclusion, such that a process, method, commodity, or apparatus that includes a list of elements includes not only those elements but also other elements not explicitly listed, or includes elements inherent to such process, method, commodity, or apparatus. In the absence of further limitations, an element defined by the phrase "comprising a..." does not preclude the presence of other identical elements in the process, method, commodity, or apparatus that includes the element.

[0189] Those skilled in the art will appreciate that the embodiments of this specification may be provided as methods, systems, or computer program products. Therefore, this specification may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware. Furthermore, this specification may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0190] This specification may be described in the general context of computer-executable instructions, such as program modules, executed by a computer. Generally, program modules include routines, programs, objects, components, data structures, and the like that perform specific tasks or implement specific abstract data types. This specification may also be practiced in distributed computing environments where tasks are performed by remote processing devices connected through a communications network. In a distributed computing environment, program modules may be located in both local and remote computer storage media, including storage devices.

[0191] The various embodiments in this specification are described in a progressive manner. Similar parts between the various embodiments can be referred to in conjunction with each other. Each embodiment focuses on the differences between the other embodiments. In particular, the system embodiments are generally similar to the method embodiments, so the description is relatively simple. For relevant parts, refer to the description of the method embodiments.

[0192] The foregoing is merely an embodiment of the present invention and is not intended to limit the present invention. Various modifications and variations are possible for those skilled in the art. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention are intended to be included within the scope of the claims of this application.< / graph> < / graph>

Claims

1. A large model training method for gene regulatory network prediction, characterized in that: include: Constructing a gene regulatory network based on gene pairs with regulatory relationships, wherein nodes in the gene regulatory network represent genes and edges represent regulatory relationships between genes; In the gene regulatory network, a one-hop subgraph centered on a node of each gene is determined respectively; For each subgraph, perform structural perturbation on the subgraph to generate positive and negative samples of the subgraph; Obtaining, by means of a preset graph encoding model, graph embedding features of the subgraph, a positive sample of the subgraph, and a negative sample of the subgraph, respectively, and obtaining a contrast loss based on the graph embedding feature similarity between the subgraph and the positive sample and the graph embedding feature similarity between the subgraph and the negative sample; For each subgraph, using the first prompt information, the large model learns the correspondence between the graph embedding features of the subgraph and the text embedding features of each gene in the subgraph, and calculates the cross entropy loss based on the first result output by the large model, wherein the text embedding features are obtained by feature extraction of pre-acquired text description information of the genes; Adjusting the parameters of the large model according to the contrast loss and the cross entropy loss; Obtaining a sample gene pair, inputting the sample gene pair and second prompt information into the large model, and obtaining a second result output by the large model, wherein the second prompt information is used to indicate whether the large model predicts whether there is a regulatory relationship between the input gene pair; According to the difference between the second result and the true regulatory relationship of the sample gene pair, an inference loss is determined, and based on the inference loss, the parameters of the large model are adjusted again.

2. The method according to claim 1, wherein For each subgraph, perform structural perturbation on the subgraph to generate positive samples and negative samples of the subgraph, specifically including: For each subgraph, delete the edges in the subgraph at a ratio less than a first set value to obtain a positive sample of the subgraph; The edge connection relationship in the subgraph is changed at a ratio greater than a second set value to obtain a negative sample of the subgraph, wherein the first set value is less than the second set value.

3. The method according to claim 1, wherein The first prompt information includes a graph feature placeholder, and the first prompt information is used to instruct the large model to sort the gene description information in a disrupted order; The first prompt information enables the large model to learn the correspondence between the graph embedding feature of the subgraph and the text embedding feature of each gene in the subgraph, and calculates the cross entropy loss according to the first result output by the large model, specifically including: Obtaining text embedding features extracted by the large model from the first prompt information, wherein the text embedding features include a graph feature placeholder, and the graph feature placeholder is not encoded; Replacing the graph feature placeholder with the graph embedding feature to obtain a graph-text joint feature; The large model performs reasoning based on the joint features of the graph and text, and outputs a ranking prediction result of the gene description information corresponding to each node contained in the subgraph; A cross entropy loss is determined based on the difference between the ranking prediction result and the correct ranking result.

4. The method according to claim 1, wherein The first prompt information includes description information of each node in the subgraph in order, and gene description information of each node in a disrupted order.

5. The method according to claim 1, wherein The second result is a discrete confidence level, and the true regulatory relationship is represented by an existence binary label and an existence confidence level label; Determining an inference loss based on a difference between the second result and the true regulatory relationship of the sample gene pair, specifically comprising: Obtaining the predicted probability of the regulatory relationship between the sample gene pairs obtained by the large model; Determining a first loss based on a difference between the predicted probability and the binary label of the presence of the sample gene pair; mapping the presence confidence level label of the sample gene pair to a continuous value, and determining a second loss based on a difference between the predicted probability and the mapped continuous value; An inference loss is determined based on the first loss and the second loss.

6. The method according to any one of claims 1 to 5, wherein The method further comprises: The gene pairs with unknown regulatory relationships are input into the large model, and the large model outputs a prediction result representing the regulatory relationship between the gene pairs with unknown regulatory relationships.

7. A large model training device for gene regulatory network prediction, characterized in that: include: A graph construction module constructs a gene regulatory network based on gene pairs with regulatory relationships, wherein nodes in the gene regulatory network represent genes and edges represent regulatory relationships between genes; a subgraph extraction module, which determines, in the gene regulatory network, a one-hop subgraph centered on the node of each gene; The sample generation module performs structural perturbations on each subgraph to generate positive and negative samples of the subgraph. A contrastive learning module, which obtains graph embedding features of the subgraph, a positive sample of the subgraph, and a negative sample of the subgraph through a preset graph encoding model, and obtains a contrast loss based on the graph embedding feature similarity between the subgraph and the positive sample and the graph embedding feature similarity between the subgraph and the negative sample; The graph-text combination module, for each subgraph, uses the first prompt information to enable the large model to learn the correspondence between the graph embedding features of the subgraph and the text embedding features of each gene in the subgraph, and calculates the cross entropy loss based on the first result output by the large model, wherein the text embedding features are obtained by feature extraction of the pre-acquired text description information of the genes; A parameter adjustment module, adjusting the parameters of the large model according to the contrast loss and the cross entropy loss; Obtaining a sample gene pair, inputting the sample gene pair and second prompt information into the large model, and obtaining a second result output by the large model, wherein the second prompt information is used to indicate whether the large model predicts whether there is a regulatory relationship between the input gene pair; According to the difference between the second result and the true regulatory relationship of the sample gene pair, an inference loss is determined, and based on the inference loss, the parameters of the large model are adjusted again.

8. A computer-readable storage medium, characterized in that The storage medium stores a computer program, and when the computer program is executed by a processor, the method according to any one of claims 1 to 6 is implemented.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the method according to any one of claims 1 to 6 is implemented.

Images