Protein structure prediction method, system and device based on optical quantum computer

Abstract

The present invention belongs to the field of protein structure prediction technology and relates to a protein structure prediction method, system, and device based on an optical quantum computer. The method comprises: 1) obtaining a multiple sequence alignment (MSA) matrix of the target sequence based on the target sequence alignment of the protein to be predicted; 2) encoding the obtained MSA matrix into a {0,1} matrix; 3) converting the protein's amino acid interactions into an undirected graph model with {0,1} state nodes and training the model using an optical quantum computer using a Boltzmann machine training mechanism to obtain weight coefficients for the edges connecting the nodes in the undirected graph model; and 4) obtaining interaction coefficients between different amino acids in the protein based on the weight coefficients. This method improves the computational efficiency and solution results of protein structure prediction, and solves the problem that the existing technology is difficult to calculate and cannot obtain accurate solutions for complex scheduling problems.

Description

Technical Field

[0001] The present invention belongs to the technical field of protein structure prediction, and relates to a protein structure prediction method, system and device, and in particular to a protein structure prediction method, system and device based on an optical quantum computer. Background Art

[0002] Protein structure prediction is of immense importance in life science research. In basic biology, proteins are the most functionally important class of biomolecules. They are transcribed from DNA sequences into mRNA, which is then translated into proteins to perform their primary functions. In biology, structure often determines the function of a biomolecule. When mRNA is translated into the corresponding amino acid chain, it folds into a specific 3D structure. Therefore, obtaining a precise spatial structure is fundamental to understanding protein function and developing drugs.

[0003] Existing methods for obtaining protein structures include experimental methods and AI-based predictions. Experimental methods primarily include X-ray crystallography, nuclear magnetic resonance (NMR), and cryo-electron microscopy (cryo-EM). These methods purify and crystallize protein samples, then use these experimental setups for structural analysis. While these methods can yield accurate three-dimensional structures, they are labor-intensive and resource-intensive, and some protein purification processes are subject to uncertainty, making it difficult to obtain relatively pure protein samples. AI-based structure prediction methods represented by AlphaFold and RoseTTAFold have shown higher efficiency than traditional experimental methods and can obtain the three-dimensional structure of proteins in a short time. However, there are still many shortcomings, mainly reflected in: 1. Insufficient sensitivity to mutations. These AI tools are not very sensitive to point mutations (single amino acid changes); 2. It is more difficult to predict dynamic regions. These AI tools are aimed at regions with fixed structures in their natural state. Therefore, current AI tools cannot accurately predict the structure of dynamic and disordered regions. The disordered structure of proteins plays an important role in protein functions, such as phase separation processes. Abnormal protein phase separation processes are significantly related to the occurrence of Alzheimer's disease and cancer; 3. The prediction of some rare proteins is not accurate. These proteins have low similarity with currently known structural proteins, so the prediction results are not accurate; 4. These AI tools still have limitations in simulating the physical interaction between drugs and proteins.

[0004] Therefore, in order to address the defects in the above-mentioned prior art, it is necessary to develop a new protein structure prediction method, system and device. Summary of the Invention

[0005] In order to overcome the shortcomings of the existing technology, the present invention proposes a protein structure prediction method, system and device based on an optical quantum computer, which improves the computational efficiency and solution results of protein structure prediction, and solves the problem that the existing technology is difficult to calculate and cannot obtain accurate solutions for complex scheduling problems.

[0006] In order to achieve the above object, the present invention provides the following technical solutions:

[0007] A protein structure prediction method based on an optical quantum computer, characterized by comprising the following steps:

[0008] 1) Obtaining a multiple sequence alignment (MSA) matrix of the target sequence based on the target sequence alignment of the protein to be predicted;

[0009] 2) Encoding the obtained multiple sequence alignment MSA matrix into a {0,1} matrix;

[0010] 3) converting the protein's amino acid interactions into an undirected graph model with {0,1} state nodes and training the model using a Boltzmann machine training mechanism using an optical quantum computer to obtain weight coefficients for the edges between the nodes in the undirected graph model;

[0011] 4) Obtaining the interaction coefficients between different amino acids in the protein based on the weight coefficients of the edges between the nodes in the undirected graph model.

[0012] Preferably, the step 1) specifically includes:

[0013] 11) Download the target sequence of the protein to be predicted from the protein sequence database;

[0014] 12) Using the downloaded target sequence as input, the multiple sequence alignment tool mmseqs2 and the protein database Swiss-Prot database were used to align the sequences to obtain the multiple sequence alignment (MSA) matrix.

[0015] Preferably, the step 2) is specifically as follows: one-hot encoding the multiple sequence alignment MSA matrix encoded by the string, there are 21 choices of amino acids at each position, for matrix alignment, the space character is represented by “-”, and an L*21-dimensional {0,1} matrix is ​​used to represent the information of a row in the multiple sequence alignment MSA matrix, and the multiple sequence alignment MSA matrix is ​​converted to obtain M L*21-dimensional {0,1} matrices, wherein L is the length of the amino acid sequence and M is the number of protein strips obtained by alignment.

[0016] Preferably, in step 3), when converting the amino acid interactions of the protein into an undirected graph model of {0,1} state nodes, first, the amino acids of the protein sequence of the protein to be predicted are taken as vertices, and the interaction relationships between the amino acids are taken as edges to establish a fully connected undirected graph; then, since each amino acid is a 21-dimensional one-hot vector, each amino acid is converted into 21 {0,1} state nodes, and the connection between any two {0,1} state nodes is taken as an edge to establish the undirected graph model of the {0,1} state nodes, wherein the state of each state node in the undirected graph model of the {0,1}} state node is 0 or 1, and the weight is h i , the weight coefficient of the edge between any two state nodes is J ij .

[0017] Preferably, in step 3), when the Boltzmann machine training mechanism is used to train using an optical quantum computer to obtain the weight coefficients of the edges between the nodes in the undirected graph model, the training loss function C is:

[0018]

[0019] Where D kl represents KL divergence, N(θ) represents negative log-conditional likelihood, N DS represents the number of {0,1}} state nodes in the undirected graph model, α is a coefficient and α∈[0,1].

[0020] Preferably, the step 4) specifically includes:

[0021] 41) Based on the weight coefficient J of the edges between the nodes in the undirected graph model ij , establish the weight coefficient matrix J;

[0022] 42) extracting a submatrix J′=[i*21:i*21+21,j*21:j*21+21] from the weight coefficient matrix J, where the submatrix represents the coupling information between the 21*21 {0,1} state nodes corresponding to the i-th amino acid and the 21*21 {0,1} state nodes corresponding to the j-th amino acid;

[0023] 43) Mean-center the submatrix as follows to obtain a mean-centered matrix J″:

[0024] J″=J′-μ0-μ1+μ

[0025] Where μ0 is the column mean, μ1 is the row mean, and μ is the overall mean of the submatrix;

[0026] 44) Calculate the interaction coefficient S between the i-th amino acid and the j-th amino acid using the obtained mean-centered matrix J″ ij :

[0027] S ij =||J″||2.

[0028] In addition, the present invention also provides a protein structure prediction system based on an optical quantum computer, which is characterized by comprising:

[0029] An alignment module, which is used to obtain a multiple sequence alignment (MSA) matrix of the target sequence based on the target sequence of the protein to be predicted;

[0030] An encoding module, which is used to encode the obtained multiple sequence alignment MSA matrix into a {0,1} matrix;

[0031] A weight coefficient acquisition module is used to convert the amino acid interactions of the protein into an undirected graph model with {0,1} state nodes and train it using a Boltzmann machine training mechanism using an optical quantum computer to obtain the weight coefficients of the edges between the nodes in the undirected graph model;

[0032] The interaction coefficient acquisition module is used to obtain the interaction coefficients between different amino acids in the protein based on the weight coefficients of the edges between the nodes in the undirected graph model.

[0033] Furthermore, the present invention also provides a protein structure prediction device based on an optical quantum computer, which is characterized by comprising:

[0034] one or more processors;

[0035] a memory for storing one or more programs;

[0036] When the one or more programs are executed by the one or more processors, the one or more processors implement the protein structure prediction method based on optical quantum computer as described above.

[0037] Finally, the present invention provides a computer-readable storage medium having a computer program stored thereon, characterized in that when the program is executed by a processor, the steps of the protein structure prediction method based on an optical quantum computer as described above are implemented.

[0038] Compared with the prior art, the protein structure prediction method, system, and device based on optical quantum computers of the present invention have one or more of the following beneficial technical effects:

[0039] 1. This invention uses a computational approach to predict protein structure from sequences. Compared with traditional experimental methods such as X-ray crystallography, nuclear magnetic resonance (NMR), and cryo-electron microscopy (Cryo-EM), it can obtain results in a shorter time, avoids the complicated steps of protein purification and expensive observation experimental equipment, and lowers the threshold for protein structure analysis.

[0040] 2. Compared with AI-based methods such as AlphaFold and RoseTTAFold, the present invention does not require a large number of high-quality training sets. Since some low-similarity proteins and high-complexity proteins do not have known structures that can be trained, these AI-based methods cannot accurately predict the structures of these proteins. The present invention does not rely on known structural data, but only needs to combine protein evolution-related information to train and obtain relevant parameters.

[0041] 3. All current technical methods, whether experimental or AI-based, only analyze protein structures for segments with fixed structures, but lack structural analysis of disordered segments of proteins. Disordered segments are crucial for proteins to function normally, and they can promote processes such as protein phase separation, which is related to various diseases such as Alzheimer's disease, diabetes, and cardiovascular disease. Therefore, the limitations of existing methods make it impossible to analyze the structure of disordered segments, so that no related drugs have been developed targeting disordered segments. The present invention can sample the structural distribution of disordered segments through energy minimization, thereby obtaining the structural state of the disordered segments.

[0042] 4. During the training process of the Boltzmann machine, the traditional method based on contrastive divergence uses other distributions to approximate the Boltzmann distribution. At the same time, due to the parameter complexity brought by full connection, the Boltzmann machine model has not been widely used. The present invention uses an optical quantum computer and directly performs energy sampling based on the quantum computer to obtain the true Boltzmann distribution, and increases the number of nodes of the Boltzmann machine model to thousands. The traditional method is difficult to converge at twenty nodes. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 It is a flow chart of the protein structure prediction method based on optical quantum computer of the present invention.

[0044] Figure 2 Figure 2 is an MSA matrix obtained by partial exemplary alignment.

[0045] Figure 3 A schematic diagram showing the encoding of an MSA matrix into a {0,1} matrix is ​​shown.

[0046] Figure 4 It is a schematic diagram of an exemplary undirected graph model of the present invention for converting amino acid interactions into {0, 1} state nodes.

[0047] Figure 5 This is a schematic diagram of extracting the interaction coefficients between amino acids from the weight coefficient matrix of the edges between nodes in the trained undirected graph model.

[0048] Figure 6 A schematic diagram of the structure of the protein structure prediction system based on optical quantum computer of the present invention is shown. DETAILED DESCRIPTION

[0049] Before describing in detail any embodiment of the present invention, it should be understood that the present invention is not limited in its application to the construction and arrangement details of the components set forth in the following description or illustrated in the following figures. The present invention is capable of other embodiments and can be practiced or carried out in various ways. In addition, it should be understood that the words and terms used herein are for descriptive purposes and should not be considered restrictive. The use of "including" or "having" and their variations herein is intended to cover the items and their equivalents set forth below and additional items. Unless otherwise specified or limited, the terms "mounted", "connected", "supported" and "coupled" and their variations are used broadly and cover direct mounting and indirect mounting, connection, support and coupling. In addition, "connected" and "coupled" are not limited to physical or mechanical connections or couplings.

[0050] Furthermore, in the disclosure of the present invention, the term "one" should be understood as "at least one" or "one or more", that is, in one embodiment, the number of an element may be one, while in another embodiment, the number of the elements may be multiple, and the term "one" should not be understood as a limitation on the quantity.

[0051] During protein evolution, amino acid mutations can occur, disrupting interactions between residues and leading to structural instability. If residues interacting with the mutated residues also mutate, forming new interactions that stabilize the protein structure, this change is called protein coevolution. Therefore, by leveraging protein evolutionary information and aligning the sequences of numerous homologous proteins to identify shared evolutionary amino acid pairs, we can infer the proximity of amino acid residues in the protein's 3D structure. Based on this global information about amino acid interactions, we can derive the protein's spatial structure.

[0052] Based on the above principles, the present invention provides a protein structure prediction method, system and device based on an optical quantum computer. By binarizing protein sequence alignment data and introducing a Boltzmann machine training mechanism, a mathematical model suitable for quantum computing training is designed. Then, a quantum computer is used to sample and solve the problem, thereby improving the computational efficiency and solution results of protein structure prediction, and solving the problem that the existing technology is difficult to calculate and cannot obtain accurate solutions for complex scheduling problems.

[0053] Figure 1 The flowchart of the protein structure prediction method based on optical quantum computer of the present invention is shown. Figure 1 As shown, the protein structure prediction method based on optical quantum computer of the present invention includes the following steps:

[0054] 1. Based on the target sequence alignment of the protein to be predicted, a multiple sequence alignment (MSA) matrix of the target sequence is obtained.

[0055] In the present invention, the target protein sequence to be predicted is first obtained. For example, a .fa file containing the amino acid sequence information of the protein to be predicted is downloaded from a protein sequence database such as SWISS-PROT. As an example, the format of the full-length amino acid sequence file for the human FIND3A protein with the Fibronection III domain is shown below.

[0056] >sp|Q9Y2H6|FND3A_HUMAN Fibronectin type-III domain-containinggprotein 3A OS=Homo sapiens OX=9606 GN=FNDC3A PE=1 SV=4

[0057]

[0058] Next, the protein sequence obtained above was used as input and compared with the protein database Swiss-Prot database using the multiple sequence alignment tool mmseqs2s to obtain the multiple sequence alignment (MSA) matrix, as shown in the following example: Figure 2 As shown. Figure 2 In , each row represents the protein sequence of a species. Figure 2 The length of the amino acid sequence is L, that is, it has L amino acids, and the number of proteins obtained by alignment is M, that is, there are M homologous proteins.

[0059] 2. Encode the obtained multiple sequence alignment MSA matrix into a {0,1} matrix.

[0060] The multiple sequence alignment MSA matrix is ​​a string code. In the present invention, the string-coded multiple sequence alignment MSA matrix needs to be one-hot encoded. There are 21 choices of amino acids at each position (that is, there are 21 amino acid species to choose from). For matrix alignment, the space character is represented by "-". An L*21-dimensional {0,1} matrix is ​​used to represent the information of a row in the multiple sequence alignment MSA matrix. The multiple sequence alignment MSA matrix is ​​converted to obtain M L*21-dimensional {0,1} matrices, where L is the length of the amino acid sequence and M is the number of protein strips obtained by alignment. For example, Figure 3 In the left half of the MSA matrix, each row in the MSA matrix is ​​represented by an L*21 {0,1} matrix, and the one-hot matrix in the right half is obtained, that is, the {0,1} matrix.

[0061] In the L*21-dimensional {0,1} matrix, one dimension in each row represents an amino acid type. During conversion, the type of amino acid corresponding to each row is determined, so the state of the dimension in the row corresponding to the amino acid type is 1, and the remaining 20 dimensions are 0.

[0062] 3. Convert the amino acid interactions of the protein into an undirected graph model with {0, 1} state nodes and use a Boltzmann machine training mechanism to train it using an optical quantum computer to obtain the weight coefficients of the edges between the nodes in the undirected graph model.

[0063] For a stable protein, based on the thermodynamic statistical model, it can be assumed that the protein amino acid sequence obeys the Boltzmann distribution:

[0064]

[0065] Where P represents the global probability of amino acid distribution, a1,…,a L represents the specific types of amino acids from position 1 to position L, and Z represents the partition function.

[0066]

[0067] Where Z is the partition function; β is the inverse temperature; E i It is the energy of a system in a certain state. In this invention, it refers to the energy corresponding to a certain amino acid distribution.

[0068] H(a1…a L ) represents the sum of the energies of all amino acids under the Boltzmann distribution, including the energy of individual amino acids and the interaction energy between any two amino acids. The specific expression can be written as:

[0069]

[0070] Where a i represents the amino acid at position i, a j represents the amino acid at position j, h i is the energy of the amino acid at position i, J ij For amino acid pairs {a i ,a j}, that is, the interaction energy between the amino acid at position i and the amino acid at position j.

[0071] The goal of the present invention is to learn the parameter θ={h i ,J ij}, so that

[0072] arg maxp(x|θ)=arg maxΠp(x i |θ)=argmax∑logp(x i |θ)(4)

[0073] Where p(x|θ) represents the parameter θ={h i ,J ij}, that is, the joint probability distribution of M data; p(x i |θ) means that the parameter θ={h i ,J ij}Next x i The probability of occurrence, in the present invention, x i Indicates the state distribution corresponding to an amino acid data; in this parameter θ={h i ,J ij}H(a1,...,a L ) to obtain the minimum value.

[0074] In order to train the above parameters θ={h i ,J ij}, the present invention uses a fully connected undirected graph model, that is, an undirected graph model with {0,1} state nodes, to represent the connection and interaction rules of protein amino acid residues, and maps the overall energy function of formula (3) into the exp space.

[0075] Specifically, when converting the amino acid interactions of proteins into an undirected graph model of {0,1} state nodes, first, the amino acids of the protein sequence of the protein to be predicted are taken as vertices, and the interaction relationships between the amino acids are taken as edges to establish a fully connected undirected graph; then, since each amino acid is a 21-dimensional one-hot vector, each amino acid is converted into 21 {0,1} state nodes, and the connection between any two {0,1} state nodes is taken as an edge to establish the undirected graph model of the {0,1} state nodes, wherein the state of each state node in the undirected graph model of the {0,1} state nodes is 0 or 1, and the weight is h i , the weight coefficient of the edge between any two state nodes is J ij .

[0076] For example, assuming the protein sequence is ABCD, that is, there are four amino acids, namely A, B, C and D, it can be converted into a fully connected undirected graph, such as Figure 4 shown. Figure 4 The four vertices in the left graph correspond to the four amino acids in the protein sequence. Figure 4 The six edges between the vertices in the left figure represent the interaction between amino acids. The amino acid at each vertex is a 21-dimensional one-hot vector, which can be converted into Figure 4 The visible layer in the middle right picture (for simplicity, Figure 4 The middle right figure shows only some nodes in the visible layer). Figure 4 The value of each node in the visible layer of the middle right figure is {0,1}, for example, the node like but If 1 is set, it means the type of the amino acid at the i-th node is α, otherwise it is set to 0. Then the interaction relationship between the four amino acids is finally converted into an undirected graph model with 4*21=84 nodes in the {0,1} state. The state of each node in the undirected graph model is 0 or 1, corresponding to the one-hot matrix encoded by each amino acid in the MSA matrix. The final parameter θ={h i ,J ij}. Among them, h i Represents the weight of the i-th node in the undirected graph model, J ij Represents the edge weight, which corresponds to the interaction coefficient between the i-th node and the j-th node in the undirected graph model.

[0077] In order to solve the parameter θ={h i ,J ij}, the present invention introduces the training mechanism of Boltzmann machine, that is, the undirected graph model of {0,1} state nodes is trained as the Boltzmann machine model. During training, the nodes in the undirected graph model of {0,1} state nodes can be regarded as visible layer nodes, and a small number of hidden layer nodes are additionally introduced, such as Figure 4 As shown in the figure on the right, based on the characteristics of the converted one-hot encoded input data, the visible layer nodes are randomly divided into half input nodes and half output nodes. For example, if the number of nodes in the visible layer is 1000, 500 nodes can be used as visible layer input nodes, another 500 nodes as visible layer output nodes, and approximately 20 additional hidden layer nodes are introduced. Due to the increased parameter complexity brought by full connectivity, the training effect of traditional Boltzmann machine models is not ideal. Furthermore, because the partition function is difficult to handle, other methods are often used to approximate the node state value probabilities to the Boltzmann distribution, resulting in unsatisfactory network parameter training results.

[0078] To this end, the present invention proposes to perform parameter training of the Boltzmann machine model based on an optical quantum computer. The training loss function C uses weighted KL divergence and negative log-likelihood, specifically:

[0079]

[0080] Where D kl represents KL divergence, N(θ) represents negative log-conditional likelihood, N DS It represents the number of visible layer nodes, that is, the number of {0,1} state nodes in the undirected graph model. α is a coefficient and α∈[0,1], and its specific value can be set as needed.

[0081] The specific training process is as follows: First, initialize the parameter θ = {h i ,J ij}, that is, manually given parameter h i ,J ij And the {0,1} matrix obtained by converting the MSA matrix is ​​used as input, and the initial parameters θ={h i ,J ij The Ising matrix obtained by encoding the Ising model using the following formula (6) is sampled on an optical quantum computer to obtain the state of the {0,1} node with the lowest energy of the undirected graph model of the {0,1} state node and the probability of each state combination obtained through energy distribution.

[0082] H=min(-∑h i x i +∑J ij xi x j ) (6)

[0083] Where H represents the energy state corresponding to the amino acid combination; x i represents the state of the amino acid at position i, x j Indicates the state of the amino acid at position j, with a value of {0,1}; h i is the energy of the amino acid at position i; J ij For amino acid pairs {a i ,a j}, that is, the interaction energy between the amino acid at position i and the amino acid at position j.

[0084] The Ising matrix is ​​obtained by encoding the above Ising model, and is input into the optical quantum computer for calculation and sampling. The output is the state of different amino acid combinations and the corresponding probabilities of different state combinations are obtained through energy distribution. The KL divergence is calculated by the state combination distribution output by the Ising model in formula (6) and the input data distribution. The calculation formula is as follows:

[0085]

[0086] Among them, D kl represents the KL divergence to be calculated, that is, the KL divergence of the probability distribution of the original data of the {0,1} matrix and the probability distribution of the data output by the optical quantum computer through the Ising model, p(x) is the probability distribution of the original data of the {0,1} matrix, and q(x) is the probability distribution of the data output by the optical quantum computer through the Ising model, that is, the state combination distribution.

[0087] The negative logarithmic conditional likelihood N(θ) can be obtained by iteratively calculating the {0,1} matrix obtained from the input M MSA matrices, and using an optical quantum computer to infer the {0,1} state of the output from the {0,1} state of the visible layer input and the probability of different state combinations, thereby calculating the negative logarithmic conditional likelihood N(θ). The calculation method is as follows:

[0088]

[0089] Where N(θ) is the parameter θ={h i ,J ij}, the negative log-conditional likelihood value, v I ,v O Represent the state of the visible layer input node and the state of the visible layer output node, p(v O |v O ; θ) means that the parameter θ={h i ,J ij}Under the undirected graph model, the input is v I v appears O The conditional probability of v D Represents the state of the D-th input node in the visible layer.

[0090] The gradient is calculated by the state combination of the sampled nodes and their corresponding probabilities. The specific calculation formula is:

[0091]

[0092] Where α is a coefficient and α∈[0,1], its specific value can be set as needed, which is the same as α in formula (5); represents the expectation of the data distribution of all nodes (including visible layer nodes and hidden layer nodes) output by the optical quantum computer based on formula (6), which is the product of the data distribution of all nodes output and the probability of the state distribution; D represents the input data set, that is, M one-hot matrices; v represents a single input data, that is, a single one-hot matrix data; q(v) represents the probability distribution of a single one-hot input data; N DS Indicates the number of visible layer nodes; It represents the expectation of the visible layer node data distribution output by the optical quantum computer based on formula (6), which is the product of the output visible layer node data distribution and the probability of the state distribution; It represents the expectation of the visible layer input node data distribution output by the optical quantum computer based on formula (6), which is the product of the output visible layer input node data distribution and the probability of the state distribution.

[0093] Based on the above calculated gradient update parameter θ={h i ,J ij}:

[0094]

[0095] Among them, θ new is the updated parameter, γ is the learning rate, which is an adjustable parameter.

[0096] Then, the training is iterated continuously through M one-hot data input until convergence, and the optimal parameter θ = {h i ,J ij}, thereby obtaining the weight coefficient J of the edges between the nodes in the undirected graph model ij .

[0097] 4. Based on the weight coefficients of the edges between the nodes in the undirected graph model, the interaction coefficients between different amino acids in the protein are obtained.

[0098] Through the training method of step 3, the weight coefficient J of the edges between 21*L state nodes in the undirected graph model is obtained ij ,Depend on Figure 5 As shown, finally the weight coefficient J between the 21*L state nodes is ij Translated into interaction coefficients between L amino acids.

[0099] In the present invention, the specific conversion process is:

[0100] In order to get Figure 5 The interaction coefficients between amino acids in the right figure, such as the interaction coefficient S between the i-th amino acid and the j-th amino acid ij First, based on the weight coefficient J of the edges between nodes in the undirected graph model ij ,Establish Figure 5 The weight coefficient matrix J is shown in the left figure.

[0101] Secondly, from Figure 5 In the undirected graph model on the left, the submatrix J′=[i*21:i*21+21,j*21:j*21+21] is extracted from the weight coefficient matrix J of the edges between nodes. This submatrix represents the coupling information between the 21*21 one-hot encoded nodes corresponding to the i-th amino acid and the 21*21 one-hot encoded nodes corresponding to the j-th amino acid.

[0102] Then, the submatrix J′ is mean-centered as follows:

[0103] J″=J′-μ0-μ1+μ

[0104] Where μ0 is the column mean, μ1 is the row mean, and μ is the overall mean of the submatrix.

[0105] Finally, the interaction coefficient between the i-th amino acid and the j-th amino acid is calculated using the mean-centered matrix J″. The specific calculation method is:

[0106] S ij =||J″||2.

[0107] The present invention uses the co-evolutionary information of proteins and finds common evolutionary amino acid pairings through sequence alignment of a large number of homologous proteins. It can infer the amino acid residues that are close to each other in the 3D structure of the protein, and the spatial structure of the protein can be obtained based on the global amino acid interaction information. At the same time, the present invention encodes the amino acid position and type selection into a {0,1} state and converts it into a fully connected Boltzmann machine model (i.e., an undirected graph model). By establishing an Ising model and using an optical quantum computer for parameter training, the interaction relationship between amino acid residues is finally obtained. The present invention can sample the structural state of the disordered region of the protein without relying on the known protein structure, and at the same time increase the scale of nodes that can be trained by the Boltzmann machine model to thousands.

[0108] In addition, the present invention also provides a protein structure prediction system based on an optical quantum computer.

[0109] like Figure 6 As shown, the protein structure prediction system based on optical quantum computer of the present invention includes:

[0110] 1. Comparison module.

[0111] The alignment module is used to obtain a multiple sequence alignment (MSA) matrix of the target sequence based on the target sequence of the protein to be predicted.

[0112] 2. Encoding module.

[0113] The editing module is used to encode the obtained multiple sequence alignment MSA matrix into a {0,1} matrix.

[0114] 3. Weight coefficient acquisition module.

[0115] The weight coefficient acquisition module is used to convert the amino acid interactions of the protein into an undirected graph model of {0,1} state nodes and use a Boltzmann machine training mechanism to train using an optical quantum computer to obtain the weight coefficients of the edges between the nodes in the undirected graph model.

[0116] 4. Interaction coefficient acquisition module.

[0117] The interaction coefficient acquisition module is used to obtain the interaction coefficients between different amino acids in the protein based on the weight coefficients of the edges between the nodes in the undirected graph model.

[0118] Moreover, the present invention also provides a protein structure prediction device based on an optical quantum computer, which includes: one or more processors; a memory for storing one or more programs; when the one or more programs are executed by the one or more processors, the one or more processors implement the protein structure prediction method based on the optical quantum computer as described above.

[0119] Finally, the present invention also provides a computer storage medium having a computer program stored thereon, characterized in that when the program is executed by a processor, the steps of the protein structure prediction method based on an optical quantum computer as described above are implemented.

[0120] Finally, it should be noted that the above embodiments are intended only to illustrate the technical solutions of the present invention and are not intended to limit the scope of protection of the present invention. Those skilled in the art may, based on the principles of the present invention, modify or replace the technical solutions of the present invention with equivalents without departing from the essence and scope of the technical solutions of the present invention.

Claims

1. A protein structure prediction method based on optical quantum computers, characterized in that: The following steps are involved: 1) Obtaining a multiple sequence alignment (MSA) matrix of the target sequence based on the target sequence alignment of the protein to be predicted; 2) Encoding the obtained multiple sequence alignment MSA matrix into a {0,1} matrix; 3) Converting the amino acid interactions of the protein into an undirected graph model of {0,1} state nodes and training it using an optical quantum computer using a Boltzmann machine training mechanism to obtain the weight coefficients of the edges between the nodes in the undirected graph model; wherein, when converting the amino acid interactions of the protein into an undirected graph model of {0,1} state nodes, first, the amino acids of the protein sequence of the protein to be predicted are used as vertices, and the interaction relationship between the amino acids is used as an edge to establish a fully connected undirected graph; then, since each amino acid is a 21-dimensional one-hot vector, each amino acid is converted into 21 {0,1} state nodes, and the connection between any two {0,1} state nodes is used as an edge to establish the undirected graph model of the {0,1} state node, wherein the state of each state node in the undirected graph model of the {0,1} state node is 0 or 1, and the weight is h i , the weight coefficient of the edge between any two state nodes is J ij When the Boltzmann machine training mechanism is used to train using an optical quantum computer to obtain the weight coefficients of the edges between nodes in the undirected graph model, the training loss function C is: Where D kl represents KL divergence, N(θ) represents negative log-conditional likelihood, N DS represents the number of {0,1} state nodes in the undirected graph model, α is a coefficient and α∈[0,1]; 4) Obtaining the interaction coefficients between different amino acids in the protein based on the weight coefficients of the edges between the nodes in the undirected graph model, which specifically includes: 41) Based on the weight coefficient J of the edges between the nodes in the undirected graph model ij , establish the weight coefficient matrix J; 42) Extract submatrix J from the weight coefficient matrix J ′ =[i*21:i*21+21,j*21:j*21+21], this submatrix represents the coupling information between the 21*21 {0,1} state nodes corresponding to the i-th amino acid and the 21*21 {0,1} state nodes corresponding to the j-th amino acid; 43) Mean-center the submatrix as follows to obtain a mean-centered matrix J″: J″=J′-μ0-μ1+μ Where μ0 is the column mean, μ1 is the row mean, and μ is the overall mean of the submatrix; 44) Calculate the interaction coefficient S between the i-th amino acid and the j-th amino acid using the obtained mean-centered matrix J″ ij : S ij =||J″||2。 2. The protein structure prediction method based on optical quantum computer according to claim 1, characterized in that: The step 1) specifically includes: 11) Download the target sequence of the protein to be predicted from the protein sequence database; 12) Using the downloaded target sequence as input, the multiple sequence alignment tool mmseqs2 and the protein database Swiss-Prot database were used to obtain the multiple sequence alignment MSA matrix.

3. The protein structure prediction method based on optical quantum computer according to claim 1, characterized in that: The step 2) is specifically as follows: one-hot encoding is performed on the multiple sequence alignment MSA matrix encoded by the string, and there are 21 choices of amino acids at each position. For matrix alignment, the space character is represented by "-", and an L*21-dimensional {0,1} matrix is ​​used to represent the information of a row in the multiple sequence alignment MSA matrix. The multiple sequence alignment MSA matrix is ​​converted to obtain M L*21-dimensional {0,1} matrices, where L is the length of the amino acid sequence and M is the number of protein strips obtained by alignment.

4. A protein structure prediction system based on optical quantum computers, characterized in that: include: An alignment module, which is used to obtain a multiple sequence alignment (MSA) matrix of the target sequence based on the target sequence of the protein to be predicted; An encoding module, which is used to encode the obtained multiple sequence alignment MSA matrix into a {0,1} matrix; A weight coefficient acquisition module is used to convert the amino acid interactions of proteins into an undirected graph model of {0,1} state nodes and adopt a Boltzmann machine training mechanism to train using an optical quantum computer to obtain the weight coefficients of the edges between the nodes in the undirected graph model; wherein, when converting the amino acid interactions of proteins into an undirected graph model of {0,1} state nodes, first, the amino acids of the protein sequence of the protein to be predicted are used as vertices, and the interaction relationship between the amino acids is used as an edge to establish a fully connected undirected graph; then, since each amino acid is a 21-dimensional one-hot vector, each amino acid is converted into 21 {0,1} state nodes, and the connection between any two {0,1} state nodes is used as an edge to establish the undirected graph model of the {0,1} state node, wherein the state of each state node in the undirected graph model of the {0,1} state node is 0 or 1, and the weight is h i , the weight coefficient of the edge between any two state nodes is J ij ; When the Boltzmann machine training mechanism is used to train using an optical quantum computer to obtain the weight coefficients of the edges between the nodes in the undirected graph model, the training loss function C is: Where D kl represents KL divergence, N(θ) represents negative log-conditional likelihood, N DS represents the number of {0,1} state nodes in the undirected graph model, α is a coefficient and α∈[0,1]; The interaction coefficient acquisition module is used to obtain the interaction coefficients between different amino acids in the protein based on the weight coefficients of the edges between the nodes in the undirected graph model, specifically including: 1) Based on the weight coefficient J of the edges between nodes in the undirected graph model ij , establish the weight coefficient matrix J; 2) Extract submatrix J from the weight coefficient matrix J ′ =[i*21:i*21+21,j*21:j*21+21], this submatrix represents the coupling information between the 21*21 {0,1} state nodes corresponding to the i-th amino acid and the 21*21 {0,1} state nodes corresponding to the j-th amino acid; 3) Mean-centered the submatrix as follows to obtain a mean-centered matrix J″: J″=J′-μ0-μ1+μ Where μ0 is the column mean, μ1 is the row mean, and μ is the overall mean of the submatrix; 4) Calculate the interaction coefficient S between the i-th amino acid and the j-th amino acid using the obtained mean-centered matrix J″ ij : S ij =||J″||2。 5. A protein structure prediction device based on an optical quantum computer, characterized in that: include: one or more processors; a memory for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the protein structure prediction method based on an optical quantum computer according to any one of claims 1 to 3.

6. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps of the protein structure prediction method based on an optical quantum computer as described in any one of claims 1 to 3 are implemented.

Images