The invention discloses a method for aided analysis of 5apos based on a machine learning model. Method for influencing mRNA translation efficiency by UTR (Untranslated Region)

Abstract

The invention discloses a method for aided analysis of influence of 5 'UTR on mRNA translation efficiency based on a machine learning model, and relates to the technical field of synthetic biology and computational biology, the method comprises the following steps: S1, constructing a 5' UTR plasmid library; s2, transcribing the 5 'UTR plasmid into an independent mRNA (messenger ribonucleic acid) with a luciferase-opt luciferase, and taking the luminous intensity of the luciferase-opt luciferase as the translation intensity of the mRNA; s3, analyzing characteristics related to protein expression intensity of the mRNA; and S4, constructing a random forest algorithm model on the basis of the features analyzed in the step S3 to predict the influence of the 5 'UTR on the mRNA translation efficiency. According to the method for auxiliary analysis of the influence of the 5 'UTR on the mRNA translation efficiency based on the machine learning model, a feasible technical route reference is provided for sequence optimization of mRNA in subsequent mRNA vaccine development, so that the mRNA vaccine can efficiently express target protein in a body.

Description

Technical Field

[0001] The present invention relates to the technical field of synthetic biology and computational biology, and in particular to a method for assisted analysis of the effect of 5' UTR on mRNA translation efficiency based on a machine learning model. Background Art

[0002] mRNA vaccine is a new type of vaccine. Its working principle is to use RNA molecules to carry information encoding pathogen-related proteins, so that human cells produce corresponding antigenic proteins, thereby inducing an immune response. Compared with traditional vaccines, mRNA vaccines have the advantages of rapid preparation, low production cost, and high safety. mRNA sequence design is a key step in the development of mRNA vaccines. When designing the mRNA sequence, it is necessary to determine the optimal antigen protein sequence as well as the optimal RNA sequence and morphological structure to improve the immunogenicity and stability of the vaccine. In addition, RNA molecules are easily degraded, so chemical modification and other means are needed to improve RNA to improve its stability and immune tolerance. The conversion efficiency of RNA is also one of the key factors, because only RNA that can be efficiently converted into protein can produce a sufficient immune response.

[0003] In recent years, studies have found that there are important regulatory elements in the 5' untranslated region (5' UTR) of mRNA, including the internal ribosome entry site (IRES), 5' UTR secondary structure, G-quadruplexes (G4), 5' cap structure, upstream open reading frame (u ORF), Kozak sequence, upstream start codon ATG (upstream ATG, u ATG), 5' UTR length, 18 base pairs after the translation start ATG, stem-loop structure and mRNA secondary structure. The regulation of gene expression by 5' UTR involves multiple levels and plays an important role in maintaining the stability of mRNA, nuclear transport, RNA splicing and processing, and cell proliferation. Some of the more popular transgenic technologies currently use u ORF to micro-regulate the translation level.

[0004] In addition, the nucleotide composition of 5'UTR affects the initiation of translation, and there may be random epistatic effects between nucleotides. Gene expression mainly involves four levels: transcriptional regulation, post-transcriptional regulation, translational regulation, and post-translational regulation. Its influencing factors mainly include the length of the coding sequence and UTR, nucleotide composition and structure, whether the sequence near ATG conforms to the Kozak rule, the presence of u ORF, possible target sites of mi RNA, codon usage, amino acid composition, and protein degradation signals. The initiation process of gene expression is the key to studying the regulatory mechanism.

[0005] The regulation of 5'UTR at the transcription and translation levels is mainly reflected in the normal progress of life activities under cell proliferation, differentiation, growth and development, apoptosis and stress conditions, and plays an important role in studying the tumorigenesis process and the mechanism of action of cancer. Although the sequence design of mRNA vaccines can greatly increase the stability and translation efficiency of mRNA by adding chemical modification groups, optimizing codon usage and optimizing 5' and 3' UTR. However, compared with other optimization methods, the optimization method of 5'UTR is a highly targeted and stable optimization method that does not require changing the RNA coding sequence, will not affect the immunogenicity of the vaccine, and can significantly improve the translation efficiency and stability of RNA. However, there is no effective method to optimize the sequence composition of 5'UTR, and many cis-regulatory elements contained in 5'UTR have a very large impact on the translation efficiency of mRNA.

[0006] Traditional experimental methods for analyzing the effect of 5' UTR on mRNA translation efficiency have the following disadvantages: 1. TE value (Translation Efficiency) or MRL value (Mean Ribosome Load) is difficult to reflect the protein expression of mRNA vaccines in vivo: Current experiments basically use TE value or MRL to reflect protein expression. The TE value is the ratio of ribosome protected fragments (RPF, Ribosome Protected Fragments) to the corresponding mRNA level, which reflects the relative translation efficiency of mRNA and does not directly reflect the actual number of proteins produced by a single mRNA molecule. The MRL value is the average number of ribosomes on mRNA at a given time point, ignoring the dynamics and complexity of translation.

[0007] 2. Complexity of 5'UTR regulatory mechanism: Many cis-regulatory elements contained in 5'UTR have a great influence on the efficiency of mRNA translation. Rational design of 5'UTR sequence to enhance protein translation efficiency requires precise control of the combination and position of these cis-elements. However, due to the interaction between cis-elements within 5'UTR, arbitrary changes may reduce rather than enhance translation efficiency.

[0008] 3. Unable to analyze multiple 5'UTR sequences simultaneously: Traditional experimental methods usually analyze a single 5'UTR sequence and are unable to analyze the changes in translation efficiency of multiple 5'UTR sequences simultaneously, thus limiting the depth of the research.

[0009] 4. High time and cost: Traditional experimental methods require a lot of experimental operations and data analysis, which is time-consuming and costly, thus limiting the scale and scope of the research. Summary of the invention

[0010] To solve the above problems, the present invention provides a method for analyzing the effect of 5'UTR on mRNA translation efficiency based on a machine learning model, which provides a feasible technical route reference for mRNA sequence optimization in the subsequent development of mRNA vaccines, so that mRNA vaccines can efficiently express target proteins in the body.

[0011] To achieve the above object, the present invention provides a method for analyzing the effect of 5'UTR on mRNA translation efficiency based on a machine learning model, comprising the following steps: S1, construction of 5'UTR plasmid library; S2, transcribe the 5'UTR plasmid in the plasmid library into a separate mRNA carrying luciferase-opt luciferase, and use the luminescence intensity of luciferase-opt luciferase as the translation intensity of mRNA; S3. Characteristic analysis of mRNA transcribed from 5'UTR plasmid, including k-mer frequency, codon usage frequency, GC content percentage, GC content ratio, AT content ratio, Kozak sequence, minimum free energy of RNA molecule, and average base pair distance; S4. Based on the features analyzed in step S3, a random forest algorithm model is constructed to predict the effect of 5'UTR on mRNA translation efficiency.

[0012] Preferably, in step S1, constructing a 5'UTR plasmid library comprises: S11, screening 5'UTR sequences with a length of 100 bp-200 bp, and performing similarity clustering on the screened 5'UTR sequences; S12, construct a T7 promoter-5'UTR-luciferase-opt-3'UTR_Pfizer pUC57 plasmid vector by gene synthesis; S13, obtaining the 5'UTR gene by PCR, connecting the 5'UTR and the pUC57 plasmid vector into a plasmid by Gibson seamless ligation technology, and constructing a plasmid library; S14. The plasmid library was transformed into DH5α competent cells by chemical method, cultured overnight to obtain single colonies, single colonies were picked for expansion culture, plasmids were extracted, and plasmids with 5'UTR were obtained.

[0013] Preferably, in step S2, the determination of the luminescence intensity of luciferase-opt luciferase comprises: (1) Transfect individual mRNAs carrying luciferase-opt into HEK293 cells for expression, and accurately measure the fluorescence intensity of each mRNA using a high-throughput automated microplate reader. (2) The mRNA expression intensity of the BNT-162b2-vaccine human α-hemoglobin truncated 5' UTR was used as a standard, and the measured mRNA fluorescence intensity was normalized to quantify the mRNA translation intensity.

[0014] Preferably, in step S4, the specific steps of model construction include: S41. Feature processing (1) k-mer frequency: This is achieved through a simple sliding window algorithm, which traverses the entire sequence, records the number of times each k-mer appears, traverses the sequence, takes a subsequence of length K each time (called a K-mer), and counts the number of times each K-mer appears; If the K-mer is already in the kmer_counts dictionary, increase its count; otherwise, add it to the dictionary and set the count to 1; (2) Codon usage frequency: Translate the input sequence into a codon sequence, and convert the codon sequence into a string; then traverse the string, count the number of times each codon appears, and store the results in a dictionary feature_map; The number of start codons and stop codons was calculated as the basis for the evaluation of upstream open reading frames and upstream start codons, and this information was also added to the dictionary feature_map; (3) GC content percentage: traverse the input sequence string, count the number of times the bases G and C appear, divide the total number of G and C by the sequence length, and store the result in a dictionary N_count; (4) GC content ratio: traverse the input sequence string, count the number of occurrences of bases G and C, and obtain the number by dividing the number of Cs by the absolute value of the difference between the number of Gs and 1, and store the result in a dictionary N_count; (5) AT content ratio: traverse the input sequence string, count the number of occurrences of bases A and T, and obtain the number of Ts divided by the absolute value of the difference between the number of A and 1, and store the result in a dictionary N_count; (6) Kozak sequence: perform a regular sequence match at the start codon, 1 means it matches the Kozak sequence, and 0 means there is no Kozak sequence; (7) Minimum free energy of RNA molecules: input DNA sequence information, use RNAfold package to predict secondary structure, and output minimum free energy; (8) Average base pair distance: obtained by calling the mean_bp_distance method after predicting the secondary structure using RNAfold; (9) Stability energy of the G-quadruplex structure within the 5' UTR; S42, using the random sequence samples and random features in step S41 to construct a decision tree, and using the advance feature information to construct a least squares regression tree. The model has a total of 100 least squares regression trees.

[0015] Preferably, the least squares regression tree construction formula includes: and in, R is the division of the input space, select j Features and its specific value s As the segmentation variable and segmentation point, two regions are defined: and , in order to find the optimal segmentation variable and the optimal segmentation point s ; in, It is the first feature space after segmentation The mean of It is the second feature space after segmentation The mean of is the eigenvalue of each sample in the feature space, j For the selected j features, and s is the specific feature segmentation value.

[0016] The present invention provides a method for analyzing the effect of 5'UTR on mRNA translation efficiency based on a machine learning model, which has the following beneficial effects: (1) The technical route adopted is based on the combination of dry and wet experiments. The data generated by wet experiments are accurate, and the dry experiment data analysis method is accurate and high-throughput; (2) The mRNA sequence optimization method is based on the 5' UTR region with a length of 100bp-200bp. This length range not only covers most of the 5' UTR in the human genome, but also ensures sequence diversity. By streamlining the data set by clustering, the problem of requiring a large data set to build a machine learning algorithm model can be alleviated, so that the model can achieve similar results to the model trained with big data after being trained with a smaller data set. (3) The quantitative mRNA expression intensity of the present invention is measured by a high-throughput automated microplate reader, and the relative expression intensity of each 5'UTR is obtained by homogenization. This measurement method is accurate and high-throughput, and can automatically process large-scale samples; (4) The input of the random forest model established in the present invention is the 5'UTR sequence and its annotated fluorescence intensity. Compared with the traditional TE value or MRL, the fluorescence intensity can better correspond to the protein expression intensity information, and the prediction accuracy of the established model is higher; (5) The present invention has carried out a relatively systematic analysis of the characteristics of the cis-regulatory elements of the 5'UTR, such as 5'UTR length distribution analysis, upstream open reading frame UORFs site analysis, upstream start codon UAUGs base distribution analysis, mRNA secondary structure and its minimum free energy prediction, Kozak sequence analysis and expression intensity correlation analysis, etc. These characteristics are related to protein expression intensity, and the random forest model established using these data has a higher prediction accuracy; (6) The present invention adopts a random forest algorithm model and uses ExtraTreeRegressor as the core algorithm, which is an extreme random tree regressor and a type of ensemble learning method. Unlike the traditional decision tree algorithm, the extreme random tree introduces a higher degree of randomness. It not only performs random sampling when selecting a subset of the data set for training, but also randomly selects features during the division of each node. This step is significantly different from the conventional decision tree method of selecting the optimal feature when splitting a node. The introduction of this randomness is intended to reduce the variance of the model and improve the generalization ability of the model to obtain better optimization results.

[0017] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1are the clustering test results, where A is the test result of training set 1 (5838 items); B is the test result of training set 2 (2468 items); Figure 2 A technical route to establish a machine learning model to assist in analyzing the effect of 5' UTR on mRNA translation efficiency; Figure 3 Flowchart for plasmid library construction; Figure 4 The pUC57 plasmid vector is T7 promoter-5'UTR-luciferase-opt-3'UTR_Pfizer; Figure 5 Length distribution of the constructed pre-experimental library (5'UTR+18cds); Figure 6 An example of a scoring scheme for RNAfold; Figure 7 A diagram of the random forest framework and one of the decision trees in action; Figure 8 The ten-fold cross correlation comparison between the model established in Example 2 and the random forest (Cao); Fig. 9 The results of model training comparison between the public dataset and the dataset established in Example 2, where A is the ten-fold cross correlation of the public dataset, and B is the ten-fold cross correlation of the dataset established in Example 2; Fig.10 The model training comparison results are for 400 clustered public test set data and 1326 unclustered data. Part A establishes the correlation between the predicted value and the true value of the model for the 1326 unclustered data, and Part B establishes the correlation between the predicted value and the true value of the model for the 400 clustered data sets. DETAILED DESCRIPTION

[0019] In order to make the purpose, technical scheme and advantages disclosed in the embodiments of the present invention more clearly understood, the embodiments of the present invention are further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the embodiments of the present invention and are not intended to limit the embodiments of the present invention. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of this application.

[0020] Embodiment 1 Determine the clustering scheme (1) Clustering method: In order to improve the efficiency and accuracy of data processing, the cd-hit software was used to cluster sequences with 50% similarity. The conditional command used was: cd-hit-i 5' UTR -o 5' UTR_cluster -c 0.6 -n4 (cd-hit was used because the lower limit of clustering threshold of cd-hit-est was 0.8).

[0021] (2) Clustering scheme test: First, the 6721 5'UTR data in the GSE database (https: / / www.ncbi.nlm.nih.gov / geo / query / acc.cgi?acc=GSE176581) were clustered (only the 5'UTR sequence data derived from the wild type were considered; secondly, the sequence data with missing TE values ​​were excluded).

[0022] In order to simulate the data from a larger data set such as the 5' UTR of the human genome as much as possible in the 6721 5' UTR data set, it is necessary to select as many clusters as possible to train the model. However, the learning resources of the cluster machine learning model with count 1 are limited and it is not suitable to be added to the test set. Therefore, a sequence from the 883 clusters with count ≥ 2 is selected as the representative sequence and listed as the test set, with a total of 883 sequences. The test results in this test set can test the generalization ability of the model to the greatest extent.

[0023] The remaining 5838 5'UTR data out of the 6721 5'UTR data are used as training set 1. The 5838 5'UTR data are clustered by similarity in step (1), and a total of 2468 5'UTR data are obtained, which are listed as training set 2.

[0024] Apply training set 1 and training set 2 to the same random forest model, and use the test set to test how much data is needed to achieve the highest possible model accuracy and good generalization ability.

[0025] (3) Clustering test results: The test results of training set 1 (5838 items) are as follows Figure 1 As shown in Part A, the test results of training set 2 (2468 items) are as follows Figure 1 As shown in part B. It can be seen that after clustering, the training set data is compressed by half, and the Pearson correlation of nearly 0.6 and the mean square error of 0.98 can still be achieved, which shows the feasibility of the clustering scheme.

[0026] Embodiment 2 A method for analyzing the effect of 5' UTR on mRNA translation efficiency based on machine learning model, the technical route is as follows Figure 2 As shown, the specific steps include: S1. Construction of 5'UTR pre-experimental plasmid library (1) 5'UTR data screening: In the GSE database (https: / / www.ncbi.nlm.nih.gov / geo / query / acc.cgi?acc=GSE176581), data were selected by clustering, and the 5'UTR in the PC3 cell line was finally selected as the experimental data. The TE value of the sequence was comprehensively considered to make it as close to the normal distribution as possible. Finally, 400 5'UTR sequences with a length ranging from 100bp to 200bp were selected.

[0027] Then, the corresponding 5'UTR sequence in the Hg19 genome data was downloaded from the USCS database (https: / / genome.ucsc.edu / ) (https: / / genome.ucsc.edu / cgi-bin / hgTables), aiming to find the first 18 bases of the CDS (protein-coding region sequence) of the gene corresponding to the 5'UTR, and splice this base segment to the end of the 5'UTR sequence.

[0028] (2) Construction of plasmid library: Figure 3 As shown, including: Plasmid vector construction: A T7 promoter-5' UTR-luciferase-opt-3' UTR_Pfizer pUC57 plasmid vector was constructed by gene synthesis, and the luminescence intensity of luciferase-opt luciferase was used as the translation intensity of mRNA. Figure 4 Shown is the pUC57 plasmid vector of T7 promoter-5' UTR-luciferase-opt-3' UTR_Pfizer.

[0029] Primer design: Design the front and back primers of each 5'UTR according to the primer design guidelines, and design the pUC57 vector according to the primer design guidelines.

[0030] PCR (polymerase chain reaction): Prepare PCR reaction solution, including DNA template (plasmid with 5' UTR sequence or pUC57 plasmid vector), primers (forward and reverse primers), dNTP (deoxynucleoside triphosphate), buffer and DNA polymerase. Mix these components in proportion according to the experimental needs and the formula provided by the reagent manufacturer.

[0031] Prepare clean PCR reaction tubes and dispense the PCR reaction solution into the reaction tubes. Amplify the 5'UTR insert and plasmid vector using the set PCR amplification program, and collect the two gene fragments through gel purification.

[0032] Seamless ligation: Prepare the mixture for the ligation reaction. The mixture includes the fragment assembly products (insert fragment and plasmid vector), DNA polymerase, dNTP (deoxynucleoside triphosphate) and buffer. Mix them in the appropriate molar ratio and incubate the reaction at the appropriate temperature.

[0033] Transformation: Use chemical methods to transform the connected plasmid into DH5α competent cells. The transformed cells are inoculated on a selective medium containing kanamycin antibiotics for screening. Only cells carrying the 5' UTR gene can grow. Finally, sequencing is used to verify whether the 5' UTR gene sequence has been successfully inserted into the plasmid.

[0034] S2. High-throughput screening Each plasmid in the plasmid library is transcribed into a separate mRNA with luciferase-opt through enzyme digestion, in vitro transcription and other technologies, and then transfected into HEK293 cells for expression through cell transfection technology, and the fluorescence intensity of each mRNA is accurately measured by a high-throughput automated microplate reader. The mRNA expression intensity with the truncated 5' UTR of human α-hemoglobin (HBA1) of the Pfizer vaccine is used as the standard, and the fluorescence intensities of all measured mRNAs are normalized to quantify the expression intensity of the mRNA.

[0035] The obtained mRNA was characterized and the RNAfold toolkit was used to predict the secondary structure of mRNA and the minimum free energy visualization analysis, as well as the influence of the first 18 base sequences of the CDS (protein coding sequence) region on the translation intensity. Python was used to analyze the 5' UTR length distribution and the correlation between length and expression intensity, and the upstream start codons UAUGs and upstream open reading frames UORFs were analyzed.

[0036] like Figure 5 Shown is the length distribution of the constructed pre-experimental library (5'UTR+18cds).

[0037] S3. Build a prediction model A random forest algorithm model was constructed using python+sklearn+RNAfold to predict the effect of 5'UTR on the expression intensity of mRNA, and a script was used to extract DNA sequence information. A series of biologically significant features were extracted using the biopython toolkit. These features are as follows: Codon Usage Frequency, which covers the frequency of occurrence of each codon in the sequence; K-mer frequency includes the frequency of one to six base combinations in the sequence; The number of untranslated regions before the start codon (uAUG Count), which provides statistics of the region before the start codon AUG; The percentage of C and G content (CG Percentage) indicates the ratio of the total ratio of G and C in the DNA sequence to the entire sequence length. In the genome, a higher CG content is usually associated with the presence of CpG islands. CpG islands are DNA regions rich in CpG dinucleotides, often located near gene promoters. The high CG content of these regions is usually associated with the transcriptional activity of genes because CpG islands can affect the expression level of genes. CpG islands are often regions where transcription factors bind, playing a key role in the initiation and regulation of genes.

[0038] The CG content ratio (CG Ratio) indicates the ratio between the number of C and the number of G in a DNA sequence. If the CG Ratio is close to 1, it means that the number of C and G is roughly equal; if the CG Ratio is greater than 1, it means that the number of C is greater than the number of G; if the CG Ratio is less than 1, it means that the number of G is greater than the number of C.

[0039] AT Ratio indicates the ratio between the number of A and the number of T in a DNA sequence. If the AT Ratio is close to 1, it means that the number of A and T is roughly equal; if the AT Ratio is greater than 1, it means that the number of A is greater than the number of T; if the AT Ratio is less than 1, it means that the number of T is greater than the number of A.

[0040] Kozak sequence is a nucleic acid sequence located behind the cap structure at the 5' end of eukaryotic mRNA, usually GCCACCAUGG, which can bind to the translation initiation factor and mediate the translation initiation of mRNA containing the 5' cap structure. It corresponds to the SD sequence of prokaryotes. It is a sequence present in eukaryotic mRNA and plays an important role in the initiation of translation.

[0041] The minimum free energy of RNA molecules (Energy of Whole RNA Molecule) indicates the stability of RNA structure. The lower the free energy, the more stable the structure. Calculating the free energy of the structure helps to evaluate the stability of the predicted RNA secondary structure in the natural environment. The stability of RNA secondary structure is related to the binding of substances such as ribosomes and translation initiation complexes to RNA, thereby affecting the efficiency of protein translation.

[0042] Mean base pair distance (mean_distance), which calculates the average base pair distance of mRNA secondary structure, is an indicator to measure the stability of RNA secondary structure.

[0043] Energy of G-quadruplex in 5' UTR. G-quadruplex is a high-order structure formed by folding DNA or RNA rich in tandem repeats of guanine (G). G-quartet is the structural unit of quadruplex, with four Gs connected by Hoogsteen hydrogen bonds to form a ring-shaped plane. Two or more layers of tetrads form a quadruplex through π-π stacking.

[0044] These features are used as the basis for building random forest decision trees, and the ExtraTreeRegressor extreme decision tree is used to control the randomness of each decision tree. In this way, the model can benefit from the different rules learned by each tree. The final prediction result is obtained by averaging the prediction results of all decision trees. This ensemble learning method effectively utilizes the independence of multiple trees, reduces the risk of overfitting, and also enhances the model's ability to fit complex data structures.

[0045] The specific steps of model construction include: S1. Feature processing (1) k-mer frequency: This is achieved through a simple sliding window algorithm that traverses the entire sequence, records the number of times each k-mer appears, traverses the sequence, takes a subsequence of length K each time (called a K-mer), and counts the number of times each K-mer appears.

[0046] If the K-mer is already in the kmer_counts dictionary, increment its count; otherwise, add it to the dictionary and set the count to 1.

[0047] K-mer is a concept commonly used in bioinformatics to describe continuous subsequences in DNA, RNA or protein sequences. The length of K-mer is determined by the parameter k.

[0048] (2) Codon usage frequency: The input sequence is translated into a codon sequence, which is then converted into a string. Next, it traverses the string, counts the number of times each codon appears, and stores the results in a dictionary (feature_map). Finally, the number of start codons (AUG) and stop codons (ORF) is also calculated as a basis for evaluating the upstream open reading frame (uORF) and the upstream start codon (uAUG), and this information is also added to the dictionary.

[0049] (3) GC content percentage (CGperc): Traverse the input sequence string, count the number of times the bases G and C appear, divide the total number of G and C by the sequence length, and store the result in a dictionary (N_count).

[0050] (4) GC content ratio (CG Ratio): CG Ratio can provide information about the stability of the genome of an organism. Traverse the input sequence string, count the number of times the bases G and C appear, and obtain the number of Cs divided by the absolute value of the difference between the number of Gs and 1, and store the result in a dictionary (N_count).

[0051] (5) AT content ratio (AT Ratio): AT Ratio can also provide information about the stability of the organism's genome. Traverse the input sequence string, count the number of bases A and T, and obtain the number of Ts divided by the absolute value of the difference between the number of A and 1, and store the result in a dictionary (N_count).

[0052] (6) Kozak sequence: Perform a regular sequence match at the start codon. 1 indicates that the Kozak sequence is met, and 0 indicates that there is no Kozak sequence.

[0053] (7) Energy of Whole RNA Molecule: Input DNA sequence information and use RNAfold package to predict secondary structure and output MFE (minimum free energy). RNAfold uses Zuker algorithm, which is based on dynamic programming algorithm and first requires a scoring scheme for base pair matching.

[0054] like Figure 6 The figure below shows an example of the scoring scheme for RNAfold. The hydrogen bonds produced by the pairing of AU and CG will reduce the free energy of the secondary structure, so it is set to a negative number. The algorithm aims to find the lowest overall free energy, which requires finding the most suitable secondary structure with the largest number of pairings.

[0055] The Zuker algorithm also considers variants that include stacking energy to calculate RNA structures. The structure enclosed by base pairs can be a hairpin, an internal loop, or multiple loops. The base pair matching table used will also be two-dimensional. The pairing scoring scheme for each substructure (hairpin, internal loop, or multiple loop) is different. RNAfold will distinguish different structures internally. For details, see the RNAfold manual (https: / / bio.libretexts.org / Bookshelves / Computational_Biology / Book%3A_Computational_Biology_-_Genomes_Networks_and_Evolution_(Kellis_et_al.) / 10%3A_RNA_Folding / 10.05%3A_RNA_Folding_Problem_and_Approaches).

[0056] (8) Mean base pair distance (mean_distance): Calculates the average base pair distance of the mRNA secondary structure, which is an indicator of the stability of the RNA secondary structure. After the secondary structure is predicted by RNAfold, the mean_bp_distance method is called to obtain it. This method calculates the distance between each pair of bases in the optimal structure and returns the average of these distances.

[0057] (9) Energy of G-quadruplex in 5' UTR: This feature is a more accurate version of the minimum free energy of RNA molecules, taking into account the quadruplex situation, while the minimum free energy of RNA molecules only considers hairpin loops, internal loops, multiple loops, etc. Considering the G-quadruplex structure may result in a lower minimum free energy, and therefore is more accurate.

[0058] S2. Model calculation The random forest is composed of many weak predictors, and the final calculation result is obtained by outputting results from each predictor and weighting them. The weak predictor is composed of decision trees. The random forest decision tree used in this embodiment is a least squares regression tree. The model has a total of 100 least squares regression trees, and each least squares regression tree uses random samples and features.

[0059] Input: training dataset D (sequences and corresponding annotations (TE values ​​or mean fluorescence intensity)); The input training data set D first passes through the feature processing module in step S1 to extract various feature information, and then uses the feature information to construct a least squares regression tree.

[0060] Output: Least Squares Regression Tree .

[0061] The construction formula of the least squares regression decision tree is as follows: and ; in, R It is a division of the input space. Here we use a heuristic method to select the variables and its value s , as the segmentation variable and segmentation point, thus defining two regions: and Then we can find the optimal segmentation variable and the optimal segmentation point. s Here s is the specific value of a feature. It is Features. , It is used The two feature spaces are divided.

[0062] in, It is the first feature space after segmentation The mean of It is the second feature space after segmentation The purpose of the algorithm is to find the segmentation feature with the minimum mean square error in the current feature space. j and the specific feature values ​​of the segmentation s . Eventually, a decision tree will be formed, such as Figure 7 shown.

[0063] Prediction results (1) Random Forest (Cao) is a 5' UTR random forest prediction model developed by Jicong Cao et al., which comes from High-throughput 5′ UTR engineering for enhanced protein production in non-viral gene therapies (https: / / doi.org / 10.1038 / s41467-021-24436-7) and is used here for comparison with the model of this example.

[0064] like Figure 8 As shown, the ten-fold cross correlation comparison between the model established in this embodiment and the random forest (Cao) is shown.

[0065] It can be seen that after the optimization of the feature extraction module and the algorithm in this embodiment, the model correlation is higher, and the average Pearson correlation is higher by 0.1.

[0066] (2) The dataset annotated with fluorescence intensity established in this example and the existing publicly available dataset annotated with TE value (from the GSE database (https: / / www.ncbi.nlm.nih.gov / geo / query / acc.cgi?acc=GSE176581)) were used for model training to establish a 5' UTR random forest prediction model.

[0067] like Fig. 9 As shown, part A is the ten-fold cross correlation of the public data set, and part B is the ten-fold cross correlation of the data set established in this embodiment.

[0068] It can be seen that the experimental data set using fluorescence intensity as an annotation in this embodiment is of higher quality than the data set using TE value in the current academic field, and the model trained using the data set of this embodiment can have higher prediction accuracy. This is because fluorescence intensity can better correspond to protein expression intensity information, and the features extracted by the model in this embodiment are also features related to protein expression intensity.

[0069] (3) Model training was compared with 400 clustered public test set data and 1326 unclustered data (the 400 sequences were obtained by clustering the 1326 PC3 data in the Jicong Cao training set).

[0070] The results are as follows Fig.10 As shown, part A is the correlation between the predicted value and the true value of the model established by 1326 unclustered data sets, and part B is the correlation between the predicted value and the true value of the model established by 400 clustered data sets.

[0071] It can be seen that the Pearson correlation of the clustered model is further improved because the noise generated by some redundant data is reduced.

[0072] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that they can still modify or replace the technical solution of the present invention with equivalents, and these modifications or equivalent replacements cannot cause the modified technical solution to deviate from the spirit and scope of the technical solution of the present invention.

Claims

1. A method for analyzing the effect of 5'UTR on mRNA translation efficiency based on a machine learning model, characterized in that: The steps include: S1, construction of 5'UTR plasmid library; S2, transcribe the 5'UTR plasmid in the plasmid library into a separate mRNA carrying luciferase-opt luciferase, and use the luminescence intensity of luciferase-opt luciferase as the translation intensity of mRNA; S3. Characteristic analysis of mRNA transcribed from 5'UTR plasmid, including k-mer frequency, codon usage frequency, GC content percentage, GC content ratio, AT content ratio, Kozak sequence, minimum free energy of RNA molecule, and average base pair distance; S4. Based on the features analyzed in step S3, a random forest algorithm model is constructed to predict the effect of 5'UTR on mRNA translation efficiency.

2. The method for analyzing the effect of 5'UTR on mRNA translation efficiency based on a machine learning model according to claim 1, characterized in that: In step S1, constructing a 5'UTR plasmid library comprises: S11, screening 5'UTR sequences with a length of 100 bp-200 bp, and performing similarity clustering on the screened 5'UTR sequences; S12, construct a T7 promoter-5'UTR-luciferase-opt-3'UTR_Pfizer pUC57 plasmid vector by gene synthesis; S13, obtaining the 5'UTR gene by PCR, connecting the 5'UTR and the pUC57 plasmid vector into a plasmid by Gibson seamless ligation technology, and constructing a plasmid library; S14. The plasmid library was transformed into DH5α competent cells by chemical method, cultured overnight to obtain single colonies, single colonies were picked for expansion culture, plasmids were extracted, and plasmids with 5'UTR were obtained.

3. The method for analyzing the effect of 5'UTR on mRNA translation efficiency based on a machine learning model according to claim 1, characterized in that: In step S2, the determination of the luminescence intensity of luciferase-opt luciferase comprises: (1) Transfect individual mRNAs carrying luciferase-opt into HEK293 cells for expression, and accurately measure the fluorescence intensity of each mRNA using a high-throughput automated microplate reader. (2) The mRNA expression intensity of the BNT-162b2-vaccine human α-hemoglobin truncated 5' UTR was used as a standard, and the measured mRNA fluorescence intensity was normalized to quantify the mRNA translation intensity.

4. The method for analyzing the effect of 5'UTR on mRNA translation efficiency based on a machine learning model according to claim 1, characterized in that: In step S4, the specific steps of model construction include: S41. Input sequence samples and their fluorescence intensity annotations, and perform feature processing on the input, including: (1) k-mer frequency: This is achieved through a simple sliding window algorithm, which traverses the entire sequence, records the number of times each k-mer appears, traverses the sequence, takes a subsequence of length K each time (called a K-mer), and counts the number of times each K-mer appears; If the K-mer is already in the kmer_counts dictionary, increase its count; otherwise, add it to the dictionary and set the count to 1; (2) Codon usage frequency: Translate the input sequence into a codon sequence, and convert the codon sequence into a string; then traverse the string, count the number of times each codon appears, and store the results in a dictionary feature_map; The number of start codons and stop codons was calculated as the basis for the evaluation of upstream open reading frames and upstream start codons, and this information was also added to the dictionary feature_map; (3) GC content percentage: traverse the input sequence string, count the number of times the bases G and C appear, divide the total number of G and C by the sequence length, and store the result in a dictionary N_count; (4) GC content ratio: traverse the input sequence string, count the number of occurrences of bases G and C, and obtain the number by dividing the number of Cs by the absolute value of the difference between the number of Gs and 1, and store the result in a dictionary N_count; (5) AT content ratio: traverse the input sequence string, count the number of occurrences of bases A and T, and obtain the number of Ts divided by the absolute value of the difference between the number of A and 1, and store the result in a dictionary N_count; (6) Kozak sequence: perform a regular sequence match at the start codon, 1 means it matches the Kozak sequence, and 0 means there is no Kozak sequence; (7) Minimum free energy of RNA molecules: input DNA sequence information, use RNAfold package to predict secondary structure, and output minimum free energy; (8) Average base pair distance: obtained by calling the mean_bp_distance method after predicting the secondary structure using RNAfold; (9) Stability energy of the G-quadruplex structure within the 5' UTR; S42, using the random sequence samples and random features in step S41 to construct a decision tree, and using the advance feature information to construct a least squares regression tree. The model has a total of 100 least squares regression trees.

5. The method for analyzing the effect of 5'UTR on mRNA translation efficiency based on machine learning model according to claim 4, characterized in that: The least squares regression tree construction formula includes: and in, R is the division of the input space, select j Features and its specific value s As the segmentation variable and segmentation point, two regions are defined: and , in order to find the optimal segmentation variable and the optimal segmentation point s ; in, It is the first feature space after segmentation The mean of It is the second feature space after segmentation The mean of is the eigenvalue of each sample in the feature space, j For the selected j features, and s is the specific feature segmentation value.

Images