Method and device for predicting synthetic lethal genes based on two-stage multi-modal fusion, terminal and storage medium

Abstract

The application relates to the technical field of bioinformatics. The application discloses a two-stage multi-modal fusion synthetic lethal gene prediction method and device, a terminal and a storage medium, which can improve the accuracy of predicting synthetic lethal gene pairs. The method comprises the following steps: acquiring multiple omics data of two genes; inputting the multiple omics data into a multi-modal prediction model to perform synthetic lethal gene pair prediction processing, obtaining a prediction result of the two genes, and using the prediction result to display whether the two genes belong to a synthetic lethal gene pair.

Description

Technical Field

[0001] This application relates to the field of bioinformatics technology. More specifically, this application relates to a method, apparatus, terminal, and storage medium for predicting synthetic lethal genes using a two-stage multimodal fusion approach. Background Technology

[0002] Cancer is a complex polygenic disease whose occurrence and development involve the interaction of multiple genes. Identifying gene interactions is crucial for understanding cancer pathogenesis, developing novel treatments, and discovering effective drug targets. Synthetic lethality is a phenomenon of gene interaction where a single gene abnormality is harmless to the cell, but when two genes mutate or become inactive simultaneously, it leads to cell death. Synthetic lethality, as a promising targeted anticancer therapy, has the core advantage of selectively killing cancer cells without damaging normal cells. Traditional methods for predicting synthetic lethal genes typically involve: obtaining structural features of proteins based on protein structure data; obtaining sequence features of proteins based on protein sequence data; obtaining functional features of proteins based on protein-protein interaction networks; merging and standardizing the structural, sequence, and functional features of proteins to obtain the gene features of the major gene generating the protein; obtaining gene interactions; training a synthetic lethal gene pair prediction model based on graph convolutional neural networks using gene interactions and gene features; obtaining the final feature representation of each gene based on the trained synthetic lethal gene pair prediction model; and predicting whether two genes are a synthetic lethal gene pair based on the final feature representation. The data used in this method is protein structure data, which cannot capture the cancer specificity of synthetic lethal effects, thus reducing the accuracy of predicting synthetic lethal gene pairs. Summary of the Invention

[0003] The purpose of this application is to provide a method, apparatus, terminal, and storage medium for predicting synthetic lethal genes using a two-stage multimodal fusion method, which can improve the accuracy of predicting synthetic lethal gene pairs. This application is mainly achieved through the following technical solutions: A first aspect of this application provides a method for predicting synthetic lethal genes based on a two-stage multimodal fusion, comprising: Obtain multi-omics data for two genes; The multi-omics data are input into a multimodal prediction model to predict synthetic lethal gene pairs, and the prediction results of the two genes are obtained. The prediction results are used to show whether the two genes belong to a synthetic lethal gene pair.

[0004] According to one embodiment of this application, the multiple omics data for each gene include gene expression data, gene mutation data, DNA methylation data, and copy number variation data.

[0005] According to one embodiment of this application, the step of inputting the multi-omics data into a multimodal prediction model for synthetic lethal gene pair prediction processing to obtain the prediction results of the two genes, wherein the prediction results are used to indicate whether the two genes belong to a synthetic lethal gene pair, includes: Multiple omics data corresponding to each gene are input into the omics prediction model of the multimodal prediction model for omics feature extraction processing to obtain the target omics features corresponding to each gene. The target omics features corresponding to all genes are input into the first feature fusion module of the multimodal prediction model for fusion processing to obtain the first fused feature; The first fusion feature is input into the first classifier of the multimodal prediction model to perform synthetic lethal gene pair prediction processing to obtain the prediction result.

[0006] According to one embodiment of this application, the step of inputting multiple omics data corresponding to each gene into the omics prediction model of the multimodal prediction model for omics feature extraction processing to obtain the target omics features corresponding to each gene includes: Each omics data corresponding to each gene is input into the cross encoder of the omics prediction model for mean and log variance calculation to obtain the first mean and first log variance of each omics data corresponding to each gene. Generate cross-coding features for each gene based on the first mean and first log variance of each omics data corresponding to each gene; Each omics data corresponding to each gene is input into the autoencoder of the omics prediction model for mean and log variance calculation, to obtain the second mean and second log variance of each omics data corresponding to each gene; Generate autoencoded features for each omics data corresponding to each gene based on the second mean and second log variance of each omics data corresponding to each gene; All autoencoded features and all cross-coding features are input into the splicing module of the omics prediction model for splicing processing to obtain the target splicing features corresponding to each gene; The target splicing features corresponding to each gene are input into the linear layer of the omics prediction model for dimensionality reduction to obtain the target omics features corresponding to each gene.

[0007] According to one embodiment of this application, the method for predicting synthetic lethal genes through two-stage multimodal fusion further includes: Set the first loss function; The model parameters of the original omics model are adjusted based on the first loss function to form the omics prediction model.

[0008] According to one embodiment of this application, the step of inputting the multi-omics data into a multimodal prediction model for synthetic lethal gene pair prediction processing to obtain the prediction results of the two genes, wherein the prediction results are used to indicate whether the two genes belong to a synthetic lethal gene pair, includes: The multi-omics data corresponding to each gene are input into the knowledge graph prediction model of the multimodal prediction model for graph feature extraction processing to obtain the knowledge graph features corresponding to each gene. All knowledge graph features corresponding to the genes are input into the second feature fusion module of the multimodal prediction model for fusion processing to obtain the second fused features; The second fusion feature is input into the second classifier of the multimodal prediction model to perform synthetic lethal gene pair prediction processing, and the prediction result is obtained.

[0009] According to one embodiment of this application, the knowledge graph prediction model is trained by a relational graph convolutional network.

[0010] A second aspect of this application provides a device for predicting synthetic lethal genes using a two-stage multimodal fusion method, comprising: The omics data acquisition module is used to acquire multiple omics data for two genes; The prediction module is used to input the multi-omics data into a multimodal prediction model to perform synthetic lethal gene pair prediction processing, and obtain the prediction results of the two genes. The prediction results are used to show whether the two genes belong to a synthetic lethal gene pair.

[0011] A third aspect of this application provides a terminal device, including a processor and a memory, the memory being used to store a computer program, the processor being used to call and run the computer program stored in the memory, and to execute the steps of the two-stage multimodal fusion synthetic lethal gene prediction method provided in the first aspect of this application.

[0012] A fourth aspect of this application provides a computer-readable storage medium for storing a computer program that causes a computer to perform the steps of the two-stage multimodal fusion synthetic lethal gene prediction method provided in the first aspect of this application.

[0013] The beneficial effects of the embodiments of this application include: This application embodiment acquires multi-omics data of two genes; inputs the multi-omics data into a multimodal prediction model for synthetic lethal gene pair prediction, and obtains prediction results for the two genes. These prediction results indicate whether the two genes belong to a synthetic lethal gene pair. Compared with existing technologies, this application embodiment uses multi-omics data of genes to replace protein structure data in existing technologies, thereby capturing the cancer specificity of synthetic lethal effects and significantly improving the accuracy of predicting synthetic lethal gene pairs. Attached Figure Description

[0014] To more clearly illustrate the technical solutions in the embodiments of this application or the conventional technology, the drawings used in the description of the embodiments or the conventional technology will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0015] Figure 1 The flowcharts for some embodiments of the two-stage multimodal fusion method for predicting synthetic lethal genes according to this application are shown below. Figure 2 The following is a reference diagram of the overall architecture of the two-stage multimodal fusion synthetic lethal gene prediction method of this application in some embodiments; Figure 3 The schematic diagram of the two-stage multimodal fusion synthetic lethal gene prediction device of this application is shown in some embodiments. Figure 4 This is a schematic block diagram of the terminal device of this application in some embodiments. Detailed Implementation

[0016] To make the above-mentioned objectives, features, and advantages of this application more apparent and understandable, the specific embodiments of this application are described in detail below with reference to the accompanying drawings. Many specific details are set forth in the following description to provide a thorough understanding of this application. However, this application can be implemented in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of this application. Therefore, this application is not limited to the specific embodiments disclosed below.

[0017] It should be noted that the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "multiple" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0018] The terms "exemplary" or "for example" are used to indicate that something is an example, illustration, or illustration. Any embodiment or design described as "exemplary" or "for example" in the embodiments of this application should not be construed as being more preferred or advantageous than other embodiments or designs. Specifically, the use of terms such as "exemplary" or "for example" is intended to present the relevant concepts in a specific manner.

[0019] The terms “comprising,” “including,” or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, system, product, or apparatus that includes a series of steps or units is not necessarily limited to those steps or units that are expressly listed, but may include other steps or units that are not expressly listed or that are inherent to such process, method, product, or apparatus.

[0020] Unless otherwise defined, all technical and scientific terms used in this application have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used in this application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. The term "and / or" as used in this application includes any and all combinations of one or more of the associated listed items.

[0021] The specific embodiments of this application will be further described below with reference to the accompanying drawings.

[0022] refer to Figure 1 The diagram shown is a flowchart of a two-stage multimodal fusion method for predicting synthetic lethal genes, provided in the first aspect of an embodiment of this application. Figure 1 The method for predicting synthetic lethal genes through two-stage multimodal fusion includes: S1. Obtain multi-omics data for two genes.

[0023] More specifically, each gene corresponds to multiple omics data. These multiple omics data for each gene include gene expression data, gene mutation data, DNA methylation data, and copy number variation data.

[0024] Each gene corresponds to multiple omics data, which can be expressed as... ;in, It is the first of two genes. One gene; It is the first The first type of omics data corresponding to each gene; It is the first The second type of omics data corresponding to each gene; It is the first The corresponding gene is the first Genomics data.

[0025] S2. Input the multi-omics data into a multimodal prediction model to perform synthetic lethal gene pair prediction processing, and obtain the prediction results of the two genes. The prediction results are used to show whether the two genes belong to a synthetic lethal gene pair.

[0026] Further, step S2 includes: inputting the multiple omics data corresponding to each gene into the omics prediction model of the multimodal prediction model for omics feature extraction processing to obtain the target omics features corresponding to each gene; inputting the target omics features corresponding to all genes into the first feature fusion module of the multimodal prediction model for fusion processing to obtain the first fusion feature; inputting the first fusion feature into the first classifier of the multimodal prediction model for synthetic lethal gene pair prediction processing to obtain the prediction result.

[0027] The fusion process can be understood as a merging process or a splicing process.

[0028] The first classifier can be a pre-trained support vector machine or decision tree classifier. In other embodiments, the first classifier can also be other classifiers, which can be set by those skilled in the art according to actual needs.

[0029] The prediction result is selected from either two genes that are synthetic lethal gene pairs or two genes that are not synthetic lethal gene pairs.

[0030] Further, the step of inputting the multiple omics data corresponding to each gene into the omics prediction model of the multimodal prediction model for omics feature extraction processing to obtain the target omics features corresponding to each gene includes: inputting each type of omics data corresponding to each gene into the cross encoder of the omics prediction model for mean and log-variance calculation processing to obtain the first mean and first log-variance of each type of omics data corresponding to each gene; generating cross-coding features of each type of omics data corresponding to each gene based on the first mean and first log-variance of each type of omics data corresponding to each gene; inputting each type of omics data corresponding to each gene into the cross encoder of the omics prediction model for mean and log-variance calculation processing to obtain the first mean and first log-variance of each type of omics data corresponding to each gene; and inputting the cross encoder of the omics prediction model for each gene into the target omics features. The autoencoder of the omics prediction model is processed by calculating the mean and log-variance to obtain the second mean and second log-variance of each omics data corresponding to each gene; based on the second mean and second log-variance of each omics data corresponding to each gene, autoencoded features of each omics data corresponding to each gene are generated; all autoencoded features and all cross-coding features are input into the splicing module of the omics prediction model for splicing processing to obtain the target spliced ​​features corresponding to each gene; the target spliced ​​features corresponding to each gene are input into the linear layer of the omics prediction model for dimensionality reduction processing to obtain the target omics features corresponding to each gene.

[0031] The autoencoder is the diagonal part of the VAE encoder matrix, and the cross encoder is the off-diagonal part. Each encoder is implemented identically, using a multilayer perceptron to reduce dimensionality and encode the mean and log-variance.

[0032] The target omics feature corresponding to each gene is the final omics latent variable for each gene.

[0033] Further, the step of inputting the data for each omics corresponding to each gene into the cross-encoder of the omics prediction model for mean and log-variance calculation to obtain the first mean and first log-variance of the data for each omics corresponding to each gene includes: inputting the data for each omics corresponding to each gene into the cross-encoder of the omics prediction model for mean calculation to obtain the first mean of the data for each omics corresponding to each gene; and inputting the data for each omics corresponding to each gene into the cross-encoder of the omics prediction model for log-variance calculation to obtain the first log-variance of the data for each omics corresponding to each gene.

[0034] Furthermore, the formula for calculating the first mean of each omics data corresponding to each gene by inputting it into the cross-encoder of the omics prediction model for mean calculation is as follows: ; in, It is the first The corresponding gene is the first The first mean of the omics data; It is the first The corresponding gene is the first The first log-variance of omics data; It is the first The total number of types of omics data in each gene; It is the first Each gene is processed by a cross-omics encoder The first mean obtained afterward; It is the first Each gene is processed by a cross-omics encoder The first logarithmic variance was obtained afterward.

[0035] Furthermore, and The value can be calculated using the following formula: ; in, It is the first In the gene, the _ ... omics data for the first Cross-omics encoders generated from omics data; It is the first The corresponding gene is the first Genomics data.

[0036] Furthermore, the formula for calculating the first log-variance of each omics data corresponding to each gene by inputting it into the cross-encoder of the omics prediction model for log-variance calculation is as follows: .

[0037] It should be understood that for each type of omics data corresponding to each gene, the PoE (product of experts) module can be used to fuse the variational posteriors of other omics into a joint variational posterior, and obtain the first mean and the first log variance after fusion.

[0038] Furthermore, the formula for calculating the cross-coding features of each gene's omics data based on the first mean and first log-variance of each omics data is as follows: ; in, It is the first The corresponding gene is the first Cross-coding features of omics data It represents cross-encoding (or inter-encoding); It is the first standard normal distribution noise; It is the element-wise multiplication symbol; ; Represents a distribution symbol; Represents a normal distribution; The covariance matrix is ​​represented by the identity matrix.

[0039] Further, the step of inputting the data for each omics corresponding to each gene into the autoencoder of the omics prediction model for mean and log-variance calculation to obtain the second mean and second log-variance of the data for each omics corresponding to each gene includes: inputting the data for each omics corresponding to each gene into the autoencoder for mean calculation to obtain the second mean of the data for each omics corresponding to each gene; and inputting the data for each omics corresponding to each gene into the autoencoder for log-variance calculation to obtain the second log-variance of the data for each omics corresponding to each gene.

[0040] Furthermore, the formula for calculating the second mean of each omics data corresponding to each gene by inputting it into the autoencoder for mean calculation is as follows: ; in, It is the first The corresponding gene is the first The second mean of the omics data; It is the first The corresponding gene is the first Mean weighting of omics data; It is the first The corresponding gene is the first Mean bias in omics data; It is the first in the self-encoder Embedding of multilayer perceptrons.

[0041] The calculation method can be achieved through the first autoencoder. Embedding of multilayer perceptrons The calculation formula was derived from this.

[0042] Furthermore, The calculation formula is: ; ; in, It is a linear correction unit, which is an activation function; It is the first in the self-encoder Weights of a multilayer perceptron; It is the first in the self-encoder Embedding of multilayer perceptrons; It is the first in the self-encoder Bias of multilayer perceptrons; It is the total number of layers in the multilayer perceptron of the autoencoder.

[0043] Furthermore, ;in, yes A 3D real vector space. For example, The value can be 5; in other implementations, The specific value can be set by those skilled in the art according to actual needs.

[0044] It should also be noted that the first The number of hidden units in a multilayer perceptron is denoted as . Therefore, the number of hidden units in all layers of the multilayer perceptron constitutes the hyperparameter list. .

[0045] Furthermore, the formula for calculating the second log-variance of each omics data corresponding to each gene by inputting it into the autoencoder for log-variance calculation is as follows: ; ; in, It is the first The corresponding gene is the first The second log-variance of the omics data; yes The square of; It is the first The corresponding gene is the first Log-variance weighting of omics data; It is the first The corresponding gene is the first Log-variance bias of omics data.

[0046] Furthermore, the formula for calculating the autoencoded features of each gene's corresponding omics data based on the second mean and second log-variance of each omics data is as follows: ; in, It is the first The corresponding gene is the first Autoencoding features of omics data It represents the meaning of self-encoding; It is the second standard normal distribution noise. .

[0047] The first standard normal distribution noise and the second standard normal distribution noise can be the same or different, and can be set by those skilled in the art according to actual needs.

[0048] The embodiments of this application use multi-omics data of genes to replace protein structure data in the prior art, thereby capturing the cancer specificity of synthetic lethal effects and significantly improving the accuracy of predicting synthetic lethal gene pairs.

[0049] In some implementations, the method for predicting synthetic lethal genes through two-stage multimodal fusion further includes: setting a first loss function; and adjusting the model parameters of the original omics model based on the first loss function to form the omics prediction model.

[0050] The original omics model and the omics prediction model have the same structure; the only difference between them is the model parameters.

[0051] Furthermore, the formula for calculating the first loss function is as follows: ; in, It is the first loss function; It is a binary cross-entropy loss function used to measure classification loss to ensure that the learned representation can accurately predict synthetic lethal relationships; It is the first hyperparameter, used to control the regularization strength of the autoencoder; It is the KL divergence loss function corresponding to the autoencoder; It is the second hyperparameter, used to control the regularization strength of the cross encoder; It is the KL divergence loss function corresponding to the cross encoder. and It can constrain the encoder distribution to approximate a standard normal distribution, thereby achieving representation regularization and compression. and It can be balanced and right . contributions.

[0052] It is also understood that the loss function of the omics prediction model can adopt the Variational Information Bottleneck (VIB) framework. VIB is an information-theoretic representation learning method that aims to learn feature representations that retain task-related information while being highly compressible. Its core idea is to maximize the accuracy of the prediction task while minimizing the amount of input information contained in the representation, thereby obtaining more robust and generalizable features. In this embodiment, the objective function of variational inference is transformed into the specific loss used by the omics prediction model, namely the first loss function.

[0053] In some implementations, step S2 includes: inputting the multi-omics data corresponding to each gene into the knowledge graph prediction model of the multimodal prediction model for graph feature extraction processing to obtain the knowledge graph features corresponding to each gene; inputting the knowledge graph features corresponding to all genes into the second feature fusion module of the multimodal prediction model for fusion processing to obtain the second fused features; and inputting the second fused features into the second classifier of the multimodal prediction model for synthetic lethal gene pair prediction processing to obtain the prediction result.

[0054] The knowledge graph features corresponding to each gene can be understood as the latent variables of the knowledge graph for each gene. For example, the first... The latent variables of the knowledge graph of a gene can be represented as: .

[0055] The knowledge graph prediction model is trained using a Relational Graph Convolutional Network (RGCN). In other embodiments, the knowledge graph prediction model can also be trained using a Heterogeneous Graph Attention Network (HAN) or a Graph Transformer, which can be configured by those skilled in the art according to actual needs.

[0056] The relational graph convolutional network is a deep learning model specifically designed for processing multi-relational graph data (i.e., knowledge graphs). It is an extension of the standard graph convolutional network, and its core innovation lies in its ability to distinguish different types of relations or edges in the graph.

[0057] The heterogeneous graph attention network is a graph neural network based on an attention mechanism, designed to process heterogeneous graph data.

[0058] The Graph Transformer is a neural network model that applies the Transformer architecture to graph-structured data. By integrating the basic principles of Graph Neural Networks (GNNs) with the self-attention mechanism of Transformers, it enables the processing of relationship information between nodes in the graph and the effective capture of long-range dependencies.

[0059] The knowledge graph prediction model updates node representations in the following ways: ; in, It is a node In the Representation of a multilayer perceptron; It is an activation function; It is a set of relation types, that is, all relations corresponding to all nodes; Through relationships Connect to node The set of neighboring nodes; It is a normalization constant; It is a relationship The weight matrix; It is a node In the Representation of a multilayer perceptron; It is a node Its own weight matrix; It is a node In the Representation of a multilayer perceptron.

[0060] In this embodiment of the application, the knowledge graph can be represented as ,in, It is a set of nodes; It is a set of edges.

[0061] The second classifier and the first classifier are the same type of classifier.

[0062] In some implementations, the method for predicting synthetic lethal genes using a two-stage multimodal fusion further includes: setting a second loss function; and adjusting the model parameters of the relational graph convolutional network based on the second loss function to form the knowledge graph prediction model.

[0063] The second loss function is the standard binary cross-entropy loss function.

[0064] In some implementations, the multimodal prediction model can be understood as a synthetic cancer lethality prediction model. The synthetic cancer lethality prediction model employs a two-stage training strategy. The first stage involves training an omics-based unimodal prediction model and a knowledge graph-based unimodal prediction model separately to obtain a pre-trained unimodal encoder, where the omics encoder performs early fusion of multiple omics features. The second stage, based on the pre-trained unimodal encoder, guides the training of the multimodal prediction model through knowledge distillation techniques and performs late fusion using a multimodal classifier. This part of the model is called the Uni-Modal Teacher (UMT). Figure 2 As shown in the example, this application also uses the direct averaging of single-modal prediction results as the result of the Uni-Modal Ensemble (UME) model, and finally selects the model with the higher AUPR (Area Under the Precision-Recall Curve) metric on the validation set from both UMT and UME as the final prediction model.

[0065] UMT aims to address the modality inertia problem, ensuring that multimodal prediction models can fully learn the unimodal features of each modality. The core idea of ​​UMT is to use a pre-trained and frozen unimodal model as a teacher, guiding the training of the multimodal student model through feature-level knowledge distillation. Specifically, UMT uses a frozen pre-trained encoder (the teacher) to obtain the multimodal pre-trained embeddings of genes. and The multimodal embeddings of the genes are obtained using the newly initialized encoder (student). and .in, and These are all multimodal pre-trained embeddings of genes; more specifically... It is the first The corresponding gene is the first omics data in The coding features, It is the first The corresponding gene is the first omics data in The coding features; It is the first The corresponding gene is the first omics data; In the pre-trained encoder, the first The corresponding gene is the first Transformation functions corresponding to omics data This represents the pre-trained encoder (teacher model). In the pre-trained encoder, the first The corresponding gene is the first Transformation functions corresponding to omics data; It is the first of two genes. One gene; and These are all multimodal embeddings of genes; more specifically, It is the first The corresponding gene is the first omics data in The coding features, It is the first The corresponding gene is the first omics data in The coding features; This represents the newly initialized encoder (student model). In the newly initialized encoder, the first... The corresponding gene is the first Transformation functions corresponding to omics data; In the newly initialized encoder, the first... The corresponding gene is the first Transformation functions corresponding to omics data.

[0066] Since the embodiments of this application use a frozen pre-trained encoder, the third loss function of UMT is generated by the student model.

[0067] The formula for calculating the third loss function is as follows: ; in, It is the third loss function; It is the third hyperparameter, used to balance the magnitude of the loss; It is the characteristic-level distillation loss function.

[0068] Furthermore, The calculation formula is: ; in, It stands for Mean Squared Error. It is the first The corresponding gene is the first omics data in The coding features; It is the first The corresponding gene is the first omics data in The coding features; It is the first The corresponding gene is the first omics data in The coding features; It is the first The corresponding gene is the first omics data in The coding features.

[0069] UME adopts a completely different approach from UMT. Specifically, it obtains the target value by directly combining the prediction results of a single mode to minimize cross-modal interactions. The specific formula is as follows: ;in, It is the target value; It is the first The first gene and the first The corresponding gene is the first Single-modal prediction results from omics data; It is the first The first gene and the first The corresponding gene is the first The single-modal prediction results for multi-omics data. UME essentially freezes the entire model, including the encoder and classifier, so there is no need to retrain it. UME completely avoids the problem of modal inertia because each modality is trained independently, making it particularly suitable for scenarios where cross-modal interactions may have negative effects.

[0070] Considering the varying effectiveness of cross-modal interactions across different datasets, this application employs a data-driven adaptive selection strategy. After training, the AUPR metrics of UMT and UME on the validation set are evaluated separately, and the method with better performance is selected as the final model. UMT typically performs better when both omics and knowledge graph features are strong and their interactions provide additional information. When one modality is clearly dominant, or when cross-modal interactions introduce noise, UME's simple ensemble strategy is more effective.

[0071] Compared to traditional complex fusion methods, the two-stage multimodal fusion method for predicting synthetic lethal genes provided in this application is simple to implement and easy to optimize. UMT only requires tuning one additional hyperparameter, also known as the third hyperparameter, while UME does not even require additional training. This simplicity not only improves the practicality of the method but also enhances its portability on new datasets.

[0072] It should also be understood that the embodiments of this application, by introducing an omics encoder to perform early fusion of multi-omics data, can enhance the model's ability to extract omics features. Simultaneously, the embodiments of this application design a multimodal encoder to fuse heterogeneous data from multiple sources to obtain richer representations. Furthermore, since there are differences in the convergence speed between omics data and knowledge graphs during joint training, this imbalance can lead to the fast-converging modal dominating the learning process, thereby inhibiting the full learning of the slow-converging modalities. Therefore, the embodiments of this application significantly improve the overall performance of the model by optimizing the convergence strategy of joint training and balancing the learning efficiency among modalities.

[0073] refer to Figure 3 The diagram shown is a schematic block diagram of a two-stage multimodal fusion synthetic lethal gene prediction device provided in the second aspect of an embodiment of this application. Figure 3 The two-stage multimodal fusion synthetic lethal gene prediction device 100 includes: The omics data acquisition module 101 is used to acquire multiple omics data of two genes; The prediction module 102 is used to input the multi-omics data into a multimodal prediction model to perform synthetic lethal gene pair prediction processing, and obtain the prediction results of the two genes. The prediction results are used to show whether the two genes belong to a synthetic lethal gene pair.

[0074] In some embodiments, the prediction module 102 includes: an omics feature extraction and processing module, a first fusion processing module, and a first prediction submodule. The omics feature extraction and processing module is used to input multiple omics data corresponding to each gene into the omics prediction model of the multimodal prediction model for omics feature extraction processing to obtain target omics features corresponding to each gene. The first fusion processing module is used to input the target omics features corresponding to all genes into the first feature fusion module of the multimodal prediction model for fusion processing to obtain a first fused feature. The first prediction submodule is used to input the first fused feature into the first classifier of the multimodal prediction model for synthetic lethal gene pair prediction processing to obtain the prediction result.

[0075] In other embodiments, the prediction module 102 includes a graph feature extraction and processing module, a second fusion processing module, and a second prediction submodule. The graph feature extraction and processing module is used to input multiple omics data corresponding to each gene into the knowledge graph prediction model of the multimodal prediction model for graph feature extraction processing to obtain knowledge graph features corresponding to each gene. The second fusion processing module is used to input the knowledge graph features corresponding to all genes into the second feature fusion module of the multimodal prediction model for fusion processing to obtain a second fused feature. The second prediction submodule is used to input the second fused feature into the second classifier of the multimodal prediction model for synthetic lethal gene pair prediction processing to obtain the prediction result.

[0076] In some embodiments, the omics feature extraction and processing module includes a first computational processing module, a first generation module, a second computational processing module, a second generation module, a splicing processing module, and a dimensionality reduction processing module. The first computational processing module is used to input the data for each omics corresponding to each gene into the cross-encoder of the omics prediction model for mean and log-variance calculation, obtaining a first mean and a first log-variance of the data for each omics corresponding to each gene. The first generation module is used to generate cross-coding features for each omics corresponding to each gene based on the first mean and the first log-variance of the data for each omics corresponding to each gene. The second computational processing module is used to input the data for each omics corresponding to each gene into the cross-encoder of the omics prediction model for mean and log-variance calculation, obtaining a first mean and a first log-variance of the data for each omics corresponding to each gene. The autoencoder of the omics prediction model is input for mean and log-variance calculation to obtain the second mean and second log-variance of each omics data corresponding to each gene; the second generation module is used to generate autoencoded features of each omics data corresponding to each gene based on the second mean and second log-variance of each omics data corresponding to each gene; the splicing processing module is used to input all autoencoded features and all cross-coding features into the splicing module of the omics prediction model for splicing processing to obtain the target spliced ​​features corresponding to each gene; the dimensionality reduction processing module is used to input the target spliced ​​features corresponding to each gene into the linear layer of the omics prediction model for dimensionality reduction processing to obtain the target omics features corresponding to each gene.

[0077] In some embodiments, the two-stage multimodal fusion synthetic lethal gene prediction device further includes a setting module and an adjustment module, wherein the setting module is used to set a first loss function; and the adjustment module is used to adjust the model parameters of the original omics model based on the first loss function to form the omics prediction model.

[0078] A third aspect of this application provides a terminal device, the schematic diagram of which is as follows: Figure 4As shown. The terminal device includes a processor, memory, network interface, display screen, and temperature sensor connected via a system bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The network interface of the terminal device is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements a method for predicting synthetic lethal genes using a two-stage multimodal fusion approach. The display screen can be a liquid crystal display or an e-ink display, and the temperature sensor is pre-installed inside the terminal device to detect the operating temperature of the internal components.

[0079] Those skilled in the art will understand that Figure 4 The schematic diagram shown is only a partial structural diagram related to the present invention and does not constitute a limitation on the terminal device to which the present invention is applied. The specific terminal device may include more or fewer components than shown in the figure, or combine certain components, or have different component arrangements.

[0080] In some embodiments, this application provides a terminal device including a processor and a memory. The memory stores a computer program, and the processor calls and runs the computer program stored in the memory to perform the steps of the two-stage multimodal fusion synthetic lethal gene prediction method provided in the first aspect of this application.

[0081] A fourth aspect of this application provides a computer-readable storage medium for storing a computer program that causes a computer to perform the steps of the two-stage multimodal fusion synthetic lethal gene prediction method provided in the first aspect of this application.

[0082] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided by this invention can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0083] The technical features of the above embodiments can be combined without changing the basic principles of this application. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0084] The above embodiments merely illustrate several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the patent protection scope of this application should be determined by the appended claims.

Claims

1. A method for predicting synthetic lethal genes through a two-stage multimodal fusion, characterized in that, include: Obtain multi-omics data for two genes; The multi-omics data are input into a multimodal prediction model to predict synthetic lethal gene pairs, and the prediction results of the two genes are obtained. The prediction results are used to show whether the two genes belong to a synthetic lethal gene pair.

2. The method for predicting synthetic lethal genes based on two-stage multimodal fusion according to claim 1, characterized in that, Multiple omics data for each gene include gene expression data, gene mutation data, DNA methylation data, and copy number variation data.

3. The method for predicting synthetic lethal genes based on two-stage multimodal fusion according to claim 1, characterized in that, The step of inputting the multi-omics data into a multimodal prediction model for synthetic lethal gene pair prediction to obtain the prediction results of the two genes, wherein the prediction results are used to indicate whether the two genes belong to a synthetic lethal gene pair, includes: Multiple omics data corresponding to each gene are input into the omics prediction model of the multimodal prediction model for omics feature extraction processing to obtain the target omics features corresponding to each gene. The target omics features corresponding to all genes are input into the first feature fusion module of the multimodal prediction model for fusion processing to obtain the first fused feature; The first fusion feature is input into the first classifier of the multimodal prediction model to perform synthetic lethal gene pair prediction processing, and the prediction result is obtained.

4. The method for predicting synthetic lethal genes based on two-stage multimodal fusion according to claim 3, characterized in that, The steps of inputting the multiple omics data corresponding to each gene into the omics prediction model of the multimodal prediction model for omics feature extraction processing to obtain the target omics features corresponding to each gene include: Each omics data corresponding to each gene is input into the cross encoder of the omics prediction model for mean and log variance calculation to obtain the first mean and first log variance of each omics data corresponding to each gene. Generate cross-coding features for each gene based on the first mean and first log variance of each omics data corresponding to each gene; Each omics data corresponding to each gene is input into the autoencoder of the omics prediction model for mean and log variance calculation, to obtain the second mean and second log variance of each omics data corresponding to each gene; Generate autoencoded features for each omics data corresponding to each gene based on the second mean and second log variance of each omics data corresponding to each gene; All autoencoded features and all cross-coding features are input into the splicing module of the omics prediction model for splicing processing to obtain the target splicing features corresponding to each gene; The target splicing features corresponding to each gene are input into the linear layer of the omics prediction model for dimensionality reduction to obtain the target omics features corresponding to each gene.

5. The method for predicting synthetic lethal genes based on two-stage multimodal fusion according to claim 3, characterized in that, The method for predicting synthetic lethal genes through two-stage multimodal fusion also includes: Set the first loss function; The model parameters of the original omics model are adjusted based on the first loss function to form the omics prediction model.

6. The method for predicting synthetic lethal genes based on two-stage multimodal fusion according to claim 1, characterized in that, The step of inputting the multi-omics data into a multimodal prediction model for synthetic lethal gene pair prediction to obtain the prediction results of the two genes, wherein the prediction results are used to indicate whether the two genes belong to a synthetic lethal gene pair, includes: The multi-omics data corresponding to each gene are input into the knowledge graph prediction model of the multimodal prediction model for graph feature extraction processing to obtain the knowledge graph features corresponding to each gene. All knowledge graph features corresponding to the genes are input into the second feature fusion module of the multimodal prediction model for fusion processing to obtain the second fused features; The second fusion feature is input into the second classifier of the multimodal prediction model to perform synthetic lethal gene pair prediction processing, and the prediction result is obtained.

7. The method for predicting synthetic lethal genes based on two-stage multimodal fusion according to claim 6, characterized in that, The knowledge graph prediction model is trained using a relational graph convolutional network.

8. A two-stage, multimodal fusion-based predictive device for synthetic lethal genes, characterized in that, include: The omics data acquisition module is used to acquire multiple omics data for two genes; The prediction module is used to input the multi-omics data into a multimodal prediction model to perform synthetic lethal gene pair prediction processing, and obtain the prediction results of the two genes. The prediction results are used to show whether the two genes belong to a synthetic lethal gene pair.

9. A terminal device, characterized in that, include: A processor and a memory, the memory for storing a computer program, the processor for calling and running the computer program stored in the memory to perform the steps of the method for predicting synthetic lethal genes by two-stage multimodal fusion as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, Used to store a computer program that causes a computer to perform the steps of the method for predicting synthetic lethal genes by two-stage multimodal fusion as described in any one of claims 1 to 7.

Images