Gene function prediction method based on semantic correspondence of regulatory region

Abstract

The invention discloses a gene function prediction method based on semantic correspondence of a regulatory region. The method comprises the following steps: firstly, constructing an inter-species regulation semantic correspondence relationship data set, constructing an artificial intelligence model structure, then, constructing a cross-species semantic correspondence network, and finally, carrying out function annotation on a target gene or identifying a candidate gene with a specific function in a target species. The accuracy of the PhytoBabel model constructed by the method is obviously higher than that of other model structures. By utilizing the method disclosed by the invention, the genes ZmERF104 and ZmGRF16 for promoting the regeneration of the corn somatic embryos and the gene ZmNAC17 for inhibiting the regeneration of the somatic embryos, which cannot be found by the traditional method, are successfully identified.

Description

Technical Field

[0001] This invention belongs to the fields of biology and computer technology, and specifically relates to a gene function prediction method based on the semantic correspondence of regulatory regions. Background Technology

[0002] Gene function prediction has always been a fundamental goal of plant genetics research. Ideally, gene function is defined by the biochemical or physiological effects of its natural or artificial gene perturbations. However, relatively few plant genes meet this stringent standard. Taking maize, one of the major crops, as an example, three studies recorded 1111, 642, and 693 functional validation genes, respectively, totaling only about 210 maize genes (0.5% of all genes). Even in Arabidopsis thaliana, a major model plant, the number of functional validation genes is approximately 7700 (28% of all genes). These statistics indicate a need for new strategies to transfer cross-species gene function knowledge for reverse genetics research.

[0003] Gene function research has always been a fundamental goal of plant genetics research. In crops, the function of specific genes is often directly related to core agronomic traits such as yield composition, nutritional quality, disease resistance, and stress tolerance (drought / salt tolerance). Accurately analyzing the biological functions of crop genes is of irreplaceable importance for a deeper understanding of their growth and development mechanisms, optimizing genetic improvement strategies, and accelerating the breeding process. Ideally, gene function is defined by the biochemical or physiological effects of natural or artificial gene perturbations. However, relatively few plant genes meet this stringent standard. Taking maize, one of the major crops, as an example, three studies recorded 1111, 642, and 693 functional validation genes, respectively, totaling only about 210 maize genes (0.5% of all genes). Therefore, high-throughput methods are needed to predict the function of a large number of unknown crop genes to assist in optimizing agronomic traits.

[0004] Currently, gene function prediction is mainly based on two strategies. The first relies on network biology, including co-expression networks, transcriptional regulatory networks, protein-protein interaction networks, and co-essence networks. Among these, driven by the rapid growth of publicly available transcriptome datasets, co-expression networks are the most widely used gene function prediction method, based on the hypothesis that "co-expression equals functional correlation." This method utilizes transcriptome data to construct a network with genes as nodes and co-expression relationships as edges by calculating the expression similarity of gene pairs under various conditions (such as the Pearson correlation coefficient). Highly interconnected genes are often clustered into modules, which are often enriched with specific biological functions. If an unknown functional gene is highly co-expressed with a known functional gene, especially if it is located at the core of a functionally defined module, it can be inferred that it participates in similar biological processes. Methods such as Weighted Co-expression Network Analysis (WGCNA) further enhance the robustness and biological interpretability of the network.

[0005] The second strategy is based on protein homology retrieval. This method is based on the principle of evolutionary conservation: homologous proteins derived from a common ancestor are often highly conserved in both sequence and function. Therefore, if a protein with an unknown function has significant sequence similarity to a protein with a known function, it can be reasonably inferred that they are functionally similar. In application, tools such as BLAST are typically used to compare the target sequence with annotation databases to identify statistically significant homology relationships. To improve accuracy, researchers further distinguish between orthologous and paralogous homology, with orthologous homology being considered a reliable basis for functional transfer due to its stronger functional conservation.

[0006] Traditional methods for gene function prediction based on network biology are typically limited to single species. When the number of known functional genes in a species is small (e.g., only 0.5% of the genes in maize have known functions), it is difficult to effectively annotate the functions of unknown genes based on clustering results. Furthermore, traditional networks usually rely on medium to large-scale laboratory experiments, which significantly increases the cost of gene function prediction.

[0007] While protein homology searches can identify homologous genes with significant sequence similarity, they struggle to discover non-homologous genes within the same biological pathway that are functionally similar or redundant. Furthermore, paralogous genes generated after gene duplication events, despite maintaining high overall sequence similarity, may have undergone functional differentiation (e.g., neofunctionalization or subfunctionalization). Indiscriminately transferring known functions to such genes can easily lead to erroneous functional inferences. Therefore, relying solely on homology searches may not only miss crucial functional associations but also introduce systematic annotation biases, limiting the comprehensive analysis of complex biological processes. Summary of the Invention

[0008] The purpose of this invention is to provide a method for constructing a cross-species semantic network of regulatory regions using only gene DNA sequences, thereby enabling cross-species gene function prediction.

[0009] A gene function prediction method based on the semantic correspondence of regulatory regions is performed according to the following steps: (1) Construction of a dataset of regulatory semantic correspondences between species: Select representative transcripts of genes, perform BLASTP comparison analysis between species for the whole proteome of each species, extract non-coding region sequences with semantic correspondences of regulatory regions before translation initiation sites and after translation termination sites to form a positive set of models, and generate a negative set by re-pairing sequence pairs; (2) Artificial intelligence model structure construction: construct a PhytoBabel model for sequence semantic matching, process the input features of the paired control region sequences, input the two sequences to a pair of shared weight feature extraction branches, concatenate the features from the two sequences in the channel dimension to obtain a comprehensive feature matrix, and input it into a cross-sequence Attention layer. The output of Attention is further compressed into a one-dimensional feature vector by a multilayer perceptron and connected to a three-layer fully connected network. Softmax is used as the final activation function to output a binary classification result of whether the two control region sequences have semantic similarity. (3) Construction of cross-species semantic correspondence network: The reference gene set is the gene with known functional annotation in the reference species, and the query gene set is the whole gene of the query species; the trained PhytoBabel model is used to predict the semantic score of the regulatory region sequence of all reference gene-query gene pairs. (4) Gene function inference: including functional annotation of target genes and identification of candidate genes with specific functions in target species.

[0010] The selection criteria for representative transcripts of the gene in step (1) include: if multiple transcripts correspond to the same protein, the longest transcript is selected; if multiple protein isoforms exist, BLASTP is used to compare each protein isoform with all homologous proteins of the species, and the representative isoform is selected based on the highest cumulative position score, and the representative transcript and its corresponding protein are selected.

[0011] The process of generating the negative set in step (1) is as follows: keep each control region sequence unchanged, but randomly shuffle its pairing relationship, thereby destroying its control semantic correspondence, while keeping the sequence composition and length distribution consistent.

[0012] The input to the PhytoBabel model in step (2) includes the regulatory region sequences of two genes and the information of their corresponding species. The sequences are one-hot encoded into a 4×4000 matrix. The species information is mapped to a vector of length 4 through an embedding layer and expanded into a 4×4000 matrix of the same size through replication. The sequence matrix is ​​added to the species embedding matrix to form a joint representation of each sequence.

[0013] Each branch in the shared weight feature extraction branch in step (2) contains a convolutional neural network layer, a self-attention layer, and three ResNet blocks for extracting local and global sequence features; the feature output of this branch is represented as a 24×64 matrix.

[0014] The comprehensive feature matrix mentioned in step (2) is a 48×64 comprehensive feature matrix.

[0015] Step (3) involves predicting the semantic score of the regulatory region sequence. Specifically, for gene pairs with a model prediction score higher than 0.99, an edge is established between the genes of the two species to construct a cross-species semantic correspondence network, where nodes represent genes and edges represent that their regulatory region sequences have a high degree of semantic correspondence in the model.

[0016] The functional annotation of the target gene in step (4) is as follows: if the user wants to predict the function of a gene in a target species, the node corresponding to the gene is located in the cross-species semantic correspondence network constructed in step three, and all the reference species genes connected to it are extracted; then, based on the existing GO annotations of these reference genes, or by integrating their functional descriptions with the help of a large language model, the potential function of the target gene is inferred.

[0017] Step (4) involves identifying candidate genes with specific functions in the target species. Specifically, this involves finding genes in the target species that are related to a certain target function, selecting all reference species gene nodes with annotations for that function in the semantic correspondence network, counting the number of edges between each gene in the target species and these reference genes as a measure of semantic relevance, and sorting them from high to low according to the number of edges, thereby obtaining a set of candidate genes for the target function.

[0018] The function and because ZmERF104 , ZmGRF16 and ZmNAC17 Genes; the stated ZmERF104 , ZmGRF16 Genes promote the regeneration of maize somatic embryos.

[0019] The beneficial effects of this invention: This invention proposes for the first time a method for transferring gene function knowledge across species and predicting gene function based solely on the coding information of cis-regulatory region sequences. This method overcomes several limitations of existing technologies in function transfer, specifically including: 1. Breaking through the limitation of network biology methods that can only transfer knowledge within a single species. Traditional gene co-expression networks or regulatory networks rely on a large number of known functional genes, but the proportion of known functional genes in staple crops such as maize is extremely low (approximately 0.5% in maize), resulting in limited gene function prediction capabilities. In contrast, approximately 28% of genes in the model plant Arabidopsis thaliana have functional annotations, but existing network methods cannot directly transfer them to crops. This invention constructs a cross-species semantic correspondence network of regulatory regions, enabling the effective transfer of rich gene function knowledge from Arabidopsis thaliana to distantly related crops. 2. Breaking through the limitations of traditional homologous gene transfer by introducing a transfer mechanism based on the semantics of regulatory regions. Previous cross-species function inference relied on protein sequence homology, but a large number of genes between distantly related species do not have a reliable one-to-one homology relationship. This invention, by analyzing the semantic correspondences in regulatory region sequences of different species, enables cross-species transfer of gene function without homology requirements, significantly expanding the applicability of functional prediction. The PhytoBabel model constructed in this invention exhibits significantly higher accuracy than other model structures. Using the method of this invention, genes promoting maize somatic embryo regeneration that could not be found using traditional methods were successfully identified. ZmERF104 , ZmGRF16 and genes that inhibit somatic embryo regeneration ZmNAC17 . Attached Figure Description

[0020] Figure 1 This is a flowchart of the technical method of the present invention.

[0021] Figure 2 This is a schematic diagram of the PhytoBabel model structure.

[0022] Figure 3 This is a comparison of model structures.

[0023] Figure 4 This is a semantic correspondence network graph.

[0024] Figure 5 The diagram shows the effect of candidate genes found for PhytoBabel on the regeneration of maize somatic embryos and a schematic diagram of representative somatic embryos produced by each gene. Detailed Implementation

[0025] To facilitate understanding of the present invention, a more comprehensive description will be provided below. However, the present invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided to provide a thorough and complete understanding of the disclosure of the present invention. Example 1

[0026] like Figure 1-4 As shown, the gene function prediction method based on the semantic correspondence of regulatory regions in this invention is carried out according to the following steps: (1) Construction of a dataset of semantic correspondences of regulation among species: 1) Selection of representative transcripts: For genes with alternative splicing, the following criteria are used: 1) If multiple transcripts correspond to the same protein, the longest transcript is selected; 2) If multiple protein isoforms exist, BLASTP is used to compare each protein isoform with all homologous proteins of 15 plant species (Arabidopsis thaliana, Populus tomentosa, cucumber, Alfalfa, soybean, sugar beet, grape, potato, tomato, sorghum, maize, Setaria viridis, millet, rice, and Brachypodium distichum), and the representative isoform is selected based on the highest cumulative locus score, and the representative transcript and its corresponding protein are selected.

[0027] 2) Pairwise BLASTP alignment analysis was performed on the whole proteomes of 15 representative plant species. For any two species A and B, the best hit of a protein x in species A in species B is defined as follows: First, calculate the BLASTP alignment results of protein x with all protein sequences in species B; then, sum the bit-scores of each aligned fragment (if there are multiple high-scoring fragments) to obtain the total bit-score of x with each protein in species B; the protein y with the highest score is considered the best match of x in species B.

[0028] If a pair of proteins (a, b) (where a ∈ Species A, b ∈ Species B) simultaneously satisfy the following two conditions: 1. In a forward alignment, the best match for a is b; 2. In a reverse alignment (i.e., searching in species A with b as the query sequence), the best match for b is a; then the protein pair (a, b) is said to constitute a candidate gene pair with semantic correspondence to a regulatory region. 3) To further improve the reliability of candidate gene pairs, the above RBH gene pairs were further screened, and only gene pairs whose length of the comparable region (homological fragments shown by BLASTP, including mismatches and gaps) accounted for more than 95% of the length of both protein sequences were retained.

[0029] 4) Extract the non-coding region sequences before the translation start site and after the translation end site to form the positive set of the model. The negative set is generated by re-pairing sequence pairs: keep each regulatory region sequence itself unchanged, but randomly shuffle its pairing relationship (for example, if the original positive set is x–y, a–b, then the negative set can be x–b, a–y), thereby destroying its regulatory semantic correspondence.

[0030] (2) Construction and Training of the Artificial Intelligence Model: To handle the input features of paired regulatory region sequences, we constructed a PhytoBabel model specifically for sequence semantic matching. The model input includes the regulatory region sequences of two genes and information about their corresponding species. The sequences are one-hot encoded into a 4×4000 matrix; the species information is mapped to a vector of length 4 through an embedding layer and expanded into a 4×4000 matrix of the same size through replication. The sequence matrix and the species embedding matrix are added together to form a joint representation of each sequence.

[0031] The two sequences are input into a pair of feature extraction branches with shared weights. Each branch contains a convolutional neural network (CNN) layer, a self-attention layer, and three ResNet blocks to extract local and global sequence features. The feature output of this branch is represented as a 24×64 matrix.

[0032] Subsequently, the features from the two sequences are concatenated along the channel dimension to obtain a 48×64 comprehensive feature matrix, which is then input into a cross-sequence Attention layer to explicitly model the semantic correspondence between the sequences. The output of the Attention layer is further compressed into a one-dimensional feature vector by a multilayer perceptron (MLP) and then connected to a three-layer fully connected (Dense) network with softmax as the final activation function, outputting a binary classification result indicating whether the two control region sequences have semantic similarity.

[0033] After comparison, the model trained using the dataset constructed in step one outperforms the Enformer, TM-VEC, and improved-Enformer and improved-TM-VEC models optimized for this task in the literature in terms of prediction accuracy. (3) Construction of cross-species semantic correspondence network: Genes with known functional annotations in the reference species (Arabidopsis thaliana) are used as the reference gene set, and all genes in the query species (maize) are used as the query set. The trained PhytoBabel model is used to predict the semantic scores of regulatory region sequences for all reference gene-query gene pairs. For gene pairs with model prediction scores higher than 0.99, edges are established between the genes of the two species to construct a cross-species semantic correspondence network, where nodes represent genes and edges represent that their regulatory region sequences have a high degree of semantic correspondence in the model.

[0034] (4) To meet different application needs, we designed two gene function inference schemes: (1) Functional annotation of the target gene. If the user wants to predict the function of a gene in a target species, the node corresponding to the gene is located in the cross-species semantic correspondence network constructed in step three, and all the reference species genes connected to it are extracted. Then, based on the existing GO annotations of these reference genes, or by integrating their functional descriptions with the help of a large language model, the potential function of the target gene is inferred.

[0035] (2) Identify candidate genes with specific functions in the target species. If the user wants to find genes related to a certain target function in the target species, first select all reference species gene nodes with that function annotation in the semantic correspondence network. Then count the number of edges between each gene in the target species and these reference genes as a measure of semantic relevance, and sort them from high to low according to the number of edges to obtain the set of candidate genes for the target function.

[0036] Using this method, we obtained a list of candidate genes for promoting somatic embryo regeneration in maize from genes in Arabidopsis thaliana, and selected four of them for experimental verification. The results showed that... ZmGRF16 and ZmERF104 Both significantly increased the number of somatic embryos per immature embryo. Additionally, ZmNAC17+WU S co-transformation led to a significant reduction in somatic embryo regeneration.

[0037] Example 2 ZmERF104、ZmGRF16、ZmNAC17 Gene function verification Immature ears of the fast-flowering maize variety Mini-Maize, used for transformation experiments, were harvested 8 days after pollination. The ears were surface-sterilized in 1.65% sodium hypochlorite for 20 minutes, followed by rinsing three times (5 minutes each) with sterile distilled water. Immature embryos were removed and transferred to 2 mL centrifuge tubes containing 700A liquid medium (4.4 g / L MS basal medium, 68.5 g / L sucrose, 36 g / L glucose, 1.5 mg / L 2,4-D; pH adjusted to 5.8 with KOH, filtered and sterilized), and washed three additional times with fresh 700A medium.

[0038] Containing recombinant plasmids ( WUS; WUS+ZmGRF16; WUS+ZmNAC17; WUS+ZmERF104 ,in WUS The Agrobacterium tumefaciens strain EHA105 (as a control) was cultured overnight at 28°C with shaking in DYT medium (16 g / L tryptone, 10 g / L yeast extract, 5 g / L NaCl) supplemented with 100 mg / L kanamycin, 30 mg / L rifampin, and 200 μM acetylsylgenone. After collection, the bacterial cells were resuspended in 700A liquid medium until the optical density (OD600) reached 0.4. Immature embryos were immersed in the Agrobacterium suspension for 5 minutes and then placed in 562V medium (4.0 g / L N6 basal salt mixture, 30 g / L sucrose, 2.0 mg / L 2,4-D, 1 mg / L silver nitrate, 100 μM acetylsylgenone, pH adjusted to 5.8 with KOH) for co-culture at 21°C. After overnight incubation, infected embryos were placed on 605T medium (4.3 g / L basal salt mixture, 60 ml / L N6 macrosalt, 0.6 ml / L B5 microsalt, 0.4 ml / L Eriksson vitamins, 0.6 g / L S&H vitamins, 6 ml / L sodium ferrous sulfate stock solution, 1.68 g / L potassium nitrate, 0.2 mg / L thiamine hydrochloride, 0.3 g / L casein hydrolysate, 0.8 mg / L 2,4-D, 2 g / L L-proline, 20 g / L sucrose, 0.6 g / L glucose, 100 mg / L termethin, 3.4 mg / L silver nitrate, pH adjusted to 5.8 with KOH), scutellum facing upwards, and cultured in the dark at 26°C. After 6 days of culture, somatic embryo development was observed on the scutellum surface.

[0039] Limited by the number of somatic embryos obtained at one time, we performed the process in two stages. ZmERF104、ZmGRF16 and ZmNAC17 Experimental verification was conducted based on the above method. The results show that, compared to using alone... WUS (On average, 0.17 somatic embryos are produced). WUS+ ZmERF104 (Average production of 2.24 somatic embryos) and WUS+ZmGRF16(On average, 1.27 somatic embryos were produced) This can significantly improve the efficiency of somatic embryo production. WUS+ZmNAC17 (On average, 0.02 individual cell embryos are produced) can significantly inhibit the efficiency of cell embryo production.

[0040] The above experimental results prove that ( Figure 5 ), ZmERF104 , ZmGRF16 Genes promote somatic embryo regeneration in maize. ZmNAC17 Genes suppress somatic embryo regeneration.

[0041] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this invention patent should be determined by the appended claims.

Claims

1. A gene function prediction method based on the semantic correspondence of regulatory regions, characterized in that, Follow these steps: (1) Construction of a dataset of regulatory semantic correspondences between species: Select representative transcripts of genes, perform BLASTP comparison analysis between species for the whole proteome of each species, extract non-coding region sequences with semantic correspondences of regulatory regions before translation initiation sites and after translation termination sites to form a positive set of models, and generate a negative set by re-pairing sequence pairs; (2) Artificial intelligence model structure construction: construct a PhytoBabel model for sequence semantic matching, process the input features of the paired control region sequences, input the two sequences to a pair of shared weight feature extraction branches, concatenate the features from the two sequences in the channel dimension to obtain a comprehensive feature matrix, and input it into a cross-sequence Attention layer. The output of Attention is further compressed into a one-dimensional feature vector by a multilayer perceptron and connected to a three-layer fully connected network. Softmax is used as the final activation function to output a binary classification result of whether the two control region sequences have semantic similarity. (3) Construction of cross-species semantic correspondence network: The reference gene set is the gene with known functional annotation in the reference species, and the query gene set is the whole gene of the query species; the trained PhytoBabel model is used to predict the semantic score of the regulatory region sequence of all reference gene-query gene pairs. (4) Gene function inference: including functional annotation of target genes and identification of candidate genes with specific functions in target species.

2. The gene function prediction method based on semantic correspondence of regulatory regions according to claim 1, characterized in that, The selection criteria for representative transcripts of the gene in step (1) include: if multiple transcripts correspond to the same protein, the longest transcript is selected; if multiple protein isoforms exist, BLASTP is used to compare each protein isoform with all homologous proteins of the species, and the representative isoform is selected based on the highest cumulative position score, and the representative transcript and its corresponding protein are selected.

3. The gene function prediction method based on the semantic correspondence of regulatory regions according to claim 1, characterized in that, The process of generating the negative set in step (1) is as follows: keep each control region sequence unchanged, but randomly shuffle its pairing relationship, thereby destroying its control semantic correspondence, while keeping the sequence composition and length distribution consistent.

4. The gene function prediction method based on semantic correspondence of regulatory regions according to claim 1, characterized in that, The input to the PhytoBabel model in step (2) includes the regulatory region sequences of two genes and the information of their corresponding species. The sequences are one-hot encoded into a 4×4000 matrix. The species information is mapped to a vector of length 4 through an embedding layer and expanded into a 4×4000 matrix of the same size through replication. The sequence matrix is ​​added to the species embedding matrix to form a joint representation of each sequence.

5. The gene function prediction method based on semantic correspondence of regulatory regions according to claim 1, characterized in that, Each branch in the shared weight feature extraction branch in step (2) contains a convolutional neural network layer, a self-attention layer, and three ResNet blocks for extracting local and global sequence features; the feature output of this branch is represented as a 24×64 matrix.

6. The gene function prediction method based on semantic correspondence of regulatory regions according to claim 1, characterized in that, The comprehensive feature matrix mentioned in step (2) is a 48×64 comprehensive feature matrix.

7. The gene function prediction method based on semantic correspondence of regulatory regions according to claim 1, characterized in that, Step (3) involves predicting the semantic score of the regulatory region sequence. Specifically, for gene pairs with a model prediction score higher than 0.99, an edge is established between the genes of the two species to construct a cross-species semantic correspondence network, where nodes represent genes and edges represent that their regulatory region sequences have a high degree of semantic correspondence in the model.

8. The gene function prediction method based on the semantic correspondence of regulatory regions according to claim 1, characterized in that, The functional annotation of the target gene in step (4) is as follows: if the user wants to predict the function of a gene in a target species, the node corresponding to the gene is located in the cross-species semantic correspondence network constructed in step three, and all the reference species genes connected to it are extracted; then, based on the existing GO annotations of these reference genes, or by integrating their functional descriptions with the help of a large language model, the potential function of the target gene is inferred.

9. The gene function prediction method based on the semantic correspondence of regulatory regions according to claim 1, characterized in that, Step (4) involves identifying candidate genes with specific functions in the target species. Specifically, this involves finding genes in the target species that are related to a certain target function, selecting all reference species gene nodes with annotations for that function in the semantic correspondence network, counting the number of edges between each gene in the target species and these reference genes as a measure of semantic relevance, and sorting them from high to low according to the number of edges, thereby obtaining a set of candidate genes for the target function.

10. The functional gene obtained by the method of claim 1, characterized in that, The function and because ZmERF104 , ZmGRF16 and ZmNAC17 Genes; the stated ZmERF104 , ZmGRF16 Genes promote the regeneration of maize somatic embryos.

Images