Protein domain identification method and system, storage medium, and device

Abstract

Provided are a protein domain identification method and system, a storage medium, and a device. The method comprises: acquiring a density map embedding and an atom structure embedding by means of a dual-tower network based on a density tower and an atom tower, and performing cross-modal alignment; constructing a density map-aware structure embedding database on the basis of the density map embedding and the atom structure embedding which have undergone the cross-modal alignment; and querying the density map-aware structure embedding database on the basis of an inputted density map to obtain a closest atom structure embedding, and identifying a protein domain. The method improves the protein domain identification efficiency and the identification accuracy under a low-resolution density map.

Description

Methods, systems, storage media and devices for protein domain identification Technical Field

[0001] This invention belongs to the field of bioinformatics technology, and in particular relates to a method, system, storage medium and device for identifying protein domains. Background Technology

[0002] Protein structure determination is crucial for understanding biological processes and molecular mechanisms. In recent years, cryo-electron microscopy (cryo-EM) has gradually become an important tool for determining the structure of endogenous macromolecules, and the number of cryo-EM density maps deposited in the cryo-EM database (EMDB) is growing exponentially. Furthermore, an emerging branch of cryo-EM—cryo-electron tomography (cryo-ET)—can reveal in-situ nanoscale structural information within the cellular environment, information that might be overlooked in traditional single-particle analysis (SPA) cryo-EM.

[0003] Obtaining atomic-resolution protein structures from cryo-electron microscopy requires high-resolution density maps, and typically also the primary amino acid sequence. As the resolution of the density map decreases, density regions corresponding to protein side chains become blurred and difficult to identify. In extreme cases, density regions containing multiple amino acids may be completely lost, making it difficult or impossible to confirm the chemical properties of individual amino acids. In the absence of prior knowledge such as protein type or sequence, traditional structure determination methods are generally unsuitable for low-resolution density maps. Therefore, from a broader perspective, protein identification has become an important alternative for interpreting density maps.

[0004] Currently, sequence-search-based methods first identify individual amino acids from density maps and then construct amino acid chains for subsequent sequence searches. On the other hand, structure-search-based methods can also achieve protein identification, provided that a sufficiently large main chain structure can be constructed from the density map. However, both methods are limited at low resolutions (greater than 1000 N / A). The performance of these molecules on density maps drops sharply because they are highly dependent on accurate amino acid recognition.

[0005] To improve the accuracy and efficiency of identifying short sequences or broken backbones, these methods typically limit the search space by combining them with other biological experiments, such as tandem mass spectrometry (MS / MS) that can detect all proteins in a target sample. Furthermore, recent advances in sequence-based structure prediction offer another option: predicting the structure of every protein in the search space and ranking them according to their correlation with the density map. However, this process is extremely time-consuming and limited by the quality of the predicted structures. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a protein domain identification method, system, storage medium, and device. Based on a dual-tower network of density and atomic towers, it extracts density map embeddings and atomic structure embeddings, performs cross-modal alignment, and directly maps the density map embeddings to the atomic structure embeddings containing domain information. A density map-aware structure embedding database is constructed, and the closest atomic structure embedding can be obtained through density map queries to identify protein domains, thereby improving the efficiency of protein domain identification and the accuracy of identification under low-resolution density maps.

[0007] This invention is achieved through the following technical solution:

[0008] Density map embedding and atomic structure embedding are obtained through a dual-tower network based on density tower and atomic tower, and cross-modal alignment is performed.

[0009] Based on cross-modal aligned density map embedding and atomic structure embedding, a structure embedding database with known density maps is constructed.

[0010] Based on the input density map, the system queries a database of known density map structure embeddings to obtain the closest atomic structure embeddings and identify protein domains.

[0011] Optional,

[0012] The process of obtaining density map embeddings and atomic structure embeddings through a dual-tower network based on density towers and atomic towers, and performing cross-modal alignment, includes:

[0013] Density map embeddings are extracted from the input density map using density towers;

[0014] Atomic structure embeddings are extracted from the input atomic structure using atomic towers.

[0015] Optional,

[0016] The density map embedding and atomic structure embedding are aligned across modes.

[0017] Optional,

[0018] The construction of a density map-aware structure embedding database based on cross-modal aligned density map embedding and atomic structure embedding includes:

[0019] The atomic structure embeddings of all structural domains extracted from the atomic tower are stored in the known density map structure embedding database, and the atomic structure embeddings are associated with the corresponding atomic structures and other metadata.

[0020] Optional,

[0021] The step of identifying protein domains by querying a structure embedding database based on the input density map to obtain the closest atomic structure embedding includes:

[0022] The input density map is processed using a density tower to generate a density map embedding;

[0023] Based on the density map embedding, query the structure embedding database that knows the density map to obtain the closest atomic structure embedding;

[0024] Protein domain recognition is based on atomic structure embedding.

[0025] Optional,

[0026] The density tower includes a residual U-Net branch, a Swin-Conv U-Net branch, a convolution module, and a compression module;

[0027] The atomic tower comprises an atomic encoder, a structural module, and a fusion module.

[0028] Optional,

[0029] The density tower and the atom tower are each subjected to self-supervised pre-training.

[0030] The present invention also provides a protein domain recognition system for implementing the aforementioned method, the system comprising:

[0031] The dual-tower network module is used to obtain density map embeddings and atomic structure embeddings through a dual-tower network based on density towers and atomic towers, and to perform cross-modal alignment.

[0032] The database module is used to build a density-map-aware structure embedding database based on cross-modal aligned density map embedding and atomic structure embedding.

[0033] The query and identification module is used to query the structural embedding database of the input density map to obtain the closest atomic structure embedding and identify protein domains.

[0034] Optional,

[0035] The dual-tower network module is also configured to:

[0036] Density map embeddings are extracted from the input density map using density towers;

[0037] Atomic structure embeddings are extracted from the input atomic structure using atomic towers;

[0038] The density map embedding and atomic structure embedding are aligned across modes.

[0039] Optional,

[0040] The database module is also configured to:

[0041] The atomic structure embeddings of all structural domains extracted from the atomic tower are stored in the known density map structure embedding database, and the atomic structure embeddings are associated with the corresponding atomic structures and other metadata.

[0042] Optional,

[0043] The query identification module is also configured to:

[0044] The input density map is processed using a density tower to generate a density map embedding;

[0045] Based on the density map embedding, query the structure embedding database that knows the density map to obtain the closest atomic structure embedding;

[0046] Protein domain recognition is based on atomic structure embedding.

[0047] Optional,

[0048] The system is also configured to:

[0049] The density tower and the atom tower are respectively subjected to self-supervised pre-training.

[0050] The present invention also provides a computer-readable storage medium storing one or more programs, which, when executed, can implement the aforementioned protein domain identification method.

[0051] The present invention also provides a device, including a processor, a communication interface, a computer-readable storage medium, and a communication bus; wherein the processor, the communication interface, and the computer-readable storage medium communicate with each other through the communication bus;

[0052] The processor is used to execute programs stored in a computer-readable storage medium.

[0053] Compared with the prior art, the present invention has the following advantages:

[0054] 1. The protein domain identification method proposed in this invention is based on a dual-tower network of density tower and atomic tower. It extracts density map embeddings and atomic structure embeddings, and performs cross-modal alignment. The density map embeddings are directly mapped to the atomic structure embeddings containing domain information, and a database of structure embeddings that know the density map is constructed. The closest atomic structure embedding can be obtained by querying the density map, thereby identifying protein domains, improving the efficiency of protein domain identification and the accuracy of identification under low-resolution density maps.

[0055] 2. A cross-modal alignment strategy is adopted to align density map embedding and atomic structure embedding, integrating the two modalities of density map and atomic structure. This solves the problem of data richness difference between density map and atomic structure, maximizes data utilization, solves the problem of data richness difference between different modalities, and enables structural domain identification through embedding retrieval, thereby improving the accuracy of identification.

[0056] 3. Through self-supervised pre-training and cross-modal alignment, knowledge learned from large atomic structure datasets can be transferred to sparse density map datasets. Protein domains can be identified from low-resolution density maps without prior knowledge of protein sequences, thus improving the recognition accuracy in low-resolution density maps.

[0057] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures pointed out in the description, claims and drawings. Attached Figure Description

[0058] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0059] Figure 1 shows a schematic flowchart of the protein domain identification method;

[0060] Figure 2 shows a schematic block diagram of the protein domain recognition system;

[0061] Figure 3 shows a schematic diagram of the CryoDiscovery framework according to an embodiment of the present invention;

[0062] Figure 4 shows a schematic diagram of a dual-tower network structure according to an embodiment of the present invention;

[0063] Figure 5 shows a schematic diagram of the construction of the knowledge density graph embedded in the database DateDB according to an embodiment of the present invention;

[0064] Figure 6 shows a schematic diagram of the CryoDiscovery retrieval framework according to an embodiment of the present invention;

[0065] Figure 7 shows a schematic diagram of the protein domain classification hierarchy in CATH / SCOPe according to an embodiment of the present invention;

[0066] Figure 8 shows a schematic diagram of the embedding distribution of structural domain types according to an embodiment of the present invention;

[0067] Figure 9 shows a schematic diagram of protein domain identification using a low-resolution density map;

[0068] Figure 10 shows a schematic diagram illustrating the resolution effect of the low-pass filter spectrum;

[0069] Figure 11 is a schematic diagram of the structure of a device according to an embodiment of the present invention. Detailed Implementation

[0070] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0071] Referring to Figure 1, the method of the present invention includes:

[0072] S1. Obtain density map embedding and atomic structure embedding through a dual-tower network based on density tower and atomic tower, and perform cross-modal alignment;

[0073] Among them, density map embeddings are extracted from the input density map using a density tower;

[0074] Atomic structure embeddings are extracted from the input atomic structure using atomic towers.

[0075] Specifically, the density map embedding and atomic structure embedding are aligned across modes.

[0076] S2. Construct a structure embedding database that knows the density map based on density map embedding and atomic structure embedding after cross-modal alignment;

[0077] Specifically, the atomic structure embeddings of all structural domains extracted from the atomic tower are stored in the known density map structure embedding database, and the atomic structure embeddings are associated with the corresponding atomic structures and other metadata.

[0078] S3. Based on the input density map, the closest atomic structure embedding is obtained by querying the structure embedding database of the density map to identify the protein domain.

[0079] Among them, the density map is processed using a density tower to generate a density map embedding;

[0080] Based on the density map embedding, query the structure embedding database that knows the density map to obtain the closest atomic structure embedding;

[0081] Based on atomic structure embedding to identify protein domains,

[0082] The density tower includes a residual U-Net branch, a Swin-Conv U-Net branch, a convolution module, and a compression module.

[0083] The atomic tower comprises an atomic encoder, a structural module, and a fusion module.

[0084] The density tower and the atom tower are respectively subjected to self-supervised pre-training.

[0085] Specifically,

[0086] This invention proposes a cross-modal method based on protein domains, CryoDiscovery. It utilizes a dual-tower deep neural network to identify protein domains from low-resolution cryo-electron microscopy density maps. By performing cross-modal alignment in a shared low-dimensional space, it integrates the density map and atomic structure. Based on the aligned density map and atomic structure, a density-aware structure embedding database containing atomic structure embeddings is constructed. Using this database, protein domains can be identified from low-resolution density maps through embedding retrieval.

[0087] I. Density map embedding and atomic structure embedding are obtained through a dual-tower network based on density tower and atomic tower, and cross-modal alignment is performed.

[0088] 1. Dual-tower network.

[0089] The dual-tower network consists of a density tower and an atom tower, used to simultaneously learn the embeddings of protein domain density maps and atomic structures in a shared low-dimensional space.

[0090] Density map embeddings and atomic structure embeddings are obtained through a dual-tower network based on density towers and atomic towers, and cross-modal alignment is performed, including:

[0091] Density map embeddings are extracted from the input density map using density towers;

[0092] Atomic structure embeddings are extracted from the input atomic structure using atomic towers.

[0093] The density tower contains a residual U-Net branch, a Swin-Conv U-Net branch, a convolutional module (Conv module), and a compression module (Compress Module) for extracting density map embeddings (Zd) from the input density map.

[0094] The density tower employs a U-Net-like architecture, progressively learning the spatial and semantic features of cryo-electron microscopy density maps. The input density map is processed through residual U-Net branches and Swin-Conv U-Net branches to combine its local and non-local semantic patterns. It is then reconstructed via a convolutional module and projected into a density map embedding Zd via a compression module.

[0095] The density tower integrates two parallel branches: the Residual U-Net branch and the Swin-Conv U-Net branch, and also includes convolutional and compression modules. The Residual U-Net branch consists of multiple convolutional layers, max-pooling layers, and downsampling / upsampling convolutional layers; the Swin-Conv U-Net branch consists of multiple basic modules and downsampling / upsampling convolutional layers. The basic modules combine a Swin transformer and convolution.

[0096] The input features of the basic module are processed by convolutional layers and then divided into convolutional features and WMSA features. The convolutional features are further processed by two convolutional layers and combined with residual connections to obtain the final convolutional features. The WMSA features are processed by the WMSA (Window Multi-Head Self-Attention) module, combined with skip connections, and further processed by two linear layers to obtain the final WMSA features.

[0097] The final convolutional features and the final WMSA features are concatenated together, and then passed through another convolutional layer to obtain the intermediate features of the basic module. Finally, the final output is obtained by combining the intermediate features and the original input features.

[0098] The Atom Tower comprises an AtomEncoder, a Structure Module, and a Fusion Module for extracting atomic structure embeddings (Za) from the input atomic structure. The Atom Tower embeds the input atomic structure into single and paired representations via the AtomEncoder, and then fuses them using the Fusion Module to form the atomic structure embedding (Za). The Structure Module is suitable for recovering the input atomic structure given single and paired representations.

[0099] The Atom Tower comprises three modules: an AtomEncoder, a Structure Module, and a Fusion Module, taking the atomic structure of a protein domain as input. The AtomEncoder obtains single and paired representations in parallel.

[0100] For a single representation, the atomic tower extracts the torsion angle from the input atomic structure and obtains a single representation through two linear layers.

[0101] For pairwise representations, the atomic coordinates of the input atomic structure are extracted: the input atomic structure first passes through two identical combinations of "one linear layer + one multi-attention layer + residual," and then sequentially through two identical modules. Within each module, features are first passed in parallel through another two identical combinations of "one linear layer + one linear layer," and the two parallel output features are fused together via matrix multiplication. The fused features then pass through a linear layer and are element-wise multiplied with features obtained by passing the input features of the current module to another linear layer. The current features then pass through two linear layers and, after passing through the residual, obtain the pairwise representation.

[0102] Single and paired representations are combined with the backbone frame and input to the structural module to reconstruct the atomic structure of the input. Simultaneously, a 128-dimensional atomic structure embedding (Za) is obtained through the fusion module. In the structural module, the single representation is first passed through a linear layer and then fused with the paired representation using invariant point attention. The resulting fused features are passed through three linear layers that combine residuals to update the single representation and the backbone frame. Then, the updated single representation and the updated backbone frame are combined with the input single representation of the structural module, and the reconstructed protein domain atomic structure is obtained through ResNet. The fusion module consists of a combination of three convolutional layers + max pooling, three linear layers, and one global pooling layer.

[0103] The density tower and atomic tower are used to perform self-supervised pre-training on the original cryo-electron microscopy density map and atomic structure, respectively, to extract their modality-specific features: density map embedding and atomic structure embedding. The extracted density map embedding and atomic structure embedding are then integrated into a unified representation space through cross-modal alignment.

[0104] 2. Cross-modal alignment.

[0105] Cross-modal alignment enables the integration of heterogeneous data types, such as cryo-electron microscopy density maps and atomic structures. However, due to the sparse pairs of data between the two modalities, traditional cross-modal alignment methods that rely on paired data for training cannot fully utilize the unpaired data portions of each modality. Therefore, in this embodiment of the invention, the alignment process is decomposed into structural domain semantic learning and direct alignment of the two modalities.

[0106] The training process aligns density map embeddings and atomic structure embeddings of the same domain type. The training process includes:

[0107] (1) Self-supervised learning.

[0108] Both the density tower and the atom tower in the dual-tower network are trained in a self-supervised manner, allowing the network to independently learn the complex details of each input modality during the training phase.

[0109] In this embodiment, for density towers, the MSE loss function can be used to train the network; for atomic towers, a combination of FAPE loss, torsion angle loss, distance graph loss and pLDDT loss can be used to train the network.

[0110] Formula 1 is a combination of four loss functions: Loss atom =Loss FAPE +λtorsion-angle×Losstorsion-angle+λ distogram ×Lossdiistogram+λ pLDDT ×Loss pLDDT (1)

[0111] Among them, λtorsion-angle, λ distogram , λ pLDDT These are the weights of each loss function.

[0112] The FAPE loss function, as shown in Equation 2, is a loss function used to calculate the difference in atomic positions during protein structure reconstruction. It can overcome the effects of translation and rotation when processing protein amino acid structures. It calculates the difference in amino acid coordinates between the reconstructed atomic structure and the real atomic structure in each local reference frame, avoiding the loss explosion that may be caused by the accumulation of errors from the current amino acid to the next amino acid.

[0113] Among them, the local framework and three-dimensional coordinates of amino acids in the reconstructed atomic structures of T and X; These are coordinates derived from the actual atomic structure.

[0114] The torsion angle loss function is shown in Equation 3:

[0115] Where α is the torsion angle of the reconstructed atomic structure; It is the torsion angle of the real atomic structure; N res It represents the number of amino acids in the actual atomic structure.

[0116] The distance graph loss function, as shown in Equation 4, maps pairwise representations to values ​​from 2.3 to... The 64 distance intervals are determined, and their corresponding probabilities are obtained.

[0117] Among them, y b It is the binning distance of amino acids after one-hot encoding; p b It is an interval probability, N res It represents the number of amino acids in the actual atomic structure.

[0118] As shown in Equation 5, the pLDDT loss can measure the local distance difference between two atomic structures.

[0119] Loss pLDDT The cross-entropy loss between the predicted LDDT and the true LDDT is calculated. The pLDDT and true LDDT are calculated based on the reconstructed atomic structure and the true atomic structure, respectively. Here, we divide the pLDDT into 50 intervals.

[0120] in, It is a true LDDT after one-hot encoding; It is the pLDDT probability of the interval; N res It represents the number of amino acids in the actual atomic structure.

[0121] (2) Comparative learning.

[0122] To embed protein domain information into the atomic structure embeddings Za, the atomic tower is also trained through contrastive learning to enhance the uniqueness of embeddings for different domain types. This ensures that atomic structure embeddings for the same protein domain are closer together, while others are further apart, which is the basis for alignment in the shared low-dimensional space of specific domains. The contrastive learning of the atomic tower is performed based on rich annotations of protein domain types in the atomic structures.

[0123] In this embodiment, the contrast loss function as shown in Equation (6) can be used to improve the discriminativeness of embeddings between different structural domains by comparing the mean squared error (MSE) between positive and negative samples.

[0124] in, Indicates the embedding of the reference atomic structure; Z i Indicates the atomic structure embedding of positive samples; Z j Indicates the atomic structure embedding of negative samples; N p N is the number of positive samples; n It represents the number of negative samples.

[0125] (3) Cross-modal alignment.

[0126] Cross-modal alignment refers to the process of creating models capable of integrating and understanding multiple types of data. By leveraging complementary information from these two modalities, the versatility and adaptability of the model across various visual or language processing tasks are enhanced. In this embodiment, aligning cryo-electron microscopy density maps with atomic structures in a shared low-dimensional space enables protein domain identification through embedding retrieval.

[0127] In this embodiment, mean squared error (MSE) can be used as the alignment loss to align the density tower and the atomic tower. The density map embedding Zd is directly mapped to the atomic structure embedding Za, which contains domain information, thereby aligning the two modes in a shared low-dimensional space of a specific domain.

[0128] In this embodiment, the training data comes from the publicly available protein databases EMDB and PDB, and the domain information comes from the CATH / SCOPe database and KluDo predictions.

[0129] II. Based on density map embedding and atomic structure embedding after cross-modal alignment, construct a structure embedding database that knows the density map.

[0130] Based on cross-modal aligned density map embeddings and atomic structure embeddings, a density map-aware structure embedding database is constructed, including:

[0131] The atomic structure embeddings of all structural domains extracted from the atomic tower are stored in the known density map structure embedding database, and the atomic structure embeddings are associated with the corresponding atomic structures and other metadata.

[0132] In this embodiment, after cross-modal alignment, a pre-trained atomic tower is applied to extract the atomic structure embeddings Za of all domain structures. The extracted atomic structure embeddings Za, along with pointers to their corresponding atomic structures and other metadata, are stored in DateDB. The pointers are operators stored in DateDB used to associate the atomic structure embeddings Za with their corresponding atomic structures and other metadata, thus enabling DateDB to store not only embedding information but also additional context and background information.

[0133] In this embodiment, other metadata may include, but is not limited to: experimental conditions, data sources, structural classification, biological functions, and other relevant information.

[0134] Knowing the density map structure embedding database DateDB can be used to support the subsequent retrieval process, effectively matching protein domains based on atomic structure features.

[0135] Third, based on the input density map, the closest atomic structure embedding is obtained by querying the structure embedding database of the density map to identify the protein domain.

[0136] Based on the input density map, the closest atomic structure embedding is obtained by querying a density map-knowledgeable structure embedding database to identify protein domains, including:

[0137] The input density map is processed using a density tower to generate a density map embedding;

[0138] Based on the density map embedding, query the structure embedding database that knows the density map to obtain the closest atomic structure embedding;

[0139] Protein domain recognition is based on atomic structure embedding.

[0140] In this embodiment, the density tower receives the density map as input to generate a density map embedding Zd as a query embedding. Then, the query embedding is searched in the density map-aware structural embedding database DateDB to obtain the nearest atomic structure embedding Za, thereby achieving protein domain identification.

[0141] In this embodiment, after cross-modal alignment in the shared low-dimensional space of a specific structural domain, the density map embedding Zd obtained by the density tower can be used to retrieve the corresponding atomic structure embedding. By evaluating the similarity between Zd and each Za stored in DateDB, the top k atomic structure embeddings most similar to Zd are retrieved. Using the relational data in DateDB, the corresponding amino acid region of the corresponding chain of the atomic structure of Za can be retrieved, thereby obtaining the protein structural domain information of Za. The results are deduplicated by retaining only the representative with the highest cosine similarity to the query in each structural domain type, and the final ranking is calculated.

[0142] In this embodiment, the Faiss algorithm can be used to implement the approximate nearest neighbor search (ANN) algorithm for retrieval.

[0143] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0144] Referring to Figure 2, the structure of a protein domain identification system for implementing the above method is shown, including a dual-tower network module, a database module, and a query identification module.

[0145] The dual-tower network module is used to obtain density map embedding and atomic structure embedding through a dual-tower network based on density tower and atomic tower, and to perform cross-modal alignment.

[0146] The database module is used to construct a density-map-aware structure embedding database based on cross-modal aligned density map embedding and atomic structure embedding.

[0147] The query and recognition module is used to obtain the closest atomic structure embedding by querying the structure embedding database of the density map based on the input density map, and to identify the protein domain.

[0148] Figure 3 is a schematic diagram of the CryoDiscovery framework according to an embodiment of the present invention, which includes a density tower (DensityTower) and an atom tower (AtomTower) for simultaneously learning the embeddings of protein domain density maps and atomic structures in a shared low-dimensional space. Density map embeddings (Zd) and atomic structure embeddings (Za) of the same domain type are aligned during the training process. After training, all atomic structure embeddings (Za) are stored in the DateDB database. During the inference process for identifying protein domains, CryoDiscovery has two modes: Mode 1 uses a user-specified density region as input; Mode 2 uses the original density map as input. Mode 2 integrates the density map segmentation tool Segger. The density map is processed by the density tower to generate density map embeddings Zd, and then the DateDB database is queried to find the nearest atomic structure embeddings Za, ultimately achieving the identification of the protein domain.

[0149] Figure 4 is a schematic diagram of the dual-tower network structure according to an embodiment of the present invention.

[0150] DensityTower contains a residual U-Net branch, a Swin-Conv U-Net branch, a convolutional module, and a compression module, which extracts density map embeddings (Zd) from the input density map.

[0151] Data preprocessing for density towers:

[0152] Cryo-electron microscopy density maps of protein domains were used as input to the density tower. After obtaining the density map, the density tower normalized the density map using the average density value and standard deviation. The density map was then filled using its minimum density value. Due to the limitation of the "attention" mechanism, the side length of the resulting cube must be a multiple of 48.

[0153] The AtomTower consists of an AtomEncoder, a Structure Module, and a Fusion Module, which extracts the atomic structure embedding (Za) from the input atomic structure.

[0154] Here, the twist angle represents the rotation of one atomic group relative to another within an amino acid; the main chain frame records the Euclidean transformation from the local coordinate system defined for each amino acid to the global coordinate system. Specifically, the local coordinate system is defined by C... α Centered on the atom, N and C atoms define the X and Y axes.

[0155] Data preprocessing for atomic towers:

[0156] The atomic structure of protein domains is used as input to the atomic tower. After obtaining the atomic structure, the atomic tower extracts the three-dimensional coordinates of the atom, the torsion angle of the amino acid, and the skeletal framework.

[0157] The three-dimensional coordinates are stored in the atomic structure file, and the skeleton framework records the Euclidean transformation from the local coordinate system to the global coordinate system. Each amino acid is a local coordinate system, and each amino acid is represented as a coordinate system in C0. α A triangle centered on the atom, with N and C as the other two vertices.

[0158] Each amino acid has seven torsion angles, which are encoded as sine and cosine. Three of the seven torsion angles come from the amino acid's backbone [ω, φ, ψ], and the remaining four come from the amino acid's side chain [χ1, χ2, χ3, χ4]. In proteins, the χ angle is typically a dihedral angle that the side chain can rotate freely, used to describe the rotational conformation of the amino acid's side chain. The side chain is the part attached to the amino acid backbone. Different amino acids have different numbers of χ angles in their side chains and are usually classified as χ1, χ2, χ3, and χ4.

[0159] Both the density tower and the atom tower underwent self-supervised pre-training. Furthermore, the atom tower was trained using contrastive learning.

[0160] Figure 5 is a schematic diagram illustrating the construction of the Known Density Map structure embedding database DateDB according to an embodiment of the present invention. Atomic structure embeddings are extracted from the protein domain atomic structure dataset using atomic towers. These extracted atomic structure embeddings, along with their corresponding atomic structures and other metadata, are systematically stored in DateDB.

[0161] Figure 6 is a schematic diagram of the CryoDiscovery retrieval framework according to an embodiment of the present invention. Density map embeddings (Zd) are extracted using a density tower, and the closest atomic structure embeddings (Za) are retrieved from DateDB using the density map embeddings (Zd). The retrieval results are then deduplicated and sorted based on similarity.

[0162] The following describes the embodiments of the present invention in detail with reference to experiments.

[0163] 1. Dataset construction.

[0164] 1) Prepare training data: Download the data from EMDB at a resolution of [resolution value missing]. The original cryo-electron microscopy density map; the corresponding protein atomic structure as of April 2023 downloaded from the PDB; and the protein structure resolved by other determination methods (e.g., X-ray crystallography).

[0165] 2) Obtaining protein domain structure data: Since proteins may contain multiple domains, it is necessary to segment the atomic structure into individual domains. CATH and SCOPe are two partially overlapping bioinformatics databases that provide domain information. Using CATH and SCOPe databases, the type and location of the domains can be specified.

[0166] CATH uses computational methods and manual validation to annotate the three-dimensional protein structures in the PDB, classifying protein domains into four levels: class (level 1), framework (level 2), topology / folding (level 3), and homology superfamily (level 4). Two protein domains are more similar when they share higher levels. Evolutionary relationships exist between protein domains at the same homology superfamily level.

[0167] SCOP comprises protein domains manually classified based on the similarity of their atomic structures and sequences, and also categorizes protein domains into four levels: class (level 1), fold (level 2), superfamily (level 3), and family (level 4). Protein domains within the same family level have similar atomic structures, sequences, and functions, and are considered to have a closer ancestor. SCOPe was introduced as an extension of SCOP after adding new atomic structures and correcting some errors in SCOP.

[0168] Atomic structures can be segmented into protein domains using metadata from CATH / SCOPe. This step generates approximately 400,000 domain structures labeled with CATH and 250,000 domain structures labeled with SCOPe. For approximately 50% of the atomic structures without CATH / SCOPe labels, KluDo can be used to segment them into protein domains. Since KluDo only predicts the location of the domains and does not provide the domain type, the domain atomic structures segmented by KluDo are considered unlabeled domains.

[0169] 3) Obtaining domain density maps: Using the protein domain structures described above, the original cryo-electron microscopy density map was cropped into density regions and redundancy was removed to obtain approximately 200,000 domain density maps.

[0170] In this embodiment, cryo-electron microscopy density maps and atomic structures are downloaded from EMDB and PDB for use in self-supervised training. Atomic structures of protein domains can be obtained through CATH / SCOPe labeling or prediction using KluDo: 400,000 protein domain atomic structures with CATH labels and 250,000 with SCOPe labels are obtained for comparative learning. Density maps are cropped using protein domain atomic structures with or without protein domain type labels. Since most paired data (protein domain density maps and their corresponding atomic structures) lack protein domain type labels, to ensure sufficient CATH / SCOPe test set size, paired data with CATH / SCOPe labels can be selected first to evaluate CryoDiscovery. Then, data similar to the test data are removed by similarity, and the remaining paired data are used as the training set.

[0171] Figure 7 is a schematic diagram of the protein domain classification hierarchy in CATH / SCOPe according to an embodiment of the present invention. CATH / SCOPe independently classifies protein domains into four hierarchical levels to indicate the structural and functional similarity between protein domains, wherein higher hierarchical levels indicate closer associations.

[0172] To evaluate the method of this embodiment, atomic structure and density map pairs with CATH / SCOPe labels obtained in the above experiments can be used, and at least ten pairs of domain types are retained to ensure statistical significance. In this embodiment, a total of 8780 atomic structures containing 159 CATH domain types and 3038 atomic structures containing 79 SCOPe domain types are used to construct the database DateDB. Low-resolution maps can be used. Two test sets were constructed, containing 121 domain density maps labeled CATH and 116 domain density maps labeled SCOPe, respectively.

[0173] In this embodiment, it is ensured that the sequence identity between the structural domains in the test set and the structural domains used for training does not exceed 30%.

[0174] To verify the effectiveness of the method in the embodiments of the present invention, three different methods—CryoDiscovery, cryoID, and ModelAngelo—were used to identify protein domains from low-resolution density maps and compared. CryoID and ModelAngelo are both methods for identifying proteins from density maps.

[0175] Sequences with embedded protein domains were collected from CATH / SCOPe DataDB, converted to sequence databases provided by cryoID and ModelAngelo, and the domain type of the returned sequences was specified as the identification result.

[0176] Figure 8 is a schematic diagram of the embedding distribution of the structural domain types according to an embodiment of the present invention, reflecting the alignment of the density map and the atomic structure. The figure shows the embedding distribution of the four CATH structural domain types after each training step. After each training step, the atomic structure and density map embedding of the four CATH structural domain types are visualized using t-SNE.

[0177] As shown in parts A and B, after self-supervised training of the density tower and atomic tower, clusters of different domain types are almost indistinguishable. In contrast, as shown in parts A and C, contrastive learning can separate the atomic structure embeddings of different domain types. As shown in parts D and C, after the final cross-modal alignment, the embeddings of the density map tend to converge to the embeddings of the corresponding atomic structures. As shown in parts D and B, the density map embeddings of different domains no longer mix, indicating that cross-modal alignment unifies the embeddings of two modes in the same low-dimensional space, while ensuring that the embeddings of different domain types are different, while the embeddings of the same domain type are similar.

[0178] Figure 9 is a schematic diagram of protein domain identification using a low-resolution density map.

[0179] Figure 9A shows the Top-1, Top-5, and Top-10 accuracies and false negatives for protein domain identification in the 4-level hierarchy on the CATH / SCOPe test set.

[0180] Figure 9B shows the Top-1 accuracy across four hierarchical levels on the CATH / SCOPe test set. For the #ALL row, columns 1 through 4 indicate that the target domain was hit in the Top-1; for the #ADD row, level 3 indicates that the target protein domain was hit in the Top-1 of the level 3 hierarchy, but not in the level 4 hierarchy; the cases for levels 4, 2, and 1 are similar; WrongType indicates that the target protein domain was not hit in the Top-1 even in the level 1 hierarchy; Failure indicates that the detection result was empty.

[0181] In this embodiment, Top-K accuracy and mean precision (mAP) were calculated. As shown in Figure 7, CryoDiscovery outperforms cryoID and ModelAngelo in all metrics. CryoDiscovery's mAP is also higher than cryoID and ModelAngelo. The CATH / SCOPe hierarchy reflects the evolutionary relationships of protein domains. Although identification in the 4th level hierarchy may result in missed detections, hits in lower levels can still provide useful information. Therefore, this embodiment also calculated the Top-1 accuracy in the 3rd, 2nd, and 1st levels, finding that CryoDomian outperforms cryoID and ModelAngelo, further illustrating the accuracy of CryoDomian. Table 1 shows the Top-K accuracy and mean precision (mAP) of different methods on the CATH / SCOPe test set.

[0182] Table 1

[0183] In this embodiment, to investigate the impact of density map resolution, as shown in Table 1, the test set was divided into... and There are two groups. It can be seen that in... and Within its resolution range, cryoID fails to accurately identify protein domains in most cases. While ModelAngelo, however, does... At the lower resolution, half of the protein domains in the density maps were identified, but ModelAngelo's performance dropped sharply compared to the slight decrease in CryoDiscovery. To further investigate the relationship between resolution and recognition accuracy, 45 successful cases of ModelAngelo in the Top-5 of the CATH test set were low-pass filtered to generate low-resolution density maps. After removing 11 density maps with execution errors, the remaining 34 density maps were processed using resolution parameters of 5, 6, 7, 8, 9, and 1000 respectively. The cryoSPARC low-pass filter tool was used for processing.

[0184] Figure 10 is a schematic diagram illustrating the resolution impact of low-pass filtering. As shown in the figure, CryoDiscovery... Stable recognition performance was achieved in the high-resolution spectral maps, but... and It showed a gradual downward trend. In contrast, cryoID's performance has consistently been relatively poor, while ModelAngelo... Performance drops sharply at higher resolutions, with only 11 out of 34 cases succeeding in the Top-5, and failing completely at lower resolutions. These results demonstrate that CryoDiscovery has better accuracy and robustness over a wider resolution range, significantly outperforming cryoID and ModelAngelo.

[0185] Furthermore, embodiments of the present invention also provide a protein domain recognition device, comprising:

[0186] The dual-tower network module obtains density map embeddings and atomic structure embeddings through a dual-tower network based on density towers and atomic towers, and performs cross-modal alignment.

[0187] The database module constructs a structure embedding database that knows the density map based on density map embedding and atomic structure embedding after cross-modal alignment.

[0188] The query and recognition module, based on the input density map, queries the structural embedding database of the density map to obtain the closest atomic structure embedding and identify the protein domain.

[0189] Based on the same inventive concept, the present invention also provides a computer-readable storage medium storing one or more programs, which, when executed, can implement the aforementioned protein domain method.

[0190] As shown in Figure 11, this embodiment of the invention also provides a device, including a processor, a communication interface, a memory, and a communication bus. The processor, communication interface, and memory communicate with each other via the communication bus.

[0191] The memory is a computer-readable storage medium used to store one or more programs.

[0192] The processor is configured to execute a program stored in a computer-readable storage medium.

[0193] The computer-readable storage medium may be included in the device / apparatus described in the above embodiments; or it may exist independently and not assembled into the device / apparatus.

[0194] Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Images