Protein complex structure prediction model training method, equipment and medium

Abstract

The invention relates to a training method and device of a protein complex structure prediction model, electronic equipment, a computer readable medium and a program product. The training method comprises the steps of obtaining a real distance map corresponding to current training data; obtaining an input feature corresponding to the current training data; model prediction; determining an optimal docking posture of the first protein and the second protein under the current parameters; updating the current parameter according to the outer layer loss function; updating the current training data; and repeatedly executing the steps of acquiring a real distance map corresponding to the current training data, acquiring input features corresponding to the current training data, predicting a model, determining an optimal docking posture of the first protein and the second protein under the current parameter, updating the current parameter according to the outer layer loss, and updating the current training data. And obtaining a trained protein complex structure prediction model until the cycle index reaches a cycle index threshold value or the protein complex structure prediction model converges.

Description

Technical Field

[0001] The present disclosure relates to the field of artificial intelligence, and in particular to a training method for a protein complex structure prediction model, a protein complex structure prediction method, an electronic device, a computer-readable storage medium, and a computer program product. Background Art

[0002] Recently, deep learning technology has shown significant improvements in computational speed and accuracy in the field of protein complex structure prediction. For example, EquiDock is a pioneering method for applying deep learning to rigid protein docking, and AlphaFold Multimer is hailed as a breakthrough in directly predicting complex structures from amino acid sequences. However, when predicting the structure of protein complexes, EquiDock only relies on structural modal information and does not use sequence modal information; AlphaFold Multimer takes sequence modal information into account, but fails to effectively use the characteristic that the protein monomers to be complexed can be regarded as rigid bodies, resulting in inaccurate prediction results in some cases.

[0003] The methods described in this section are not necessarily methods that have been previously conceived or employed. Unless otherwise indicated, it should not be assumed that any method described in this section is considered to be prior art simply because it is included in this section. Similarly, unless otherwise indicated, the issues mentioned in this section should not be considered to have been recognized in any prior art. Summary of the invention

[0004] According to a first aspect of the present disclosure, a method for training a protein complex structure prediction model is provided, comprising: obtaining a real distance map corresponding to current training data; wherein the current training data comprises a first amino acid sequence of a first protein, a second amino acid sequence of a second protein, monomer three-dimensional structure information of the first protein, monomer three-dimensional structure information of the second protein, and three-dimensional structure information of a real protein complex composed of the first protein and the second protein; wherein the monomer three-dimensional structure information of the first protein comprises the coordinates of reference atoms included in each amino acid residue thereof; the monomer three-dimensional structure information of the second protein comprises the coordinates of reference atoms included in each amino acid residue thereof; the three-dimensional structure information of the protein complex comprises, in the real protein complex, The first real coordinates of the reference atom included in each amino acid residue of the first protein and the second real coordinates of the reference atom included in each amino acid residue of the second protein; the real distance map is determined based on the first real coordinates and the second real coordinates; the input features corresponding to the current training data are obtained; wherein the input features are determined based on the first amino acid sequence, the second amino acid sequence, the monomer three-dimensional structure information of the first protein, and the monomer three-dimensional structure information of the second protein; model prediction, including: inputting the input features into a protein complex structure prediction model with current parameters to obtain a prediction result; the prediction result includes a predicted distance map; the predicted distance map and the real distance map are two-dimensional matrices, and the element of the i-th row and j-th column of the predicted distance map represents the protein complex structure using The prediction model predicts the distance between the i-th reference atom of the first protein and the j-th reference atom of the second protein when the first protein and the second protein form a protein complex; the element of the i-th row and j-th column of the true distance map represents the true distance between the i-th reference atom of the first protein and the j-th reference atom of the second protein in the true protein complex; the optimal docking posture of the first protein and the second protein under the current parameters is determined, wherein when the first protein and the second protein are docked into a protein complex with the optimal docking posture, the distance between the reference atom included in each amino acid residue of the first protein and the reference atom included in each amino acid residue of the second protein in the protein complex is closest to the predicted distance map; and the outer layer loss function is updated according to the outer layer loss function. Current parameters; the outer loss function includes a weighted sum of a distance loss function and an inner loss function when the docking posture is the optimal docking posture; the distance loss function is used to characterize the mapping relationship between the first difference and the parameters of the protein complex structure prediction model, the first difference being the difference between the predicted distance map and the true distance map; the inner loss function is used to characterize the mapping relationship between the second difference and the parameters of the protein complex structure prediction model and the docking posture; the second difference is the difference between the distance between the reference atom included in each amino acid residue of the first protein and the reference atom included in each amino acid residue of the second protein in the protein complex when the first protein and the second protein are docked into a protein complex in the docking posture and the predicted distance map; update the current training data;Repeat the steps of obtaining the true distance map corresponding to the current training data, obtaining the input features corresponding to the current training data, model prediction, determining the optimal docking posture of the first protein and the second protein under the current parameters, updating the current parameters according to the outer layer loss, and updating the current training data until the number of cycles reaches the cycle number threshold or the protein complex structure prediction model converges, and a trained protein complex structure prediction model is obtained; wherein the initial value of the current parameter is the initial parameter of the protein complex structure prediction model. ;

[0005] According to a second aspect of the present disclosure, a method for predicting a protein complex structure is provided, the method comprising: obtaining input data, the input data comprising a first target amino acid sequence of a first target protein, a second target amino acid sequence of a second target protein, monomer three-dimensional structure information of the first target protein, and monomer three-dimensional structure information of the second target protein; determining input features corresponding to the input data according to the first target amino acid sequence, the second target amino acid sequence, the monomer three-dimensional structure information of the first target protein, and the monomer three-dimensional structure information of the second target protein; inputting the input features corresponding to the input data into a protein complex structure prediction model trained according to the method of the first aspect of the present disclosure to obtain a prediction result; the prediction result comprises a prediction distance graph; the prediction distance graph is a two-dimensional matrix, and the elements of the i-th row and j-th column of the prediction distance graph represent the protein The method comprises the following steps: determining the distance between the i-th reference atom of the first target protein and the j-th reference atom of the second target protein when the first target protein and the second target protein form a protein complex predicted by a protein complex structure prediction model; determining the optimal docking posture of the first target protein and the second target protein, wherein the docking posture of the first target protein and the second target protein is the optimal docking posture, and when the docking posture of the first target protein and the second target protein is docked into a protein complex, the distance between the reference atoms included in each amino acid residue of the first target protein and the reference atoms included in each amino acid residue of the second target protein in the protein complex is closest to the predicted distance map; and taking the three-dimensional structure of the protein complex obtained by docking the first target protein and the second protein target with the optimal docking posture as the predicted three-dimensional structure of the protein complex composed of the first target protein and the second target protein.

[0006] According to a third aspect of the present disclosure, a training device for a protein complex structure prediction model is provided, the device comprising: a processing module, for obtaining a real distance map corresponding to current training data; wherein the current training data comprises a first amino acid sequence of a first protein, a second amino acid sequence of a second protein, monomer three-dimensional structure information of the first protein, monomer three-dimensional structure information of the second protein, and three-dimensional structure information of a real protein complex composed of the first protein and the second protein; wherein the monomer three-dimensional structure information of the first protein comprises the coordinates of the reference atoms included in each amino acid residue thereof; the monomer three-dimensional structure information of the second protein comprises the coordinates of the reference atoms included in each amino acid residue thereof; the three-dimensional structure information of the protein complex comprises, in the real protein complex, the first real coordinates of the reference atoms included in each amino acid residue of the first protein and the second real coordinates of the reference atoms included in each amino acid residue of the second protein; the real distance map is determined based on the first real coordinates and the second real coordinates; further used for obtaining input features corresponding to the current training data; wherein the input features are determined based on the first amino acid sequence, the second amino acid sequence, the monomer three-dimensional structure information of the first protein, and the monomer three-dimensional structure information of the second protein; further used for model prediction, comprising: inputting the input features into the protein complex structure prediction model with current parameters to obtain a prediction result; The prediction results include a predicted distance map; the predicted distance map and the true distance map are two-dimensional matrices, the elements of the i-th row and j-th column of the predicted distance map represent the distance between the i-th reference atom of the first protein and the j-th reference atom of the second protein when the first protein and the second protein form a protein complex predicted by the protein complex structure prediction model; the elements of the i-th row and j-th column of the true distance map represent the true distance between the i-th reference atom of the first protein and the j-th reference atom of the second protein in the true protein complex; and are also used to determine the optimal docking posture of the first protein and the second protein under current parameters, wherein the first protein and the second protein are in the optimal docking posture When docking is a protein complex, the distance between the reference atom included in each amino acid residue of the first protein and the reference atom included in each amino acid residue of the second protein in the protein complex is closest to the predicted distance map; it is also used to update the current parameters according to the outer loss function; the outer loss function includes the weighted sum of the distance loss function and the inner loss function when the docking posture is the optimal docking posture; the distance loss function is used to characterize the mapping relationship between the first difference and the parameters of the protein complex structure prediction model, the first difference is the difference between the predicted distance map and the true distance map; the inner loss function is used to characterize the mapping relationship between the second difference and the parameters of the protein complex structure prediction model and the docking posture;The second difference is the difference between the distance between the reference atom included in each amino acid residue of the first protein and the reference atom included in each amino acid residue of the second protein in the protein complex and the predicted distance map when the first protein and the second protein are docked into a protein complex in a docking posture; it is also used to update the current training data; it is also used to repeatedly execute the steps of obtaining the real distance map corresponding to the current training data, obtaining the input features corresponding to the current training data, model prediction, determining the optimal docking posture of the first protein and the second protein under the current parameters, updating the current parameters according to the outer layer loss, and updating the current training data until the number of cycles reaches the cycle number threshold or the protein complex structure prediction model converges, and a trained protein complex structure prediction model is obtained; wherein the initial value of the current parameter is the initial parameter of the protein complex structure prediction model. ;

[0007] According to a fourth aspect of the present disclosure, a protein complex structure prediction device is provided, the device comprising: an acquisition module, used to acquire input data, the input data comprising a first target amino acid sequence of a first target protein, a second target amino acid sequence of a second target protein, monomer three-dimensional structure information of the first target protein, and monomer three-dimensional structure information of the second target protein; a first determination module, used to determine input features corresponding to the input data according to the first target amino acid sequence, the second target amino acid sequence, the monomer three-dimensional structure information of the first target protein, and the monomer three-dimensional structure information of the second target protein; a model prediction module, used to input the input features corresponding to the input data into a protein complex structure prediction model trained according to the first aspect of the present disclosure to obtain a prediction result; the prediction result comprises a prediction distance map; the prediction distance map is a two-dimensional matrix, and the element of the i-th row and j-th column of the prediction distance map represents The invention relates to a method for determining an optimal docking posture of the first target protein and the second target protein, wherein the distance between the i-th reference atom of the first target protein and the j-th reference atom of the second target protein when the first target protein and the second target protein form a protein complex predicted by a protein complex structure prediction model; a second determination module, used to determine the optimal docking posture of the first target protein and the second target protein, wherein when the docking posture of the first target protein and the second target protein is the optimal docking posture and the distance between the reference atoms included in each amino acid residue of the first target protein and the reference atoms included in each amino acid residue of the second target protein in the protein complex is closest to the predicted distance map; a structure determination module, used to use the three-dimensional structure of the protein complex obtained by docking the first target protein and the second protein target with the optimal docking posture as the predicted three-dimensional structure of the protein complex composed of the first target protein and the second target protein.

[0008] According to another aspect of the present disclosure, an electronic device is provided, including: a processor; and a memory, wherein the memory stores instructions executable by the processor, and when the instructions are executed by the processor, the processor executes any one of the above methods.

[0009] According to another aspect of the present disclosure, a non-transitory computer-readable storage medium storing instructions is provided. When the instructions are executed by a processor, the processor executes any of the above methods.

[0010] According to another aspect of the present disclosure, a computer program product is provided, comprising: instructions, wherein when the instructions are executed by a processor, the processor is caused to execute any of the above methods.

[0011] These and other aspects of the disclosure will be apparent from and elucidated with reference to the embodiments described hereinafter. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] The accompanying drawings exemplarily illustrate the embodiments and constitute a part of the specification, and together with the text description of the specification, are used to explain the exemplary implementation of the embodiments. The embodiments shown are for illustrative purposes only and do not limit the scope of the claims. In all drawings, the same reference numerals refer to similar but not necessarily identical elements.

[0013] Figure 1 is a flowchart of a method for training a protein complex structure prediction model according to an embodiment of the present disclosure;

[0014] Figure 2 is a schematic diagram of protein rigid body docking according to an embodiment of the present disclosure;

[0015] Figure 3 is a schematic diagram of a method for processing input features by a protein complex structure prediction model in a training method for a protein complex structure prediction model according to an embodiment of the present disclosure;

[0016] Figure 4 is a flow chart of a method for predicting a protein complex structure according to an embodiment of the present disclosure;

[0017] Figure 5 is a block diagram of a training device for a protein complex structure prediction model according to an embodiment of the present disclosure;

[0018] Figure 6 is a block diagram of a protein complex structure prediction apparatus according to an embodiment of the present disclosure;

[0019] Figure 7 is a block diagram of an exemplary electronic device that can be used to implement the embodiments of the present disclosure. DETAILED DESCRIPTION

[0020] In the present disclosure, unless otherwise specified, the use of the terms "first", "second", etc. to describe various elements is not intended to limit the positional relationship, timing relationship, or importance relationship of these elements, and such terms are only used to distinguish one element from another element. In some examples, the first element and the second element may refer to the same example of the element, and in some cases, based on the description of the context, they may also refer to different examples.

[0021] The terms used in the description of various examples described in this disclosure are only for the purpose of describing specific examples and are not intended to be limiting. Unless the context clearly indicates otherwise, if the number of elements is not specifically limited, the element can be one or more. As used herein, the term "plurality" means two or more, and the term "based on" should be interpreted as "based at least in part on". In addition, the terms "and / or" and "at least one of..." cover any one of the listed items and all possible combinations.

[0022] Unless otherwise defined, all terms (including technical and scientific terms) used herein have the same meaning as commonly understood by those of ordinary skill in the art to which the present disclosure belongs. It will be further understood that terms such as those defined in commonly used dictionaries should be interpreted as having a meaning consistent with their meaning in the relevant art and / or the context of this specification, and will not be interpreted in an idealized or overly formal sense, unless explicitly defined as such herein.

[0023] Figure 1 FIG. 1 is a flow chart of a method 100 for training a protein complex structure prediction model according to an embodiment of the present disclosure. Figure 1 As shown, the method 100 includes steps 110 to 140 .

[0024] In step 110, a true distance map corresponding to the current training data is obtained.

[0025] The current training data may be one or more pieces. If the current training data is multiple pieces, a real distance graph corresponding to each piece of the current training data is obtained.

[0026] The current training data includes the first amino acid sequence of the first protein and the second amino acid sequence of the second protein, the monomer three-dimensional structure information of the first protein, the monomer three-dimensional structure information of the second protein, and the three-dimensional structure information of the real protein complex composed of the first protein and the second protein. The three-dimensional structure information is used to characterize the three-dimensional structure of the protein or protein complex. The monomer three-dimensional structure information usually includes the coordinates of the atoms contained in the protein, and the three-dimensional structure information of the protein complex includes the coordinates of the atoms contained in the protein complex. In other words, the training data includes both the amino acid sequence of the protein and the three-dimensional structure information of the protein, providing the model with multimodal information including sequence mode and structural mode. In addition, the training data includes the three-dimensional structure information of the real protein complex composed of the first protein and the second protein. The three-dimensional structure information of the real protein complex can be considered as the true value of the three-dimensional structure information of the protein complex formed by the protein monomers, so that the predicted value can be compared with the true value during the training process, thereby performing supervised training. The training data can come from a public data set such as the PDB database, or can be obtained by processing a public data set.

[0027] Protein in the present disclosure should be understood in a broad sense, and polypeptides also belong to the proteins defined in the present disclosure.

[0028] The monomer three-dimensional structure information of the first protein includes the coordinates of the reference atoms included in each amino acid residue of the first protein; the monomer three-dimensional structure information of the second protein includes the coordinates of the reference atoms included in each amino acid residue of the second protein. For the convenience of calculation, the center of mass of the first protein and / or the center of mass of the second protein can be located at the origin of the coordinate. For example, if in the subsequent protein docking, the first protein is fixed, and the second protein adjusts the docking posture to dock with the first protein as a protein complex, the original monomer three-dimensional structure information of the first protein obtained from the PDB database can be coordinate transformed so that the center of mass of the first protein is used as the origin of the coordinate; the original monomer three-dimensional structure information of the second protein obtained from the PDB database can be coordinate transformed so that the center of mass of the second protein is used as the origin of the coordinate, or the center of mass of the second protein can be used as the origin of the coordinate. If, in subsequent protein docking, the second protein is fixed, and the first protein adjusts the docking posture to dock with the second protein as a protein complex, the original monomer three-dimensional structure information of the second protein obtained from the PDB database can be transformed to use the center of mass of the second protein as the origin of the coordinates; the original monomer three-dimensional structure information of the first protein obtained from the PDB database can be transformed to use the center of mass of the first protein as the origin of the coordinates, or the center of mass of the first protein can be used as the origin of the coordinates. Exemplarily, the center of mass of the protein can be the average value of the coordinates of the reference atoms included in each amino acid residue of the protein.

[0029] The three-dimensional structural information of the real protein complex includes the first real coordinates of the reference atoms included in each amino acid residue of the first protein and the second real coordinates of the reference atoms included in each amino acid residue of the second protein when the first protein and the second protein form a protein complex. The first real coordinate and the second real coordinate can be located in the same coordinate system as the coordinates of the reference atoms included in each amino acid residue of the first protein and the coordinates of the reference atoms included in each amino acid residue of the second protein. For the convenience of calculation, if the first protein is fixed during subsequent protein docking, and the second protein adjusts the docking posture to dock with the first protein as a protein complex, the center of mass of the first protein in the real protein complex can be located at the origin of the coordinates. It can be understood that if the first protein is regarded as a rigid body, that is, its coordinates remain unchanged when it is a monomer and as a component of the real protein complex, the coordinates of the reference atoms included in each amino acid residue of the first protein included in the monomer three-dimensional structural information of the first protein and the first real coordinates included in the three-dimensional structural information of the real protein complex should be consistent. If the second protein is fixed during subsequent protein docking, and the first protein adjusts the docking posture to dock with the second protein as a protein complex, the center of mass of the second protein in the real protein complex can be located at the origin of the coordinates. It can be understood that if the second protein is regarded as a rigid body, that is, its coordinates remain unchanged when it is a monomer and as a component of a real protein complex, then the coordinates of the reference atoms included in each amino acid residue of the second protein included in the monomer three-dimensional structural information of the second protein and the second real coordinates included in the three-dimensional structural information of the real protein complex should be consistent.

[0030] It is understandable that when predicting the three-dimensional structure of a protein complex by protein docking, the first protein and the second protein can be regarded as rigid bodies, the position and posture of one of the first protein and the second protein are fixed, and the other adjusts its position and posture in a docking posture (the docking posture includes translation and rotation), and the two are docked to form a protein complex. Figure 2 The first protein 1 is fixed and unchanged, and the second protein 2 adjusts its docking posture to dock with the first protein 1 to form a protein complex 3.

[0031] It is understandable that the three-dimensional structural information of protein monomers and protein complexes in the current training data can be determined by experimental means such as X-ray crystallography and cryo-electron microscopy. The reference atoms can be at least part of the backbone atoms in the amino acid residues, such as Cβ atoms (for amino acids that do not contain Cβ atoms, Cα atoms can be used instead). Exemplarily, in addition to the coordinates of the reference atoms, the three-dimensional structural information can also include the coordinates of other backbone atoms and side chain atoms that are not used as reference atoms.

[0032] The real distance map is determined based on the first real coordinate and the second real coordinate. When the first protein and the second protein contain m and n amino acid residues respectively, the real distance map is a two-dimensional matrix of m*n. The element of the i-th row and j-th column (i=1..m, j=1..n) of the real distance map represents the real distance between the i-th reference atom of the first protein and the j-th reference atom of the second protein in the real protein complex.

[0033] Exemplarily, when calculating the element value in the real distance map, the first real coordinate of the i-th reference atom of the first protein and the second real coordinate of the j-th reference atom of the second protein can be found from the three-dimensional structure information of the real protein complex, and then the distance between the two real coordinates is calculated as the element value of the element in the real distance map. For example, the coordinates of the first real coordinate of the i-th reference atom of the first protein monomer in the real protein complex are [x1, y1, z1], and the coordinates of the second real coordinate of the j-th reference atom of the second protein monomer are [x2, y2, z2], then the element value of the element in the i-th row and j-th column of the real distance map is the distance between [x1, y1, z1] and [x2, y2, z2].

[0034] Exemplarily, the element value of the element in the i-th row and j-th column of the real distance map is not a specific distance value, but a distance category to which the distance value belongs. Before calculating the element value in the real distance map, the mapping relationship between the distance value and the distance category is preset. For example, the specified distance interval can be divided into P categories, and the mapping relationship between the distance interval and the distance category can be obtained, and the distance value falling into the distance interval corresponds to the corresponding distance category. When calculating the element value in the real distance map, the first real coordinates of the i-th reference atom of the first protein and the second real coordinates of the j-th reference atom of the second protein can be found from the three-dimensional structure information of the real protein complex, and then the specific distance value between the two real coordinates is calculated; then, according to the mapping relationship between the distance value and the distance category, the distance category corresponding to the distance value is determined; finally, the distance category is used as the element value of the element in the real distance map. The use of distance categories instead of specific distance values ​​is because the three-dimensional structure information of the real protein complex is determined by experimental means. Due to the experimental resolution, the distance directly calculated according to the real coordinates in the three-dimensional structure information of the real protein complex is noisy. Using distance categories instead of specific distance values ​​can reduce the impact of noise.

[0035] Exemplarily, on the basis that the element value of the true distance map represents the distance category, further, the element value of the true distance map can be a vector or a group of vectors used to represent the distance category. At this time, the element value of the i-th row and j-th column of the true distance map can be determined by the following steps: first determine the distance category corresponding to the distance value between the i-th reference atom of the first protein and the j-th reference atom of the second protein in the manner described above, and then determine the one-hot encoding vector corresponding to the distance category according to the distance category, and use the vector as the element value of the i-th row and j-th category of the true distance map. In this way, it is convenient to calculate the loss value between the predicted distance map and the true distance map in the subsequent calculation.

[0036] Exemplarily, step 110 includes: acquiring current training data, and determining a real distance map from first real coordinates and second real coordinates in the current training data.

[0037] In step 120, input features corresponding to the current training data are obtained.

[0038] Each current training data has its corresponding input feature. The input feature corresponding to the current training data is determined based on the first amino acid sequence, the second amino acid sequence, the monomer three-dimensional structure information of the first protein, and the monomer three-dimensional structure information of the second protein. In other words, the input feature of the current training data reflects both the information of the protein sequence modality and the information of the protein structure modality, so that the protein can be characterized from multiple dimensions. Compared with using only single-modality information, using multi-modality information can obtain more accurate prediction results.

[0039] Exemplarily, the current training data can be input into a trained feature extraction model to obtain the input features of the current training data; or a preset feature extraction operation can be performed based on the current training data to obtain the input features of the current data. For specific methods for determining input features, reference can be made to existing technologies such as AlphaFold Multimer.

[0040] In step 130, model prediction includes: inputting the input features into the protein complex structure prediction model with current parameters to obtain a prediction result. The prediction result includes a prediction distance map. When the first protein monomer and the second protein monomer contain m and n amino acid residues respectively, the prediction distance map is a two-dimensional matrix of m*n.

[0041] Exemplarily, the protein complex structure prediction model in the embodiment of the present disclosure includes a multilayer perceptron (MLP) and an evolution module, wherein the multilayer perceptron is used to capture the nonlinear relationship between input features, and the evolution module iteratively evolves the features input to the evolution module to obtain evolution features. The structure of the evolution module can be found in Alphafold Multimer. Then, the evolution features are processed through the post-processing layer to obtain a prediction result. The multilayer perceptron, the evolution module and the post-processing layer contain learnable parameters. It is understandable that when the parameters of the protein complex structure prediction model change, the prediction results obtained by the protein complex structure prediction model with the parameters also change accordingly.

[0042] The prediction result includes a predicted distance map, and the elements of the i-th row and j-th column (i=1..m, j=1..n) of the predicted distance map represent the distance between the i-th reference atom of the first protein and the j-th reference atom of the second protein when the first protein and the second protein form a protein complex, as predicted by the protein complex structure prediction model. That is, the embodiment of the present disclosure does not directly predict the three-dimensional structural information of the protein complex, but predicts the distance between the reference atoms located on different proteins when the protein complex is formed.

[0043] Exemplarily, when the element value of the element in the real distance map represents the distance category, the element in the predicted distance map also represents the distance category. At this time, the element in the i-th row and j-th column of the predicted distance map is the distance category where the distance between the i-th reference atom of the first protein and the j-th reference atom of the second protein, as predicted by the model, belongs when the first protein and the second protein form a protein complex. That is to say, at this time, the prediction of the predicted distance map by the protein complex structure prediction model is a classification problem. It is understandable that the distance category predicted by the model can be represented by a vector, and each element in the vector represents the probability that the predicted distance belongs to each distance category.

[0044] The distance graph in which the element value represents the distance category and the distance graph in which the element value represents the distance value can be converted to each other. When the former is converted to the latter, the specific distance value corresponding to the center of the distance category can be used as the distance value corresponding to the distance category. For example, the distance interval from 0 angstroms to 64 angstroms is divided into 64 categories in equal parts to obtain 64 distance categories. The distance interval corresponding to the 23rd category is 22 angstroms-23 angstroms, and the specific distance value corresponding to the center of the 23rd category is 22.5 angstroms. The method of converting the latter to the former has been described in the previous article and will not be repeated here.

[0045] In step 140, the optimal docking posture of the first protein and the second protein under the current parameters is determined. When the first protein and the second protein are docked into a protein complex in the optimal docking posture, the distance between the reference atom included in each amino acid residue of the first protein and the reference atom included in each amino acid residue of the second protein in the protein complex is closest to the predicted distance graph.

[0046] The docking posture can be represented by a translation relationship and a rotation relationship. The following example is taken as an example in which the position and posture of the first protein are fixed, and the second protein adjusts its own position and posture in the docking posture, and docks with the first protein to form a protein complex. The monomer three-dimensional structure of the first protein and the monomer three-dimensional structure of the second protein are known. According to the docking posture, it can be determined what kind of translation and rotation the second protein has undergone relative to the monomer three-dimensional structure of the second protein when docking. Therefore, the three-dimensional structure of the second protein when docking can be obtained, that is, the coordinates of each reference atom in the second protein when docking. In this way, the distance between each reference atom of the first protein and each reference atom of the second protein when docking in a certain docking posture can be determined. It can be understood that when docking in different docking postures, the distance between each reference atom of the first protein and each reference atom of the second protein is different, which is different from the difference in the predicted distance map element value.

[0047] In the entire docking posture space, there is such a docking posture: when docking a protein complex with this docking posture, the distance between the reference atom included in each amino acid residue of the first protein and the reference atom included in each amino acid residue of the second protein in the protein complex is closest to the predicted distance map. Such a docking posture is called the optimal docking posture.

[0048] The difference between the distances between the reference atoms of the first protein and the reference atoms of the second protein and the predicted distance map can be expressed by formula (1). Under the model parameter Φ, the optimal docking posture is the translation relationship t and rotation relationship R that minimize the formula (1).

[0049]

[0050] Where X i is the coordinate of the reference atom included in the i-th amino acid residue in the monomer three-dimensional structure information of the first protein, i=1, 2, ...m; Y i Represents the coordinates of the reference atom included in the j-th amino acid residue in the monomer three-dimensional structure information of the second protein, where j=1, 2, ...n. Prediction distance map showing distance classes.

[0051] It is understandable that when the parameter Φ of the protein complex structure prediction model changes, the predicted distance map obtained by the protein complex structure prediction model with the parameter Φ also changes accordingly, and the optimal docking posture solved according to the predicted distance map also changes accordingly. In other words, the predicted distance map is a function of the model parameter Φ, and the optimal docking posture is also a function of the model parameter Φ.

[0052] In step 150, the current parameters are updated according to the outer layer loss function.

[0053] The outer layer loss function includes the weighted sum of the distance loss function and the inner layer loss function when the docking posture is the optimal docking posture.

[0054] The distance loss function is used to characterize the mapping relationship between the first difference and the parameters of the protein complex structure prediction model, that is, the distance loss function takes the model parameters as independent variables and the first difference as the function value. The first difference is the difference between the predicted distance map and the true distance map. The model is trained according to the distance loss function to make the predicted distance map close to the true distance map. It is understandable that the element values ​​in the predicted distance map and the true distance map used when calculating the distance loss function can characterize the distance category. Specifically, the element value of the predicted distance map can be a vector, each element in the vector represents the probability that the predicted distance belongs to each distance category; the element value in the true distance map is also a vector, which is a unique hot encoding vector of the distance category. The element values ​​of the predicted distance map and the true distance map both include K components, where K is the total number of distance categories.

[0055] For example, when the total number of distance categories k is K, the distance loss function L dist The form is as formula (2).

[0056]

[0057] in, is the kth component of the element value of the element in the i-th row and j-th column of the true distance map, It is the kth component of the element value in the i-th row and j-th column of the discretized prediction distance map.

[0058] The inner loss function is used to characterize the mapping relationship between the second difference and the parameters and docking posture of the protein complex structure prediction model, that is, the inner loss function takes the model parameters and the docking posture as independent variables, and the second difference is the function value. The second difference is the difference between the distance between the reference atoms included in each amino acid residue of the first protein and the reference atoms included in each amino acid residue of the second protein in the protein complex when the first protein and the second protein are docked into a protein complex in the docking posture and the predicted distance map. The model is trained according to the inner loss function to make the distance between each reference atom pair (the i-th reference atom of the first protein and the j-th reference atom of the second protein constitute a reference atom pair) under the optimal docking posture close to the predicted distance map. Compared with the outer loss function used to optimize the model, which only includes the distance loss function, or includes the distance loss function and other loss functions unrelated to the docking result, the inner loss function that reflects the docking result is added to the outer loss function, so that the model takes the improvement of the docking result as the direct optimization target, which can improve the efficiency and effect of model training.

[0059] The inner loss function takes the model parameters and docking posture as independent variables. In step 140, given the predicted distance map, i.e., the model parameters, as the independent variable, are fixed, the optimal translation relationship and rotation relationship are found; in step 150, the docking posture as the independent variable is fixed, and the model parameters are optimized according to the optimal translation relationship and rotation relationship, and then a better predicted distance map is predicted.

[0060] It is understandable that the outer layer loss function is a function of the parameters of the protein prediction model. In step 150, the current parameters can be updated based on the gradient of the outer layer loss function at the current parameters based on a conventional optimizer.

[0061] In step 160, the current training data is updated.

[0062] It is understandable that a training data pool can be prepared, and each time a piece of data in the training data pool is used as the current training data. After executing steps 110-150 based on the current training data, another piece of training data is obtained from the training data pool as the current training data, that is, the current training data is updated. For example, the training data pool includes 1000 pieces of training data, and the entire training process includes 5 epochs, and 1000 pieces of training data are used in each epoch. In the first epoch, one piece of training data that has not been used in this epoch is taken out from the 1000 pieces of training data in the training data pool at each iteration, and 1000 iterations are performed; in the second epoch, one piece of training data that has not been used in this epoch is taken out from the 1000 pieces of training data in the training data pool at each iteration, and 1000 iterations are performed; ... This is repeated for 5 epochs, and the training is completed.

[0063] Steps 110 to 160 are repeatedly performed until the number of cycles reaches a cycle number threshold or the protein structure prediction model converges, thereby obtaining a trained protein complex structure prediction model.

[0064] The disclosed embodiments provide the sequence modal information and structural modal information of protein monomers to a protein complex structure prediction model, and the model can use multimodal information to predict a prediction distance map; effectively utilize the rigid body properties of protein monomers to transform the protein complex structure prediction problem into a rigid body docking problem; and when updating the model parameters, use the optimization of the docking structure as a training goal, thereby improving the efficiency and effectiveness of model training.

[0065] In a specific embodiment, step 140 includes: based on the initial value of the docking posture, according to the inner loss function when the parameters of the protein complex structure prediction model are current parameters, through T gradient descent iterative updates, to obtain an approximate solution of the optimal docking posture.

[0066] As mentioned above, the inner layer loss function L in It is a function of the docking posture and model parameters. When the model parameters are the current parameters, the inner loss function is a function of the docking posture. The outer loss function L out In the equation, the distance loss function is a function of the model parameters, and the inner loss function when the docking posture is the optimal docking posture is a function of the model parameters. The optimal docking posture is the docking posture that minimizes the inner loss function, which is a function of the model parameters. Therefore, optimizing the model parameters and solving the optimal docking posture under the current model parameters constitute a nested structure, and the model training problem is transformed into a two-level optimization problem that can be expressed as formula (3).

[0067]

[0068] The outer loop is to update the model parameters of the protein complex structure prediction model, and the inner loop is dedicated to solving the docking posture problem.

[0069] To solve the two-layer optimization problem, we approximate the inner loop with T gradient descents (T needs to be large enough to obtain the optimal docking posture R * ,t * A sufficiently approximate solution, T is, for example, 2000). For example, in the t+1th gradient descent, the docking posture (expressed by the translation relationship t and the rotation relationship R, it should be noted that t representing the translation relationship has directionality and needs to be distinguished from t representing the number of gradient descents) is iteratively updated as shown in formula (4).

[0070]

[0071]

[0072] Wherein, t = 0, 1, 2, ... T. OPT represents an optimization algorithm, such as stochastic gradient descent (SGD) and ADAM algorithm. is the gradient of the inner loss function relative to R, is the gradient of the inner layer loss function with respect to t.

[0073] R0 and t0 may be random initial values ​​or initial values ​​determined by other means.

[0074] For example, since the orthogonality of the rotation matrix representing the rotation relationship R cannot be guaranteed, R is converted into an orthogonal vector q, using replace

[0075] and It is determined by formula (5):

[0076]

[0077]

[0078]

[0079] Among them, q t YesR t Orthogonal vector obtained by matrix transformation; X i is the coordinate of the i-th reference atom in the monomeric three-dimensional structure of the first protein, Y j The coordinates of the jth reference atom in the monomeric three-dimensional structure of the second protein, R t is the rotation relationship obtained by the t-th gradient descent update, t t is the translation relationship obtained by the t-th gradient descent update, The first protein monomer and the second protein monomer are in docking posture R t and t t The distance between the i-th reference atom of the first protein and the j-th reference atom of the second protein in the docked protein complex; is the element in the i-th row and j-th column of the predicted distance map in the form of distance numbers. and The element value of can be a specific distance value. ⊙ represents the Hadamard product.

[0080] When performing outer layer optimization, the outer layer loss function L out The differential expression of the model parameter Ф can be expressed as formula (6)

[0081]

[0082] R* ,t * Approximately R T ,t T After that, some terms in formula (6) can be transformed as in formula (7), and then d can be calculated. Φ L out Used to update model parameters Φ.

[0083]

[0084] In one embodiment, the initial values ​​of the docking poses of the first protein and the second protein are determined by a spectral initialization method.

[0085] Due to the complexity of the optimization problem and the rugged optimization situation, the gradient descent in the inner loop often requires a large number of iterations to converge and faces the challenge of finding the global minimum. Therefore, in a specific embodiment, a numerical solution of R and t is solved by the spectral initialization method, providing a more favorable starting point for the convergence of the inner optimization problem.

[0086] We order 1 of them n is a vector of all 1s. Then, replace

[0087]

[0088] According to these variable combinations, we define four variables and two center matrices as shown in Formula 9.

[0089]

[0090]

[0091]

[0092] Where I represents the identity matrix and J represents the matrix of all 1s. If it holds, then there exists the equation shown in formula 10:

[0093]

[0094] For H m X T and YH n Perform singular value decomposition (SVD) as shown in formula 11 respectively,

[0095]

[0096] The solution of the rotation relationship shown in Formula 12 can be obtained, and R solved by Formula 12 is used as R0.

[0097]

[0098] Given the rotation relationship R, the translation relationship t is solved using formula 13:

[0099]

[0100] The solution of t is shown in Formula 14, and the solution of t in Formula 14 is used as t0.

[0101]

[0102] By replacing random initial values ​​as R0 and t0 with the values ​​obtained by the above spectrum initialization, gradient descent will approach the global minimum more efficiently and accurately.

[0103] In a specific implementation, step 120 includes: determining input features of current training data.

[0104] Among them, the input features include sequence modal features and structural modal features. Sequence modal features mainly include information obtained from the sequence itself and co-evolution information obtained from multiple sequence alignment (MSA). Sequence modal features include type features F typ , sequence pairing feature F pp , and can also include multiple sequence features F msa Among them, the type feature F typ Used to characterize the type of amino acids contained in each position in the first amino acid sequence and the second amino acid sequence; multiple sequence feature F msa It is used to characterize the amino acid type of the multiple sequence alignment result of the first amino acid sequence and the second amino acid sequence, and can also characterize the amino acid covering situation; sequence pairing feature F pp Used to characterize the positional relationship between amino acids. Structural modal features include structural pairing features F p and angle feature F ang . Structural pairing feature F p Used to characterize the spatial distance between reference atoms; angle feature F ang Used to characterize the torsion angles of the main chain backbone and side chains of the first protein and the second protein.

[0105] The acquisition of sequence modal features and structural modal features can refer to the description in Alphafold Multimer.

[0106] When extracting sequence modal features and structural modal features, the first protein and the second protein can be spliced ​​together (specific connection symbols can be used to connect them in the middle) for calculation, and Nres represents the number of amino acids after splicing.

[0107] The extraction of sequence modal features and structural modal features may refer to the prior art, and the following is only an exemplary description.

[0108] For example, F typ ∈R Nres×21 Includes a unique hot representation of the amino acid types, including the 20 known amino acids and one unknown type.

[0109] Exemplary, sequence pairing feature F pp ∈R Nres×Nres×73 Contains the location information within or between chains, including the following three components: (1) size [N res , N res , 66] is used to represent the relative amino acid residue index, which is mapped to [-32, 32]. The 66th index is used to indicate that two amino acids are located in different proteins.

[0110] (2) The size is [N res , N res , 1] indicates whether two amino acid residues are derived from the same protein monomer.

[0111] (3) The size is [N res , N res , 6], the first 5 components of the relative index feature indicate the relative sym_id index mapped to [-2, 2], and the sixth component is used to indicate a pair of two residues containing different sym_ids. For the meaning of sym_id, see the description of AlphaFold Multimer.

[0112] Exemplarily, the multiple sequence feature F MSA ∈R Ncls×Nres×49 It consists of five components (N cls represents the number of multiple sequence clustering centers, i.e., the number of MSA clusters): (1) the size is [N cluster , N res , 23], including 20 amino acids, one unknown type, one gap or missing residue, and one mask mark. (2) Size [N cluster , N res , 23], which is used to characterize the distribution of amino acid types within each MSA cluster. (3) Size [N cluster , N res , 1], indicating whether there is a cover on the left side of each residue. (4) Using the formula The calculated size [N cluster , N res , 1], where c represents the number of covers to the left of each position. (5) Size [Ncluster , N res , 1] is calculated as where c' represents the average amount of cover to the left of each position.

[0113] The structural modal features are determined based on the monomeric three-dimensional structural information of the first protein and the monomeric three-dimensional structural information of the second protein. ang It can be obtained from the monomer 3D structure information of the first protein and the monomer 3D structure information of the second protein. ang ∈R Nres×57 It consists of three components. (1) The size is [N res , 22], including 20 amino acids, one unknown type and one gap or missing residue. (2) The size is [N res , 28], which uses sines and cosines to encode three main chain torsion angles, four side chain torsion angles, and an optional torsion angle with 180° rotational symmetry for the local framework of each residue. (3) The size is [N res ,7], indicating the presence or absence of a torsion angle.

[0114] Structural pairing feature F p ∈R Nres×Nres×88 Contains five components. (1) The size is [N res , N res , 39] represents the discrete distances between Cβ atoms. In the case of glycine lacking Cβ atoms, Cα atoms are used instead. The distances are discretized into 38 categories of equal width, ranging from arrive Use a category to represent greater than (2) The size is [N res , N res ,44] is extended by the one-hot representation of residue types, with a size of [N res ,1,22] and [N res ,22,1]. (3) The size is [N res , N res ,3] is obtained by constructing the unit vector of the local skeleton based on the Gram-Schmidt process of the original N-Cα-C coordinates. (4) The size is [N res , N res , 1] is an extension of the residue presence indicator. (5) The size is [N res , N res ,1] indicates whether the pair is covered.

[0115] Exemplarily, the protein complex structure prediction model includes a multilayer perceptron, an evolution module, and may further include a post-processing layer. The current parameters include the parameters of the multilayer perceptron and the parameters of the evolution module, and may further include the parameters of the post-processing layer. The process of inputting the input features into the protein complex structure prediction model with the current parameters to obtain the prediction results is as follows: Figure 3 Step 130 includes:

[0116] Step a, using a multi-layer perceptron to process the sequence pairing features and the structure pairing features to obtain the integrated first feature.

[0117] Exemplarily, the first multilayer perceptron is used to process the sequence pairing feature F pp , and obtain the transformed sequence pairing feature P pp ; Use the second multi-layer perceptron to process the structural pairing feature F p , and obtain the transformed structural pairing feature P p .

[0118] The transformed first pairing feature and the transformed pairing feature are added together (as shown in Formula 15) to obtain the integrated first feature P.

[0119] P=P pp +P p . (15)

[0120] Step b, using a multi-layer perceptron to process the type feature, the multi-sequence feature and the angle feature to obtain an integrated second feature;

[0121] Exemplarily, the third multilayer perceptron is used to process the type feature F typ , get the transformed type feature M typ , Use the fourth multi-layer perceptron to process multiple sequence features F msa , and obtain the transformed multi-sequence feature M msa , use the fifth multi-layer perceptron to process the angle feature F ang , get the transformed angle feature M ang .

[0122] The transformed type feature and the transformed multi-sequence feature are added together, and then concatenated with the transformed angle feature (as shown in Formula 16) to obtain the integrated second feature M.

[0123] M=[M typ +M msa ||M ang ] (16)

[0124] Step c: Input the integrated first feature P and the integrated second feature M into the evolution module evoformer to obtain the evolved first feature and the second characteristic of evolution

[0125] Exemplarily, the integrated first feature and the integrated second feature can be input into the evolution module for multiple iterations to obtain the evolved first feature after multiple evolutions. and the second characteristic of evolution

[0126] Step d, based on the first characteristic of evolution Determine the prediction distance map.

[0127] Through the post-processing layer, the first feature of the evolution is transformed nonlinearly to obtain the predicted distance map.

[0128] Exemplarily, the first characteristic of evolution is Through the post-processing layer, the first feature of the evolution is transformed nonlinearly as shown in formula 17 to obtain the predicted distance map of the element value representing the distance category.

[0129]

[0130] Where W is the learnable parameter in the post-processing layer and σ is the activation function.

[0131] It is understandable that the current parameters also include W at this time.

[0132] In a specific embodiment, the multiple sequence features are obtained by the following steps: performing multiple sequence alignment on the first amino acid sequence and the second amino acid sequence to obtain a multiple sequence alignment result, performing a mask operation on the multiple sequence alignment result, and performing feature extraction on the result of the mask operation to obtain the multiple sequence features. In this case, the prediction result also includes reconstructing the multiple sequence.

[0133] In addition to steps a to d, step 130 also includes:

[0134] Step e, determining and reconstructing multiple sequences according to the second characteristic of evolution;

[0135] Exemplarily, the second characteristic of evolution is Step e includes: through the post-processing layer, Linear transformation and softmax classification are performed to obtain the reconstructed multi-sequence M. It can be understood that the current parameters also include the post-processing layer The parameters to use when performing the transformation.

[0136] Correspondingly, the outer layer loss includes the weighted sum of distance loss, inner layer loss and reconstruction loss; the outer layer loss function can be expressed as Formula 18:

[0137]

[0138] The reconstruction loss is used to characterize the difference between the reconstruction of multiple sequences and the multiple sequence alignment results. Specifically, the reconstruction loss L msa It can be expressed as formula 19:

[0139]

[0140] Among them, N mask is the number of masked words (the masking ratio and method can be preset), A is the true value of the masked word, Represents the kth component of the predicted value at the position of the i-th row and j-th column.

[0141] Similar to the prediction of distance graphs, the prediction of masked tokens can also be regarded as a classification problem. It is predicted that the masked token is one of multiple (for example, 23, corresponding to 20 amino acids, an unknown type of amino acid, a gap or missing residue, and a mask mark) candidate tokens. The prediction result is represented by the probability of the 23 categories. The kth component of the prediction result represents the probability of predicting that the token is the kth token among the 23 candidate tokens. The true value is represented by the unique hot encoding of the category. For A ij Actually, it is the wth word, then A ij The wth component of is 1, and the rest are 0.

[0142] Figure 4 FIG. 6 is a flow chart of a protein complex structure prediction method 600 according to an embodiment of the present disclosure. Figure 4 As shown, the method 600 includes steps 610 to 640 .

[0143] Step 610, obtaining input data, the input data including a first target amino acid sequence of a first target protein, a second target amino acid sequence of a second target protein, monomer three-dimensional structure information of the first target protein, and monomer three-dimensional structure information of the second target protein.

[0144] The first target protein and the second protein are proteins to be composed of protein complexes. The first target amino acid sequence and the second target amino acid sequence need to be known, and the monomer three-dimensional structure information of the target protein also needs to be known. If the monomer three-dimensional structure information of the target protein is unknown, the existing protein monomer structure prediction models such as AlphaFold, Openfold, etc. can be used to predict the monomer three-dimensional structure information of the protein. If the target protein is a chain in an antibody, the existing antibody structure prediction model such as xTrimoABFold can be used to predict its three-dimensional structure information.

[0145] Step 620 , determining input features corresponding to the input data according to the first target amino acid sequence, the second target amino acid sequence, the monomer three-dimensional structure information of the first target protein, and the monomer three-dimensional structure information of the second target protein.

[0146] The method for extracting the input features corresponding to the input data is the same as the method for extracting the input features of the current training data, and will not be repeated here.

[0147] Step 630, input the input features corresponding to the input data into the protein complex structure prediction model trained by method 100 to obtain a prediction result; the prediction result includes a prediction distance map; the prediction distance map is a two-dimensional matrix, and the elements in the i-th row and j-th column of the prediction distance map represent the distance between the i-th reference atom of the first target protein and the j-th reference atom of the second target protein when the first target protein and the second target protein form a protein complex, as predicted by the protein complex structure prediction model.

[0148] The description of step 630 refers to the description of step 130 and will not be repeated here.

[0149] Step 640, determining the optimal docking posture of the first target protein and the second target protein, wherein the docking posture of the first target protein and the second target protein is the optimal docking posture. When the docking is a protein complex, the distance between the reference atom included in each amino acid residue of the first target protein and the reference atom included in each amino acid residue of the second target protein in the protein complex is closest to the predicted distance graph.

[0150] Exemplarily, step 640 includes:

[0151] Step A: determining an initial value of the docking posture between the first target protein and the second target protein;

[0152] For example, the initial value may be determined by a spectral initialization method. For details, please refer to the above description and will not be repeated here.

[0153] Step B, based on the initial value of the docking posture between the first target protein and the second target protein, iterate through T gradient descents to obtain an approximate solution of the optimal docking posture.

[0154] The gradient used in the t+1th gradient descent is the inner loss function in the docking posture R t ,t t The inner layer loss function is used to characterize the difference between the distance between the reference atoms included in each amino acid residue of the first target protein and the reference atoms included in each amino acid residue of the second target protein and the predicted protein-protein distance map in the protein complex formed by the first target protein and the second target protein in the docking posture.

[0155] By developing the gradient descent, the present application embodiment can control the computational cost in the reasoning process. When the spectral initialization method is used to determine the initial value of the docking posture between the first target protein and the second target protein, the computational cost in the reasoning process can be further reduced.

[0156] Step 650 , the three-dimensional structure of the protein complex obtained by docking the first target protein and the second target protein in the optimal docking posture is used as the predicted three-dimensional structure of the protein complex composed of the first target protein and the second target protein.

[0157] As mentioned above, when the protein is regarded as a rigid body, the protein complex structure prediction problem can be converted into a rigid body docking problem. In step 640, the optimal docking posture R is determined. * ,t * After that, in the protein complex formed by docking the first target protein and the second target protein in the optimal docking posture, the coordinates of the i-th reference atom of the first protein are Xi, and the coordinates of the j-th reference atom of the second target protein are Y j *R * +t * , thus determining the three-dimensional structural information of the protein complex.

[0158] Figure 5 is a block diagram of a training device for a protein complex structure prediction model according to an embodiment of the present disclosure, such as Figure 5 As shown, the training device 300 for the protein complex structure prediction model includes a processing unit 310 .

[0159] A processing module is used to obtain a real distance map corresponding to current training data; wherein the current training data includes a first amino acid sequence of a first protein, a second amino acid sequence of a second protein, a monomer three-dimensional structure information of the first protein, a monomer three-dimensional structure information of the second protein, and a three-dimensional structure information of a real protein complex composed of the first protein and the second protein; wherein the monomer three-dimensional structure information of the first protein includes the coordinates of the reference atoms included in each amino acid residue thereof; the monomer three-dimensional structure information of the second protein includes the coordinates of the reference atoms included in each amino acid residue thereof; the three-dimensional structure information of the protein complex includes the first real coordinates and The method comprises the following steps: the first real coordinates of the reference atom included in each amino acid residue of the second protein; the real distance map is determined based on the first real coordinates and the second real coordinates; and the method is also used to obtain input features corresponding to the current training data; wherein the input features are determined based on the first amino acid sequence, the second amino acid sequence, the monomer three-dimensional structure information of the first protein, and the monomer three-dimensional structure information of the second protein; and the method is also used for model prediction, including: inputting the input features into a protein complex structure prediction model with current parameters to obtain a prediction result; the prediction result includes a predicted distance map; the predicted distance map and the real distance map are two-dimensional matrices, and the elements of the i-th row and the j-th column of the predicted distance map represent the first protein and the second protein predicted by the protein complex structure prediction model. The distance between the i-th reference atom of the first protein and the j-th reference atom of the second protein when the proteins form a protein complex; the element of the i-th row and j-th column of the true distance map represents the true distance between the i-th reference atom of the first protein and the j-th reference atom of the second protein in the true protein complex; it is also used to determine the optimal docking posture of the first protein and the second protein under the current parameters, wherein when the first protein and the second protein are docked into a protein complex with the optimal docking posture, the distance between the reference atom included in each amino acid residue of the first protein and the reference atom included in each amino acid residue of the second protein in the protein complex is closest to the predicted distance map; it is also used to update the current parameters according to the outer layer loss function; the outer The layer loss function includes a weighted sum of a distance loss function and an inner layer loss function when the docking posture is the optimal docking posture; the distance loss function is used to characterize the mapping relationship between the first difference and the parameters of the protein complex structure prediction model, the first difference being the difference between the predicted distance map and the true distance map; the inner layer loss function is used to characterize the mapping relationship between the second difference and the parameters of the protein complex structure prediction model and the docking posture; the second difference is the difference between the distance between the reference atom included in each amino acid residue of the first protein and the reference atom included in each amino acid residue of the second protein in the protein complex when the first protein and the second protein are docked into a protein complex in the docking posture and the predicted distance map; it is also used to update the current training data;It is also used to repeatedly execute the steps of obtaining the real distance map corresponding to the current training data, obtaining the input features corresponding to the current training data, model prediction, determining the optimal docking posture of the first protein and the second protein under the current parameters, updating the current parameters according to the outer layer loss, and updating the current training data until the number of cycles reaches the cycle number threshold or the protein complex structure prediction model converges, and a trained protein complex structure prediction model is obtained; wherein the initial value of the current parameter is the initial parameter of the protein complex structure prediction model. ;

[0160] It should be understood that Figure 5 The various units of the device 300 shown in FIG. 3 can be the same as those described above with reference to Figure 1 The steps in the method 100 described above correspond to each other. Therefore, the operations, features and advantages described above for the method 100 are also applicable to the device 300 and the units included therein. For the sake of brevity, some operations, features and advantages are not repeated here.

[0161] Figure 6 FIG. 4 is a block diagram of a protein complex structure prediction apparatus 400 according to an embodiment of the present disclosure. Figure 6 As shown, the protein complex structure prediction apparatus 400 includes an acquisition module 410 , a first determination module 420 , a model prediction module 430 , a second determination module 440 , and a structure determination module 450 .

[0162] An acquisition module 410 is used to acquire input data, where the input data includes a first target amino acid sequence of a first target protein, a second target amino acid sequence of a second target protein, monomer three-dimensional structure information of the first target protein, and monomer three-dimensional structure information of the second target protein;

[0163] A first determination module 420, configured to determine an input feature corresponding to the input data according to the first target amino acid sequence, the second target amino acid sequence, the monomer three-dimensional structure information of the first target protein, and the monomer three-dimensional structure information of the second target protein;

[0164] The model prediction module 430 is used to input the input features corresponding to the input data into the protein complex structure prediction model trained according to the first aspect of the present disclosure to obtain a prediction result; the prediction result includes a prediction distance map; the prediction distance map is a two-dimensional matrix, and the elements in the i-th row and j-th column of the prediction distance map represent the distance between the i-th reference atom of the first target protein and the j-th reference atom of the second target protein when the first target protein and the second target protein form a protein complex, as predicted by the protein complex structure prediction model;

[0165] A second determination module 440 is used to determine the optimal docking posture of the first target protein and the second target protein, wherein when the docking posture of the first target protein and the second target protein is the optimal docking posture and the distance between the reference atom included in each amino acid residue of the first target protein and the reference atom included in each amino acid residue of the second target protein in the protein complex is closest to the predicted distance map;

[0166] The structure determination module 450 is used to use the three-dimensional structure of the protein complex obtained by docking the first target protein and the second target protein in the optimal docking posture as the predicted three-dimensional structure of the protein complex composed of the first target protein and the second target protein.

[0167] It should be understood that Figure 6 The various units of the device 400 shown in FIG. 4 can be the same as those in the above reference Figure 4 The steps in the method 600 described above correspond to each other. Therefore, the operations, features and advantages described above for the method 600 are also applicable to the apparatus 400 and the units included therein. For the sake of brevity, some operations, features and advantages are not described in detail here.

[0168] Although specific functions are discussed above with reference to specific units, it should be noted that the functions of the various units discussed herein can be divided into multiple units, and / or at least some functions of multiple units can be combined into a single unit. The specific unit discussed herein performs an action including the specific unit itself performing the action, or alternatively the specific unit calls or otherwise accesses another component or unit that performs the action (or performs the action in conjunction with the specific unit). Therefore, the specific unit that performs an action can include the specific unit itself that performs the action and / or another unit that the specific unit calls or otherwise accesses and performs the action.

[0169] It should also be understood that various techniques may be described herein in the general context of software hardware elements or program units. Individual units may be implemented in hardware or in hardware in combination with software and / or firmware. For example, these units may be implemented as computer program codes / instructions that are configured to be executed in one or more processors and stored in a computer-readable storage medium. Alternatively, these units may be implemented as hardware logic / circuits.

[0170] According to one aspect of the present disclosure, an electronic device is provided, comprising: at least one processor; and at least one memory communicatively connected to the at least one processor, wherein the at least one memory stores instructions, and when the instructions are executed by the at least one processor, the at least one processor executes the above method.

[0171] According to another aspect of the present disclosure, a non-transitory computer-readable storage medium storing instructions is provided. When the instructions are executed by at least one processor of a computer, the computer executes the above method.

[0172] According to another aspect of the present disclosure, a computer program product is provided, including a computer program, and the computer program implements the above method when executed by a processor.

[0173] Figure 7 is a block diagram of an exemplary electronic device that can be used to implement the embodiments of the present disclosure.

[0174] The electronic device 1200 can be a variety of different types of devices. Examples of the electronic device 1200 include, but are not limited to, desktop computers, server computers, laptop or netbook computers, mobile devices (e.g., tablet computers, cellular or other wireless phones (e.g., smart phones), notepad computers, mobile stations), wearable devices (e.g., glasses, watches), entertainment devices (e.g., entertainment appliances, set-top boxes communicatively coupled to display devices, game consoles), televisions or other display devices, automotive computers, and the like.

[0175] The electronic device 1200 may include at least one processor 1202, memory 1204, communication interface(s) 1206, a display device 1208, other input / output (I / O) devices 1210, and one or more mass storage devices 1212, which can communicate with each other, such as via a system bus 1214 or other appropriate connection.

[0176] The processor 1202 may be a single processing unit or multiple processing units, all of which may include a single or multiple computing units or multiple cores. The processor 1202 may be implemented as one or more microprocessors, microcomputers, microcontrollers, digital signal processors, central processing units, state machines, logic circuits, and / or any device that manipulates signals based on operating instructions. Among other capabilities, the processor 1202 may be configured to retrieve and execute computer-readable instructions stored in the memory 1204, mass storage device 1212, or other computer-readable media, such as program code of an operating system 1216, program code of an application program 1218, program code of other programs 1220, and the like.

[0177] The memory 1204 and the mass storage device 1212 are examples of computer-readable storage media for storing instructions that are executed by the processor 1202 to implement the various functions described above. For example, the memory 1204 may generally include both volatile memory and non-volatile memory (e.g., RAM, ROM, etc.). In addition, the mass storage device 1212 may generally include a hard drive, a solid-state drive, a removable medium, including external and removable drives, a memory card, a flash memory, a floppy disk, an optical disk (e.g., a CD, a DVD), a storage array, a network attached storage, a storage area network, etc. The memory 1204 and the mass storage device 1212 may all be collectively referred to herein as memory or computer-readable storage media, and may be a non-transitory medium capable of storing computer-readable, processor-executable program instructions as computer program code, which may be executed by the processor 1202 as a specific machine configured to implement the operations and functions described in the examples herein.

[0178] A number of programs may be stored on mass storage device 1212. These programs include operating system 1216, one or more application programs 1218, other programs 1220, and program data 1222, and they may be loaded into memory 1204 for execution. Examples of such applications or program modules may include, for example, computer program logic (e.g., computer program code or instructions) for implementing the following components / functions: method 100 (including any suitable steps of method 100), method 400 (including any suitable steps of method 400), method 600 (including any suitable steps of method 600), and / or additional embodiments described herein.

[0179] Although in Figure 7 1204 of the electronic device 1200, but modules 1216, 1218, 1220, and 1222, or portions thereof, may be implemented using any form of computer-readable media accessible by the electronic device 1200. As used herein, "computer-readable media" includes at least two types of computer-readable media, namely, computer-readable storage media and communication media.

[0180] Computer-readable storage media include volatile and non-volatile, removable and non-removable media implemented by any method or technology for storing information, such as computer-readable instructions, data structures, program modules or other data. Computer-readable storage media include but are not limited to RAM, ROM, EEPROM, flash memory or other memory technology, CD-ROM, digital versatile disk (DVD), or other optical storage device, magnetic cassette, magnetic tape, magnetic disk storage device or other magnetic storage device, or any other non-transmission medium that can be used to store information for access by electronic devices. In contrast, communication media can embody computer-readable instructions, data structures, program modules or other data in a modulated data signal such as a carrier wave or other transmission mechanism. Computer-readable storage media defined herein do not include communication media.

[0181] One or more communication interfaces 1206 are used to exchange data with other devices, such as through a network, direct connection, etc. Such communication interfaces can be one or more of the following: any type of network interface (e.g., a network interface card (NIC)), a wired or wireless (such as IEEE 802.11 wireless LAN (WLAN)) wireless interface, a Worldwide Interoperability for Microwave Access (Wi-MAX) interface, an Ethernet interface, a Universal Serial Bus (USB) interface, a cellular network interface, a Bluetooth™ interface, a Near Field Communication (NFC) interface, etc. The communication interface 1206 can facilitate communication within a variety of network and protocol types, including wired networks (e.g., LAN, cable, etc.) and wireless networks (e.g., WLAN, cellular, satellite, etc.), the Internet, etc. The communication interface 1206 can also provide communication with external storage devices (not shown) such as storage arrays, network attached storage, storage area networks, etc.

[0182] In some examples, a display device 1208 such as a monitor may be included for displaying information and images to the user. Other I / O devices 1210 may be devices that receive various inputs from the user and provide various outputs to the user, and may include a touch input device, a gesture input device, a camera, a keyboard, a remote control, a mouse, a printer, an audio input / output device, and the like.

[0183] The technology described herein can be supported by these various configurations of the electronic device 1200, and is not limited to the specific examples of the technology described herein. For example, the function can also be implemented in whole or in part on the "cloud" by using a distributed system. The cloud includes and / or represents a platform for resources. The platform abstracts the underlying functions of the hardware (e.g., server) and software resources of the cloud. Resources can include applications and / or data that can be used when performing computing processing on a server away from the electronic device 1200. Resources can also include services provided through the Internet and / or through a subscriber network such as a cellular or Wi-Fi network. The platform can abstract resources and functions to connect the electronic device 1200 to other electronic devices. Therefore, the implementation of the functions described herein can be distributed throughout the cloud. For example, the functions can be implemented partially on the electronic device 1200 and partially through a platform that abstracts the functions of the cloud.

[0184] Although the present disclosure has been illustrated and described in detail in the drawings and the foregoing description, such illustration and description should be considered illustrative and exemplary rather than restrictive; the present disclosure is not limited to the disclosed embodiments. By studying the drawings, the disclosure and the appended claims, those skilled in the art will be able to understand and implement variations to the disclosed embodiments when practicing the claimed subject matter. In the claims, the word "comprising" does not exclude other elements or steps that are not listed, the indefinite article "a" or "an" does not exclude a plurality, and the term "plurality" means two or more. The mere fact that certain measures are recited in mutually different dependent claims does not indicate that a combination of these measures cannot be used to advantage.

Claims

1. A method for training a protein complex structure prediction model, characterized in that: The method comprises: Obtain a true distance map corresponding to current training data; wherein the current training data includes a first amino acid sequence of a first protein, a second amino acid sequence of a second protein, monomer three-dimensional structural information of the first protein, monomer three-dimensional structural information of the second protein, and three-dimensional structural information of a real protein complex composed of the first protein and the second protein; wherein the monomer three-dimensional structural information of the first protein includes the coordinates of the reference atoms included in each amino acid residue thereof; the monomer three-dimensional structural information of the second protein includes the coordinates of the reference atoms included in each amino acid residue thereof; the three-dimensional structural information of the protein complex includes the first real coordinates of the reference atoms included in each amino acid residue of the first protein and the second real coordinates of the reference atoms included in each amino acid residue of the second protein in the real protein complex; the true distance map is determined based on the first real coordinates and the second real coordinates; Obtaining input features corresponding to the current training data; wherein the input features are determined based on the first amino acid sequence, the second amino acid sequence, the monomer three-dimensional structure information of the first protein, and the monomer three-dimensional structure information of the second protein; Model prediction, including: inputting the input features into a protein complex structure prediction model with current parameters to obtain a prediction result; the prediction result includes a prediction distance map; the prediction distance map and the real distance map are two-dimensional matrices, the elements of the i-th row and j-th column of the prediction distance map represent the distance between the i-th reference atom of the first protein and the j-th reference atom of the second protein when the first protein and the second protein form a protein complex predicted by the protein complex structure prediction model; the elements of the i-th row and j-th column of the real distance map represent the real distance between the i-th reference atom of the first protein and the j-th reference atom of the second protein in the real protein complex; determining an optimal docking posture of the first protein and the second protein under current parameters, wherein when the first protein and the second protein are docked into a protein complex in the optimal docking posture, a distance between a reference atom included in each amino acid residue of the first protein and a reference atom included in each amino acid residue of the second protein in the protein complex is closest to the predicted distance map; The current parameters are updated according to the outer loss function; the outer loss function includes a weighted sum of a distance loss function and an inner loss function when the docking posture is the optimal docking posture; the distance loss function is used to characterize the mapping relationship between the first difference and the parameters of the protein complex structure prediction model, and the first difference is the difference between the predicted distance map and the true distance map; the inner loss function is used to characterize the mapping relationship between the second difference and the parameters of the protein complex structure prediction model and the docking posture; the second difference is the difference between the distance between the reference atom included in each amino acid residue of the first protein and the reference atom included in each amino acid residue of the second protein in the protein complex when the first protein and the second protein are docked into a protein complex in the docking posture and the predicted distance map; Update current training data; Repeating the steps of obtaining the true distance map corresponding to the current training data, obtaining the input features corresponding to the current training data, model prediction, determining the optimal docking posture of the first protein and the second protein under the current parameters, updating the current parameters according to the outer layer loss, and updating the current training data until the number of cycles reaches a cycle number threshold or the protein complex structure prediction model converges, thereby obtaining a trained protein complex structure prediction model; Wherein, the initial value of the current parameter is the initial parameter of the protein complex structure prediction model.

2. The method according to claim 1, characterized in that The determining of the optimal docking posture of the first protein and the second protein under the current parameters includes: based on the initial value of the docking posture, according to the inner layer loss function when the parameters of the protein complex structure prediction model are the current parameters, through T times of gradient descent iterative update, to obtain an approximate solution of the optimal docking posture.

3. The method according to claim 2, characterized in that The initial value of the docking posture is determined according to a spectral initialization method.

4. The method according to any one of claims 1 to 3, characterized in that: The input features include sequence modal features and structural modal features; The sequence modal features include type features and sequence pairing features, and the sequence modal features are determined based on the first amino acid sequence and the second amino acid sequence; the type features are used to characterize the types of amino acids at each position in the first amino acid sequence and the second amino acid sequence; the sequence pairing features are used to characterize the positional relationship between amino acids in the first amino acid sequence and the positional relationship between amino acids in the second amino acid sequence; The structural modal features include structural pairing features and angle features; The structural modal feature is determined based on the monomer three-dimensional structural information of the first protein and the monomer three-dimensional structural information of the second protein; the structural pairing feature is used to characterize the distance between the amino acids in the first amino acid sequence and the distance between the amino acids in the second amino acid sequence; The angle feature is used to characterize the torsion angle of the main chain backbone and side chain of the first protein and the torsion angle of the main chain backbone and side chain of the second protein; The protein complex structure prediction model includes a multi-layer perceptron and an evolutionary module; Inputting the input features into a protein complex structure prediction model with current parameters to obtain prediction results, including: Processing the sequence pairing feature and the structure pairing feature according to the multilayer perceptron to obtain an integrated first feature; Processing the type feature and the angle feature according to the multilayer perceptron to obtain an integrated second feature; Inputting the integrated first feature and the integrated second feature into the evolution module to obtain an evolved first feature and an evolved second feature; The predicted distance graph is determined based on the evolved first feature.

5. The method according to claim 4, characterized in that The input features also include multiple sequence features; the multiple sequence features are determined in the following manner: including performing multiple sequence alignment on the first amino acid sequence and the second amino acid sequence to obtain a multiple sequence alignment result, performing a mask operation on the multiple sequence alignment result, and performing feature extraction on the result of the mask operation to obtain the multiple sequence features; the multiple sequence features are used to characterize the amino acid type and amino acid deletion in the multiple sequence alignment result of the first amino acid sequence and the second amino acid sequence; The step of processing the type feature and the angle feature according to the multi-layer perceptron to obtain an integrated second feature includes: The type feature, the multiple sequence feature and the angle feature are processed by the multi-layer perceptron to obtain an integrated second feature.

6. The method according to claim 4 or 5, characterized in that: The prediction results also include reconstructing multiple sequences; Inputting the input features into a protein complex structure prediction model with current parameters to obtain a prediction result, further comprising: Determine and reconstruct multiple sequences according to the second characteristic of the evolution; Among them, the outer layer loss function includes the weighted sum of the distance loss function, the inner layer loss function and the reconstruction loss function; the reconstruction loss function is used to characterize the mapping relationship between the difference between the reconstructed multiple sequences and the multiple sequence alignment results and the parameters of the protein complex structure prediction model.

7. A method for predicting protein complex structure, characterized in that: The method comprises: Acquire input data, wherein the input data includes a first target amino acid sequence of a first target protein, a second target amino acid sequence of a second target protein, monomer three-dimensional structure information of the first target protein, and monomer three-dimensional structure information of the second target protein; Determining an input feature corresponding to the input data according to the first target amino acid sequence, the second target amino acid sequence, the monomer three-dimensional structure information of the first target protein, and the monomer three-dimensional structure information of the second target protein; Inputting the input features corresponding to the input data into a protein complex structure prediction model trained by the method according to any one of claims 1 to 6 to obtain a prediction result; the prediction result includes a prediction distance map; the prediction distance map is a two-dimensional matrix, and the elements in the i-th row and j-th column of the prediction distance map represent the distance between the i-th reference atom of the first target protein and the j-th reference atom of the second target protein when the first target protein and the second target protein form a protein complex predicted by the protein complex structure prediction model; Determining an optimal docking posture of the first target protein and the second target protein, wherein the docking posture of the first target protein and the second target protein is the optimal docking posture. When the first target protein and the second target protein are docked into a protein complex, the distance between the reference atom included in each amino acid residue of the first target protein and the reference atom included in each amino acid residue of the second target protein in the protein complex is closest to the predicted distance map; The three-dimensional structure of the protein complex obtained by docking the first target protein and the second target protein in the optimal docking posture is used as the predicted three-dimensional structure of the protein complex composed of the first target protein and the second target protein.

8. A training device for a protein complex structure prediction model, characterized in that: The device comprises: A processing module, used to obtain a real distance map corresponding to current training data; wherein the current training data includes a first amino acid sequence of a first protein, a second amino acid sequence of a second protein, a monomer three-dimensional structure information of the first protein, a monomer three-dimensional structure information of the second protein, and a three-dimensional structure information of a real protein complex composed of the first protein and the second protein; wherein the monomer three-dimensional structure information of the first protein includes the coordinates of the reference atoms included in each amino acid residue; the monomer three-dimensional structure information of the second protein includes the coordinates of the reference atoms included in each amino acid residue; the three-dimensional structure information of the protein complex includes the coordinates of the first protein in the real protein complex. The method comprises the following steps: determining the first real coordinates of the reference atom included in each amino acid residue of the first protein and the second real coordinates of the reference atom included in each amino acid residue of the second protein; the real distance map is determined based on the first real coordinates and the second real coordinates; and is also used to obtain input features corresponding to the current training data; wherein the input features are determined based on the first amino acid sequence, the second amino acid sequence, the monomer three-dimensional structure information of the first protein, and the monomer three-dimensional structure information of the second protein; and is also used for model prediction, including: inputting the input features into a protein complex structure prediction model with current parameters to obtain a prediction result; the prediction result includes a predicted distance map; the predicted distance map and the real distance map The distance map is a two-dimensional matrix, wherein the elements of the i-th row and j-th column of the predicted distance map represent the distance between the i-th reference atom of the first protein and the j-th reference atom of the second protein when the first protein and the second protein form a protein complex predicted by the protein complex structure prediction model; the elements of the i-th row and j-th column of the real distance map represent the real distance between the i-th reference atom of the first protein and the j-th reference atom of the second protein in the real protein complex; and the distance map is also used to determine the optimal docking posture of the first protein and the second protein under the current parameters, wherein the first protein and the second protein are docked in the optimal docking posture as the protein complex. When a protein complex is formed, the distance between the reference atom included in each amino acid residue of the first protein and the reference atom included in each amino acid residue of the second protein in the protein complex is closest to the predicted distance map; it is also used to update the current parameters according to the outer layer loss function; the outer layer loss function includes a weighted sum of a distance loss function and an inner layer loss function when the docking posture is the optimal docking posture; the distance loss function is used to characterize the mapping relationship between the first difference and the parameters of the protein complex structure prediction model, the first difference being the difference between the predicted distance map and the true distance map; the inner layer loss function is used to characterize the mapping relationship between the second difference and the parameters of the protein complex structure prediction model and the docking posture;The second difference is the difference between the distance between the reference atom included in each amino acid residue of the first protein and the reference atom included in each amino acid residue of the second protein in the protein complex and the predicted distance map when the first protein and the second protein are docked into a protein complex in a docking posture; it is also used to update the current training data; it is also used to repeatedly perform the steps of obtaining the real distance map corresponding to the current training data, obtaining the input features corresponding to the current training data, model prediction, determining the optimal docking posture of the first protein and the second protein under the current parameters, updating the current parameters according to the outer layer loss, and updating the current training data until the number of cycles reaches the cycle number threshold or the protein complex structure prediction model converges, and a trained protein complex structure prediction model is obtained; wherein the initial value of the current parameter is the initial parameter of the protein complex structure prediction model. ; 9. A protein complex structure prediction device, characterized in that: The device comprises: an acquisition module, used to acquire input data, wherein the input data includes a first target amino acid sequence of a first target protein, a second target amino acid sequence of a second target protein, monomer three-dimensional structure information of the first target protein, and monomer three-dimensional structure information of the second target protein; A first determination module, configured to determine an input feature corresponding to the input data according to the first target amino acid sequence, the second target amino acid sequence, the monomer three-dimensional structure information of the first target protein, and the monomer three-dimensional structure information of the second target protein; A model prediction module, used for inputting the input features corresponding to the input data into a protein complex structure prediction model trained by the method according to any one of claims 1 to 6 to obtain a prediction result; the prediction result includes a prediction distance map; the prediction distance map is a two-dimensional matrix, and the elements in the i-th row and j-th column of the prediction distance map represent the distance between the i-th reference atom of the first target protein and the j-th reference atom of the second target protein when the first target protein and the second target protein form a protein complex, as predicted by the protein complex structure prediction model; A second determination module is used to determine the optimal docking posture of the first target protein and the second target protein, wherein the docking posture of the first target protein and the second target protein is the optimal docking posture. When the docking posture of the first target protein and the second target protein is a protein complex, the distance between the reference atom included in each amino acid residue of the first target protein and the reference atom included in each amino acid residue of the second target protein in the protein complex is closest to the predicted distance map; A structure determination module is used to use the three-dimensional structure of the protein complex obtained by docking the first target protein and the second target protein in the optimal docking posture as the predicted three-dimensional structure of the protein complex composed of the first target protein and the second target protein.

10. A computing device comprising: a memory, a processor and a computer program stored on the memory, The processor is configured to execute the computer program to implement the steps of the method according to any one of claims 1 to 6.

11. A non-transitory computer-readable storage medium having a computer program stored thereon, wherein: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

12. A computer program product comprising computer instructions, wherein: When the computer instructions are executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

Images