Protein data conversion method

By extracting and labeling protein structure data, converting it into three-dimensional electron density data available in deep learning models, it solves the problem that existing tools cannot process protein structure data, realizes data conversion and labeling for deep learning, and promotes the field of biopharmaceuticals progress.

CN119069030APending Publication Date: 2024-12-03SHENYUAN PHARMACEUTICAL BIOTECHNOLOGY (BEIJING) CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202310644106.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2023-06-01
Publication Date
2024-12-03

AI Technical Summary

Technical Problem

Existing tools cannot convert protein structure data into three-dimensional electron density data required for deep learning models, and cannot add rich annotation information to meet the data needs of machine learning models.

Method used

A protein data conversion method is proposed, including obtaining the original data of the protein, performing specific structure extraction and standardization processing, converting it into deep learning data of electron density mode, and adding labeling information to store it as data morphology available in machine learning models. The method includes extracting the frequency domain information of atomic Cartesian coordinates and diffraction data from PDB and MTZ data, performing standardization processing and annotation, and converting it into PLY and H5 data formats.

Benefits of technology

Converting data from biological research into three-dimensional electron density data available in deep learning has promoted the advancement of machine learning in the biopharmaceutical field, providing rich and accurate data for deep learning model training, satisfying data preprocessing and Mark the requirements.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119069030A_ABST
    Figure CN119069030A_ABST
Patent Text Reader

Abstract

The invention provides a method for converting disclosed protein structure data into protein data for an electron density modal deep learning model. The method comprises two processes of processing and labeling electron density data. According to the method, a large amount of public data in the field of biological research can be converted into three-dimensional electron density data which can be used by deep learning, so that the progress of machine learning in the field of biopharmacy is promoted.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present disclosure relates to protein drug development, and more particularly to a protein data conversion method for protein drug development based on machine learning. Background Art

[0002] In recent years, machine learning, including deep learning, has been widely studied and paid attention to in the field of protein pharmaceuticals. Deep learning models require a large amount of high-quality and diverse protein-related data for training. The history of protein databases can be traced back to the 1980s. For example, the International Protein Structural Database (Protein Data Bank, PDB for short) is a database that specializes in collecting three-dimensional structural data of biological macromolecules such as proteins, nucleic acids and their complexes. PDB can be accessed free of charge via the Internet and is an important resource in structural biology research. Therefore, when deep learning entered the field of protein research, PDB became a natural choice. Various models, such as AlphaFold, RoseTTAFold, etc., have used data based on PDB for training.

[0003] In the PDB database, the experimental data of structure determination are stored together with the atomic three-dimensional coordinate data of the protein structure model. The structural model usually only records the three-dimensional coordinates of the nucleus of each atom in the protein; the experimental data varies depending on the experimental method. For example, the experimental data of three-dimensional reconstruction of cryo-electron microscopy is the electron density function, and the experimental data of X-ray crystallography is the diffraction data that has been processed in a certain way. The electron density function can also be calculated based on the diffraction data. The electron density function gives the electron density value at any given coordinate point in the space where the protein structure is located, which constitutes the electron density data of the structure. Compared with the structural model that only retains the atomic coordinates, the electron density data contains richer and more realistic structural information and interaction information, and is therefore more suitable for the application of deep learning in the field of protein pharmaceuticals. However, the above-mentioned deep learning models do not use electron density information.

[0004] In the traditional field of structural biology, researchers use a series of visualization tools such as Chimera, PyMOL, and Coot to display and adjust electron density data, but these tools cannot generate the three-dimensional electron density data required for deep learning model training. In the field of deep learning, three-dimensional point cloud data with color and depth information is often used in applications such as robotics, computer vision, and three-dimensional modeling, but related visualization and modeling tools such as Meshlab and Open3D cannot directly process electron density data in PDB. In addition, the data required for machine learning generally requires rich annotation information and must undergo data preprocessing. These requirements cannot be met by existing tools.

[0005] Therefore, there is a need to provide a protein data processing tool suitable for machine learning to convert the original protein structure data into protein data for electron density modal deep learning models, especially to add annotation information so that it can be directly applied to machine learning models. Summary of the invention

[0006] This disclosure proposes a method for converting public protein structure data into protein data for electron density modal deep learning models, including two processes: processing and annotation of electron density data. This method can convert a large amount of public data in the field of biological research into three-dimensional electron density data that can be used for deep learning, thereby promoting the progress of machine learning in the field of biopharmaceuticals.

[0007] According to the first aspect of the present disclosure, a protein data conversion method is provided. The method may include: obtaining raw data of the protein; performing specific structure extraction and standardization based on the raw data of the protein to form a specific structure description of the protein; performing data conversion and annotation on the specific structure description of the protein; and storing the converted and annotated data in a data form that can be used by a machine learning model.

[0008] In the method according to the first aspect of the present disclosure, the original data of the protein includes atomic rectangular coordinate system information and diffraction data frequency domain information of the protein.

[0009] Preferably, the atomic rectangular coordinate system information of the protein is derived from PDB data, and the diffraction data frequency domain information of the protein is derived from MTZ data.

[0010] In the method according to the first aspect of the present disclosure, for the atomic rectangular coordinate system information of the protein, the specific structure extraction and standardization processing based on the original data of the protein may include: extracting atomic information; based on the extracted atomic information, identifying the specific structure, surrounding environment and range of the protein, and standardizing the spatial position conversion.

[0011] Preferably, the specific target structure position and surrounding range in each subunit of the protein are determined according to the secondary structure information provided in the PDB data; the rules for segmenting and annotating the protein are set according to user input or preset parameters; and the PDB data are segmented and annotated according to the rules.

[0012] Preferably, the specific target structure may include any identifiable secondary structure, such as an α-helix, a β-fold, a loop region, etc.

[0013] In the method according to the first aspect of the present disclosure, for the frequency domain information of the diffraction data of the protein, the specific structure extraction and standardization processing based on the original protein data may include: performing a transformation from the frequency domain to the spatial domain to form crystal space fractional coordinate data for representing a 3D density image; segmenting the electron density information corresponding to the fractional coordinate data; converting the space group of the segmented data into a specific space group, and representing the original fractional coordinates with the fractional coordinates of the specific space group to obtain the electron density fractional coordinate information.

[0014] Preferably, the crystal space fractional coordinate data used to represent the 3D density image is MRC data.

[0015] Preferably, the electron density information corresponding to the fractional coordinate data is segmented with reference to the specific structure extraction and identification performed on the atomic rectangular coordinate system information of the protein.

[0016] Preferably, the transformation from the frequency domain to the spatial domain is an inverse Fourier transform.

[0017] In the method according to the first aspect of the present disclosure, the data conversion and annotation of the specific structural description of the protein may include: performing rotational isomer identification on the atomic rectangular coordinate system information of the protein to form annotated spatial data.

[0018] In the method according to the first aspect of the present disclosure, the data conversion and annotation of the specific structural description of the protein may include: converting fractional coordinates into Cartesian coordinates; annotating rotational isomers; and transforming the annotated data from the spatial domain to the frequency domain to form annotated frequency domain data.

[0019] Preferably, the Cartesian coordinate data is annotated with respect to rotational isomers with reference to the rotational isomer identification performed on the atomic rectangular coordinate system information.

[0020] Preferably, the transformation from the spatial domain to the frequency domain is a Fourier transform.

[0021] In the method according to the first aspect of the present disclosure, storing the converted and labeled data as a data form usable by the machine learning model may include: storing the labeled spatial data as PLY data; and / or storing the labeled frequency domain data as H5 data.

[0022] According to the second aspect of the present disclosure, a protein data conversion system is provided, which can be implemented by a computer, preferably, can be implemented by executing a computer program to implement the following operations: obtaining raw data of the protein; performing specific structure extraction and standardization processing based on the raw data of the protein to form a specific structural description of the protein; performing data conversion and annotation on the specific structural description of the protein; and storing the converted and annotated data in a data form usable by a machine learning model.

[0023] According to a third aspect of the present disclosure, a method for training a machine learning model for protein drug development is provided. The method may include: using the protein data conversion method according to the first aspect of the present disclosure to convert the original protein data of the protein sample for training into a machine learning training sample; and using the machine learning training sample converted from the original protein data to train the machine learning model for protein drug development.

[0024] According to a fourth aspect of the present disclosure, a method for developing a protein drug based on a machine learning model is provided. The method may include: using the protein data conversion method according to the first aspect of the present disclosure to convert the original protein data of the protein to be tested into input data of the machine learning model; inputting the input data into the machine learning model to obtain an output result of the machine learning model; and realizing protein drug development according to the output result.

[0025] According to a fifth aspect of the present disclosure, a non-transitory computer-readable storage medium is provided for storing a computer program. The computer program includes instructions. When the instructions are executed by a processor of an electronic device, the electronic device implements the protein data conversion method according to the first aspect of the present disclosure.

[0026] According to a sixth aspect of the present disclosure, a computer system is provided. The computer system comprises: a processor, a memory, and a computer program. The computer program is stored in the memory and is configured to be executed by the processor. The computer program comprises instructions for implementing the protein data conversion method according to the first aspect of the present disclosure.

[0027] In addition to the application directions proposed above, the scheme proposed in the present disclosure can be used to perform multi-resolution and multi-dimensional representation of protein electron density. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] The present disclosure will be more fully understood from the following detailed description taken in conjunction with the accompanying drawings, in which like elements are numbered in a similar manner, and in which:

[0029] Figure 1 is a flowchart of a protein data conversion method according to an embodiment of the present disclosure.

[0030] Figure 2 is a more detailed data flow diagram of the protein data conversion method according to an embodiment of the present disclosure.

[0031] Figure 3 is a schematic diagram of a dihedral angle. DETAILED DESCRIPTION

[0032] Unless otherwise defined, technical and scientific terms used in the present disclosure have the same meaning as commonly understood by one of ordinary skill in the art to which the present disclosure belongs.

[0033] The technical solution of the present disclosure is further described in detail below through examples and in conjunction with the accompanying drawings. Unless otherwise specified, the methods, tools and materials of the embodiments described below are conventional methods, tools and products that can be obtained through public channels or purchased on the market. Those skilled in the art of the present disclosure will understand that the methods, tools and materials described below are only exemplary and should not be regarded as limiting the scope of the present disclosure.

[0034] Definitions of Abbreviations and Key Terms

[0035] Protein Data Bank (PDB) database: international protein structure database

[0036] PDB file format: A file format used to store structural information of proteins and other biological molecules. PDB files are stored in text format and include a file header and atomic structure data records, which contain structural information such as atomic coordinates, connections, structural features, temperature factors, etc. of biological molecules.

[0037] MTZ file format: A file format used to store and process X-ray diffraction data, which is usually used for crystal structure analysis and precise atomic location. The data stored in MTZ files usually include crystallographic structure factors (amplitude and phase angle), electron density mapping, precise atomic location and temperature factors, etc.

[0038] MRC file format: A file format commonly used to store electron microscope three-dimensional image data. It is also used to store low-resolution 3D density images, tomography images, and other types of scientific data.

[0039] PLY file format: A file format used to describe the geometry and color information of 3D objects. It is usually used to store 3D scanning data, 3D rendering results, CAD models, etc.

[0040] Rotamer: Rotamers refer to the situation where the amino acid side chains in proteins can exist in different conformations under the free rotation of covalent single bonds. Due to the interaction and energy constraints between atoms in the side chains, some conformations are more stable than others. These more stable conformations are called rotamers, and they have different rotor angles at different positions on the amino acid side chains.

[0041] H5 file format: A file format used to store and manage large amounts of data. It can store multiple types of data, including numbers, images, audio, and video. Since H5 files have a hierarchical structure and support compression and encryption, they are widely used in scientific computing, machine learning, and deep learning. It is one of the important tools for storing and processing large-scale data.

[0042] A protein structural model is a digital model used to describe the three-dimensional structure of a protein. The protein structural model is based on the atomic coordinate data of the protein and contains the connection information of all covalent bonds and coordination bonds therein. Protein structural models are usually constructed using X-ray diffraction, cryo-electron microscopy, nuclear magnetic resonance or other techniques. Protein structural models are crucial to understanding how proteins perform their biological functions and can help researchers design new drugs and develop treatments.

[0043] Existing artificial intelligence (AI) algorithms are generally based on protein structure models and sequence models when processing protein data.

[0044] The protein structure model represents the real physical form of the protein in a modeled way, simplifying the complex electron cloud of interactions and mutual influences between atoms into the coordinates of the nucleus at the center of the atom. Various non-covalent interactions between atoms, such as hydrogen bonds, ionic bonds, hydrophobic interactions, etc., can only be judged based on the distance between the nuclei and the type of atoms. The protein structure model is derived from real structural biology experimental data, such as X-ray crystallography, cryo-electron microscopy 3D reconstruction, or nuclear magnetic resonance and other methods, but the structural model is not equivalent to the experimental data, but only one of the models that best matches the experimental data.

[0045] The protein sequence model is a model that only describes the one-dimensional information of the protein. The vast majority of natural proteins are composed of only one one-dimensional peptide chain, and the amino acid sequence that makes up the peptide chain (i.e., the sequence model of the protein) can be used for deep learning training. Various AI models that imitate natural language processing can use protein sequence models to study biological problems such as the composition rules of protein sequences, the correspondence between sequences and their biological functions, and the correspondence between sequences and three-dimensional protein structures. Protein sequence models can be directly generated from genome sequencing results, but they do not necessarily have corresponding protein structure models from experimental data.

[0046] Both structural models and sequence models simplify protein structures at different levels and cannot accurately express protein structural information or fully represent the real structure. The specific problems are as follows:

[0047] 1) The electron functional cannot be calculated when the number of atoms is large, so the coordinates of the nucleus alone are not enough to calculate the complex electron cloud that interacts with each other inside the protein molecule;

[0048] 2) The experiment measures the electron density, while the structural model is a set of atomic coordinates that best matches it through calculation, so there must be an irreconcilable difference between the two.

[0049] 3) When analyzing the interaction between atoms based on the structural model, since only the coordinate information of the atomic nucleus is available, the interaction can only be analyzed based on the distance between atoms, which is not accurate and may miss some complex or weak interactions that are not yet mastered by the current knowledge system;

[0050] 4) The electron density information contains some complex structural features, such as the multiple conformational phenomena of some amino acid residues, which may be unintentionally ignored or intentionally deleted when processing using atomic coordinates.

[0051] The electron density function of a protein refers to the electron cloud distribution function of all the atoms that make up the protein, which can be measured by experimental methods such as X-ray diffraction or cryo-electron microscopy three-dimensional reconstruction. The electron density function, combined with the known chemical composition information of the protein (i.e., the amino acid sequence information determined by the gene), can roughly determine the position coordinates of most of the important atoms in the protein, as well as the activity level of these atoms and other information. Based on this information, a structural model of the protein can be constructed, and the function of the protein, the interaction between the protein and the ligand, etc. can be analyzed accordingly. The electron density contains rich structural information and interaction information, which cannot be fully represented by the structural model. Therefore, using electron density information as learning data for artificial intelligence algorithms will provide richer and more accurate information than structural models or sequence models, and will help the algorithm learn more accurate information about protein molecular structure, conformation, bonds, and interactions.

[0052] Protein electron density visualization refers to the process of converting the electron density data of a protein into an image. Visualization techniques usually include computer graphics methods for generating three-dimensional images. The purpose of protein electron density visualization is to make the structure and shape of the protein visible and easy to understand for humans, thereby facilitating the study of the biological functions and interactions of proteins. Common visualization techniques include surface rendering, heat maps, transparency images, etc. The result of protein electron density visualization is usually a realistic three-dimensional image, which can help biologists understand the conformation and dynamics of proteins. Structural biology researchers use a series of visualization tools such as Chimera, PyMOL, and Coot to better understand electron density images, but, as mentioned earlier, these tools cannot directly produce annotated three-dimensional electron density data required for deep learning model training.

[0053] A 3D point cloud is a mathematical model that represents 3D spatial information. It consists of discrete points, each of which represents a 3D coordinate (x, y, z). Point cloud is a very flexible way of representing data and can be used to record the 3D shape, appearance, and surface features of an object. 3D point cloud data can be obtained from a variety of data sources, such as laser scanning, cameras, or other sensors. In the field of deep learning, 3D point cloud data with color and depth information are often used in applications such as robotics, computer vision, and 3D modeling, but related tools such as Meshlab and Open3D cannot directly process 3D electron density data of proteins. In addition, the data required for machine learning generally requires rich annotation information and must undergo data preprocessing. These requirements cannot be met by existing tools.

[0054] Taking the above factors into consideration, the present disclosure proposes a solution for protein data conversion.

[0055] From an overall perspective, Figure 1 A flowchart of a protein data conversion method according to an embodiment of the present disclosure is shown.

[0056] like Figure 1 As described in the foregoing, the protein data conversion method 100 according to the embodiment of the present disclosure starts at step S110, in which the original data of the protein is obtained.

[0057] Specifically, the original data of the protein may include the atomic rectangular coordinate system information and diffraction data frequency domain information of the protein. For example, as described above, the atomic rectangular coordinate system information of the protein may be derived from PDB data. Similarly, the diffraction data frequency domain information of the protein may be derived from MTZ data.

[0058] Next, in step S120, specific structure extraction and standardization are performed based on the original data of the protein to form a specific structure description of the protein.

[0059] Then, in step S130, data conversion and annotation are performed on the specific structural description of the protein.

[0060] Finally, in step S140, the converted and annotated data is stored in a data form that can be used by the machine learning model.

[0061] For example, according to a preferred embodiment of the present disclosure, spatial data converted from atomic rectangular coordinate system information and annotated can be stored as PLY data. On the other hand, frequency domain data converted from diffraction data frequency domain information and annotated can be stored as H5 data. As mentioned above, both PLY and H5 data formats can be directly used for machine learning, especially deep learning models.

[0062] Figure 2 is a more detailed data flow diagram of the protein data conversion method according to an embodiment of the present disclosure. Figure 2 The protein data conversion method according to an embodiment of the present disclosure is described in more detail.

[0063] Download source data of PDB database

[0064] As mentioned earlier, in Figure 1 In step S110 of the protein data conversion method 100 shown, the original data of the protein is obtained. According to a preferred embodiment of the present disclosure, the original data of the protein can be downloaded from an open source (eg, PDB database).

[0065] like Figure 2As shown in , source data can be downloaded from the public PDB database, with proteins as units, and each sample corresponds to two files, namely, a PDB data format file and an MTZ data format file. The PDB file contains atomic coordinates and various annotation information, and is an approximate structural model of the protein. The structure analyzed by X-ray crystallography comes with an MTZ file, which contains experimentally obtained diffraction data, as well as information such as the phase calculated after correction. The electron density information describing the protein structure can be calculated from the data in the MTZ file. As previously mentioned, according to an embodiment of the present disclosure, the atomic rectangular coordinate system information of the protein can be derived from PDB data. Similarly, the frequency domain information of the diffraction data of the protein can be derived from the MTZ data.

[0066] Subsequent processing, including specific structure extraction, standardization, data conversion, and annotation, will be performed on the two types of data separately.

[0067] Extraction and standardization of specific structures

[0068] The specific protein structure required for the project needs to be extracted, identified and standardized. Figure 1 In step S120 of the protein data conversion method 100 shown, specific structure extraction and standardization are performed based on the original data of the protein to form a specific structure description of the protein.

[0069] Specifically, if Figure 2 As shown in , the atomic information of the protein's atomic rectangular coordinate system information, i.e., PDB data, is extracted. Then, based on the extracted atomic information, the specific structure, surrounding environment, and range of the protein are identified, and the spatial position conversion is standardized.

[0070] More specifically, the segmentation and normalization of proteins with specific structures can be achieved through a data analysis module. The data analysis module can determine the specific target structure position and surrounding range in each subunit of the protein based on the secondary structure information provided in the PDB data. Then, the rules for segmenting and annotating the protein are set according to user input or preset parameters. The PDB data is segmented and annotated according to the rules.

[0071] Taking the α-helix structure as an example, the data analysis module is used to determine the α-helix position and surrounding range in each subunit of the protein based on the α-helix information provided in the PDB data, and to set the rules for segmenting and annotating the protein based on user input or preset parameters, and to segment and annotate the PDB data according to the rules, and output the segmentation and annotation results. Of course, in practical applications, the specific target structure position and surrounding range that we need to determine not only involve α-helix structures, but also other secondary structures such as β-folds or loop regions. That is, the specific target structure can include any identifiable secondary structure, such as α-helix, β-fold, loop region, etc.

[0072] Those skilled in the art should understand that the data analysis module can be implemented entirely by a computer program or by a pre-programmed functional module in a computer system. The module can effectively, flexibly and controllably segment and annotate PDB data to further study protein structural features.

[0073] On the other hand, Figure 2 As shown in , the frequency domain information of the diffraction data of the protein is transformed from the frequency domain to the space domain to form crystal space fractional coordinate data for representing the 3D density image. Then, the electron density information corresponding to the fractional coordinate data is segmented. The space group of the segmented data is converted to a specific space group, and the original fractional coordinates are represented by the fractional coordinates of the specific space group to obtain the electron density fractional coordinate information.

[0074] The transformation from the frequency domain to the spatial domain described here may be an inverse Fourier transform, which is generally implemented by an inverse fast Fourier transform (IFFT).

[0075] Therefore, the solution of the present disclosure may include a device for performing inverse Fourier transform based on an open source algorithm to obtain space coordinate electron density information. Specifically, the device further includes: an input module for receiving an MTZ file; an inverse Fourier transform module for performing inverse Fourier transform on the MTZ file received by the input module to obtain space coordinate electron density information; and an output module for outputting an MRC file obtained by the inverse Fourier transform module. That is, in a preferred embodiment according to the present disclosure, the crystal space fractional coordinate data used to represent the 3D density image is MRC data.

[0076] Those skilled in the art should understand that the module for performing inverse Fourier transform can be completely implemented by a computer program, or can be implemented by a pre-programmed functional module in a computer system.

[0077] Since we only need to extract the electron density of a given protein in the target area of ​​interest, we need to segment the electron density information corresponding to the fractional coordinate data. At this point, we can refer to the extraction and identification of the specific target structure based on the atomic rectangular coordinate system information of the protein, and use the obtained regional coordinate information of the specific target structure to segment its electron density. In other words, the results of the above-mentioned protein segmentation and normalization processing of specific structures are a necessary prerequisite for segmenting the electron density information corresponding to the fractional coordinate data.

[0078] In the process of converting from the space group of the segmented data to the specific space group, an algorithm step for rotating the crystal can be used. Specifically, the arbitrary direction adjustment of the crystal coordinate system can be achieved through this step. The algorithm step includes a first rotating mechanism fixed on a base, a second rotating mechanism fixed on the first rotating mechanism, and a third rotating mechanism fixed on the second rotating mechanism. The first, second, and third rotating mechanisms rotate along the Z axis, X axis, and Y axis of the default rectangular coordinate system, respectively, and their rotation angles are controlled by corresponding drivers. By adjusting the relative positions of the three rotating mechanisms, the Euler angle transformation of the crystal coordinate system can be achieved, thereby achieving the desired crystal orientation.

[0079] Those skilled in the art should understand that the algorithm steps and specific rotation mechanism for rotating the crystal can be completely implemented through a computer program, or can be implemented through a pre-programmed functional module in a computer system.

[0080] As mentioned above, after the specific structure extraction and standardization, a specific structural description of the protein is formed. That is, the original PDB data and MTZ data are converted into PDB data with atomic rectangular coordinate information of the specific structural description and MRC data with electron density fractional coordinate information.

[0081] Data conversion and annotation

[0082] exist Figure 1 In step S130, the specific structural description of the protein is then converted and annotated.

[0083] The atomic information description and electron density information description of a specific structure and its surrounding space can be further analyzed by multi-dimensional information calculation (residue sequence, rotational isomers, skeleton, etc.), providing more annotation information for the final electron density data. Among them, the space where the electron density is located can contain a series of labels such as atom attribution information, amino acid attribution information, structure attribution information, rotational isomer attribution information, etc.

[0084] like Figure 2As shown in , the atomic rectangular coordinate system information of the protein, that is, specifically the PDB data with structural description, is used to identify rotational isomers to form annotated spatial data.

[0085] The following explains the problem of rotamer identification.

[0086] In structural chemistry, dihedral angles can usually be specified unambiguously using four consecutive atoms. Figure 3 As shown, the dihedral angle ABCD represents the rotation angle α from plane ABC to plane BCD.

[0087] If you know the four-dimensional coordinates of four points A, B, C, and D, for example, p A 、p B 、p C 、p D , we can calculate the vector:

[0088] u1=u AB =p B -p A

[0089] u2=u BC =p C -p B

[0090] u3=u CD =p D -p C

[0091] At this point, from trigonometry we know that:

[0092]

[0093]

[0094] Therefore, we can use the atan2(y,x) function to get the unambiguous angle:

[0095] α=atan2(|u2|u1·(u2×u3),(u1×u2)·(u2×u3))

[0096] The atan2(y,x) function is a commonly used function in computer code. It can return the arc tangent value of the input (x,y) coordinate. Its unit is radians. The return value is between -π and π, which represents the angle required to rotate counterclockwise from the x-axis to the point (x,y).

[0097] On the other hand, Figure 2As shown in , for MRC data with electron density fractional coordinate information, i.e., in the preferred embodiment of the present disclosure, the following data conversion and annotation processing is required: first, the fractional coordinates are converted to Cartesian coordinates. Then, rotational isomer annotation is performed. Finally, the annotated data is transformed from the space domain to the frequency domain to form annotated frequency domain data.

[0098] Let's first look at the coordinate conversion process (fractional coordinates → Cartesian coordinates).

[0099] The conversion from fractional coordinates to Cartesian coordinates is performed according to the following formula. Assuming the fractional coordinates are (f1, f2, f3), the Cartesian coordinates are (x, y, z), and the unit cell parameters are (a, b, c, α, β, γ), where α is the angle between b and c, β is the angle between a and c, and γ is the angle between a and b, then the conversion matrix from fractional coordinates to Cartesian coordinates is:

[0100]

[0101] in:

[0102] c1=ccos(β)

[0103]

[0104]

[0105] So, using the transformation matrix, we can get the Cartesian coordinates:

[0106]

[0107] In the subsequent rotational isomer annotation, the rotational isomer identification results for the atomic rectangular coordinate system information (PDB data) mentioned above can be referred to to perform rotational isomer annotation on the Cartesian coordinate data. Specifically, a digital identifier is attached to each data grid point converted to the Cartesian coordinates to indicate the rotational isomer type of the position of the data grid point.

[0108] Then, the labeled data needs to be transformed from the spatial domain to the frequency domain to form labeled frequency domain data.

[0109] Simply put, if Figure 2 As shown in , the transformation from the spatial domain to the frequency domain described here can be a Fourier transform, which is generally implemented by a fast Fourier transform (FFT).

[0110] Specifically, in the process of forming the frequency domain data with annotations, a method of sampling in the reciprocal lattice space can be used. This method can be used to solve the properties of the crystal such as energy state density and electronic structure. More specifically, the method includes: obtaining the reciprocal lattice space by Fourier transform according to the positive lattice of the crystal, determining the basis vector of the reciprocal lattice and the range of the Brillouin zone; selecting a number of representative momentum points (K points) in the Brillouin zone, making them as evenly distributed as possible and covering all positions of different symmetries in the Brillouin zone, and using the Monkhorst-Pack method to generate K points; calculating the Hamiltonian matrix and eigenvalues ​​formed by the interaction between electrons and atomic nuclei and other electrons in the crystal at each K point; calculating the energy state density according to the eigenvalue and the number of sampling points, the energy state density indicates how many occupiable states exist in the unit energy range, reflects the law of electron motion and physical properties in the crystal, and the energy state density is obtained by multiplying the eigenvalue by the weight and then summing them, and the weight is determined by the number of sampling points. As an embodiment, the method introduced in the present disclosure may be applied to calculate the energy state density of a one-dimensional crystal, where two atoms constitute a unit cell and each atom has one electron.

[0111] Next, the mutual transformation from the spatial domain to the frequency domain described in the present disclosure also includes a device module that can be applied to machine learning and uses the Fourier method to encode and decode information in the space. The device module may include: an encoder module, which is used to convert the spatial domain function f(x, y, z) into a frequency domain function F(u, v, w) through discrete Fourier transform or fast Fourier transform, and output a spectrum diagram and a phase diagram; a decoder module, which is used to restore the frequency domain function F(u, v, w) to the spatial domain function f(x, y, z) through discrete Fourier inverse transform or fast Fourier inverse transform, and output the reconstructed spatial information; a processing module, which is used to perform some operations on the spectrum diagram and phase diagram to achieve filtering, enhancement, compression and other functions.

[0112] Those skilled in the art should understand that the device module for encoding and decoding information in space using the Fourier method that can be applied to machine learning can be completely implemented through a computer program, or through a pre-programmed functional module in a computer system.

[0113] Data storage

[0114] exist Figure 1 In step S140, the converted and labeled data is stored in a data form that can be used by the machine learning model.

[0115] Thus, the above data containing electron density and multiple annotations can be stored as data represented in the spatial and frequency domains, and the data contains rich labels. Different machine learning models can be used to process certain types of data.

[0116] Specifically, for example, Figure 2 As shown in , for the spatial data with annotations, which comes from PDB data, it can be stored as PLY data. For the frequency domain data with annotations, which comes from MTZ data and MRC data, it can be stored as H5 data. As mentioned above, the PLY and H5 data formats can be directly used for some machine learning, especially some deep learning models.

[0117] In addition to the technical effects described above, the technical solution proposed in the present disclosure may also produce the following beneficial effects:

[0118] 1. Reflect multidisciplinary and multidimensional characteristics in the data set;

[0119] 2. It has a positive effect on machine learning tasks.

[0120] The present disclosure relates to a protein data conversion method, and in particular, according to a preferred embodiment of the present disclosure, at least relates to the conversion of PDB data format, MTZ or MRC data format to PLY and H5 data format. It should be noted that, although there are some toolkits that can realize the conversion from MTZ or MRC to PLY format, they do not support the interception and standardization of specific protein structures, nor do they support the multi-directional annotation of training data, and therefore cannot solve the technical problem to be solved by the technical solution of the present disclosure.

[0121] Those skilled in the art should understand that based on the protein data conversion method disclosed in the present invention, a protein data conversion system can be developed. Those skilled in the art should understand that such a protein data conversion system can be implemented by a computer. For example, the following operations can be implemented by executing a computer program: obtaining the original data of the protein; performing specific structure extraction and standardization based on the original data of the protein to form a specific structural description of the protein; performing data conversion and annotation on the specific structural description of the protein; storing the converted and annotated data in a data form that can be used by a machine learning model. On the other hand, the protein data conversion system can also be described as a corresponding module or sub-module with an operating function corresponding to each method step, but those skilled in the art should understand that such a module or sub-module may not be composed of circuit components or other physical components, but a functional module constructed by a computer program.

[0122] The protein data is converted according to the scheme disclosed in the present invention and can be widely used in the process of protein drug development using machine learning models (especially deep learning models).

[0123] For example, in a training method for a machine learning model for protein drug development, the protein data conversion method taught in the present disclosure can be used to convert the original protein data of the protein sample used for training into a machine learning training sample. Those skilled in the art should recognize that the converted machine learning training sample can be directly applied to certain machine learning models (deep learning models) because it is in, for example, PLY or H5 format. Therefore, next, the machine learning training sample converted from the original protein data can be used to train the machine learning model for protein drug development.

[0124] On the other hand, in a method for protein drug development based on a machine learning model, the protein data conversion method taught in the present disclosure can be used to convert the original protein data of the protein to be tested into input data for the machine learning model. Those skilled in the art should recognize that the converted protein data can be directly used as input data for certain machine learning models (deep learning models) because it is in PLY or H5 format, for example. Therefore, next, the input data converted from the original protein data can be directly input into the machine learning model to obtain the output result of the machine learning model. The output result of the machine learning model can be a prediction of various properties of the protein, such as solubility, immunogenicity, viscosity, etc., in order to determine whether the target protein can be developed for drug development; it can also be based on the target protein to generate other new substances and new drugs. The result, such as the generation of binding proteins, etc. According to such an output result, protein drug development can be achieved.

[0125] In addition, it should be recognized by those skilled in the art that the method of the present disclosure can be implemented as a computer program. As described above in conjunction with the accompanying drawings, the method of the above embodiment is executed by one or more programs, and the instructions in the program cause the computer or processor to execute the algorithm described in conjunction with the accompanying drawings. These programs can be stored and provided to the computer or processor using various types of non-transient computer-readable media. Non-transient computer-readable media include various types of tangible storage media. Examples of non-transient computer-readable media include magnetic recording media (such as floppy disks, tapes, and hard disk drives), magneto-optical recording media (such as magneto-optical disks), CD-ROMs (compact disk read-only memories), CD-Rs, CD-R / Ws, and semiconductor memories (such as ROMs, PROMs (programmable ROMs), EPROMs (erasable PROMs), flash ROMs, and RAMs (random access memories)). Further, these programs can be provided to the computer using various types of transient computer-readable media. Examples of transient computer-readable media include electrical signals, optical signals, and electromagnetic waves. Transient computer-readable media can be used to provide programs to the computer through wired communication paths or wireless communication paths such as wires and optical fibers.

[0126] For example, according to one embodiment of the present disclosure, a non-transitory computer-readable storage medium may be provided for storing a computer program, wherein the computer program includes instructions that, when executed by a processor of an electronic device, enable the electronic device to implement the protein data conversion method as described above.

[0127] In addition, according to the content disclosed in the present disclosure, a computer system can also be provided, the computer system comprising: a processor; a memory; and a computer program. The computer program is stored in the memory and is configured to be executed by the processor. The computer program includes instructions for implementing the protein data conversion method described above.

[0128] The implementation methods of the present disclosure are not limited to the above-mentioned embodiments. Without departing from the spirit and scope of the present disclosure, ordinary technicians in this field can make various changes and improvements to the present disclosure in form and details, and these are considered to fall within the protection scope of the present disclosure.

Claims

1. A protein data conversion method, characterized in that: The method comprises: Obtain raw data of proteins; Performing specific structure extraction and standardization processing based on the original data of the protein to form a specific structural description of the protein; Performing data conversion and annotation on the specific structural description of the protein; Store the converted and labeled data in a data form that can be used by machine learning models.

2. The method according to claim 1, characterized in that The original data of the protein includes the atomic rectangular coordinate system information and diffraction data frequency domain information of the protein.

3. The method according to claim 2, characterized in that The atomic rectangular coordinate system information of the protein comes from PDB data, and the diffraction data frequency domain information of the protein comes from MTZ data.

4. The method according to claim 2, characterized in that: The specific structure extraction and standardization processing based on the original data of proteins includes: Extracting atomic information from the atomic rectangular coordinate system information of the protein; Based on the extracted atomic information, the specific structure, surrounding environment and scope of the protein are identified, and the spatial position transformation is standardized.

5. The method according to claim 4, characterized in that The method of identifying the specific structure, surrounding environment and range of the protein based on the extracted atomic information and standardizing the spatial position conversion includes: Determine the specific target structure position and surrounding range in each subunit of the protein based on the secondary structure information provided in the PDB data; Setting rules for segmenting and labeling proteins based on user input or preset parameters; PDB data is segmented and annotated according to the rules.

6. The method according to claim 5, characterized in that The specific target structure includes any identifiable secondary structure.

7. The method according to claim 2, characterized in that The specific structure extraction and standardization processing based on the original data of proteins includes: Performing frequency domain to space domain transformation on the frequency domain information of the diffraction data of the protein to form crystal space fractional coordinate data for representing a 3D density image; Segment the electron density information corresponding to the fractional coordinate data; The space group of the segmented data is converted into a specific space group, and the original fractional coordinates are represented by the fractional coordinates of the specific space group to obtain the electron density fractional coordinate information.

8. The method according to claim 7, characterized in that The crystal space fractional coordinate data used to represent the 3D density image is MRC data.

9. The method according to claim 7, characterized in that: The segmenting of the electron density information corresponding to the fractional coordinate data further includes: The electron density information corresponding to the fractional coordinate data is segmented with reference to the specific structure extraction and identification performed on the atomic rectangular coordinate system information of the protein.

10. The method according to claim 7, characterized in that The transformation from frequency domain to space domain is an inverse Fourier transform.

11. The method according to claim 4, characterized in that The data conversion and annotation of the specific structural description of the protein includes: Rotational isomer identification is performed on the atomic rectangular coordinate system information of the protein to form annotated spatial data.

12. The method according to claim 7, characterized in that The data conversion and annotation of the specific structural description of the protein includes: Convert fractional coordinates to Cartesian coordinates; Carry out rotational isomer annotation; The labeled data is transformed from the spatial domain to the frequency domain to form labeled frequency domain data.

13. The method according to claim 12, characterized in that The performing rotational isomer annotation further includes: performing rotational isomer annotation on the Cartesian coordinate data with reference to the rotational isomer identification performed on the atomic rectangular coordinate system information.

14. The method according to claim 12, characterized in that The transformation from the spatial domain to the frequency domain is Fourier transformation.

15. The method according to claim 11, characterized in that The storing of the converted and annotated data into a data form usable by the machine learning model includes: Stores labeled spatial data as PLY data.

16. The method according to claim 12, characterized in that The storing of the converted and annotated data into a data form usable by the machine learning model includes: Store the annotated frequency domain data as H5 data.

17. A protein data conversion system implemented by a computer, wherein the computer-implemented system implements the following operations by executing a computer program: Obtain raw data of proteins; Performing specific structure extraction and standardization processing based on the original data of the protein to form a specific structural description of the protein; Performing data conversion and annotation on the specific structural description of the protein; Store the converted and labeled data in a data form that can be used by machine learning models.

18. A method for training a machine learning model for protein drug development, characterized in that: The method comprises: Using the protein data conversion method according to any one of claims 1 to 16, converting the original protein data of the protein sample used for training into a machine learning training sample; Train machine learning models for protein drug development using machine learning training samples converted from raw protein data.

19. A method for protein drug development based on a machine learning model, characterized in that: The method comprises: Using the protein data conversion method according to any one of claims 1 to 16, converting the original protein data of the protein to be tested into input data of the machine learning model; Inputting the input data into the machine learning model to obtain an output result of the machine learning model; According to the output results, protein drug development is achieved.

20. A non-transitory computer-readable storage medium for storing a computer program, the computer program comprising instructions, which, when executed by a processor of an electronic device, cause the electronic device to implement the protein data conversion method according to any one of claims 1 to 16.

21. A computer system, comprising: processor; Memory; and A computer program, wherein the computer program is stored in the memory and is configured to be executed by the processor, the computer program comprising instructions for implementing the protein data conversion method according to any one of claims 1 to 16.