Target gene prediction method for traditional Chinese medicines tRNA, tRFs and t-halves based on artificial intelligence

Abstract

A target gene prediction method of traditional Chinese medicine tRNA, tRFs and t-halves based on artificial intelligence comprises the following steps: extracting and separating the tRNA of the traditional Chinese medicine, and carrying out next-generation sequencing to obtain a tRNA sequence; according to a sequencing result, tRFs and t-halves sequences derived from the traditional Chinese medicine are designed, the modification variety and site of tRNA are analyzed through LC-MS / MS, and a database of tRNA fragments of the traditional Chinese medicine is constructed; based on a sequence of a human gene mRNA 3 'UTR, a potential target gene of a traditional Chinese medicine tRNA fragment is preliminarily screened by using a base complementary pairing tool; verifying the inhibition efficiency of the traditional Chinese medicines tRFs and t-halves on the target gene through experiments, and constructing a sample data set in combination with the screening result in the previous step; and constructing a deep learning convolutional neural network (CNN) model, carrying out training and optimization by using the sample data set, and outputting a prediction result of the target gene and the inhibition efficiency thereof.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of computer-aided drug target screening, in particular to a method for predicting target genes of traditional Chinese medicine tRNA, tRFs and t-halves based on artificial intelligence, which is suitable for screening target genes related to specific diseases according to traditional Chinese medicine tRNA fragments. BACKGROUND

[0002] As an important part of traditional Chinese medicine, the analysis of active ingredients and the study of mechanisms of traditional Chinese medicine have always been the focus of modern drug development. In recent years, with the development of molecular biology technology, researchers have found that nucleic acid components (such as tRNA and its fragments tRFs, t-halves) in traditional Chinese medicine may play an important therapeutic role by regulating gene expression. These non-coding RNA fragments can mimic the mechanism of siRNA and miRNA, bind to target mRNA and regulate its expression level, thereby producing various pharmacological activities such as anti-inflammatory and anti-tumor activities.

[0003] In the field of traditional Chinese medicine target gene screening, traditional methods mainly rely on molecular biology experimental techniques, including gene knockout, protein interaction analysis, Western blotting and qPCR. Although these methods are reliable, they have obvious limitations: first, the experimental period is long and the cost is high, which makes it difficult to meet the needs of large-scale screening; second, the experimental process is complex and requires high technical requirements for the operator; most importantly, these methods cannot quickly predict potential target genes, and researchers need to pre-set candidate target points for verification, which greatly limits the efficiency of discovering new target points.

[0004] With the development of bioinformatics technology, some computational prediction tools based on the principle of base complementary pairing have emerged. These tools can quickly predict potential target genes by analyzing the complementarity of RNA sequences, significantly improving the screening efficiency. However, existing bioinformatics prediction methods still have many shortcomings:

[0005] 1. Strong data dependence: existing algorithms are mainly trained based on limited known databases, which leads to biased prediction results and makes it difficult to adapt to tRNA fragments of different sources or special modifications;

[0006] 2. Limited scope of application: most tools are developed for specific species or cell types, lack of universality, and are difficult to meet the prediction needs of complex components of traditional Chinese medicine;

[0007] 3. Insufficient prediction accuracy: traditional algorithms lack self-learning ability and cannot optimize the prediction model through iterative training, making it difficult to continuously improve the accuracy of prediction results;

[0008] 4. Single function: existing tools mainly focus on target gene prediction, lack of regulatory efficiency evaluation function, and cannot provide more comprehensive reference for subsequent experiments.

[0009] These technical bottlenecks seriously restrict the in-depth study of nucleic acid components of traditional Chinese medicine, and hinder the development of innovative drugs based on tRNA fragments of traditional Chinese medicine. Therefore, developing an efficient, accurate and self-learning traditional Chinese medicine tRNA fragment target gene prediction method has important theoretical and application value for promoting the modernization of traditional Chinese medicine research. SUMMARY

[0010] The technical problem solved by the present application is to overcome the shortcomings of current small RNA sequence target gene screening, and a traditional Chinese medicine tRNA, tRFs, t-halves target gene prediction method based on artificial intelligence is disclosed, which can help quickly find the target gene of traditional Chinese medicine tRNA fragments, and provide inhibition efficiency analysis, thereby helping drug development for different diseases.

[0011] Technical scheme: A traditional Chinese medicine tRNA, tRFs, t-halves target gene prediction method based on artificial intelligence, comprising the following steps: (1) extracting and separating tRNA of traditional Chinese medicine, performing second generation sequencing to obtain tRNA sequence; (2) designing tRFs and t-halves sequences derived from traditional Chinese medicine according to the sequencing results, and analyzing the modification types and sites of tRNA through LC-MS / MS to construct a database of traditional Chinese medicine tRNA fragments; (3) based on the sequence of human gene mRNA 3'UTR, using base complementary pairing tool to preliminarily screen potential target genes of traditional Chinese medicine tRNA fragments; (4) verifying the inhibition efficiency of traditional Chinese medicine tRFs and t-halves on target genes through experiments, and constructing a sample data set combining the screening results of step (3); (5) constructing a deep learning convolutional neural network (CNN) model, training and optimizing using the sample data set, and outputting the prediction results of target genes and their inhibition efficiency.

[0012] The extraction method of tRNA in step (1) comprises: (1.1) using a plant RNA extraction kit to extract total RNA of traditional Chinese medicine, and separating small RNA with a length of less than 200 nt through a small RNA enrichment kit; (1.2) separating tRNA through 6% urea-PAGE gel electrophoresis, cutting the target region and recovering through electrophoresis; (1.3) using TruSeq Small RNA Sample Preparation Kit to construct a cDNA library, and performing second generation sequencing through an Illumina platform.

[0013] The design of tRFs and t-halves sequences in step (2) includes: (2.1) defining 5'-t-half, 3'-t-half, 5'-tRF and 3'-tRF sequences based on the break position of the tRNA anticodon loop; (2.2) analyzing the modification sites of the purified tRNA by digesting it with RNA restriction endonucleases or exonucleases and generating fragments; (2.3) optimizing the modification type based on the mass spectrometry results, including pseudouracil, dihydrouracil, methylation modification and 2'-oxymethylation modification.

[0014] The base pairing tool mentioned in step (3) is the miRanda program, with the screening threshold set to Score≥140 and Energy≤-1.

[0015] The experimental verification in step (4) includes: (4.1) selecting cell models related to the treatment of diseases with traditional Chinese medicine and setting the concentration gradient of tRNA fragments of traditional Chinese medicine (0-400nM); (4.2) detecting the mRNA content of target genes by qPCR, calculating the inhibition efficiency = (expression level of control group - expression level of experimental group) / expression level of control group × 100%, and including target genes with inhibition efficiency ≥ 70% in the sample dataset.

[0016] The construction of the CNN model in step (5) includes: (5.1) converting the tRNA sequence into a binary vector using one-hot encoding, and numerically encoding the gene name using LabelEncoder; (5.2) the model structure includes an input layer, a convolutional layer, a pooling layer, and an output layer, with the output layer simultaneously predicting the target gene name and the inhibition efficiency; (5.3) the optimizer uses Adam with an initial learning rate of 0.001, and the loss function is a combination of sparse_categorical_crossentropy and mean squared error (MSE), with a loss weight range of 0.3 and 0.7.

[0017] The training parameters of the above CNN model include: (7.1) The ratio of training set to validation set is 8:2; (7.2) The number of epochs is set to 50 and the batch size is 32; (7.3) The model performance is evaluated by hierarchical K-fold cross-validation (K=5), and the evaluation indicators include the MAE value of target gene prediction accuracy and inhibition efficiency.

[0018] Preferably, the sequence of the above-mentioned human gene mRNA3'UTR is obtained from the GeneCode database.

[0019] Preferably, the small RNA enrichment kit mentioned above is Ambion mirVana. TM miRNA Isolation Kit.

[0020] Preferably, the above concentration gradient includes 400, 200, 100, 50, 25, 12.5, 6.25, 3.125, 1.5625 and 0 nM.

[0021] Beneficial effects: 1. Compared with traditional experiments, this method can efficiently screen potential target genes of TCM tRNA, especially suitable for large-scale screening; 2. Based on the previous base complementary pairing principle, this method adds experimental verification for building predictive models, and also incorporates the mRNA inhibition efficiency, thus improving the accuracy of gene target prediction; 3. By building predictive models through artificial intelligence, accuracy and adaptability can be increased through continuous training and optimization. These features effectively enable target screening of TCM tRNA fragments, especially disease-related targets, providing a new auxiliary method for the development of TCM nucleic acid drugs. Attached Figure Description

[0022] Figure 1 A flowchart for building a CNN model.

[0023] Figure 2 This is a schematic diagram of the structure of a CNN model.

[0024] Figure 3 This is a commonly used combination modification strategy in this invention. Detailed Implementation

[0025] To make the objectives, technical solutions, and beneficial effects of this invention clearer, the technical solutions of this invention will be further described below in conjunction with the accompanying drawings.

[0026] Example 1

[0027] This invention can be achieved through the following technical route ( Figure 1 To achieve:

[0028] 1. An artificial intelligence-based method for predicting target genes of tRNA, tRFs, and t-halves in traditional Chinese medicine, comprising the following steps:

[0029] 1.1 Step 1: Extract and isolate tRNA from traditional Chinese medicine for next-generation sequencing;

[0030] Step 2: Based on the next-generation sequencing results, design tRFs and t-halves sequences from various traditional Chinese medicine sources. Purify tRNA, use LC-MS to determine the types and sites of modifications on natural tRNA, design modified tRFs and t-halves sequences, and construct a database of traditional Chinese medicine tRNA fragments;

[0031] The above modification types include:

[0032] (1) Phosphodiester backbone modification: One non-bridging oxygen in the phosphate backbone of the nucleotide is replaced with sulfur or methyl, including phosphate thioate (PS) modification and methylphosphonate modification. This modification does not significantly affect the activity of nucleic acid drugs, but it can resist degradation by nucleases.

[0033] (2) Ribose modification: Modification of the hydroxyl group at the 2-position of the ribose structure, including 2'-O-methylation (2'-OMe), 2'-fluorination (2'-F), 2'-O-(2-methoxyethyl)ation, and 2'-deoxidation. Additionally, there is the locked nucleic acid (LNA) method, which introduces an additional bridging ring onto the ribose ring, placing the nucleic acid in a locked conformation. Furthermore, there is glycol nucleic acid (GNA). These modifications can further enhance resistance to nucleases, thereby improving serum stability, and can enhance its binding affinity to complementary nucleotide chains, thus reducing off-target effects.

[0034] (3) Base modification: including pseudonucleoside (Ψ), 2-thiouridine (s2U) and 5-methylcytidine (5-Methylcytidine), which can enhance the stability of nucleic acids and reduce immune response, and can also increase the melting temperature of nucleic acids to reduce off-target effects.

[0035] (4) End modification: including adding phosphate groups, (E)-vinylphosphonate (VP), lipid molecules, inverted thymidine, etc. to the 3' and / or 5' ends of the sequence, which can prevent the nucleotide sequence from being degraded by exonuclease, thereby improving nucleic acid stability and promoting Ago2 loading.

[0036] The above modification sites need to be adjusted according to the nucleic acid sequence and modification type. The following shows one of the commonly used combination modification strategies: the optimal configuration of modification can be achieved by modifying 2'-OMe, 2'-F, PS and 5'-VP, which can effectively improve nucleic acid stability, form RISC and target recognition. Specifically: (1) Guide Strand: 5' end 5'-VP modification; all or part of the nucleosides 2'-OMe modification at positions 1, 3, 9, 11, 13, 15, 17, 18, 19, 21; all or part of the nucleosides 2'-F modification at positions 2, 4, 5, 6, 8, 10, 12, 14, 16, 20; all or part of the phosphodiester backbone PS modification at positions 1, 2, 14, 15, 16, 17, 18, 19, 20. (2) Passenger Strand: all or part of the 2'-OMe nucleosides at positions 1, 2, 3, 5, 7, 9, 11, 12, 13, 15, and 16 are modified with 2'-OMe nucleosides; all or part of the 2'-F nucleosides at positions 4, 6, 8, 10, and 14 are modified with 2'-F nucleosides; and all or part of the PS phosphodiester skeleton at positions 1, 2, 14, and 15 are modified with PS.

[0037] 1.3 Step 3: Preliminary establishment of a screening library of Chinese herbal medicine tRNA fragments and human gene mRNA 3'UTR.

[0038] 1.4 Step 4: Conduct gene targeting experiments to verify the effects of traditional Chinese medicine tRFs and t-halves, and summarize and optimize the screening results with the initial screening results obtained in Step 3, and organize them into a sample dataset for subsequent prediction model construction.

[0039] Step 5: Construct a deep learning convolutional neural network. Randomly divide the sample dataset into training and test sets for model training and learning, and then evaluate the model's predictive performance.

[0040] Step one above includes the following steps:

[0041] 2.1 Based on literature reports, search for the top 5 Chinese medicinal herbs with medicinal value that have been reported the most.

[0042] 2.2 A commercially available plant RNA extraction kit was used to extract RNA from the traditional Chinese medicine, and the RNA was extracted using the commercially available Ambion mirVana kit. TM The miRNA Isolation Kit was used to isolate and enrich small RNA species (<200 mere). The content and quality of total RNA and small RNA samples were analyzed and evaluated using a Thermo Fisher Nanodrop instrument and an Agilent Bioanalyzer instrument.

[0043] 2.3 Small RNAs in traditional Chinese medicine were separated using 6% uremic-page gel. Then, the region containing tRNA in the gel was cut off and placed into a 3kDa cutoff dialysis tube, and the tRNA was recovered by electrophoresis.

[0044] 2.4 The recovered tRNA was used to construct a cDNA library using the TruSeq Small RNA Sample Preparation Kit, and sequenced using the Illumina next-generation sequencing platform. Quality control was then performed, and the tRNA sequence was analyzed and identified to obtain the final tRNA sequence of the traditional Chinese medicine.

[0045] Step two above includes the following steps:

[0046] 3.1 Design various tRFs and t-halves sequences from Chinese herbal medicine sources, including (1) 5'-t-half: the sequence from the break point to the 5' end of the tRNA after any tRNA anticodon loop is broken at any position; (2) 3'-t-half: the sequence from the break point to the 3' end of the tRNA after any tRNA anticodon loop is broken at any position; (3) 5'-tRF: the sequence formed by any tRNA from the 5' end to a break point before the anticodon loop; (4) 3'-tRF: the sequence formed by any tRNA from the 3' end to a break point before the anticodon loop.

[0047] 3.2 RNA was extracted from traditional Chinese medicine and analyzed using Ambion mirVana. TM The miRNA Isolation Kit is used to isolate and enrich small RNA species (<200 mere nucleotides). Then, quality control of RNA and small RNA content and quality is performed.

[0048] 3.3 Based on the results of next-generation sequencing, biotinylated DNA probes were designed to capture target tRNA molecules using probe specificity, thereby obtaining small RNA samples for further purification.

[0049] 3.4 The purified tRNA was broken down by RNA restriction endonucleases or exonucleases to form different tRNA fragments. The tRNA fragments were then qualitatively analyzed by LC-MS / MS to confirm the modification type and site of the tRNA.

[0050] 3.5 Based on the obtained mass spectrometry results containing modification and site information, tRFs and t-halves sequences derived from traditional Chinese medicine (TCM) were designed, and a TCM tRNA sample library was constructed. The design principle of the modification sequences was based on the natural modifications of TCM, while adjusting and optimizing the modification types and sites to ensure the stability and efficacy of tRFs and t-halves. The main types of modifications included pseudouracil, dihydrouracil, various monomethylation modifications, various dimethylation modifications, various trimethylation modifications, and various 2'-oxymethylation modifications of nucleotide pentose sugars.

[0051] Step three above includes the following steps:

[0052] 4.1 Obtain the FASTA sequence and annotation file of the human gene from the GeneCode website, and extract the FASTA file of the human gene mRNA3'UTR.

[0053] 4.2 Using the miRanda online package (with thresholds set to Score: 140 and Energy: -1), we screened human genes that different Chinese medicine tRNA fragments might target, and initially established a screening library of Chinese medicine sequences and target genes.

[0054] Step four above includes the following steps:

[0055] 5.1 Based on the diseases mainly treated by the selected TOP5 traditional Chinese medicines, corresponding cells were selected to verify their gene inhibition efficiency.

[0056] 5.2 As described in 5.1, a concentration gradient of the traditional Chinese medicine tRNA fragment was established (main concentrations included: 400, 200, 100, 50, 25, 12.5, 6.25, 3.125, 1.5625, 0 nM), and validated by qPCR. The inhibition efficiency was expressed as the ratio of (control group expression level - transfection group expression level) / control group expression level. The concentration with the best inhibitory effect was selected as the targeting efficiency of the traditional Chinese medicine tRNA fragment and the human gene. An inhibition efficiency ≥70% was considered to have good targeting performance, between 50% and 70% was considered to have moderate targeting performance, and <50% was classified as poor targeting performance.

[0057] 5.3 The obtained Chinese herbal medicine tRNA fragments-human genes-inhibition efficiency are summarized to construct the feature set of the prediction model.

[0058] Step five above includes the following steps:

[0059] 6.1 Based on Python, feature variables are extracted from the sample dataset, mainly including tRNA fragment sequences, target gene names, and inhibition efficiencies from step four.

[0060] 6.2 The text variables in the feature variables are numerically transformed. The base sequence of tRNA is characterized using one-hot encoding. The "A", "C", "G", and "U" in the original nucleic acid sequence data can be converted into binary vectors, which can be encoded as (1, 0, 0, 0):(0, 1, 0, 0):(0, 0, 1, 0):(0, 0, 0, 1), forming an L×4 matrix (L is the sequence length). Gene names are assigned numerical codes using LabelEncoder, converting them into discrete integer features.

[0061] 6.3 A deep learning convolutional neural network (CNN) model was constructed using Keras. The deep learning neural network model consists of an input layer, convolutional layers, pooling layers, and an output layer (as shown in Table 1). The input layer mainly contains the feature variables transformed in 6.1 and 6.2, with an input dimension of (L, 4), which is the one-hot encoding matrix of each tRNA sequence. The convolutional layers use 3 convolutions with a filter size of 3×3, a stride of 1, and the activation function is ReLU. The pooling layers are the feature extraction module of the CNN and are connected to the shared feature layer. Each convolutional layer is followed by a 2×2 max pooling layer. The output layer of the model contains the corresponding target gene (Softmax activation function, output dimension is the number of all gene categories) and the corresponding inhibition efficiency (Linear activation function, output dimension is 1 dimension).

[0062] Table 1 CNN Model Parameter Settings

[0063]

[0064] 6.4 Model Compilation and Training Parameter Settings: ① The optimizer used in the optimizer configuration is the Adam optimizer, with an initial learning rate of 0.001. This value is commonly used in literature and experiments for the initial learning rate of deep neural networks and exhibits good convergence stability. In this invention, the learning rate is further optimized within the range [0.0001, 0.001, 0.01] using a grid search method. ② A multi-task learning strategy is adopted, with different loss functions set for the model's dual outputs: sparse categorical cross-entropy for the target gene classification subtask and mean squared error (MSE) for the inhibition efficiency regression subtask. Loss weights are explored and combined using a grid search method, including: [0.3, 0.7]: biased towards inhibition efficiency prediction; [0.5, 0.5]: balanced task weights; [0.7, 0.3]: biased towards target gene identification. ③ The evaluation metrics are classification accuracy, which measures the proportion of the model that correctly predicts the target gene; and mean absolute error (MAE), which measures the average absolute deviation between the predicted and actual values ​​of the inhibition efficiency.

[0065] 6.5 Then, model training is performed, with 20% of the input dataset selected as the validation set and 80% as the training set. The hyperparameters are tuned using the grid search method, with the epoch number ranging from [10, 50, 100, 200, 500] and the batch size ranging from [10, 32, 64, 128].

[0066] 6.6 A stratified K-fold cross-validation model was used, with K set to 5. Performance evaluation metrics primarily included: accuracy of target gene prediction and MAE value of inhibition efficiency. The closer the accuracy of the validation set prediction is to the target gene prediction accuracy of the training set (with a difference not exceeding 5%), the better the prediction performance. The smaller the difference between the MAE values ​​of the validation and test sets (with a difference not exceeding 3%), the better the model's generalization ability for inhibition efficiency prediction. A smaller standard deviation indicates a more stable model.

[0067] Example 2

[0068] A specific technical solution for a method to predict target genes of traditional Chinese medicine tRNA, tRFs, and t-halves based on artificial intelligence (taking Ginkgo biloba as an example):

[0069] A total of 67 tRNA sequences were isolated and identified from Ginkgo biloba. Through bioinformatics screening, Ginkgo tRFs and t-halves were screened to identify human genes that can be targeted. These target genes were verified by qPCR, and modeling samples were constructed. 80% of all modeling sample data were used as the training set for model construction, and 20% were used as the test set.

[0070] Disease-related cells were divided into control and experimental groups. After transfecting cells with the sequenced tRNA sequences of traditional Chinese medicine (TCM), total RNA was extracted and validated by qPCR. Inhibition efficiency was characterized as (expression level in control group - expression level in transfected group) / expression level in control group. Textual variables such as tRNA fragment sequence, target gene name, and inhibition efficiency were converted into numerical variables, and these characteristic variables were then incorporated into subsequent model construction.

[0071] This prediction model is primarily built using Python 3.6, employing the Keras package to construct a deep learning convolutional neural network (CNN) model. The deep learning neural network model comprises an input layer, convolutional layers, pooling layers, and an output layer. The input layer mainly contains numerically processed feature variables; the convolutional and pooling layers are the feature extraction modules of the CNN and are connected to a shared feature layer. The model's output layer contains the corresponding target gene and its corresponding inhibition efficiency, such as... Figure 2 As shown.

[0072] The model training parameters were configured, with the Adam optimizer used and an initial learning rate of 0.001. Sparse categorical cross-entropy and MSE were chosen as the loss functions; the loss weights were set between 0.3 and 0.7. Evaluation metrics included classification accuracy (measuring the proportion of correct target gene predictions) and mean absolute error (MAE) (evaluating the average absolute deviation between predicted and true values ​​for suppression efficiency).

[0073] Then, model training was performed, selecting 20% ​​of the input dataset as the validation set and 80% as the training set. The parameter optimization range was set, with the epoch number ultimately set to 50 and the batch size to 32. The model was validated using stratified cross-validation with K set to 5. Performance evaluation metrics primarily included: the accuracy of target gene prediction and the MAE value of inhibition efficiency. The closer the accuracy of the validation set to the target gene prediction accuracy of the training set (with a difference not exceeding 5%), the better the prediction performance. The smaller the difference between the MAE values ​​of the validation and test sets (with a difference not exceeding 3%), the better the model's generalization ability for inhibition efficiency prediction. As shown in Table 2, this demonstrates that the model has superior prediction performance.

[0074] Table 2 Model Performance Evaluation

[0075] Evaluation Metrics Training Set Test Set Gene Accuracy 0.8923±0.0112 0.8541±0.0186 Efficient MAE 0.0721±0.0034 0.0895±0.0051

[0076] Explanation of the scope and universality of the method in traditional Chinese medicine

[0077] Although the above examples are represented by ginkgo, the modeling framework proposed in this invention has broad applicability and is applicable to the mining of small RNAs and the prediction of target genes in a variety of Chinese medicinal materials. Simple transformations and substitutions of these RNAs are also within the scope of protection of this invention.

Claims

1. An artificial intelligence-based traditional Chinese medicine tRNA, tRF, t-half target gene prediction method, characterized in that, The method comprises the following steps: (1) extracting and separating tRNA of traditional Chinese medicine, performing second-generation sequencing to obtain tRNA sequences; (2) designing tRFs and t-halves sequences from traditional Chinese medicine according to the sequencing results, analyzing modification types and sites of tRNA through LC-MS / MS, and constructing a database of tRNA fragments of traditional Chinese medicine; (3) based on the sequence of human gene mRNA 3' UTR, using a base complementary pairing tool to preliminarily screen potential target genes of tRNA fragments of traditional Chinese medicine; (4) verifying the inhibition efficiency of tRFs and t-halves of traditional Chinese medicine on target genes through experiments, and constructing a sample data set in combination with the screening results of step (3); (5) constructing a deep learning convolutional neural network (CNN) model, training and optimizing the sample data set, and outputting the prediction results of target genes and their inhibition efficiency.

2. The target gene prediction method of claim 1, wherein, The extraction method of tRNA in step (1) comprises: (1.1) extracting total RNA of traditional Chinese medicine by using a plant RNA extraction kit, and separating small RNA with a length of less than 200 nt by using a small RNA enrichment kit; (1.2) separating tRNA by 6% urea-PAGE gel electrophoresis, cutting the target region, and recovering by electrophoresis; (1.3) constructing a cDNA library by using TruSeq Small RNA Sample Preparation Kit, and performing second-generation sequencing by using an Illumina platform.

3. The target gene prediction method of claim 1, wherein, The design of tRFs and t-halves sequences in step (2) comprises: (2.1) based on the cleavage position of the anticodon loop of tRNA, defining 5'-t-half, 3'-t-half, 5'-tRF and 3'-tRF sequences; (2.2) decomposing purified tRNA by using RNA restriction endonuclease or exonuclease, analyzing modification sites after fragment generation by LC-MS / MS; (2.3) optimizing modification types according to mass spectrometry results, including pseudouracil, dihydrouracil, methylation modification and 2'-O-methylation modification.

4. The target gene prediction method of claim 1, wherein, The base complementary pairing tool in step (3) is miRanda program, and the screening threshold is set to Score≥140 and Energy≤-1.

5. The target gene prediction method of claim 1, wherein, The experimental verification in step (4) comprises: (4.1) selecting a cell model related to the treatment of diseases by traditional Chinese medicine, and setting a concentration gradient (0-400 nM) of tRNA fragments of traditional Chinese medicine; (4.2) detecting the mRNA content of the target gene by qPCR, calculating the inhibition efficiency = (expression amount of the control group-expression amount of the experimental group) / expression amount of the control group × 100%, and adding the target gene with an inhibition efficiency of ≥70% into the sample data set.

6. The target gene prediction method of claim 1, wherein, The construction of the CNN model in step (5) comprises: (5.1) converting tRNA sequences into binary vectors by one-hot encoding, and numerically encoding gene names by LabelEncoder; (5.2) the model structure comprises an input layer, a convolution layer, a pooling layer and an output layer, and the output layer simultaneously predicts the target gene name and the inhibition efficiency; (5.3) The optimizer uses Adam with an initial learning rate of 0.001, a combination of sparse_categorical_crossentropy and mean squared error (MSE) as the loss function, and a loss weight interval of 0.3 to 0.

7.

7. The target gene prediction method of claim 6, wherein, The training parameters of the CNN model include: (7.1) The ratio of the training set to the validation set is 8:2; (7.2) The number of epochs is set to 50, and the batch size is 32; (7.3) The model performance is evaluated by stratified K-fold cross-validation (K=5), and the evaluation indicators include the prediction accuracy of the target gene and the MAE value of the inhibition efficiency.

8. The target gene prediction method of claim 1, wherein, The sequence of the human gene mRNA 3' UTR is derived from the GeneCode database.

9. The target gene prediction method of claim 2, wherein, The small RNA enrichment kit is Ambion mirVana™ miRNA Isolation Kit.

10. The target gene prediction method of claim 5, wherein, The concentration gradient includes 400, 200, 100, 50, 25, 12.5, 6.25, 3.125, 1.5625, and 0 nM.

Images