Generative ai protein backbone generation method fusing multi-modal data

Abstract

The application discloses a fusion multi-modal data generation AI protein skeleton generation method, comprising the following steps: step one, training data set construction and standardization coding; step two, residue level local feature extraction and global context coding; step three, structure geometry feature extraction and secondary structure tendency coding; step four, functional semantic feature embedding and cross-modal attention fusion; step five, conditional probability path construction and network training based on flow matching generator; step six, conditional input acquisition in the reasoning stage; step seven, iterative denoising and skeleton coordinate trajectory tracking; step eight, main chain atom complete coordinate reconstruction and skeleton conformation optimization; step nine, multi-dimensional quality evaluation and conditional feedback screening of the generated skeleton. The technical scheme adopted by the application can effectively improve the diversity and designability of the generated skeleton, significantly reduce the calculation cost, and has clear industrial application value.

Description

Technical Field

[0001] This invention relates to the field of drug design technology, and more specifically to a generative AI method for generating protein backbones by integrating multimodal data. Background Technology

[0002] Protein backbone design is a core component of protein engineering and drug discovery, aiming to generate amino acid sequences that can fold into specific three-dimensional conformations and possess desired functions. Traditional protein design methods mainly rely on physical simulations and energy function optimization, such as molecular dynamics simulations and Monte Carlo searches. These methods face significant challenges in terms of computational accuracy and efficiency, making it difficult to effectively explore the vast sequence-conformation space.

[0003] In recent years, deep learning-based generative models, such as variational autoencoders and generative adversarial networks, have been applied to protein structure generation. However, such models usually only use a single type of data as input, such as only using sequence information or only using structural coordinates, ignoring multi-dimensional information such as protein functional annotation, residue conservation, and secondary structure dynamics.

[0004] This single-modal input method makes it difficult for generative models to accurately capture the complex mapping relationship between sequences and structures. The generated skeleton structures have significant deficiencies in terms of the rationality of local conformation and the designability of global folding. In addition, existing generative frameworks lack an active intervention mechanism for the generation process and cannot dynamically adjust generation preferences according to the target functional requirements, resulting in a large gap between the generated results and the actual experimental requirements. Summary of the Invention

[0005] To address this issue, the present invention provides a generative AI protein backbone generation method that integrates multimodal data, thereby solving the problem in the prior art that there is a lack of active intervention mechanism for the generation process, and that the generation preferences cannot be dynamically adjusted according to the target functional requirements, resulting in a large gap between the generation results and the actual experimental requirements.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] A generative AI method for protein backbone generation that integrates multimodal data includes the following steps:

[0008] Step 1: Training Dataset Construction and Standardized Encoding: Obtain sequence information, three-dimensional structural coordinates, and functional annotation labels of multiple protein samples from a known structural protein database, and perform standardized encoding on each sample to obtain the sequence feature matrix, structural feature tensor, and functional feature vector of each sample, thus forming the training dataset.

[0009] Step 2, Residue-level Local Feature Extraction and Global Context Encoding: Local convolutional encoding and bidirectional recurrent global encoding are performed on the sequence feature matrix respectively, and the local features and global features are concatenated to obtain the comprehensive sequence encoding features;

[0010] Step 3: Extraction of structural geometric features and encoding of secondary structure tendency: Calculate the inter-residue distance matrix and intra-residue geometric features based on the structural feature tensor. After radial basis mapping, the distance matrix is ​​fused with the geometric features by attention to obtain the comprehensive structural coding features.

[0011] Step 4: Functional semantic feature embedding and function-guided cross-modal attention fusion: The functional feature vector is mapped to a functional embedding vector through a multilayer perceptron. The comprehensive sequence coding feature and the comprehensive structural coding feature are concatenated to form a content feature matrix. The functional embedding vector is used as the query matrix, and the content feature matrix is ​​used as both the key matrix and the value matrix. The cross-modal attention output is calculated and added to the comprehensive sequence coding feature to obtain the conditional coding feature.

[0012] Step 5: Construction of conditional probability path and network training based on flow matching generator: During the training phase, with the real skeleton coordinates of the training samples as the endpoint and Gaussian noise as the starting point, construct the noise conditional probability path and corresponding vector field from the starting point to the endpoint. Train the flow matching generator network based on the conditional coding features to estimate the vector field. The training loss is the mathematical expectation of the conditional flow matching loss function.

[0013] Step 6: Obtaining Conditional Inputs for the Inference Stage: Based on the target functional requirements of the skeleton to be generated, obtain the target functional label vector as the conditional input for the inference stage. If there is an initial skeleton fragment that can be referenced, its structural coordinates are also used as auxiliary conditional inputs.

[0014] Step 7: Iterative Denoising and Skeleton Coordinate Trajectory Tracking: In the inference phase, starting from the initial Gaussian noise, the noise, time variables, and conditional coding features are input into the trained stream matching generator network. The ordinary differential equation is iteratively integrated and solved to obtain the predicted... Atomic coordinate tensors are used to record intermediate states during iterations, forming a coordinate trajectory tracking sequence.

[0015] Step 8: Reconstruction of complete coordinates of main chain atoms and optimization of skeletal structure: based on prediction Atomic coordinate tensors are used to add the main chain to each residue using a geometric conformation reconstruction algorithm. Atoms and Atomic coordinates are used to obtain the complete main chain atomic coordinate tensor, and energy minimization optimization is performed to obtain the optimized skeleton structure;

[0016] Step 9: Multi-dimensional quality assessment and conditional feedback screening of the generated skeleton: Calculate the Laplace plot score and template modeling score of the optimized skeleton structure, identify the Laplace plot not allowed areas and the low-matching segment intervals of the template modeling, construct a comprehensive evaluation index, compare the comprehensive evaluation index with a preset threshold, and regenerate after accepting or selecting the corresponding position state superimposed with perturbation from the coordinate trajectory tracking sequence.

[0017] Preferably, in step three, a distance matrix between residues is calculated, where each element of the distance matrix represents the distance between any two residues. Euclidean distance between atomic coordinates; calculation of the dihedral angle of each residue and Pseudo-bond angles are used to form local geometric feature vectors, and the local geometric feature vectors of all residues are concatenated into a local structural feature matrix. The distance matrix is ​​input into the radial basis function encoding layer, and each distance value is mapped to a feature vector of fixed dimension to obtain global distance encoding features. The local structural feature matrix and the global distance encoding features are fused using a self-attention mechanism to obtain comprehensive structural encoding features.

[0018] Preferably, in step four, the integrated sequence coding features and the integrated structure coding features are concatenated along the feature dimension and linearly projected to form the content feature matrix; the functional embedding vector is expanded to the same length as the content feature matrix to form the query matrix; the content feature matrix is ​​used as both the key matrix and the value matrix; the scaled dot product attention output is calculated; and the output is added element-wise to the integrated sequence coding features to obtain the conditional coding features.

[0019] Preferably, the noisy conditional probability path constructed in step five ensures that the latent variables follow a standard Gaussian distribution at the start of time and a Gaussian distribution with minimal variance centered on the true skeleton coordinates at the end of time; the vector field is jointly determined by the current latent variables, the true skeleton coordinates, and the time variable; the flow matching generator network outputs an estimated vector field based on conditional coding features and the time variable; the conditional flow matching loss function is the expected value of the squared error between the estimated vector field and the true vector field; and no preset endpoint coordinates are required during the inference phase after training.

[0020] Preferably, in step seven, the Euler method is used to solve the ordinary differential equation, and iterative integration is performed from the start time to the end time. The update method for each step is to add the current state to the product of the step size and the output value of the flow matching generator network, wherein the step size is between 0.001 and 0.05. The intermediate states of the time variable in the interval of 0.7 to 0.9 during the iteration process are recorded to form a coordinate trajectory tracking sequence, which is used for the selection of the initial state for subsequent retry.

[0021] Preferably, in step eight, for each residue, according to the residue's Coordinates, the previous residue Coordinates and standard peptide bond geometry parameters are determined. Atoms and The spatial position of the atom; The atomic position is determined by the current residue. Coordinates, the previous residue The coordinates and the first standard key angle are determined. The atomic position is determined by the current residue. Coordinates, current residue Atomic coordinates and second standard bond angles are determined; the above reconstruction yields a result containing... The complete main chain atomic coordinate tensor of three types of atoms is input into the energy minimization module based on molecular force field to perform conjugate gradient optimization and obtain the optimized skeleton structure.

[0022] Preferably, in step nine, a comprehensive evaluation index is constructed. The comprehensive evaluation index is equal to the first weighting coefficient multiplied by the Laplace plot score plus the second weighting coefficient multiplied by the template modeling score, where the first weighting coefficient is 0.4 and the second weighting coefficient is 0.6. An acceptable threshold is set. When the comprehensive evaluation index is greater than or equal to the acceptable threshold, the currently generated skeleton is accepted as the final output skeleton. When the comprehensive evaluation index is less than the acceptable threshold, based on the residue position index of the disallowed region in the Laplace plot and the low-matching fragment interval in the template modeling, the corresponding residue position state with the time variable in the interval of 0.7 to 0.9 is selected from the coordinate trajectory tracking sequence. This part of the state is perturbed by Gaussian noise with an amplitude of 0.005 and used as a new initial state. The iterative integration of step seven is executed again from the beginning to generate a new skeleton candidate. If the number of regenerations reaches the preset maximum number of retries and still does not meet the quality threshold, the skeleton candidate with the highest current comprehensive evaluation index is output.

[0023] Preferably, in the iterative solution process of step seven, during the time interval from the starting point to 0.3, Gaussian perturbation noise with an amplitude of 0.01 is added to the current state. This perturbation only takes effect within the time interval. The cross-modal attention mechanism adopts a multi-head attention form, with 8 attention heads and 32 output dimensions for each head. The outputs of all heads are concatenated on the feature dimension and passed through a linear projection layer to obtain the final cross-modal attention output.

[0024] This invention offers the following advantages: By aligning and fusing protein sequence features, structural geometry features, and functional semantic features across modalities, it constructs a generative AI model based on conditional diffusion or flow matching, achieving protein backbone generation with high structural rationality and functional customizability. The technical solution employed in this invention effectively enhances the diversity and designability of the generated backbone while significantly reducing computational costs, demonstrating clear industrial application value. Attached Figure Description

[0025] To more intuitively illustrate the prior art and this application, exemplary drawings are provided below. It should be understood that the specific shapes and structures shown in the drawings should not generally be regarded as limiting conditions for implementing this application; for example, based on the technical concept disclosed in this application and the exemplary drawings, those skilled in the art are able to easily make conventional adjustments or further optimizations to the addition / reduction / classification, specific shapes, positional relationships, connection methods, size ratios, etc. of certain units (components).

[0026] Figure 1 A flowchart illustrating the generative AI protein backbone generation method that integrates multimodal data, provided in an embodiment of this application. Detailed Implementation

[0027] The following specific embodiments illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. It should be understood that these embodiments are merely for further explanation of the present invention and should not be construed as limiting the scope of protection of the present invention. Technical engineers in the field can make some non-essential improvements and adjustments to the present invention based on the above-described content. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0028] Please see Figure 1 A generative AI protein backbone generation method integrating multimodal data includes the following steps:

[0029] Step 1: Training Dataset Construction and Standardized Encoding: Obtain sequence information, three-dimensional structural coordinates, and functional annotation labels of multiple protein samples from a known structural protein database, and perform standardized encoding on each sample to obtain the sequence feature matrix, structural feature tensor, and functional feature vector of each sample, thus forming the training dataset.

[0030] Step 2, Residue-level Local Feature Extraction and Global Context Encoding: Local convolutional encoding and bidirectional recurrent global encoding are performed on the sequence feature matrix respectively, and the local features and global features are concatenated to obtain the comprehensive sequence encoding features;

[0031] Step 3: Extraction of structural geometric features and encoding of secondary structure tendency: Calculate the inter-residue distance matrix and intra-residue geometric features based on the structural feature tensor. After radial basis mapping, the distance matrix is ​​fused with the geometric features by attention to obtain the comprehensive structural coding features.

[0032] Step 4: Functional semantic feature embedding and function-guided cross-modal attention fusion: The functional feature vector is mapped to a functional embedding vector through a multilayer perceptron. The comprehensive sequence coding feature and the comprehensive structural coding feature are concatenated to form a content feature matrix. The functional embedding vector is used as the query matrix, and the content feature matrix is ​​used as both the key matrix and the value matrix. The cross-modal attention output is calculated and added to the comprehensive sequence coding feature to obtain the conditional coding feature.

[0033] Step 5: Construction of conditional probability path and network training based on flow matching generator: During the training phase, with the real skeleton coordinates of the training samples as the endpoint and Gaussian noise as the starting point, construct the noise conditional probability path and corresponding vector field from the starting point to the endpoint. Train the flow matching generator network based on the conditional coding features to estimate the vector field. The training loss is the mathematical expectation of the conditional flow matching loss function.

[0034] Step 6: Obtaining Conditional Inputs for the Inference Stage: Based on the target functional requirements of the skeleton to be generated, obtain the target functional label vector as the conditional input for the inference stage. If there is an initial skeleton fragment that can be referenced, its structural coordinates are also used as auxiliary conditional inputs.

[0035] Step 7: Iterative Denoising and Skeleton Coordinate Trajectory Tracking: In the inference phase, starting from the initial Gaussian noise, the noise, time variables, and conditional coding features are input into the trained stream matching generator network. The ordinary differential equation is iteratively integrated and solved to obtain the predicted... Atomic coordinate tensors are used to record intermediate states during iterations, forming a coordinate trajectory tracking sequence.

[0036] Step 8: Reconstruction of complete coordinates of main chain atoms and optimization of skeletal structure: based on prediction Atomic coordinate tensors are used to add the main chain to each residue using a geometric conformation reconstruction algorithm. Atoms and Atomic coordinates are used to obtain the complete main chain atomic coordinate tensor, and energy minimization optimization is performed to obtain the optimized skeleton structure;

[0037] Step 9: Multi-dimensional quality assessment and conditional feedback screening of the generated skeleton: Calculate the Laplace plot score and template modeling score of the optimized skeleton structure, identify the Laplace plot not allowed areas and the low-matching segment intervals of the template modeling, construct a comprehensive evaluation index, compare the comprehensive evaluation index with a preset threshold, and regenerate after accepting or selecting the corresponding position state superimposed with perturbation from the coordinate trajectory tracking sequence.

[0038] In one specific embodiment of the present invention, a generative AI protein backbone generation method integrating multimodal data is provided. This method first requires constructing a training dataset, and then performs inference generation after training is complete.

[0039] Step 1: Training Dataset Construction and Standardization Encoding. This involves using a database of known protein structures (e.g., a protein structure database). This method retrieves the sequence information, 3D structural coordinates, and functional annotation labels of multiple protein samples. For each protein sample, the length is [length missing]. One-hot encoding is performed on its sequence information to obtain the sequence feature matrix. , dimension ×20, where 20 represents the number of standard amino acid types; its three-dimensional structural coordinates are used to extract the main chain atoms ( The Cartesian coordinate tensor of the entire skeleton structure is normalized to ensure that the geometric center of the entire skeleton structure is located at the origin, and then uniformly scaled to a unit scale to obtain standardized structural features. , dimension ×3×3; The functional annotation tags (including gene ontology molecular function tags and biological pathway tags) are binarized to obtain functional feature vectors. , dimension This yields the sequence feature matrix, structural feature tensor, and functional feature vector for each sample, forming the training dataset.

[0040] Step 2: Residue-level local feature extraction and global context encoding. This involves processing the sequence feature matrix... A one-dimensional convolutional neural network is used to extract local sequence patterns. The kernel size is 5, and the number of output channels is 128. Activation yields local features of the sequence. , dimension ×128. Simultaneously, the sequence feature matrix... The input is fed into a bidirectional long short-term memory network for global context encoding to obtain sequence global context features. , dimension ×256. (This appears to be a number or formatting error.) and By concatenating the features along the feature dimension, a comprehensive sequence coding feature is obtained. , dimension ×384.

[0041] Step 3: Structural geometric feature extraction and secondary structure tendency coding. Based on standardized structural features... Calculate the value of each residue Pairwise distance matrix between atomic coordinates , dimension ,in Calculate the dihedral angle corresponding to each residue. as well as pseudo-bond corner , forming local geometric feature vectors The dimension is 6. The local geometric feature vectors of all residues are concatenated into a local structural feature matrix. , dimension ×6. The pairwise distance matrix The input is fed into a radial basis function encoding layer, which maps each distance value to a 64-dimensional feature vector to obtain the global distance encoding features. , dimension ×64. A self-attention mechanism is used to analyze the local structural feature matrix. Global distance encoding features By fusing the features, a comprehensive structural coding feature is obtained. , dimension ×256.

[0042] Step 4: Functional semantic feature embedding and function-guided cross-modal attention fusion. The functional feature vectors are then... The input is fed into a multilayer perceptron, which contains two hidden layers with 512 and 256 neurons respectively, and outputs a fixed-length function embedding vector. The dimension is 256. The integrated sequence encoding features will be used. With integrated structural coding features The feature matrix is ​​obtained by concatenating along the feature dimensions and compressing it to 256 dimensions through a linear projection layer. , dimension ×256. A cross-modal attention mechanism is employed to embed the function into the vector. Expand to length Later used as a query matrix (dimension is) ×256), the content feature matrix Simultaneously serving as a bond matrix Sum matrix (All dimensions are) (×256), calculate cross-modal attention output. :

[0043]

[0044] in The feature dimension of the key matrix is ​​256. The output is a structure-sequence joint code guided by functional semantics. Combined with sequence coding features By adding elements one by one, we obtain the conditional coding features. , dimension ×384.

[0045] Step 5: Conditional probability path construction and network training based on the flow matching generator. During the training phase, the true skeleton coordinates of a sample are sampled from the training dataset. (i.e., the sample) Atomic coordinate tensor, dimension ×3) as the endpoint sample. Initialize a standard Gaussian noise latent variable. Dimensions and Same. Define a continuous-time variable. , Construct a noisy conditional probability path. , making exist The time follows a standard Gaussian distribution. Time follows the endpoint sample Centered, variance Gaussian distribution, The value is 0.01. Defined from... arrive vector field :

[0046]

[0047] Based on the conditional coding features corresponding to the training samples Using based Stream matching generator network architecture To estimate the vector field, These are the trainable parameters of the network. The conditional flow matching loss function is used during training. :

[0048]

[0049] By minimizing the loss function To optimize generator network parameters After training, the inference phase starts from standard Gaussian noise and does not require preset endpoint coordinates. The generated skeleton can be obtained simply by integrating along the estimated vector field using the trained generator network.

[0050] Step Six: Obtaining Conditional Inputs for the Inference Stage. Based on the target functional requirements of the skeleton to be generated, obtain the target functional label vector, and after binarization encoding, use it as the conditional input features for the inference stage. If there is a referenceable initial skeleton fragment or known partial structural information, the corresponding... Atomic coordinate tensors serve as auxiliary structural condition inputs, and together with the target function label vector, they form the multimodal condition input set for the reasoning stage.

[0051] Step 7: Iterative Denoising and Skeleton Coordinate Trajectory Tracking. During the inference phase, initial noise is sampled from a standard Gaussian distribution. .Will Time variables and the conditional coding features obtained in step six Input to the trained generator network Euler's method is used to solve ordinary differential equations To solve, from arrive Perform iterative integration to obtain the prediction. Atomic coordinate tensor During the iteration process, the update formula for each step is:

[0052]

[0053] in The integration step size ranges from [0.001, 0.05]. The time variable is recorded during the iteration process. All intermediate states within the range of 0.7 to 0.9 This forms a coordinate trajectory tracking sequence, which is used to select the initial state for retry when the subsequent quality assessment fails to meet the standards.

[0054] Step 8: Reconstruction of complete main chain atom coordinates and optimization of the skeletal structure. Based on the predicted coordinates... Atomic coordinate tensor A conformational reconstruction algorithm based on geometry is used to add a main chain to each residue. Atoms and The coordinates of the atom. For the first... Each residue, according to its Coordinates, previous residue Coordinates and standard peptide bond geometry parameters (bond length, bond angle) are determined. Atoms and The spatial location of atoms. Specifically, Atomic position Depend on , Determined by the standard bond angle of 109.5°; C atom position Depend on , The standard bond angle of 111° is determined. After the above reconstruction, the following is obtained: , , Complete main chain atomic coordinate tensor of three types of atoms , dimension ×3×3. (The rest of the text appears to be incomplete and cannot be translated.) The input is fed into a short-time energy minimization module based on molecular force fields, where 100 steps of conjugate gradient optimization are performed to eliminate unreasonable interatomic contacts, resulting in the optimized framework structure. .

[0055] Step Nine: Generate a multi-dimensional quality assessment and conditional feedback screening of the skeleton. Calculate the optimized skeleton structure. Laplace score The proportion of residues located in the permitted and preferred regions is statistically analyzed, and the set of indexes of residues falling into the prohibited regions is recorded; the optimized backbone structure is then calculated. Modeling score based on templates of known structures in the natural skeleton library Record consecutive residue intervals with matching scores below a preset fragment threshold of 0.5. Construct a comprehensive evaluation index. :

[0056]

[0057] in and For preset weighting coefficients, The value is 0.4. The value is 0.6. This sets the acceptable threshold. The value is 0.75, when the comprehensive evaluation index is... When the current generated skeleton is used, it is accepted as the final output skeleton. When comprehensive evaluation metrics are considered... At that time, based on the set of residue position indices and the continuous residue interval, the corresponding residue position state with the time variable in the interval of 0.7 to 0.9 is selected from the coordinate trajectory tracking sequence recorded in step seven. This part of the state is then subjected to Gaussian noise perturbation with an amplitude of 0.005 as a new initial state. The iterative integration in step seven is then re-executed to generate a new backbone candidate. If the number of regenerations reaches the preset maximum number of retries (preset to 5 times) and still does not meet the quality threshold, the backbone candidate with the highest current comprehensive evaluation index is output.

[0058] Thus, through the construction of the training dataset, multimodal data fusion, flow matching-based conditional generation, iterative denoising and tracking, and quality screening with trajectory feedback, high-precision and highly controllable protein backbone generation has been achieved.

[0059] In another specific embodiment of the present invention, for the one-dimensional convolutional neural network in step two, multi-scale sequence patterns can be extracted in parallel using convolutional kernels of different sizes. For example, three convolutional layers with kernel sizes of 3, 5, and 7 can be used simultaneously, and the outputs of each channel are concatenated after max pooling to enhance the richness of local feature expression. For the radial basis function encoding layer in step three, its center point can be set to 64 values ​​uniformly distributed in the range of 0.5 to 20.0 angstroms, and the width parameter can be set to 1.0 to cover the conventional protein structure. The range of distances between atoms.

[0060] In yet another specific embodiment of the present invention, the conformational reconstruction algorithm in step eight can further incorporate sidechains. The positional information of atoms is used to assist in the dihedral angle correction of the main chain. Specifically, in determining... When referring to atomic positions Atoms and The direction of the lines connecting atoms, for The spatial orientation of the atoms is fine-tuned to adjust the dihedral angles of the reconstructed main chain. It more closely matches the Laplace plot distribution of the natural conformation. The fine-tuning range is controlled within ±5°, and convergence is achieved in three iterations.

[0061] In another specific embodiment of the present invention, to improve the diversity of the generated skeleton, during the iterative solution process in step seven, Within the time interval [0, 0.3], move towards the current state. Add Gaussian perturbation noise with an amplitude of 0.01. The specific implementation of adding the perturbation is as follows: ,in , The perturbation magnitude is set to 0.01. This perturbation only applies to the first 30% of iterations, maintaining a deterministic trajectory in later iterations to balance generative diversity with structural convergence.

[0062] In another specific embodiment of the present invention, the cross-modal attention mechanism can be extended to a multi-head attention form, with 8 attention heads, each head having an output dimension of 32. The outputs of all heads are concatenated along the feature dimension and passed through a linear projection layer to obtain a final cross-modal attention output A_cross_multi with a dimension of 256. This multi-head mechanism can capture the interaction relationships between sequence-structure-function from different semantic subspaces, thereby improving the expressive capacity of conditional coding features.

[0063] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A generative AI protein backbone generation method integrating multimodal data, characterized in that, Includes the following steps: Step 1: Training Dataset Construction and Standardized Encoding: Obtain sequence information, three-dimensional structural coordinates, and functional annotation labels of multiple protein samples from a known structural protein database, and perform standardized encoding on each sample to obtain the sequence feature matrix, structural feature tensor, and functional feature vector of each sample, thus forming the training dataset. Step 2, Residue-level Local Feature Extraction and Global Context Encoding: Local convolutional encoding and bidirectional recurrent global encoding are performed on the sequence feature matrix respectively, and the local features and global features are concatenated to obtain the comprehensive sequence encoding features; Step 3: Extraction of structural geometric features and encoding of secondary structure tendency: Calculate the inter-residue distance matrix and intra-residue geometric features based on the structural feature tensor. After radial basis mapping, the distance matrix is ​​fused with the geometric features by attention to obtain the comprehensive structural coding features. Step 4: Functional semantic feature embedding and function-guided cross-modal attention fusion: The functional feature vector is mapped to a functional embedding vector through a multilayer perceptron. The comprehensive sequence coding feature and the comprehensive structural coding feature are concatenated to form a content feature matrix. The functional embedding vector is used as the query matrix, and the content feature matrix is ​​used as both the key matrix and the value matrix. The cross-modal attention output is calculated and added to the comprehensive sequence coding feature to obtain the conditional coding feature. Step 5: Construction of conditional probability path and network training based on flow matching generator: During the training phase, with the real skeleton coordinates of the training samples as the endpoint and Gaussian noise as the starting point, construct the noise conditional probability path and corresponding vector field from the starting point to the endpoint. Train the flow matching generator network based on the conditional coding features to estimate the vector field. The training loss is the mathematical expectation of the conditional flow matching loss function. Step 6: Obtaining Conditional Inputs for the Inference Stage: Based on the target functional requirements of the skeleton to be generated, obtain the target functional label vector as the conditional input for the inference stage. If there is an initial skeleton fragment that can be referenced, its structural coordinates are also used as auxiliary conditional inputs. Step 7: Iterative Denoising and Skeleton Coordinate Trajectory Tracking: In the inference phase, starting from the initial Gaussian noise, the noise, time variables, and conditional coding features are input into the trained stream matching generator network. The ordinary differential equation is iteratively integrated and solved to obtain the predicted... Atomic coordinate tensors are used to record intermediate states during iterations, forming a coordinate trajectory tracking sequence. Step 8: Reconstruction of complete coordinates of main chain atoms and optimization of skeletal structure: based on prediction Atomic coordinate tensors are used to add the main chain to each residue using a geometric conformation reconstruction algorithm. Atoms and Atomic coordinates are used to obtain the complete main chain atomic coordinate tensor, and energy minimization optimization is performed to obtain the optimized skeleton structure; Step 9: Multi-dimensional quality assessment and conditional feedback screening of the generated skeleton: Calculate the Laplace plot score and template modeling score of the optimized skeleton structure, identify the Laplace plot not allowed areas and the low-matching segment intervals of the template modeling, construct a comprehensive evaluation index, compare the comprehensive evaluation index with a preset threshold, and regenerate after accepting or selecting the corresponding position state superimposed with perturbation from the coordinate trajectory tracking sequence.

2. The generative AI protein backbone generation method based on multimodal data according to claim 1, characterized in that, In step three, the distance matrix between residues is calculated, where each element of the distance matrix represents any two residues. Euclidean distance between atomic coordinates; calculation of the dihedral angle of each residue and Pseudo-bond angles are used to form local geometric feature vectors, and the local geometric feature vectors of all residues are concatenated into a local structural feature matrix. The distance matrix is ​​input into the radial basis function encoding layer, and each distance value is mapped to a feature vector of fixed dimension to obtain global distance encoding features. The local structural feature matrix and the global distance encoding features are fused using a self-attention mechanism to obtain comprehensive structural encoding features.

3. The generative AI protein backbone generation method based on multimodal data according to claim 2, characterized in that, In step four, the integrated sequence coding features and the integrated structure coding features are concatenated along the feature dimension and linearly projected to form the content feature matrix. The functional embedding vector is expanded to the same length as the content feature matrix and used as the query matrix. The content feature matrix is ​​used as both the key matrix and the value matrix. The scaled dot product attention output is calculated and then added element-wise to the comprehensive sequence coding features to obtain the conditional coding features.

4. The generative AI protein backbone generation method based on multimodal data according to claim 3, characterized in that, The noisy conditional probability path constructed in step five ensures that the latent variables follow a standard Gaussian distribution at the start of time and a Gaussian distribution with minimal variance centered on the true skeleton coordinates at the end of time. The vector field is jointly determined by the current latent variables, the true skeleton coordinates, and the time variable. The flow matching generator network outputs an estimated vector field based on conditional coding features and the time variable. The conditional flow matching loss function is the expected value of the squared error between the estimated vector field and the true vector field. Once training is complete, there is no need to preset the endpoint coordinates during the inference phase.

5. The generative AI protein backbone generation method based on multimodal data according to claim 4, characterized in that, In step seven, the Euler method is used to solve the ordinary differential equation. Iterative integration is performed from the start time to the end time. Each update method is to add the current state to the product of the step size and the output value of the flow matching generator network, where the step size is between 0.001 and 0.

05. The intermediate states of the time variable in the interval of 0.7 to 0.9 during the iteration process are recorded to form a coordinate trajectory tracking sequence. This sequence is used for the selection of the initial state for subsequent retry.

6. The generative AI protein backbone generation method based on multimodal data according to claim 5, characterized in that, In step eight, for each residue, according to the residue... Coordinates, the previous residue Coordinates and standard peptide bond geometry parameters are determined. Atoms and The spatial position of the atom; The atomic position is determined by the current residue. Coordinates, the previous residue The coordinates and the first standard key angle are determined. The atomic position is determined by the current residue. Coordinates, current residue Atomic coordinates and second standard bond angles are determined; the above reconstruction yields a result containing... The complete main-chain atomic coordinate tensor of three types of atoms; The complete main chain atomic coordinate tensor is input into the energy minimization module based on molecular force field to perform conjugate gradient optimization, resulting in the optimized skeleton structure.

7. The generative AI protein backbone generation method based on multimodal data according to claim 6, characterized in that, Step nine involves constructing comprehensive evaluation indicators. The comprehensive evaluation index is equal to the first weighting coefficient multiplied by the Laplace score plus the second weighting coefficient multiplied by the template modeling score, where the first weighting coefficient is 0.4 and the second weighting coefficient is 0.

6. Set an acceptable threshold. When the comprehensive evaluation index is greater than or equal to the acceptable threshold, accept the currently generated skeleton as the final output skeleton. When the comprehensive evaluation index is less than the acceptable threshold, based on the residue position index of the disallowed region in the Laplace plot and the low-matching fragment interval modeled by the template, the corresponding residue position state with the time variable in the range of 0.7 to 0.9 is selected from the coordinate trajectory tracking sequence. This part of the state is then perturbed with Gaussian noise of amplitude 0.005 and used as a new initial state. The iterative integration of step seven is then performed again from the beginning to generate a new backbone candidate. If the number of regenerations reaches the preset maximum number of retries and still does not meet the quality threshold, the backbone candidate with the highest current comprehensive evaluation index is output.

8. The generative AI protein backbone generation method based on multimodal data according to claim 7, characterized in that, In the iterative solution process of step seven, during the time interval from the starting point to 0.3, Gaussian perturbation noise with an amplitude of 0.01 is added to the current state. This perturbation only takes effect within the time interval. The cross-modal attention output mechanism adopts a multi-head attention form, with 8 attention heads and 32 output dimensions for each head. The outputs of all heads are concatenated on the feature dimension and passed through a linear projection layer to obtain the final cross-modal attention output.

Images