A noise adaptive protein backbone generation method, device, equipment and medium

Abstract

The application discloses a noise adaptive protein skeleton generation method, device, equipment and medium. The method comprises the following steps: constructing an attention mask according to a preset length, sampling an initial internal angle sequence from a periodic noise distribution, calling a diffusion model to denoise the initial internal angle sequence to obtain an internal angle sequence, and reconstructing the internal angle sequence into a protein skeleton. Wherein, at each time step, the internal angle sequence of the current time step is extended with a periodic feature to obtain an extended angle sequence; the current time step is encoded to generate a time step conditional vector; the time step conditional vector, the attention mask and the extended angle sequence are input into the diffusion model, and the extended angle sequence is denoised. In the embodiment of the application, the time step is encoded into a time step conditional vector, and the time step conditional vector is input into the diffusion model as conditional information, so that the diffusion model dynamically adjusts the feature transformation for different diffusion stages according to the time step, and the generation quality of the protein skeleton is improved.

Description

Technical Field

[0001] This application relates to the interdisciplinary fields of computational biology and bioinformatics, and in particular to a noise-adaptive protein backbone generation method, apparatus, device, and medium. Background Technology

[0002] The protein backbone is the basic geometric framework of a protein's three-dimensional structure. A protein backbone that conforms to geometric rules can provide fundamental support for subsequent amino acid sequence configuration, candidate structure screening, and protein functional design. Because the protein backbone contains a large number of interrelated spatial degrees of freedom, traditional generation schemes using fragment splicing, energy function search, or rule-driven methods require repeated searches in a vast conformational space, and the results are easily limited by fragment libraries, initial configurations, and optimization strategies.

[0003] With the development of generative machine learning techniques, a generative scheme that progressively adds noise and then learns to denoise inversely has begun to be applied to protein backbone construction. Currently, when using diffusion models to generate protein backbones, the internal angles of the protein backbone are generally used as the generation target. The relative configuration between residues is described by the dihedral angles and bond angles formed by adjacent backbone atoms. After generation, the three-dimensional coordinates of the protein backbone are recovered based on the internal angles and bond lengths. However, in the process of generating protein backbones using diffusion models, existing schemes only add the time-step embedding and sequence embedding at the network input, and subsequent processing always uses fixed normalization and fixed residual paths. However, the input signal-to-noise ratio and the structural information to be recovered differ in high-noise, medium-noise, and low-noise stages. The processing method of using fixed normalization and fixed residual paths cannot dynamically adjust feature transformations for different diffusion stages, which will ultimately affect the quality of protein backbone generation.

[0004] Therefore, the existing technology still needs to be improved and enhanced. Summary of the Invention

[0005] The technical problem to be solved by this application is to provide a noise-adaptive protein backbone generation method, apparatus, device and medium to address the shortcomings of the prior art.

[0006] To address the aforementioned technical problems, the first aspect of this application provides a noise-adaptive protein backbone generation method, wherein the noise-adaptive protein backbone generation method specifically includes: Construct an attention mask of a preset length and sample an initial internal angle sequence from a periodic noise distribution; The initial internal corner sequence is denoised using a diffusion model to obtain the internal corner sequence, and the internal corner sequence is reconstructed into a protein backbone. The denoising process of the diffusion model at each time step includes: Periodic feature expansion is performed on each angle vector in the internal angle sequence of the current time step to obtain the expanded angle sequence; Encode the current time step to generate a time step condition vector; The time-step conditional vector, attention mask, and extended angle sequence are input into the diffusion model, and the diffusion model is used to denoise the extended angle sequence to generate the internal angle sequence of the next time step.

[0007] The noise-adaptive protein backbone generation method, wherein the step of periodically extending each angle vector in the internal angle sequence of the current time step to obtain the extended angle sequence specifically includes: For each angle vector in the internal angle sequence of the current time step, obtain the sine and cosine terms of each internal angle in the angle vector; The extension characteristics of each interior angle are determined by concatenating the sine and cosine terms of each interior angle, or by concatenating the sine and cosine terms of each interior angle. The extended features of each interior angle in the angle vector are concatenated to form an extended angle vector to obtain an extended angle sequence.

[0008] The noise-adaptive protein backbone generation method, wherein the step of denoising the extended corner sequence using the diffusion model to generate the internal corner sequence for the next time step specifically includes: The time-step condition vector is encoded as an adaptive scaling offset parameter, which includes a global scaling offset parameter, a local scaling offset parameter, and a feedforward scaling offset parameter. Based on the global scaling offset parameter and attention mask, the global attention branch is used to process the extended angle sequence to obtain global features; Based on the local scaling offset parameter, the global features are processed using a local blending branch to obtain blended features; The feedforward operation is performed on the blended features based on the feedforward scaling offset parameters to obtain the hidden features of the next layer, and the prediction noise is determined based on the hidden features output from the last layer. The internal angle sequence for the next time step is determined based on the predicted noise and the internal angle sequence of the current time step.

[0009] The aforementioned noise-adaptive protein backbone generation method, wherein encoding the time-step condition vector into an adaptive scaling offset parameter specifically includes: The time-step conditional vector is transformed into a non-linear time feature by using an activation function; The nonlinear time features are linearly mapped to obtain a one-dimensional parameter vector; The one-dimensional parameter vector is divided into global scaling offset parameters, local scaling offset parameters, and feedforward scaling offset parameters to obtain adaptive scaling offset parameters.

[0010] The aforementioned noise-adaptive protein backbone generation method, wherein the step of processing the extended angle sequence using a global attention branch to obtain global features based on a global scaling offset parameter and an attention mask specifically includes: The hidden features corresponding to the extended angle sequence are normalized to obtain normalized features; Based on the global scaling parameter and global offset parameter in the global scaling offset parameter, the normalized feature is subjected to an affine transformation to obtain an adaptive normalized feature; Based on the adaptive normalized features and the attention mask, attention features are determined through a multi-head attention mechanism. The attention features are gated and modulated based on the global gating parameters in the global scaling offset parameters, and the gated and modulated attention features are residually connected with the hidden features to obtain global features.

[0011] The noise-adaptive protein backbone generation method, wherein the step of processing global features using local blending branches based on local scaling offset parameters to obtain blended features specifically includes: The global features corresponding to the extended angle sequence are normalized to obtain normalized global features; Based on the local scaling parameter and local offset parameter in the local scaling offset parameter, the normalized global features are subjected to an affine transformation to obtain adaptive normalized global features. The adaptive normalized global features are processed using local modules to obtain local features, wherein the local modules include depthwise convolution, nonlinear activation, and pointwise convolution along the residue sequence direction; The local features are gated and modulated based on the local gate control parameters in the local scaling offset parameters, and the gated and modulated local features are residually concatenated with the global features to obtain hybrid features.

[0012] The noise-adaptive protein backbone generation method further includes, during the training process of the diffusion model: To train the protein backbone, a training internal angle sequence is constructed. The training internal angle sequence includes the angle vector of each residue in the protein backbone, and the angle vector includes three dihedral angles and three backbone bond angles. Forward diffusion with periodic constraints is applied to the training interior angle sequences to construct a set of noisy interior angle sequences and a set of real noise sequences, wherein the set of noisy interior angle sequences includes the noisy interior angle sequence at each time step and the set of real noise sequences includes the real noise at each time step. In the forward diffusion time step sequence, a time step is selected, and the noise internal angle sequence corresponding to the time step is denoised using the initial diffusion model to obtain the predicted noise. The initial diffusion model is trained based on the predicted noise and the actual noise at the time step to obtain the diffusion model.

[0013] A second aspect of this application provides a noise-adaptive protein backbone generation device, wherein the noise-adaptive protein backbone generation device specifically includes: The construction module is used to construct an attention mask according to a preset length and sample the initial internal angle sequence from the periodic noise distribution. The calling module is used to invoke the diffusion model to denoise the initial internal corner sequence to obtain the internal corner sequence, and to reconstruct the internal corner sequence into a protein backbone. The denoising process of the diffusion model at each time step includes: Periodic feature expansion is performed on each angle vector in the internal angle sequence of the current time step to obtain the expanded angle sequence; Encode the current time step to generate a time step condition vector; The time-step conditional vector, attention mask, and extended angle sequence are input into the diffusion model, and the diffusion model is used to denoise the extended angle sequence to generate the internal angle sequence of the next time step.

[0014] A third aspect of this application provides a computer-readable storage medium storing one or more programs that can be executed by one or more processors to implement the steps in the noise-adaptive protein backbone generation method as described above.

[0015] A fourth aspect of this application provides a terminal device, which includes: a processor and a memory; The memory stores a computer-readable program that can be executed by the processor; When the processor executes the computer-readable program, it implements the steps in any of the noise-adaptive protein backbone generation methods described above.

[0016] Beneficial effects: 1. This application encodes each time step as a time step condition vector, and then inputs the time step condition vector as condition information into the diffusion model, so that the diffusion model dynamically adjusts the feature transformation according to the time step for different diffusion stages, thereby improving the quality of protein backbone generation.

[0017] 2. This application combines global self-attention, one-dimensional depth convolution / pointwise convolution along the residue sequence and gated feedforward network into a local-global angle denoising block, and then uses this local-global angle denoising block as the denoising module of the diffusion model for denoising processing, which can improve the backbone angle modeling ability of the diffusion model, thereby improving the quality of protein backbone generation.

[0018] 3. This application adds sine and cosine features to the interior angles, thus providing geometrically adjacent but scalar-valued features. and Adding continuous periodic representation channels to the dihedral angles on both sides helps reduce the difficulty for the network to learn the periodic relationship of the angle.

[0019] 4. This application performs zero initialization on the adaptive parameter generation layer, so that the newly added residual branches have less disturbance to the backbone features during initialization. Then, the data-driven learning of the branch contributions at different diffusion stages helps to stabilize the optimization process of the deep denoising network.

[0020] 5. This application directly diffuses on the internal angle representation that is insensitive to overall translation and rotation, and reconstructs the 3D skeleton using a deterministic internal coordinate algorithm, without requiring the denoising network itself to have a complex 3D rigid body isomorphic structure. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 A flowchart of a noise-adaptive protein backbone generation method provided in an embodiment of this application.

[0023] Figure 2 This is a schematic flowchart of the noise-adaptive protein backbone generation method provided in the embodiments of this application.

[0024] Figure 3 This is a schematic diagram of the overall structure of the diffusion model.

[0025] Figure 4 This is a detailed architecture diagram of the diffusion model.

[0026] Figure 5 This is a detailed architecture diagram of the denoising module in the diffusion model.

[0027] Figure 6 This is a flowchart illustrating the overall training process of the diffusion model.

[0028] Figure 7 This is a flowchart illustrating the training process for a single sample in a diffusion model.

[0029] Figure 8 A flowchart of single-step noise prediction and parameter update for each training batch.

[0030] Figure 9 This is a schematic diagram of the noise-adaptive protein backbone generation device provided in the embodiments of this application.

[0031] Figure 10 A schematic block diagram of the terminal device provided in the embodiments of this application. Detailed Implementation

[0032] This application provides a method, apparatus, device, and medium for generating a noise-adaptive protein backbone. To make the objectives, technical solutions, and effects of this application clearer and more explicit, the following detailed description, with reference to the accompanying drawings and embodiments, further illustrates this application. It should be understood that the specific embodiments described herein are merely illustrative and are not intended to limit the scope of this application.

[0033] Those skilled in the art will understand that, unless specifically stated otherwise, the singular forms “a,” “an,” “the,” and “the” used herein may also include the plural forms. It should be further understood that the term “comprising” as used in this application means the presence of the stated features, integers, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof. It should be understood that when we say an element is “connected” or “coupled” to another element, it can be directly connected or coupled to the other element, or there may be intermediate elements. Furthermore, “connected” or “coupled” as used herein can include wireless connections or wireless coupling. The term “and / or” as used herein includes all or any units and all combinations of one or more associated listed items.

[0034] It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. It should also be understood that terms such as those defined in general dictionaries should be understood to have the same meaning as in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless specifically defined as herein.

[0035] It should be understood that the sequence number and size of each step in this embodiment do not imply the order of execution. The execution order of each process is determined by its function and internal logic, and should not constitute any limitation on the implementation process of this application embodiment.

[0036] Research has revealed that the protein backbone is the fundamental geometric framework of a protein's three-dimensional structure. A protein backbone that conforms to geometric rules can provide basic support for subsequent amino acid sequence configuration, candidate structure screening, and protein functional design. Because the protein backbone contains a large number of interrelated spatial degrees of freedom, traditional generation schemes using fragment splicing, energy function search, or rule-driven methods require repeated searches in a vast conformational space, and the results are easily limited by fragment libraries, initial configurations, and optimization strategies.

[0037] With the development of generative machine learning techniques, a generative scheme that progressively adds noise and then learns to denoise inversely has begun to be applied to protein backbone construction. This involves describing the modeling and generation process of protein structures as gradually removing noise signals from a set of randomly generated noisy data (but conforming to a specific probability distribution function) to ultimately obtain a noise-free protein structure. Early generation methods using diffusion models directly used the three-dimensional coordinates of backbone atoms as the processing object. While three-dimensional coordinates can intuitively reflect atomic positions, the same protein backbone will yield completely different coordinate values ​​after overall rotation or translation. To enable the diffusion model to recognize that the coordinates actually correspond to the same or similar structures, additional coordinate alignment, data augmentation, rigid body transformation processing, geometrically variable networks, or geometric constraints are needed. This undoubtedly increases the complexity of the network design and training process.

[0038] To address the issue of the three-dimensional coordinates of backbone atoms, another type of generation method utilizing diffusion models chooses to generate the protein backbone's internal angles. This type of method describes the relative configuration between residues using dihedral angles and bond angles formed by adjacent backbone atoms. After generation, the three-dimensional coordinates of the protein backbone are reconstructed based on the internal angles and bond lengths. Since the internal angles do not change with the overall translation or rotation of the protein, spatial invariance can be directly embedded into the data representation, reducing the burden on diffusion models in handling three-dimensional rigid body transformations.

[0039] However, existing methods that use the internal corners of the protein backbone as the generation target only add the time-step embedding and the sequence embedding at the network input, and subsequent processing always uses fixed normalization and fixed residual paths. The input signal-to-noise ratio and the structural information to be recovered differ at high-noise, medium-noise, and low-noise stages. Using fixed normalization and fixed residual paths cannot dynamically adjust feature transformations for different diffusion stages, ultimately affecting the quality of the generated protein backbone.

[0040] To address this, this application provides a noise-adaptive protein backbone generation method. The method includes constructing an attention mask of a preset length, sampling an initial internal angle sequence from a periodic noise distribution, calling a diffusion model to denoise the initial internal angle sequence to obtain an internal angle sequence, and reconstructing the internal angle sequence into a protein backbone. The denoising process of the diffusion model at each time step includes: periodically expanding each angle vector in the internal angle sequence of the current time step to obtain an expanded angle sequence; encoding the current time step to generate a time step condition vector; inputting the time step condition vector, the attention mask, and the expanded angle sequence into the diffusion model, and using the diffusion model to denoise the expanded angle sequence to generate the internal angle sequence of the next time step. This application encodes each time step as a time step condition vector, and then inputs this time step condition vector as condition information into the diffusion model, enabling the diffusion model to dynamically adjust feature transformations according to different diffusion stages based on the time step, thereby improving the quality of protein backbone generation.

[0041] It should be noted that the executing entity of each embodiment of this application can be a computing service system with data processing, network communication, and program execution functions, such as an electronic system capable of realizing the above functions, a noise-adaptive protein backbone generation system, etc. The following uses a noise-adaptive protein backbone generation system (hereinafter referred to as "the system") as an example to describe the following embodiments.

[0042] The application content will be further explained below with reference to the accompanying drawings and the description of the embodiments.

[0043] This embodiment provides a noise-adaptive protein backbone generation method, such as... Figure 1 and Figure 2 As shown, the noise-adaptive protein backbone generation method specifically includes: S10. Construct an attention mask according to a preset length and sample the initial internal angle sequence from the periodic noise distribution.

[0044] Specifically, the preset length is the target length of the protein backbone to be generated. This target length is the preset total number of residues in the protein backbone to be generated, which can be directly set by the user according to the generation requirements of the target protein backbone. In other words, the system will obtain the preset length in advance, such as the number of target residues input by the user, the number of target residues sent by an external device, or the number of target residues pre-configured in the system.

[0045] It is understandable that proteins may exist in multiple chains (e.g., antibodies contain heavy and light chains), each composed of a variable number of amino acid residues. All 20 standard amino acids are included. Heavy atoms, only in relation to The side chain residues connected by atoms differ, and these side chain residues also correspond to the protein's sequence information. In this application, the protein backbone is represented by an angle vector formed by the internal angles of the backbone atomic chains formed by adjacent residues; that is, the protein backbone can be represented by an internal angle sequence composed of all angle vectors. This angle vector includes three dihedral angles. and three skeleton key corners This allows the internal angle sequence to have more complete geometric information.

[0046] For example, for those with For a protein backbone with ___ residues, sampling the angle vectors of the backbone atoms formed by any two residues in the protein backbone, the initial internal angle sequence formed by all combinations of angle vectors can be represented as: , , in, Represents the initial interior angle sequence. Both represent angle vectors. Indicates the number of residues (i.e., the preset length). Represents angle vector The three dihedrals in Represents angle vector Three skeletal bond angles in the middle.

[0047] The embodiments of this application use backbone atomic chains formed by adjacent residues as units, which can eliminate invalid boundary angles. Each segment has a complete and computable angle vector, and there are no physically non-existent angles. There is no need to fill invalid placeholders, so that all rows of the internal angle sequence are real and valid geometric descriptions, reducing the noise caused by the model learning invalid placeholder markers.

[0048] Based on this, after obtaining the preset length, an initial internal corner sequence can be sampled based on the preset length of the protein backbone to be generated. Each angle vector in the initial internal corner sequence is sampled from a periodic noise distribution. For example, initial noise values ​​are randomly sampled from the angle vectors corresponding to the backbone atomic chains formed by any adjacent residues in the periodic noise distribution, and combined to form the initial internal corner sequence. The periodic noise distribution follows a standard Gaussian distribution, and the distribution parameters are preset during the diffusion model training phase.

[0049] Since the length of natural proteins is typically variable, padding can be performed based on the initial internal corner sequence to accommodate different length requirements. For example, zero-padding can be used to make the length of the padded initial internal corner sequence equal to the sequence length configured in the diffusion model. Simultaneously, an attention mask is generated based on this preset length to distinguish between valid residue positions and padded positions in the padded initial internal corner sequence. For example, assuming the preset length is... The diffusion model is configured with a sequence length of 1000. , Then the initial interior angle sequence is filled. Fill positions, and Set the mask value corresponding to each padding position to 0, and The mask value for each valid residue position is set to 1.

[0050] S20. Use the diffusion model to denoise the initial internal corner sequence to obtain the internal corner sequence, and reconstruct the internal corner sequence into a protein backbone.

[0051] Specifically, after obtaining the initial interior angle sequence, it can be done according to... The trained diffusion model (i.e., the denoising network) is called sequentially. This diffusion model is used to determine the predicted noise. Then, the internal angle sequence for the next time step is obtained according to the backdiffusion update formula, until the last time step. This process yields a denoised internal angle sequence, which is then used to reconstruct the protein backbone. After each update, a periodic wraparound is performed on the angle dimension to prevent the angle from deviating from the preset periodic range.

[0052] The reverse diffusion update formula can be expressed as: , , , in, Indicates time step The internal angle sequence, Indicates time step The internal angle sequence, Let represent random noise that follows a standard Gaussian distribution, and when No random noise is added. The parameters are determined based on the posterior variance. Represents a diffusion model. Represents the network parameters of the diffusion model. Indicates time step The variance of the prediction noise, and Both represent intermediate parameters.

[0053] Furthermore, such as Figure 3 and Figure 4 As shown, at each time step, the diffusion model denoises the input internal angle sequence of the current time step. This denoising process includes: H10. Perform periodic feature expansion on each angle vector in the internal angle sequence of the current time step to obtain the expanded angle sequence; H20. Encode the current time step to generate a time step condition vector; H30. Input the time step condition vector, attention mask, and extended angle sequence into the diffusion model, and use the diffusion model to denoise the extended angle sequence to generate the internal angle sequence of the next time step.

[0054] In step H10, since the dihedral angles in the internal angles of the protein have definite periodic boundaries, their numerical values ​​are close to... and Two dihedral angles are geometrically adjacent. If the interior angles are directly input into the diffusion model as ordinary continuous scalars, the diffusion model needs to learn the periodic relationship of the angle variables, which can easily lead to discontinuous representations at the periodic boundaries. Therefore, in this embodiment, after obtaining the interior angle sequence, periodic feature expansion is performed on each interior angle in each angle vector of the interior angle sequence, so that the interior angles contain periodic representations.

[0055] In one embodiment, the step of periodically extending each angle vector in the internal angle sequence of the current time step to obtain the extended angle sequence specifically includes: For each angle vector in the internal angle sequence of the current time step, obtain the sine and cosine terms of each internal angle in the angle vector; The extension characteristics of each interior angle are determined by concatenating the sine and cosine terms of each interior angle, or by concatenating the sine and cosine terms of each interior angle. The extended features of each interior angle in the angle vector are concatenated to form an extended angle vector to obtain an extended angle sequence.

[0056] Specifically, each angle vector includes 3 dihedral angles and 3 skeleton bond angles, for any one of the interior angles... Calculate the sine term corresponding to the interior angle directly. and cosine term , sine term and cosine term As a periodic representation, it is added to the interior angle. This allows us to obtain extended features of interior angles. For example, for interior angles... ,use As this interior angle Extended features; interior angles can also be omitted. Directly adopt As this interior angle The extended features. Both methods allow the diffusion model to directly obtain a periodically invariant angular representation, enabling the diffusion model to explicitly identify crossings. The angular adjacency relationship of the boundary does not require the diffusion model to learn the geometric relationship of the periodic boundary itself.

[0057] Next, after completing the periodic expansion of each interior angle, the expansion features of all interior angles in the angle vector are concatenated sequentially to obtain the expansion angle vector corresponding to that angle vector. Then, combining the expansion angle vectors corresponding to all angle vectors yields the expansion angle sequence. Furthermore, as... Figure 3 As shown, in practical applications, when determining the extended angle sequence, a sequence position embedding can be added, that is, the periodic feature extension and the initial internal angle sequence are positionally encoded to form the sequence position embedding, and then the periodic feature extension and the sequence position embedding are feature fused to obtain the extended angle sequence.

[0058] Furthermore, in step H20, the existing scheme only adds the time-step embedding and the sequence embedding at the network input, and subsequent layers use fixed normalization and fixed residual paths. However, the input signal-to-noise ratio and the information of the structure to be recovered are different in the high-noise, medium-noise, and low-noise stages, which makes it difficult for the static calculation path (i.e., fixed normalization and fixed residual paths) to dynamically adjust the feature transformation for different diffusion stages. Therefore, as... Figure 3 As shown, in this embodiment of the application, the time step is encoded in each time step to obtain the time step condition vector.

[0059] Specifically, the current time step can be... Input the time step encoder, and use the time step encoder to measure the current time step. Encode to obtain the time-step condition vector The time-step encoder can employ Gaussian Fourier projection, fixed sine / cosine encoding, or learnable embedding, among other methods. In this embodiment, the time-step encoder uses Gaussian Fourier projection to map different noise stages to a continuous hidden condition space.

[0060] Wherein, the time-step condition vector vector dimension Not related to the number of residues And the angle vector dimension binding is not required to have a fixed value in all implementations. However, the time-step condition vector vector dimension It needs to be consistent with the input dimension of the subsequent modulation mapping. For example, setting the hidden dimension of the diffusion model to be... The diffusion model Modulation mapping of each noise block Achieve from arrive The mapping, where the vector dimension It can be equal to hidden dimension ,like = =384, if the vector dimension Not equal to hidden dimension , can The linear layers in the process complete the dimensional transformation.

[0061] Furthermore, when the time step encoder uses Gaussian Fourier projection, it can measure the current time step. Perform normalization to obtain the normalized time step. and normalize the time step The mapping is to pairwise features, which are: , in, Representing paired features, Indicates inclusion A frequency vector of frequency components.

[0062] If we directly apply this pair of features As a time-step condition vector Then the time-step condition vector vector dimension 2 If paired features Use linear projection or multilayer perceptron to determine the time-step condition vector. Then the time-step condition vector vector dimension The dimensionality can be selected based on the dimensionality requirements of the diffusion model, and is not limited to an even number. For example, paired features... As a time-step condition vector The number of frequency components =192, then the time-step condition vector vector dimension The value is 384.

[0063] Furthermore, in step H30, the diffusion model may include A series of sequentially connected denoising modules, where the input of the subsequent denoising module includes the hidden features output by the preceding denoising module. The inputs to the foremost denoising module include the expansion angle sequence corresponding to the previous time step, the time step conditional vector, and the attention mask. The processing procedure of each of the sequentially connected noise reduction modules is the same.

[0064] Since the continuous changes in the angles of adjacent residues form local geometric patterns, and the mutual organization between distant locations affects the overall skeleton structure, when using a diffusion model to predict noise, if the model mainly relies on a single sequence relationship modeling approach and does not explicitly configure a local angle mixing path, it will increase the difficulty for the network to simultaneously learn local continuous patterns and long-range dependencies. Therefore, the diffusion model in this embodiment combines a global attention branch and a local mixing branch when predicting noise. The global attention branch models the long-distance relationships between effective residues, while the local mixing branch extracts the local angle patterns of adjacent residues from each hidden channel. These patterns are then jointly adjusted through the same time-step condition (i.e., the time-step condition vector), enabling the diffusion model to achieve a local-global coordinated skeleton angle modeling capability.

[0065] For example, such as Figure 5 As shown, the step of using the diffusion model to denoise the extended angle sequence to generate the internal angle sequence for the next time step specifically includes: H31. Encode the time-step condition vector into an adaptive scaling offset parameter; H32. Based on the global scaling offset parameter and attention mask, the global attention branch is used to process the extended angle sequence to obtain global features; H33. Based on the local scaling offset parameter, the global features are processed using the local blending branch to obtain the blended features; H34. Perform a feedforward operation on the hybrid features based on the feedforward scaling offset parameters to obtain the hidden features of the next layer, and determine the prediction noise based on the hidden features output by the last layer. H35. Determine the internal angle sequence for the next time step based on the predicted noise and the internal angle sequence of the current time step.

[0066] In step H31, the adaptive scaling offset parameters include global scaling offset parameters, local scaling offset parameters, and feedforward scaling offset parameters. The global scaling offset parameters include global scaling parameters, global offset parameters, and global gating parameters. The local scaling offset parameters include local scaling parameters, local offset parameters, and local gating parameters. The feedforward scaling offset parameters include feedforward scaling parameters, feedforward offset parameters, and feedforward gating parameters. The global scaling parameters, global offset parameters, global gating parameters, local scaling parameters, local offset parameters, local gating parameters, feedforward scaling parameters, feedforward offset parameters, and feedforward gating parameters all have the same parameter dimension, which is equal to the hidden dimension of the diffusion model. .

[0067] It is understandable that, since the global scaling offset parameters, local scaling offset parameters, and feedforward scaling offset parameters are used in different network layers, they can be different. In other words, the global attention branch, local fusion branch, and feedforward branch do not share the same set of parameters. Specifically, the global scaling offset parameters, local scaling offset parameters, and feedforward scaling offset parameters are different parameter sets generated using the time-step conditional vector. Each set of parameters is only responsible for feature modulation of the corresponding network layer, adapting to the differentiated time-step information requirements of different modules.

[0068] For example, encoding the time-step condition vector into an adaptive scaling offset parameter specifically includes: The time-step conditional vector is transformed into a non-linear time feature by using an activation function; The nonlinear time features are linearly mapped to obtain a one-dimensional parameter vector; The one-dimensional parameter vector is divided into global scaling offset parameters, local scaling offset parameters, and feedforward scaling offset parameters to obtain adaptive scaling offset parameters.

[0069] Specifically, firstly, the time-step conditional vector is input into an activation function, such as the SiLU activation function, to perform a nonlinear transformation and obtain nonlinear time features. Then, a linear fully connected layer maps the nonlinear time features into a one-dimensional parameter vector. This one-dimensional parameter vector has a dimension nine times that of the hidden dimension of the diffusion model, which precisely allows for the decomposition of the nine sub-vectors, including the global scaling offset parameters, local scaling offset parameters, and feedforward scaling offset parameters, into nine dimensions that are part of the hidden dimension. Finally, decomposing the aforementioned one-dimensional parameter vector yields the global scaling offset parameters, local scaling offset parameters, and feedforward scaling offset parameters. The nine sets of modulation parameters included in the global scaling offset parameters, local scaling offset parameters, and feedforward scaling offset parameters can be expressed as: , in, This represents the global scaling parameter. This represents the global offset parameter. Indicates global gating parameters. This represents the local scaling parameter. Indicates the local offset parameter. Indicates the local control parameters, This represents the feedforward scaling parameter. This represents the feedforward offset parameter. Indicates the feedforward gating parameters. In the diffusion model, the first... The time-step modulation mapping of each denoising module is used to modulate the time-step conditional vector into adaptive global scaling offset parameters, local scaling offset parameters, and feedforward scaling offset parameters. This represents the time-step condition vector.

[0070] It should be noted that although the global scaling offset parameter, the local scaling offset parameter, and the feedforward scaling offset parameter are the same in dimension, they are all... The modulation parameters are 3D vectors, but in practical applications, the values ​​and mapping weights in the global scaling offset parameters, local scaling offset parameters, and feedforward scaling offset parameters are independent of each other and do not share the same set of parameters. That is, all nine sets of modulation parameters are derived from the same time-step conditional vector. Each denoising module generates a different output row block, but these blocks correspond to different output rows in the linear layers of the modulation map. Furthermore, the time-step modulation maps in each denoising module can be independent of each other.

[0071] For example, for the first The noise reduction module first uses the SiLU activation function to apply the time-step conditional vector. Processing yields nonlinear time characteristics Then, the nonlinear time characteristics Mapped to a one-dimensional parameter vector Finally Divided into 9 equal parts dimensional vector Among them, nonlinear time characteristics and one-dimensional parameter vector It can be represented as: , , in, This represents the SiLU activation function. The mapping matrix can be decomposed into nine shapes. Independent submatrices, Time step condition vector vector dimension, This is the bias vector.

[0072] Furthermore, in step H32, the process of using a global attention branch to process the extended angle sequence based on the global scaling offset parameter and the attention mask to obtain global features specifically includes: The hidden features corresponding to the extended angle sequence are normalized to obtain normalized features; Based on the global scaling parameter and global offset parameter in the global scaling offset parameter, the normalized feature is subjected to an affine transformation to obtain an adaptive normalized feature; Based on the adaptive normalized features and the attention mask, attention features are determined through a multi-head attention mechanism. The attention features are gated and modulated based on the global gating parameters in the global scaling offset parameters, and the gated and modulated attention features are residually connected with the hidden features to obtain global features.

[0073] Specifically, for the hidden features input to the current global attention branch, if the denoising module is the first denoising module, then the hidden feature is the input extended angle sequence; if the denoising module is not the first denoising module, then the hidden feature is the hidden feature output by the previous denoising module. First, layer normalization (LN) is performed on the hidden feature to obtain normalized features. Then, an affine transformation is performed on the normalized features using global scaling parameters and global offset parameters to obtain adaptive normalized features. The process of obtaining these adaptive normalized features can be represented as follows: , in, This represents adaptive normalized features. This represents the adaptive layer-normalized modulation function based on layer normalization. Representation layer normalization, This represents the global scaling parameter. This represents the global offset parameter. This represents element-wise multiplication. Indicates the first Hidden features of each noise reduction module.

[0074] It should be noted that in practical applications, an adaptive layer-normalized modulation function based on layer normalization can be pre-constructed. Then, when executing the global attention branch, this adaptive layer-normalized modulation function can be directly called to determine the adaptive normalized features. These adaptive normalized features can be expressed as: , in, Indicates hidden features, This represents the scaling parameter. This represents the offset parameter.

[0075] Furthermore, after obtaining the adaptively normalized features, these features are mapped to the query vectors required by the multi-head attention mechanism. Key vector Sum value vector Then, attention scores between each residue position are calculated using the attention mask. These scores are then weighted and summed to obtain the multi-head attention output. The multi-head attention outputs are concatenated and fused to obtain the attention features. Next, channel-wise gating weighting is applied to the attention features using global gating parameters to obtain the weighted attention features. This result is then added to the hidden features of the original input to complete the residual connection, finally outputting the global features of the current denoising module. These global features are represented as follows: , in, Represents global features. Indicates global gating parameters. This indicates the bulls' self-attention. This represents an attention mask.

[0076] In step H33, the step of processing the global features using a local blending branch based on the local scaling offset parameter to obtain the blended features specifically includes: The global features corresponding to the extended angle sequence are normalized to obtain normalized global features; Based on the local scaling parameter and local offset parameter in the local scaling offset parameter, the normalized global features are subjected to an affine transformation to obtain adaptive normalized global features. The adaptive normalized global features are processed using local modules to obtain local features, wherein the local modules include depthwise convolution, nonlinear activation, and pointwise convolution along the residue sequence direction; The local features are gated and modulated based on the local gate control parameters in the local scaling offset parameters, and the gated and modulated local features are residually concatenated with the global features to obtain hybrid features.

[0077] Specifically, firstly, layer normalization is performed on the input global features to obtain normalized global features. Then, an affine transformation is applied to the normalized global features using local scaling and local offset parameters to obtain adaptive normalized global features. The adaptive normalized global features can be directly determined by calling the adaptive layer normalization modulation function; the only difference is that the scaling parameter used is the local scaling parameter, and the offset parameter is the local offset parameter. That is, the adaptive normalized global features can be represented as... .

[0078] Next, the adaptively normalized global features are input into the local module. First, a one-dimensional depthwise convolution (e.g., with a kernel size of 5) along the residue sequence direction is used to extract the local angular dependencies between adjacent residues. Then, a nonlinear transformation is introduced through nonlinear activation (e.g., the SiLU activation function). Finally, the channel-dimensional features are integrated through pointwise convolution to obtain the processed local features. Then, local gated parameters are used to perform channel-wise gated weighting on the local features. The gated and modulated local features are added to the original input global features to complete the residual connection, ultimately obtaining the hybrid features of the current denoising module. This hybrid feature can be represented as: , in, Indicates mixed features, Indicates a local module. Represents global features. This represents the local scaling parameter. Indicates the local offset parameter. This indicates the local control parameters.

[0079] Furthermore, in step H34, after obtaining the mixed features, a feed-forward network (FFN) can be used to perform a feed-forward operation on the mixed features to obtain the hidden features of the next layer. Before performing the feed-forward operation, an adaptive layer normalization modulation function is also called to adaptively normalize the mixed features. The feed-forward network can adopt a Swish-Gated Linear Unit (SwiGLU) structure and use a Sigmoid Linear Unit (SiLU) as the gate activation. The hidden features of the next layer and the feed-forward operation performed by the feed-forward network can be represented as follows: , , in, Indicates the hidden features of the next layer. Indicates a feedforward network. Indicates mixed features, This represents the feedforward scaling parameter. This represents the feedforward offset parameter. Indicates the feedforward gating parameters. This indicates the feedforward operation performed by the feedforward network. This represents the hidden input features of the feedforward network. The learnable linear mapping parameters represent the value branches. The learnable linear mapping parameters representing the gated branches, This represents the learnable parameters of the output projection.

[0080] Next, after obtaining the hidden features of the next layer, it is determined whether the hidden features of the next layer are the same as the hidden features output by the last layer, that is, to detect all... After each noise reduction module has completed the above processing in sequence, if all Each denoising module sequentially completes the above processing (i.e., the hidden features output from the last layer). Then, final-layer normalization is performed on the hidden features output from the last layer, and the decoder outputs predicted noise with the same shape as the input internal angle sequence, providing a basis for calculating the internal angle sequence at the next time step. If a denoising module has not completed the above processing, the hidden features of that next layer are input into the next denoising module, so that the next denoising module performs the above processing based on the hidden features of that next layer, until all... Each denoising module performs the above processing sequentially. The decoding head includes local convolution along the sequence direction, nonlinear activation, pointwise convolution, and output convolution; linear layers or multilayer perceptrons can also be used.

[0081] It should be noted that, in order to improve the stability in the initial stage of training, the final linear layer that generates nine sets of modulation parameters can be initialized to zero during training, so that the gating parameters of each residual branch are initially zero or close to zero. Then, the branch strength is learned according to different diffusion time steps during training.

[0082] This application's embodiments combine global self-attention, one-dimensional depthwise convolution / pointwise convolution along the residue sequence, and a gated feedforward network into a local-global angle denoising block. This local-global angle denoising block is then used as the denoising module of the diffusion model for denoising processing, which improves the diffusion model's ability to model the backbone angles, thereby improving the quality of protein backbone generation. Simultaneously, each denoising module in the diffusion model independently generates global scaling offset parameters, local scaling offset parameters, and feedforward scaling offset parameters based on the current time step. This allows global attention, local mixing, and feedforward branches to adjust feature transformation methods according to the input signal-to-noise ratio, rather than using time step information only once at the network entry point. This effectively improves the modeling accuracy of noise distribution differences at different diffusion stages, resulting in a protein structure that more accurately fits the given constraints. It also reduces the difficulty for the network to simultaneously learn local patterns and long-range dependencies, accelerating model convergence and improving generation performance.

[0083] Furthermore, in step S20 above, after obtaining the internal corner sequence, the internal corner sequence can be reconstructed into a protein backbone. Specifically, the data center offset of the finally generated internal corner sequence is restored and wrapped back to the legal angle range. Based on the preset backbone bond length and the initial three atomic coordinates, the Natural Extension Reference Frame (NeRF) or an equivalent internal coordinate reconstruction algorithm is used to iteratively calculate the internal coordinates in residue order. , and The three-dimensional coordinates of atoms are used to output a protein backbone structure file. The generated protein backbone can be further input into a reverse folding model to obtain candidate amino acid sequences, or into a structure prediction model to perform self-consistency verification, etc.

[0084] In one embodiment, during the training process of the diffusion model, such as Figure 6 and Figure 7 As shown, the method further includes: To train the protein backbone, a training internal angle sequence is constructed. The training internal angle sequence includes the angle vector of each residue in the protein backbone, and the angle vector includes three dihedral angles and three backbone bond angles. Forward diffusion with periodic constraints is applied to the training interior angle sequences to construct a set of noisy interior angle sequences and a set of real noise sequences, wherein the set of noisy interior angle sequences includes the noisy interior angle sequence at each time step and the set of real noise sequences includes the real noise at each time step. In the forward diffusion time step sequence, a time step is selected, and the noise internal angle sequence corresponding to the time step is denoised using the initial diffusion model to obtain the predicted noise. The initial diffusion model is trained based on the predicted noise and the actual noise at the time step to obtain the diffusion model.

[0085] Specifically, one or more protein backbone structure files are obtained, and the main chain nitrogen atoms are extracted from the protein backbone in the protein backbone structure files. , carbon atom and carbonyl carbon atoms The spatial coordinates are used to calculate three dihedral angles and three backbone bond angles for the backbone atomic chains formed by adjacent residues to obtain training internal angle sequences. An attention mask is generated for each training internal angle sequence to distinguish between effective residue positions and filling positions. The training internal angle sequences have the same form as the internal angle sequences described above, so they will not be repeated here.

[0086] Next, set the number of time steps to... The noise variance sequence is ,make , , and All represent intermediate parameters. From the time step sequence Extraction time step And extract noise from the preset distribution. Construct a noisy interior angle sequence, which is represented as: , in, Represents the internal angle sequence of noise. This indicates the training of the internal angle sequence.

[0087] Perform a periodic wraparound operation on the feature dimension that is an angle, so that both the noisy internal angle sequence and the noise are within a preset period interval. This periodic wraparound operation is expressed as: , in, This indicates a periodic wrap-around operation. This indicates a feature dimension belonging to the angle. This indicates a modulo operation.

[0088] Then, the noise internal angle sequence corresponding to the time step is denoised using the initial diffusion model to obtain the predicted noise. The denoising process of the initial diffusion model is the same as that of the trained diffusion model, so it will not be repeated here. For details, please refer to the denoising process of the trained diffusion model.

[0089] After denoising the noise internal angle sequence corresponding to the time step using the initial diffusion model to obtain the predicted noise, a loss function is constructed based on the predicted noise and the actual noise of the time step, and the initial diffusion model is trained using this loss function. The loss function can be: , in, Represents the loss function. This represents the model parameters of the initial diffusion model. Indicates the first The true noise of each internal angle Indicates the first Predicted noise for each interior angle, Indicates the first The loss function used for each interior angle, such as L1 loss, L2 loss, or a combination thereof, etc. It expresses expectation.

[0090] Furthermore, in practical applications, the denoised interior angles in the predicted denoised interior angle sequence can be converted into three-dimensional coordinates for calculation. The distance loss between atoms is calculated, and then this distance loss is combined with the aforementioned loss function based on prediction noise to jointly optimize the initial diffusion model.

[0091] For example, such as Figure 8 As shown, assume each training batch contains For the nth sample in the training batch Each sample is independently derived from the time-step sequence. Extract a time step And independently using real noise Based on this Forward construction of noisy interior angle sequence Therefore, each sample undergoes a noisy prediction at a random time step only once in a training iteration, and the time steps corresponding to different samples in the same training batch are different. They can be different.

[0092] The training batch is input into the initial diffusion model all at once to obtain the predicted noise in parallel. The batch loss of this training batch can be expressed as: , in, This represents the batch loss of the training batch. Indicates the first The first sample An effective mask for an angle vector. Indicates the first The first sample The th angle vector in the th angle vector The true noise of each internal angle Indicates the first The first sample The th angle vector in the th angle vector Predicted noise for each interior angle.

[0093] After averaging the batch loss, perform a backpropagation and update the optimizer. Once. As the training batches iterate, randomly sampled... It covers all diffusion steps in a statistical sense; the training phase does not need to be executed sequentially within each batch. A reverse diffusion step, only the generation stage follows this process. The denoising network is called repeatedly in sequence. For periodic smoothing L1 loss, it can be calculated first. Then calculate the smoothed L1 value.

[0094] In summary, this embodiment provides a noise-adaptive protein backbone generation method. The method includes constructing an attention mask based on a preset length, sampling an initial internal angle sequence from a periodic noise distribution, calling a diffusion model to denoise the initial internal angle sequence to obtain an internal angle sequence, and reconstructing the internal angle sequence into a protein backbone. The denoising process of the diffusion model at each time step includes: periodically expanding each angle vector in the internal angle sequence of the current time step to obtain an expanded angle sequence; encoding the current time step to generate a time step condition vector; inputting the time step condition vector, the attention mask, and the expanded angle sequence into the diffusion model, and using the diffusion model to denoise the expanded angle sequence to generate the internal angle sequence of the next time step. This embodiment encodes each time step as a time step condition vector, and then inputs this time step condition vector as condition information into the diffusion model, enabling the diffusion model to dynamically adjust feature transformations according to different diffusion stages based on the time step, thereby improving the quality of protein backbone generation.

[0095] Based on the above-described noise-adaptive protein backbone generation method, this embodiment provides a noise-adaptive protein backbone generation device, such as... Figure 9 As shown, the noise-adaptive protein backbone generation device specifically includes: The construction module 100 is used to construct an attention mask according to a preset length and sample an initial internal angle sequence from a periodic noise distribution. Module 200 is used to invoke a diffusion model to denoise the initial internal corner sequence to obtain an internal corner sequence, and to reconstruct the internal corner sequence into a protein backbone. The denoising process of the diffusion model at each time step includes: Periodic feature expansion is performed on each angle vector in the internal angle sequence of the current time step to obtain the expanded angle sequence; Encode the current time step to generate a time step condition vector; The time-step conditional vector, attention mask, and extended angle sequence are input into the diffusion model, and the diffusion model is used to denoise the extended angle sequence to generate the internal angle sequence of the next time step.

[0096] Based on the above-described noise-adaptive protein backbone generation method, this embodiment provides a computer-readable storage medium storing one or more programs that can be executed by one or more processors to implement the steps in the noise-adaptive protein backbone generation method as described in the above embodiment.

[0097] Based on the above-described noise-adaptive protein backbone generation method, this application also provides a terminal device, such as... Figure 10As shown, it includes at least one processor 20; a display screen 21; and a memory 22, and may also include a communications interface 23 and a bus 24. The processor 20, display screen 21, memory 22, and communications interface 23 can communicate with each other via the bus 24. The display screen 21 is configured to display a preset user guide interface in the initial setup mode. The communications interface 23 can transmit information. The processor 20 can invoke logical instructions in the memory 22 to execute the methods described in the above embodiments.

[0098] Furthermore, the logical instructions in the aforementioned memory 22 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium.

[0099] The memory 22, as a computer-readable storage medium, can be configured to store software programs, computer-executable programs, such as program instructions or modules corresponding to the methods in the embodiments of this disclosure. The processor 20 executes functional applications and data processing by running the software programs, instructions, or modules stored in the memory 22, thereby implementing the methods in the above embodiments.

[0100] The memory 22 may include a program storage area and a data storage area. The program storage area may store the operating system and application programs required for at least one function; the data storage area may store data created based on the use of the terminal device. Furthermore, the memory 22 may include high-speed random access memory (RAM) and non-volatile memory. Examples include various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks, as well as transient storage media.

[0101] Furthermore, the specific process of loading and executing multiple instruction processors in the aforementioned storage medium and terminal device has been described in detail in the above method, and will not be repeated here.

[0102] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application 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. Such 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 this application.

Claims

1. A noise-adaptive protein backbone generation method, characterized in that, The noise-adaptive protein backbone generation method specifically includes: Construct an attention mask of a preset length and sample an initial internal angle sequence from a periodic noise distribution; The initial internal corner sequence is denoised using a diffusion model to obtain the internal corner sequence, and the internal corner sequence is reconstructed into a protein backbone. The denoising process of the diffusion model at each time step includes: Periodic feature expansion is performed on each angle vector in the internal angle sequence of the current time step to obtain the expanded angle sequence; Encode the current time step to generate a time step condition vector; The time step conditional vector, attention mask, and extended angle sequence are input into the diffusion model, and the diffusion model is used to denoise the extended angle sequence to generate the internal angle sequence of the next time step.

2. The noise-adaptive protein backbone generation method according to claim 1, characterized in that, The step of performing periodic feature expansion on each angle vector in the internal angle sequence of the current time step to obtain the expanded angle sequence specifically includes: For each angle vector in the internal angle sequence of the current time step, obtain the sine and cosine terms of each internal angle in the angle vector; The extension characteristics of each interior angle are determined by concatenating the sine and cosine terms of each interior angle, or by concatenating the sine and cosine terms of each interior angle. The extended features of each interior angle in the angle vector are concatenated to form an extended angle vector to obtain an extended angle sequence.

3. The noise-adaptive protein backbone generation method according to claim 1, characterized in that, The step of using the diffusion model to denoise the extended angle sequence to generate the internal angle sequence for the next time step specifically includes: The time-step condition vector is encoded as an adaptive scaling offset parameter, which includes a global scaling offset parameter, a local scaling offset parameter, and a feedforward scaling offset parameter. Based on the global scaling offset parameter and attention mask, the global attention branch is used to process the extended angle sequence to obtain global features; Based on the local scaling offset parameter, the global features are processed using a local blending branch to obtain blended features; The feedforward operation is performed on the blended features based on the feedforward scaling offset parameters to obtain the hidden features of the next layer, and the prediction noise is determined based on the hidden features output from the last layer. The internal angle sequence for the next time step is determined based on the predicted noise and the internal angle sequence of the current time step.

4. The noise-adaptive protein backbone generation method according to claim 3, characterized in that, Encoding the time-step condition vector into adaptive scaling offset parameters specifically includes: The time-step conditional vector is transformed into a non-linear time feature by using an activation function; The nonlinear time features are linearly mapped to obtain a one-dimensional parameter vector; The one-dimensional parameter vector is divided into global scaling offset parameters, local scaling offset parameters, and feedforward scaling offset parameters to obtain adaptive scaling offset parameters.

5. The noise-adaptive protein backbone generation method according to claim 3, characterized in that, The process of processing the extended angle sequence using a global attention branch to obtain global features based on global scaling offset parameters and attention masks specifically includes: The hidden features corresponding to the extended angle sequence are normalized to obtain normalized features; Based on the global scaling parameter and global offset parameter in the global scaling offset parameter, the normalized feature is subjected to an affine transformation to obtain an adaptive normalized feature; Based on the adaptive normalized features and the attention mask, attention features are determined through a multi-head attention mechanism. The attention features are gated and modulated based on the global gating parameters in the global scaling offset parameters, and the gated and modulated attention features are residually connected with the hidden features to obtain global features.

6. The noise-adaptive protein backbone generation method according to claim 3, characterized in that, The process of processing global features using local blending branches based on local scaling offset parameters to obtain blended features specifically includes: The global features corresponding to the extended angle sequence are normalized to obtain normalized global features; Based on the local scaling parameter and local offset parameter in the local scaling offset parameter, the normalized global features are subjected to an affine transformation to obtain adaptive normalized global features. The adaptive normalized global features are processed using local modules to obtain local features, wherein the local modules include depthwise convolution, nonlinear activation, and pointwise convolution along the residue sequence direction; The local features are gated and modulated based on the local gate control parameters in the local scaling offset parameters, and the gated and modulated local features are residually concatenated with the global features to obtain hybrid features.

7. The noise-adaptive protein backbone generation method according to claim 1, characterized in that, During the training process of the diffusion model, the method further includes: To train the protein backbone, a training internal angle sequence is constructed. The training internal angle sequence includes the angle vector of each residue in the protein backbone, and the angle vector includes three dihedral angles and three backbone bond angles. Forward diffusion with periodic constraints is applied to the training interior angle sequences to construct a set of noisy interior angle sequences and a set of real noise sequences, wherein the set of noisy interior angle sequences includes the noisy interior angle sequence at each time step and the set of real noise sequences includes the real noise at each time step. In the forward diffusion time step sequence, a time step is selected, and the noise internal angle sequence corresponding to the time step is denoised using the initial diffusion model to obtain the predicted noise. The initial diffusion model is trained based on the predicted noise and the actual noise at the time step to obtain the diffusion model.

8. A noise-adaptive protein backbone generation device, characterized in that, The noise-adaptive protein backbone generation device specifically includes: The construction module is used to construct an attention mask according to a preset length and sample the initial internal angle sequence from the periodic noise distribution. The calling module is used to invoke the diffusion model to denoise the initial internal corner sequence to obtain the internal corner sequence, and to reconstruct the internal corner sequence into a protein backbone. The denoising process of the diffusion model at each time step includes: Periodic feature expansion is performed on each angle vector in the internal angle sequence of the current time step to obtain the expanded angle sequence; Encode the current time step to generate a time step condition vector; The time step conditional vector, attention mask, and extended angle sequence are input into the diffusion model, and the diffusion model is used to denoise the extended angle sequence to generate the internal angle sequence of the next time step.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores one or more programs, which can be executed by one or more processors to implement the steps in the noise-adaptive protein backbone generation method as described in any one of claims 1-7.

10. A terminal device, characterized in that, include: Processor and memory; The memory stores a computer-readable program that can be executed by the processor; When the processor executes the computer-readable program, it implements the steps in the noise-adaptive protein backbone generation method as described in any one of claims 1-7.

Images