Protein processing method, device, storage medium and computer program product

Abstract

This invention relates to the field of protein processing technology, and more particularly to a protein processing method, apparatus, storage medium, and computer program product. The method includes: acquiring a protein to be processed; generating at least one self-generated protein corresponding to the protein to be processed, wherein the protein to be processed and the self-generated protein are homologous proteins; storing at least one self-generated protein in a protein database, and performing structure prediction on the protein to be processed based on the protein to be processed and / or the self-generated protein in the protein database. This invention addresses the deficiency in existing technologies where the number of homologous proteins is small, leading to inaccurate structural prediction results for structural proteins, and improves the performance of protein structure prediction.

Description

Technical Field

[0001] This invention relates to the field of protein processing technology, and more particularly to a protein processing method, apparatus, storage medium, and computer program product. Background Technology

[0002] Homologous proteins are sets of protein sequences that evolved from a common ancestor in different organisms. Generally speaking, two protein sequences are considered homologous if their similarity exceeds 30%. Homologous protein sequences contain co-evolutionary information; that is, two amino acid sites that mutate together are usually in spatial contact, which is crucial for protein structure prediction. Currently, commonly used protein structure prediction models include algorithms such as AlphaFold2. However, existing protein structure prediction models require a large number of homologous proteins to complete the prediction, and their performance is poor when there are few or no homologous proteins. Furthermore, in the homologous protein retrieval stage before structure prediction, existing protein retrieval methods cannot solve the problem of retrieving homologous proteins with limited resources. When the number of homologous proteins is small, these sequences are insufficient to provide enough information for building a statistical model. Therefore, these retrieval algorithms quickly terminate the search process, failing to obtain enough homologous protein sequences, thus hindering the protein structure prediction process. Summary of the Invention

[0003] This invention provides a protein processing method, apparatus, storage medium, and computer program product to address the shortcomings of existing technologies where the number of homologous proteins is small, leading to inaccurate structural protein prediction results, and to improve protein structure prediction performance.

[0004] This invention provides a protein processing method, comprising: obtaining a protein to be processed; generating at least one self-generated protein corresponding to the protein to be processed, wherein the protein to be processed and the self-generated protein are homologous proteins; storing at least one self-generated protein in a protein database, and performing structural prediction on the protein to be processed based on the protein to be processed and / or the self-generated protein in the protein database.

[0005] According to a protein processing method provided by the present invention, the step of generating at least one self-generated protein corresponding to the protein to be processed includes: obtaining the sequence information of each amino acid site arrangement in the protein to be processed; obtaining the site arrangement probability corresponding to each amino acid site according to the sequence information to be processed; and generating at least one self-generated protein corresponding to the protein to be processed according to each site arrangement probability.

[0006] According to a protein processing method provided by the present invention, generating at least one self-generated protein corresponding to the protein to be processed based on the permutation probability of each of the said sites includes: inputting the permutation probability of each of the said sites into a preset sequence generation model to obtain at least one self-generated protein corresponding to the protein to be processed output by the sequence generation model; wherein, the sequence generation model is obtained by training an autoregressive model based on self-attention using a first sample homologous protein, the first sample homologous protein includes A sample proteins to be processed, and at least one first sample self-generated protein corresponding to each of the A sample proteins to be processed, wherein A is an integer greater than 1.

[0007] According to a protein processing method provided by the present invention, the step of generating at least one self-generated protein corresponding to the protein to be processed based on the permutation probability of each of the sites includes: obtaining a preset control factor; weighting the permutation probability of each of the sites using the control factor; and generating at least one self-generated protein corresponding to the protein to be processed based on the weighted permutation probability of each of the sites.

[0008] According to a protein processing method provided by the present invention, the step of generating at least one self-generated protein corresponding to the protein to be processed includes: inputting the protein to be processed into a preset latent variable processing model, wherein the latent variable processing model includes an encoder module, a prior network module, and a decoder module; the processing procedure of the latent variable processing model is as follows: inputting the protein to be processed into the encoder module to obtain the expected original latent variable distribution corresponding to the protein to be processed output by the encoder; inputting the expected original latent variable distribution into the prior network module to obtain the expected sampled latent variable distribution output by the prior network module; inputting the expected sampled latent variable distribution into the decoder module to obtain one self-generated protein output by the decoder module; obtaining the number of self-generated proteins corresponding to the protein to be processed; when it is determined that the number does not reach a preset number threshold, the self-generated protein is used as a new protein to be processed and re-input into the latent variable processing model to obtain a new self-generated protein re-output by the latent variable processing model.

[0009] According to a protein processing method provided by the present invention, the latent variable processing model is trained on the original latent variable processing model using a second sample homologous protein. The original latent variable processing model includes an original encoder module, an original prior network module, and an original decoder module. The second sample homologous protein includes B second proteins, each of which is homologous to the others, where B is an integer greater than 1. The training process of the original latent variable processing model is as follows: the B second proteins are sequentially input into the original encoder module to obtain the expected distributions of the predicted original latent variables of the B second proteins output by the original encoder module; the expected distributions of each predicted original latent variable are then sequentially processed... The input is given to the original prior network module to obtain the expected distribution of the predicted sampled latent variables output by the original prior network module; each expected distribution of the predicted sampled latent variables is sequentially input to the original decoder module to obtain the second predicted self-generated protein output by the original decoder module; information divergence is calculated based on the expected distribution of the original predicted latent variables and the expected distribution of the predicted sampled latent variables; the parameters of the original latent variable processing model are adjusted using the second predicted self-generated protein and the information divergence as supervision signals until both the second predicted self-generated protein and the information divergence meet the preset supervision conditions, at which point the original latent variable processing model is determined to be the latent variable processing model.

[0010] According to a protein processing method provided by the present invention, after generating at least one self-generated protein corresponding to the protein to be processed based on the protein to be processed, and before storing at least one self-generated protein in a protein database, the method further includes: screening each self-generated protein to obtain at least one target self-generated protein corresponding to the screened protein to be processed; storing at least one self-generated protein in a protein database includes: storing at least one target self-generated protein corresponding to the screened protein to be processed in the protein database.

[0011] According to a protein processing method provided by the present invention, the step of screening each self-generated protein to obtain at least one target self-generated protein corresponding to the screened protein to be processed includes: performing the following processing on each self-generated protein: obtaining a first sequence probability density of the self-generated protein; when the first sequence probability density is greater than a preset density threshold, determining the self-generated protein as the target self-generated protein.

[0012] According to a protein processing method provided by the present invention, the step of screening each self-generated protein to obtain at least one target self-generated protein corresponding to the screened protein to be processed includes: obtaining a first sequence vector of the protein to be processed; performing the following processing on each self-generated protein: obtaining a second sequence vector of the self-generated protein; and determining the self-generated protein as the target self-generated protein when the difference between the first sequence vector and the second sequence vector is less than a preset vector threshold.

[0013] According to a protein processing method provided by the present invention, the step of screening each self-generated protein to obtain at least one target self-generated protein corresponding to the screened protein to be processed includes: obtaining a quality reference value for each self-generated protein by dual sampling; sorting each self-generated protein in descending order of the quality reference values ​​to generate a protein sequence; and determining at least one target self-generated protein from the protein sequence according to a preset diversity selection method, wherein at least one homologous protein is not arranged in a continuous order in the protein sequence.

[0014] According to a protein processing method provided by the present invention, the step of screening each self-generated protein to obtain at least one target self-generated protein corresponding to the screened protein to be processed includes: obtaining a second sequence probability density of the protein to be processed; generating a target probability density ratio based on the second sequence probability density; and performing the following processing on each self-generated protein respectively: obtaining a first sequence probability density of the self-generated protein; obtaining the ratio of the first sequence probability density to the second sequence probability density; and determining the self-generated protein as the target self-generated protein when the ratio is greater than the target probability density ratio.

[0015] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of any of the protein processing methods described above.

[0016] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of any of the protein processing methods described above.

[0017] The present invention also provides a computer program product, comprising a computer program that, when executed by a processor, implements the steps of any of the protein processing methods described above.

[0018] The protein processing method, apparatus, storage medium, and computer program product provided by this invention, after obtaining the protein to be processed, generate at least one self-generated protein corresponding to the protein to be processed, wherein the protein to be processed and the self-generated protein are homologous proteins; store at least one self-generated protein in a protein database, and perform structure prediction on the protein to be processed based on the protein to be processed and / or the self-generated protein in the protein database. In the case of no homologous proteins or a small number of homologous proteins, the above process generates at least one self-generated protein corresponding to the protein to be processed, increasing the number of homologous proteins. This facilitates subsequent use of protein structure prediction models to predict the structure of homologous proteins, avoiding inaccurate structure predictions due to the absence or scarcity of homologous proteins, and improving the accuracy of protein structure prediction. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0020] Figure 1 This is a schematic flowchart of the protein processing method provided by the present invention;

[0021] Figure 2 This is a schematic diagram of the autoregressive model structure based on self-attention provided by the present invention;

[0022] Figure 3 This is a schematic diagram illustrating the principle of adding control factors to the sequence generation model provided by this invention;

[0023] Figure 4 This is a schematic diagram illustrating the principle of the latent variable processing model provided by this invention;

[0024] Figure 5 This is a schematic diagram of the training of the original latent variable processing model provided by the present invention;

[0025] Figure 6 This is a schematic diagram of the structure of the protein processing method apparatus provided by the present invention;

[0026] Figure 7 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0027] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0028] The following is combined Figures 1-5 The protein processing method of the present invention is described.

[0029] In one embodiment, such as Figure 1 As shown, the process steps for implementing the protein processing method are as follows:

[0030] Step 101: Obtain the protein to be processed.

[0031] In this embodiment, the protein to be processed refers to a protein with a small number of homologous proteins, making structural prediction impossible. In this case, it is necessary to generate at least one self-generated protein corresponding to the protein to be processed using the method provided by this invention.

[0032] In this embodiment, when predicting the structure of a protein, it is first necessary to search for homologous proteins in the protein database using a preset protein retrieval method before performing structure prediction. However, if the number of homologous proteins for a protein in the protein database is too small, the retrieved homologous proteins may not provide sufficient information for structure prediction. In this case, at least one of the proteins with a small number of homologous proteins can be selected as the protein to be processed and processed using the method provided by this invention.

[0033] Step 102: Based on the protein to be processed, generate at least one self-generated protein corresponding to the protein to be processed, wherein the protein to be processed and the self-generated protein are homologous proteins.

[0034] In this embodiment, after obtaining the protein to be processed, at least one self-generated protein corresponding to the protein to be processed is generated. The protein to be processed and each self-generated protein are homologous proteins. Homologous proteins refer to a set of sequences that evolved from protein sequences of the same ancestor in different organisms. For example, two proteins with a sequence similarity of more than 30% are homologous proteins.

[0035] Step 103: Store at least one self-generated protein in a protein database to perform structure prediction on the protein to be processed based on the protein to be processed in the protein database and / or the self-generated protein.

[0036] In this embodiment, after obtaining the self-generated protein, the self-generated protein is saved to the protein database, which increases the number of homologous proteins of the protein to be processed in the protein database. Then, by using the preset protein search method, the structure prediction can be completed with sufficient information provided by the retrieved homologous proteins.

[0037] In one embodiment, a specific protein generation method is provided. Specifically, based on the protein to be processed, at least one self-generated protein corresponding to the protein to be processed is generated. The process is as follows: obtaining the sequence information of each amino acid site in the protein to be processed; obtaining the site arrangement probability corresponding to each amino acid site based on the sequence information; and generating at least one self-generated protein corresponding to the protein to be processed based on the site arrangement probability.

[0038] In this embodiment, a protein is composed of at least one amino acid arranged in a specific sequence. The position of each amino acid in the sequence is its location within the protein, and these amino acid sites constitute the protein's sequence information. The sequence information of the protein to be processed is obtained, and a self-generated protein corresponding to the protein to be processed is generated using this sequence information.

[0039] In this embodiment, after obtaining the sequence information to be processed, the site arrangement probability corresponding to each amino-terminal site is obtained. This site arrangement probability refers to the probability that a certain amino acid is arranged at this position. For example, if the second amino acid in the protein sequence to be processed is 'a', the amino acids related to the second site are obtained based on amino acid 'a', including 'a', 'b', and 'c'. Then, the site arrangement probabilities are: the arrangement probability of amino acid 'a' is 50%, the arrangement probability of amino acid 'b' is 30%, and the arrangement probability of amino acid 'c' is 10%. Then, at least one self-generated protein corresponding to the protein to be processed is generated based on the arrangement probability of each amino acid.

[0040] In this embodiment, after obtaining the permutation probability of each site, at least one self-generated protein corresponding to the protein to be processed is generated according to the permutation order of the site.

[0041] In one embodiment, when generating each self-generated protein, the amino acids at each site are determined sequentially according to the sequence information of the protein to be processed. For example, if the protein to be processed consists of 5 amino acids, and according to the sequence information, amino acid a is at site 1, amino acid b is at site 2, amino acid c is at site 3, amino acid d is at site 4, and amino acid e is at site 5, then at least one self-generated protein is generated by sequentially arranging the probabilities of the sites corresponding to a, b, c, d, and e according to the order of positions 1 to 5.

[0042] In one embodiment, the generation of self-generated proteins corresponding to the proteins to be processed is achieved through a preset sequence generation model. Specifically, generating at least one self-generated protein corresponding to the protein to be processed based on the permutation probability of each site includes: inputting the permutation probability of each site into the preset sequence generation model to obtain at least one self-generated protein corresponding to the protein to be processed output by the sequence generation model; wherein, the sequence generation model is obtained by training a self-attention-based autoregressive model using a first sample homologous protein, the first sample homologous protein including A sample proteins to be processed, and at least one self-generated protein corresponding to each of the A sample proteins to be processed, where A is an integer greater than 1.

[0043] In this embodiment, an autoregressive model based on self-attention (also known as the Transformer model) uses an attention mechanism instead of a recurrent neural network (RNN) to build the entire model framework. The Transformer model proposes a multi-headed attention mechanism, which extensively uses multi-headed self-attention in both the encoder and decoder, greatly accelerating the model's training speed.

[0044] In this embodiment, the overall structure of the Transformer model is as follows: Figure 2 As shown, the network consists of an encoder and a decoder. Each block of the encoder comprises a multi-headed attention layer and a feedforward neural network layer. The entire encoder is stacked with N blocks, where N is an integer greater than 1. The decoder has a similar structure to the encoder, except that each block contains an additional multi-headed attention layer (masked multi-headed attention). To better optimize the deep network, residual connections and layer normalization (Add & Norm) are used throughout the network.

[0045] The encoder has N identical layers, each containing two sub-layers. The first sub-layer is a multi-headed attention layer, followed by a simple fully connected layer. There is also a residual connection, and on top of that, there is a layer normalization process.

[0046] The decoder also has N identical layers, but these layers differ from those in the encoder. The decoder's layers include three sub-layers: a self-attention layer, an encoder-decoder attention layer, and a fully connected layer. The first two sub-layers are based on multi-head attention layers. One of the multi-head attention layers in the decoder includes masking. The purpose of masking is to prevent the use of future output words during training. For example, during training, the first word cannot reference the generation result of the second word. Masking sets this information to 0 to ensure that the information at prediction position i can only reference outputs smaller than i, where i is an integer greater than 0.

[0047] Figure 2 In this context, "shifted right" means right shift, "Output probabilities" means output probabilities, "Softmax" means normalized exponential function, and "Linear" means linear regression.

[0048] In this embodiment, the input sentence is set to (x1,x2,x3,......,xn); the sequence after input embedding and positional encoding is Z = (z1,z2,z3,......,zn); the sequence after encoding by the encoder and inputting into the decoder, and then outputting is (y1,y2,y3,.......,yn).

[0049] In one embodiment, the Transformer model is trained using homologous proteins from a first sample. Specifically, the training process of the Transformer model is as follows:

[0050] Obtain the protein sample to be processed; extract the site permutation probability corresponding to each amino acid site in the protein sample to be processed; input each site permutation probability into the original Transformer model to obtain at least one predicted self-generated protein output by the original Transformer model; compare the predicted homologous protein with the self-generated protein of the first sample; if the comparison result is outside the preset error range, adjust the parameters of the original Transformer model, and repeatedly input each site permutation probability of the protein sample to be processed into the original Transformer model until the predicted homologous protein and the self-generated protein of the first sample are within the preset error range, and use the original Transformer model as the final sequence generation model.

[0051] In one embodiment, the Transformer model can also use other intelligent models to implement this method, such as Generative Adversarial Networks (GANs) and Variational Auto-encoders (VAEs).

[0052] In one embodiment, when generating self-generated proteins, the influence of control factors is incorporated. Specifically, based on the permutation probability of each site, at least one self-generated protein corresponding to the protein to be processed is generated. The process is as follows: obtain a preset control factor; weight the permutation probability of each site using the control factor; and generate at least one self-generated protein corresponding to the protein to be processed based on the weighted permutation probability of each site.

[0053] In this embodiment, to generate homologous proteins more flexibly and efficiently, a control factor is added to influence the generation of the self-generated protein corresponding to the protein to be treated. The control factor refers to a numerical value determined based on a certain relationship between the homologous protein and the protein to be treated. For example, the control factor can be determined based on similarity and / or modified regions. Similarity refers to the degree of overlap between the amino acid sequences of the homologous protein and the protein to be treated after alignment. Modified regions refer to restrictions that the self-generated protein can only modify the protein to be treated in specified segments, while non-modified regions must remain consistent with the protein to be treated.

[0054] In one specific embodiment, when generating a self-generated protein corresponding to the protein to be processed through a sequence generation model, also known as a multiple sequence alignment (MSA) generator, the permutation probability of each site is input into the preset sequence generation model, and the control factor is also input into the sequence generation model.

[0055] Specifically, such as Figure 3 As shown, the protein to be processed, also known as the target protein sequence, is input into the bidirectional encoder along with a control factor, denoted as μ. In the figure, M, D, S, R, and T represent amino acids, cls represents the global representation of the protein to be processed, and s and e are special markers used in the computer program to identify the protein sequence; s represents the start character, and e represents the end character. The control factor μ acts on each amino acid site of the protein to be processed, that is, it weights the probability of each site's arrangement. The bidirectional encoder encodes the protein to be processed, transforming its information into information that is easy to analyze and calculate.

[0056] The sequence generation model is designed around three learning objective functions. Each objective function has a corresponding loss function. Specifically, firstly, the sequence generation objective. This objective function is basically consistent with the commonly used text generation objective functions in natural language processing. It requires that the self-generated protein (also known as the homology protein sequence) be as consistent as possible with the given protein to be processed (also known as the target protein sequence). Secondly, the MSA classification task. After the protein to be processed (also known as the target protein sequence) is encoded by the encoder, a global representation based on the entire sequence is obtained. We require that this global representation be as close as possible between different protein sequences within the same homology protein set (MSA), and as far apart as possible between protein sequences in different source protein sets (MSAs). Thirdly, the cycle consistency objective. This objective requires that, in addition to directly generating self-generated proteins from the protein to be processed, if the self-generated proteins are input into the sequence generation model, new proteins to be processed can also be generated, and the generated new proteins to be processed should be as consistent as possible with the original proteins to be processed.

[0057] In this embodiment, by adding control factors, the process of homologous protein generation becomes more flexible and controllable. It is possible to flexibly influence the homologous protein generation process through the required control factors, thereby improving the efficiency of homologous protein generation.

[0058] In one embodiment, another specific protein generation method is provided. Specifically, based on the protein to be processed, at least one self-generated protein corresponding to the protein to be processed is generated. The specific process is as follows: The protein to be processed is input into a preset latent variable processing model, wherein the latent variable processing model includes an encoder module, a prior network module, and a decoder module. The processing process of the latent variable processing model is as follows: The protein to be processed is input into the encoder module to obtain the expected distribution of the original latent variable corresponding to the protein to be processed, output by the encoder; the expected distribution of the original latent variable is input into the prior network module to obtain the expected distribution of the sampled latent variable, output by the prior network module; the expected distribution of the sampled latent variable is input into the decoder module to obtain a self-generated protein output by the decoder module; the number of self-generated proteins corresponding to the protein to be processed is obtained; when the number does not reach a preset threshold, the self-generated protein is used as a new protein to be processed and re-input into the latent variable processing model to obtain a new self-generated protein re-output by the latent variable processing model.

[0059] In this embodiment, latent variables refer to auxiliary variables, which indicate which Gaussian distribution the sample belongs to. A latent variable is a discrete random variable. For example... Figure 4 As shown, this illustrates the process of treating the protein using a latent variable processing model. Figure 4 In this model, x represents the protein sequence as the input variable, q(z|x) is the posterior distribution defined by the encoder network, and c is the mean of the posterior distribution; p(z|c) is the prior distribution defined by the conditional prior network, with c as its input; r(x|z) is the reconstruction network, with z as its input from the sampled conditional prior distribution; and sample refers to the sampling process. The protein to be processed is input to the encoder network, which encodes and samples the protein, outputting the expected value of the original latent variable distribution. This expected value is then input to the prior network, which samples the distribution curve corresponding to the expected value and outputs the sampled latent variable distribution. This sampled latent variable distribution is then input to the decoder network, which generates and outputs a self-generated protein based on the sampled latent variable distribution.

[0060] In this embodiment, as Figure 4As shown, the process of generating self-generated proteins through the latent variable processing model is a finite loop. When a protein to be processed is input into the latent variable processing model, it undergoes one processing step through the encoder module, prior network module, and decoder module to generate a self-generated protein. At this point, the number of self-generated proteins corresponding to the protein to be processed is counted. When the decoder module outputs a self-generated protein, the count is incremented by one and then compared with a preset count threshold. If the count does not reach the threshold, the self-generated protein output in this loop is saved, and the self-generated protein output in this loop is re-input into the encoder module to start the next loop processing process, whereby the decoder module outputs a new self-generated protein. If the count reaches the count threshold, the loop of the latent variable model stops. Thus, the number of self-generated proteins corresponding to the protein to be processed output by the latent variable processing model is the count threshold.

[0061] In one embodiment, the latent variable handling model is a model constructed based on latent variable theory. This latent variable handling model can be implemented based on the original Transformer model mentioned in the above embodiments, for example, Figure 2 The Transformer model given in the document. Specifically, the latent variable processing model consists of three parts: an encoder module, a prior network module, and a decoder module. Each module can use a separate Transformer model as its skeleton structure, namely the original encoder module, the original prior network module, and the original decoder module. These are constructed using Transformer models as their skeleton structures, and then the entire original latent variable processing model is trained to obtain the latent variable processing model.

[0062] In this embodiment, the latent variable processing model is trained on the original latent variable processing model using the second sample homologous protein. The original latent variable processing model includes an original encoder module, an original prior network module, and an original decoder module. The second sample homologous protein includes B second proteins, each of which is a homologous protein to the others, where B is an integer greater than 1.

[0063] The training process of the original latent variable processing model is as follows: B second proteins are sequentially input into the original encoder module to obtain the expected distributions of the predicted original latent variables for each of the B second proteins output by the original encoder module; each expected distribution of the predicted original latent variables is sequentially input into the original prior network module to obtain the expected distribution of the predicted sampled latent variables output by the original prior network module; each expected distribution of the predicted sampled latent variables is sequentially input into the original decoder module to obtain the second predicted self-generated protein output by the original decoder module; information divergence is calculated based on the expected distributions of the predicted original latent variables and the expected distributions of the predicted sampled latent variables; the parameters of the original latent variable processing model are adjusted using the second predicted self-generated protein and the information divergence as supervision signals until both the second predicted self-generated protein and the information divergence meet the preset supervision conditions, at which point the original latent variable processing model is determined to be the latent variable processing model.

[0064] In this embodiment, as Figure 5 The training process of the original latent variable processing model is shown, where ABHPHLSLQY, ACHPHLSLPY, and CBHQHLSLPY are examples of different protein sequences. It is assumed that the homologous protein set (the set of homologous proteins of the second sample) includes at least three protein sequences (i.e., the second protein): ABHPHLSLQY, ACHPHLSLPY, and CBHQHLSLPY. The ABHPHLSLQY sequence from the homologous protein set is input into the original encoder module (i.e.,...). Figure 5 The encoder on the left side of the image obtains the predicted raw latent variable distribution (i.e., the posterior distribution of the latent variables) of ABHPHLSLQY from the encoder module output. The expected value of the predicted raw latent variable distribution is sampled from this distribution, denoted as p_1. Then, the predicted raw latent variable distribution is passed through the original decoder module. (i.e., ...) Figure 5 The decoder outputs the second predicted self-generated protein, thus reconstructing the loss function.

[0065] In this embodiment, ACHPHLSLPY from the homologous protein set is input into another preset encoder (i.e., Figure 5 In the encoder on the right, the latent variable vector of ACHPHLSLPY (i.e. the expectation of the latent variable distribution of ACHPHLSLPY) is obtained by the encoder, and the expectation of the latent variable is input into the conditional prior network to obtain the predicted sampled latent variable distribution (i.e. the conditional prior distribution). The expected value of the predicted sampled latent variable distribution is obtained by sampling from the predicted sampled latent variable distribution. The expected value of the predicted sampled latent variable distribution is represented as p_2.

[0066] In this training process, information divergence is calculated using p_2 and p_1. Information divergence is also known as Kullback-Leibler divergence, or KL divergence for short. KL divergence is used as one supervisory signal for the original latent variable processing model, and the second predicted self-generated protein is used as another supervisory signal. The original latent variable processing model is trained using homologous proteins from the second sample until both the second predicted self-generated protein and the information divergence meet the preset supervisory conditions. At this point, the original latent variable processing model is determined to be the latent variable processing model.

[0067] In this embodiment, the preset monitoring conditions can be set according to actual conditions and needs, and the scope of protection of this application is not limited by the specific implementation form of the monitoring conditions.

[0068] In one embodiment, after generating at least one self-generated protein corresponding to the protein to be processed based on the protein to be processed, before storing the at least one self-generated protein in the protein database, each self-generated protein is screened to obtain at least one target self-generated protein corresponding to the screened protein to be processed; storing the at least one self-generated protein in the protein database includes: storing the at least one target self-generated protein corresponding to the screened protein to be processed in the protein database.

[0069] In this embodiment, after storing at least one target self-generated protein in a protein database, when performing structure prediction on the protein to be processed based on the protein to be processed and / or the self-generated protein in the protein database, the structure prediction can be performed based on the protein to be processed and / or the target self-generated protein in the protein database. Screening at least one self-generated protein corresponding to the protein to be processed can avoid the situation where a self-generated protein is of poor quality or does not meet the requirements due to accidental factors, which would cause chaos in the protein database.

[0070] In one embodiment, depending on the actual situation and specific needs, the screening method can be adopted to screen the self-generated proteins. In one approach, each self-generated protein is screened to obtain at least one target self-generated protein corresponding to the screened protein to be processed. The specific implementation process is as follows: Each self-generated protein is processed as follows: The first sequence probability density of the self-generated protein is obtained; when the first sequence probability density is greater than a preset density threshold, the self-generated protein is determined to be the target self-generated protein.

[0071] In this embodiment, the probability density of a protein can be obtained based on the permutation probability of amino acids at each amino acid site of the protein. For example, if a protein includes x1, x2, x3, ..., xn amino acids, and p(x|y) represents the permutation probability corresponding to an amino acid site, then the principle for calculating the probability density of the protein is shown in equation (1):

[0072] P(x1,x2,x3,…..,xn|y)=p(x1|y)p(x2|x1,y)…p(xn|x(n-1)…x1,y) (1).

[0073] In this embodiment, the preset density threshold can be set according to the probability density of the protein to be processed, combined with the actual situation and needs.

[0074] In one embodiment, each self-generated protein is screened to obtain at least one target self-generated protein corresponding to the screened protein to be processed. The specific implementation process is as follows: obtain the first sequence vector of the protein to be processed; perform the following processing on each self-generated protein: obtain the second sequence vector of the self-generated protein; when the difference between the first sequence vector and the second sequence vector is less than a preset vector threshold, determine the self-generated protein as the target self-generated protein.

[0075] In this embodiment, a pre-trained vector processing model can be used to process the protein to be processed and each self-generated protein, obtaining corresponding first sequence vectors and second sequence vectors. By comparing the two vectors, the similarity between each self-generated protein and the protein to be processed can be determined. The smaller the difference between the two vectors, the more similar the self-generated protein and the protein to be processed are.

[0076] In one embodiment, each self-generated protein is screened to obtain at least one target self-generated protein corresponding to the screened protein to be processed. The specific implementation process is as follows: the quality reference value of each self-generated protein is obtained by dual sampling; each self-generated protein is sorted in descending order of quality reference value to generate a protein sequence; at least one target self-generated protein is determined from the protein sequence according to a preset diversity selection method, wherein at least one homologous protein is arranged in a non-continuous order in the protein sequence.

[0077] In this embodiment, after obtaining the quality reference value for each self-generated protein using the antithetic sampling method, each self-generated protein is sorted according to its quality, from largest to smallest. In the protein sequence, the earlier the self-generated protein appears, the higher its quality and the closer it is to the protein to be processed.

[0078] In this embodiment, based on a preset diversity selection method, such as a discrete selection method based on the quality reference value of the protein to be processed, the difference between the protein to be processed and the target self-generated protein can be maximized by selecting the target self-generated protein, thereby improving the diversity of homologous proteins and avoiding the problem of insufficient diversity caused by only selecting self-generated proteins with larger quality reference values. This improves the accuracy of structural prediction of the protein to be processed based on the target self-generated protein.

[0079] In one embodiment, each self-generated protein is screened to obtain at least one target self-generated protein corresponding to the screened protein to be processed. The specific implementation process is as follows: obtain the second sequence probability density of the protein to be processed; generate a target probability density ratio based on the second sequence probability density; perform the following processing on each self-generated protein: obtain the first sequence probability density of the self-generated protein; obtain the ratio of the first sequence probability density to the second sequence probability density; when the ratio is greater than the target probability density ratio, determine the self-generated protein as the target self-generated protein.

[0080] In this embodiment, the first sequence probability density and the second sequence probability density can be obtained based on equation (1) in the above embodiment. According to the ratio of the first sequence probability density to the second sequence probability density and the target probability density ratio, target self-generated proteins with higher similarity to the homologous protein to be treated can be screened out, that is, target self-generated proteins with higher similarity to the homologous protein to be treated are accepted, and target self-generated proteins with low similarity to the homologous protein to be treated are removed, that is, target self-generated proteins with low similarity to the homologous protein to be treated are rejected, thereby improving the accuracy of structural prediction of the protein to be treated based on the target self-generated proteins.

[0081] In this embodiment, the target probability density ratio is set based on the second sequence probability density of the protein to be processed, combined with the actual situation and needs. The scope of protection of this application is not limited by the specific value of the target probability density ratio.

[0082] The screening process for self-generated proteins in the above embodiments can remove some homologous proteins that are unfavorable to structure prediction, and identify target self-generated proteins that are more in line with the requirements for structure prediction. This is equivalent to a noise removal process, which further improves the accuracy of structure prediction of the protein to be treated based on the target self-generated protein.

[0083] In one embodiment, after generating the target self-generated protein using the method described above, the target self-generated protein is stored in the protein database corresponding to the protein to be processed. The increased number of homologous proteins is then searched using a protein retrieval method. The models for homologous protein retrieval are primarily based on the retrieval framework HHblist and many-to-many sequence search (i.e., MMseq). HHblist uses a Hidden Markov Model (HMM) built from intermediate results of existing sequence searches to perform fuzzy sequence matching. HHblist can achieve high-quality and fast searches in large-scale databases like Uniprot. MMSeq has a speed advantage over HHblist; it achieves faster sequence searches based on sequence clustering and statistics.

[0084] In this embodiment, after searching for homologous proteins using the above method, the structure of the protein is predicted using a preset protein structure prediction model, such as the AlphaFold2 model.

[0085] In this embodiment, a deep learning sequence generation model is used to incorporate the target self-generated protein generated by the model into the entire search process, thereby strengthening the search signal and finding high-quality homologous protein sequences that can improve downstream tasks.

[0086] The protein processing method provided by this invention involves obtaining the protein to be processed, and then generating at least one self-generated protein corresponding to the protein to be processed, wherein the protein to be processed and the self-generated protein are homologous proteins. The at least one self-generated protein is stored in a protein database, and the structure of the protein to be processed is predicted based on the protein to be processed and / or the self-generated protein in the protein database. In cases where there are no homologous proteins or the number of homologous proteins is small, generating at least one self-generated protein corresponding to the protein to be processed through the above process increases the number of homologous proteins, facilitating subsequent use of a protein structure prediction model to predict the structure of homologous proteins. This avoids inaccurate structure predictions due to the absence or scarcity of homologous proteins, thereby improving the accuracy of protein structure prediction.

[0087] The protein processing apparatus provided by this invention is described below. The protein processing apparatus described below corresponds to the protein processing method described above. Repeated points will not be repeated. Figure 6 As shown, the protein processing apparatus includes:

[0088] Acquisition module 601 is used to acquire the protein to be processed;

[0089] The generation module 602 is used to generate at least one self-generated protein corresponding to the protein to be processed, wherein the protein to be processed and the self-generated protein are homologous proteins.

[0090] Storage module 603 is used to store at least one self-generated protein to a protein database for structural prediction of the protein to be processed based on the protein to be processed and / or the self-generated protein in the protein database.

[0091] In one embodiment, the generation module 602 is specifically used to obtain the sequence information of each amino acid site in the protein to be processed; obtain the site arrangement probability corresponding to each amino acid site according to the sequence information to be processed; and generate at least one self-generated protein corresponding to the protein to be processed according to the site arrangement probability.

[0092] In one embodiment, the generation module 602 is specifically used to input the permutation probability of each site into a preset sequence generation model to obtain at least one self-generated protein corresponding to the protein to be processed output by the sequence generation model; wherein, the sequence generation model is obtained by training an autoregressive model based on self-attention using a first sample homologous protein, the first sample homologous protein includes A sample proteins to be processed, and at least one first sample self-generated protein corresponding to each of the A sample proteins to be processed, wherein A is an integer greater than 1.

[0093] In one embodiment, the generation module 602 is specifically used to obtain a preset control factor; weight the permutation probability of each site using the control factor; and generate at least one self-generated protein corresponding to the protein to be processed based on the weighted permutation probability of each site.

[0094] In one embodiment, the generation module 602 is specifically used to input the protein to be processed into a preset latent variable processing model, wherein the latent variable processing model includes an encoder module, a prior network module, and a decoder module; the processing procedure of the latent variable processing model is as follows: inputting the protein to be processed into the encoder module to obtain the expected original latent variable distribution corresponding to the protein to be processed, output by the encoder; inputting the expected original latent variable distribution into the prior network module to obtain the expected sampled latent variable distribution, output by the prior network module; inputting the expected sampled latent variable distribution into the decoder module to obtain a self-generated protein, output by the decoder module; obtaining the number of self-generated proteins corresponding to the protein to be processed; when it is determined that the number does not reach a preset number threshold, the self-generated protein is used as a new protein to be processed and re-input into the latent variable processing model to obtain a new self-generated protein re-output by the latent variable processing model.

[0095] In one embodiment, the generation module 602 is specifically used to train the original latent variable processing model using second sample homologous proteins to obtain a latent variable processing model. The original latent variable processing model includes an original encoder module, an original prior network module, and an original decoder module. The second sample homologous proteins include B second proteins, each of which is homologous to the others, where B is an integer greater than 1. The training process of the original latent variable processing model is as follows: the B second proteins are sequentially input into the original encoder module to obtain the expected distribution of the predicted original latent variables of the B second proteins output by the original encoder module; the expected distribution of each predicted original latent variable is then sequentially input into the original encoder module. The expected distribution of latent variables is input into the original prior network module to obtain the expected distribution of the predicted sampled latent variables output by the original prior network module. Each expected distribution of the predicted sampled latent variables is then input into the original decoder module to obtain the second predicted self-generated protein output by the original decoder module. Based on the expected distribution of the predicted original latent variables and the expected distribution of the predicted sampled latent variables, the information divergence is calculated. The second predicted self-generated protein and the information divergence are used as supervision signals to adjust the parameters of the original latent variable processing model until both the second predicted self-generated protein and the information divergence meet the preset supervision conditions. At this point, the original latent variable processing model is determined to be the latent variable processing model.

[0096] In one embodiment, the protein processing apparatus further includes a screening module 604.

[0097] The screening module 604 is used to screen each self-generated protein after generating at least one self-generated protein corresponding to the protein to be processed, and before storing at least one self-generated protein in the protein database, to obtain at least one target self-generated protein corresponding to the screened protein to be processed.

[0098] The storage module 603 is specifically used to store at least one target self-generated protein corresponding to the screened protein to be processed into the protein database.

[0099] In one embodiment, the screening module 604 is specifically used to perform the following processing on each self-generated protein: obtain the first sequence probability density of the self-generated protein; when the first sequence probability density is greater than a preset density threshold, determine the self-generated protein as the target self-generated protein.

[0100] In one embodiment, the screening module 604 is specifically used to obtain a first sequence vector of the protein to be processed; and to perform the following processing on each self-generated protein: obtain a second sequence vector of the self-generated protein; and when the difference between the first sequence vector and the second sequence vector is less than a preset vector threshold, determine the self-generated protein as the target self-generated protein.

[0101] In one embodiment, the screening module 604 is specifically used to obtain the quality reference value of each self-generated protein by dual sampling; sort each self-generated protein in descending order of quality reference value to generate a protein sequence; and determine at least one target self-generated protein from the protein sequence according to a preset diversity selection method, wherein at least one homologous protein is not arranged in a continuous order in the protein sequence.

[0102] In one embodiment, the screening module 604 is specifically used to obtain the second sequence probability density of the protein to be processed; generate a target probability density ratio based on the second sequence probability density; and perform the following processing on each self-generated protein: obtain the first sequence probability density of the self-generated protein; obtain the ratio of the first sequence probability density to the second sequence probability density; and determine the self-generated protein as the target self-generated protein when the ratio is greater than the target probability density ratio.

[0103] Figure 7 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 7 As shown, the electronic device may include a processor 701, a communications interface 702, a memory 703, and a communication bus 704, wherein the processor 701, communications interface 702, and memory 703 communicate with each other via the communication bus 704. The processor 701 can invoke logical instructions stored in the memory 703 to execute a protein processing method, which includes: generating at least one self-generated protein corresponding to the protein to be processed, wherein the protein to be processed and the self-generated protein are homologous proteins; storing the at least one self-generated protein in a protein database, and performing structural prediction of the protein to be processed based on the protein to be processed and / or the self-generated protein in the protein database.

[0104] Furthermore, the logical instructions in the aforementioned memory 703 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, essentially, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0105] On the other hand, the present invention also provides a computer program product, the computer program product comprising a computer program that can be stored on a non-transitory computer-readable storage medium, wherein when the computer program is executed by a processor, the computer is able to execute the protein processing method provided by the above methods, the method comprising: generating at least one self-generated protein corresponding to the protein to be processed, wherein the protein to be processed and the self-generated protein are homologous proteins; storing at least one self-generated protein in a protein database, and performing structural prediction of the protein to be processed based on the protein to be processed and / or the self-generated protein in the protein database.

[0106] In another aspect, the present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, is implemented to perform the protein processing methods provided by the above methods, the method comprising: generating at least one self-generated protein corresponding to the protein to be processed, wherein the protein to be processed and the self-generated protein are homologous proteins; storing at least one self-generated protein in a protein database, and performing structural prediction of the protein to be processed based on the protein to be processed and / or the self-generated protein in the protein database.

[0107] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0108] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0109] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method of protein processing, characterized by, include: Obtain the protein to be processed; Based on the protein to be processed, at least one self-generated protein corresponding to the protein to be processed is generated, wherein the protein to be processed and the self-generated protein are homologous proteins. At least one of the self-generated proteins is stored in a protein database to perform structure prediction of the protein to be processed based on the protein to be processed in the protein database and / or the self-generated protein. After generating at least one self-generated protein corresponding to the protein to be processed based on the protein to be processed, and before storing at least one of the self-generated proteins into the protein database, the method further includes: Each of the self-generated proteins is screened to obtain at least one target self-generated protein corresponding to the screened protein to be processed; The step of storing at least one of the self-generated proteins to a protein database includes: At least one target self-generated protein corresponding to the screened protein to be processed is stored in the protein database; The step of screening each of the self-generated proteins to obtain at least one target self-generated protein corresponding to the screened protein to be processed includes: For each of the self-generated proteins, the following processing is performed: a first sequence probability density of the self-generated protein is obtained; when the first sequence probability density is greater than a preset density threshold, the self-generated protein is determined to be the target self-generated protein; or Obtain the first sequence vector of the protein to be processed; perform the following processing on each of the self-generated proteins: obtain the second sequence vector of the self-generated protein; when the difference between the first sequence vector and the second sequence vector is less than a preset vector threshold, determine the self-generated protein as the target self-generated protein; or A quality reference value for each self-generated protein is obtained using a dual sampling method; each self-generated protein is sorted in descending order of the quality reference values ​​to generate a protein sequence; at least one target self-generated protein is determined from the protein sequence according to a preset diversity selection method, wherein at least one homologous protein is not arranged in a continuous order in the protein sequence; or Obtain the second sequence probability density of the protein to be processed; generate a target probability density ratio based on the second sequence probability density; perform the following processing on each of the self-generated proteins: obtain the first sequence probability density of the self-generated protein; obtain the ratio of the first sequence probability density to the second sequence probability density; when the ratio is greater than the target probability density ratio, determine the self-generated protein as the target self-generated protein.

2. The protein processing method according to claim 1, wherein, The step of generating at least one self-generated protein corresponding to the protein to be processed includes: Obtain the sequence information of each amino acid site in the protein to be processed; Based on the sequence information to be processed, the site permutation probability corresponding to each amino acid site is obtained; Based on the permutation probability of each of the sites, at least one self-generated protein corresponding to the protein to be processed is generated.

3. The protein processing method according to claim 2, wherein, The step of generating at least one self-generated protein corresponding to the protein to be processed based on the permutation probability of each of the said sites includes: The permutation probability of each site is input into a preset sequence generation model to obtain at least one self-generated protein corresponding to the protein to be processed output by the sequence generation model. The sequence generation model is obtained by training a self-attention-based autoregressive model using a first sample homologous protein. The first sample homologous protein includes A sample proteins to be processed and at least one first sample self-generated protein corresponding to each of the A sample proteins to be processed, where A is an integer greater than 1.

4. The protein processing method according to claim 2, wherein, The step of generating at least one self-generated protein corresponding to the protein to be processed based on the permutation probability of each of the said sites includes: Obtain the preset control factor; The probability of each of the sites is weighted by the control factor; Based on the weighted probability of each of the sites, at least one self-generated protein corresponding to the protein to be processed is generated.

5. The protein processing method according to claim 1, characterized in that, The step of generating at least one self-generated protein corresponding to the protein to be processed includes: The protein to be processed is input into a preset latent variable processing model, wherein the latent variable processing model includes an encoder module, a prior network module, and a decoder module. The processing procedure of the latent variable handling model is as follows: The protein to be processed is input into the encoder module to obtain the expected distribution of the original latent variables corresponding to the protein to be processed, which is output by the encoder; the expected distribution of the original latent variables is input into the prior network module to obtain the expected distribution of the sampled latent variables output by the prior network module; the expected distribution of the sampled latent variables is input into the decoder module to obtain a self-generated protein output by the decoder module. Obtain the number of self-generated proteins corresponding to the protein to be processed; When the quantity is determined to be less than the preset quantity threshold, the self-generated protein is taken as the new protein to be processed and re-input into the latent variable processing model to obtain the new self-generated protein re-output by the latent variable processing model.

6. The protein processing method according to claim 5, characterized in that, The latent variable processing model is trained on the original latent variable processing model using the second sample homologous protein. The original latent variable processing model includes an original encoder module, an original prior network module, and an original decoder module. The second sample homologous protein includes B second proteins, each of which is a homologous protein to the others, where B is an integer greater than 1. The training process of the original latent variable processing model is as follows: B second proteins are sequentially input into the original encoder module to obtain the expected distribution of the predicted original latent variables of the B second proteins output by the original encoder module. Each expected distribution of the original latent variable is sequentially input into the original prior network module to obtain the expected distribution of the predicted sampled latent variable output by the original prior network module; Each of the predicted sampled latent variable distribution expectations is sequentially input into the original decoder module to obtain the second predicted self-generated protein output by the original decoder module; Calculate the information divergence based on the expected distribution of the original latent variables and the expected distribution of the sampled latent variables. Using the second predicted self-generated protein and the information divergence as supervision signals, the parameters of the original latent variable processing model are adjusted until both the second predicted self-generated protein and the information divergence meet the preset supervision conditions. Then, the original latent variable processing model is determined to be the latent variable processing model.

7. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the protein processing method as described in any one of claims 1 to 6.

8. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the protein processing method as described in any one of claims 1 to 6.

9. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the protein processing method as described in any one of claims 1 to 6.

Images