Method for training amino acid sequence generation model and application thereof

Abstract

The application relates to a training method of an amino acid sequence generation model and application thereof. The method comprises the following steps: obtaining training data, wherein the training data comprises sample structure information of sample proteins; performing feature coding on the sample structure information to obtain node features and edge features corresponding to the sample proteins; alternately updating the node features and the edge features through multiple message passing layers of a graph neural network model to be trained to obtain updated target node features and target edge features; decoding and predicting each node in a random order through the graph neural network model to be trained according to the corresponding target node features and target edge features, and outputting predicted sequence information; and iteratively updating network parameters of the graph neural network model according to a loss value calculated based on the predicted sequence information and sample sequence information to obtain a trained amino acid sequence generation model. The scheme provided in the application can improve the generalization and robustness of the prediction result of the generation model.

Description

Technical Field

[0001] This application relates to the field of sequence prediction technology, and in particular to a training method for an amino acid sequence generation model and its application. Background Technology

[0002] Protein reverse folding refers to the computational process of deriving the amino acid sequence that can stably fold into a given three-dimensional protein backbone structure. Understanding the sequence-structure relationship of proteins is crucial for protein sequence prediction and design, and accurate amino acid sequence design is the core of creating new functional proteins.

[0003] In related technologies, the scarcity of training data and the limitations of existing models have led to a generalized problem of insufficient generalization ability. How to design an amino acid sequence generation model with strong generalization ability is a problem that needs to be solved. Summary of the Invention

[0004] To address or partially address the problems existing in related technologies, this application provides a training method for an amino acid sequence generation model and its application, which can improve the model's generalization and robustness.

[0005] The first aspect of this application provides a training method for an amino acid sequence generation model, comprising:

[0006] Acquire training data, which includes sample protein structure information and corresponding sample sequence information;

[0007] The sample structure information is feature-encoded to obtain the node features and edge features corresponding to the sample protein; wherein, the node features are used to represent residue information, and the edge features are used to represent the spatial geometric relationship between adjacent residues;

[0008] The node features and edge features are alternately updated by multiple message passing layers of the graph neural network model to be trained, so as to obtain the updated target node features and target edge features; wherein, the number of neighbor nodes aggregated by each message passing layer increases in sequence.

[0009] Based on the corresponding target node features and target edge features, the graph neural network model to be trained decodes and predicts each node in a random order, and outputs prediction sequence information.

[0010] The network parameters of the graph neural network model are iterated based on the loss value calculated from the predicted sequence information and the sample sequence information to obtain a trained amino acid sequence generation model.

[0011] A second aspect of this application provides a method for generating an amino acid sequence, comprising:

[0012] Obtain the main chain structure information of the target protein;

[0013] The main chain structure information is input into the amino acid sequence generation model described in the first aspect, and the corresponding target amino acid sequence is output.

[0014] A third aspect of this application provides a training apparatus for an amino acid sequence generation model, comprising:

[0015] The sample acquisition module is used to acquire training data, which includes sample protein structure information and corresponding sample sequence information.

[0016] The feature embedding module is used to encode the sample structure information to obtain the node features and edge features corresponding to the sample protein; wherein, the node features are used to represent residue information, and the edge features are used to represent the spatial geometric relationship between adjacent residues;

[0017] The message passing module is used to alternately update the node features and edge features through multiple message passing layers of the graph neural network model to be trained, so as to obtain the updated target node features and target edge features; wherein the number of neighbor nodes aggregated by each message passing layer increases sequentially.

[0018] The node decoding module is used to decode and predict each node in a random order according to the corresponding target node features and target edge features, and output the prediction sequence information.

[0019] The backpropagation module is used to iterate the network parameters of the graph neural network model based on the loss value calculated from the predicted sequence information and the sample sequence information to obtain a trained amino acid sequence generation model.

[0020] A fourth aspect of this application provides an apparatus for generating an amino acid sequence, comprising:

[0021] The information acquisition module is used to acquire the main chain structure information of the target protein;

[0022] The sequence generation module is used to input the main chain structure information into the amino acid sequence generation model trained by the training method described in the first aspect, and output the corresponding target amino acid sequence.

[0023] The fifth aspect of this application provides an electronic device, comprising:

[0024] Processor; and

[0025] A memory having executable code stored thereon, which, when executed by the processor, causes the processor to perform the method described in the first or second aspect above.

[0026] A sixth aspect of this application provides a computer-readable storage medium having executable code stored thereon, which, when executed by a processor of an electronic device, causes the processor to perform the method described in the first or second aspect above.

[0027] A seventh aspect of this application provides a computer program product comprising computer instructions that, when executed by a processor, implement the method described in the first or second aspect above.

[0028] The technical solution provided in this application may include the following beneficial results:

[0029] The training method for the amino acid sequence generation model in this application, during protein structure modeling, is based on a graph neural network framework. Nodes in the graph represent amino acid residues, and edges represent the spatial geometric relationships between residues. By introducing a multi-dimensional geometric encoding system, it overcomes the limitations of traditional distance threshold encoding, enabling the modeling of complex geometric systems during protein folding. In the graph neural network encoding process, the encoder takes defined graph node and edge features as input and updates these features alternately through multiple message-passing layers. During this process, the perceptual range of nodes is gradually expanded layer by layer by increasing the neighbor range. This hierarchical perception method effectively captures both local and global information in the graph, enabling the graph neural network to perceive protein structural features at multiple levels, thereby improving the model's ability to learn and predict protein sequences. In the graph neural network decoding process, a random mask matrix is ​​used to combine structural confidence with random noise, achieving order-independent decoding. Finally, by introducing structural confidence as a loss weight adjustment factor into the loss function, the contribution of high-confidence regions to model training is enhanced. At the same time, a hierarchical masking mechanism is used to filter low-confidence noise data, enabling the model to dynamically adjust the loss weight according to the confidence of each predicted amino acid residue. This ensures that the model pays more attention to high-confidence regions during training, avoids the impact of low-confidence regions on loss calculation, and effectively balances the model's prediction accuracy and generalization performance.

[0030] In addition, in terms of training data construction, an enhancement strategy of data fusion of real and predicted structures was adopted to expand the amount, diversity and representativeness of the data required for training, while ensuring the quality of the training data.

[0031] The amino acid sequence generation method of this application achieves amino acid sequence prediction based on the three-dimensional protein backbone by encoding the structural features of the target protein and updating the model based on a hierarchical sensing approach, thus ensuring the robustness of the prediction results and the generalizability of the application.

[0032] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application. Attached Figure Description

[0033] The above and other objects, features and advantages of this application will become more apparent from the more detailed description of exemplary embodiments thereof in conjunction with the accompanying drawings, wherein the same reference numerals generally represent the same components in the exemplary embodiments thereof.

[0034] Figure 1 This is a schematic flowchart illustrating the training method of the amino acid sequence generation model shown in the embodiments of this application;

[0035] Figure 2 This is another schematic flowchart illustrating the training method of the amino acid sequence generation model shown in the embodiments of this application;

[0036] Figure 3 This is a schematic flowchart illustrating the method for generating an amino acid sequence generation model according to an embodiment of this application;

[0037] Figure 4 This is a schematic diagram of the structure of the training device for the amino acid sequence generation model shown in the embodiments of this application;

[0038] Figure 5 This is a schematic diagram of the structure of the device for generating an amino acid sequence generation model as shown in the embodiments of this application;

[0039] Figure 6 This is a schematic diagram of the structure of an electronic device shown in an embodiment of this application. Detailed Implementation

[0040] Embodiments of this application will now be described in more detail with reference to the accompanying drawings. While embodiments of this application are shown in the drawings, it should be understood that this application may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to make this application more thorough and complete, and to fully convey the scope of this application to those skilled in the art.

[0041] The terminology used in this application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. The singular forms “a,” “the,” and “the” used in this application and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any or all possible combinations of one or more of the associated listed items.

[0042] It should be understood that although the terms "first," "second," "third," etc., may be used in this application to describe various information, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, without departing from the scope of this application, initial information may also be referred to as second information, and similarly, second information may also be referred to as initial information. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0043] Among related technologies, Graph Neural Networks (GNNs) are deep learning models based on graph-structured data, capable of effectively processing graphical data. In protein structure modeling, GNNs can be used to simulate the amino acid residues of proteins and their interactions. However, the scarcity of training data is one of the key bottlenecks in the current field of protein sequence design. Although a large number of protein structures have been resolved, experimentally obtained protein structures account for only 0.1% of the known protein sequence space. This lack of data volume and diversity significantly limits the performance of deep learning models. Furthermore, current protein sequence design methods oversimplify structural features during modeling. Limited by the scarcity of training data and the limitations of existing models, current protein sequence generation methods generally suffer from insufficient generalization ability, resulting in poor performance when handling unknown data.

[0044] To address the aforementioned issues, this application provides a training method for an amino acid sequence generation model and its application, which can improve the generalization and robustness of the model's prediction results.

[0045] The technical solutions of the embodiments of this application are described in detail below with reference to the accompanying drawings.

[0046] See Figure 1 and Figure 2 An embodiment of this application illustrates a training method for an amino acid sequence generation model, which includes:

[0047] S110, Obtain training data, which includes sample protein structure information and corresponding sample sequence information.

[0048] In this application, the training data comes from real experimental data and simulated prediction data. It is understood that since real experimental data is limited (e.g., over 100,000 data points), this application enriches the data using millions of high-confidence prediction data points. The sample protein's structural information includes the three-dimensional coordinates of the protein backbone, and the sample sequence information includes the corresponding amino acid sequence. Each residue in the amino acid sequence is selected from 20 natural amino acids (i.e., standard amino acids); optionally, some residues in the amino acid sequence of the sample protein can also be selected from non-standard amino acids to enrich the quantity and diversity of the sample proteins as much as possible.

[0049] To obtain a large amount of high-quality training data, some implementations may acquire the training data in the following manner: Wherein:

[0050] S111, according to the first preset percentage of sequence consistency, the real proteins are clustered to obtain multiple first protein clusters.

[0051] In this step, existing protein structures can be screened from various known protein databases or publicly available literature. For example, protein structures determined by X-ray crystallography or cryo-electron microscopy can be screened from the PDB (Protein Data Bank) database. To improve the data quality of proteins, protein structures with a resolution smaller than a preset value and / or a sequence length (i.e., the number of residues) smaller than a preset length threshold can be selected. The preset resolution value can be, for example, [missing value]. Or other values, the preset length threshold can be selected from 800 to 1000. For example, selecting a resolution smaller than... And protein structures with fewer than 1,000 residues. It can be understood that the smaller the resolution of a protein, the higher the accuracy of its structure, and consequently the higher the quality of the generated training data; an appropriate number of residues can improve computational efficiency and reduce the memory space occupied.

[0052] Furthermore, relevant tools can be used to cluster the selected real proteins according to a preset sequence identity, obtaining multiple first protein clusters. Sequence identity refers to the percentage of sites in any two aligned amino acid sequences that are identical. The preset percentage can be selected from 25% to 35%. For example, when the preset percentage is 30%, real proteins containing at least 30% of the same sites are selected and clustered together to form a corresponding first protein cluster. This process can be repeated to obtain over 100,000 first protein clusters, each containing multiple real proteins.

[0053] S112, the simulated proteins are clustered according to the dual clustering rules to obtain multiple second protein clusters. The dual clustering rules include clustering according to a second preset percentage of sequence overlap and / or a third preset percentage of sequence consistency, and clustering according to a fourth preset percentage of structural overlap.

[0054] In this step, simulated proteins can be selected from known protein structure prediction platforms (such as AlphaFold, Swiss-Model, I-TASSER, etc.) as the source of training data. The structures of these simulated proteins are generally predicted by artificial intelligence models; for example, the AlphaFold Database contains hundreds of millions of predicted simulated proteins. In this application, high-quality training data can be obtained by aggregating and filtering these predicted data using a dual clustering rule.

[0055] The dual clustering rules of this application include clustering based on sequence overlap and / or consistency, and clustering based on structural overlap. When multiple simulated proteins simultaneously meet the dual clustering rules of sequence clustering and structural clustering, they can be clustered into one cluster. In some specific embodiments, the dual clustering rules include: clustering simulated proteins with a sequence overlap of not less than a second preset percentage and / or a sequence consistency of not less than a third preset percentage, and a structural overlap of not less than a fourth preset percentage. Sequence overlap refers to the percentage of the total length of the shorter sequence that can be aligned with the longer sequence during the alignment of any two amino acid sequences. Optionally, the second preset percentage can be selected from 80% to 90%, the third preset percentage can be selected from 40% to 60%, and the fourth preset percentage can be selected from 80% to 90%. For example, multiple simulated proteins with a sequence overlap of not less than 90%, a sequence consistency of not less than 50%, and a structural overlap of not less than 90% are clustered to obtain corresponding second protein clusters, i.e., each second protein cluster contains multiple simulated proteins. Through clustering, more than 200 million simulated proteins can be clustered into 18 million second protein clusters.

[0056] S113, Identify sample proteins from each of the first and second protein clusters.

[0057] The sample proteins in the training data of this application consist of a mixture of real proteins and simulated proteins. In some embodiments, a preset number of real proteins are selected as sample proteins from each of the first protein clusters, and simulated proteins with a confidence level greater than a preset confidence threshold are selected as sample proteins from each of the second protein clusters.

[0058] Specifically, a fixed preset quantity can be set, and a preset quantity of real proteins can be selected from each of the first protein clusters. Alternatively, a random preset quantity can be set, and at least one real protein can be randomly selected from each of the first protein clusters.

[0059] After obtaining multiple second protein clusters, to improve the data quality of sample proteins, simulated proteins with low confidence can be filtered out for data augmentation, i.e., simulated proteins with higher confidence are selected as sample proteins. To obtain simulated proteins with confidence levels meeting a preset confidence threshold, in some embodiments, the confidence level of each simulated protein is obtained separately; simulated proteins in each second protein cluster with confidence levels greater than the preset confidence threshold are retained as sample proteins. Optionally, the confidence level in this application can be characterized using the average pLDDT score of the protein. Of course, in other embodiments, other indicators can also be selected to represent the confidence level, and there is no limitation here. pLDDT (Predicted Local Distance Difference Test) is a confidence index for evaluating protein structure prediction, with a value range of 0 to 100. A higher value indicates higher prediction accuracy, and its value is automatically evaluated by the AlphaFold Database. The average pLDDT score is the average of the pLDDT scores of each residue in a single protein.

[0060] In this embodiment, the average pLDDT value of each protein can be obtained as an evaluation index of the corresponding confidence level. Proteins in each second protein cluster whose average pLDDT value is greater than a preset confidence threshold are then retained. The preset confidence threshold can be selected from 80 to 90; for example, simulated proteins with an average pLDDT value greater than 80 are retained as sample proteins. It is understood that the number of sample proteins determined in each second protein cluster may differ. To obtain more balanced structural data, in some embodiments, a maximum and / or minimum value for the number of sample proteins selected in each second protein cluster can be preset. That is, the number of qualified sample proteins selected from each second protein cluster does not exceed the maximum value, and / or the number of qualified sample proteins selected from each second protein cluster is not less than the minimum value. Furthermore, the simulated proteins in the second protein clusters can be sorted according to their confidence level, and simulated proteins with higher confidence levels can be selected as sample proteins.

[0061] By filtering based on the aforementioned confidence level, 3.2 million high-quality simulated protein data points can be obtained, effectively expanding the quality and diversity of model training data.

[0062] In this step, the first protein clusters obtained in S111 and the second protein clusters in S112 can be deduplicated in advance, that is, redundant sequences and proteins with the same structure are deleted before sample proteins are selected from them.

[0063] It is understandable that the sample proteins are selected from both real and simulated proteins, meaning that the training data contains both real and simulated proteins, making the training data both reliable and rich. As the model training continues, after multiple epochs, such as 8 to 10 epochs, proteins can be reselected to form new training data.

[0064] Furthermore, the training data is pre-divided into training, validation, and test sets, with different protein samples in each set. This avoids data leakage and overfitting, ensuring the generalization ability and evaluation accuracy of the trained model in practical applications.

[0065] S120: Feature encoding is performed on the sample structure information to obtain the node features and edge features corresponding to the sample protein; where node features are used to represent residue information and edge features are used to represent the spatial geometric relationship between adjacent residues.

[0066] For each sample protein in the training data, its structural information can be encoded according to a preset feature dimension. The amino acid sequence generation model in this application is based on a graph neural network architecture, where the nodes of the graph represent each amino acid residue in the protein, and the edges represent the spatial geometric relationships between adjacent residues.

[0067] In some implementations, each node feature includes the periodic trigonometric function values ​​corresponding to the torsion angles φ and ψ of the residues, the secondary structure, and the structural confidence of the residues. Each node feature includes 25 preset feature dimensions, and each node feature is a feature vector composed of 25 features, specifically including:

[0068] The periodic trigonometric function values ​​corresponding to the torsion angles φ and ψ of the residues include: 16-dimensional periodic features obtained by expanding the torsion angles φ and ψ using a fourth-order Fourier transform, namely sin(nφ), cos(nφ), and sin(nψ), cos(nψ), where n = 1 to 4. When n = 1, they are sin(φ), cos(φ), sin(φ), and cos(φ) of the residues, respectively; when n = 2, they are sin(2φ), cos(2φ), sin(2φ), and cos(2φ) of the residues, respectively, and so on, resulting in a total of 16 feature values.

[0069] The secondary structures of residues include: α-helix, 3... 10There are a total of 8 secondary structures, including spiral, π-spiral, extended β-chain, isolated β-bridge, turn, bend, and coil / irregular structure, corresponding to 8-dimensional features. The corresponding features of each secondary structure can be represented by one-hot encoding.

[0070] The structural confidence score of a residue is expressed as either a pLDDT score or 100. When the sample protein is a simulated protein, the structural confidence score can be represented by the pLDDT score of each residue automatically evaluated by the AlphaFold Database. When the sample protein is a real protein, the structural confidence score can be set to 100 by default. The structural confidence score of a residue can be represented after normalization.

[0071] In some implementations, a single edge feature includes distance information between two residues based on predetermined atoms and a confidence level of the relative positions of the two residues. Taking residues i and j in a protein as examples, each residue has N atoms and C atoms on its main chain. α Atoms, C atoms, and O atoms, including a virtual C atom simulating a side chain position. β Atoms are obtained by pairing any two atoms from two residues, resulting in 25 atomic combinations. The distance information corresponding to each atomic combination is then obtained. Furthermore, the distance information of each atomic combination can be calculated using multiple Radial Basis Functions (RBFs) to obtain the distance features calculated for each RBF. Taking 16 RBFs as an example, a total of 25 × 16 = 400 dimensions of distance features can be obtained. Among these, virtual C... β Atoms are not real atoms, but rather artificially defined virtual atoms. Each residue contains a virtual C. β The coordinates of an atom can be determined according to a preset formula. Virtual C β The formula for calculating atomic coordinates is:

[0072] C β =-0.58273431×a+0.56802827×b-0.54067466×c+C α

[0073] Where, vector b = C α -N, vector c = CC α The normal vector a = cross(b,c).

[0074] In some implementations, the confidence level of the relative positions of two residues can be represented using PAE (Predicted Alignment Error). PAE is a metric used in protein structure prediction tools such as AlphaFold to evaluate the accuracy of protein structure predictions. A higher PAE value indicates a larger deviation between the predicted and actual structures, and vice versa. When the sample protein is a simulated protein, the PAE is automatically evaluated, for example, by the AlphaFold Database. When the sample protein is a real protein, the default PAE value is 0.01.

[0075] In summary, a single node feature can be a feature vector composed of 25-dimensional features, and a single edge feature can be a feature vector composed of 401-dimensional features. In this step, for each sample protein, corresponding 25-dimensional node features and 401-dimensional edge features are encoded based on the corresponding sample structure information, and these serve as the initial node features and initial edge features input to the encoder in subsequent steps.

[0076] This application captures the periodicity of the torsion angle of residues through fourth-order Fourier expansion and employs virtual C β A spatial reference frame for constructing virtual side chains is used, and multiple RBFs are combined to finely encode the interatomic spacing, integrating secondary structure and confidence features. This design, through a multi-scale geometric feature fusion mechanism, overcomes the limitations of traditional distance threshold encoding, enabling the modeling of complex geometric systems during protein folding. This improves the modeling ability for complex geometric constraints of proteins and enhances the model's understanding of protein structure.

[0077] S130, the node features and edge features are alternately updated through multiple message passing layers of the graph neural network model to be trained, so as to obtain the updated target node features and target edge features; wherein, the number of neighbor nodes aggregated by each message passing layer increases in sequence.

[0078] The graph neural network model in this application includes an encoder and a decoder. In this step, the encoder comprises multiple message-passing layers. The encoder takes the features of each node and edge in the graph as input and updates these features alternately through multiple message-passing layers to obtain updated target node and edge features. Each message-passing layer aggregates the node features of a predetermined number of neighboring nodes and the edge features connected to each neighboring node. The next message-passing layer aggregates more neighboring nodes than the previous one. By progressively increasing the number of aggregated neighboring nodes, the perceptual range of the nodes is expanded, thereby helping to capture structural information of different scales. This hierarchical perception method effectively captures local and global information in the graph, enabling the graph neural network to perceive the structural features of proteins at multiple levels, thus improving the model's ability to learn and predict protein sequences.

[0079] This application has at least two message passing layers. In some embodiments, the multiple message passing layers include at least a first message passing layer and a second message passing layer. The first message passing layer is used to aggregate the X neighboring nodes of the current node and their corresponding connecting edges, and the second message passing layer is used to aggregate the Y neighboring nodes of the current node and their corresponding connecting edges. Preferably, the message passing layer further includes a third message passing layer, which is used to aggregate the Z neighboring nodes of the current node and their corresponding connecting edges; 0 < X ​​< Y < Z, where X, Y, and Z are all natural numbers. For example, X can be selected from 3 to 12, Y can be selected from 13 to 24, and Z can be selected from 25 to 48. In some specific embodiments, X can be 12, Y can be 24, and Z can be 48. That is, the first message passing layer is used to aggregate the 12 nearest neighboring nodes of the current node in space, the second message passing layer is used to aggregate the 24 nearest neighboring nodes of the current node in space, and the third message passing layer is used to aggregate the 48 nearest neighboring nodes of the current node in space. The number of neighbor nodes in each message passing layer can increase randomly or in a certain regular pattern, such as arithmetic or geometric progression, without restriction.

[0080] In some specific implementations, the initial node features of the current node are updated by aggregating X neighboring nodes and their corresponding connecting edges through a first message passing layer, thus obtaining the first node features corresponding to the current node. Based on the first node features of the current node and the current node features of the neighboring nodes, the initial edge features of the connecting edges are updated, thus obtaining the first edge features corresponding to the current connecting edge. Similarly, the first node features of the current node are updated by aggregating Y neighboring nodes and their corresponding connecting edges through a second message passing layer, thus obtaining the second node features corresponding to the current node. Finally, the first edge features of the connecting edges are updated based on the second node features of the current node and the current node features of the neighboring nodes, thus obtaining the second edge features corresponding to the current connecting edge. When a third message passing layer exists, the second node features of the current node are updated by aggregating Z neighboring nodes and their corresponding connecting edges through the third message passing layer, thus obtaining the target node features corresponding to the current node. Finally, the second edge features of the connecting edges are updated based on the target node features of the current node and the current node features of the neighboring nodes, thus obtaining the target edge features corresponding to the current connecting edge.

[0081] In other words, the only difference between each message passing layer is the number of neighbor nodes aggregated. Multiple message passing layers sequentially update node features. The node features updated by the first message passing layer serve as the input data for the second message passing layer, and the node features updated by the second message passing layer serve as the input data for the third message passing layer. The node features updated by the third message passing layer then serve as the target node features. Simultaneously, for each message passing layer, the node features are updated first, followed by the corresponding edge features. This alternating updating of node and edge features continues until the last message passing layer outputs the final target node features.

[0082] For ease of understanding, let's take any node i as the current node to be updated as an example. Combining the node feature update formula (1) below, the initial node feature V is updated through the first message passing layer. i Perform an update to obtain the updated first node feature V' i Then, combining the following edge feature update formula (2) with its initial edge feature E ij Perform an update to obtain the updated first-side feature E' ij The node feature update formula is shown below:

[0083]

[0084] Where i represents the current node, j represents the neighboring node, V represents the node feature, and E represents the edge feature.

[0085] As can be understood, a neighboring node refers to the node closest to the current node in the current space. Taking the aggregation of 12 neighboring nodes in the first message passing layer as an example, the node features V are first processed by a multilayer perceptron (MLP1). i Node characteristics V of neighboring nodes j and the edge features E between the current node and its neighboring nodes. ij The process generates 12 edge information entries for each node i. These 12 edge information entries are then aggregated to obtain aggregated information, which is further processed using a Multilayer Perceptron (MLP2). Preferably, Dropout is applied to the processed aggregated information to prevent overfitting. Specifically, the Dropout layer randomly zeros out 10%–20% of the aggregated information, and then the dropped information is compared with the initial node features V. i Add them together, and then normalize them to obtain the updated node features V' of the current message passing layer. i Optionally, the hidden dimension of each multilayer perceptron in this application is 512 dimensions.

[0086] To facilitate differentiation, the node features updated by the first message passing layer are defined as the first node features. After obtaining the first node features corresponding to the current node, the update formula (2) for the edge features is used for understanding, and the edge features E are... ij Perform an update to obtain the first edge feature E' after the current first message passing layer update. ij .

[0087] E′ ij =LayerNorm(E ij +Dropout(MLP3([V′ i , V′ j E ij ]))) (2)

[0088] Among them, the updated first node feature V' is obtained through a multilayer perceptron (MLP3). i The current node feature V' corresponding to neighbor node j j and the initial edge features E ij Perform concatenation and aggregation. Preferably, similarly, Dropout (random discarding) can be applied to the processed aggregated information to prevent overfitting. Then, the Dropout-enhanced information is combined with the initial edge features E. ij Add them together, and then normalize the layers to obtain the updated first-side feature E' of the current message passing layer. ij .

[0089] It should be noted that the current node feature V' corresponding to neighbor node j jThis refers to the latest feature of node j. The current feature of neighboring node j may be the initial node feature that has not yet been updated, or it may be the first node feature that has been updated, depending on the actual situation.

[0090] After obtaining the first node features corresponding to the current node i and the 12 first edge features corresponding to the current node i and its 12 neighboring nodes j through the first message passing layer in sequence, the updated second node features and 24 second edge features are obtained sequentially through the second message passing layer based on the latest feature information as input, following the same principle. If there is no third message passing layer, feature updating stops, and the second node features and 24 second edge features are used as the target node features and target edge features corresponding to the current node i to be output by the encoder. If there is a third message passing layer, the second node features and 24 second edge features are used as input to the third message passing layer, and the updated third node features and 48 third edge features are obtained sequentially according to the same principle. This process continues until the last message passing layer outputs the target node features and target edge features corresponding to the current node i. Through this multi-level perception update method, the features of each node and edge are updated, providing more accurate input for the subsequent decoding process.

[0091] It is understandable that if the number of residues in a single protein does not exceed 1000, then the corresponding number of nodes does not exceed 1000. After updating the node features and edge features of the first node, the node features and edge features of the second node are updated in the same way as the multi-layer message passing layer update described above, until the feature updates of all nodes are completed, so as to serve as the input data for the subsequent decoder.

[0092] To further reduce the negative impact of the quality of the training data itself on the reliability of the training results, some implementations employ a masking matrix to mask abnormal residues in each sample protein during the information aggregation process of each message passing layer. Specifically, for simulated proteins in the training data, abnormal residues refer to residues with a structural confidence score (e.g., a pLDDT score below 90); for real proteins in the training data, abnormal residues refer to residues with non-standard amino acids or missing coordinates. It is understood that these masked abnormal residues do not need to participate in the updating of corresponding node features. For example, in the process of processing aggregated information using the multilayer perceptron (MLP2) described above, these abnormal residues are masked, and therefore no message passing occurs. These masked abnormal nodes also do not need to participate in the prediction of node residue types and the calculation of the loss function in subsequent decoding steps.

[0093] To distinguish between abnormal and normal residues in a sample protein, some implementations use corresponding preset codes to label abnormal and normal residues in the mask matrix. Abnormal residues are labeled as 0, and normal residues are labeled as 1.

[0094] S140: Based on the corresponding target node features and target edge features, the graph neural network model to be trained decodes and predicts each node in a random order, and outputs the prediction sequence information.

[0095] In this step, all the updated target node features and target edge features from the previous steps are input into the decoder of the graph neural network model for decoding. Based on the autoregressive masking mechanism, the residue type corresponding to each node is predicted one by one. The residue types include 20 natural amino acids.

[0096] Furthermore, the decoder of this application has a symmetrical design with the encoder, that is, the number of message passing layers in the decoder is the same as the number of message passing layers in the encoder, and the number of neighbor nodes aggregated in each layer corresponds to the same value. The operating mechanism of each message passing layer in the decoder is the same as that in the encoder. For example, if the encoder has three message passing layers and aggregates 12, 24, and 48 neighbor nodes sequentially, then similarly the decoder has a first message passing layer, a second message passing layer, and a third message passing layer, each corresponding to aggregate 12, 24, and 48 neighbor nodes sequentially.

[0097] In some implementations, the current node to be predicted is determined based on a random mask matrix; the target node features of the node to be predicted are updated sequentially and cyclically according to multiple message passing layers in the decoder to obtain the corresponding node update features; and the corresponding residue type is predicted and the corresponding prediction sequence information is generated based on the residue type probability distribution transformed from the node update features.

[0098] It should be noted that traditional autoregressive decoding mechanisms typically rely on a fixed predicted order of residues to generate an amino acid sequence. This approach fails to adequately consider the structural characteristics of proteins and lacks a dynamic adjustment mechanism. Due to the high complexity of protein folding, fixed-order decoding strategies cannot adapt to their variable folding paths, thus limiting the flexibility and adaptability of the model.

[0099] Specifically, unlike related technologies that use a fixed prediction order based on the amino acid sequence from beginning to end, this application employs a random mask matrix to randomly determine the nodes to be predicted at each step. The node prediction order for different sample proteins is randomly determined. In the node prediction of the current step, the target node features input from the encoder are updated through multiple message passing layers of the decoder. The mechanism of each message passing layer of the decoder is the same as that of the encoder, the only difference being the number of neighbor nodes aggregated. The input data of the next message passing layer is the output data of the previous message passing layer. The last message passing layer only needs to output the updated node features of the last update and does not need to output updated edge features.

[0100] For example, taking the mechanism of the first message passing layer of the decoder as an example, the target node features corresponding to the determined node to be predicted, the target edge features of the 12 connecting edges, and the sequence relative position features corresponding to the current node are concatenated to obtain the corresponding concatenated features. Then, the concatenated features are aggregated with the target node features corresponding to the 12 neighboring nodes to obtain the corresponding node aggregated features. Next, the node aggregated features are further updated through layer normalization, Dropout, and a feedforward network to obtain the node updated features after the first message passing layer. After the node feature update is completed in the first message passing layer, the target edge features are alternately updated in the same way as the encoder to obtain the corresponding 12 edge updated features. The node updated features and the 12 edge updated features are input into the second message passing layer for further updates to obtain new node updated features and 24 new edge updated features. The node updated features and the 24 edge updated features are input into the third message passing layer for further updates to obtain the final node updated features and 48 new edge updated features. Based on the final node updated features, the probability distribution of the 20 natural amino acids corresponding to the current node is output through a linear layer and a softmax function. Based on these probability distributions, the amino acid with the highest probability is selected as the prediction result for the current node. Based on the predicted residue types for each node, the complete amino acid sequence of the sample protein is constructed, thus obtaining the predicted sequence information.

[0101] Furthermore, to overcome the limitation of autoregressive decoding relying on a fixed generation order (from left to right), the random mask matrix T' of this application is obtained by random permutation based on the original triangular mask T. In some embodiments, the random mask matrix T' is obtained according to the following formula (3), specifically:

[0102] T' = PTP T (3)

[0103] Here, P is a random mask matrix with the same size as T. T is an L×L square matrix, where L represents the length of the protein's amino acid sequence. The role of matrix T is to indicate which nodes at each step in the decoding process need to be masked, so that the decoder cannot directly access the masked portion of the input sequence. The elements of the mask matrix are 0 or 1, where 1 indicates that the node at that position is not masked, and 0 indicates that the node at that position is masked.

[0104] As shown above, by using a random mask matrix T', the rows and columns of the mask are shuffled simultaneously, so that the positions of nodes that should be adjacent in order are not generated in a fixed order, thus achieving order-independent decoding.

[0105] S150: Based on the loss value calculated from the predicted sequence information and the sample sequence information, iterate the network parameters of the graph neural network model to obtain the trained amino acid sequence generation model.

[0106] In this step, the loss value is calculated based on the generated predicted sequence information and the real sample sequence information, where the real sample sequence information is the training label. To improve the accuracy of the prediction results, the loss function used in the graph neural network model of this application is a hybrid loss function that combines the structural confidence of each node in the sample protein for weighted adjustment. In some specific implementations, the loss value of the predicted sequence information and the sample sequence information is calculated according to the following formula (4), as shown below:

[0107]

[0108] Among them, -log(probs(S) i The negative log-likelihood loss (probs(S)) measures the log difference between the predicted sequence information and the sample sequence information. i ) is the probability value of the residue class predicted for node i, pLDDT i Mask represents the structural confidence level corresponding to each residue. i It can be 0 or 1, used to ensure that only the loss of normal nodes that are not shaded is calculated.

[0109] By employing the aforementioned hybrid loss function that includes structural confidence weighting, and using the structural confidence represented by pLDDT as the loss weight adjustment factor, the contribution of high-confidence regions to model training is enhanced. At the same time, a hierarchical masking mechanism is used to filter low-confidence noise data. This dynamic optimization strategy effectively balances prediction accuracy and generalization performance.

[0110] After calculating the loss value, backpropagation can be performed based on the loss value to adjust the network parameters of each hidden layer of the encoder and decoder. This process is repeated iteratively until the calculated loss value converges, and the trained graph neural network model is obtained as an amino acid sequence generation model for application in the generation and prediction of amino acid sequences.

[0111] It is understandable that using validation and test sets that are completely different from the training set to validate and test the trained graph neural network model, and strictly performing sequence and structure deduplication, ensures that the training set, validation set, and test set are independent of each other at the sequence and structure level, thereby effectively avoiding data leakage and overfitting problems.

[0112] In summary, the training method for the amino acid sequence generation model in this application employs an enhancement strategy of fusing real and predicted structures in the construction of training data. This expands the amount, diversity, and representativeness of the data required for training while ensuring the quality of the training data. In the protein structure modeling process, a graph neural network framework is used, where nodes represent amino acid residues and edges represent the spatial geometric relationships between residues. By introducing a multi-dimensional geometric encoding system, the limitations of traditional distance threshold encoding are overcome, enabling the modeling of complex geometric systems during protein folding. During the graph neural network encoding process, the encoder takes the defined graph node and edge features as input and updates these features alternately through multiple message passing layers. The perceptual range of nodes is gradually expanded through three layers by progressively increasing the neighbor range. This hierarchical perception method effectively captures both local and global information in the graph, enabling the graph neural network to perceive protein structural features at multiple levels, thereby improving the model's learning and prediction capabilities for protein sequences. In the graph neural network decoding process, a random mask matrix is ​​used to combine structural confidence with random noise, achieving order-independent decoding. Finally, by introducing structural confidence as a loss weight adjustment factor into the loss function, the contribution of high-confidence regions to model training is enhanced. At the same time, a hierarchical masking mechanism is used to filter low-confidence noise data, enabling the model to dynamically adjust the loss weight according to the confidence of each predicted amino acid residue. This ensures that the model pays more attention to high-confidence regions during training, avoids the impact of low-confidence regions on loss calculation, and effectively balances the model's prediction accuracy and generalization performance.

[0113] See Figure 3 An embodiment of this application also provides a method for generating an amino acid sequence, comprising:

[0114] S210: Obtain the main chain structure information of the target protein.

[0115] The amino acid sequence generation model of this application is applied to the protein reverse folding task. Based on the three-dimensional structure of the target protein, it obtains the main chain structure information and then predicts the amino acid sequence that can stably fold into that three-dimensional structure. The main chain structure information includes the three-dimensional coordinates of each heavy atom on the backbone. Optionally, the target protein can be a known real protein or a virtual protein; there is no limitation on this.

[0116] Based on the main chain structure information, node features corresponding to each residue in the target protein can be generated by encoding the embedding module, and edge features connecting each node to 12–48 neighboring nodes can also be generated. In some implementations, each node feature includes:

[0117] The torsion angles φ and ψ of the residues are obtained as 16-dimensional periodic features through fourth-order Fourier expansion: sin(nφ), cos(nφ), and sin(nψ), cos(nψ), where n = 1 to 4;

[0118] Secondary structure of 8-dimensional residues: α-helix, 3 10 - Helices, π-helices, extended β-chains, isolated β-bridges, turns, bends, and coiled / irregular structures;

[0119] Structural confidence of 1-dimensional residues: pLDDT score.

[0120] In other words, each node feature contains at most 25 dimensions.

[0121] In some implementations, each edge feature includes:

[0122] The N atom and C atom between residue i and residue j respectively α Atoms, C atoms, O atoms and virtual C β The distance information between any two atoms of an atom;

[0123] Confidence level of the predicted relative positions of residue i and residue j.

[0124] Preferably, the distance information between any two atoms can be distance information encoded by multiple radial basis functions (RBFs). When the number of radial basis functions is 16, each edge feature contains at most 401-dimensional features. Calculated using 16 RBFs, a total of 25 × 16 = 400-dimensional geometric distance features can be obtained.

[0125] S220: Input the main chain structure information into the trained amino acid sequence generation model, and output the corresponding target amino acid sequence.

[0126] The node features and related edge features corresponding to each residue of the target protein are input into the amino acid sequence generation model. Following a random decoding order, the model predicts the amino acid type corresponding to each residue one by one, sequentially assembling the target amino acid sequence. The amino acid sequence generation model updates the node features hierarchically according to the multiple message passing layers in the encoder and decoder described above; this will not be elaborated further here.

[0127] As can be seen from this example, the amino acid sequence generation method of this application, by encoding the structural features of the target protein and updating the amino acid sequence generation model in a hierarchical sensing manner, achieves the prediction of the amino acid sequence based on the three-dimensional protein backbone, ensuring the robustness of the prediction results and the generalizability of the application.

[0128] Corresponding to the aforementioned application function implementation method embodiments, this application also provides a training device, generation device, electronic device, and corresponding embodiments for an amino acid sequence generation model.

[0129] Figure 4 This is a schematic diagram of the structure of the training device for the amino acid sequence generation model shown in the embodiments of this application.

[0130] See Figure 4 An embodiment of this application illustrates a training apparatus for an amino acid sequence generation model, which includes a sample acquisition module 310, a feature embedding module 320, a message passing module 330, a node decoding module 340, and a backpropagation module 350. Wherein:

[0131] The sample acquisition module 310 is used to acquire training data, which includes sample protein structure information and corresponding sample sequence information.

[0132] The feature embedding module 320 is used to encode the sample structure information to obtain the node features and edge features corresponding to the sample protein; wherein, the node features are used to represent residue information, and the edge features are used to represent the spatial geometric relationship between adjacent residues.

[0133] The message passing module 330 is used to alternately update the node features and edge features through multiple message passing layers of the graph neural network model to be trained, so as to obtain the updated target node features and target edge features; wherein, the number of neighbor nodes aggregated by each message passing layer increases sequentially.

[0134] The node decoding module 340 is used to decode and predict each node in a random order according to the corresponding target node features and target edge features, and output the prediction sequence information.

[0135] The backpropagation module 350 is used to iterate the network parameters of the graph neural network model based on the loss value calculated from the predicted sequence information and the sample sequence information to obtain the trained amino acid sequence generation model.

[0136] In some specific implementations, the sample acquisition module 310 is used to cluster each real protein according to a first preset percentage of sequence consistency to obtain multiple first protein clusters; to cluster the simulated protein according to a dual clustering rule to obtain multiple second protein clusters, wherein the dual clustering rule includes clustering according to a second preset percentage of sequence overlap and / or a third preset percentage of sequence consistency, and clustering according to a fourth preset percentage of structural overlap; and to determine the sample protein from each of the first protein clusters and the second protein clusters.

[0137] In some specific implementations, the sample acquisition module 310 determines sample proteins from each first protein cluster and second protein cluster, including: selecting a preset number of real proteins as sample proteins from each first protein cluster, and selecting simulated proteins with a confidence level greater than a preset confidence threshold from each second protein cluster as sample proteins.

[0138] In some specific implementations, the message passing module 330 is used to aggregate X neighboring nodes and their corresponding connecting edges through a first message passing layer, update the initial node features of the current node, and obtain the first node features corresponding to the current node; update the initial edge features of the connecting edges based on the first node features of the current node and the current node features of the neighboring nodes, and obtain the first edge features corresponding to the current connecting edge; aggregate Y neighboring nodes and their corresponding connecting edges through a second message passing layer, update the first node features of the current node, and obtain the second node features corresponding to the current node; update the first edge features of the connecting edges based on the second node features of the current node and the current node features of the neighboring nodes, and obtain the second edge features corresponding to the current connecting edge; when a third message passing layer exists, aggregate Z neighboring nodes and their corresponding connecting edges through the third message passing layer, update the second node features of the current node, and obtain the target node features corresponding to the current node; update the second edge features of the connecting edges based on the target node features of the current node and the current node features of the neighboring nodes, and obtain the target edge features corresponding to the current connecting edge.

[0139] In some specific implementations, the node decoding module 340 is used to determine the node to be predicted in the current step based on the random mask matrix; based on the corresponding target node features and target edge features, the target node features of the node to be predicted are updated sequentially and cyclically according to multiple message passing layers in the decoder to obtain the corresponding node update features; based on the residue type probability distribution transformed by the node update features, the corresponding residue type is predicted and the corresponding prediction sequence information is generated.

[0140] In some specific implementations, the backpropagation module 350 is used to calculate the loss value calculated from the predicted sequence information and the sample sequence information based on the mixed loss function, and to backpropagate the network parameters of the iterative graph neural network model based on the loss value to obtain the trained amino acid sequence generation model.

[0141] As this example demonstrates, the training device for the amino acid sequence generation model in this application obtains large-scale and high-quality training data through data augmentation strategies and extracts data features through a multi-dimensional geometric feature encoding system, thereby improving the modeling ability for complex geometric constraints of proteins and enhancing the model's understanding of protein structures. By utilizing a random permutation matrix and combining structural confidence with random noise, order-independent decoding is achieved. A hybrid loss function with confidence weighting is constructed to overcome the generalization ability deficiencies caused by data scarcity and model limitations. This design, combining the expansion of the training dataset, the hierarchical awareness update method, and the construction of the weighted loss function, effectively enhances the robustness and generalization ability of the amino acid sequence generation model on unknown data.

[0142] Figure 5 This is a schematic diagram of the structure of the device for generating an amino acid sequence generation model as shown in the embodiments of this application.

[0143] See Figure 5 This application embodiment illustrates an amino acid sequence generation device, which includes: an information acquisition module 410 and a sequence generation module 420. Wherein:

[0144] The information acquisition module 410 is used to acquire the main chain structure information of the target protein;

[0145] The sequence generation module 420 is used to input the main chain structure information into the trained amino acid sequence generation model and output the corresponding target amino acid sequence.

[0146] As can be seen from this example, the amino acid sequence generation device of this application can be widely used to predict the amino acid sequence of unknown proteins, ensuring the generalizability and robustness of the prediction results.

[0147] Regarding the apparatus in the above embodiments, the specific manner in which each module performs its operation has been described in detail in the embodiments related to the method, and will not be elaborated further here.

[0148] Figure 6 This is a schematic diagram of the structure of an electronic device shown in an embodiment of this application.

[0149] See Figure 6 The electronic device 1000 includes a memory 1010 and a processor 1020.

[0150] The processor 1020 can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor.

[0151] Memory 1010 may include various types of storage units, such as system memory, read-only memory (ROM), and permanent storage devices. ROM may store static data or instructions required by processor 1020 or other modules of the computer. Permanent storage devices may be read-write storage devices. Permanent storage devices may be non-volatile storage devices that retain stored instructions and data even when the computer is powered off. In some embodiments, permanent storage devices use mass storage devices (e.g., magnetic or optical disks, flash memory) as permanent storage devices. In other embodiments, permanent storage devices may be removable storage devices (e.g., floppy disks, optical drives). System memory may be a read-write storage device or a volatile read-write storage device, such as dynamic random access memory. System memory may store some or all of the instructions and data required by the processor during operation. Furthermore, memory 1010 may include any combination of computer-readable storage media, including various types of semiconductor memory chips (e.g., DRAM, SRAM, SDRAM, flash memory, programmable read-only memory), and disks and / or optical disks may also be used. In some embodiments, memory 1010 may include a removable storage device that is readable and / or writable, such as a laser disc (CD), a read-only digital multifunction optical disc (e.g., DVD-ROM, dual-layer DVD-ROM), a read-only Blu-ray disc, an ultra-high-density optical disc, a flash memory card (e.g., SD card, mini SD card, Micro-SD card, etc.), a magnetic floppy disk, etc. Computer-readable storage media do not contain carrier waves or transient electronic signals transmitted wirelessly or via wired connections.

[0152] The memory 1010 stores executable code, which, when processed by the processor 1020, can cause the processor 1020 to execute part or all of the methods described above.

[0153] Furthermore, the method according to this application can also be implemented as a computer program or computer program product, which includes computer program code instructions for performing some or all of the steps in the method described above.

[0154] Alternatively, this application may also be implemented as a computer-readable storage medium (or a non-transitory machine-readable storage medium or a machine-readable storage medium) storing executable code (or computer program or computer instruction code) that, when executed by a processor of an electronic device (or server, etc.), causes the processor to perform part or all of the steps of the methods described above according to this application.

[0155] This application also provides a computer program product, which includes computer instructions that, when executed by a processor, implement the method described above.

[0156] The various embodiments of this application have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or improvement of the technology in the market, or to enable others skilled in the art to understand the embodiments disclosed herein.

Claims

1. A training method for an amino acid sequence generation model, characterized in that, include: Acquire training data, which includes sample protein structure information and corresponding sample sequence information; The sample structure information is feature-encoded to obtain the node features and edge features corresponding to the sample protein; wherein, the node features are used to represent residue information, and the edge features are used to represent the spatial geometric relationship between adjacent residues; The node features and edge features are alternately updated by multiple message passing layers of the graph neural network model to be trained, so as to obtain the updated target node features and target edge features; wherein, the number of neighbor nodes aggregated by each message passing layer increases in sequence. Based on the corresponding target node features and target edge features, the graph neural network model to be trained decodes and predicts each node in a random order, and outputs prediction sequence information. The network parameters of the graph neural network model are iterated based on the loss value calculated from the predicted sequence information and the sample sequence information to obtain a trained amino acid sequence generation model.

2. The training method according to claim 1, characterized in that: The individual node features include the periodic trigonometric function values ​​corresponding to the torsion angles φ and ψ of the residue, the secondary structure, and the structural confidence of the residue; Each edge feature includes the distance information between two residues based on a preset atom and the confidence level of the relative positions of the two residues.

3. The training method according to claim 1, characterized in that, In the encoder of the graph neural network model, the plurality of message passing layers include at least a first message passing layer and a second message passing layer; The first message passing layer is used to aggregate the X neighboring nodes of the current node and their corresponding connection edges, and the second message passing layer is used to aggregate the Y neighboring nodes of the current node and their corresponding connection edges.

4. The training method according to claim 3, characterized in that, The message passing layer also includes a third message passing layer, which is used to aggregate the Z neighbor nodes of the current node and their corresponding connecting edges; 0 < X ​​< Y < Z, where X, Y, and Z are all natural numbers.

5. The training method according to claim 4, characterized in that, The process involves alternately updating node features and edge features through multiple message-passing layers of the graph neural network model to be trained, to obtain updated target node features and target edge features, including: By aggregating X neighboring nodes and their corresponding connection edges through the first message passing layer, the initial node features of the current node are updated to obtain the first node features corresponding to the current node; based on the first node features of the current node and the current node features of the neighboring nodes, the initial edge features of the connection edges are updated to obtain the first edge features corresponding to the current connection edge. By aggregating Y neighboring nodes and their corresponding connecting edges through the second message passing layer, the first node feature of the current node is updated to obtain the second node feature corresponding to the current node; based on the second node feature of the current node and the current node features of the neighboring nodes, the first edge feature of the connecting edge is updated to obtain the second edge feature corresponding to the current connecting edge. When the third message passing layer exists, the third message passing layer aggregates Z neighboring nodes and their corresponding connecting edges, updates the second node feature of the current node, and obtains the target node feature corresponding to the current node; based on the target node feature of the current node and the current node features of the neighboring nodes, the second edge feature of the connecting edge is updated, and the target edge feature corresponding to the current connecting edge is obtained.

6. The training method according to claim 3, characterized in that, The number of message passing layers in the decoder of the graph neural network model is the same as the number of message passing layers in the encoder; wherein: The process involves decoding and predicting each node in a random order using a graph neural network model to be trained, outputting predicted sequence information, including: The current node to be predicted is determined based on the random mask matrix; The target node features of the node to be predicted are updated sequentially and cyclically according to multiple message passing layers in the decoder to obtain the corresponding node update features; Based on the residue type probability distribution of the node update feature transformation, the corresponding residue type is predicted and the corresponding predicted sequence information is generated.

7. The training method according to any one of claims 1-6, characterized in that, The loss value is calculated based on the following hybrid loss function, including: Among them, -log(probs( S i The negative log-likelihood loss (probs()) measures the log difference between the predicted sequence information and the sample sequence information. S i ) is a node i The probability value of the predicted residue class. pLDDT i Represents residues i Confidence level of prediction, mask i It can be 0 or 1.

8. The training method according to any one of claims 1-6, characterized in that, The acquisition of training data includes: Based on the first preset percentage of sequence consistency, each real protein is clustered to obtain multiple first protein clusters. The simulated proteins are clustered according to a dual clustering rule to obtain multiple second protein clusters. The dual clustering rule includes clustering according to a second preset percentage of sequence overlap and / or a third preset percentage of sequence consistency, and clustering according to a fourth preset percentage of structural overlap. Sample proteins were identified from each of the first and second protein clusters.

9. The training method according to claim 8, characterized in that, The real protein has a resolution of less than 3.5 Å and / or fewer than 1,000 residues.

10. The training method according to claim 8, characterized in that, The step of determining sample proteins from each of the first and second protein clusters includes: A preset number of real proteins are selected as sample proteins from each of the first protein clusters, and simulated proteins with a confidence level greater than a preset confidence threshold are selected as sample proteins from each of the second protein clusters.

11. A method for generating an amino acid sequence, characterized in that, include: Obtain the main chain structure information of the target protein; The main chain structure information is input into the amino acid sequence generation model trained by the training method of any one of claims 1 to 10, and the corresponding target amino acid sequence is output.

12. The generation method according to claim 11, characterized in that, The step of obtaining the main chain structure information of the target protein includes: encoding and obtaining the node features and edge features corresponding to the target protein based on the main chain structure information; wherein: The node features include some or all of the following features: The torsion angles φ and ψ of the residues are obtained as 16-dimensional periodic features through fourth-order Fourier expansion, namely sin(nφ), cos(nφ), and sin(nψ), cos(nψ), where n = 1~4; The secondary structures of residues include: α-helix, 3... 10 - Helices, π-helices, extended β-chains, isolated β-bridges, turns, bends, and coiled / irregular structures; Structural confidence of residues; The edge features include some or all of the following features: The N atom and C atom between residue i and residue j respectively α Atoms, C atoms, O atoms, and virtual C β The distance information between any two atoms of an atom; Confidence level of the predicted relative positions of residue i and residue j.

13. A training device for an amino acid sequence generation model, characterized in that, include: The sample acquisition module is used to acquire training data, which includes sample protein structure information and corresponding sample sequence information. The feature embedding module is used to encode the sample structure information to obtain the node features and edge features corresponding to the sample protein; wherein, the node features are used to represent residue information, and the edge features are used to represent the spatial geometric relationship between adjacent residues; The message passing module is used to alternately update the node features and edge features through multiple message passing layers of the graph neural network model to be trained, so as to obtain the updated target node features and target edge features; wherein the number of neighbor nodes aggregated by each message passing layer increases sequentially. The node decoding module is used to decode and predict each node in a random order according to the corresponding target node features and target edge features, and output the prediction sequence information. The backpropagation module is used to iterate the network parameters of the graph neural network model based on the loss value calculated from the predicted sequence information and the sample sequence information to obtain a trained amino acid sequence generation model.

14. An apparatus for generating an amino acid sequence, characterized in that, include: The information acquisition module is used to acquire the main chain structure information of the target protein; The sequence generation module is used to input the main chain structure information into the amino acid sequence generation model trained by the training method of any one of claims 1-10, and output the corresponding target amino acid sequence.

15. An electronic device, characterized in that, include: processor; as well as A memory having executable code stored thereon, which, when executed by the processor, causes the processor to perform the training method as described in any one of claims 1-10 or the generation method as described in any one of claims 11-12.

16. A computer-readable storage medium having executable code stored thereon, which, when executed by a processor of an electronic device, causes the processor to perform a training method as claimed in any one of claims 1-10 or a generation method as claimed in any one of claims 11-12.

17. A computer program product, characterized in that, The computer program product includes computer instructions that, when executed by a processor, implement the training method as described in any one of claims 1-10 or the generation method as described in any one of claims 11-12.

Images