Method and device for determining protein structure

Abstract

The present invention discloses a method and apparatus for determining protein structure. First structural data is obtained, where the first structural data is modeled based on cryo-electron microscopy observation data. Second structural data is obtained, where the second structural data is predicted based on the amino acid sequence. The protein structure is determined based on the first and second structural data. This method effectively combines the advantages of two different protein structure prediction methods and overcomes the shortcomings of each method.

Description

Technical Field

[0001] The present disclosure relates to the field of bioinformatics, and in particular to a method and device for determining protein structure. Background Art

[0002] A protein's function is determined by its structure, which is determined by its unique amino acid sequence and its three-dimensional (3D) arrangement. Researchers can draw conclusions about a protein's function based solely on its molecular structure. This insight can be used to develop new vaccines and drugs, as viral fusion proteins play a central role in the invasion of host cells. To prevent COVID-19, for example, researchers are trying to develop vaccines and drugs targeting these fusion proteins. Structural information about fusion proteins is crucial for researchers to predict their behavior and ultimately find suitable vaccines.

[0003] Recent advances in protein structural modeling are based on observations using cryo-electron microscopy (cryo-EM). Cryo-EM enables researchers to capture 3D images of macromolecules at near-atomic resolution. Due to its improved imaging quality and model-building efficiency, the technique has gained popularity in recent years as an alternative to structure-determination methods such as X-ray crystallography.

[0004] Currently, it is important to deploy cryo-EM alongside X-ray crystallography to support the search for drugs and vaccines to combat the current COVID-19 pandemic.To deduce the structure of a protein from a 3D cryo-EM map, researchers currently have to either manually fit atoms or resort to existing template-based or homology modeling methods.

[0005] Therefore, there is an urgent need for a method to automatically and accurately determine the molecular structure of proteins from cryo-EM images. Summary of the Invention

[0006] A technical problem to be solved by the present disclosure is to provide a solution that can automatically and accurately determine the molecular structure of a protein from a cryo-electron microscopy image.

[0007] According to a first aspect of the present disclosure, a method for determining a protein structure is provided, comprising: obtaining first structure data, the first structure data being obtained by modeling the protein structure based on cryo-electron microscopy observation data of the protein; obtaining second structure data, the second structure data being obtained by predicting the protein structure based on the amino acid sequence; and determining the protein structure based on the first structure data and the second structure data.

[0008] Optionally, the step of determining the protein structure based on the first structure data and the second structure data includes: optimizing the main chain and / or residues in the first structure data using the second structure data; and determining the optimized first structure data as the protein structure.

[0009] Optionally, the step of optimizing the main chain in the first structure data using the second structure data includes: using the main chain in the second structure data to fill in the main chain atoms missing in the main chain in the first structure data; and / or using the main chain in the second structure data to correct the position of the helical structure existing in the main chain in the first structure data.

[0010] Optionally, the step of optimizing the residues in the first structure data using the second structure data includes: comparing the second residue structure in the second structure data corresponding to the same residue position with the first residue structure in the first structure data to determine the degree of matching between the second residue structure and the first residue structure; if the degree of matching is lower than a first threshold, replacing the residue structure of the inconsistent part in the first residue structure with the residue structure corresponding to the inconsistent part in the second residue structure.

[0011] Optionally, before using the second structure data to optimize the main chain and / or residues in the first structure data, the method also includes: matching the main chain in the first structure data with the main chain in the second structure data, and determining the correspondence between the main chain in the first structure data and the main chain in the second structure data, so as to optimize the main chain and / or residues in the first structure data using the second structure data based on the correspondence.

[0012] Optionally, the step of obtaining the first structural data includes: dividing the cryo-electron microscopy observation data into tensors of a predetermined size; inputting the tensors into multiple branch neural networks respectively to obtain prediction results output by multiple branch neural networks; obtaining the first structural data based on the prediction results output by the multiple branch neural networks, wherein the multiple branch neural networks respectively correspond to one of a plurality of protein structure prediction indicators, and the branch neural network is used to obtain the prediction results of the protein structure prediction indicators corresponding to the branch neural network at different positions in the space corresponding to the tensor.

[0013] Optionally, the plurality of branch neural networks include: a first branch neural network for predicting atom types, the atom types including C α atoms, common C atoms, N atoms, O atoms and non-atoms; a second branch neural network for predicting whether it belongs to the main chain; a third branch neural network for predicting the type of amino acids; and a fourth branch neural network for predicting the secondary structure.

[0014] Optionally, the step of obtaining the first structural data based on the prediction results output by the plurality of branch neural networks includes: retaining the main chain atoms having a first confidence greater than or equal to a second threshold value based on the prediction results output by the second branch neural network; determining C among the retained main chain atoms based on the prediction results output by the first branch neural network; α Atom; C-based α The second confidence of the atom and Cα The distance between atoms, connecting C α atoms to obtain one or more fragments; based on other atoms in the main chain atoms, the missing atoms in the fragments are filled in to obtain a preliminary construction result of the main chain, which includes one or more main chains; based on the prediction results of the third branch neural network and the prediction results of the first branch neural network, the residues connected to the main chain are determined; based on the prediction results of the fourth branch neural network, the secondary structure of the polypeptide chain composed of the main chain and the residues connected to the main chain is determined.

[0015] Optionally, the method further comprises: merging all fragments into one or more main chains using a depth-first search algorithm; correcting the merged main chains to remove residue paths, loop paths, and terminal node paths in the main chains, wherein a residue path refers to a C-terminal node with a connection number greater than 2. α A path in an atom has a path length less than 3, a loop path refers to different paths ending at the same end point, and an end node path refers to a path with a path length less than 3 and the end point is the end of the peptide chain.

[0016] According to a second aspect of the present disclosure, a device for determining a protein structure is provided, comprising: a first acquisition module for acquiring first structure data, wherein the first structure data is obtained by modeling the protein structure based on cryo-electron microscopy observation data; a second acquisition module for acquiring second structure data, wherein the second structure data is obtained by predicting the protein structure based on the amino acid sequence; and a determination module for determining the protein structure based on the first structure data and the second structure data.

[0017] According to a third aspect of the present disclosure, a computing device is provided, comprising: a processor; and a memory on which executable code is stored, and when the executable code is executed by the processor, the processor executes the method described in the first aspect above.

[0018] According to a fourth aspect of the present disclosure, a computer program product is provided, comprising an executable code. When the executable code is executed by a processor of an electronic device, the processor is caused to execute the method as described in the first aspect above.

[0019] According to a fifth aspect of the present disclosure, a non-transitory machine-readable storage medium is provided, on which executable code is stored. When the executable code is executed by a processor of an electronic device, the processor executes the method described in the first aspect above.

[0020] Therefore, the present invention obtains first structure data obtained by predicting the protein structure based on protein cryo-electron microscopy observation data, and second structure data obtained by predicting the protein structure based on the amino acid sequence of the protein, and determines the protein structure based on the first structure data and the second structure data, so that the determined protein structure can effectively combine the advantages of two different protein structure prediction methods and make up for the shortcomings of each single method. BRIEF DESCRIPTION OF THE DRAWINGS

[0021] The above and other objects, features and advantages of the present disclosure will become more apparent through a more detailed description of exemplary embodiments of the present disclosure with reference to the accompanying drawings, wherein like reference numerals generally represent like components in the exemplary embodiments of the present disclosure.

[0022] Figure 1 A schematic diagram showing the principle of the method for determining protein structure of the present disclosure is shown.

[0023] Figure 2 A schematic diagram showing the principle of a method for predicting protein structure based on cryo-electron microscopy observation data to obtain first structure data according to an embodiment of the present disclosure is shown.

[0024] Figure 3 A schematic diagram showing the principle of the method for predicting protein structure based on amino acid sequence.

[0025] Figure 4 A schematic diagram of the principle of a protein structure joint modeling method according to an embodiment of the present disclosure is shown.

[0026] Figure 5 A schematic structural diagram of an apparatus for determining protein structure according to an embodiment of the present disclosure is shown.

[0027] Figure 6 A schematic structural diagram of a computing device according to an embodiment of the present disclosure is shown. DETAILED DESCRIPTION

[0028] The preferred embodiments of the present disclosure will be described in more detail below with reference to the accompanying drawings. Although preferred embodiments of the present disclosure are shown in the accompanying drawings, it should be understood that the present disclosure can be implemented in various forms and should not be limited by the embodiments described herein. Rather, these embodiments are provided to make the present disclosure more thorough and complete, and to fully convey the scope of the present disclosure to those skilled in the art.

[0029] Existing tools such as Rosetta (comprehensive software for simulating macromolecular structures), MAINMAST (main-chain model tracing from a spanning tree), and Phenix (protein crystallography software) can only determine fragments of protein complexes or require extensive manual processing steps. Because cryo-EM can capture multiple large proteins in a single study, designing a method that can fully automate and efficiently determine protein structure is critical to improving data processing throughput and accelerating drug development.

[0030] The composition of a protein can be a single polypeptide chain or a complex composed of multiple polypeptide chains. The structure of a single polypeptide chain is divided into the main chain and residues. The main chain is composed of multiple peptide bonds (stable structures formed by adjacent amino acids) and adjacent carbon atoms (i.e. C α Atoms) are connected front and back to form peptide units. In addition to the main chain structure, each C α The remaining structure of an amino acid attached to an atom is called a residue.

[0031] For the structure of a single polypeptide chain, protein cryo-electron microscopy (i.e., cryo-EM) data are based on actual observations of real samples. Therefore, the overall authenticity and accuracy of protein structure modeling based on this data are high. However, due to limitations in observation noise and model expressiveness, there are often drawbacks such as misjudgments of amino acids or omissions of atoms. This high accuracy corresponds to smaller errors in atomic position modeling.

[0032] Methods for predicting protein structure based on amino acid sequences (e.g., sequence prediction technologies like Alphafold2) often have better recall or coverage due to the availability of complete amino acid sequence information. However, since they are not based on actual sample observations, atomic positions may be distorted or have large errors. High recall or coverage corresponds to modeling as many atoms as possible.

[0033] In addition, for complex structures composed of multiple main chains, the above two methods also face the following new challenges: the modeling method based on cryo-electron microscopy observation data can obtain the basic building blocks of all sequences, such as peptide bonds and residues, but it is easy to classify these basic building blocks into the wrong sequence; the method based on amino acid sequences can only obtain the modeling results of multiple sequences separately, and cannot model the relative position information between multiple peptide chains.

[0034] In view of the respective advantages and disadvantages of the above two methods, the present disclosure makes a targeted fusion design for the above two methods to effectively combine the advantages of the above two methods and make up for the shortcomings of each single method.

[0035] Figure 1A schematic diagram showing the principle of the method for determining protein structure of the present disclosure is shown.

[0036] like Figure 1 As shown, the present disclosure may first acquire the first structure data and the second structure data.

[0037] The first structural data is structural data obtained by modeling the protein structure based on cryo-electron microscopy observation data of the protein. The second structural data is structural data obtained by predicting the protein structure based on the amino acid sequence of the protein. The present disclosure does not restrict the order in which the first structural data and the second structural data are obtained.

[0038] Pre-generated first structure data can be obtained externally, or the protein structure can be modeled based on cryo-electron microscopy observation data to obtain first structure data generated in real time. Correspondingly, pre-generated second structure data can be obtained externally, or the protein structure can be predicted based on the amino acid sequence to obtain second structure data generated in real time.

[0039] After obtaining the first and second structure data, the protein structure can be determined based on the first and second structure data. Determining the protein structure based on the first and second structure data involves fusing the structure data obtained using the two prediction methods to compensate for the shortcomings of each method, allowing the fused data to accurately represent the protein structure.

[0040] The method for predicting protein structure based on cryo-electron microscopy data can be referred to as "Method 1." The method for predicting protein structure based on the amino acid sequence can be referred to as "Method 2." Specifically, the prediction results of Method 1 (i.e., the first structure data) and the prediction results of Method 2 (i.e., the second structure data) can be fused using various fusion methods. For example, weighted averaging can be used for fusion.

[0041] Considering the high overall fidelity of the modeling results from Method 1 and the fact that Method 2 precisely compensates for the shortcomings of Method 1, the first structural data obtained using Method 1 can be used as the data to be optimized, and the second structural data obtained using Method 2 can be used to optimize the first structural data. This allows the optimized first structural data to effectively combine the advantages of the two aforementioned methods and compensate for the shortcomings of each method. The optimized first structural data can accurately represent the protein structure, and therefore can be determined as the protein structure.

[0042] The following further explains the generation of the first structure data, the generation of the second structure data, and the fusion of the first structure data and the second structure data.

[0043] 1. Generation of the First Structure Data

[0044] Modeling can be performed based on U-Net (an algorithm that uses a fully convolutional network for semantic segmentation) or cascaded CNNs (Convolutional Neural Networks) to obtain a predictive model that can predict protein structure based on cryo-electron microscopy observation data, and the protein structure can be determined using the output of the predictive model.

[0045] As an example, a custom deep convolutional neural network can be constructed from multiple branched neural networks. Each branched neural network corresponds to one of several protein structural metrics. Each branched neural network is used to obtain predictions for the protein structural metrics corresponding to each branched neural network at different locations in the space corresponding to the input tensor.

[0046] The cryo-electron microscopy observation data can be divided into tensors of a predetermined size (such as 64*64*64), which can be input into these multiple branch neural networks respectively to obtain the prediction results output by these multiple branch neural networks.

[0047] Each branched neural network is used to predict a protein structural indicator. The sum of the prediction results output by each of the multiple branched neural networks constitutes structural data capable of characterizing the protein structure. Therefore, first structural data can be obtained based on the prediction results output by these multiple branched neural networks. For example, the prediction results output by these multiple branched neural networks can be fused to obtain first structural data capable of characterizing the protein structure.

[0048] Figure 2 A schematic diagram showing the principle of a method for predicting protein structure based on cryo-electron microscopy observation data to obtain first structure data according to an embodiment of the present disclosure is shown. Figure 2 The method shown is also the method 1 mentioned above.

[0049] like Figure 2 As shown in the figure, a deep convolutional neural network consisting of four branch neural networks can be customized. These four branch neural networks can be respectively called the first branch neural network (U-Net1), the second branch neural network (U-Net2), the third branch neural network (U-Net3), and the fourth branch neural network (U-Net4).

[0050] These four U-Nets are used to perform four levels of end-to-end prediction.

[0051] U-Net1 is used to predict the atomic type of each position in the space corresponding to the input tensor. The atomic type can include C α There are five types: atoms, ordinary C atoms, N atoms, O atoms and non-atoms.

[0052] U-Net2 is used to predict whether each position in the space corresponding to the input tensor belongs to the main chain. The prediction categories can specifically include belonging to the main chain, belonging to the residue, and belonging to the non-protein structure.

[0053] U-Net3 is used to predict the amino acid type at each position in the space corresponding to the input tensor. The amino acid type can include 21 categories, including 20 amino acids and non-amino acids.

[0054] U-Net4 is used to predict the secondary structure of each position in the space corresponding to the input tensor, which can include four categories: helix, sheet, loop, and no structure.

[0055] After the cryo-EM data is divided into tensors of predetermined sizes, they are fed into four U-Nets, which are responsible for predicting the presence of protein atoms, atom types, amino acid types, and secondary structures such as helices and tiled structures at each position.

[0056] Based on the prediction results of the above four U-Nets, information fusion can be performed to obtain the preliminary modeling results of the protein structure, that is, the first structure data mentioned above.

[0057] An exemplary process of fusing the prediction results of the above four U-Nets to obtain the first structure data is as follows.

[0058] Based on the prediction results output by U-Net2, we can determine the first confidence level of each position in the space belonging to the main chain. Positions with a first confidence level greater than or equal to a second threshold (e.g., 0.5) are considered main chain positions, and the atoms at those main chain positions are considered main chain atoms. Therefore, based on the prediction results output by U-Net2, we can retain main chain atoms with a first confidence level greater than or equal to the second threshold. This way, only unconnected main chain fragments are retained throughout the space.

[0059] The prediction results based on the U-Net1 output can further determine (i.e. locate) the C atoms in the retained main chain α atom.

[0060] C-based α The second confidence of the atom and C α The distance between atoms can connect C α The atom gets one or more fragments. Specifically, U-Net1 will output that the atom belongs to C α The confidence of the atom (i.e., the second confidence), so the C with higher confidence (such as greater than or equal to the second threshold) can be αAtoms are connected into fragments as the basic components of the main chain. As an example, a fragmented main chain can be constructed based on the path walking method. The path walking method is a type of tabu search algorithm. The specific process is as follows: First, C can be sorted in descending order of confidence. α Atom sorting; then select an initial C in sequence α Atom; according to the proximity of the initial C α Other C atoms α Atom confidence and distance are used for searching and fragment construction, and exceeding twist and search radius thresholds are used as stopping signals.

[0061] In connection C α After obtaining one or more fragments from atoms, the missing atoms in the fragments can be filled based on other atoms in the main chain atoms to obtain a preliminary main chain construction result, which includes one or more main chains.

[0062] Based on the predictions from U-Net3 and U-Net1, residues connected to the backbone can be determined. Protein modeling algorithms such as Scwrl4 and OPUS-Rota3 can be used to fill in residues connected to the backbone and missing atoms in the backbone.

[0063] After constructing the main chain and the residues connected to the main chain, the secondary structure of the polypeptide chain composed of the main chain and the residues connected to the main chain can be further determined based on the prediction results of U-Net4.

[0064] In this way, the first structural data capable of characterizing the protein structure can be obtained.

[0065] The fragments constructed during the generation of the first structural data may be inaccurate and may contain false positive fragments that are actually residues, as well as incorrect shortcuts. Therefore, further optimization can be performed. The optimization process is as follows.

[0066] First, you can connect C α All fragments obtained from the atoms are sequentially merged and connected into one or more complete main chains using depth-first search. Each merged main chain is then corrected (i.e., optimized) to remove residue paths, loop paths, and terminal node paths in the main chain.

[0067] Residue paths are C residues with a connection number greater than 2. α A path in an atom with a path length less than 3. Normal C α Atoms will have 1-2 connections, if C α The number of atoms connected is greater than 2, then the C αA depth-first path search is performed starting from each connection. If some of the searched paths have a length less than 3, while others are much longer than that, then the paths with a length less than 3 can be considered as incorrectly introduced residue paths and need to be deleted.

[0068] Loop paths are different paths that end at the same end point. Similar to residue path removal, when removing residues from the same C α If multiple different path searches starting from an atom end at the same endpoint, it can be considered a loop and needs to be deleted.

[0069] The end node path is a path with a path length less than 3 and a terminal end at the end of the peptide chain. If a search path length is less than 3 and the terminal end is the end of the peptide chain, the path is deleted.

[0070] 2. Generation of Second Structure Data

[0071] Using technologies such as AlphaFold, protein structure can be predicted based on the amino acid sequence of the protein to obtain secondary structure data. Figure 3 A schematic diagram showing the principle of the method for predicting protein structure based on amino acid sequence. Figure 3 The method shown is also the method 2 mentioned above.

[0072] See also Figure 3 The amino acid sequence can be searched separately in a sequence database and a structure database. The sequence database is also known as an MSA (Multiple Sequence Alignment) database, and the structure database is also known as an amino acid distance matrix template library. After obtaining the corresponding expression (embedding), the resulting information is fed into a transformer module (a model based on the attention mechanism). Through multiple iterations, a better structure prediction result (i.e., the second structure data) is obtained.

[0073] Method 2 requires the construction of an extremely large library of sequences and templates. Essentially, predictions require the MSA of an entire protein family, along with cross-correlation and covariance information. Therefore, only a subset of residues in the human proteome are reliable. Protein structure is dynamic, and most proteins change in response to their environment. Without actual observational data, Method 2 cannot predict these structural changes. Furthermore, modeling the structure of protein complexes is often necessary, but Method 2 can only provide sequence-level predictions.

[0074] Therefore, the second structural data is the structural data of each polypeptide chain constituting the protein.

[0075] 3. Fusion of the First Structure Data and the Second Structure Data

[0076] The parts of the first structure data that need to be optimized using the second structure data mainly include the main chain and residues. Therefore, the second structure data can be used to optimize the main chain and / or residues in the first structure data.

[0077] Before optimizing the backbone and / or residues in the first structural data using the second structural data, it is necessary to match the backbone in the first structural data with each backbone in the second structural data (i.e., overall structural alignment) and determine the correspondence between the backbone in the first structural data and the backbone in the second structural data, so as to optimize the backbone and / or residues in the first structural data using the second structural data based on the correspondence. The correspondence reflects the pairing relationship between the backbone in the first structural data and the backbone in the second structural data, as well as the specific structural correspondence between each pair of matching backbones. For example, it can be a pairwise pairing relationship that maximizes the overall structural match of all backbones, as well as the correspondence between the CN bonds in each pair of backbones.

[0078] The first structure data may include multiple main chains or fragments corresponding to different amino acid sequences (caused by modeling errors). For each main chain (referred to as the first main chain for ease of distinction) or fragment in the first structure data, a main chain (referred to as the second main chain for ease of distinction) may be selected from the second structure data in sequence, and the first main chain or fragment may be matched with the second main chain in terms of the overall structure. Then, based on the matching result (i.e., the second main chain and the first main chain fragment or fragment that match in terms of the overall structure), the two may be aligned at the matching position to determine the corresponding relationship between the first main chain or fragment and the second main chain.

[0079] As an example, the first main chain or fragment and the second main chain can be matched one by one according to the position of the CN bond. For example, for each CN bond position in the second main chain, the Qprob algorithm (a single model quality assessment method) can be used to evaluate the overall structure match between the first main chain or fragment and the second main chain, and the second main chain with the maximum match is selected as the paired main chain, and the position of the CN bond with the maximum match is selected as the actual position of the match between the first main chain or fragment and the second main chain.

[0080] 1. Main chain completion and spiral optimization

[0081] The main chain in the first structure data may refer to the main chain structure after optimization processes such as path merging and residue removal, loop removal, and terminal node removal. The optimized main chain structure in the first structure data already has a high degree of accuracy, but there may be missing atoms. Therefore, the results of method 2 (i.e., the second structure data) can be used to fill in the missing atoms (i.e., main chain atoms) in the main chain in the first structure data.

[0082] As an example, based on the correspondence between the first structure data and the second structure data, the first structure data and the second structure data can be matched to the same pose (aligned) using a transformation matrix, and then the missing main chain part can be filled with the second structure data.

[0083] An important part of the secondary structure of protein is the helical structure. The helix is ​​a relatively complex structure. Due to observation noise or model expression capability limitations, the accuracy of directly fitting the positions of atoms on the helix using the prediction results of U-Net4 of method 1 is poor. However, method 2 can give accurate results based on the template for complex structures. Therefore, the position of the helical structure in the main chain of the first structure data obtained by method 1 can be corrected by using the aligned second structure data obtained by method 2 (i.e., the initial predicted position of the helix), that is, the C at the helix determined by method 1 can be corrected. α The atoms are moved to the helix determined by method 2.

[0084] 2. Residue Optimization

[0085] The second residue structure in the second structure data corresponding to the same residue position can be compared with the first residue structure in the first structure data to determine the degree of match between the second residue structure and the first residue structure. If the degree of match is lower than a first threshold (e.g., 0.8), the residue structure of the inconsistent portion in the first residue structure is replaced with the residue structure corresponding to the inconsistent portion in the second residue structure.

[0086] As an example, the structures of each residue position in the aligned results of Method 1 and Method 2 (i.e., the first and second structural data) can be compared. If the match exceeds 0.8, the two are considered to be consistent, and the observation-based modeling results of Method 1 can be directly used. If the match is less than 0.8, the residue structures at that location are considered to conflict. Based on the maximum match, the results of Method 1 are used for the atoms that agree, and the results of Method 2 are used to fill in the residue atomic structure for the remaining atoms. Considering that the results of Method 1 are likely to contain missing residue structures, in this case, the Scrrl4 algorithm can be used to fill in the atomic structure and adjust the pose to connect to the atoms at this location, and then perform the above residue optimization steps.

[0087] Figure 4 A schematic diagram of the principle of a joint protein structure modeling method that integrates a modeling method based on cryo-electron microscopy observation data (i.e., method 1) and a prediction method based on amino acid sequence (i.e., method 2) according to an embodiment of the present disclosure is shown.

[0088] like Figure 4As shown, amino acid sequences are effective structural controls. Cryo-EM data allow for modeling of all observed sequences. Method 2, based purely on sequence files, results in separate modeling of each sequence.

[0089] First, sequence matching can be achieved between the results obtained by these two methods using a dynamic programming algorithm. In this step, the observation-based method helps to couple the independent sequence results obtained by sequence prediction according to their actual structure. Sequence prediction, on the other hand, effectively denoises the modeling results, which contain varying noise, ultimately resulting in well-matched sequence results. Sequence matching, in other words, aligns the results obtained by the two methods based on their overall structure.

[0090] Next, the backbone is the primary structure-determining component. The two perform complementary corrections to determine the positions of the primary amino acids that bind to the backbone. Here, sequence prediction results are still responsible for recall, while observation-based results are responsible for determining the backbone's position based on actual data.

[0091] Finally, the residue optimization module performs a joint optimization based on observation-based and sequence prediction methods. The sequence prediction method fills in missing atoms, while the observation-based method adjusts conflicting residues in the sequence prediction results. Ultimately, the optimized result is obtained.

[0092] In summary, the joint modeling approach disclosed herein exploits the high recall characteristics of sequence prediction-based methods, but lacks detailed modeling of spatial locations and inter-sequence coupling, which can be corrected using observation-based methods. Furthermore, the shortcomings of cryo-electron microscopy-based methods, which can model realistic spatial structures but often struggle to handle noise interference, can be corrected using sequence prediction-based methods. This effectively leverages the respective strengths of the two approaches to achieve complementary results.

[0093] The method for determining protein structure disclosed herein may also be implemented as a device for determining protein structure. Figure 5 The following is a schematic diagram of a device for determining protein structure according to an embodiment of the present disclosure. The functional units of the device for determining protein structure can be implemented by hardware, software, or a combination of hardware and software that implements the principles of the present disclosure. It will be understood by those skilled in the art that Figure 5 The functional units described can be combined or divided into sub-units to implement the principles of the above invention. Therefore, the description herein can support any possible combination, division, or further limitation of the functional units described herein.

[0094] The following is a brief description of the functional units that a device for determining protein structure may have and the operations that each functional unit may perform. For the details involved, please refer to the relevant description above and will not be repeated here.

[0095] See also Figure 5 The apparatus 500 for determining protein structure includes a first acquisition module 510, a second acquisition module 520, and a determination module 530. The first acquisition module 510 is used to acquire first structure data, which can be obtained by modeling the protein structure based on cryo-electron microscopy observation data of the protein; the second acquisition module 520 is used to acquire second structure data, which can be obtained by predicting the protein structure based on the amino acid sequence of the protein; and the determination module 530 is used to determine the protein structure based on the first structure data and the second structure data.

[0096] The determination module 530 may include an optimization module and a determination submodule. The optimization module is configured to optimize the main chain and / or residues in the first structure data using the second structure data. The determination submodule is configured to determine the optimized first structure data as a protein structure.

[0097] The optimization module can use the main chain in the second structure data to fill the main chain atoms missing in the main chain in the first structure data; and / or use the main chain in the second structure data to correct the position of the helical structure in the main chain in the first structure data.

[0098] The optimization module can compare the second residue structure in the second structure data corresponding to the same residue position with the first residue structure in the first structure data to determine the degree of matching between the second residue structure and the first residue structure; if the degree of matching is lower than a first threshold, the residue structure of the inconsistent part in the first residue structure is replaced with the residue structure corresponding to the inconsistent part in the second residue structure.

[0099] The apparatus 500 for determining protein structure may further include a matching module. Before optimizing the main chain and / or residues in the first structure data using the second structure data, the matching module may be used to match the main chain in the first structure data with the main chain in the second structure data to determine the corresponding relationship between the main chain in the first structure data and the main chain in the second structure data, so as to optimize the main chain and / or residues in the first structure data using the second structure data based on the corresponding relationship.

[0100] First acquisition module 510 can divide cryo-electron microscopy observation data into tensors of predetermined sizes, input each of the tensors into multiple branched neural networks, obtain prediction results output by the multiple branched neural networks, and obtain first structure data based on the prediction results output by the multiple branched neural networks. Each of the multiple branched neural networks corresponds to one of multiple protein structure prediction indicators, and the branched neural networks are used to obtain prediction results for the protein structure prediction indicators corresponding to the branched neural networks at different locations in the space corresponding to the tensor. The specific process for generating the first and second structure data can be found in the relevant description above.

[0101] Figure 6 A schematic diagram of the structure of a computing device that can be used to implement the above-mentioned method for determining protein structure according to an embodiment of the present disclosure is shown.

[0102] See also Figure 6 , the computing device 600 includes a memory 610 and a processor 620 .

[0103] The processor 620 may be a multi-core processor or may include multiple processors. In some embodiments, the processor 620 may include a general-purpose main processor and one or more special coprocessors, such as a graphics processing unit (GPU) or a digital signal processor (DSP). In some embodiments, the processor 620 may be implemented using customized circuits, such as application-specific integrated circuits (ASICs) or field-programmable gate arrays (FPGAs).

[0104] The memory 610 may include various types of storage units, such as system memory, read-only memory (ROM), and permanent storage devices. ROM may store static data or instructions required by the processor 620 or other modules of the computer. The permanent storage device may be a readable and writable storage device. The permanent storage device may be a non-volatile storage device that retains stored instructions and data even when the computer is powered off. In some embodiments, the permanent storage device uses a large-capacity storage device (such as a magnetic or optical disk, flash memory) as the permanent storage device. In other embodiments, the permanent storage device may be a removable storage device (such as a floppy disk, optical drive). The system memory may be a readable and writable storage device or a volatile readable and writable storage device, such as dynamic random access memory. The system memory may store some or all instructions and data required by the processor during operation. In addition, the memory 610 may include any combination of computer-readable storage media, including various types of semiconductor memory chips (DRAM, SRAM, SDRAM, flash memory, programmable read-only memory), and magnetic disks and / or optical disks may also be used. In some embodiments, the memory 610 may include a readable and / or writable removable storage device, such as a compact disc (CD), a read-only digital versatile disc (e.g., DVD-ROM, dual-layer DVD-ROM), a read-only Blu-ray disc, an ultra-density optical disc, a flash memory card (e.g., SD card, mini SD card, Micro-SD card, etc.), a magnetic floppy disk, etc. Computer-readable storage media do not include carrier waves and transient electronic signals transmitted wirelessly or wired.

[0105] The memory 610 stores executable codes. When the executable codes are processed by the processor 620 , the processor 620 can execute the above-mentioned method for determining protein structure.

[0106] The method, device and apparatus for determining protein structure according to the present disclosure have been described above in detail with reference to the accompanying drawings.

[0107] In addition, the method according to the present disclosure may also be implemented as a computer program or a computer program product, which includes computer program code instructions for executing the above steps defined in the above method of the present disclosure.

[0108] Alternatively, the present disclosure may also be implemented as a non-transitory machine-readable storage medium (or computer-readable storage medium, or machine-readable storage medium) on which executable code (or computer program, or computer instruction code) is stored. When the executable code (or computer program, or computer instruction code) is executed by a processor of an electronic device (or computing device, server, etc.), the processor is caused to perform the various steps of the above-mentioned method according to the present disclosure.

[0109] Those skilled in the art will further appreciate that the various illustrative logical blocks, modules, circuits, and algorithm steps described in connection with the disclosure herein may be implemented as electronic hardware, computer software, or combinations of both.

[0110] The flowcharts and block diagrams in the accompanying drawings show the possible architecture, functions and operations of the systems and methods according to multiple embodiments of the present disclosure. In this regard, each box in the flowchart or block diagram can represent a module, program segment or part of code, and the part of the module, program segment or code contains one or more executable instructions for realizing the specified logical function. It should also be noted that in some alternative implementations, the functions marked in the box can also occur in an order different from that marked in the accompanying drawings. For example, two consecutive boxes can actually be executed substantially in parallel, and they can sometimes be executed in the opposite order, depending on the functions involved. It should also be noted that each box in the block diagram and / or flowchart, and the combination of the boxes in the block diagram and / or flowchart, can be implemented with a dedicated hardware-based system that performs the specified function or operation, or can be implemented with a combination of dedicated hardware and computer instructions.

[0111] While various embodiments of the present disclosure have been described above, the above descriptions are illustrative, non-exhaustive, and not intended to be limiting of the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is selected to best explain the principles of the embodiments, their practical applications, or improvements to existing technologies, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A method for determining protein structure, comprising: Acquiring first structural data, where the first structural data is obtained by modeling the protein structure based on cryo-electron microscopy observation data; obtaining second structural data, where the second structural data is obtained by predicting the protein structure based on the amino acid sequence; For each CN bond position in the second main chain, the matching degree of the overall structure of the first main chain and the second main chain is evaluated using a single model quality assessment method, the second main chain with the maximum matching degree is selected as the paired main chain, and the position of the CN bond with the maximum matching degree is selected as the actual matching position between the first main chain and the second main chain; wherein the first main chain is the main chain in the first structure data, and the second main chain is the main chain in the second structure data; Using the second main chain in the second structural data to fill in the missing main chain atoms in the first main chain in the first structural data; and / or using the second main chain in the second structural data to correct the position of the helical structure in the first main chain in the first structural data; comparing a second residue structure in the second structure data corresponding to the same residue position with a first residue structure in the first structure data to determine a degree of match between the second residue structure and the first residue structure; If the matching degree is lower than a first threshold, replacing the residue structure of the inconsistent portion in the first residue structure with the residue structure corresponding to the inconsistent portion in the second residue structure; The optimized first structural data is determined as the protein structure.

2. The method according to claim 1, wherein The step of obtaining the first structure data includes: Dividing the cryo-electron microscopy observation data into tensors of predetermined sizes; Inputting the tensors into a plurality of branch neural networks respectively to obtain prediction results output by the plurality of branch neural networks; The first structure data is obtained based on the prediction results output by the multiple branch neural networks, wherein the multiple branch neural networks respectively correspond to one of a plurality of protein structure prediction indicators, and the branch neural network is used to obtain the prediction results of the protein structure prediction indicators corresponding to the branch neural network at different positions in the space corresponding to the tensor.

3. The method according to claim 2, wherein: The multiple branch neural networks include: A first branch neural network for predicting atom types, wherein the atom types include Atom, ordinary atom, atom, Atoms and nonatoms; The second branch neural network is used to predict whether it belongs to the main chain; A third branch neural network for predicting amino acid type; and The fourth branch of the neural network for predicting secondary structure.

4. The method according to claim 3, wherein: The step of obtaining the first structure data based on the prediction results output by the multiple branch neural networks includes: Retaining the main chain atoms having a first confidence greater than or equal to a second threshold based on the prediction result of the second branch neural network; Determine the number of retained main chain atoms based on the prediction results output by the first branch neural network. atom; Based on the The second confidence of the atom and the The distance between atoms, connecting the The atom gets one or more fragments; Based on other atoms in the main chain atoms, fill in the missing atoms in the fragment to obtain a preliminary main chain construction result, wherein the preliminary main chain construction result includes one or more main chains; Determining the residue connected to the main chain based on the prediction result of the third branch neural network and the prediction result of the first branch neural network; Based on the prediction result of the fourth branch neural network, the secondary structure of the polypeptide chain consisting of the main chain and the residues connected to the main chain is determined.

5. The method according to claim 4, further comprising: Merging all of the fragments into one or more backbones using a depth-first search algorithm; Correct the merged main chain to remove the residue path, loop path and terminal node path in the main chain, wherein the residue path refers to the one with a connection number greater than 2. The path length in the atom is less than 3, the loop path refers to different paths ending at the same end point, and the terminal node path refers to a path length less than 3 and the end point is the end of the peptide chain.

6. A device for determining protein structure, comprising: A first acquisition module is used to acquire first structural data, where the first structural data is obtained by modeling the protein structure based on cryo-electron microscopy observation data; A second acquisition module is used to acquire second structure data, where the second structure data is obtained by predicting the protein structure based on the amino acid sequence; a matching module for evaluating the overall structural match between the first main chain and the second main chain using a single model quality assessment method for a CN bond position in each second main chain, selecting the second main chain with the maximum matching degree as the paired main chain, and selecting the position of the CN bond with the maximum matching degree as the actual matching position between the first main chain and the second main chain; wherein the first main chain is the main chain in the first structure data, and the second main chain is the main chain in the second structure data; A determination module is configured to use the main chain in the second structure data to fill in the main chain atoms missing in the main chain in the first structure data; and / or use the main chain in the second structure data to correct the position of the helical structure existing in the main chain in the first structure data; compare the second residue structure in the second structure data corresponding to the same residue position with the first residue structure in the first structure data to determine the degree of match between the second residue structure and the first residue structure; if the degree of match is lower than a first threshold, replace the residue structure of the inconsistent part in the first residue structure with the residue structure in the second residue structure corresponding to the inconsistent part; and determine the optimized first structure data as the protein structure.

7. A computing device comprising: processor; as well as A memory having executable codes stored thereon, which, when executed by the processor, causes the processor to perform the method according to any one of claims 1 to 5.

8. A computer program product comprising executable codes, which, when executed by a processor of an electronic device, causes the processor to perform the method according to any one of claims 1 to 5.

9. A non-transitory machine-readable storage medium having executable codes stored thereon, which, when executed by a processor of an electronic device, causes the processor to perform the method according to any one of claims 1 to 5.

Images