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

Abstract

The invention 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 invention is how to screen essential genes of non-model organisms. The method comprises the following steps: respectively obtaining a label corresponding to each protein from a protein sequence of a target species and protein sequences of n model biological species on the basis of annotations; embedding the protein sequence and the tag into the protein language model, and outputting an embedded vector; constructing a residue contact graph for each protein based on the embedded vector; modeling, integrating and outputting the features of each amino acid residue in the residue contact graph by using a graph neural network to obtain an integral protein embedding vector; inputting all the protein embedded vectors into a classifier model for training, and outputting a necessary probability value of a protein gene by using an activation function; and obtaining a necessary probability value P1 of the gene of the protein based on the classifier model. The invention can be applied to the prevention and control of agricultural diseases and the development of biopesticides.

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: 1) Many non-model organisms are difficult to culture stably or have their genes edited under laboratory conditions; 2) Lack of high-quality genome annotation, protein-protein interaction information, and other omics resources; 3) The experimental operation is costly and time-consuming, making it difficult to scale up for systematic screening of multiple species.

[0003] 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: 1) Its high dependence on manually extracted features and high-quality protein interaction network data limits its applicability in non-model species; 2) While sequence-based models improve versatility, they generally lack the ability to provide biological explanations for the prediction mechanisms. 3) Most models only perform well within the training species, and have poor generalization ability for cross-species "zero-shot" prediction; 4) Existing models have not yet met the high reliability and practicality standards required for pesticide target screening.

[0004] 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

[0005] The technical problem to be solved by the present application is how to screen essential genes of non-model organisms and / or how to screen essential genes of organisms with incomplete gene annotation and / or how to screen target genes of RNA pesticides and / or how to screen essential genes of pests or plant pathogens.

[0006] To solve the above technical problem, the present application first provides a data processing device comprising a memory, a processor and a computer program stored on the memory, the processor executing the computer program to implement the following steps: A1) obtaining a label corresponding to each protein based on the annotation of the protein sequence of the target species and the protein sequence of n model organism species respectively; A2) embedding each of the protein sequences and the corresponding labels into a (pre-trained) protein language model to output a context embedding vector of the amino acid residues in the protein; constructing a residue contact map for each protein based on the context embedding vector; A3) using a graph neural network to model the residue contact map to integrate the features of each amino acid residue in the residue contact map and output a protein overall embedding vector; A4) inputting all the protein overall embedding vectors into a classifier model for training, using an activation function to output the classifier model as an essential probability value of the gene corresponding to the protein; obtaining an essential 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.

[0007] In one specific embodiment of the present application, the data processing device is a device for predicting essential genes of non-model organisms (or organisms with incomplete gene annotation).

[0008] In one specific embodiment of the present application, n is 5. The labels include two categories of "essential" and "non-essential". "Essential" means that the gene encoding the protein is an essential gene, and "non-essential" means that the gene encoding the protein is a non-essential gene. Essential genes refer to genes that are indispensable for the growth, development and reproduction of organisms, and are essential for maintaining the basic functions and survival of organisms. These genes are essential for maintaining the basic functions and survival of organisms, and their absence or mutation often leads to serious physiological defects or even death of organisms. Non-essential genes refer to genes other than essential genes in the genome of an organism.

[0009] The labels of the model organism species can be obtained based on the OGEE and DEG databases. The labels of the target species can be obtained based on the PHI-base database. The labels of the protein genes annotated as “lethal” or “reduced virulence” in the PHI-base database can be essential (genes), and the labels of the protein genes not annotated and without functional records in the PHI-base database can be non-essential (genes).

[0010] In the above data processing apparatus, the steps can further include: A5) Multi-omics feature fusion: constructing a protein interaction network of the target species, fusing omics features of nodes of the protein interaction network based on omics data to obtain a multi-omics protein interaction network; using a graph neural network model to model the multi-omics protein interaction network to nonlinearly aggregate omics features of each node in the multi-omics protein interaction network; and using an activation function to output the graph neural network model as an essential probability value P2 of a gene corresponding to a protein; A6) Result output: determining whether each gene corresponding to a protein in the target species is an essential gene according to the essential probability value P1 and the essential probability value P2.

[0011] In the above data processing apparatus, the graph neural network in A3) and A5) can be a graph attention network (GAT).

[0012] A3) The graph neural network can also be a graph convolutional neural network (GCN) or other graph neural networks such as Graph Sample and Aggregated (GraphSAGE).

[0013] In A4) above, before the training, an adversarial sample can also be introduced into the protein embedding vector to enhance the robustness of the classifier model to unknown perturbations.

[0014] The omics data can include gene expression profile data corresponding to the nodes, conservation score data of the nodes, and / or subcellular localization data of the nodes. The omics data can also include other omics data such as proteome and / or metabolome.

[0015] In the above data processing apparatus, the classifier model is optimized using a cross-entropy loss function.

[0016] In the data processing apparatus, the protein interaction network can be constructed using an existing database. In an embodiment of the present application, the database is STRING. The database can also be BioGRID or IntAct.

[0017] The residue contact map 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 vectors. The edges can represent pairs of (amino acid) residues with high attention scores, and the edges can be determined based on the parameters of the protein language model, which can be an attention weight matrix and a residue sequence distance.

[0018] In an embodiment of the present application, the protein language model is ESM-2. The protein language model can also be other protein language models such as ProtTrans or ESM-C.

[0019] The probability value ranges between 0 and 1.

[0020] In A3), the feature of each node (amino acid residue) in the residue contact map can be the embedding vector in A2). Each layer of the graph neural network aggregates neighbor node information through trainable attention weights and outputs multi-level embeddings. The graph neural network finally generates the protein embedding vector through global pooling.

[0021] The data processing apparatus described above can be an apparatus for predicting essential genes of non-model organisms or organisms with incomplete gene annotations.

[0022] To solve the above technical problems, the present application also provides a method for predicting essential genes of non-model organisms or organisms with incomplete gene annotations, which can include: B1) obtaining the corresponding label of each protein based on the annotation of the protein sequence of the target species and the protein sequence of n model organism species, respectively; B2) embedding each protein sequence and its corresponding label into a (pre-trained) protein language model to output the context embedding vector of the amino acid residues in the protein; and constructing a residue contact map for each protein based on the context embedding vector; B3) using a graph neural network to model the residue contact map to integrate and output the features of each amino acid residue in the residue contact map to obtain a protein overall embedding vector; B4) embedding all the proteins into a vector input classifier model for training, using an activation function to output the classifier model as a probability value of essentiality of the gene corresponding to the protein; obtaining a probability value P1 of essentiality 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.

[0023] In a specific embodiment of the present application, n is 5. The label includes two types of "essential" and "non-essential". "Essential" means that the coding gene of the protein is an essential gene, and "non-essential" means that the coding gene of the protein is a non-essential gene. Essential gene refers to a gene that is indispensable for the growth, development and reproduction of an organism. These genes are essential for maintaining the basic functions and survival of an organism, and once lost or mutated, they often cause serious physiological defects or even death of the organism. Non-essential gene refers to a gene in the genome of an organism other than an essential gene.

[0024] The label of the model organism species can be obtained based on the OGEE and DEG databases. The label of the target species can be obtained based on the PHI-base database. The label of the protein gene annotated as "lethal" or "reduced virulence" in the PHI-base database can be essential (gene), and the label of the protein gene not annotated and without functional record in the PHI-base database can be non-essential (gene).

[0025] In some embodiments of the present application, the method is not directly aimed at obtaining a disease diagnosis result or a health condition.

[0026] In some embodiments of the present application, the method does not include the step of obtaining a biological sample from an animal body. The method is not directed to a living human or animal body, but only to data. The method can be an information processing method in which all steps are implemented by a computer or other data processing device.

[0027] The method described above can further include the following steps: B5) Multi-omics feature fusion: constructing a protein interaction network of the target species, fusing (adding) omics features to the nodes of the protein interaction network based on omics data to obtain a multi-omics protein interaction network; using a graph neural network model to model the multi-omics protein interaction network to nonlinearly aggregate the omics features of each node in the multi-omics protein interaction network; using an activation function to output the graph neural network model as a probability value P2 of essentiality of the gene corresponding to the protein; B6) result output: determining whether the gene corresponding to each protein in the target species is an essential gene according to the essential probability value P1 and the essential probability value P2.

[0028] In the above method, the graph neural network in B3) and B5) can be a graph attention network (GAT).

[0029] The graph neural network in B3) can also be a graph convolutional network (GCN) or other graph neural network such as Graph Sample and Aggregated (GraphSAGE).

[0030] In B4) above, before the training, an adversarial sample can also be introduced into the protein embedding vector to enhance the robustness of the classifier model to unknown perturbations.

[0031] The omics data can include gene expression profile data corresponding to the nodes, conservation score data of the nodes, and / or subcellular localization data of the nodes. The omics data can also include other omics data such as proteome and / or metabolome.

[0032] In the above method, the classifier model is optimized using a cross-entropy loss function.

[0033] In the above method, the protein interaction network can be constructed using an existing database. In one specific embodiment of the present application, the database is STRING. The database can also be BioGRID or IntAct.

[0034] The above-described residue contact map 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 vectors. The edges can represent pairs of (amino acid) residues with high attention scores, and the edges can be determined based on parameters of the protein language model, which can be an attention weight matrix and a residue sequence distance.

[0035] The residue contact map is composed of nodes and edges, where the nodes represent amino acid residues, and the features of the nodes are based on residue-level embedding vectors generated by a protein language model; the edges represent pairs of residues with high attention scores, which are constructed based on an attention weight matrix and a residue sequence distance in the protein language model, and are used to reflect potential structural associations between residues.

[0036] In one specific embodiment of the present application, the protein language model is ESM-2. The protein language model can also be other protein language models such as ProtTrans or ESM-C.

[0037] The probability value is in the range of 0-1.

[0038] In B3), the feature of each node (amino acid residue) in the residue contact graph can be the embedding vector in B2). Each layer of the graph neural network aggregates neighbor node information through trainable attention weights and outputs multi-level embeddings. The graph neural network finally generates the protein embedding vector through global pooling.

[0039] In one embodiment of the present application, the target species is Fusarium graminearum.

[0040] To solve the above technical problems, the present application further provides a computer program product comprising a computer program which, when executed by a processor, can implement the steps of A1)-A4) or A1)-A6) above.

[0041] To solve the above technical problems, the present application further provides a computer readable storage medium having stored thereon a computer program which, when executed by a processor, can implement the steps of A1)-A4) or A1)-A6) above.

[0042] Any of the following applications of the above data processing device and / or the above computer program product and / or the above computer readable storage medium also belongs to the protection scope of the present application: C1) in predicting pest or pathogen RNA interference target sites; C2) in pesticide target screening; C3) in agricultural disease prevention and / or biological pesticide development or preparation.

[0043] To break through the above technical bottleneck, the present application proposes a prediction workflow ProGATE (Protein Graph Attention Transformer) that fuses protein language model and graph attention network Pro tein Language Model + G raph At tention + EThe system integrates the sequence context embedding capability of the protein large language model ESM-2, combines the structural modeling advantages of the graph attention network (GAT), and can realize the cross-species prediction of essential genes in non-model pathogenic organisms (such as Fusarium graminearum) without relying on high-quality gene annotation or PPI networks. Especially in the "zero sample" setting, ProGATE shows good generalization performance. In addition, ProGATE introduces a multi-head attention mechanism and an adversarial training strategy, not only improving the robustness of the model, but also providing explanatory results based on key regions such as amino acid residues and domains, providing high credibility support for target gene screening of RNA biological pesticides.

[0044] The present application aims at the problems of strong label dependence, poor generalization ability and lack of biological explanation in the process of essential gene identification of non-model plant pathogenic fungi such as Fusarium graminearum, and proposes a general prediction workflow ProGATE that integrates protein language model and multi-omics graph attention network to solve the following key technical problems: 1) The problem of insufficient essential gene prediction ability in non-model species scenarios: a computational prediction method that does not rely on high-quality genome annotation and protein interaction network is constructed, which can effectively and accurately predict in the context of scarce omics data.

[0045] 2) The problem of cross-species generalization prediction under "zero sample" conditions: based on protein sequences themselves, pre-trained protein language model (ESM-2) is used to extract residue-level context semantic embedding from amino acid sequences, realizing "zero sample" generalization prediction without prior knowledge of target species.

[0046] 3) The problem of insufficient model explainability: traditional deep learning models are mostly "black box" structures, making it difficult to trace the biological mechanisms behind the prediction results. The present application introduces the attention mechanism in graph neural networks, so that the model judgment results can be traced back to specific amino acid residues, domains or functional regions, improving the biological explainability of the prediction results and the scientific credibility of target screening.

[0047] 4) The problem of insufficient multi-omics information fusion capability: protein interaction networks, expression profiles, conservation indices, and subcellular localization are integrated as multi-dimensional inputs to construct a graph attention network model, realizing deep modeling of gene "external attributes" and improving the comprehensiveness of the model.

[0048] 5) The problem of low credibility of target screening results: the present application integrates the results of internal protein representation prediction and external multi-omics graph attention network to construct an end-to-end essential gene prediction and screening process, enhancing the prediction reliability and consistency of the results, and providing high credibility support for the mining of RNA biological pesticide target genes.

[0049] The present application proposes a modular prediction system named ProGATE, which is particularly suitable for cross-species prediction in the context of missing or insufficient omics data (such as Fusarium graminearum) by fusing protein language models and graph attention neural networks. The overall structure of the system includes the following five core modules and their implementation steps: A. Data processing module.

[0050] Training species selection: The present application takes Fusarium graminearum as the target prediction species, and introduces human (Homo sapiens), Saccharomyces cerevisiae, Drosophila melanogaster, Caenorhabditis elegans, and Escherichia coli, etc. as training and validation samples of model organisms with standard annotations, to enhance the cross-species generalization ability of the model.

[0051] 1) Protein sequence collection and standardization: The protein sequences of Fusarium graminearum are derived from FungiDB 20, totaling 14,143; The protein sequences of model species are derived from UniProt database21; All sequences are uniformly formatted and de-redundant to ensure the diversity and representativeness of the training samples.

[0052] 2) Label construction: For Fusarium graminearum, based on the PHI-base database22, the genes annotated as "lethal" or "reduced virulence" are used as weakly supervised positive classes (such as Chs3b and TRI5); Proteins without annotations and functional records are used as negative classes (non-essential); The essential / non-essential label pairs of model species are constructed according to the OGEE and DEG databases23,24, as supervised training samples.

[0053] B. Protein language model feature generation module.

[0054] 1) Sequence embedding extraction: The processed FASTA sequence is input into the ESM-2 pre-trained protein language model; Output the context embedding vector of each amino acid residue, capturing its structural and functional semantic information.

[0055] 2) Residue contact map inference: Based on the model attention weight and the distance between residues, an undirected graph structure is constructed: the nodes are residues, and the edges represent the residue pairs with high attention scores; The contact map reflects the potential internal structural information of the protein, providing input basis for graph neural network modeling.

[0056] C. Graph neural network modeling module.

[0057] 1. Network structure and modeling strategy: Residue contact map is modeled using graph attention network (GAT), introducing multi-head attention mechanism to capture heterogeneous connection strength between residues.

[0058] 2. Information propagation and embedding representation: Node features in the graph are ESM embedding vectors; Each layer of the network aggregates neighbor information through trainable attention weights, outputting multi-level embeddings; Finally, a global pooling is used to generate protein-level embedding representation.

[0059] 3. Biological interpretability: After model training, the distribution of attention weights is analyzed to identify key residues that contribute most to the prediction results, assisting in target screening and structural interpretation.

[0060] D. Classification and adversarial training module.

[0061] 1. Classifier structure design: A two-layer perceptron (MLP) is used as the classifier, with the input being the protein embedding output by the graph neural network; The output is a "required" probability value ranging from 0 to 1, with a sigmoid activation function.

[0062] 2. Training loss function design: The standard binary cross-entropy loss function is used for training, without introducing class weighting or focal loss, to ensure the generality of the model in multiple species scenarios.

[0063] 3. Introduction of adversarial training mechanism: Adversarial samples are generated based on the gradient direction in the embedding space, enhancing the model's robustness to unknown perturbations; The total loss function is the weighted sum of the loss of the original sample and the adversarial sample.

[0064] E. Multisomic feature fusion prediction module.

[0065] 1. Integration of graph structure and external features: The PPI network of F. graminearum is constructed using the STRING database25, with nodes representing proteins and edges representing functional or physical interactions between proteins; Each node is integrated with the following omics features: gene expression profiles (RNA-Seq TPM values), cross-species conservation scores (such as based on protein homology), and subcellular localization information (such as nucleus, mitochondria, etc.) from UniProt, COMPARTMENTS, and other databases.

[0066] 2) Graph Attention Network modeling: Using GAT model to propagate information on PPI graph, realizing nonlinear aggregation of multi-omics features; Finally output the predicted probability P_Pro of the protein, P_Pro [0, 1].

[0067] 3) Double path cross-validation and fusion mechanism: The system integrates two independent prediction methods: 1) Protein language model: based on the contact graph inside the single sequence, output the predicted score P_Pro; 2) Multi-omics graph attention network: based on the biological interaction graph structure and multi-omics features, output P_GAT.

[0068] The following decision strategy is adopted: if the scores of both paths are higher than the dynamic threshold (such as 0.7), it is determined as a high-confidence essential gene; if the difference between the scores of the two paths is large (such as more than 0.5), it is marked as model prediction divergence, and further experimental confirmation is suggested; if both scores are lower than the threshold (such as 0.3), it is determined as a non-essential gene.

[0069] Alternative technical solutions: Pre-trained protein language model replacement: ProtTrans or ESM-C can be used instead of ESM-2; Graph neural network model replacement: GCN or GraphSAGE can be used; Protein interaction expansion: BioGRID, IntAct network structure can be used instead; Multi-omics source replacement: such as introducing proteomics, metabolomics and other data types.

[0070] The present application belongs to the field of cross-application of bioinformatics and artificial intelligence technology, and specifically relates to an essential gene prediction method based on the combination of protein language model and graph neural network. The method is mainly applied to target gene screening of RNA biological pesticides, and is especially suitable for high-throughput identification of gene function and precise control target mining in non-model pathogenic fungi (such as Fusarium graminearum) or agricultural pests with insufficient gene annotation, and can be widely applied in the fields of agricultural disease control, biological pesticide development, etc.

[0071] Advantages of the present application: (1) Wide applicability, breaking through the modeling limitations of non-model organisms.

[0072] The pre-trained protein language model method adopted in the application only relies on the amino acid sequence of the target gene for modeling, without the need for three-dimensional structure analysis, experimental knockout data or protein interaction network information, significantly reducing the dependence on prior data. This feature makes the method widely applicable to non-model organisms such as Fusarium graminearum, and is particularly suitable for plant pathogenic fungi and agricultural pest targets that lack high-throughput functional annotation. Compared with traditional PPI network and experimental label driven modeling methods, the application has more practicality and application potential.

[0073] (2) The prediction accuracy is significantly improved, which is better than that of mainstream deep learning models.

[0074] The prediction method based on the protein language model has excellent performance advantages in multiple model species, including Caenorhabditis elegans, Homo sapiens and Saccharomyces cerevisiae. Experimental results show that the average AUC value of the method is more than 0.87, which is significantly better than traditional deep learning models (such as CNN, BiLSTM) and standard Transformer architecture (the performance improvement is 15-22%). It is particularly noteworthy that the systematic evaluation on Fusarium graminearum and Homo sapiens data shows that the graph attention network (GAT) module integrated with multi-omics information outperforms graph convolution network (GCN), node embedding model (Node2Vec) and traditional machine learning methods (SVM or RF) in multiple key indicators. These results fully demonstrate the stability and accuracy advantages of the method in multi-species applications.

[0075] (3) It has cross-species generalization ability and is suitable for "zero sample" prediction scenarios.

[0076] The method of the application can completely exclude the target species label in the model construction process, and only rely on the training data of other species (such as humans, nematodes, yeast, etc.), and still can be applied to the essential gene prediction of Fusarium graminearum, and in the test, an AUC of more than 0.7 is achieved. In addition, for non-mammalian eukaryotes (such as Saccharomyces cerevisiae), an AUC>0.7 performance can also be achieved under completely "zero sample" conditions, reflecting good structure semantic embedding migration ability, and is particularly suitable for multi-species unlabeled gene function prediction tasks.

[0077] (4) Good interpretability, key feature role clear and quantifiable.

[0078] By introducing the attention mechanism of the graph neural network, the model can be explicitly weighted modeling at the level of residue or protein node, so as to mine the structural region with higher contribution to the prediction result. The attention visualization heat map shows that the model focuses on the high-conservative functional domain and the core residue area, providing theoretical support for subsequent RNAi target design and verification. In addition, the application constructs an independent multi-omics graph attention network, which integrates features such as expression profile (EXP), conservation score (ORT) and subcellular localization (SUB). Ablation experiments show that removing EXP or ORT features will cause a 6%-8% decrease in AUC, and the GAT model integrating the three information can obtain the highest AUC approximately 0.78 in Fusarium graminearum, further verifying the role of the multi-feature fusion mechanism in improving the interpretability and performance of the model.

[0079] (5) The double-path cross-validation mechanism improves the reliability and biological interpretability.

[0080] The application fuses protein language modeling and multi-omics graph attention network to construct a double-path scoring mechanism, which can effectively improve the reliability and interpretability of the final prediction result. When the prediction results of the two paths are consistent and the scores are both higher than the set threshold (for example, greater than or equal to 0.7), the reliability of the essential gene is significantly improved; if the scores of the two paths are significantly different (for example, the absolute difference is greater than 0.5), the system is automatically marked as “prediction inconsistent” state, prompting experimental verification. The double-path scoring mechanism helps to analyze the differences in attention of different models in terms of internal structural properties and external functional interactions of proteins, thereby enhancing the prediction robustness and decision controllability of the overall system.

[0081] (6) The method is repeatable and open in structure, and has scientific research transformation and popularization value.

[0082] The model used in the application is realized based on the PyTorch Geometric framework, has a good modular structure and openness, supports GPU parallel acceleration, and supports reproduction and expansion. The data (protein sequence, label file, interaction network) used for training and testing are all from public databases, ensuring the standardization and consistency of data sources. Based on the framework, it can be quickly migrated and deployed to other plant pathogenic fungi or agricultural pest species, providing a technical support platform for high-throughput screening of RNAi target genes and developing green pesticides, and has good scientific research transformation prospects. BRIEF DESCRIPTION OF DRAWINGS

[0083] Figure 1 is a schematic diagram of the overall process of the method of the application.

[0084] Figure 2To 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. elegans A 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).

[0085] 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. elegans A 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).

[0086] 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.

[0087] 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.

[0088] Figure 6 This is a flowchart for predicting essential genes based on a multi-omics graph attention network.

[0089] Figure 7 This 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*.

[0090] 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*.

[0091] 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).

[0092] 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*.

[0093] 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 DESCRIPTION

[0094] The application will be further described in conjunction with the specific embodiments. The examples given are only to illustrate the application, and are not intended to limit the scope of the application. The examples provided below can serve as a guide for further improvement by those of ordinary skill in the art, and do not in any way constitute a limitation on the application.

[0095] In the following examples, the experimental methods are conventional methods, and are performed according to the techniques or conditions described in the literature in the art or according to the product instructions, unless otherwise specified. The materials, reagents, etc. used in the following examples can be obtained commercially, unless otherwise specified.

[0096] Example 1. Prediction process of essential genes of non-model organisms based on protein language model and graph attention network.

[0097] This embodiment proposes a system for predicting essential genes of non-model organisms by fusing protein language model and graph attention network—ProGATE (Protein Language Model + Graph Attention + Essentiality) Figure 1 ).

[0098] 1. Data processing module (A).

[0099] This module provides raw data support for model training and reasoning, 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, etc.

[0100] 1.1 Training species selection and target species setting.

[0101] In this embodiment, Fusarium graminearum is taken as the core prediction object. To enhance the generalization ability of the model, five model organisms with relatively complete gene essentiality annotations are additionally selected as training and validation species: Escherichia coli (E. coli); Saccharomyces cerevisiae (S. cerevisiae); Drosophila melanogaster (D. melanogaster); Caenorhabditis elegans (C. elegans); Homo sapiens; This design enables the model to learn protein functional features across species, improving the transfer prediction effect for non-model organisms.

[0102] 1.2 Protein sequence data acquisition and processing.

[0103] F. graminearum protein sequence data is from the FungiDB database, containing a total of 14,143 protein sequences; The protein sequences of the training species are uniformly from the UniProt database; All protein FASTA format files are standardized by uniform format encoding and de-redundancy processing, eliminating redundant records, non-standard amino acid residues, and short sequence noise samples. The cleaned protein sequences are saved under the / data / directory according to the species name for subsequent module calls.

[0104] 1.3 Label construction and weak supervision strategy.

[0105] The labels of the model species in the training species are divided into "essential" and "non-essential" categories according to the OGEE (Online GE Essentiality) and DEG (Database of Essential Genes) database annotation information, serving as supervised training samples; For F. graminearum, based on the functional annotations of "lethal" and "reduced virulence" in the PHI-base database as weakly supervised positive samples (i.e. "essential", such as Chs3b and TRI5 genes), other proteins without annotations and no functional records are considered as negative class (i.e. "non-essential") by default, constituting weakly supervised training labels; All labels are stored in CSV format, with fields including gene ID and binary label (0 or 1: 0 corresponds to "non-essential", 1 corresponds to "essential"), and the file structure is as follows: Gene Label; FGRAMPH1_01G04077 0; FGRAMPH1_01G26781 1; 2. Protein language model feature generation module (B).

[0106] This module is used to extract residue features with semantic context association from protein primary sequences, and to infer potential empty structure contact information, providing support for building input graph structure for subsequent graph neural network modules.

[0107] 2.1 Pre-training (language) model embedding generation.

[0108] ESM-2 (Evolutionary Scale Modeling version 2) is used as a pre-trained protein language model; Each cleaned protein sequence in FASTA format obtained in step 1.2 and the corresponding label of the protein obtained in step 1.3 are input into the ESM-2 model, and the context embedding vector of each amino acid residue is output, which captures its structural and functional semantic information. The context embedding vector is a high-dimensional matrix vector of residue length x feature dimension (1280 in this embodiment), which is used to set the initial features of the graph neural network nodes.

[0109] 2.2 Residue Contact Graph Generation.

[0110] Based on the attention weight matrix of the ESM-2 model and the distance between residue sequences, the attention strength between each pair of residues of the protein is calculated, a undirected graph structure is constructed, and a residue contact graph is obtained, which is composed of nodes and edges: the nodes are residues, and the edges represent pairs of residues with high attention scores, reflecting the potential spatial adjacency relationship between the residue pairs: If a pair of residues meets the double conditions of "attention weight greater than or equal to a set threshold (0.2 in this embodiment)" and "residue sequence distance L≥5", a undirected edge is established between the two residues.

[0111] The graph structure is saved in.pt format for direct reading by the GNN module.

[0112] The residue contact graph reflects the internal potential structural information of the protein and provides input basis for graph neural network modeling.

[0113] 3. Graph Neural Network Module (C).

[0114] This module uses a graph attention network (GAT) to process the residue contact graph of each protein, models the residue contact graph, introduces a multi-head attention mechanism to capture the heterogeneous connection strength between residues, and extracts a protein embedding vector that integrates semantic and structural information to provide input features for subsequent classification tasks.

[0115] 3.1 Graph Structure Construction and Input Definition.

[0116] After step 2, each protein is constructed into a residue contact graph: the nodes are amino acid residues, and the node features are the context embedding vectors of the amino acid residues obtained by ESM-2; the edges between nodes are obtained from the contact relationship inferred by the attention weight and the sequence position, forming a set of undirected edges.

[0117] 3.2 GAT Architecture and Propagation Mechanism.

[0118] GAT assigns different aggregation weights to neighbor nodes in the process of handling the residue contact map of each protein through a learning attention mechanism: each layer of the network aggregates neighbor information through trainable attention weights and outputs multi-level embeddings; GAT supports multi-head attention mechanism to improve model stability and feature expression diversity; finally, the node features are integrated into the overall protein embedding vector through multi-layer stacking and global pooling layer operations.

[0119] 3.3 Output structure and downstream interface.

[0120] Each protein graph is encoded by the graph neural network, and a fixed-length protein embedding vector is output; The embedding vector is used as the input feature of the subsequent classification and adversarial training module.

[0121] 3.4 Structural explainability analysis.

[0122] After the model is trained, the attention weight distribution is extracted and analyzed to identify the key amino acid residues that contribute most to the prediction results, which can assist in target screening and protein structure interpretation.

[0123] Visualize the attention heat map to identify the key residue regions relied on by the model prediction; provide theoretical support for subsequent protein structure function annotation, biological verification and target design.

[0124] 4. Classification and adversarial training module (D).

[0125] This module is responsible for using the classifier to classify the overall protein embedding vector extracted by the graph neural network to determine 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-model species and perturbation inputs.

[0126] 4.1 Classifier structure design.

[0127] The classifier uses a multi-layer perceptron (MLP) structure, which includes: Input layer: receives the overall protein embedding vector output by the graph neural network in step 3; Hidden layer: includes a fully connected operation, a ReLU activation function, and a Dropout operation (to prevent overfitting); Output layer: uses a single neuron for logistic regression, with a Sigmoid activation function, and the output range is normalized to [0, 1], representing the predicted probability that the protein is an "essential gene" (P_Pro).

[0128] 4.2 Loss function and optimization process.

[0129] The classification loss function is the standard binary cross entropy (BCE): , In formula 1 is the true label, is the model prediction probability; To maintain inter-species generalization, no class weight or focal loss is set; The model optimizer uses Adam and combines the learning rate dynamic scheduling mechanism (such as Cosine Decay). The standard binary cross entropy loss function is used for training, and no class weight or focal loss is introduced to ensure the generality of the multi-species scene.

[0130] 4.3 Adversarial perturbation mechanism design.

[0131] Introduce gradient direction-based perturbation to generate adversarial samples in the embedding space to enhance the robustness of the model to unknown perturbations: Adversarial perturbation is only enabled during training; Fast Gradient Sign Method (FGSM) is used to apply perturbation to the embedding vector: Calculate the gradient of the loss function with respect to the input embedding; Add a perturbation with amplitude ε∈ [0.001, 0.01] in the gradient direction; The final training loss is the weighted sum of the original sample loss and the adversarial sample loss: , In formula 2, λ1, λ2 are adjustable weights, usually 0.5:0.5; No perturbation is added during inference, and only the original embedding is used for prediction.

[0132] 5. Multi-omics feature fusion prediction module (E).

[0133] This module is based on external multi-omics information of proteins and protein interaction network to construct an independent + prediction path ( Figure 6 ). Combined with graph attention network modeling, it runs in parallel with the protein language model method to form a dual-path complementary prediction architecture, and enhances the robustness and reliability of the model through integrated fusion strategy.

[0134] 5.1 Multi-omics data integration and protein interaction graph construction.

[0135] Constructing the protein-protein interaction network (PPI) of F. graminearum using STRING database: nodes represent proteins; edges represent physical or functional associations between proteins, set a confidence threshold (such as score ≥ 0.5) to filter low-quality edges; Each node is attached to the following omics features, resulting in a multi-omics protein-protein interaction network (multi-omics PPI): Expression Profile: Expression profile reflects the transcription level of target genes in cells. To obtain the gene expression data of F. graminearum, the present application downloads the transcriptome data of the species F. graminearum from the FungiDB database and calculates the TPM (Transcripts Per Million) value. The TPM value as a unit can quantitatively reflect the expression level of genes in transcriptome sequencing and provide strong support for the prediction of protein function. Orthology Score: Obtained from cross-species alignment (such as based on protein homology) by using BLAST or Diamond method; Orthology Score: used to evaluate the protein homology between the target species and other species, revealing the conservation of proteins in different species. The present application uses the multi-species protein sequence data provided by the NCBI database and uses the DIAMOND tool to perform efficient cross-species alignment of the protein sequence of F. graminearum with the protein sequence of other species, and the obtained alignment result provides a cross-species conservation score for the protein of each gene of F. graminearum.

[0136] Subcellular Localization: refers to the distribution of proteins in functional regions within cells, which is crucial for understanding protein function. To obtain the subcellular localization information of F. graminearum proteins, the present application extracts relevant data from the COMPARTMENTS database and UniProtKB database, and uses One-Hot Encoding technology to label the subcellular localization (such as nucleus, mitochondria, etc.) of proteins. This data will further enrich the node features in the protein-protein interaction network, helping to better understand the biological function of proteins.

[0137] Multi-omics feature fusion into a high-dimensional vector, such as formula 3: .

[0138] 5.2 Multi-omics graph attention modeling process.

[0139] The multi-omics PPI graph constructed in step 5.1 is modeled using a graph attention network (GAT) to propagate information on the PPI graph and nonlinearly aggregate the omics features of each node in the multi-omics PPI: Each layer performs weighted aggregation of neighbor node features; Multi-head attention is introduced to enhance feature expression capability; Finally, attention pooling is used to converge into a fixed-length protein representation; The Sigmoid activation function is used to output the predicted probability value P_GAT ∈ [0,1] of the protein, representing the confidence of the gene as an essential gene.

[0140] 5.3 Dual-path prediction cross-fusion mechanism.

[0141] The system integrates two prediction paths: 1) Protein language model (Pro path): Based on the residue contact map within a single sequence extracted by the protein language model (ESM-2) and the graph neural network (GAT), the prediction probability P_Pro is output, which is the prediction probability ŷ of the "essential gene" output in step 4; 2) Multi-omics graph attention network (GAT path): Based on the PPI and multi-omics features, a graph attention network is constructed, and the prediction probability P_GAT is output, which is the prediction probability value P_GAT output in step 5.2.

[0142] The following fusion judgment strategy is adopted: If P_Pro≥T and P_GAT≥T (T is an adaptive threshold, which is 0.7 in this embodiment), it is determined as a high-confidence essential gene; 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; If both are <0.3, it is considered as a non-essential gene.

[0143] Example 2. Model training and evaluation process in the prediction process of non-model organism essential genes.

[0144] This embodiment details the complete process of the model from data preprocessing to training, adversarial optimization, and evaluation, aiming to ensure the repeatability, verifiability, and cross-species applicability of the model results.

[0145] 1. Training data preparation process.

[0146] The data of each species is stored after uniform preprocessing; Label information is uniformly mapped to "mandatory (1)" and "non-mandatory (0)", and automatic completion is achieved by using a Python script (label_mapping.py); Fusarium graminearum is trained by weak supervision, and evaluation is performed by strong label method; Data division method: training set 80%, validation set 10% and test set 10%, which can be selected as "species division" or "mixed division" to evaluate the generalization ability or single species performance.

[0147] 2. Model training process.

[0148] All model modules (protein language model feature extraction, graph attention network, classifier, and fusioner) are integrated in a unified training main program (main_train.py); Supporting independent training and joint end-to-end training in two modes; The main training parameter configuration used is shown in Table 1:

[0149] 3. Adversarial training details (for graph embedding).

[0150] The adversarial perturbation module is embedded in the front end of the classifier, and a perturbation is added to the protein embedding vector output by GAT: Perturbation method: Fast Gradient Sign Method (FGSM); The perturbation amplitude ε is controlled at 0.001 ~ 0.01; The adversarial sample loss and the original sample loss are weighted to form the final loss; Adversarial perturbation is added in the training stage, and no perturbation mechanism is used in the inference stage.

[0151] 4. Model evaluation and performance indicators.

[0152] The models used for evaluation include five models: the combination of the ESM model and the graph attention network model of the method of the application (ESM+GAT), the ESM model, the convolutional neural network (CNN, Convolutional Neural Network), the long short-term memory network (LSTM, Long Short-Term Memory), and the Transformer neural network model. The model performance evaluation indicators include: AUC-ROC (Area Under the Receiver Operating Characteristic Curve): to evaluate the overall discriminant ability of the model for positive and negative samples; AUPRC (Area Under the Precision-Recall Curve): suitable for situations where the positive and negative sample proportions are not balanced, and measures the effectiveness of the model under high confidence prediction; ACC (Accuracy): refers to the proportion of instances that the model correctly predicts out of the total number of instances; 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. 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.

[0153] 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.

[0154] 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.

[0155] 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.

[0156] 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 5In the middle A) generally get AUC > 0.7 excellent effect under different training-test combination, embody its cross-species generalization ability.

[0157] Figure 7 It is a performance comparison chart of multi-omics graph attention network (GAT) model and other machine learning models combined with multi-omics. The performance indicators include the AUC value (ROC) of the model, the AUPRC value, the specificity (Specificity, SP represents) value and the Matthew correlation coefficient (MCC) value. The comparison objects include graph attention network model (GAT), node embedding model (Node2Vec), multilayer perceptron (Multilayer Perceptron, MLP), random forest (Random Forest, RS), naive Bayes (Naive Bayes, NB) and support vector machine (SVM) and the like. Because the GAT model adopts an adaptive attention mechanism, the performance is better than that of other static adjacency structure methods, and the overall performance of ROC, AUPRC, Specificity and MCC is optimal.

[0158] Figure 8 It is a performance comparison chart of multi-omics graph attention network (GAT) and various graph learning methods in multiple species. The performance indicators include the AUC value (ROC) of the model, the AUPRC value, the specificity (Specificity, SP represents) value, the accuracy value (Accuracy, ACC) value and the Matthew correlation coefficient (MCC) value. A is the performance value comparison in fusarium graminearum; B is the performance value comparison in saccharomyces cerevisiae; C is the performance value comparison in escherichia coli; D is the performance value comparison in humans; E is the performance value comparison in drosophila melanogaster; F is the performance value comparison in caenorhabditis elegans. Various graph learning methods include graph convolution network (GCN), degree centrality (Degree Centrality, DC) and clustering centrality (Clustering Coefficient, CC).

[0159] By comparing the performance indicators of GAT, GCN, DC and CC in each species, the graph attention network GAT model adopted by the application has the advantages of Figure 8 In the middle sauce red represents) because of the adoption of adaptive attention mechanism, the performance is better than that of GCN and other static adjacency structure methods.

[0160] 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.

[0161] 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.

[0162] 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.

[0163] 5. Results visualization and report generation.

[0164] The loss curve and accuracy curve during model training are output to / outputs / plots / ; All evaluation processes are encapsulated in evaluate.py, supporting automatic multi-model parallel testing and chart (CSV / PDF format) output; The model supports attention heatmap output to reveal the key residues or adjacent proteins that the model focuses on when distinguishing the "required" label.

[0165] The above has been described in detail. For those skilled in the art, without departing from the purpose and scope of the present application, and without unnecessary experiments, the present application can be implemented in a wider range under equivalent parameters, concentrations and conditions. Although the present application gives a special example, it should be understood that further improvements can be made to the present application. In summary, according to the principle of the present application, this application intends to include any change, use or improvement of the present application, including changes made by conventional techniques known in the art, which are outside the scope disclosed in the present application.

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 the corresponding label for each protein based on the protein sequence of the target species and the protein sequences of n model organisms, respectively, using annotation. A2) Embed each protein sequence and its corresponding tag into a protein language model, and output the context embedding vector of amino acid residues in the protein; construct a residue contact map for each protein based on the context embedding vector; 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. 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 steps also include: 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; 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.

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

4. 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.

5. 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. 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 map is constructed for each protein based on the context embedding vector; 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. The target species is a non-model organism species or a species with incomplete gene annotation; n is greater than or equal to 1.

6. The method according to claim 5, characterized in that: The method further includes the following steps: B5) 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. 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.

7. The method according to claim 5 or 6, characterized in that: The graph neural networks described in B3) and B5) are graph attention networks.

8. A computer program product, comprising a computer program, characterized in that: When the computer program is executed by the processor, it implements the steps of A1)-A4) in claim 1 or the steps of A1)-A6) in any of claims 1-3.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of A1)-A4) in claim 1 or the steps of A1)-A6) in any of claims 1-3.

10. 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 8 and / or the computer-readable storage medium according to claim 9: 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