Endoglucanase de novo design method based on generative adversarial network
By generating an adversarial network design, the problems of complexity and inefficiency of protein design are solved, and the rapid generation of efficient and stable enzyme sequences are achieved, which shortens the experimental cycle and improves the efficiency of enzyme preparation development.
Patent Information
- Application Number
- CN202411066552.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-05
- Publication Date
- 2024-11-22
AI Technical Summary
Existing protein design methods are complex and inefficient, making it difficult to design enzyme sequences with expected structure and functions in a short time.
Using the de novoglucanase design method based on the generative adversarial network, the loss of generator and discriminator is optimized by constructing a neural network model containing generator and discriminator, using self-attention module and residual block, and the loss of generator and discriminator is optimized to generate Enzyme sequences with the desired properties and screen for enzyme preparations by multi-sequence alignment, clustering and three-dimensional modeling.
The number of iterations of experimental verification has been greatly shortened, the stability and catalytic efficiency of enzymes have been improved, the cost and time of experiments have been saved, the enzyme sequence space has been expanded, and the development efficiency of industrial enzyme preparations has been improved.
Smart Images

Figure CN119007818A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of biotechnology, in particular to the technical fields related to enzyme engineering and protein design. Specifically, the present invention relates to a method for de novo design of endoglucanase based on generative adversarial networks, which is used to improve the activity, stability and specificity of endoglucanase. Background Art
[0002] Protein design is a key technology in the field of modern biotechnology, and its main purpose is to create proteins with new structures and functions. With the increasing demand for human disease treatment and the urgency of environmental pollution problems, it is particularly important to efficiently and accurately design proteins with specific functions.
[0003] Traditional protein design methods mainly rely on a large number of random mutations or directed evolution of a single amino acid sequence, or use the recombination of natural homologous proteins to screen for the best performing protein. The number of protein sequence combinations increases exponentially with the increase in sequence length. For example, the number of combinations of a protein sequence containing 100 amino acids can reach a maximum of 20. 100 , so protein design and mutation analysis are highly complex. Faced with such a rich and diverse sequence combination, the designed protein is often difficult to obtain the expected, stable structure and function because the generation, screening and verification of mutants are complex and time-consuming in a wet laboratory environment. The new de novo design method for endoglucanase based on generative adversarial networks can design a large number of enzyme sequences in a very short time, and then screen out effective samples for wet experiment verification, greatly improving the efficiency of de novo enzyme design. Summary of the invention
[0004] Technical problem to be solved by the invention
[0005] In view of the problems of complexity and low efficiency of existing protein design methods, the present invention proposes a novel de novo design method for endoglucanase based on generative adversarial networks.
[0006] Technical Solution
[0007] In order to achieve the above object, the technical solution provided by the present invention is:
[0008] The method for de novo design of endoglucanase based on generative adversarial networks comprises the following steps:
[0009] Step 1: Build a generative adversarial network model. The network model consists of two neural network generators G and discriminators D. The generator and discriminator include several interconnected residual blocks and self-attention modules.
[0010] Step 2, constructing a training data set based on the real gene sequence of endoglucanase, padding the shorter gene sequences in the training set with zeros to make all sequences of equal length, and encoding the gene sequences.
[0011] Step 3: Randomly generate a batch of noise vectors from a normal random distribution. The length of the noise vector is equal to the length of the gene sequence in the training set. Input the noise vector into the generator created in step 1.
[0012] Step 4: Extract a batch of real sequences from the training set encoded in step 2 and input them into the discriminator created in step 1, and output D out , input the noise sequence generated in step 3 into the discriminator created in step 1, and output GD out ;
[0013] Calculate D out The loss d with all 1 vector loss-real , calculate GD ou The loss d with all zero vectors loss-fake .
[0014] Step 5: Based on the result of step 4, loss-real ,d loss-fake Calculate the generator loss G loss and the discriminator loss D loss , the method is:
[0015] G loss =-d loss-fake
[0016] D loss =d loss-real +d loss-fake +R1 penalty
[0017] Where R1 penalty is the penalty parameter used to constrain the discriminator to the real data.
[0018] Step 6: Calculate the target optimization function V(D,G) based on the output of the generator and the discriminator, and optimize the generator and discriminator neural networks by:
[0019]
[0020] Where P r (x) represents the distribution of the real data x in step 2, P z (z) represents the distribution of random noise z in step 3, E represents the mathematical expectation, G(z) represents the output of random noise passing through the generator, D(x) represents the output of real data passing through the discriminator, and D(G(z)) represents the processing result of random noise after being output by the generator and then by the discriminator.
[0021] Step 7: Change the parameters of the generator and discriminator and iterate steps 3-6 until the generator loss G loss and the discriminator loss D loss When the two tend to be stable and the deviation between them is less than a certain threshold, the training process stops and the generator G and discriminator D neural network are obtained.
[0022] Step 8, using the generator G obtained in step 7 to generate a new sequence for preparing endoglucanase.
[0023] Furthermore, the self-attention module in the generator G constructed in step 1 is located before the last residual block, and the self-attention module in the discriminator D is located after the first residual block.
[0024] Furthermore, the method of encoding the gene sequence in step 2 is one-hot encoding.
[0025] Furthermore, the method for calculating the loss in step 4 is:
[0026] D out The loss is calculated using the mean square error of the all-1 vector, and the result is recorded as d loss-real :
[0027]
[0028] Where n is the vector dimension, y i is the vector D out The i-th element of ;
[0029] GD out The loss is calculated using the mean square error and the all-zero vector, and the result is recorded as d loss-fake :
[0030]
[0031] Where n is the vector dimension, y i is the vector GD out The i-th element of .
[0032] Further, the step of preparing endoglucanase in step 8 is:
[0033] Step 8-1, performing multiple sequence alignment between the sequence generated by the generator and the original sequence to generate an initial sequence database;
[0034] Step 8-2, clustering the initial sequence database, and finding the generated sequence of the same class as the original sequence according to the clustering result;
[0035] Step 8-3, using three-dimensional modeling and biological methods to screen and prepare endoglucanase from the generated sequence obtained in step 8-2.
[0036] Beneficial Effects
[0037] The present invention utilizes generative adversarial networks for de novo design, which can computationally predict enzyme structures and sequences with desired properties, reduce the number of iterations of experimental verification, significantly shorten the experimental cycle in the laboratory, and save costs and time compared to traditional methods;
[0038] The present invention utilizes an artificial intelligence-assisted de novo design strategy, breaks through the limitations of traditional directed evolution methods, greatly expands the sequence space of endoglucanases, and can improve the development efficiency of industrial enzyme preparations with high stability and high catalytic efficiency. BRIEF DESCRIPTION OF THE DRAWINGS
[0039] Figure 1 is a flow chart of the method of the present invention;
[0040] Figure 2 The method of the present invention generates a model diagram of a confrontation network;
[0041] Figure 3 It is a schematic diagram of the composition of the discriminator and the generator of the method of the present invention;
[0042] Figure 4 A clustering diagram of the sequences generated for the present invention and the true sequences. DETAILED DESCRIPTION
[0043] In order to further understand the content of the present invention, the present invention is described in detail in conjunction with the accompanying drawings and specific implementation methods.
[0044] like Figure 1 , Figure 2 As shown, the method for designing endoglucanase from scratch based on the generative adversarial network of the present invention comprises the following steps:
[0045] 1) Construct a generative adversarial network model suitable for amino acid sequences, learn hidden features from the collected samples, and generate diverse new samples similar to the original amino acid sequences. The generative adversarial network model of the present invention consists of two parts: a generator G and a discriminator D. The task of the generator is to generate realistic data from random noise to "cheat" the discriminator to the greatest extent, and the discriminator is responsible for distinguishing the real sequence from the generated sequence. The two continue to optimize until the discriminator cannot distinguish the authenticity of the data and the generator can achieve the best data distribution approximation through the adversarial struggle between the generated sequence and the real sequence.
[0046] The network model uses multiple residual blocks to connect the generator and the discriminator. The use of residual blocks helps to extract features of amino acid sequences and match similar distributions. Activation functions are used in residual blocks to alleviate the problem of gradient disappearance during training. Each residual block contains three 1-D convolutional layers with a filter size of 3. In order to prevent the problem of gradient disappearance during the training of the generator and the discriminator, the LeakyRelu function is selected as the activation function, and its function form is:
[0047]
[0048] In this embodiment, α=0.2 is set to better fit the network.
[0049] In order to allow the network to capture the role of different regions of the protein in the overall protein behavior, the model includes a self-attention module, which consists of many layers, and different layers are used to highlight the features of different important regions in the sequence. Figure 3 As shown, in this embodiment, the discriminator G is composed of 6 residual blocks, and a self-attention module is added after the first block structure; the generator D is composed of 6 residual blocks, and a self-attention module is added before the last residual block. The discriminator and generator models are batch normalized after being processed by the last residual block.
[0050] 2) construct a training data set and encode the real sequence in the training data set. In the present embodiment, the whole genome data of endoglucanase is first collected from the NCBI official website, and 79045 sequences are obtained after removing redundant data, and the data length range is 156-1222. Due to the wide distribution range of data length, but the number of sequences with a length of 220-630 amino acids exceeds half, in order to ensure the model training effect, the sequence of this part is selected as the training data set, a total of 50365 sequences.
[0051] For the shorter sequences in the data set, zeros are added at the end to make all sequences of equal length. Each amino acid is then mapped to a specific number (1-20, the sequence consists of 20 amino acids). The mapped sequence is then one-hot encoded and input into the generator network. The vector is converted into a matrix of size 512×21, where 512 is the length of the sequence and 21 represents 20 typical amino acids and the symbol for the zero-filled part.
[0052] 3) Extract a batch of 128-dimensional random noise vectors from a random distribution with a mean of 0 and a standard deviation of 0.5 and input them into the generator G to generate data G with a dimension of 512×21 out .
[0053] 4) Extract a batch of real sequences from the encoded data set in step 2) and input them into the discriminator D together with the noise sequence data processed by the generator in step 3) to output D out , GD out
[0054] D out The loss is calculated using the mean square error (MSE) and the all-1 vector, and the result is recorded as d loss-real :
[0055]
[0056] Where n is the vector dimension, y i is the vector D out The i-th element of .
[0057] GD out The loss is calculated using the mean square error (MSE) and the all-zero vector, and the result is recorded as d loss-fake :
[0058]
[0059] Where n is the vector dimension, y i is the vector GD out The i-th element of .
[0060] 5) Calculate the generator loss G loss and the discriminator loss D loss :
[0061] G loss =-d loss-fake
[0062] D loss =d loss-real +d loss-fake +R1 penalty
[0063] Where R1 penalty R1 penalty is a penalty parameter used to constrain the discriminator's response to real data. It is a training regularization method for generative adversarial networks, which aims to improve the stability and generation effect of the model. It is based on imposing constraints on the discriminator gradient of real data, thereby encouraging the smoothness of the discriminator near real data points. The ultimate goal is to make the loss of the generator close to that of the discriminator, ensuring that the two functions are equivalent so that the generated sequence distribution is similar to the real one. In this embodiment, the form of the R1 penalty parameter is:
[0064] R1 penalty =log(1+e x )+penalty×(10×0.5)
[0065] The mathematical definition of the R1 penalty is:
[0066]
[0067] Where γ is a hyperparameter that controls the strength of the penalty term, Represents the L2 norm of the gradient of the discriminator output with respect to the input x.
[0068] 6) Calculate the target optimization function V(D,G) and optimize the generator and discriminator neural networks by:
[0069]
[0070] Where P r (x) represents the distribution of the real data x, P z (z) represents the distribution of random noise z, and E represents expectation. G(z) represents the output of random noise through the generator, D(x) represents the output of real data through the discriminator, and D(G(z)) represents the result of random noise after being output by the generator and then processed by the discriminator.
[0071] 7) Change the parameters of the generator and discriminator and iterate steps 3)-6) until the generator loss G loss and the discriminator loss D loss When the two tend to be stable and the deviation between them is less than a certain threshold, the training process stops and the generator and discriminator neural networks are obtained.
[0072] In this embodiment, the training step ratio of the generator and the discriminator is selected to be 1:1 in actual operation. The generator and discriminator networks are optimized using the Adam optimizer. Some sequences are selected from the training set as the validation data set. After every 1200 iterations of the training process, the sequence generated by the discriminator is compared with the training set validation data set.
[0073] 8) Use the trained generator model to generate new sequences for preparing endoglucanase.
[0074] In this embodiment, the generated sequence is first preprocessed by combining the original sequence and the generated sequence for multiple sequence alignment to obtain an initial sequence database. In this embodiment, blastdb is used for alignment.
[0075] Then, cluster analysis is performed on the initial sequence database. The method is: cluster the initial sequence database using a clustering tool with a similarity of 80% as the boundary. In this embodiment, Mmseq2 is used for clustering. The clustering results are generated as follows: Figure 4 The cluster diagram shown in Figure 2 shows how to find the generated sequences that are classified into the same category as the original sequence.
[0076] Finally, three-dimensional modeling and biological methods were used to screen and prepare endoglucanase. The specific implementation steps in this embodiment were:
[0077] (1) The obtained sequences were modeled and analyzed using a three-dimensional modeling tool to obtain the predicted structural model and the global model quality estimation score of the generated sequences, and the generated sequences with a GMQE score lower than 0.9 were eliminated. In this embodiment, SwissModel was used for modeling.
[0078] (2) According to the generated sequence synthetase protein obtained in step (1), the synthetase protein is expressed in Escherichia coli, and is treated and obtained through various purification techniques such as affinity chromatography and ion exchange chromatography to obtain a high-purity enzyme product; finally, the activity of the enzyme is measured to evaluate its catalytic efficiency and functional characteristics.
[0079] The present invention and its embodiments are described schematically above, and the description is not restrictive. The drawings show only one embodiment of the present invention, and the actual structure is not limited thereto. Therefore, if a person skilled in the art is inspired by it and designs a structural method and an embodiment similar to the technical solution without creativity without departing from the purpose of the invention, they shall all fall within the protection scope of the present invention.
Claims
1. A method for de novo design of endoglucanase based on generative adversarial networks, characterized in that: The steps include: Step S1, constructing a generative adversarial network model, wherein the network model consists of two neural network generators G and a discriminator D, wherein the generator and the discriminator include a plurality of interconnected residual blocks and self-attention modules; Step S2, constructing a training data set based on the real gene sequence of endoglucanase, padding the shorter gene sequences in the training set with zeros to make all sequences of equal length, and encoding the gene sequences; Step S3, randomly generating a batch of noise vectors from a normal random distribution, wherein the length of the noise vector is equal to the length of the gene sequence in the training set, and inputting the noise vector into the generator G created in step S1; Step S4, extracting a batch of real sequences from the training set encoded in step S2 and inputting them into the discriminator D created in step S1, and outputting D out , input the noise sequence generated in step S3 into the discriminator created in step S1, and output GD out ; Calculate D out The loss d with all 1 vector loss-real , calculate GD ou The loss d with all zero vectors loss-fake ; Step S5, based on the d obtained in step S4 loss-real ,d loss-fake Calculate the generator loss G loss and the discriminator loss D loss , the method is: G loss =-d loss-fake , D loss =d loss-real +d loss-fake +R1 penalty , Where R1 penalty To constrain the penalty parameter of the discriminator on the real data; Step S6, based on the output of the generator and the discriminator, the target optimization function V(D,G) is calculated to optimize the generator and the discriminator neural network by: Where P r (x) represents the distribution of the real gene sequence x in step S2, P z (z) represents the distribution of the random noise z in step S3, E represents the mathematical expectation, G(z) represents the output of the random noise through the generator, D(x) represents the output of the real data through the discriminator, and D(G(z)) represents the processing result of the random noise after being output by the generator and then by the discriminator; Step S7, change the parameters of the generator and the discriminator and iterate the steps S3-S6 until the generator loss G loss and the discriminator loss D loss When the training process becomes stable and the deviation between the two is less than a certain threshold, the generator G and the discriminator D neural network are obtained. Step S8, using the generator G obtained in step S7 to generate a new sequence for preparing endoglucanase.
2. The method for de novo design of endoglucanase based on generative adversarial networks according to claim 1, characterized in that: The self-attention module in the generator G constructed in step S1 is located before the last residual block, and the self-attention module in the discriminator D is located after the first residual block.
3. The method for de novo design of endoglucanase based on generative adversarial networks according to claim 1, characterized in that: The method of encoding the gene sequence in step S2 is one-hot encoding.
4. The method for de novo design of endoglucanase based on generative adversarial networks according to claim 1, characterized in that: The method for calculating the loss in step S4 is: D out The loss is calculated using the mean square error of the all-1 vector, and the result is recorded as d loss-real : Where n is the vector dimension, y i is the vector D out The i-th element of ; GD out The loss is calculated using the mean square error and the all-zero vector, and the result is recorded as d loss-fake : Where n is the vector dimension, y i is the vector GD out The ith element of .
5. The method for de novo design of endoglucanase based on generative adversarial networks according to claim 1, characterized in that: The step of preparing endoglucanase in step S8 is: Step S8-1, performing a multiple sequence alignment between the sequence generated by the generator and the original sequence to generate an initial sequence database; Step S8-2, clustering the initial sequence database, and finding the generated sequence of the same class as the original sequence according to the clustering result; Step S8-3, using three-dimensional modeling and biological means to screen and prepare endoglucanase from the generated sequence obtained in step S8-2.
