CN116978462A - Method for generating non-natural promoter based on diffusion model - Google Patents

Abstract

The invention discloses a method for generating a non-natural promoter based on a diffusion model, and belongs to the field of biological information technology. In order to generate promoters, a deep learning network based on the diffusion model was established. At the same time, this application conducted authenticity judgment and functional interval analysis on the generated promoters. The results showed that more than 40% of the generated promoters were true promoters, and the sequences had significant ‑35 and ‑10 functional regions, and had relatively high High credibility.

Description

A method to generate non-natural promoters based on diffusion model

Technical field

The invention relates to a method for generating a non-natural promoter based on a diffusion model, and belongs to the field of biological information technology.

Background technique

Promoter design can assist in building metabolic engineering networks for de novo synthesis of chemicals, drugs, and other raw materials in microorganisms. The main function of the promoter is to initiate gene transcription and translation, which has an intuitive impact on the expression level of the target gene. The latest research shows that the Pearson correlation coefficient between the transcription amount and translation amount of a promoter-initiated gene is as high as 0.8. Therefore, regulating promoters can achieve precise regulation of protein expression. In previous studies, researchers tried different methods to mine non-natural promoters, including directed evolution schemes, randomly introducing mutation sites into the target promoter, and rational design, which only targeted the conserved regions of the promoter or non-natural promoters. A small segment of the conserved region is mutated.

Although certain progress has been made in screening non-natural promoters at this stage, the constructed promoter library is still small. Normally, a promoter is 50 bases long and has 450 composition methods, which is difficult to verify using only experimental methods. The length of eukaryotic promoters is much longer than 50 bases, making experimental screening more difficult. Therefore, it is extremely important to develop methods for computationally assisted promoter generation that will facilitate promoter screening.

In 2020, Wang et al. proposed using an adversarial generative network to design promoters from scratch, converting promoter genes into one-dimensional arrays for learning, and then through the self-game of the generator and the discriminator to generate genes located in a similar distribution to natural biomolecules. New artificial molecular sequences enable de novo design of promoters. However, the training of the adversarial generative network is very unstable due to the conflict between training the optimal discriminator and the minimized generator, and the diversity of promoters generated by adversarial learning also has certain limitations, so it is not easy. Extension to modeling complex multimodal distributions. Based on the above reasons, it is necessary to study a new method for generating non-natural promoters.

Contents of the invention

In order to solve the instability problem that currently exists in the de novo design of promoters using adversarial generation networks, the present invention provides a method for generating non-natural promoters based on a diffusion model. The method includes:

Step S1: Construct a diffusion model for generating non-natural promoters. The diffusion model for generating non-natural promoters relies on UNet in the convolutional neural network. When building the coding region of UNet, a convolutional neural network is used. ; The non-coding area uses upsampling to restore image size; a normalized UNet jump connection is used between the coding area and the non-coding area for feature transfer, and a self-attention mechanism is introduced in both the coding area and the decoding area;

Step S2: Use the promoters in the public data set as training data to train the diffusion model for generating non-natural promoters;

Step S3: Generate a new promoter using the diffusion model trained to generate non-natural promoters.

Optionally, the step S2 includes:

Digitize the gene sequences of promoters from public datasets;

The diffusion model used to generate non-natural promoters is trained using the gene sequence of the digitally processed promoter, the loss value is calculated during the training process, promoter identification and conservation evaluation are performed on the output sample, and the result after training is saved model parameters;

Promoter identification uses the PromoR module based on deep learning to distinguish whether each generated sequence is true or false, and calculates the proportion of true promoters to all generated promoters;

Promoter conservation is evaluated by comparing the sequences of the generated promoters and observing the sequences of the -35 and -10 regions. When the identifiers of the -35 and -10 regions are TT and TATAAT, the generated promoter is considered to have the characteristics of a natural promoter. feature.

Optionally, when observing the -35 and -10 region sequences using the method, the tool MetaLogo is used.

Optionally, the digital processing of the gene sequence of the promoter in the public data set includes:

The one-hot encoding method is used for feature extraction, and a sequence of 50 bases in length is converted into a vector with a channel number of 1, a length of 4, and a width of 50;

After conversion, the bases A, T, C, and G are respectively: [1 0 0 0], [0 0 0 1], [0 1 0 0], [0 0 1 0].

Optionally, the method further includes setting a threshold value for the ratio of true promoters to all generated promoters.

This application also provides the application of the above-mentioned method of generating non-natural promoters based on the diffusion model in chemicals and pharmaceuticals.

The beneficial effects of the present invention are:

The genetic sequence is digitized by computer coding, and public data sets are collected to construct a training data set. Taking digital genes as input and using a diffusion model to learn their characteristics, the quality of the generated samples is evaluated and used to generate non-natural promoters. Compared with the existing technology of designing promoters through adversarial generative networks, the generative model training used in the present invention is more efficient It is stable and can effectively identify the key regions of small segments of the sequence, and can also identify the correlation between the small region and the full-length sequence. And it does not have the disadvantage of unstable training. At the same time, the diffusion model is more conducive to the stable generation of promoters with higher diversity and the mining of new promoters.

Description of the drawings

In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings needed to be used in the description of the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can also be obtained based on these drawings without exerting creative efforts.

Figure 1 is a flow chart for generating a promoter based on a diffusion model provided in one embodiment of the present invention.

Figure 2 is a sequence identification diagram of a promoter generated by the present invention using a diffusion model.

Detailed ways

In order to make the purpose, technical solutions and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

PyTorch: It is the Python version of torch. It is a neural network framework open sourced by Facebook and is specifically designed for GPU-accelerated deep neural network (DNN) programming. Torch is a classic tensor that operates on multi-dimensional matrix data. A tensor is a digital container in a machine learning program. It is essentially an array of various dimensions. The dimensions of the tensor are usually called axes. The number is called an order) library and is widely used in machine learning and other mathematics-intensive applications. Unlike Tensorflow's static calculation graph, pytorch's calculation graph is dynamic and can be changed in real time according to computing needs.

Example 1:

This embodiment provides a method for generating a non-natural promoter based on a diffusion model. The method performs diffusion model training and generates a new promoter based on the promoter gene sequence. See Figure 1. The method includes:

Step S1: Standardization of the input gene sequence.

Construct a training set: Use the promoter in the data set reported by Thomason (E. coli promoter) as training data. The training set contains a total of 11884 promoters;

Step S2: Digital processing of the gene sequence of the input promoter.

The one-hot encoding method is used for feature extraction, and a sequence of 50 bases in length is converted into a vector with a channel number of 1, a length of 4, and a width of 50. After conversion, the bases A, T, C, and G are respectively: [1 0 0 0], [0 0 0 1], [0 1 0 0], and [0 0 10].

Through this step, the gene sequence can be digitized, such as the promoter:

CCGCTCAAATATTGTTAAATTGCCGGTTTTGTATCAACTACTCACCCGGG is converted to: [[0 1 00][01 0 0][0 0 1 0][0 1 0 0][0 0 0 1][0 1 0 0][1 0 0 0][1 0 0 0 ][1 0 0 0][0 0 01][1 0 0 0][0 0 0 1][0 0 0 1][0 0 1 0][0 0 0 1][0 0 0 1][1 0 0 0][1 0 0 0][10 0 0][0 0 0 1][0 0 0 1][0 0 1 0][0 1 0 0][0 1 0 0][0 0 1 0 ][0 0 1 0][0 0 01][0 0 0 1][0 0 0 1][0 0 0 1][0 0 1 0][0 0 0 1][1 0 0 0][0 0 0 1][0 1 0 0][10 0 0][1 0 0 0][0 1 0 0][0 0 0 1][1 0 0 0][0 1 0 0][0 0 0 1 ][0 1 0 0][1 0 00][0 1 00][0 1 0 0][0 1 0 0][0 0 1 0][0 0 1 0][0 0 1 0]].

Step S3: Construct a diffusion model for sequence feature learning, specifically including:

S3.1: Convert data into tensors recognized by PyTorch;

Mainly convert Numpy array to tensor, please refer to

Convert using the conversion method introduced in https://blog.csdn.net/weixin_43728604/article/details/102679016.

S3.2: Build a deep learning network based on PyTorch. The network construction mainly relies on UNet in the convolutional neural network. When building the coding area of UNet, the convolutional neural network is used; the non-coding area uses upsampling to restore the image size. . The normalized UNet skip connection is used between the coding area and the non-coding area for feature transfer, and a self-attention mechanism is introduced in both the coding area and the decoding area;

Diffusion model: Under the condition that the diffusion process is a Markov chain, Gaussian noise is continuously added to the original information. The process of adding Gaussian noise at each step is from X _t-1 → X _t ,

Among them, X _t represents the data at time t-1, and X _t-1 represents the data at time t after adding Gaussian noise.

The inverse diffusion process is to recover the original data from Gaussian noise. Assuming that the inverse diffusion process is still a Markov chain process, what needs to be done is X _T →X ₀ , where X _T represents the data at time T after adding Gaussian noise, X ₀ refers to the original data recovered from the data at time T after adding Gaussian noise.

UNet: It is a typical encoder and decoder structure. The encoder mainly performs feature extraction, and the image size is continuously reduced. The right side corresponds to the upsampling process, which restores the image to a size close to the original image through jump links with information from different convolutional layers.

Convolutional layer: Feature extraction of digital promoters based on image information. The formula is as follows:

Activation layer: The activation layer uses the ReLU function, which can be understood as a piecewise linear function, changing all negative values to 0 while leaving the positive values unchanged.

Self-attention layer: The self-attention mechanism allows inputs to interact with each other and find out which objects they should pay more attention to. The output is the sum of these interactions and attention scores.

S3.3: Calculate the loss value during model training, perform promoter identification and conservative evaluation of the output samples.

Promoter identification uses the PromoR module (BioRxiv:doi:https://doi.org/10.1101/2023.03.05.531155) based on deep learning, that is, each generated sequence is distinguished between true and false, and the proportion of true promoters among all Proportion of generated promoters.

Promoter conservation mainly conducts sequence alignment of the generated promoters and observes the -35 and -10 region sequences, using the tool MetaLogo (http://metalogo.omicsnet.org/analysis). When the -35 and -10 regions are identified as TT and TATAAT, the generated promoter is considered to have the characteristics of a natural promoter.

The loss value is calculated using the mean squared error loss (MSELoss) to compare the output results with the real results, and the parameters are optimized through the AdamW optimizer.

S3.4: When the loss value decreases slowly during the training process, evaluate the generated samples and judge based on the promoter identification results. When the proportion of real promoters is relatively high and the sequence has a significant conservative interval, the training of the training generation Parameters are saved;

Step S4: Use the diffusion model to generate non-natural promoters based on the parameters saved after training.

In order to illustrate the superiority of the generative model constructed by the present invention, the present invention uses deep learning-based discrimination of true and false promoters and analysis of generated promoter sequence characteristics. The results are shown in Figure 2. According to Figure 2, it can be seen that the proportion of generated promoter identifications that are true exceeds 40%, and the promoter has significant -35 and -10 regions, indicating that the generated promoter has high reliability. reliability.

The method provided by this application to generate Tianre promoter based on the diffusion model realizes digitization by computer coding the gene sequence, uses the digitized gene as input and uses the diffusion model to learn its characteristics, evaluates the quality of the generated sample and uses it to generate non-natural promoter Compared with the existing technology of designing promoters through adversarial generative networks, the generative model training adopted in the present invention is more stable, can effectively identify key regions of small segments of sequences, and can simultaneously identify the association between small regions and full-length sequences. . And it does not have the disadvantage of unstable training. At the same time, the diffusion model is more conducive to the stable generation of promoters with higher diversity and the mining of new promoters.

Some steps in the embodiments of the present invention can be implemented using software, and corresponding software programs can be stored in readable storage media, such as optical disks or hard disks.

The above are only preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention shall be included in the protection of the present invention. within the range.

Claims (6)

1. A method for generating non-natural promoters based on a diffusion model, characterized in that the method includes: Step S1: Construct a diffusion model for generating non-natural promoters. The diffusion model for generating non-natural promoters relies on UNet in the convolutional neural network. When building the coding region of UNet, the convolutional neural network is used ; The non-coding area uses upsampling to restore image size; a normalized UNet jump connection is used between the coding area and the non-coding area for feature transfer, and a self-attention mechanism is introduced in both the coding area and the decoding area; Step S2: Use the promoters in the public data set as training data to train the diffusion model for generating non-natural promoters; Step S3: Generate a new promoter using the diffusion model trained to generate non-natural promoters. 2. The method according to claim 1, characterized in that step S2 includes: Digitize the gene sequences of promoters from public datasets; The diffusion model used to generate non-natural promoters is trained using the gene sequence of the digitally processed promoter, the loss value is calculated during the training process, promoter identification and conservation evaluation are performed on the output sample, and the result after training is saved model parameters; Promoter identification uses the PromoR module based on deep learning to distinguish whether each generated sequence is true or false, and calculates the proportion of true promoters to all generated promoters; Promoter conservation is evaluated by comparing the sequences of the generated promoters and observing the sequences of the -35 and -10 regions. When the identifiers of the -35 and -10 regions are TT and TATAAT, the generated promoter is considered to have the characteristics of a natural promoter. feature. 3. The method according to claim 2, characterized in that when observing the -35 and -10 region sequences, the method uses MetaLogo. 4. The method according to claim 3, characterized in that said digital processing of the gene sequence of the promoter in the public data set includes: The one-hot encoding method is used for feature extraction, and a sequence of 50 bases in length is converted into a vector with a channel number of 1, a length of 4, and a width of 50; After conversion, the bases A, T, C, and G are respectively: [1 0 0 0], [0 0 0 1], [0 1 0 0], [0 0 1 0]. 5. The method according to claim 4, further comprising setting a threshold value for the ratio of true promoters to all generated promoters. 6. Application of the method of generating non-natural promoters based on the diffusion model according to any one of claims 1 to 5 in chemicals and pharmaceuticals.