A codon optimization method and system based on a BERT model

Abstract

The application discloses a codon optimization method and system based on a BERT model, belongs to the cross technical field of bioinformatics and genetic engineering, and comprises the following steps: training a BERT model on a specific species genome; using a multi-head self-attention mechanism in the model to extract context semantic features of the genome sequence; and analyzing the context dependency relationship between codons through model output to realize host-specific codon optimization. The application can significantly improve the expression level of target proteins, and the codon optimization method can be widely applied to the heterologous expression of industrial enzymes, protein drugs and the like, and breaks through the technical bottleneck that traditional methods cannot identify the context dependency relationship of host-specific codons.

Description

Technical Field

[0001] This invention belongs to the fields of bioinformatics and genetic engineering, and in particular relates to a codon optimization method and system based on the BERT model. Background Technology

[0002] A codon, the basic unit of the genetic code, is a set of three consecutive nucleotides located on messenger RNA (mRNA). Four nucleotides (adenine-A, guanine-G, cytosine-C, and uracil-U) form a set of 64 codons depending on their sequence. Synonymous codons are those that encode the same amino acid but may differ in sequence. A single amino acid can be designated by multiple different codons; this phenomenon reflects the degeneracy of the genetic code. The presence of synonymous codons reduces the risk of protein functional changes due to gene mutations. That is, if a gene mutation occurs on mRNA and results in the production of a synonymous codon (ACA→ACG, threonine), the synthesized protein sequence will still be the same, and the final protein structure and function will remain unaffected. In different species, the frequency of use of codons encoding the same amino acid is not equal; this is called codon bias, which has a significant impact on the rate of protein synthesis, protein folding, gene expression regulation, and evolutionary adaptation. In addition, specific codons serve as initiation and termination signals in the protein synthesis process. AUG is the main initiation codon, which not only encodes methionine (Met) but also marks the start of protein synthesis in most cases. UAA, UAG, and UGA are three stop codons that mark the end of protein synthesis and do not encode any amino acids.

[0003] In the field of codon optimization, pioneering solutions typically employ a single objective (primarily CAI, requiring the replacement of rare codons with only the most common codons in the host organism). Methods exist to optimize gene expression using codon usage tables, where each codon is assigned a probability score based on the frequency distribution of codons in the genome normalized to each amino acid. Clearly, codon sequence optimization is inherently a multi-objective problem, requiring the simultaneous optimization of multiple metrics; one study developed a dynamic programming algorithm that can sequentially optimize three objectives. Given that the above approach is impractical for typical genes, other methods employ heuristics based on simulated annealing. Additionally, schemes utilize mixed-integer linear programming (MILP) for codon optimization, proposing formulas for two bi-objective problems; one optimizes one objective function, while the remaining objectives serve as constraints to create a Pareto optimal solution. In summary, the above solutions rely on heuristics and mathematical programming to solve the codon optimization problem. Existing methods only consider optimization on certain statistical metrics, neglecting the specific contextual information implicit in the species' genome, potentially leading to optimized sequences that are ill-suited to the host expression environment. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention provides a codon optimization method based on the BERT model, comprising:

[0005] The coding region sequences are extracted only from the reference genome of the host species. The dataset is constructed using the biological prior knowledge of codon bias in different species. The dataset is then randomly divided into training set, test set, and validation set.

[0006] An unsupervised learning network model is built based on the basic version of the BERT model. The network model is trained using the training set to learn the representation and distribution of the host species coding region context and obtain a codon optimization model.

[0007] The validation set sequence is optimized based on the codon optimization model, and the optimized sequence is verified by a biological wet experiment to complete the optimization.

[0008] Preferably, the process of constructing the dataset includes:

[0009] The coding region sequences are extracted from the reference genome of the host species, and the coding region sequences are filtered based on preset screening conditions;

[0010] The filtered sequences are further processed to limit the sequence length to no more than 1536, and the limited sequences are divided into training set, test set and validation set according to the proportion.

[0011] Preferably, the preset screening conditions include:

[0012] The sequence information does not contain the partial field and is a complete CDS sequence;

[0013] The value of the pseudo field in the sequence information is False;

[0014] CDS sequences begin with ATG;

[0015] The length of a CDS sequence is an integer multiple of 3.

[0016] Preferably, the parameters of the unsupervised learning network model include: architecture type BertForMaskedLM, hidden layer dimension of 768, number of attention heads of 12, number of hidden layers of 12, activation function of GELU, data type of float32, position embedding type of absolute position embedding, and layer normalization parameter of 1e-12.

[0017] The relevant parameters for the training process were set as follows: 200 training rounds, 32 batch size, 512 maximum sequence length, and cross-entropy loss function.

[0018] Based on the cross-entropy loss function, a strategy of adding intra-class distance of synonymous codons of the same amino acid and inter-class distance of synonymous codons of different amino acids as additional terms to the loss function.

[0019] Preferably, the expression for calculating the intra-class distance of synonymous codons of the same amino acid is:

[0020]

[0021] Among them, Cluster(x) k ) represents the cluster centers represented by different codon embeddings of the same amino acid, c i Let d represent the representation of the i-th synonymous codon of the k-th amino acid in the hidden space. in d(x) represents the mean intraclass distance of 20 different amino acids. k ) represents the intra-class distance of the synonymous codons of the k-th amino acid, x k N represents the k-th amino acid. k This represents the total number of synonymous codons for the k-th amino acid.

[0022] Preferably, the expression for calculating the interclass distance of synonymous codons of different amino acids is:

[0023]

[0024]

[0025] Among them, d x1x2d represents the L2 norm of two distinct amino acid cluster centers. out The distance between classes of different amino acids is represented by m, where m represents the number of pairwise combinations of the 20 amino acids.

[0026] Preferably, the process of optimizing the validation set sequence based on the trained model includes:

[0027] MiniGFP, a fluorescent protein smaller than traditional GFP protein, was selected as the research object. The codon sequence was optimized by a codon optimization model trained by deep learning, and the corresponding miniGFP gene fragment was obtained by gene synthesis. The miniGFP gene fragment was used to verify the effectiveness of the codon optimization model.

[0028] The pPIC9K-mCherry plasmid was constructed by using the weak constitutive promoter YPT1 and terminator AOX1 of Pichia pastoris to construct an mCherry expression cassette. Using pPIC9K as the original plasmid, the plasmid was recovered by gel digestion with NdeI and homologously recombined with the YPT1 fragment, the mCherry fragment and the AOX1 terminator fragment to obtain the pPIC9K-mCherry plasmid, which was then sequenced.

[0029] Preferably, the step of optimizing and validating the validation set sequence using the trained model further includes:

[0030] The pPIC9K-mCherry-miniGFPs expression plasmid was constructed, including PCR amplification of the miniGFP gene fragment with homologous arms by designing primers, gel recovery of the pPIC9K-mCherry plasmid by double digestion with BamHI and EcoRI, and homologous recombination with the miniGFP gene fragment with homologous arms to obtain the pPIC9K-mCherry-miniGFPs expression plasmid, which was then sequenced.

[0031] The miniGFP yeast strain was constructed, including the preparation of Pichia pastoris GS115 competent cells and electroporation integration of the pPIC9K-mCherry-miniGFPs expression plasmid.

[0032] The strain culture and induced expression included adding 1 mL of MGY medium to a 48-well plate, picking out single clones and inoculating them into the medium, culturing at 30°C and 800 rpm for 24 h using a real-time microbial growth analysis system, centrifuging to discard all supernatant, adding 1 mL of MGY medium, and inducing at 28°C and 800 rpm for 24 h.

[0033] Flow cytometry fluorescence detection included taking the induced bacterial culture, centrifuging to discard all supernatant, washing three times with phosphate buffer, resuspending, diluting to an OD600 of approximately 1.0, filtering the bacterial culture through a 70 μm cell sieve, and detecting cell fluorescence values ​​using a flow cytometer. The fluorescence values ​​of miniGFP and mCherry were detected using the FITC channel and PerCP channel, respectively.

[0034] On the other hand, the present invention also provides a codon optimization system based on the BERT model, comprising:

[0035] The dataset construction module is used to extract coding region sequences from only the reference genome of the host species, utilize the biological prior knowledge of codon bias in different species, construct the dataset, and randomly divide the dataset into training set, test set, and validation set;

[0036] The model building module is used to build an unsupervised learning network model based on the basic version of the BERT model. The network model is trained using the training set to learn the representation and distribution of the host species coding region context and obtain the codon optimization model.

[0037] The optimization module is used to optimize the validation set sequence based on the codon optimization model. The optimized sequence is then validated using a flow cytometry fluorescence detection method in a biological wet experiment to complete the optimization.

[0038] Compared with the prior art, the present invention has the following advantages and technical effects:

[0039] This invention presents a novel codon optimization method that uses a deep learning model to capture evolutionary information from CDS sequences, optimizing codon sequences to improve expression levels in heterologous proteins. This invention not only optimizes CDS sequences for specific species but also studies codon usage preferences across multiple species, improving codon expression levels in heterologous proteins and overcoming the genetic instability caused by directly replacing rare codons with the most commonly used codons in the target species. This invention is expected to improve expression levels in 80% of experimental groups before and after codon optimization, with some groups showing improvements of more than 2-fold. Attached Figure Description

[0040] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings:

[0041] Figure 1 This is the UMAP diagram of the BERT model after codon embedding in an embodiment of the present invention;

[0042] Figure 2 This is a UMAP diagram showing the codon embedding and the classification of amino acids according to four different properties in an embodiment of the present invention.

[0043] Figure 3 The figure shows the experimental results of an embodiment of the present invention. Detailed Implementation

[0044] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0045] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0046] Example 1

[0047] like Figure 1-2 As shown, this embodiment provides a codon optimization method based on the BERT model, including:

[0048] The coding sequence (CDS) is extracted only from the reference genome of the host species. The dataset is constructed using the biological prior knowledge of codon bias in different species. The dataset is then randomly divided into training set, test set, and validation set.

[0049] An unsupervised learning network model is built based on the basic version of the BERT model. The network model is trained using the training set to learn the representation and index of the host species coding region context, and a codon optimization model is obtained.

[0050] The validation set sequence is optimized based on the codon optimization model. The sequence to be optimized is input into the model to obtain the optimized sequence, which is then verified in a biological wet experiment to complete the optimization and verification.

[0051] 1. Dataset Construction: The selected species was Pichia pastoris. The CDS sequences collected from the NCBI database were screened and constructed. The sequences were randomly divided according to a certain ratio to obtain the training set, test set, and validation set.

[0052] 2. Model network design and training: The network model is built based on the basic version of the BERT model, and additional loss function terms are added for training.

[0053] 3. Expression of heterologous proteins from codon sequences: The model trained in step 2 was used to optimize the validation set sequences, and biowet experiments were conducted for verification.

[0054] Furthermore, the process of step 1 is as follows:

[0055] 1.1 The filtering criteria are as follows:

[0056] The sequence information does not include the partial field; it must be a complete CDS sequence.

[0057] The pseudo field of the sequence information has a value of False, indicating that it cannot be a pseudogene;

[0058] CDS sequences must begin with ATG;

[0059] The length of the CDS sequence must be an integer multiple of 3.

[0060] 1.2 Preprocessing of selected sequences: In order to conform to the input format of the network architecture, the sequences selected in step 1.1 are further processed, and the length of the sequence is limited to no more than 1536. The resulting sequences are divided into training set, test set and validation set according to the ratio.

[0061] Furthermore, the process of step 2 is as follows:

[0062] 2.1 The hyperparameter design of the BERT model references the standard configuration of BERT-base (768 hidden layer dimensions, 12 attention heads). This architecture has been fully validated in natural language processing tasks for its effectiveness in capturing long-range sequence dependencies. Given the isomorphism between genome sequences and natural language at the sequence modeling level—both are composed of linearly arranged discrete units (codons / vocabularies) forming complex semantic patterns—adopting the same parameter dimensions can inherit the technical advantages of the Transformer architecture in extracting local sequence features and modeling global associations, ensuring that the model effectively learns the multi-level patterns of codon usage preferences of the host species.

[0063] The main parameter settings are shown in Table 1:

[0064] Table 1

[0065]

[0066] 2.2 Hyperparameter design for model training: The relevant parameter settings during the training process are shown in Table 2.

[0067] Table 2

[0068]

[0069] 2.3 Model Loss Function Design: Based on the original cross-entropy loss function, a strategy was designed to include intra-class distance for synonymous codons of the same amino acid and inter-class distance for synonymous codons of different amino acids as additional terms in the loss function during model training. The specific calculation process is as follows:

[0070] 2.3.1 Calculate the cluster centers represented by different codon embeddings of the same amino acid. The calculation process is as follows:

[0071]

[0072] Where, x k Let c represent the k-th amino acid. i N represents the representation of the i-th synonymous codon of the k-th amino acid in the latent space. k v(c) represents the total number of synonymous codons for the k-th amino acid. i ) represents c i The feature vectors of the embedding space, Cluster(x) k ) represents the cluster center of the synonym codons of the k-th amino acid in the embedding space.

[0073] 2.3.2 Calculate the intra-class distance represented by different codon embeddings of the same amino acid. The calculation process is as follows:

[0074]

[0075] Where d(x) k ) represents the intra-class distance of the synonymous codons of the k-th amino acid, d in This represents the mean intraclass distance for 20 different amino acids.

[0076] 2.3.3 Calculate the interclass distance of synonymous codons for different amino acids. The specific calculation process is as follows:

[0077]

[0078] in, d represents the L2 norm of two distinct amino acid cluster centers. out denoted by , where m represents the mean interclass distance of 20 different amino acids, and m represents the number of pairwise combinations of the 20 amino acids.

[0079] 2.3.4. Through the above three steps, the loss function can be defined, and its calculation process is as follows:

[0080] loss2 = d in / d out (6)

[0081] loss=w1*loss1+w2*loss2(7)

[0082] Where loss1 is the original cross-entropy loss function, loss2 is a user-defined codon classification loss function, and w1 and w2 are the weights of the two loss functions, which are also hyperparameters, and their specific values ​​are determined during the training process. After the model is trained, the optimized sequence can be obtained by inputting the nucleic acid sequence to be optimized.

[0083] Furthermore, the process of step 3 is as follows:

[0084] 3.1 Obtaining the miniGFP sequence: After reviewing relevant literature, miniGFP, a fluorescent protein smaller than traditional GFP protein, was selected as the research object. The sequence of the miniGFP protein is as follows: MEKSFVITDPWLPDYPIISASDGFLELTEYSRDEIMGRNARFLQGPETDQATVQ KIRDAIRDRRPTTVQLINYTKSGKKFWNLLHLQPVFDGKGGLQYFIGVQLVGSDHV. The codon sequence was optimized using a codon optimization model trained by deep learning, and the corresponding miniGFP gene fragment was obtained through gene synthesis for constructing the pPIC9K-mCherry-miniGFPs expression plasmid.

[0085] 3.2 Construction of the miniGFP expression plasmid, the process is as follows:

[0086] 3.2.1 Construction of pPIC9K-mCherry plasmid: To facilitate subsequent detection and normalized characterization, a red fluorescent protein, mCherry, was introduced as an internal control. An mCherry expression cassette was constructed using the weak constitutive promoter YPT1 and terminator AOX1 of Pichia pastoris. Using pPIC9K as the original plasmid, after gel recovery via NdeI digestion, homologous recombination was performed with the YPT1 fragment, the mCherry fragment, and the AOX1 terminator fragment to obtain the pPIC9K-mCherry plasmid. Sanger sequencing was used to confirm its construction.

[0087] 3.2.2 Construction of pPIC9K-mCherry-miniGFPs expression plasmid: The miniGFP gene fragment with homologous arms was obtained by PCR amplification using designed primers. After double digestion with BamHI and EcoRI, the pPIC9K-mCherry plasmid was recovered from the gel and homologously recombinated with the miniGFP gene fragment with homologous arms to obtain the pPIC9K-mCherry-miniGFPs expression plasmid. The construction was confirmed by Sanger sequencing.

[0088] 3.3 Construction of the miniGFP yeast strain, the process is as follows:

[0089] 3.3.1 Preparation of Pichia pastoris GS115 competent cells: Streak cryovials onto plates, pick single colonies and inoculate them into YPD medium. Inoculate overnight cultured cells into a larger volume of YPD medium at a certain ratio, and culture with shaking until the OD600 value reaches 1.0–1.2. Centrifuge the culture at 4°C and collect the precipitated cells. Wash the cells several times (usually 3 times) with pre-cooled sterile water, centrifuging after each wash to collect the cells. Finally, wash the cells once with pre-cooled 1 mol / L sorbitol solution and resuspend the cells in a certain volume of pre-cooled sorbitol.

[0090] 3.3.2 Electroporation Integration: The pPIC9K-mCherry-miniGFPs expression plasmid was linearized by SacI digestion, purified using a DNA purification kit, and then electroporated into GS115 competent cells. The cells were plated on MD plates and incubated at 30°C for 2–3 days. Positive single clones were screened by colony PCR using universal primers 5AOX / 3AOX.

[0091] 3.4 Fluorescence detection, the process of which is as follows:

[0092] 3.4.1. Culture and Induction of Strains: Add 1 mL of MGY medium to a 48-well plate, pick out single clones and inoculate them into the medium, and culture at 30℃ and 800 rpm for 24 h using a real-time microbial growth analysis system. Centrifuge and discard all supernatant, add 1 mL of LMMY medium, and induce at 28℃ and 800 rpm for 24 h.

[0093] 3.4.2 Flow Cytometry Detection: After induction, the bacterial culture was centrifuged and all supernatant was discarded. The culture was washed three times with phosphate buffer and resuspended, then diluted to an OD600 of approximately 1.0. The culture was then filtered through a 70 μm cell sieve. Cell fluorescence values ​​were detected using a flow cytometer, with miniGFP and mCherry fluorescence values ​​detected via the FITC and PerCP channels, respectively.

[0094] Experimental results are as follows Figure 3 As shown, in the 12 groups of experiments, the comparative analysis of expression levels before and after codon optimization revealed that "91.7% of the groups showed improvement, and a few groups showed improvement of more than 2 times".

[0095] On the other hand, this embodiment also provides a codon optimization system based on the BERT model, including:

[0096] The dataset construction module is used to extract coding region sequences from only the reference genome of the host species, utilize the biological prior knowledge of codon bias in different species, construct the dataset, and randomly divide the dataset into training set, test set, and validation set;

[0097] The model building module is used to build an unsupervised learning network model based on the basic version of the BERT model. The network model is trained using the training set to learn the representation and distribution of the host species coding region context and obtain the codon optimization model.

[0098] The optimization module is used to optimize the validation set sequence based on the codon optimization model and perform biological wet experiment verification to complete the optimization and verification.

[0099] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A codon optimization method based on the BERT model, characterized in that, include: The coding region sequences are extracted only from the reference genome of the host species. The dataset is constructed using the biological prior knowledge of codon bias in different species. The dataset is then randomly divided into training set, test set, and validation set. An unsupervised learning network model is built based on a basic version of the BERT model. The network model is trained using the training set to learn the representation and distribution of the host species coding region context and obtain a codon optimization model. The validation set sequence is optimized based on the codon optimization model, and the optimized sequence is validated in biological experiments to complete the optimization and validation. The parameters of the unsupervised learning network model include: architecture type BertForMaskedLM, hidden layer dimension of 768, number of attention heads of 12, number of hidden layers of 12, activation function of GELU, data type of float32, position embedding type of absolute position embedding, and layer normalization parameter of 1e. -12 ; The relevant parameters for the training process were set as follows: 200 training rounds, 32 batch size, 512 maximum sequence length, and cross-entropy loss function. Based on the cross-entropy loss function, a strategy of adding intra-class distance of synonymous codons of the same amino acid and inter-class distance of synonymous codons of different amino acids as additional terms to the loss function; The expression for calculating the intra-class distance of synonymous codons of the same amino acid is: ； ； in, The cluster centers represented by different codon embeddings of the same amino acid, c i Let represent the representation of the i-th synonymous codon of the k-th amino acid in the latent space. This represents the mean intraclass distance for 20 different amino acids. This represents the intra-class distance of the synonym codons of the k-th amino acid. This represents the k-th amino acid. This represents the total number of synonymous codons for the k-th amino acid. The expression for calculating the interclass distance of synonymous codons of different amino acids is as follows: ； ， ； Among them, d x1x2 d represents the L2 norm of two distinct amino acid cluster centers. out The distance between classes of different amino acids is represented by m, where m represents the number of pairwise combinations of the 20 amino acids.

2. The method according to claim 1, characterized in that, The process of constructing the dataset includes: The coding region sequences are extracted from the reference genome of the host species, and the coding region sequences are filtered based on preset screening conditions; The filtered sequences are further processed by dividing the restricted sequences into training, testing, and validation sets according to a certain ratio.

3. The method according to claim 2, characterized in that, The preset filtering conditions include: The sequence information does not contain the partial field and is a complete CDS sequence; The value of the pseudo field in the sequence information is False; CDS sequences begin with ATG; The length of a CDS sequence is an integer multiple of 3.

4. The method according to claim 1, characterized in that, The process of optimizing the validation set sequence based on the codon optimization model includes: MiniGFP, a fluorescent protein smaller than traditional GFP protein, was selected as the research object. The codon sequence was optimized by a codon optimization model trained by deep learning, and the corresponding miniGFP gene fragment was obtained by gene synthesis. The miniGFP gene fragment was used to verify the effectiveness of the codon optimization model. Constructing the pPIC9K-mCherry plasmid, including building an mCherry expression cassette using the weakly constitutive promoter YPT1 and terminator AOX1 from Pichia pastoris, and using pPIC9K as the original plasmid. Nde After being digested with enzyme I and recovered from the gel, the plasmid was homologously recombinated with the YPT1 fragment, the mCherry fragment, and the AOX1 terminator fragment to obtain the pPIC9K-mCherry plasmid, which was then sequenced.

5. The method according to claim 1, characterized in that, The steps for optimizing the validation set sequence based on the codon optimization model also include: Constructing the pPIC9K-mCherry-miniGFPs expression plasmid includes designing primers for PCR amplification to obtain the miniGFP gene fragment with homologous arms. The pPIC9K-mCherry plasmid uses... Bam HI and Eco After RI double digestion and gel recovery, the plasmid was homologously recombinated with the miniGFP gene fragment with homologous arms to obtain the pPIC9K-mCherry-miniGFPs expression plasmid, which was then sequenced. The miniGFP yeast strain was constructed, including the preparation of Pichia pastoris GS115 competent cells and electroporation integration of the pPIC9K-mCherry-miniGFPs expression plasmid.

6. The method according to claim 1, characterized in that, The process of validating the optimized sequence in biological experiments includes: The strain culture and induced expression included adding 1 mL of MGY medium to a 48-well plate, picking out single clones and inoculating them into the medium, culturing at 30 ℃ and 800 rpm for 24 h using a real-time microbial growth analysis system, centrifuging to discard all supernatant, adding 1 mL of MMY medium, and inducing at 28 ℃ and 800 rpm for 24 h. Flow cytometry fluorescence detection included taking the induced bacterial culture, centrifuging to discard all supernatant, washing three times with phosphate buffer, resuspending, diluting to an OD600 of approximately 1.0, filtering the bacterial culture through a 70 μm cell sieve, and detecting cell fluorescence values ​​using a flow cytometer. The fluorescence values ​​of miniGFP and mCherry were detected using the FITC channel and PerCP channel, respectively.

7. A codon optimization system based on the BERT model, used to implement the codon optimization method based on the BERT model as described in any one of claims 1-6, characterized in that, include: The dataset construction module is used to extract coding region sequences from only the reference genome of the host species, utilize the biological prior knowledge of codon bias in different species, construct the dataset, and randomly divide the dataset into training set, test set, and validation set; The model building module is used to build an unsupervised learning network model based on the basic version of the BERT model. The network model is trained using the training set to learn the representation and distribution of the host species coding region context and obtain the codon optimization model. The optimization module is used to optimize the validation set sequence based on the codon optimization model, and then perform biological wet experiment verification on the optimized sequence to complete the optimization and verification.

Images