Data processing device and method for predicting essential genes of non-model organisms and application

Abstract

The application discloses a data processing device and method for predicting essential genes of non-model organisms in the field of bioinformatics and application. The technical problem to be solved by the application is how to screen essential genes of non-model organisms. In the application, the protein sequences of a target species and the protein sequences of n model organism species are respectively based on annotation to obtain the corresponding labels of each protein; then the protein sequences and the labels are embedded into a protein language model to output embedding vectors; based on the embedding vectors, a residue contact map is constructed for each protein; the features of each amino acid residue in the residue contact map are modeled and integrated using a graph neural network to output a whole protein embedding vector; all protein embedding vectors are input into a classifier model for training and using an activation function to output essential probability values of protein genes; and based on the classifier model, essential probability values P1 of the genes of the proteins are obtained. The application can be applied to agricultural disease prevention and control and biological pesticide development.

Description

Technical Field

[0001] This invention relates to data processing devices, methods, and applications for predicting essential genes in non-model organisms in the field of bioinformatics. Background Technology

[0002] Fusarium graminearum ( Fusarium graminearum Fusarium head blight (FHB) is one of the major pathogenic fungi causing wheat scab globally. Its toxins, such as deoxynivalenol (DON), produced during infection not only seriously threaten crop yields but also endanger food safety. Currently, the main control method still relies on chemical pesticides, but long-term use has led to drug resistance in the pathogen, exacerbating the difficulty of control. Therefore, developing novel RNA biopesticide targets has become an urgent technical challenge. In target screening, essential genes, due to their core role in maintaining life and reproduction, are potential ideal targets, especially valuable in tumor therapy, antimicrobial drug development, and agricultural pest and disease control. Traditionally, the identification of essential genes still relies on experimental methods, such as gene knockout or RNA interference (RNAi). However, these methods generally face the following problems in application:

[0003] 1) Many non-model organisms are difficult to culture stably or have their genes edited under laboratory conditions;

[0004] 2) Lack of high-quality genome annotation, protein-protein interaction information, and other omics resources;

[0005] 3) The experimental operation is costly and time-consuming, making it difficult to scale up for systematic screening of multiple species.

[0006] To overcome the limitations of experimental methods, researchers have recently attempted to use machine learning and deep learning techniques for computational prediction of essential genes. Typical strategies include extracting physicochemical features from gene sequences to construct classifiers or building models based on protein-protein interaction networks (PPIs) to assess gene importance. These methods have achieved some success in certain model organisms, but they generally suffer from the following shortcomings:

[0007] 1) Its high dependence on manually extracted features and high-quality protein interaction network data limits its applicability in non-model species;

[0008] 2) While sequence-based models improve versatility, they generally lack the ability to provide biological explanations for the prediction mechanisms.

[0009] 3) Most models only perform well within the training species, and have poor generalization ability for cross-species "zero-shot" prediction;

[0010] 4) Existing models have not yet met the high reliability and practicality standards required for pesticide target screening.

[0011] Currently, there are no publicly reported models that can combine cross-species generalization prediction capabilities with biological interpretability in the absence of gene annotations and PPI networks. Summary of the Invention

[0012] The technical problem to be solved by this invention is how to screen essential genes of non-model organisms and / or how to screen essential genes of organisms lacking high-quality gene annotation and / or how to screen target genes of RNA pesticides and / or how to screen essential genes of pests or plant pathogens.

[0013] To address the aforementioned technical problems, the present invention first provides a data processing apparatus, which includes a memory, a processor, and a computer program stored in the memory. The processor executes the computer program to perform the following steps:

[0014] A1) Obtain the corresponding label for each protein based on the protein sequence of the target species and the protein sequences of n model organism species, respectively, using annotation.

[0015] A2) Embed each protein sequence and its corresponding tag into a (pre-trained) protein language model, and output the context embedding vector of the amino acid residues in the protein; construct a residue contact map for each protein based on the context embedding vector;

[0016] A3) Use a graph neural network to model the residue contact map, so as to integrate the features of each amino acid residue in the residue contact map and output the overall protein embedding vector;

[0017] A4) Input the overall embedding vector of all the proteins into the classifier model for training, and use the activation function to output the necessary probability value of the gene corresponding to the protein as the necessary probability value of the gene corresponding to the protein; obtain the necessary probability value P1 of the gene corresponding to the protein of the target species based on the classifier model.

[0018] The target species is a non-model organism or a species with incomplete gene annotation; n is greater than or equal to 1. The protein sequence refers to the amino acid sequence of the protein.

[0019] In one specific embodiment of the present invention, the data processing device is a device for predicting essential genes in non-model organisms (or organisms with incomplete gene annotation).

[0020] In one specific embodiment of the present invention, n is 5. The labels include two categories: "essential" and "non-essential." "Essential" indicates that the gene encoding the protein is an essential gene, and "non-essential" indicates that the gene encoding the protein is a non-essential gene. Essential genes are genes indispensable for the growth, development, and reproduction of an organism. These genes are crucial for maintaining the basic functions and survival of an organism; their absence or mutation often leads to serious physiological defects or even death. Non-essential genes are genes in an organism's genome other than essential genes.

[0021] The tags for the model organism species can be obtained based on the OGEE and DEG databases. The tags for the target species can be obtained based on the PHI-base database. The tags for protein genes annotated as "lethal" or "reduced virulence" in the PHI-base database may be essential (genes), while the tags for protein genes that are not annotated and have no functional records in the PHI-base database may be non-essential (genes).

[0022] In the above-mentioned data processing apparatus, the steps may further include:

[0023] A5) Multi-omics feature fusion: Construct a protein-protein interaction network for the target species; based on omics data, fuse the omics features of the nodes in the protein-protein interaction network to obtain a multi-omics protein-protein interaction network; use a graph neural network model to model the multi-omics protein-protein interaction network to nonlinearly aggregate the omics features of each node in the multi-omics protein-protein interaction network; use an activation function to output the graph neural network model as the required probability value P2 of the gene corresponding to the protein;

[0024] A6) Output: Determine whether the gene corresponding to each protein in the target species is an essential gene based on the required probability values ​​P1 and P2.

[0025] In the aforementioned data processing apparatus, the graph neural network described in A3) and A5) can be a graph attention network (GAT).

[0026] A3) The graph neural network mentioned can also be a graph convolutional network (GCN) or other graph neural networks such as GraphSAGE (Graph Sample and Aggregated).

[0027] In A4 above, adversarial examples can also be introduced into the protein embedding vector before the training is performed to enhance the robustness of the classifier model to unknown perturbations.

[0028] The omics data may include gene expression profiles corresponding to the nodes, conservation scores of the nodes, and / or subcellular localization data of the nodes. The omics data may also include other omics data such as proteomics and / or metabolomics.

[0029] In the aforementioned data processing device, the classifier model is optimized using the cross-entropy loss function.

[0030] In the aforementioned data processing apparatus, the protein interaction network can be constructed using an existing database. In one specific embodiment of the present invention, the database is STRING. The database can also be a database such as BioGRID or IntAct.

[0031] The residue contact graph described above can be composed of nodes and edges between nodes. The nodes can be amino acid residues, and the features of the nodes can be obtained based on the embedding vector. The edges can represent (amino acid) residue pairs with high attention scores, and the edges can be confirmed based on the parameters of the protein language model, which can be the attention weight matrix and the residue sequence spacing.

[0032] In one specific embodiment of the present invention, the protein language model is ESM-2. The protein language model may also be other protein language models such as ProtTrans or ESM-C.

[0033] The probability value ranges from 0 to 1.

[0034] In A3) above, the features of each node (amino acid residue) in the residue contact graph can be the embedding vector described in A2). Each layer of the graph neural network aggregates neighbor node information through trainable attention weights, outputting multi-level embeddings. The graph neural network ultimately generates the protein embedding vector through global pooling.

[0035] The aforementioned data processing device can be used to predict essential genes in non-model organisms or organisms with incomplete gene annotation.

[0036] To address the aforementioned technical problems, the present invention also provides a method for predicting essential genes in non-model organisms or organisms with incomplete gene annotation, the method comprising:

[0037] B1) Obtain the corresponding label for each protein based on the protein sequence of the target species and the protein sequences of n model organism species, respectively, using annotation.

[0038] B2) Embed each protein sequence and its corresponding tag into a (pre-trained) protein language model, and output the context embedding vector of the amino acid residues in the protein; construct a residue contact map for each protein based on the context embedding vector;

[0039] B3) Use a graph neural network to model the residue contact map, so as to integrate the features of each amino acid residue in the residue contact map and output the overall protein embedding vector;

[0040] B4) Input the overall embedding vector of all the proteins into the classifier model for training, and use an activation function to output the necessary probability value of the gene corresponding to the protein as the classifier model; obtain the necessary probability value P1 of the gene corresponding to the protein of the target species based on the classifier model; the target species is a non-model organism species or a species with incomplete gene annotation; n is greater than or equal to 1. The protein sequence refers to the amino acid sequence of the protein.

[0041] In one specific embodiment of the present invention, n is 5. The labels include two categories: "essential" and "non-essential." "Essential" indicates that the gene encoding the protein is an essential gene, and "non-essential" indicates that the gene encoding the protein is a non-essential gene. Essential genes are genes indispensable for the growth, development, and reproduction of an organism. These genes are crucial for maintaining the basic functions and survival of an organism; their absence or mutation often leads to serious physiological defects or even death. Non-essential genes are genes in an organism's genome other than essential genes.

[0042] The tags for the model organism species can be obtained based on the OGEE and DEG databases. The tags for the target species can be obtained based on the PHI-base database. The tags for protein genes annotated as "lethal" or "reduced virulence" in the PHI-base database may be essential (genes), while the tags for protein genes that are not annotated and have no functional records in the PHI-base database may be non-essential (genes).

[0043] In some embodiments of the present invention, the method is not intended to obtain disease diagnosis results or health status.

[0044] In some embodiments of the present invention, the method does not include the step of obtaining biological samples from an animal. All methods do not target living human or animal bodies, but only data. All methods can be information processing methods in which all steps are implemented by a data processing device such as a computer.

[0045] The method described above may also include the following steps:

[0046] B5) Multi-omics feature fusion: Construct a protein-protein interaction network for the target species, and fuse (add) omics features to the nodes of the protein-protein interaction network based on omics data to obtain a multi-omics protein-protein interaction network; model the multi-omics protein-protein interaction network using a graph neural network model to nonlinearly aggregate the omics features of each node in the multi-omics protein-protein interaction network; use an activation function to output the graph neural network model as the required probability value P2 of the gene corresponding to the protein;

[0047] B6) Output: Determine whether the gene corresponding to each protein in the target species is an essential gene based on the required probability values ​​P1 and P2.

[0048] In the above methods, the graph neural network described in B3) and B5) can be a graph attention network (GAT).

[0049] B3) The graph neural network mentioned can also be a graph convolutional network (GCN) or other graph neural networks such as GraphSAGE (Graph Sample and Aggregated).

[0050] In B4 above, adversarial examples can also be introduced into the protein embedding vector before the training is performed to enhance the robustness of the classifier model to unknown perturbations.

[0051] The omics data may include gene expression profiles corresponding to the nodes, conservation scores of the nodes, and / or subcellular localization data of the nodes. The omics data may also include other omics data such as proteomics and / or metabolomics.

[0052] In the above method, the classifier model is optimized using the cross-entropy loss function.

[0053] In the above method, the protein interaction network can be constructed using an existing database. In one specific embodiment of the present invention, the database is STRING. The database can also be a database such as BioGRID or IntAct.

[0054] The residue contact graph described above can be composed of nodes and edges between nodes. The nodes can be amino acid residues, and the features of the nodes can be obtained based on the embedding vector. The edges can represent (amino acid) residue pairs with high attention scores, and the edges can be confirmed based on the parameters of the protein language model, which can be the attention weight matrix and the residue sequence spacing.

[0055] The residue contact graph consists of nodes and edges. Nodes represent amino acid residues, and their features are based on residue-level embedding vectors generated by the protein language model. Edges represent residue pairs with high attention scores, which are constructed based on the attention weight matrix and sequence distance between residues in the protein language model, and are used to reflect the potential structural associations between residues.

[0056] In one specific embodiment of the present invention, the protein language model is ESM-2. The protein language model may also be other protein language models such as ProtTrans or ESM-C.

[0057] The probability value is in the range of 0-1.

[0058] In B3) above, the features of each node (amino acid residue) in the residue contact graph can be the embedding vector described in B2). Each layer of the graph neural network aggregates neighbor node information through trainable attention weights, outputting multi-level embeddings. The graph neural network ultimately generates the protein embedding vector through global pooling.

[0059] In one specific embodiment of the present invention, the target species is Fusarium graminearum.

[0060] To address the aforementioned technical problems, the present invention also provides a computer program product, comprising a computer program that, when executed by a processor, can implement steps A1)-A4) or steps A1)-A6) above.

[0061] To address the aforementioned technical problems, the present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, can implement steps A1)-A4) or steps A1)-A6) as described above.

[0062] The following applications of the data processing apparatus and / or the computer program product and / or the computer-readable storage medium described above also fall within the scope of protection of this invention:

[0063] C1) Application in predicting targets of RNA interference in pests or pathogens;

[0064] Application of C2 in pesticide target screening;

[0065] C3) Application in agricultural disease control and / or development or preparation of biological pesticides.

[0066] To overcome the aforementioned technical bottlenecks, this invention proposes a prediction workflow that integrates protein language models and graph attention networks—ProGATE ( Pro tein Language Model +G raph At attention + E The ProGATE system integrates the sequence context embedding capabilities of the large-scale protein language model ESM-2 with the structural modeling advantages of graph attention networks (GAT). It enables cross-species prediction of essential genes in non-model pathogens (such as Fusarium graminearum) without relying on high-quality gene annotation or PPI networks. ProGATE exhibits particularly good generalization performance under a "zero-sample" setting. Furthermore, by introducing a multi-head attention mechanism and adversarial training strategy, ProGATE not only enhances the model's robustness but also provides interpretive results based on key regions such as amino acid residues and structural domains, offering highly reliable support for target gene screening of RNA biopesticides.

[0067] This invention addresses the challenges faced by non-model plant pathogenic fungi such as Fusarium graminearum in the identification of essential genes, including strong tag dependence, poor generalization ability, and lack of biological explanation. It proposes a general prediction workflow—ProGATE—that integrates a protein language model and a multi-omics graph attention network to solve the following key technical problems:

[0068] 1) The problem of insufficient prediction ability of essential genes in non-model species scenarios: Construct a computational prediction method that does not rely on high-quality genome annotation and protein interaction network, and can carry out effective and relatively accurate prediction in the context of scarce omics data.

[0069] 2) Cross-species generalization prediction under “zero-sample” conditions: Based on the protein sequence itself, the residue-level contextual semantic embedding is extracted from the amino acid sequence using a pre-trained protein language model (ESM-2), achieving “zero-sample” generalization prediction without prior knowledge of the target species.

[0070] 3) Insufficient model interpretability: Traditional deep learning models are mostly "black box" structures, making it difficult to trace the biological mechanisms behind the prediction results. This invention introduces the attention mechanism from graph neural networks, enabling the model's judgment results to be traced back to specific amino acid residues, structural domains, or functional regions, thereby improving the biological interpretability of the prediction results and the scientific credibility of target screening.

[0071] 4) Insufficient ability to integrate multi-omics information: Integrate protein interaction networks with external omics information such as expression profiles, conservation indices, and subcellular localization as multi-dimensional inputs, construct graph attention network models, realize in-depth modeling of gene "external attributes", and improve the comprehensiveness of the model.

[0072] 5) The problem of low reliability of target screening results: This invention integrates the results of internal protein characterization prediction with external multi-omics graph attention network to construct an end-to-end essential gene prediction and screening process, which enhances the reliability of prediction and the consistency of results, and provides high reliability support for the mining of RNA biopesticide target genes.

[0073] This invention proposes a modular prediction system called ProGATE, which integrates a protein language model and a graph attention neural network. It is particularly suitable for cross-species prediction in contexts where omics data is missing or insufficient (e.g., Fusarium graminearum). The overall structure of the system includes the following five core modules and their implementation steps:

[0074] A. Data processing module.

[0075] Training species selection: This invention uses Fusarium graminearum as the target predictive species, and introduces model organisms with standard annotations, such as humans (Homo sapiens), Saccharomyces cerevisiae, fruit flies, Caenorhabditis elegans, and Escherichia coli, as training and validation samples to enhance the model's cross-species generalization ability.

[0076] 1) Protein sequence collection and standardization:

[0077] The protein sequences of Fusarium graminearum were obtained from FungiDB 20, totaling 14,143.

[0078] The protein sequences of the model species were obtained from the UniProt database 21;

[0079] All sequences are encoded in a uniform format and redundancy is removed to ensure the diversity and representativeness of the training samples.

[0080] 2) Tag construction:

[0081] For Fusarium graminearum, genes annotated as "lethal" or "reduced virulence" based on the PHI-base database22 were used as weakly supervised positive classes (such as Chs3b and TRI5).

[0082] Proteins that are not annotated and have no functional records are classified as negative (not required).

[0083] The model species were constructed using required / non-required label pairs 23 and 24 from the OGEE and DEG databases as supervised training samples.

[0084] B. Protein Language Model Feature Generation Module.

[0085] 1) Sequence embedding extraction:

[0086] The processed FASTA sequence was input into the ESM-2 pre-trained protein language model;

[0087] Output the context embedding vector for each amino acid residue, capturing its structural and functional semantic information.

[0088] 2) Residue contact diagram inference:

[0089] Based on the model attention weights and the distance between residues, an undirected graph structure is constructed: nodes are residues, and edges represent residue pairs with high attention scores;

[0090] Contact diagrams reflect the potential structural information inside proteins, providing an input basis for graph neural network modeling.

[0091] C. Graph Neural Network Modeling Module.

[0092] 1) Network structure and modeling strategy:

[0093] Graph Attention Network (GAT) is used to model the residue contact graph, introducing a multi-head attention mechanism to capture the strength of heterogeneous connections between residues.

[0094] 2) Information dissemination and embedded representation:

[0095] The node features in the diagram are ESM embedding vectors;

[0096] Each layer of the network aggregates neighbor information through trainable attention weights and outputs multi-layered embeddings;

[0097] Finally, protein-level embedding representations are generated through global pooling.

[0098] 3) Biological explanatory ability:

[0099] After the model is trained, the key residues that contribute the most to the prediction results are identified by analyzing the distribution of attention weights, which helps in target screening and structural interpretation.

[0100] D. Classification and adversarial training module.

[0101] 1) Classifier structure design:

[0102] A two-layer perceptron (MLP) is used as the classifier, and the input is the protein embedding output by the graph neural network;

[0103] The output is the "required" probability value, ranging from 0 to 1, and the activation function is sigmoid.

[0104] 2) Training loss function design:

[0105] The standard binary cross-entropy loss function is used for training, without introducing class weighting or focus loss, to ensure versatility for multi-species scenarios.

[0106] 3) Introduction of adversarial training mechanism:

[0107] Introducing gradient-direction-based perturbations into the embedding space to generate adversarial examples enhances the model's robustness to unknown perturbations.

[0108] The total loss function is the weighted sum of the losses of the original samples and the adversarial samples.

[0109] E. Multi-omics feature fusion prediction module.

[0110] 1) Integration of graph structure and external features:

[0111] A PPI network of Fusarium graminearum was constructed using the STRING database, where nodes represent proteins and edges represent functional or physical interactions between proteins.

[0112] Each node integrates the following omics features: gene expression profile (RNA-Seq TPM value), cross-species conservation score (e.g., based on protein homology), and subcellular localization information (e.g., cell nucleus, mitochondria, etc.) from databases such as UniProt and COMPARTMENTS.

[0113] 2) Graph attention network modeling:

[0114] The GAT model is used to propagate information on the PPI graph, enabling nonlinear aggregation of multi-omics features;

[0115] The final output is the predicted probability P_Pro∈[0,1].

[0116] 3) Dual-path cross-validation and fusion mechanism:

[0117] The system integrates two independent prediction methods: 1) Protein language model: based on the internal contact graph of a single sequence, outputting a prediction score P_Pro; 2) Multi-omics graph attention network: based on the biological interaction graph structure and multi-omics features, outputting P_GAT.

[0118] The following decision-making strategy is adopted: if the scores of both paths are higher than the dynamic threshold (e.g., 0.7), they are identified as highly reliable essential genes; if the difference between the two scores is large (e.g., greater than 0.5), it is marked as a model prediction discrepancy, and further experiments are recommended for confirmation; if both are lower than the threshold (e.g., 0.3), they are identified as non-essential genes.

[0119] Alternative technologies:

[0120] Pre-trained protein language model replacement: ESM-2 can be replaced by ProtTrans or ESM-C, etc.

[0121] Graph neural network model replacement: GCN or GraphSAGE, etc., can be selected;

[0122] Protein interaction extension: can be replaced with BioGRID or IntAct network structures;

[0123] Replacement of multiple omics sources: such as introducing other data types like proteomics and metabolomics.

[0124] This invention belongs to the interdisciplinary application field of bioinformatics and artificial intelligence, specifically involving a method for predicting essential genes based on a combination of protein language models and graph neural networks. This method is mainly applied to target gene screening for RNA biopesticides, and is particularly suitable for high-throughput identification of gene functions and precise target discovery in non-model pathogens (such as Fusarium graminearum) or agricultural pests with scarce gene annotations. It can be widely applied in agricultural disease control, biopesticide development, and other fields.

[0125] The beneficial effects of this invention are:

[0126] (1) It has wide applicability and breaks through the modeling limitations of non-model organisms.

[0127] The pre-trained protein language model-based method employed in this invention relies solely on the amino acid sequence of the target gene for modeling, eliminating the need for 3D structural analysis, experimental knockout data, or protein-protein interaction network information, thus significantly reducing dependence on prior data. This characteristic allows the method to be widely applied to non-model organisms such as Fusarium graminearum, and is particularly suitable for plant pathogenic fungi and agricultural pests lacking high-throughput functional annotation. Compared to traditional modeling methods that rely on PPI networks and experimental tag-driven approaches, this invention is more practical and has greater potential for application and promotion.

[0128] (2) The prediction accuracy is significantly improved, which is better than the mainstream deep learning model.

[0129] The protein-language model-based prediction method demonstrates superior performance across multiple model species, including *Caenorhabditis elegans*, *Homo sapiens*, and *Saccharomyces cerevisiae*. Experimental results show that the method achieves an average AUC of over 0.87, significantly outperforming traditional deep learning models (such as CNN and BiLSTM) and the standard Transformer architecture (performance improvement of 15–22%). Particularly noteworthy is the systematic evaluation on *Fusarium graminearum* and *Homo sapiens* datasets, where the Graph Attention Network (GAT) module, integrating multi-omics information, surpasses Graph Convolutional Networks (GCN), Node2Vec, and traditional machine learning methods (SVM or RF) on several key metrics. These results strongly demonstrate the stability and accuracy advantages of this method in multi-species applications.

[0130] (3) It has cross-species generalization ability and is suitable for "zero sample" prediction scenarios.

[0131] The method of this invention can completely exclude the target species label during the model construction process, relying solely on training data from other species (such as humans, nematodes, and yeast), and can still be transferred to the prediction of essential genes in Fusarium graminearum, achieving a stable performance with an AUC exceeding 0.7 in testing. Furthermore, it can also achieve an AUC > 0.7 performance for non-mammalian eukaryotes (such as Saccharomyces cerevisiae) under completely "zero-sample" conditions, reflecting good structural semantic embedding transfer ability, making it particularly suitable for multi-species, label-free gene function prediction tasks.

[0132] (4) It has good interpretability and the key features are clearly defined and quantifiable.

[0133] By introducing the attention mechanism of graph neural networks, the model of this invention can perform explicit weighted modeling at the residue or protein node level, thereby identifying structural regions that contribute significantly to the prediction results. Attention visualization heatmaps show that the model focuses on highly conserved functional domains and core residue regions, providing theoretical support for subsequent RNAi target design and validation. Furthermore, this invention constructs an independent multi-omics graph attention network, fusing features such as expression profile (EXP), conservation score (ORT), and subcellular localization (SUB). Ablation experiments show that removing EXP or ORT features causes a 6%–8% decrease in AUC, while the GAT model integrating all three features achieves the highest AUC ≈ 0.78 in Fusarium graminearum, further validating the role of the multi-feature fusion mechanism in improving model interpretability and performance.

[0134] (5) The dual-path cross-validation mechanism improves credibility and biological explanatory power.

[0135] This invention constructs a dual-path scoring mechanism by integrating protein language modeling and multi-omics graph attention networks, which can effectively improve the reliability and interpretability of the final prediction results. When the prediction results of the two paths are consistent and the scores are both higher than a set threshold (e.g., ≥0.7), the credibility of it as an essential gene is significantly improved; if the scores differ greatly (e.g., the absolute difference exceeds 0.5), the system automatically marks it as a "prediction inconsistency" state, prompting experimental verification. This dual-path scoring mechanism helps to analyze the differences in the focus of different models on the internal structural properties and external functional interactions of proteins, thereby enhancing the prediction robustness and decision controllability of the overall system.

[0136] (6) The method is reproducible and has an open structure, and has the value of scientific research transformation and promotion.

[0137] The model used in this invention is implemented based on the PyTorch Geometric framework, featuring a good modular structure and openness, supporting GPU parallel acceleration, reproducibility, and extensibility. The training and testing data (protein sequences, tag files, and interaction networks) are all sourced from publicly available databases, ensuring standardization and consistency of data sources. Based on this framework, it can be rapidly deployed to other plant pathogens or agricultural pests, providing a technical support platform for high-throughput screening of RNAi target genes and the development of green pesticides, and possessing promising prospects for scientific research translation. Attached Figure Description

[0138] Figure 1 This is a schematic diagram of the overall process of the method of the present invention.

[0139] Figure 2 To demonstrate the ROC curves and AUC (Area Under Curve) of each prediction model, the models used for comparison include 1 (ESM+GAT), 2 (ESM), 3 (CNN), 4 (Transformer), and 5 (LSTM). The species evaluated includes Fusarium graminearum (…). F. graminearum ) 、 brewing yeast ( S. cerevisiae ) 、 Escherichia coli ( E. coli ) 、 Human beings ( H. sapiens ) 、 Drosophila melanogaster ( D. melanogaster ) and Caenorhabditis elegans ( C. elegansA shows the ROC curves and AUC values ​​of each model prediction result in *Fusarium graminearum* (AUC = 0.804, AUC = 0.795, AUC = 0.693, AUC = 0.685, AUC = 0.563); B shows the ROC curves and AUC values ​​of each model prediction result in *Saccharomyces cerevisiae* (AUC = 0.743, AUC = 0.766, AUC = 0.642, AUC = 0.645, AUC = 0.559); C shows the ROC curves and AUC values ​​of each model prediction result in *Escherichia coli* (AUC = 0.806, AUC = 0.768, AUC = 0.702, AUC = 0.691, AUC = 0.549). D represents the ROC curves and AUC values ​​of the prediction results of each model in humans (AUC = 0.873 for 1, 0.887 for 2, 0.707 for 3, 0.723 for 4, and 0.708 for 5); E represents the ROC curves and AUC values ​​of the prediction results of each model in Drosophila melanogaster (AUC = 0.727 for 1, 0.762 for 2, 0.597 for 3, 0.627 for 4, and 0.578 for 5); F represents the ROC curves and AUC values ​​of the prediction results of each model in Caenorhabditis elegans (AUC = 0.885 for 1, 0.899 for 2, 0.756 for 3, 0.804 for 4, and 0.78 for 5).

[0140] Figure 3 To demonstrate the AUPRC (Area Under Precision-Recall Curve) metric, the models used for comparison included 1 (ESM+GAT), 2 (ESM), 3 (CNN), 4 (Transformer), and 5 (LSTM), and the species evaluated included Fusarium graminearum (…). F. graminearum ) 、 brewing yeast ( S. cerevisiae ) 、 Escherichia coli ( E. coli ) 、 Human beings ( H. sapiens ) 、 Drosophila melanogaster ( D. melanogaster ) and Caenorhabditis elegans ( C. elegansA represents the AUPRC values ​​of each model prediction result in *Fusarium graminearum* (AUPR = 0.584, AUPR = 0.55, AUPR = 0.478, AUPR = 0.47, AUPR = 0.457); B represents the AUPRC values ​​of each model prediction result in *Saccharomyces cerevisiae* (AUPR = 0.786, AUPR = 0.808, AUPR = 0.704, AUPR = 0.696, AUPR = 0.62); C represents the AUPRC values ​​of each model prediction result in *Escherichia coli* (AUPR = 0.666, AUPR = 0.752, AUPR = 0.586, AUPR = 0.609, AUPR = 0.481). D represents the AUPRC values ​​of the prediction results of each model in humans (AUPR=0.835, AUPR=0.923, AUPR=0.604, AUPR=0.62, AUPR=0.607); E represents the AUPRC values ​​of the prediction results of each model in Drosophila melanogaster (AUPR=0.599, AUPR=0.63, AUPR=0.495, AUPR=0.51, AUPR=0.479); F represents the AUPRC values ​​of the prediction results of each model in Caenorhabditis elegans (AUPR=0.845, AUPR=0.93, AUPR=0.537, AUPR=0.644, AUPR=0.583).

[0141] Figure 4 A violin plot showing the class distribution of probabilities predicted by the ESM+GAT model. It illustrates the distribution of ESM+GAT predicted probabilities and their medians across species. The red diamonds represent the median of the required and non-required gene probabilities predicted using the ESM+GAT model of this invention for samples from various species.

[0142] Figure 5 The image shows the AUC heatmap under cross-species training-test combination. A represents the result of the ESM+GAT combined model based on this invention; B represents the result based on the ESM model alone; C represents the result based on the Transformer model; D represents the result based on the LSTM model; and E represents the result based on the (CNN) model.

[0143] Figure 6 This is a flowchart for predicting essential genes based on a multi-omics graph attention network.

[0144] Figure 7This chart compares the performance of the Multi-Group Learning Graph Attention Network (GAT) model with other machine learning models. Performance metrics include the model's AUC (ROC), AUPRC, Specificity (SP), and Matthews correlation coefficient (MCC). A compares the performance of different models in *Fusarium graminearum*; B compares the performance of different models in *Saccharomyces cerevisiae*; C compares the performance of different models in *Escherichia coli*; D compares the performance of different models in humans; E compares the performance of different models in *Drosophila melanogaster*; and F compares the performance of different models in *Caenorhabditis elegans*.

[0145] Figure 8 This chart compares the performance of the Multi-Group Graph Attention Network (GAT) with various graph learning methods across multiple species. Performance metrics include the model's AUC (ROC), AUPRC, Specificity (SP), and Matthews correlation coefficient (MCC). A represents the performance comparison in *Fusarium graminearum*; B represents the performance comparison in *Saccharomyces cerevisiae*; C represents the performance comparison in *Escherichia coli*; D represents the performance comparison in humans; E represents the performance comparison in *Drosophila melanogaster*; and F represents the performance comparison in *Caenorhabditis elegans*.

[0146] Figure 9 This is a heatmap showing the ablation experiment performance of a multi-group Learned Graph Attention Network (GAT). The gray background represents unavailable data (NA).

[0147] Figure 10 This is a graph showing the predicted probability distribution of a multi-group Learning Graph Attention Network (GAT) under different feature combinations. Exp+Ort represents the fusion expression profile and conservation score; Sub+Ort represents the fusion subcellular localization and conservation score; Exp+Sub represents the fusion expression profile and subcellular localization; Exp+Sub+Ort represents the fusion expression profile, subcellular localization, and conservation score. A represents the predicted probability distribution in *Fusarium graminearum*; B represents the predicted probability distribution in *Saccharomyces cerevisiae*; C represents the predicted probability distribution in humans; D represents the predicted probability distribution in *Drosophila melanogaster*; and E represents the predicted probability distribution in *Caenorhabditis elegans*.

[0148] Figure 11 The influence trend of different thresholds on the performance of multi-group learning graph attention networks (GAT). 1 represents Fusarium graminearum (…). F. graminearum ) 、 2 represents brewing yeast ( S. cerevisiae ) 、 3 represents Escherichia coli ( E. coli ) 、 4 represents human beings ( H. sapiens ) 、 5 represents the fruit fly with a black belly (Drosophila melanogaster) D. melanogaster ), 6 represents Caenorhabditis elegans (C. elegans)C. elegans ). Detailed Implementation

[0149] The present invention will now be described in further detail with reference to specific embodiments. The given embodiments are merely illustrative of the invention and not intended to limit its scope. The embodiments provided below can serve as a guide for further improvements by those skilled in the art and do not constitute a limitation on the invention in any way.

[0150] Unless otherwise specified, the experimental methods used in the following examples are conventional methods, performed according to the techniques or conditions described in the literature in this field or according to the product instructions. Unless otherwise specified, the materials and reagents used in the following examples are commercially available.

[0151] Example 1. Prediction process of essential genes in non-model organisms based on protein language model and graph attention network.

[0152] This embodiment proposes a system for predicting essential genes in non-model organisms—ProGATE (Protein Language Model + Graph Attention + Essentiality)—that integrates a protein language model and a graph attention network. Figure 1 ).

[0153] 1. Data processing module (A).

[0154] This module provides raw data support for model training and inference, aiming to build a cross-species transferable supervised training dataset and evaluation system, mainly including species selection, protein sequence processing, label generation and data directory structure organization.

[0155] 1.1 Selection of training species and setting of target species.

[0156] This embodiment uses *Fusarium graminearum* as the core prediction target. To enhance the model's generalization ability, five additional model organisms with relatively complete gene essentiality annotations were selected as training and validation species:

[0157] Escherichia coli;

[0158] Saccharomyces cerevisiae;

[0159] Fruit fly (Drosophila melanogaster);

[0160] Caenorhabditis elegans;

[0161] Homo sapiens;

[0162] This design enables the model to learn protein functional characteristics across species, improving migration prediction performance for non-model organisms.

[0163] 1.2 Protein sequence data acquisition and processing.

[0164] The Fusarium graminearum protein sequence data were obtained from the FungiDB database, which contains a total of 14,143 protein sequences;

[0165] The protein sequences of the training species were uniformly derived from the UniProt database;

[0166] All protein FASTA format files undergo standardized processing, including uniform format encoding and redundancy removal. Noise samples such as redundant records, non-standard amino acid residues, and excessively short sequences are eliminated to obtain cleaned protein sequences. The cleaned protein sequences are named by species and saved in the / data / directory for subsequent modules to use.

[0167] 1.3 Tag Construction and Weak Supervision Strategy.

[0168] The labels of model species in the training species were divided into "essential" and "non-essential" categories based on the annotation information of the OGEE (Online Gene Essentiality) and DEG (Database of Essential Genes) databases, and were used as supervised training samples;

[0169] For Fusarium graminearum, functional annotations labeled as "lethal" and "reduced virulence" in the PHI-base database were used as weakly supervised positive samples (i.e., "essential", such as Chs3b and TRI5 genes). Other unannotated proteins without functional records were considered negative by default (i.e., "non-essential"), thus forming weakly supervised training labels.

[0170] All tags are stored in CSV format, with fields including gene ID and binary label (0 or 1: 0 corresponds to "not required", 1 corresponds to "required"). The file structure is as follows:

[0171] Gene Label;

[0172] FGRAMPH1_01G04077 0;

[0173] FGRAMPH1_01G26781 1;

[0174] 2. Protein Language Model Feature Generation Module (B).

[0175] This module is used to extract semantically context-dependent residue features from the primary sequence of a protein and to infer potential empty structure contact information, thus providing support for the subsequent graph neural network module to construct the input graph structure.

[0176] 2.1 Pre-trained (language) model embedding generation.

[0177] ESM-2 (Evolutionary Scale Modeling version 2) was used as the pre-trained protein language model;

[0178] Each cleaned protein sequence in FASTA format obtained in step 1.2 and the corresponding tag obtained in step 1.3 are input into the ESM-2 model, which outputs a context embedding vector for each amino acid residue to capture its structural and functional semantic information. This context embedding vector is a high-dimensional matrix vector (1280 dimensions in this embodiment) of residue length × feature dimension, used for initial feature setting of graph neural network nodes.

[0179] 2.2 Residue Contact Graph Generation.

[0180] Based on the attention weight matrix and residue sequence spacing of the ESM-2 model, the attention intensity between residue pairs of each protein is calculated, and an undirected graph structure is constructed to obtain the residue contact graph, which consists of nodes and edges: nodes represent residues, and edges represent residue pairs with high attention scores, reflecting the potential spatial adjacency relationship of residue pairs.

[0181] If a pair of residues satisfies the dual conditions of "attention weight greater than or equal to a set threshold (0.2 in this embodiment)" and "residue sequence spacing L≥5", then an undirected edge is established between the two residues.

[0182] The graph structure is saved in .pt format for direct reading by the GNN module.

[0183] Residue contact diagrams reflect the potential structural information inside proteins, providing an input basis for graph neural network modeling.

[0184] 3. Graph Neural Network Module (C).

[0185] This module uses a graph attention network (GAT) to process the residue contact map of each protein, models the residue contact map, introduces a multi-head attention mechanism to capture the heterogeneous connection strength between residues, and extracts protein embedding vectors that fuse semantic and structural information to provide input features for subsequent classification tasks.

[0186] 3.1 Graph Structure Construction and Input Definition.

[0187] After step 2, each protein is constructed as a residue contact graph: nodes are amino acid residues, and node features are the context embedding vectors of amino acid residues obtained by ESM-2; the edges between nodes are obtained by the contact relationships inferred from the attention weights and sequence positions, forming a set of undirected edges.

[0188] 3.2 GAT architecture and propagation mechanism.

[0189] In processing the residue contact map of each protein, GAT assigns different aggregation weights to neighboring nodes through a learned attention mechanism: each layer of the network aggregates neighbor information through trainable attention weights and outputs multi-level embeddings.

[0190] GAT supports multi-head attention mechanisms, improving model stability and feature representation diversity; ultimately, through multi-layer stacking and global pooling layer operations, node features are integrated into a protein overall embedding vector.

[0191] 3.3 Output structure and downstream interface.

[0192] Each protein map is encoded by a graph neural network, and the output is a fixed-length protein embedding vector.

[0193] This embedding vector serves as the input feature for subsequent classification and adversarial training modules.

[0194] 3.4 Structural interpretability analysis.

[0195] After the model is trained, the key amino acid residues that contribute the most to the prediction results are identified by extracting and analyzing the attention weight distribution, which helps in target screening and protein structure interpretation.

[0196] Visualize attention heatmaps to identify key residue regions on which model predictions depend; provide theoretical support for subsequent protein structure-function annotation, biological validation, and target design.

[0197] 4. Classification and adversarial training module (D).

[0198] This module is responsible for classifying the overall protein embedding vectors extracted by the graph neural network using a classifier, and determining whether the corresponding gene is an "essential gene". At the same time, an adversarial perturbation mechanism is introduced during the training phase to enhance the model's robustness to non-pattern species and perturbation inputs.

[0199] 4.1 Classifier structure design.

[0200] The classifier uses a multilayer perceptron (MLP) structure, specifically including:

[0201] Input layer: Receives the overall protein embedding vector output by the graph neural network in step 3;

[0202] Hidden layer: includes a fully connected layer, ReLU activation function, and Dropout operation (to prevent overfitting);

[0203] Output layer: Logistic regression is performed using a single neuron with the sigmoid activation function. The output range is normalized to [0,1], representing the predicted probability that the protein is an "essential gene". (P_Pro).

[0204] 4.2 Loss Function and Optimization Process.

[0205] The classification loss function is the standard binary cross entropy (BCE):

[0206] ,

[0207] In Equation 1 For real labels, Predict probabilities for the model;

[0208] To maintain generalization among species, no category weights or focus loss were set;

[0209] The model optimizer uses Adam and incorporates a dynamic learning rate scheduling mechanism (such as Cosine Decay). Training is performed using the standard binary cross-entropy loss function, without introducing class weighting or focus loss, to ensure general applicability to multi-species scenarios.

[0210] 4.3 Design of counter-disturbance mechanism.

[0211] Introducing gradient-direction-based perturbations into the embedding space to generate adversarial examples enhances the model's robustness to unknown perturbations:

[0212] Adversarial perturbations are only enabled during the training phase;

[0213] The embedding vector is perturbed using the Fast Gradient Sign Method (FGSM):

[0214] Calculate the gradient of the loss function with respect to the input embedding;

[0215] Add a perturbation with an amplitude of ε ∈ [0.001, 0.01] along the gradient direction;

[0216] The final training loss is a weighted sum of the original sample loss and the adversarial sample loss:

[0217] ,

[0218] In Equation 2, λ1 and λ2 are adjustable weights, typically 0.5:0.5.

[0219] No perturbations are added during the inference phase; prediction is made using only the original embedding.

[0220] 5. Multi-omics feature fusion prediction module (E).

[0221] This module constructs independent + predicted pathways based on external multi-omics information of proteins and protein interaction networks. Figure 6 By combining graph attention networks for modeling and running them in parallel with protein language modeling methods, a dual-path complementary prediction architecture is formed. Furthermore, the robustness and reliability of the model are enhanced through an integration and fusion strategy.

[0222] 5.1 Integration of multi-omics data and construction of protein-protein interaction maps.

[0223] A protein-protein interaction network (PPI) for Fusarium graminearum was constructed using the STRING database: nodes represent proteins; edges represent physical or functional relationships between proteins, and low-quality edges were filtered out by setting a confidence threshold (e.g., score ≥ 0.5).

[0224] Each node is further fused with the following omics features to obtain a multi-omics protein interaction network (multi-omics PPI):

[0225] Gene expression profile: A gene expression profile reflects the transcriptional level of a target gene within a cell. To obtain gene expression data for *Fusarium graminearum*, this invention downloaded the transcriptome data of this species from the FungiDB database and calculated the TPM (Transcripts Per Million) value. The TPM value, as a unit, can quantitatively reflect the expression level of a gene in transcriptome sequencing and provides strong support for the prediction of protein function.

[0226] Orthology Score: Obtained from cross-species alignments using methods such as BLAST or Diamond (e.g., based on protein homology).

[0227] Orthology Score: Used to assess protein homology between a target species and other species, revealing the conservation of proteins across different species. This invention utilizes multi-species protein sequence data provided by the NCBI database and employs the DIAMOND tool to perform efficient cross-species alignment of Fusarium graminearum protein sequences with protein sequences from other species. The calculated alignment results provide a cross-species conservation score for each gene in Fusarium graminearum.

[0228] Subcellular localization refers to the distribution of proteins within functional regions of cells, which is crucial for understanding protein function. To obtain subcellular localization information for Fusarium graminearum proteins, this invention extracts relevant data from the COMPARTMENTS and UniProtKB databases and employs one-hot encoding technology to annotate the subcellular localization of proteins (such as the nucleus and mitochondria). This data will further enrich the node features in protein-protein interaction networks, helping to better understand the biological functions of proteins.

[0229] Multi-omics features are fused into a high-dimensional vector, as shown in Equation 3:

[0230] .

[0231] 5.2 Multi-group learning graph attention modeling process.

[0232] A graph attention network (GAT) is used to model the multi-omics PPI graph constructed in step 5.1 above, propagating information on the PPI graph and non-linearly aggregating the omics features of each node in the multi-omics PPI:

[0233] Each layer performs weighted aggregation of the features of neighboring nodes;

[0234] Introducing multi-head attention to enhance feature representation capabilities;

[0235] Ultimately, the proteins are converged into a fixed-length protein representation through attention pooling.

[0236] The predicted probability value P_GAT ∈ [0,1] of a protein is output using the Sigmoid activation function, representing the confidence that the gene is an essential gene.

[0237] 5.3 Dual-pathway prediction cross-fusion mechanism.

[0238] The system integrates two prediction paths:

[0239] 1) Protein Language Model (Pro Path): Based on the residue contact map and graph neural network (GAT) extracted from a single sequence using the protein language model (ESM-2), the prediction probability P_Pro is the prediction probability ŷ of the "essential gene" output in step 4.

[0240] 2) Multi-omics graph attention network (GAT path): A graph attention network constructed based on PPI and multi-omics features, which outputs the prediction probability P_GAT, that is, the prediction probability value P_GAT output in step 5.2.

[0241] The following fusion judgment strategy is adopted:

[0242] If P_Pro≥T and P_GAT≥T (T is an adaptive threshold, which is 0.7 in this embodiment), it is determined to be a highly reliable essential gene;

[0243] If the difference between the two path predictions, |P_Pro - P_GAT|, is greater than δ (e.g., δ=0.5), it is marked as "prediction divergence" and experimental verification is recommended.

[0244] If both are <0.3, then it is considered a non-essential gene.

[0245] Example 2. Model training and evaluation process in the prediction process of essential genes in non-model organisms.

[0246] This embodiment details the complete process of modeling from data preprocessing to training, adversarial optimization, and evaluation, aiming to ensure the repeatability, verifiability, and cross-species applicability of the model results.

[0247] 1. Training data preparation process.

[0248] Data for each species is stored after undergoing a unified preprocessing procedure;

[0249] The label information is uniformly mapped to "required (1)" and "not required (0)", which is automatically completed using a Python script (label_mapping.py); Fusarium graminearum is trained with weak supervision and evaluated using a strong label method;

[0250] Data splitting method: 80% training set, 10% validation set and 10% test set. You can choose "split by species" or "mixed split" to evaluate generalization ability or single species performance.

[0251] 2. Model training process.

[0252] All model modules (protein language model feature extraction, graph attention network, classifier, fusioner) are integrated into the unified training main program (main_train.py).

[0253] Supports both independent training and joint end-to-end training modes;

[0254] The main training parameter configurations used are shown in Table 1:

[0255]

[0256] 3. Details of adversarial training (for graph embeddings).

[0257] An adversarial perturbation module is embedded in the front end of the classifier to add perturbations to the protein embedding vectors output by GAT:

[0258] Perturbation method: Fast Gradient Sign Method (FGSM);

[0259] The disturbance amplitude ε is controlled within 0.001 ~ 0.01;

[0260] The adversarial sample loss and the original sample loss are weighted proportionally to form the final loss;

[0261] Adversarial perturbations are introduced during the training phase, but no perturbation mechanisms are used during the inference phase.

[0262] 4. Model evaluation and performance indicators.

[0263] The models used for evaluation included five types: a combination of the ESM model and the graph attention network model of this invention (ESM+GAT), an ESM model, a convolutional neural network (CNN), a long short-term memory network (LSTM), and a Transformer neural network model. Model performance evaluation metrics included:

[0264] AUC-ROC (Area Under Receiver Operating Characteristic Curve): Evaluates the model's overall ability to distinguish between positive and negative samples;

[0265] AUPRC (Area Under the Precision-Recall Curve): Applicable to situations where the ratio of positive to negative samples is unbalanced, it measures the effectiveness of the model in high-confidence predictions;

[0266] ACC (Accuracy): refers to the proportion of instances that the model correctly predicts out of the total number of instances;

[0267] SP (Specificity): Measures the ability to correctly identify negative samples, i.e., the proportion of instances correctly identified as negative out of all negative instances. Specificity is very important, especially when data is imbalanced, as it helps assess whether a model avoids false positives.

[0268] MCC (Matthews Correlation Coefficient): is a comprehensive evaluation metric that considers all four classification results: true positives, true negatives, false positives, and false negatives, and is particularly suitable for evaluating imbalanced datasets.

[0269] Generalization ability assessment: interspecies migration test (e.g., nematode training, tested on yeast / Fusarium graminearum). In this embodiment, the species evaluated include: Fusarium graminearum, Saccharomyces cerevisiae, Escherichia coli, humans (Homo sapiens), Drosophila melanogaster, and Caenorhabditis elegans; weakly supervised training and strong label validation comparison.

[0270] By comparing different machine learning algorithm models Figure 2 and Figure 3 The results show that the ESM+GAT combined model ( Figure 3 The "1" in the middle represents that it generally outperforms all species, and in non-model organisms such as Fusarium graminearum (… Figure 3 The performance was particularly outstanding in the A) (AUC=0.804, AUPR=0.584). In addition, the AUPRC of ESM+GAT was higher than 0.5 in all species, which was better than the baseline; even when the sample imbalance was severe (low proportion of essential genes), the model still maintained a high accuracy, demonstrating its advantage in modeling species without labels.

[0271] pass Figure 4 The violin plot results of the predicted probability categories of the ESM+GAT model show that the method of this invention generally has high predicted probability values ​​for essential genes in each species (concentrated in >0.65), while the predicted probability values ​​for non-essential genes are concentrated in <0.35; it exhibits a clear bimodal distribution, demonstrating the model's good discriminative ability.

[0272] Figure 5 A 6x6 matrix heatmap of AUC results under cross-species training-test combinations is presented, showing the AUC values ​​under different combinations of training species and model-test species, evaluating the performance of five models in a "zero-shot" transfer learning scenario. The results show that the ESM+GAT model of this invention (… Figure 5 A) generally achieved excellent results with AUC>0.7 under different training-test combinations, demonstrating its cross-species generalization ability.

[0273] Figure 7This graph compares the performance of the multi-omics graph attention network (GAT) model with other multi-omics machine learning models. Performance metrics include the model's AUC (ROC), AUPRC (Average Upright Convergence Ratio), Specificity (SP), and Matthews correlation coefficient (MCC). Comparisons include the graph attention network (GAT), node embedding model (Node2Vec), multilayer perceptron (MLP), random forest (RS), naive bayes (NB), and support vector machine (SVM). The GAT model, due to its adaptive attention mechanism, outperforms other static adjacency structure methods, showing the best overall performance in ROC, AUPRC, Specificity, and MCC metrics.

[0274] Figure 8 This chart compares the performance of the Multi-Group Graph Attention Network (GAT) with various graph learning methods across multiple species. Performance metrics include the model's AUC (ROC), AUPRC, Specificity (SP), Accuracy (ACC), and Matthews Correlation Coefficient (MCC). A compares performance in *Fusarium graminearum*; B compares performance in *Saccharomyces cerevisiae*; C compares performance in *Escherichia coli*; D compares performance in humans; E compares performance in *Drosophila melanogaster*; and F compares performance in *Caenorhabditis elegans*. The various graph learning methods include: Graph Convolutional Networks (GCN), Degree Centrality (DC), and Clustering Coefficient (CC).

[0275] By comparing multiple performance metrics of GAT, GCN, DC, and CC across various species, this invention employs the graph attention network GAT model (…). Figure 8 (The dark red text represents the method) Because it uses an adaptive attention mechanism, its performance is better than static adjacency structure methods such as GCN.

[0276] Figure 9This is a heatmap showing the ablation experiment performance of the Multi-Omics Graph Attention Network (GAT). It displays the AUC values ​​of the ROC under different combinations of hermics data and control models. Different hermics data combinations include: expression profile data (EXP), subcellular localization data (SUB), or conservation scoring data (ORT), and any combination thereof (including: EXP_ORT representing fused expression profile and conservation scoring, SUB_ORT representing fused subcellular localization and conservation scoring, EXP_SUB representing fused expression profile and subcellular localization, EXP_SUB_ORT representing fused expression profile, subcellular localization, and conservation scoring). The comparison benchmarks include degree centrality (NDC), node embedding model (Node2Vec, N2V), and cluster centrality (NCC). The results show that by comparing the AUC values ​​of different hermics data combinations and control models, the Multi-Omics Graph Attention Network (GAT) of this invention exhibits significantly better performance than traditional methods under multiple combinations. Especially when the model incorporates a combination of expression profile (EXP) and subcellular localization (SUB) data (EXP_SUB) and a combination of the three features (EXP_SUB_ORT), the model achieves the highest AUC value, indicating that this combination can more comprehensively integrate the multidimensional features of proteins and improve the model's discriminative ability.

[0277] Figure 10 This paper presents the probability distribution of predictions made by a multi-group Learning Graph Attention Network (GAT) under different feature combinations. The average predicted probability of each category is marked with red dots. Exp+Ort represents the fusion of expression profile and conservation score; Sub+Ort represents the fusion of subcellular localization and conservation score; Exp+Sub represents the fusion of expression profile and subcellular localization; and Exp+Sub+Ort represents the fusion of expression profile, subcellular localization, and conservation score. "E" indicates essential, and "NE" indicates non-essential. The results show that the Exp+Sub+Ort combination exhibits superior discriminative performance across multiple species, providing guidance for selecting feature combinations in practical applications.

[0278] Figure 11 The influence of different thresholds on the performance of multi-group learning graph attention networks (GAT) was investigated. The AUC performance of the GAT model in various species was evaluated under different edge weight thresholds (0.1–0.9). Most species showed stable performance to threshold changes, validating the model's robustness; Fusarium graminearum (… Figure 11 (represented by 1) and fruit flies ( Figure 11 The performance of the 5-axis (represented by the Chinese character set 5) fluctuates significantly at high thresholds, highlighting the importance of introducing multi-omics in network data sparse scenarios.

[0279] 5. Results visualization and report generation.

[0280] The loss curve and accuracy curve during model training are output to / outputs / plots / ;

[0281] All evaluation processes are encapsulated in evaluate.py, which supports automatic multi-model parallel testing and chart (CSV / PDF) output;

[0282] The model supports attention heatmap output to reveal the key residues or neighboring proteins that the model focuses on when identifying "essential" tags.

[0283] The present invention has been described in detail above. Those skilled in the art will recognize that the invention can be practiced in a wide range of ways with equivalent parameters, concentrations, and conditions without departing from its spirit and scope, and without requiring unnecessary experiments. While specific embodiments have been provided, it should be understood that further modifications can be made to the invention. In summary, according to the principles of the invention, this application is intended to include any changes, uses, or improvements to the invention, including changes made using conventional techniques known in the art that depart from the scope disclosed herein.

Claims

1. A data processing apparatus, comprising a memory, a processor, and a computer program stored in the memory, characterized in that: The processor executes the computer program to perform the following steps: A1) Obtain labels for each protein based on annotations of the protein sequences of the target species and n model organisms. The labels for the target species are obtained from the PHI-base database. Protein genes annotated as "lethal" or "reduced virulence" in the PHI-base database have required labels and are considered weakly supervised positive classes. Protein genes not annotated and without functional records in the PHI-base database have optional labels and are considered negative classes. A2) Embed each protein sequence and its corresponding label into a protein language model, outputting context embedding vectors for amino acid residues in the protein. Construct a residue contact graph for each protein based on the context embedding vectors. The protein language model is ESM-2. The residue contact graph, constructed based on the context embedding vectors and model attention weights and residue distances, consists of nodes and edges: nodes represent residues, and edges represent residue pairs with high attention scores. A3) Use a graph neural network to model the residue contact map, so as to integrate the features of each amino acid residue in the residue contact map and output the overall protein embedding vector; A4) Embed all the proteins into a vector and input it into the classifier model for training. Use an activation function to output the necessary probability value of the gene corresponding to the protein. The required probability value P1 of the gene corresponding to the protein of the target species is obtained based on the classifier model. A5) Multi-omics feature fusion: Construct the protein interaction network of the target species, and fuse the omics features of the nodes of the protein interaction network based on omics data to obtain a multi-omics protein interaction network; The multi-omics protein interaction network is modeled using a graph neural network model to nonlinearly aggregate the omics features of each node in the multi-omics protein interaction network. The graph neural network model is used to output the required probability value P2 of the gene corresponding to the protein using an activation function. A6) Output: Determine whether the gene corresponding to each protein in the target species is an essential gene based on the required probability values ​​P1 and P2. The target species is a non-model organism species or a species with incomplete gene annotation; n is greater than or equal to 1.

2. The data processing apparatus according to claim 1, characterized in that: The graph neural network described in A3) and A5) is a graph attention network.

3. The data processing apparatus according to claim 1 or 2, characterized in that: The data processing device is a device for predicting essential genes in non-model organisms or organisms with incomplete gene annotation.

4. Methods for predicting essential genes in non-model organisms or organisms with incomplete gene annotation, including: B1) Obtain the corresponding label for each protein based on the protein sequence of the target species and the protein sequences of n model organism species, respectively, using annotation. The tags for the target species were obtained based on the PHI-base database; the tags for protein genes annotated as "lethal" or "reduced virulence" in the PHI-base database were required and were classified as weakly supervised positive genes; the tags for protein genes that were not annotated and had no functional records in the PHI-base database were not required and were classified as negative genes. B2) Embed each protein sequence and its corresponding tag into a protein language model, and output the context embedding vector of the amino acid residues in the protein; A residue contact graph is constructed for each protein based on the context embedding vector; the protein language model is ESM-2; the residue contact graph is constructed based on the model attention weights and the distance between residues, and consists of nodes and edges: nodes are residues, and edges represent residue pairs with high attention scores; B3) Use a graph neural network to model the residue contact map, so as to integrate the features of each amino acid residue in the residue contact map and output the overall protein embedding vector; B4) Embed all the proteins as a whole into a vector and input it into the classifier model for training. Use an activation function to output the necessary probability value of the gene corresponding to the protein into the classifier model. The required probability value P1 of the gene corresponding to the protein of the target species is obtained based on the classifier model. A5) Multi-omics feature fusion: Construct the protein interaction network of the target species, and fuse the omics features of the nodes of the protein interaction network based on omics data to obtain a multi-omics protein interaction network; The multi-omics protein interaction network is modeled using a graph neural network model to nonlinearly aggregate the omics features of each node in the multi-omics protein interaction network. The graph neural network model is used to output the required probability value P2 of the gene corresponding to the protein using an activation function. A6) Output: Determine whether the gene corresponding to each protein in the target species is an essential gene based on the required probability values ​​P1 and P2. The target species is a non-model organism species or a species with incomplete gene annotation; n is greater than or equal to 1.

5. The method according to claim 4, characterized in that: The graph neural networks described in B3) and B5) are graph attention networks.

6. A computer program product, comprising a computer program, characterized in that: When the computer program is executed by a processor, it implements the steps A1)-A6) of any one of claims 1-3.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, it implements the steps A1)-A6) of any one of claims 1-3.

8. Any of the following applications of the data processing apparatus according to any one of claims 1-3 and / or the computer program product according to claim 6 and / or the computer-readable storage medium according to claim 7: C1) Application in predicting targets of RNA interference in pests or pathogens; Application of C2 in pesticide target screening; C3) Application in agricultural disease control and / or development or preparation of biological pesticides.

Images