Prediction method and system for transcriptional regulation mechanism of downstream pathways of ginsenoside synthesis

Abstract

The prediction method and system of the transcriptional regulatory mechanism of the downstream pathway of ginsenoside synthesis belongs to the technical field of plant regulatory mechanism, and solves the problem that the prediction of the transcriptional regulatory mechanism of the downstream pathway of the synthesis of non-model organism ginseng cannot be effectively realized. The method of the present invention includes: performing transfer learning through the BERT model in the field of artificial intelligence, for example, using a large number of plant gene expression data of model plants such as Arabidopsis thaliana and rice for pre-training (pre-train) to obtain the expression pattern of plant genes, and then fine-tuning the pre-training model using the gene expression data of ginseng to obtain the special gene expression pattern of medicinal plants, which takes into account the generality of large-scale plants and pays attention to the particularity of medicinal plants, thereby improving the credibility and reliability of the network. The present invention is applicable to the annotation method of ginseng regulatory network.

Description

Technical Field

[0001] The present application relates to the technical field of plant regulatory mechanisms, and in particular to the prediction of transcriptional regulatory mechanisms in the downstream pathways of ginsenoside synthesis. Background Art

[0002] The biosynthesis of ginsenosides can be divided into two components: upstream and downstream. Figure 1 . The upstream synthesis is mainly carried out through the mevalonate pathway (MVA pathway), and the key enzymes involved in this pathway have been successfully identified. The exact steps involved in the downstream process remain to be determined. For example, the functions of key enzymes (OSC, CYP450 and UGT) in the post-synthesis modification stage of ginsenosides in plants are still unknown, and there are few studies on the specific catalytic process, structural modification and transcriptional regulation mechanism (transcription factors, miRNA, etc.) of key enzymes. Therefore, people's understanding of the ginsenoside biosynthesis pathway is still not systematic and comprehensive enough.

[0003] Current studies have found that several steps in the downstream biosynthesis of ginsenosides need to be further elucidated, such as the formation of the ginsenoside skeleton, the synthesis and skeleton modification of sugar donors, the identification of UDP glycosyltransferases and CYP450 genes, and how various ginsenosides are produced through multiple oxidation reactions catalyzed by CYP450 enzymes and glycosylation reactions catalyzed by UGTs. Certain UGT and CYP450 genes have been identified as candidate genes involved in the downstream ginsenoside biosynthesis of American ginseng, such as UGT917, UGT913, UGT716, and UGT711, which are positively correlated with pseudoginsenoside F11, trigonosaponin R1, trigonosaponin R2 in American ginseng, and ginsenoside RT5 in American ginseng. Four cytochrome P450 enzymes (CYP450) have been identified as potential catalysts for the hydroxylation of specific positions (C-16 and C-30) of the triterpene skeleton in the biosynthesis of ginsenosides. In addition, UDP-xylose synthase has been selected as a candidate enzyme that may be involved in the biosynthesis of ginsenoside Rb3.

[0004] The relevant computational analysis studies are as follows: the literature constructed a co-expression network of transcripts of the PgCYP gene in the roots of 414 ginsengs of different ages, and found that its expression level was significantly correlated with the content of ginsenosides; the literature speculated that IspD might be the key enzyme for producing ginsenosides through the MEP pathway through the co-expression network and expression spectrum analysis of 15-year-old and 5-year-old ginsengs; the literature analyzed the differential expression patterns of genes through hierarchical clustering, and used the WGCNA co-expression network to find that CYP and UGT showed tissue-specific expression patterns; the literature pointed out that although the co-expression network itself does not have a causal relationship, it can achieve a good simulation effect when combined with the transcriptional regulation map.

[0005] At present, most of the research on transcriptional regulatory networks is based on bioinformatics computational models such as differential expression and co-expression analysis implemented by biologists using software. This type of algorithm research is still in its infancy, and the model is relatively simple. Moreover, medicinal plants are non-model organisms. Although they are plants, they have their own inherent characteristics. These characteristics require a large amount of plant data to support them. According to the current situation, computer experts are urgently needed to explore efficient large-model algorithms, transition from the perspective of plant systems to the perspective of medicinal plants, and construct molecular mechanism analysis models from general to specific. Ginseng is a non-model organism, and sequencing is difficult. There is still a lack of a complete reference genome and no unified genome annotation method. Summary of the invention

[0006] The purpose of the present invention is to solve the problem that the prediction of the transcriptional regulatory mechanism of the downstream pathway of the synthesis of the non-model organism ginseng cannot be effectively realized, and a method and system for predicting the transcriptional regulatory mechanism of the downstream pathway of ginsenoside synthesis are provided.

[0007] The present invention is achieved through the following technical solutions. On the one hand, the present invention provides a method for predicting the transcriptional regulation mechanism of the downstream pathway of ginsenoside synthesis, the method comprising:

[0008] Step 1: Convert ginseng transcriptome sequencing data SRA into ginseng gene expression data and obtain a list of ginseng transcription factors;

[0009] Step 2: Use the BERT model to pre-train the gene expression data of the model plant;

[0010] Step 3: Fine-tune the BERT model using ginseng gene expression data, including:

[0011] Input the ginseng gene expression data in step 1 into the pre-trained BERT model, perform a masking strategy on non-ginseng TF genes, predict the remaining genes, complete the fine-tuning of the BERT model, obtain the correlation changes between ginseng transcription factors and ginseng genes, and obtain the ginseng gene embedding data;

[0012] Step 4: Use the ginseng gene embedding data obtained in step 3 to calculate the correlation between ginseng transcription factors and ginseng genes, and obtain the transcriptional regulatory relationship network of ginseng transcription factors and ginseng genes.

[0013] Further, step 1 comprises:

[0014] Ginseng transcriptome sequencing samples were obtained, and the sratools toolkit was used to download, split, quality control and filter the data. The ginseng genome file and gene structure annotation file were obtained, and the reference genome was prepared using the software gffread, and the reference genome was compared using the hisat2 software.

[0015] The entire species lineage of the ginseng genome was compared with species with transcriptional regulatory relationships in the PlantTFDB database from near to far to search for closely related species. The BLAST sequence alignment was then used to align the ginseng genome with the genomes of closely related species to obtain a list of ginseng transcription factors.

[0016] Furthermore, the use of hisat2 software to compare the reference genome includes:

[0017] Build an index for the reference genome and perform alignment with the -dta parameter;

[0018] Compress and sort the comparison results and build an index;

[0019] The transcript assembly software stringtie was used to calculate the FPKM expression matrix of transcripts.

[0020] Furthermore, in step 2, the model plants include but are not limited to rice and Arabidopsis thaliana.

[0021] Furthermore, step 2 includes:

[0022] Download the gene expression data samples of the model plants, input the gene expression data samples into the BERT model, perform masking strategy on the model plant genes, predict the remaining genes, and complete the self-supervised pre-training of the BERT model.

[0023] Furthermore, step 4 includes: using the ginseng gene embedding data obtained in step 3 to calculate the correlation between TF genes and ginseng genes, selecting 0.9 as the threshold, retaining the relationship with correlation coefficient greater than 0.9 and less than 1, and obtaining the transcriptional regulation relationship network of ginseng TF and ginseng genes.

[0024] Furthermore, the correlation calculation uses the Pearson correlation coefficient:

[0025] r = Cov(X,Y) / σ x σ Y

[0026] Where r is the Pearson correlation coefficient. Cov(X,Y) is the covariance of variables X and Y. σ x is the standard deviation of variable X. Y is the standard deviation of variable Y.

[0027] In a second aspect, the present invention provides a prediction system for the transcriptional regulation mechanism of the downstream pathway of ginsenoside synthesis, the system comprising:

[0028] A data processing module, used to convert ginseng transcriptome sequencing data SRA into ginseng gene expression data and obtain a ginseng transcription factor list;

[0029] Model pre-training module, used to pre-train gene expression data of model plants using BERT model;

[0030] The model fine-tuning module is used to fine-tune the BERT model using ginseng gene expression data, including:

[0031] The ginseng gene expression data was input into the pre-trained BERT model, a masking strategy was performed on non-ginseng TF genes, the remaining genes were predicted, and the BERT model was fine-tuned to obtain the correlation changes between ginseng transcription factors and ginseng genes, and the ginseng gene embedding data was obtained;

[0032] The transcriptional regulation relationship network calculation module is used to calculate the correlation between ginseng transcription factors and ginseng genes using ginseng gene embedding data, and obtain the transcriptional regulation relationship network of ginseng transcription factors and ginseng genes.

[0033] In a third aspect, the present invention provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, and when the processor runs the computer program stored in the memory, the steps of the method for predicting the transcriptional regulation mechanism of the downstream pathway of ginsenoside synthesis as described above are executed.

[0034] In a fourth aspect, the present invention provides a computer-readable storage medium storing a plurality of computer instructions, wherein the plurality of computer instructions are used to enable a computer to execute a method for predicting a transcriptional regulation mechanism of a downstream pathway of ginsenoside synthesis as described above.

[0035] Beneficial effects of the present invention:

[0036] The present invention uses the BERT model in the field of artificial intelligence for transfer learning. For example, a large number of plant gene expression data such as Arabidopsis thaliana, rice and other model plants are used for pre-training to obtain the expression patterns of plant genes. The pre-training model is then fine-tuned using the gene expression data of ginseng to obtain the special gene expression patterns of medicinal plants. This not only takes into account the generality of large-scale plants, but also pays attention to the particularity of medicinal plants, improves the credibility and reliability of the network, and is an innovation in the annotation method of ginseng regulatory networks.

[0037] Although the present invention can refer to closely related species, they are all medicinal plants with limited data and incomplete annotations, which provide insufficient reference evidence for ginseng. Therefore, the present invention uses model plants with complete annotations and abundant data for pre-training, which can obtain a lot of plant regulation information with reference value.

[0038] The present invention predicts the gene regulatory network downstream of the saponin biosynthesis biological pathway under the growth and development conditions of ginseng through the proposed TFGPT (Transcription Factor Generative Pre-trained Transformer) algorithm model, thereby achieving the goal of analyzing the synthesis mechanism of ginsenosides. This can promote the comprehensive genome annotation of medicinal plants and is of great significance to promoting the internationalization and modernization of Chinese medicine. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] In order to more clearly illustrate the technical solution of the present application, the drawings required for use in the embodiments are briefly introduced below. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0040] Figure 1 This is a schematic diagram showing that the biosynthesis of ginsenosides is divided into two components, upstream and downstream;

[0041] Figure 2 Schematic diagram of the multi-head attention mechanism in the BERT model;

[0042] Figure 3 This is a schematic diagram of the TFGPT algorithm flow;

[0043] Figure 4 Partial schematic diagram of the transcriptional regulation mechanism of ginseng. DETAILED DESCRIPTION

[0044] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are intended to be used to explain the present invention, and should not be construed as limiting the present invention.

[0045] Embodiment 1: A method for predicting the transcriptional regulation mechanism of the downstream pathway of ginsenoside synthesis, the method comprising:

[0046] Step 1: Convert ginseng transcriptome sequencing data SRA into ginseng gene expression data and obtain a list of ginseng transcription factors;

[0047] Step 2: Use the BERT model to pre-train the gene expression data of the model plant;

[0048] Step 3: Fine-tune the BERT model using ginseng gene expression data, including:

[0049] Input the ginseng gene expression data in step 1 into the pre-trained BERT model, perform a masking strategy on non-ginseng TF genes, predict the remaining genes, complete the fine-tuning of the BERT model, obtain the correlation changes between ginseng TF and ginseng genes, and obtain the ginseng gene embedding data;

[0050] The masking strategy is used to mask the expression levels of 20% of non-TF genes, and the remaining TFs and genes are used to predict them. The purpose of this masking is to capture the correlation between TFs and genes as much as possible. The prediction process is also the process of fine-tuning the BERT model.

[0051] Step 4: Use the ginseng gene embedding data obtained in step 3 to calculate the correlation between TF genes and ginseng genes, and obtain the transcriptional regulatory relationship network of ginseng TF and ginseng genes.

[0052] Transcription Factor Generative Pre-trained Transformer (TFGPT) is a large biological network graph model developed by this project for modeling the structure of gene regulatory network graphs in bioinformatics. The model transfers the characteristics of plant transcriptional regulation to the transcriptional regulation of ginseng. Based on the general to specific theoretical framework, it first conducts large-scale pre-training on model plants, and then fine-tunes with a small sample of ginseng to grasp the regulatory characteristics of ginseng transcription factors (TFs) and their associated target gene nodes, in order to achieve a more accurate reconstruction of the transcriptional regulation graph.

[0053] This implementation method uses a large number of model plant samples for pre-training, learns the TF and intergenic knowledge of large-scale model plant data (such as rice and Arabidopsis), and then uses the trained model plant pre-trained large model for non-model plants, such as ginseng, for fine-tuning in small sample data to obtain a special large model suitable for ginseng. This model retains the regulatory characteristics of model organisms and has the particularity of non-model organisms. This model can solve the current problems of few Chinese herbal medicine samples and incomplete data, and also solves the pain point that the reference data of medicinal plants is small and cannot be fully annotated.

[0054] This implementation method realizes the analysis of the synthesis mechanism of ginsenosides, can promote the comprehensive genome annotation of medicinal plants, will help reduce the difficulty of ginsenoside synthesis, increase the yield of saponins in in vitro tissue culture, and meet the growing demand for medicinal uses; as ginsenosides are invested in large quantities in drug processing, the problems of scarcity and high cost of anti-tumor drugs can be solved, and it will play a very positive role in reducing the side effects of radiotherapy and chemotherapy for tumor patients, improving the prognosis, and developing cancer targeted drugs, which is of great significance to promoting the internationalization and modernization of Chinese medicine.

[0055] Embodiment 2: This embodiment further limits the prediction method of the transcriptional regulation mechanism of the downstream pathway of ginsenoside synthesis as described above. In this embodiment, step 1 is further limited, specifically including:

[0056] Step 1 includes:

[0057] Ginseng transcriptome sequencing samples were obtained, and the sratools toolkit was used to download, split, quality control and filter the data. The ginseng genome file and gene structure annotation file were obtained, and the reference genome was prepared using the software gffread, and the reference genome was compared using the hisat2 software.

[0058] The entire species lineage of the ginseng genome was compared with species with similar transcriptional regulatory relationships from close to distant species by BLAST, and the ginseng genome was sequenced with the genomes of closely related species to obtain a list of ginseng transcription factors.

[0059] Since the annotation of the ginseng genome is not yet complete, the ginseng transcription factors have not been fully analyzed. The present embodiment aims to obtain as much ginseng transcriptional regulatory information as possible. Since closely related species have genetic similarities in sequence, the BLAST sequence alignment method is used to compare closely related species, and a list of ginseng transcription factors is obtained from closely related species with known transcription factors.

[0060] Embodiment 3: This embodiment further limits the prediction method of the transcriptional regulation mechanism of the downstream pathway of ginsenoside synthesis as described above. In this embodiment, the use of hisat2 software to compare the reference genome is further limited, specifically including:

[0061] The use of hisat2 software to align the reference genome includes:

[0062] Build an index for the reference genome and perform alignment with the -dta parameter;

[0063] Compress and sort the comparison results and build an index;

[0064] The transcript assembly software stringtie was used to calculate the FPKM expression matrix of transcripts.

[0065] In this embodiment, because there is too little ginseng gene expression data, unlike model organisms, which have a large number of gene expression samples, the gene expression data of ginseng can only be calculated and generated from sequencing data. The purpose of this design is to obtain as much gene expression data as possible, which is conducive to subsequent bioinformatics calculations or the application of deep models.

[0066] Embodiment 4: This embodiment further limits the prediction method of the transcriptional regulation mechanism of the downstream pathway of ginsenoside synthesis as described above. In this embodiment, the model plant in step 2 is further limited, specifically including:

[0067] In step 2, the model plants include but are not limited to rice and Arabidopsis thaliana.

[0068] The two model plants preferentially used in this implementation method are agricultural crops, which are hot topics in bioinformatics research. Many laboratories around the world upload their data to the Internet, so there are enough public data sets for pre-training large models.

[0069] Embodiment 5: This embodiment further limits the prediction method of the transcriptional regulation mechanism of the downstream pathway of ginsenoside synthesis as described above. In this embodiment, step 2 is further limited, specifically including:

[0070] Step 2 includes:

[0071] Download the gene expression data samples of the model plants, input the gene expression data samples into the BERT model, and then use the mask strategy gene prediction to complete the self-supervised pre-training.

[0072] A masking strategy is used to cover up 20% of the gene expression levels of all model plants, and the remaining 80% of genes are used for prediction. The purpose of this masking is to capture the correlation between genes in model plants as much as possible. The prediction process is also the pre-training process of the BERT model, which can obtain the model plant gene embedding data between genes in massive plant samples.

[0073] This embodiment uses gene expression data of model plants to pre-train the BERT model to obtain the expression pattern of plant genes. This training method obtains a lot of plant regulation information with reference value.

[0074] Embodiment 6: This embodiment further limits the prediction method of the transcriptional regulation mechanism of the downstream pathway of ginsenoside synthesis as described above. In this embodiment, step 4 is further limited, specifically including:

[0075] Step 4 includes: using the ginseng gene embedding data obtained in step 3 to calculate the correlation between TF genes and ginseng genes, selecting 0.9 as a threshold, retaining relationships with correlation coefficients greater than 0.9 and less than 1, and obtaining a transcriptional regulatory relationship network between ginseng TF and ginseng genes.

[0076] This embodiment provides a specific scheme for calculating the correlation between TF genes and ginseng genes, which can effectively obtain the correlation between TF genes and ginseng genes, and further realize the effective acquisition of the transcriptional regulatory relationship network between ginseng TF and ginseng genes.

[0077] Embodiment 7: This embodiment further limits the prediction method of the transcriptional regulation mechanism of the downstream pathway of ginsenoside synthesis as described above. In this embodiment, the correlation calculation is further limited, specifically including:

[0078] The correlation calculation uses the Pearson correlation coefficient:

[0079] r = Cov(X,Y) / σ x σ Y

[0080] Where r is the Pearson correlation coefficient. Cov(X,Y) is the covariance of variables X and Y. σ x is the standard deviation of variable X. Y is the standard deviation of variable Y.

[0081] Embodiment 8: This embodiment is based on Example 1 of the method for predicting the transcriptional regulation mechanism of the downstream pathway of ginsenoside synthesis as described above, and specifically includes:

[0082] Step 1: Convert ginseng transcriptome sequencing data SRA to gene expression data FPKM and obtain ginseng TF list:

[0083] The project first downloaded ginseng transcriptome sequencing samples from NCBI's SRA database, used the sratools toolkit to download, split, quality control and filter the data, downloaded ginseng genome files and gene structure annotation files from the Korean ginseng genome database genome db, used the software gffread to prepare the reference genome, and used hisat2 software to align the reference genome in three steps: the first step was to build an index for the reference genome and perform an alignment with the parameter -dta, the second step was to compress and sort the alignment results and build an index, and the third step was to use the transcript assembly software stringtie to calculate the FPKM expression matrix of the transcripts. The ginseng genome in the NCBI Taxonomy Browser database was compared with the species with transcriptional regulatory relationships in the PlantTFDB database from near to far to find closely related species, and then the BLAST sequence alignment was used to align the ginseng genome with the genomes of closely related species, and then a list of ginseng transcription factors (TF) was obtained.

[0084] Step 2: Use the BERT model to pre-train rice and Arabidopsis gene expression data:

[0085] We downloaded gene expression data samples of about 100,000 model plants that are currently being studied more (such as rice and Arabidopsis) from databases such as GEO, performed gene embedding operations on them, input the data into the BERT model, and then used a masking strategy to cover 20% of the genes and predict the remaining genes for self-supervised pre-training, so that the model has gene expression knowledge of large-scale plant data in public databases.

[0086] The BERT model position embedding Position Embedding related formula is as follows:

[0087] PE(pos,2i)=sin(pos / 10000 2i / dmodel )

[0088] PE(pos,2i+1)= cos(pos / 10000 2i / dmodel )

[0089] Among them, pos represents the position index, i represents the dimension index, d model is the dimension of the model.

[0090] The formula of the multi-head attention mechanism is as follows:

[0091] MultiHead(Q,K,V)= Concat(head1,...,head h )W O

[0092] The calculation for each head is as follows:

[0093] head i =Attention(QW i Q ,KW i K ,VW i V ), where Q, K, V: represent the query, key, and value matrices respectively. QW i Q ,KW i K ,VW i V : is the weight matrix of different heads, used for linear transformation, corresponding to query, key and value respectively. Attention: self-attention mechanism, the attention weight is calculated by formula and the value is weighted summed. Concat: concatenates the outputs of all heads together. W O : is a linear transformation matrix used to map the concatenated output to the final dimension. Figure 2 .

[0094] The layer-by-layer normalization formula is as follows:

[0095] Z = LayerNorm(X+Sublayer(X))

[0096] X: Input matrix, representing the input of the current layer. Z: Output matrix, the output after layer normalization. Sublayer(X): The output of the feed-forward network of the current layer.

[0097] The relevant formula of the feed-forward network is as follows:

[0098] FFN(x)=max(0,xW1+b1)W2+b2

[0099] W1,W2: weight matrices. b1,b2: bias vectors.

[0100] Step 3: Fine-tune the BERT model using ginseng data:

[0101] The ginseng gene expression data obtained in step 1 were subjected to gene embedding operation and input into the pre-trained BERT model. Then, a masking strategy was used to mask 20% of non-ginseng TF genes to predict the remaining genes for Finetune fine-tuning, and the correlation changes between ginseng TF and ginseng genes were mined to obtain ginseng gene embedding data.

[0102] The MSE formula for finetune loss function is as follows:

[0103]

[0104] Among them, y true is the true label, y pred is the model prediction value, and n is the number of samples.

[0105] Step 4: Use ginseng gene embedding data to construct the graph:

[0106] The ginseng gene embedding data obtained in step 3 was used to calculate the correlation between TF genes and ginseng genes, and the threshold, i.e., cutoff, was selected as 0.9, and the relationship with correlation coefficient greater than 0.9 and less than 1 was retained, and finally the transcriptional regulatory relationship network of ginseng TF and ginseng genes was obtained. Figure 3 .

[0107] The Pearson correlation coefficient formula is as follows:

[0108] r = Cov(X,Y) / σ x σ Y

[0109] r: Pearson correlation coefficient. Cov(X,Y): covariance of variables X and Y. σ x : Standard deviation of variable X. σ Y : standard deviation of variable Y.

[0110] Embodiment 9: This embodiment is based on Example 2 of the method for predicting the transcriptional regulation mechanism of the downstream pathway of ginsenoside synthesis as described above, and specifically includes:

[0111] The BERT model is used for pre-training, which adopts the standard Transformer Encoder structure, including 12 layers of transformer encoders, and each layer has 12 self-attention heads. In terms of data set division: the data set is divided into 80% training set and 20% validation set to maintain the independence between gene pairs. Cross-validation is used to ensure the generalization ability of the model on different training sets.

[0112] Pre-training uses the model plants of the GEO database, namely Arabidopsis and rice samples. Download Arabidopsis transcriptome sequencing samples from the GEO database of NCBI, the platform number is GPL11221, containing 2134 samples, the platform number is GPL13222, containing 3136 samples, the platform number is GPL17639, containing 3308 samples, the platform number is GPL21785, containing 3595 samples, the platform number is GPL17639, containing 3308 samples, containing 446 samples, the platform number is GPL24270, containing 453 samples, and the total number of GEO Arabidopsis samples is 13072. From the GEO database of NCBI, the rice transcriptome sequencing samples were obtained, with the platform number GPL6864, including 2547 samples, the platform number GPL11260, including 2064 samples, the platform number GPL20755, including 1568 samples, the platform number GPL24468, including 1248 samples, the platform number GPL27660, including 576 samples, and the platform number GPL13834, including 487 samples. A total of 21562 samples of Arabidopsis and rice samples were used, and the GEO rice samples totaled 8490. This embodiment also collected Arabidopsis and rice samples from other public databases, totaling about 100000.

[0113] Ginseng transcriptome sequencing samples were downloaded from the SRA database of NCBI, with platform numbers SRP066368 and SRP049125, containing 60 and 18 samples respectively, for a total of 78 samples.

[0114] In terms of parameter selection, the model learning rate uses a warm-up strategy in the early stage, the learning rate is selected as 5e-5, the accuracy is above 95%, and the batch size is selected as 16 batches for the experiment. To ensure that the input data does not lose information, the maximum sequence length is selected as 512. Finally, we obtained the ginseng gene embedding data, calculated the correlation coefficient between TF and gene, and set the threshold to 0.9, that is, genes with a correlation coefficient between 0.9 and 1. Some ginseng TF-gene regulation diagrams are shown in Figure 4 Among them, Pg is the abbreviation of Panax Ginseng, the Latin name of ginseng, S is the abbreviation of Scaffold, and the gene marked with TF is the ginseng TF gene obtained after proximal alignment. The TF gene can regulate non-TF genes, and can also regulate TF genes. The arrow points to the regulated gene.

[0115] The most valuable pharmacological component of ginseng is ginsenoside, and the transcriptional regulation mechanism of its synthesis downstream pathway has not been fully analyzed. This paper uses advanced computational methods to establish a ginseng-specific model based on transcriptome data to accurately predict the gene regulation relationship map at different developmental stages, thereby revealing the key transcriptional regulation mechanism downstream of ginsenoside synthesis.

Claims

1. A method for predicting the transcriptional regulation mechanism of the downstream pathway of ginsenoside synthesis, characterized in that: The method comprises: Step 1: Convert ginseng transcriptome sequencing data SRA into ginseng gene expression data and obtain a list of ginseng transcription factors; Step 2: Use the BERT model to pre-train the gene expression data of the model plant; Step 3: Fine-tune the BERT model using ginseng gene expression data, including: Input the ginseng gene expression data in step 1 into the pre-trained BERT model, perform a masking strategy on non-ginseng transcription factor genes, predict the remaining genes, complete the fine-tuning of the BERT model, obtain the correlation changes between ginseng transcription factors and ginseng genes, and obtain the ginseng gene embedding data; Step 4: Use the ginseng gene embedding data obtained in step 3 to calculate the correlation between ginseng transcription factors and ginseng genes, and obtain the transcriptional regulatory relationship network between ginseng transcription factors and ginseng genes; Step 2 includes: Download gene expression data samples of model plants, input the gene expression data samples into the BERT model, perform masking strategy on the genes of model plants, predict the remaining genes, and complete the self-supervised pre-training of the BERT model; The MSE loss function formula for fine-tuning in step 3 is as follows: in, is the true value of the i-th sample, is the model prediction value of the ith sample, and n is the number of samples.

2. The method for predicting the transcriptional regulation mechanism of the downstream pathway of ginsenoside synthesis according to claim 1, characterized in that: Step 1 includes: Acquire ginseng transcriptome sequencing samples, use the sratools toolkit to download, split, quality control and filter the data, obtain ginseng genome files and gene structure annotation files, use gffread software to prepare reference genomes, and use hisat2 software to align reference genomes; The entire species lineage of the ginseng genome was compared with species with transcriptional regulatory relationships in the PlantTFDB database from near to far to search for closely related species. The BLAST sequence alignment was then used to align the ginseng genome with the genomes of closely related species to obtain a list of ginseng transcription factors.

3. The method for predicting the transcriptional regulation mechanism of the downstream pathway of ginsenoside synthesis according to claim 2, characterized in that: The use of hisat2 software to align the reference genome includes: Build an index for the reference genome and perform alignment with the -dta parameter; Compress and sort the comparison results and build an index; The transcript assembly software stringtie was used to calculate the FPKM expression matrix of transcripts.

4. The method for predicting the transcriptional regulation mechanism of the downstream pathway of ginsenoside synthesis according to claim 1, characterized in that: In step 2, the model plants include rice and Arabidopsis thaliana.

5. The method for predicting the transcriptional regulation mechanism of the downstream pathway of ginsenoside synthesis according to claim 1, characterized in that: Step 4 includes: using the ginseng gene embedding data obtained in step 3 to calculate the correlation between transcription factor genes and ginseng genes, selecting 0.9 as a threshold, retaining relationships with correlation coefficients greater than 0.9 and less than 1, and obtaining a transcriptional regulatory relationship network between ginseng transcription factors and ginseng genes.

6. The method for predicting the transcriptional regulation mechanism of the downstream pathway of ginsenoside synthesis according to claim 5, characterized in that: The correlation calculation uses the Pearson correlation coefficient: r = Cov(X,Y) / σ x s Y Among them, r is the Pearson correlation coefficient, Cov(X,Y) is the covariance of variables X and Y, σ x is the standard deviation of variable X, σ Y is the standard deviation of variable Y.

7. A prediction system for transcriptional regulation mechanisms of downstream pathways of ginsenoside synthesis, characterized in that: The system comprises: A data processing module, used to convert ginseng transcriptome sequencing data SRA into ginseng gene expression data and obtain a ginseng transcription factor list; Model pre-training module, used to pre-train gene expression data of model plants using BERT model; The model fine-tuning module is used to fine-tune the BERT model using ginseng gene expression data, including: The ginseng gene expression data was input into the pre-trained BERT model, and the non-ginseng transcription factor genes were masked to predict the remaining genes, and the BERT model was fine-tuned to obtain the correlation changes between ginseng transcription factors and ginseng genes, and the ginseng gene embedding data was obtained; A transcriptional regulatory relationship network calculation module is used to calculate the correlation between ginseng transcription factors and ginseng genes using ginseng gene embedding data to obtain a transcriptional regulatory relationship network between ginseng transcription factors and ginseng genes; The model pre-training module includes: Download gene expression data samples of model plants, input the gene expression data samples into the BERT model, perform masking strategy on the genes of model plants, predict the remaining genes, and complete the self-supervised pre-training of the BERT model; The MSE related formula of the fine-tuning loss function in the model fine-tuning module is as follows: in, is the true value of the i-th sample, is the model prediction value of the ith sample, and n is the number of samples.

8. A computer device comprising a memory and a processor, wherein a computer program is stored in the memory, wherein: When the processor runs the computer program stored in the memory, the steps of the method according to any one of claims 1 to 6 are performed.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a plurality of computer instructions, and the plurality of computer instructions are used to enable a computer to execute the method according to any one of claims 1 to 6.

Images