SE (3) isotropic diffusion and ex-situ generation-based protein function topology design method and product

By using the SE(3) isovariant diffusion and non-in situ generation method, the limitations of function-oriented and diversity expression in existing protein design are solved, and unified modeling of function and physical properties is achieved, which improves the freedom and diversity of protein design and supports personalized protein generation.

CN120877837APending Publication Date: 2025-10-31XINJIANG UNIVERSITY +1
View PDF 0 Cites 1 Cited by

Patent Information

Application Number
CN202511011762.6
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2025-07-22
Publication Date
2025-10-31

AI Technical Summary

Technical Problem

Existing protein design methods have limitations in terms of function-directedness, diverse expression, and cross-topological generalization ability, making it difficult to achieve high-degree-of-freedom de novo design, and lacking unified modeling of functional and physical properties.

Method used

By employing an SE(3) isovariant diffusion and non-in-situ generation method, and combining feature extraction, non-in-situ generation network and backbone generation network with geometric modeling, physical energy guidance and knowledge guidance strategies, we can achieve flexible insertion and reconstruction of protein functional fragments in three-dimensional space.

Benefits of technology

It improves the spatial freedom and topological diversity of protein design, achieves the unity of function and physical properties, and supports the generation of multifunctional proteins and personalized customization in biomanufacturing.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN120877837A_ABST
    Figure CN120877837A_ABST
Patent Text Reader

Abstract

The invention provides a protein function topology design method based on SE (3) isovariant diffusion and ex-situ generation and a product, and relates to the technical field of protein design. According to the method, geometric deep learning and generative artificial intelligence are fused, isovariant generation and optimization of protein function sites under the action of a three-dimensional Euclidean space (SE (3) group) are achieved, and the limitation of traditional protein design on conformation sampling efficiency, function guidance and physical realizability is broken through.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The embodiments of the present invention relate to the field of protein design technology, and in particular to a protein functional topology design method and product based on SE(3) isovariant diffusion and non-in situ generation. Background Technology

[0002] With the rapid development of computational science, artificial intelligence, and structural biology, protein design is becoming a cutting-edge area of ​​interdisciplinary research, especially in the field of de novo design. Researchers are attempting to synthesize proteins with novel structures and functions using computational methods to meet specific biological or engineering application needs. This not only helps to reveal the essence of life but also provides broad application prospects for drug development, vaccine design, and biomaterial construction.

[0003] As a core component of synthetic biology, protein design is propelling life sciences into a new era. Synthetic biology is hailed as the third revolution in life sciences, following the discovery of the DNA double helix structure and the Human Genome Project, and promises to achieve a fundamental shift from natural extraction to biomanufacturing.

[0004] Several technological approaches have been developed for protein design, mainly including the following categories: Methods based on physical energy functions, such as Rosetta and AMBER molecular force fields, find the most stable sequence and structure combinations by minimizing the energy function. While physically sound, they are computationally expensive and difficult to meet the practical needs of rapid iteration. Strategies based on statistical potential and heuristic search, such as Fragment Assembly and Monte Carlosampling, offer high design efficiency but are prone to getting trapped in local optima and struggle to escape the natural protein distribution range. Sequence generation methods based on deep learning, such as ProteinMPNN and ProGen, can directly generate sequences that satisfy specific structural or functional labels, but most rely on templates, limiting their ability to explore high-degree-of-freedom structural spaces. Iterative optimization strategies from structure to sequence, such as RoseTTAFold All-atom and AlphaDesign, improve adaptability through repeated optimization between structure and sequence, but still lack the flexibility and generalization ability of function-driven approaches. Structure generation methods based on generative models, represented by diffusion models, model continuous conformation generation processes starting from random structures, such as RFdiffusion. These methods excel in 3D spatial modeling, especially in ensuring geometric consistency. In recent years, the rapid development of deep learning, especially the success of structure prediction models such as AlphaFold and RoseTTAFold, has significantly improved the modeling accuracy of protein 3D conformations. However, these methods still focus on structure modeling and lack the systematic ability to integrate functional constraints and spatial structure from a "function-based" perspective. The introduction of diffusion models, especially combined with denaturation such as SE(3), gives the generation process rotational and translational symmetry invariance, providing a natural advantage in handling 3D protein structures. RFdiffusion is a typical example, demonstrating an efficient modeling process from a random initial state to a stable conformation.

[0005] Nevertheless, existing models still face the following challenges in function-oriented protein design: (1) In-situ structural constraints limit the degree of freedom: Most current methods model functional sites (such as ligand binding pockets, catalytic centers, etc.) in the initial structure, which limits the overall flexibility of the structure and is not conducive to exploring diverse structural realization paths.

[0006] (2) Lack of cross-topology generalization ability: Diffusion models are usually trained on fixed data distributions or templates, making it difficult to adapt to new structural topologies and functional requirements, thus limiting their versatility and innovation.

[0007] (3) Functional modeling is disconnected from physical rationality: Current methods mainly focus on geometric rationality and fail to systematically integrate key physical properties such as dynamic characteristics, flexible adjustment, and binding energy, resulting in insufficient coupling between functional performance and physical structure.

[0008] It is evident that although protein generation technology based on deep learning and geometric modeling has made significant progress in recent years, especially the diffusion model, which demonstrates good geometric consistency in three-dimensional structural modeling, existing methods still face many bottlenecks in function-oriented protein design tasks. Current mainstream technologies generally suffer from strong reliance on structural templates and rigid functional fragment localization, severely limiting the flexibility and diversity of the design space. Furthermore, their generalization ability in unseen structural topologies and functional types is insufficient, making it difficult to support true de novo design. In addition, the lack of a unified modeling framework between functional modeling and physical rationality results in limited usability of generated structures in practical applications; and most SE(3) isovariant models still rely on large-scale pre-training, lacking clear theoretical support, further weakening their practicality and interpretability. These problems collectively reflect that while current methods have made some progress in conformational modeling, they still lack a systematic and effective solution path for high-degree-of-freedom generation of functional proteins, cross-space design, and integrated function-physical modeling. Therefore, it is urgent to construct a new protein functional topology design scheme. Summary of the Invention

[0009] This invention provides a protein functional topology design method and product based on SE(3) isovariant diffusion and non-in-situ generation, which can abandon the constraints of traditional in-situ design, integrate functional topology and physical property expression, and support the flexible insertion and reconstruction of functional fragments in three-dimensional Euclidean space.

[0010] The first aspect of this invention provides a method for designing protein functional topologies based on SE(3) isovariant diffusion and non-in situ generation, the method comprising: The protein sequence information corresponding to the functional fragment of the target protein is input into the feature extraction network of the non-in situ generated protein prediction framework; the feature extraction network processes the protein sequence information to obtain the corresponding expression vector. The expression vector is input into the non-in situ generation network of the non-in situ protein prediction framework. The non-in situ generation network generates three-dimensional coordinates of a protein with a function similar to the functional fragment of the target protein based on the expression vector. The non-in situ generation network uses the three-dimensional coordinates of the sample protein as input features and the protein sequence information of the sample protein as conditional information to learn the probability distribution of the three-dimensional coordinates of the protein under specific protein sequence information. The protein's three-dimensional coordinates are converted into data in a preset format. The preset format data representing the protein's three-dimensional coordinate information is then input into a backbone generation network for generating a protein prediction framework in situ. The backbone generation network generates and predicts the preset format data. Based on geometric modeling, physical energy guidance, structural hierarchical control, and knowledge guidance strategies, a protein backbone with the target function is obtained. Based on the protein backbone, sequence design and structural modeling are performed to obtain a protein with the target function.

[0011] A second aspect of this invention provides a protein functional topology design device based on SE(3) isovariant diffusion and non-in situ generation, the device comprising: The feature extraction module is used to input the protein sequence information corresponding to the functional fragment of the target protein into the feature extraction network of the non-in situ generated protein prediction framework; the feature extraction network processes the protein sequence information to obtain the corresponding expression vector. The in-situ generation module is used to input the expression vector into the in-situ generation protein prediction framework's in-situ generation network. The in-situ generation network generates three-dimensional coordinates of a protein with a function similar to the target protein's functional fragment based on the expression vector. The in-situ generation network uses the three-dimensional coordinates of the sample protein as input features and the protein sequence information of the sample protein as conditional information, and learns the probability distribution of the protein's three-dimensional coordinates under specific protein sequence information. The backbone generation module is used to convert the three-dimensional coordinates of the protein into data in a preset format. The preset format data representing the three-dimensional coordinate information of the protein is input into the backbone generation network for generating a protein prediction framework in situ. The backbone generation network generates and predicts the preset format data and obtains a protein backbone with the target function based on geometric modeling, physical energy guidance, structural hierarchical control and knowledge guidance strategies. An optimization module is used to perform sequence design and structural modeling based on the protein backbone to obtain a protein with the target function.

[0012] A third aspect of the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes a protein functional topology design method based on SE(3) isovariant diffusion and non-in-situ generation as described in the first aspect of the present invention.

[0013] The fourth aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the protein functional topology design method based on SE(3) isovariant diffusion and non-in-situ generation as described in the first aspect of the present invention.

[0014] The fifth aspect of the present invention provides a computer program product, including a computer program / instruction, which is implemented by a processor as the steps in the protein functional topology design method based on SE(3) isovariant diffusion and non-in situ generation as described in the first aspect of the present invention.

[0015] This invention integrates geometric deep learning and generative artificial intelligence to achieve isovariant generation and optimization of protein functional sites in three-dimensional Euclidean space (SE(3) group action), breaking through the limitations of traditional protein design in conformational sampling efficiency, functional guidance and physical realizability.

[0016] Specifically, in this embodiment of the invention, by decoupling the binding of protein functional fragments to fixed spatial positions, protein functional fragments are allowed to be flexibly embedded in three-dimensional structures and adaptively adjusted, thereby improving the degree of freedom in design space and enhancing topological diversity and innovation capabilities.

[0017] In this embodiment of the invention, multi-scale functional constraints are introduced on the basis of SE(3)-isovariant diffusion, and flexible modeling, dynamic stability analysis and binding energy prediction are combined to achieve the unity of functionality and physical rationality.

[0018] This invention not only continues the powerful capabilities of diffusion models in geometric modeling, but also establishes a truly "function-driven" new paradigm for protein design by introducing a non-in-situ function-driven mechanism and a function-structure joint modeling strategy. It enables the combination of different functional fragments of any protein to generate new proteins, which still retain their corresponding functions, thus achieving protein function customization. This provides solid theoretical support and technical pathways for the generation of multifunctional proteins, the exploration of complex structures, and personalized customization in biomanufacturing. Attached Figure Description

[0019] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments of the present invention will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0020] Figure 1 This is a flowchart of the steps of the protein functional topology design method based on SE(3) isovariant diffusion and non-in situ generation provided in the embodiments of the present invention; Figure 2 This is a training flowchart of the non-in-situ generated protein prediction framework based on the SE(3) isovariant diffusion and non-in-situ generated protein functional topology design method provided in this embodiment; Figure 3This is a hardware structure diagram of any device with data processing capabilities, which is the protein functional topology design device based on SE(3) isovariant diffusion and non-in-situ generation provided in the embodiments of the present invention. Detailed Implementation

[0021] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0022] To address the problems of rigid in-situ modeling, weak function-structure coupling, and lack of generalization ability in function-based protein design, this invention proposes a non-in-situ protein functional topology design method based on the SE(3) isovariant diffusion model. The core idea of ​​this method is to abandon the traditional in-situ assumption of "fixed-position embedding of functional fragments" and instead adopt a non-in-situ strategy. Through latent variable control and spatiotemporal awareness mechanisms, the distribution and reconstruction of functional fragments are flexibly realized in three-dimensional space, thereby achieving diversified, function-oriented protein generation. This method mainly includes two key processes: a non-in-situ generation modeling process and a pre-trained isovariant SE(3) model design, which together construct a protein design scheme with high expressive power, strong spatial freedom, and significant function guidance.

[0023] Protein design and reconstruction have always been core challenges in structural and synthetic biology. A widely adopted approach in traditional design strategies is fragment substitution based on an "in-situ template," where a specific location within an existing protein structure is selected and used as a template for local modification or replacement. This method assumes that functional fragments (such as ligand binding sites, catalytic centers, or conformational critical loops) must spatially remain consistent with the original template to preserve the original functionality or structural stability.

[0024] While this "in-situ" design strategy has achieved some success in the past, its limitations have become increasingly apparent as protein design tasks become more complex and diverse, mainly in the following aspects: a) Limited design space: The in-situ template strategy relies heavily on the spatial location within a known structure, which greatly limits the possible structural combinations and innovations. Designers can only make minor adjustments within a local area of ​​the original template, making it difficult to break out of the existing framework and find new functional implementation paths.

[0025] b) Over-reliance on existing databases: The structures included in current protein databases are far from covering all possible functions and conformations. In-situ methods can only utilize existing templates, which often proves inadequate when faced with design tasks involving entirely new functions or extreme conditions.

[0026] c) Difficulty in adapting to functional drift and conformational variation: In the process of natural evolution, protein functional fragments often migrate or remodel in different contexts. Traditional in situ strategies are difficult to capture these non-conserved but functionally similar structures, thus reducing the universality and robustness of the method.

[0027] d) Lack of diversity in generated results: Because the design is restricted to a fixed location, the resulting protein sequences tend to be homogeneous, making it difficult to meet the needs for functional variants or diverse designs. This is particularly prominent in fields such as drug development and enzyme engineering, limiting the richness of candidate molecules.

[0028] It is evident that breaking free from the reliance on in-situ templates and developing more free and flexible design methods that transcend spatial limitations and operate in high-dimensional structure-function semantic spaces has become a pressing issue in the field of protein design. This represents not only a paradigm shift from existing methods but also a crucial step towards achieving more diverse, efficient, and precise protein design.

[0029] De novo design of novel protein structures has always been a core challenge in protein engineering within the fields of biomedicine and chemistry. The ability to precisely design novel proteins holds immense potential for developing new biotherapies to address major public health challenges. However, traditional protein design methods are highly dependent on domain-specific knowledge, physicochemical modeling, and cumbersome and expensive experimental validation processes, making them inefficient and unable to meet the growing demand for functional proteins.

[0030] With the rapid development of deep learning technology, there is an increasing tendency to use data-driven, automated methods to extract the implicit rules of protein design, thereby significantly improving design efficiency and reducing development costs. Against this backdrop, designing novel proteins from scratch that meet specific structural or functional requirements has become an important research direction. However, this task still faces many challenges: a) The structure-sequence coupling problem is complex: the three-dimensional backbone structure of a protein is composed of several residues, each typically consisting of N, The four main chain atoms—C, O, and C—are connected by covalent bonds to form a continuous main chain. Because the three-dimensional structure of a protein is tightly coupled with its amino acid sequence, designing a rational structure and sequence simultaneously without relying on a template presents a significant technical challenge.

[0031] b) High-dimensional geometric modeling is challenging: Each residue's atomic group constitutes a precisely representable local frame. Therefore, many successful protein structure prediction and design methods represent their main chain structure as elements in the SE(3) group (i.e., a set considering rotation and translation in three-dimensional space). Since protein function is usually determined by the relative spatial configuration between atoms, these methods often require modeling for geometric transformation invariance (geometric equivariance). Although recent studies have extended diffusion models to complex geometric spaces such as Riemannian manifolds, providing scalable, stable, and easily trainable geometric equivariance models while maintaining theoretical rigor remains a research challenge.

[0032] c) The gap between theory and practice has not yet been bridged: Current deep learning methods for the SE(3) space have achieved some experimental success, but still lack a systematic theoretical foundation, relying heavily on heuristics in model design and loss function construction. For example, the RFdiffusion model proposed in related research successfully generated novel protein ligands with high affinity using a diffusion modeling framework. However, this method heavily relies on heuristic denoising loss functions and requires large-scale pre-training of protein structures in practice, which limits its widespread application to some extent.

[0033] Therefore, how to construct a protein generation model that does not require large-scale pre-training, has geometric isomorphic expression capabilities, and is unified in theory and practice to support the de novo generation of new proteins with specified structures or functions has become a key issue in current protein generation research and urgently needs to be addressed.

[0034] This invention proposes a protein functional topology design method that combines SE(3) isovariant diffusion and off-site generation strategies. It aims to fundamentally solve the core limitations of existing technologies in terms of function-driven, diverse expression and cross-topology generalization ability, and realize the off-site de novo generation of novel protein molecules with specific structures or functions.

[0035] Specifically, such as Figure 1 As shown, it illustrates the steps of the protein functional topology design method based on SE(3) isovariant diffusion and non-in-situ generation provided by an embodiment of the present invention. The method includes the following steps: S101, the protein sequence information corresponding to the functional fragment of the target protein is input into the feature extraction network of the non-in situ protein prediction framework; the feature extraction network processes the protein sequence information to obtain the corresponding expression vector.

[0036] In this embodiment of the invention, the feature extraction network encodes the protein sequence into an expression vector that can be used for subsequent generation.

[0037] In this embodiment of the invention, the feature extraction network extracts the sequence information contained in the protein sequence from the dataset, and extracts the structural pattern of the protein sequence through sequence encoding.

[0038] S102, the expression vector is input into the non-in situ generation network of the non-in situ generation protein prediction framework, and the non-in situ generation network generates three-dimensional coordinates of a protein with a function similar to the target protein functional fragment based on the expression vector.

[0039] The non-in-situ generation network uses the three-dimensional coordinates of the sample protein as input features and the protein sequence information of the sample protein as conditional information to learn the probability distribution of the three-dimensional coordinates of the protein under specific protein sequence information.

[0040] In this embodiment of the invention, the non-in-situ generation module uses the three-dimensional coordinates and sequence features of the protein as the input and output of the conditional variational autoencoder (CVAE), respectively, to achieve non-in-situ mapping from structure to sequence, providing a structural basis for scaffold modeling.

[0041] S103, the three-dimensional coordinates of the protein are converted into data in a preset format. The preset format data representing the three-dimensional coordinate information of the protein is input into the backbone generation network for generating a protein prediction framework in situ. The backbone generation network generates and predicts the preset format data. Based on geometric modeling, physical energy guidance, structural hierarchical control and knowledge guidance strategies, a protein backbone with target function is obtained.

[0042] S104. Based on the protein backbone, sequence design and structural modeling are performed to obtain a protein with the target function.

[0043] In this embodiment of the invention, during the backbone generation stage, a diffusion model based on geometric isovariability is introduced to generate a main backbone in three-dimensional space that satisfies structural or functional constraints, ensuring atomic-level geometric rationality and physical feasibility. Based on this, in step S104, existing high-performance protein sequence optimization tools (such as ProteinMPNN) can be used to perform sequence matching and stability optimization on the backbone. Finally, mature structural modeling tools in related technologies (such as AlphaFold2 and AlphaFold3) are used to predict the three-dimensional structure of the designed sequence, thereby verifying whether it can fold into the target structure.

[0044] In this embodiment of the invention, the preset format data is data that the backbone generation network, which characterizes the three-dimensional coordinate information of the protein, can recognize (e.g., mmCIF file format data or PDB file format data).

[0045] Therefore, in this embodiment of the invention, the non-in situ protein generation prediction framework effectively decouples the generation process of structure and sequence within the non-in situ framework, reduces model complexity, and improves the accuracy and reliability of structure generation by introducing geometric isovariant diffusion modeling, providing a systematic, efficient and scalable technical route for function-oriented de novo protein design.

[0046] In this embodiment of the invention, step S101 includes the following sub-steps: S1011, The protein sequence is positionally encoded using a positional encoding method.

[0047] Considering that protein sequences are essentially ordered data, in order to preserve the relative and absolute positional information of amino acids in the sequence, this embodiment of the invention uses a positional encoding method similar to that in the Transformer model to enhance the amino acid sequence.

[0048] For the first in the protein sequence pos The position and the first i The positional encoding method for the dimensional feature is as follows: .

[0049] in, The dimension representing the feature.

[0050] By using this alternating encoding method of sine and cosine functions, not only can positional information be captured, but it also has good discriminative ability at different frequencies, enabling the model to learn the relative relationships between positional information during training.

[0051] S1012, The protein sequence information is encoded using the Doc2Vec model to obtain a global vector representation of the protein sequence.

[0052] In this embodiment of the invention, in order to extract the global semantic information of the entire protein sequence, the protein sequence is treated as a "segment," where each amino acid is considered a "word." The Doc2Vec model is used to encode these "documents" (protein sequences) to obtain a global vector representation of the sequence.

[0053] The training objective of Doc2Vec is to enhance the relevance between words and their context by maximizing the objective function. ; Where: d∈D represents a protein sequence (i.e., a document); w∈ This represents an atom of an amino acid (i.e., a word). context(w) represents the context window of the current word w.

[0054] By optimizing the objective function described above, the Doc2Vec model can learn the semantic vector representation of each protein sequence, thereby capturing its overall semantic features.

[0055] S1013, The output vector of the Doc2Vec model is transformed based on the linear mapping layer to obtain the embedding vector.

[0056] To ensure that the output vector of Doc2Vec conforms to the feature dimensions required by subsequent models, this embodiment of the invention designs a linear mapping layer to transform it. Let the output vector of Doc2Vec be... The linear mapping parameters are The vector obtained after embedding is represented as: ; in This is a bias term. This transformation not only improves the matching between the feature vector and the downstream model structure, but also enhances its expressive power.

[0057] S1014, The embedding vector is input into the Transformer encoder to capture the potential structural relationships and biological function correlations within the protein sequence, and the expression vector corresponding to the protein sequence information is obtained.

[0058] After embedding, this embodiment of the invention inputs the embedding vector into a Transformer encoder to further extract deep features of the protein sequence and capture the complex dependencies between amino acids.

[0059] The core module in the Transformer architecture is the self-attention mechanism, which is calculated as follows: The core mechanism is self-attention, and the formula is: ; Where Q, K, and V represent query, key, and value, respectively, all of which are obtained through linear transformation of the input feature vector; This represents the dimension of the key vector; softmax is used to normalize the attention weights. Through this mechanism, the model can weight and aggregate information from different positions based on contextual relationships, thereby achieving dynamic adjustment and aggregation of information and effectively capturing potential structural relationships and biological function relevance within protein sequences.

[0060] like Figure 2 As shown, Figure 2The diagram illustrates the training flowchart of the non-in-situ generated protein prediction framework based on the SE(3) isovariant diffusion and non-in-situ generated protein functional topology design method provided in this embodiment. In this embodiment, the non-in-situ generated protein prediction framework is trained using the following steps: S1, the protein sequence information of the sample protein is input into the feature extraction network, which processes the protein sequence information to obtain the corresponding sample expression vector.

[0061] In this embodiment of the invention, the feature extraction network processes the protein sequence information to obtain the corresponding sample expression vector, which serves as the conditional information for the subsequent non-in situ generation network.

[0062] S2, the three-dimensional coordinates of the sample protein are used as input features, and the sample expression vector of the sample protein is used as conditional information and input into the in-situ generation network. The in-situ generation network learns the probability distribution of the three-dimensional coordinates of the protein under specific protein sequence information and generates the prediction result of the three-dimensional coordinates of the sample protein.

[0063] In this embodiment of the invention, each training sample includes a sample protein sequence and sample protein three-dimensional coordinates.

[0064] In this embodiment of the invention, to effectively achieve conditional generation of protein structures, a non-in-situ generation network based on a Conditional Variational Autoencoder (CVAE) is designed. This model can model complex distributions under given conditions, and is particularly suitable for generation tasks that require the introduction of prior knowledge or auxiliary information. In this embodiment, the three-dimensional coordinates of the protein are used as the main input feature, and the protein sequence (first-order sequence) is used as the conditional information. By learning the probability distribution of the protein structure (protein three-dimensional coordinates) x under a specific sequence condition c, non-autoregressive protein structure generation based on a given protein sequence is finally achieved. CVAE mainly includes three key modules: encoder, reparameterization, and decoder, which together constitute the forward propagation process of the conditional generation model.

[0065] In this embodiment of the invention, step S2 includes the following sub-steps: S21, the three-dimensional coordinates of the sample protein are used as input features, and the sample expression vector of the sample protein is used as sequence condition information and input to the encoder of the non-in situ generation network. The encoder obtains combined features through splicing operation, performs nonlinear mapping, and obtains the parameters of the latent space: mean and standard deviation.

[0066] The encoder's role is to map input data to a latent space to capture a low-dimensional representation of the input structure. Specifically, the encoder accepts the protein's three-dimensional coordinates (sample protein coordinates) x and conditional information (sample expression vectors obtained from other networks) c as joint input. This combined input is then concatenated to obtain combined features, which are then input into the neural network for nonlinear mapping to obtain the parameters of the latent space: the mean. μ and standard deviation σ : ; in, For encoder parameters, z It is a latent variable.

[0067] S22, sampling latent variables from the latent space based on reparameterized sampling. z The process makes latent variables z It becomes a differentiable function of the mean and standard deviation.

[0068] In order to sample latent variables from the latent space z To preserve the differentiability of the network, a reparameterization technique is introduced to separate the random sampling operation from the non-differentiable Gaussian distribution. The specific form is as follows: ; in, Let z be the noise sampled from a standard normal distribution, and ⊙ denote an element-wise multiplication operation. This method allows the latent variable z to become a differentiable function of the mean and standard deviation, thus ensuring that the model can be backpropagated to the encoder parameters. Optimize.

[0069] S23, the decoder of the non-in-situ generation network takes the latent variable z and sequence condition information as common inputs to generate corresponding protein three-dimensional coordinate predictions, learns the differences in protein three-dimensional coordinates under different sequence condition information, and realizes protein structure modeling with sequence conditions.

[0070] The decoder's task is to convert latent variables z With conditional information c A common input is used to generate corresponding protein 3D structure predictions through a neural network: ; in, This represents the parameters of the decoder network. The output protein 3D structure prediction is the network-generated protein 3D coordinates (usually represented as backbone or full atomic structure). The decoder input contains conditional information. c The model can learn the differences in generating structures under different sequence conditions, thereby achieving conditional structural modeling.

[0071] Therefore, in this embodiment of the invention, a trained in-situ generation network is used to generate proteins with similar functions based on protein sequence information of a given protein functional fragment.

[0072] S3, the three-dimensional coordinate prediction results of the sample protein are converted into sample preset format data, and the sample preset format data is input into the backbone generation network. The backbone generation network adds noise and removes noise to the sample preset format data based on the diffusion probability model to approximate the real protein backbone of the sample protein. During the training process, each step maintains the equivariance of the SE(3) transformation, introduces an external energy function as a physical constraint, controls the generation process in stages, first global topology, then local folding, and finally atomic refinement to improve the rationality of the structure, and introduces a dynamic alignment strategy to ensure that the generated protein backbone and the real protein backbone are consistent in space frame to frame.

[0073] In this embodiment of the invention, a novel diffusion generation model framework (backbone generation network) is designed to achieve high-quality, structurally sound, and physically feasible protein backbone structure generation. In this embodiment, the backbone generation network is based on a diffusion generation model (e.g., the Framediff architecture), introducing geometric energy constraint injection, multi-scale generation control, conformation library guidance, and dynamic matching mechanisms. This aims to improve the accuracy and stability of protein backbone structure generation, especially demonstrating significant advantages in modeling specific regions. Furthermore, the backbone generation network maintains SE(3) isovariability (i.e., invariance to spatial rotation and translation), ensuring theoretical completeness while also considering generation performance.

[0074] In this embodiment of the invention, the backbone generation network is trained using mmCIF file format data or PDB file format data to generate protein backbones.

[0075] In this embodiment of the invention, the protein backbone can be optimized, specifically including: sequence design using a sequence design tool (e.g., ProteinMPNN), structure prediction using a structure prediction tool (e.g., AlphaFold2), to complete the optimization of the protein backbone, obtain the final target protein, and further verify whether the target protein has the target function.

[0076] Diffusion probability models are a powerful class of generative models that approximate the data distribution by constructing a forward (noise-adding) process and a backward (denoising) process, ultimately achieving the ability to reconstruct target data from random noise. The core idea of ​​this method is to learn a parameterized backward process during training, enabling it to progressively restore Gaussian noise to samples conforming to the data distribution. The forward process then converts the real data... By gradually adding Gaussian noise, a series of intermediate states are obtained. Ultimately, it becomes pure noise. The calculation formula is as follows: ; in, It is the noise intensity at each step.

[0077] The reverse process trains a neural network to learn how to process noise. By gradually removing noise, the data can be restored. The calculation formula is as follows: ; in, and Predicted by a neural network.

[0078] Since the protein backbone naturally exists in three-dimensional Euclidean space, its modeling process must satisfy spatial symmetry. Therefore, the backbone generation network proposed in this embodiment is based on the SE(3) diffusion model, which operates on the SE(3) group, i.e., a combination of rotation and translation operations in three-dimensional space. The model ensures the robustness of the network in modeling structures under rotation and translation by constructing a diffusion process that satisfies isovariability constraints.

[0079] During the training phase, the model adds noise to the pose (frame) of the protein backbone structure to maintain the equivariance of the SE(3) transformation at each step; during the denoising phase, the denoising direction of each step is predicted by the equivariant network to gradually restore the backbone structure.

[0080] To improve the physical plausibility of the generated structures, in this embodiment of the invention, an external energy function (such as the Rosetta or AlphaFold energy function) is introduced as a physical constraint during training to guide the generated samples. The gradient of the energy function E(x) is used to guide the sampling direction and correct the sampling path in the reverse process. This process can be formalized as an energy-guided Langevin dynamic: ; in This represents the gradient of the energy function with respect to the structure, where λ is the weight. This mechanism encourages the generation of structures that better conform to the physical laws and energy stability of natural proteins.

[0081] In this embodiment of the invention, the generation process is controlled in stages during training: first global topology, then local folding, and finally atomic refinement, to improve structural rationality. Each stage has a different loss function and weights, and the generated results are gradually refined by dynamically adjusting the weights.

[0082] In this embodiment of the invention, during the training and generation process, the model also introduces a dynamic alignment strategy to ensure frame-to-frame consistency between the generated structure and the real structure in space. This mechanism employs different matching criteria at different stages, such as backbone RMSD, torsion angle consistency, or local contact graph similarity, to provide multi-dimensional structural feedback and guide model convergence.

[0083] In this embodiment of the invention, by training the noise addition and denoising process, geometric modeling, physical energy guidance, structural hierarchical control and knowledge guidance strategies are integrated, so that the backbone generation network can take into account geometric modeling, physical energy guidance, structural hierarchical control and knowledge guidance strategies when performing further generation prediction (equivalent to supplementary checking) of protein structures generated by non-in-situ generation networks.

[0084] In this embodiment of the invention, during the skeleton generation modeling process, an improved method based on a diffusion generation model (e.g., FrameDiff) is proposed, which introduces geometric constraint injection and multi-scale generation control to improve the structural rationality and local-global consistency of the generated sequence.

[0085] FrameDiff, as an efficient temporal diffusion generation model, excels at modeling the dynamic evolution process between sequence frames. However, when handling structure-sensitive tasks (such as protein, 3D skeleton, or other geometric data generation), the original FrameDiff lacks explicit geometric prior embeddings, which can easily lead to spatial distortion or physical inconsistencies in the generated results. To address this issue, this invention introduces a geometric constraint injection mechanism on top of the diffusion generation model. Explicit geometric information (such as distance matrices, angular constraints, topological graph structures, etc.) is incorporated into each stage of the diffusion process and injected into the skeleton generation network as guiding signals through a geometry-aware conditional encoder. This mechanism effectively captures key geometric relationships in the target structure and maintains local and global spatial consistency during generation, ensuring that the output results maintain biological or physical plausibility in terms of geometric structure.

[0086] Meanwhile, this invention proposes a multi-scale generation control strategy, dividing the generation process into three scale stages: coarse-scale, medium-scale, and fine-scale, corresponding to global framework construction, local detail completion, and micro-level adjustment, respectively. At each scale, the skeleton generation network independently models based on the geometric information and contextual conditions of the corresponding resolution, and ensures information flow and structural consistency through a cross-scale fusion module. This multi-scale design allows the model to gradually refine from macroscopic structure to microscopic precision, effectively mitigating the problems of overfitting, ambiguity, or loss of key geometric features present in single-scale modeling.

[0087] In this embodiment of the invention, to improve the accuracy and diversity of molecular conformation generation and protein structure modeling, a conformation library guidance and dynamic matching strategy are proposed in the scaffold generation network. This method integrates prior structural knowledge with the temporal modeling capabilities of the diffusion model, introducing high-quality conformation samples as references during the generation process to guide the results towards a physically plausible and biologically reliable conformation space. FrameDiff excels at modeling the dynamic evolution of structures, but in structure-sensitive tasks, purely data-driven denoising can easily lead to generated structures deviating from natural energy troughs, resulting in local anomalies or global breaks. To address this issue, this embodiment constructs a diverse conformation library and dynamically selects the conformation most consistent with the current state as a guiding signal through similarity matching during the diffusion process, improving generation stability and reliability. Furthermore, this embodiment designs a dynamic matching strategy that adaptively adjusts the guidance method according to the diffusion stage: emphasizing global framework construction in high-noise stages and focusing on local detail repair in low-noise stages, enhancing the consistency between local and global aspects. The matching strength and guidance weight are dynamically adjusted based on the similarity between the current structure and prior conformations, balancing the model's exploratory capabilities and physical plausibility. Experiments have shown that this strategy significantly improves the accuracy, stability, and diversity of structure generation, providing a more efficient and controllable generation scheme for tasks such as conformation sampling, protein folding prediction, and molecular modeling.

[0088] Based on this, in this embodiment of the invention, a theoretically complete and practically efficient protein backbone generation network is constructed by integrating geometric modeling, physical energy guidance, structural hierarchical control, and knowledge guidance strategies. This network not only improves the accuracy and rationality of the generated structures, but also achieves strong controllability and interpretability while maintaining SE(3) and other denaturation properties, providing a powerful generative foundation for protein structure design and functional prediction.

[0089] In this embodiment of the invention, in the feature extraction part, for the protein structure generation task, since the number of atoms in each sample may be different, a mask is used to ignore invalid atoms, and the loss is calculated only for valid atoms. Therefore, a masked version of RMSE is used as the loss function. Let... For the first i The first sample j The actual coordinates of each atom (three-dimensional vector); The generated coordinates; This is a mask, where 1 indicates valid and 0 indicates invalid. The masked version of RMSE is defined as follows: ;in, N It is the sample batch size. A This represents the maximum number of atoms. The numerator is the sum of the squared errors of all effective atoms, and the denominator is the total number of effective atoms.

[0090] In this embodiment of the invention, in order to improve the accuracy of the generation task in the non-in-situ generation part, the CVAE loss function consists of two parts: reconstruction loss and KL divergence loss.

[0091] Reconstruction Loss is used to measure the three-dimensional coordinates of the generated protein. with actual coordinates x i The difference between the two is commonly represented by mean squared error (MSE) or root mean square error (RMSE). The goal of reconstruction loss is to ensure that the generated structure closely approximates the real structure. The reconstruction loss is defined as: Where N is the number of samples, Coordinates generated for the model, These are the actual coordinates.

[0092] KL divergence loss is used to constrain the latent spatial distribution of the encoder output of an in-situ generated network. Approximately standard normal distribution p(z)=N(0,I) ,Right now ,in, Here are the encoder parameters, and z is a latent variable. c Represents sequence condition information. x This represents the three-dimensional coordinates of the protein. The KL divergence loss guarantees the regularity of the latent spatial distribution. Its analytical expression is: ; in, d For which is the potential spatial dimension, , denoted as the mean and standard deviation of the k-th dimension, respectively.

[0093] In the backbone generation section, to address SE(3) diffusion modeling and the physical plausibility of protein structures, several auxiliary losses were designed, including Translation & Rotation Score Loss, Backbone Atom Loss, Pairwise Distance Matrix Loss, Energy Guidance Loss, Multiscale Loss, Fragment Guidance Loss, and others. By rationally designing and weighting these losses, the model can generate diverse and physically plausible protein backbone structures, meeting the needs of scientific research and engineering applications.

[0094] The core of the SE(3) diffusion model is to model the spatial position (translation) and orientation (rotation) of each residue in the protein backbone. The goal of training the diffusion model is to enable the neural network to predict the "denoising direction" (score function), that is, how to adjust the current structure to be closer to the true distribution in each step of back diffusion. For this purpose, a translation and rotation score loss was designed.

[0095] Translation fractional loss: Let This is the true translation fraction (usually the gradient of the true distribution after adding noise). The loss is the score predicted by the model: ; in is the fraction scaling factor, and N is the number of residues.

[0096] Rotation fractional loss: Similar to translation fractional loss, let... This represents the true rotation fraction (usually the gradient of the true distribution after adding noise). The scores predicted by the model are typically parameterized using Lie algebras, with the loss being: ; The rotation fractions here are usually vectors on the SO(3) Lie algebra.

[0097] The protein backbone consists of N, Composed of atoms such as C and O, the spatial positions of the backbone atoms determine the basic framework of a protein. The model needs to ensure that the generated backbone atom coordinates are as consistent as possible with the actual structure. Therefore, a backbone atom reconstruction loss was designed. Specific details are as follows: set up For the first i The residue of the first a The actual coordinates of each main chain atom The loss is used to predict the coordinates: ; In practice, the loss is typically calculated only for valid atoms with a mask of 1. This loss helps the model reconstruct the main chain geometry at both global and local scales. It is often used in conjunction with fractional loss to improve the physical plausibility of the structure.

[0098] The spatial rationality of protein structures is reflected not only in the coordinates of individual atoms but also in the distance relationships between atomic pairs. Therefore, a pairing distance matrix loss is designed to constrain the local spatial relationships of the generated structures. This loss primarily focuses on local structures, preventing unreasonable distortions or breaks in the generated structures. Effective atomic pairs are selected through a mask and a distance threshold, thereby improving training efficiency.

[0099] The specific details are as follows: set up and The actual and predicted numbers are respectively i, j The distance between main chain atom pairs, the loss is: ; in I The indicator function applies the loss only to atom pairs with a distance of less than 6 Å, where M is the effective number of atom pairs.

[0100] The physical plausibility of protein structures can be measured by energy functions (such as Rosetta energy). Energy-guided loss introduces an energy gradient during the sampling / generation process, encouraging the generation of structures with lower energies. Energy guidance can significantly improve the physical feasibility of generated structures, and by adjusting the weight λ, it can be prevented from being too strong, which would lead to a decrease in diversity. Specific details are as follows: set up E(x) Let λ be the energy function, λ be the weight, and the loss be: ; In practice, the energy gradient is added to the score function of the diffusion model to correct the sampling trajectory.

[0101] Protein structure generation is a coarse-to-fine process. Multi-scale loss divides the generation process into stages such as global topology, local folding, and atomic refinement, with each stage focusing on different structural levels. Specific details are as follows: The total loss is the weighted sum of the losses at each stage: ; , and These represent the global topology loss, local folding loss, and atomic refinement loss, respectively. , and These represent the weights of the global topology loss, local folding loss, and atomic refinement loss, respectively.

[0102] By dynamically adjusting weights, coarse-to-fine optimization is achieved, improving the final structural quality. Multi-scale loss helps the model first learn the global skeleton and then refine it to the atomic level.

[0103] Common local conformations in protein structures can be obtained through fragment library searches. Fragment library-guided loss encourages the generation of structures that are locally similar to known fragments. Specific details are as follows: set up To generate fragments, For the retrieved fragment, the loss is: ; The fragment library guides the generation of more diverse and locally reasonable structures. Parameters such as retrieval frequency, fragment length, and weight need to be adjusted according to the actual task.

[0104] In this embodiment of the invention, a multimodal loss function is introduced during training, including reconstruction error, KL divergence, and function preservation loss, to ensure that the generated protein is highly consistent with the input conditions at both the structural and functional levels. This non-in-situ generation framework not only possesses strong generalization ability, capable of adapting to unseen functional and structural combinations, but also exhibits good interpretability and practical application potential, making it suitable for tasks such as protein engineering, functional prediction, and novel protein design.

[0105] To accelerate the optimization process and improve the model's generalization ability, the Adam optimizer is used in this embodiment of the invention. Adam (Adaptive Moment Estimation) is a gradient-based optimization algorithm that combines the advantages of momentum and adaptive learning rate, making it suitable for optimization problems with large-scale data and high-dimensional parameter spaces. The Adam optimizer dynamically adjusts the learning rate of each parameter based on historical gradient information, effectively avoiding the vanishing and exploding gradient problems.

[0106] By combining a loss function and the Adam optimizer, the model can automatically adjust parameters during training to minimize prediction error and continuously improve prediction accuracy. This optimization strategy has demonstrated good performance in the task of non-in situ protein generation, effectively accelerating training convergence and robustly approximating the optimal parameter set, thus achieving good prediction results.

[0107] In this embodiment of the invention, modular modeling, geometry guidance, and multi-model collaboration significantly improve modeling efficiency, generation accuracy, and system scalability, demonstrating significant advantages in several key dimensions: (1) Non-in-situ generation achieves decoupling of structure and sequence: With the help of conditional variational autoencoder (CVAE), structure and sequence are flexibly mapped in the latent space, making structure-driven sequence generation more flexible and controllable, and improving the generalization ability and combinatorial diversity of the model.

[0108] (2) Geometric isovariant diffusion enhances the physical rationality of the skeleton: The skeleton generation network of the non-in-situ protein generation prediction framework has SE(3) isovariant properties, which ensures that the skeleton generation process is geometrically consistent and physically feasible in three-dimensional space, laying the foundation for subsequent structural stability and folding correctness.

[0109] (3) Modular design enhances system flexibility and interpretability: The system is clearly decoupled at each stage, which facilitates optimization and replacement, and can be seamlessly connected to sequence design tools (e.g., ProteinMPNN) and structure prediction tools (e.g., AlphaFold2), supporting multi-task expansion such as function scoring and target design.

[0110] (4) Multi-module collaboration enhances overall robustness and adaptability: Efficient information flow is achieved through geometric representation and feature interface, which alleviates the accumulation of local errors. Combined with joint training strategy, it shows good stability and transferability in multi-task environment.

[0111] Based on the same inventive concept, this invention also provides a protein functional topology design device based on SE(3) isovariant diffusion and non-in situ generation. The device is used to implement the protein functional topology design method based on SE(3) isovariant diffusion and non-in situ generation described in any of the above embodiments. The device includes: The feature extraction module is used to input the protein sequence information corresponding to the functional fragment of the target protein into the feature extraction network of the non-in situ generated protein prediction framework; the feature extraction network processes the protein sequence information to obtain the corresponding expression vector. The in-situ generation module is used to input the expression vector into the in-situ generation protein prediction framework's in-situ generation network. The in-situ generation network generates three-dimensional coordinates of a protein with a function similar to the target protein's functional fragment based on the expression vector. The in-situ generation network uses the three-dimensional coordinates of the sample protein as input features and the protein sequence information of the sample protein as conditional information, and learns the probability distribution of the protein's three-dimensional coordinates under specific protein sequence information. The backbone generation module is used to convert the three-dimensional coordinates of the protein into data in a preset format. The preset format data representing the three-dimensional coordinate information of the protein is input into the backbone generation network for generating a protein prediction framework in situ. The backbone generation network generates and predicts the preset format data and obtains a protein backbone with the target function based on geometric modeling, physical energy guidance, structural hierarchical control and knowledge guidance strategies. An optimization module is used to perform sequence design and structural modeling based on the protein backbone to obtain a protein with the target function.

[0112] The embodiment of the protein functional topology design device based on SE(3) isovariant diffusion and non-in-situ generation provided by this invention can be applied to any device with data processing capabilities, such as a computer. The device embodiment can be implemented through software, hardware, or a combination of both. Taking software implementation as an example, as a logical device, it is formed by the processor of any data processing device reading the corresponding computer program instructions from non-volatile memory into memory and running them. From a hardware perspective, such as... Figure 3 The diagram shown is a hardware structure diagram of any data processing-capable device that houses the protein functional topology design device based on SE(3) isovariant diffusion and non-in-situ generation provided by this invention. Except for... Figure 3 In addition to the processor, memory, network interface, and non-volatile memory shown, any data processing device in the embodiment may also include other hardware depending on the actual function of the data processing device, which will not be described in detail here.

[0113] Based on the same inventive concept, embodiments of the present invention also provide an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the steps in the protein functional topology design method based on SE(3) isovariant diffusion and non-in-situ generation as described in any of the above embodiments.

[0114] Based on the same inventive concept, embodiments of the present invention also provide a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps in the protein functional topology design method based on SE(3) isovariant diffusion and non-in-situ generation described in any of the above embodiments.

[0115] Based on the same inventive concept, this invention provides a computer program product, including a computer program / instruction, which, when executed by a processor, implements the steps in the protein functional topology design method based on SE(3) isovariant diffusion and non-in situ generation described in any of the above embodiments.

[0116] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0117] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, apparatus, or computer program products. Therefore, embodiments of the present invention can take the form of entirely hardware embodiments, entirely software embodiments, or embodiments combining software and hardware aspects. Furthermore, embodiments of the present invention can take the form of computer program products implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0118] Embodiments of the present invention are described with reference to flowchart illustrations and / or block diagrams of methods, terminal devices (apparatus), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable terminal device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable terminal device, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0119] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable terminal device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0120] These computer program instructions can also be loaded onto a computer or other programmable terminal device to cause a series of operational steps to be performed on the computer or other programmable terminal device to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable terminal device for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0121] Although preferred embodiments of the present invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the present invention.

[0122] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or terminal device. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or terminal device that includes said element.

[0123] The above provides a detailed description of a protein functional topology design method based on SE(3) isovariant diffusion and non-in situ generation provided by the present invention. Specific examples are used in this article to illustrate the principle and implementation of the present invention. The above description of the embodiments is only for the purpose of helping to understand the method and core idea of ​​the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation and application scope based on the idea of ​​the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A protein functional topology design method based on SE(3) isovariant diffusion and ex-situ generation, characterized in that, The method includes: The protein sequence information corresponding to the functional fragment of the target protein is input into the feature extraction network of the non-in situ generated protein prediction framework; the feature extraction network processes the protein sequence information to obtain the corresponding expression vector. The expression vector is input into the non-in situ generation network of the non-in situ protein prediction framework. The non-in situ generation network generates three-dimensional coordinates of a protein with a function similar to the functional fragment of the target protein based on the expression vector. The non-in situ generation network uses the three-dimensional coordinates of the sample protein as input features and the protein sequence information of the sample protein as conditional information to learn the probability distribution of the three-dimensional coordinates of the protein under specific protein sequence information. The protein's three-dimensional coordinates are converted into data in a preset format. The preset format data representing the protein's three-dimensional coordinate information is then input into a backbone generation network for generating a protein prediction framework in situ. The backbone generation network generates and predicts the preset format data. Based on geometric modeling, physical energy guidance, structural hierarchical control, and knowledge guidance strategies, a protein backbone with the target function is obtained. Based on the protein backbone, sequence design and structural modeling are performed to obtain a protein with the target function.

2. The protein functional topology design method based on SE(3) isovariant diffusion and ex-situ generation according to claim 1, characterized in that, The feature extraction network processes protein sequence information to obtain corresponding expression vectors, including: The protein sequence is positionally encoded using a positional encoding method. For the first position in the protein sequence... pos The position and the first i The positional encoding method for the dimensional feature is as follows: ;in, Dimensions representing features; The protein sequence information is encoded using the Doc2Vec model to obtain a global vector representation of the protein sequence; The output vector of the Doc2Vec model is transformed using a linear mapping layer to obtain the embedding vector. The embedding vector is input into the Transformer encoder to capture the potential structural relationships and biological function correlations within the protein sequence, thereby obtaining the expression vector corresponding to the protein sequence information.

3. The protein functional topology design method based on SE(3) isovariant diffusion and ex-situ generation according to claim 1 or 2, characterized in that, The non-in-situ protein generation prediction framework was trained using the following steps: The protein sequence information of the sample protein is input into a feature extraction network, which processes the protein sequence information to obtain the corresponding sample expression vector. The three-dimensional coordinates of the sample protein are used as input features, and the sample expression vector of the sample protein is used as conditional information to input into the off-site generation network. The off-site generation network learns the probability distribution of the three-dimensional coordinates of the protein under specific protein sequence information and generates the prediction result of the three-dimensional coordinates of the sample protein. The predicted three-dimensional coordinates of the sample protein are converted into sample preset format data. The sample preset format data is then input into the backbone generation network. The backbone generation network adds noise and removes noise to the sample preset format data based on the diffusion probability model to approximate the real protein backbone of the sample protein. During the training process, each step maintains the equivariance of the SE(3) transformation. An external energy function is introduced as a physical constraint to control the generation process in stages. First, global topology is performed, then local folding is performed, and finally atomic refinement is performed to improve the rationality of the structure. A dynamic alignment strategy is introduced to ensure that the generated protein backbone and the real protein backbone are consistent in space frame by frame.

4. The protein functional topology design method based on SE(3) isovariant diffusion and ex-situ generation according to claim 3, characterized in that, The three-dimensional coordinates of the sample protein are used as input features, and the sample expression vector of the sample protein is used as conditional information. These are then input into an in-situ generation network. The in-situ generation network learns the probability distribution of the protein's three-dimensional coordinates under specific protein sequence information, including: The three-dimensional coordinates of the sample protein are used as input features, and the sample expression vector of the sample protein is used as sequence conditional information. The encoder of the non-in situ generation network is input. The encoder obtains combined features through splicing operation, performs nonlinear mapping, and obtains the parameters of the latent space: mean and standard deviation. The process of sampling latent variables z from the latent space based on reparameterized sampling makes the latent variables z a differentiable function of the mean and standard deviation; The decoder of the non-in-situ generation network takes the latent variable z and sequence condition information as common inputs to generate corresponding protein 3D coordinate predictions, learns the differences in protein 3D coordinates under different sequence condition information, and realizes protein structure modeling with sequence conditions.

5. The protein functional topology design method based on SE(3) isovariant diffusion and ex-situ generation according to claim 3, characterized in that, The backbone generation network uses a diffusion probability model to add and remove noise from the sample's preset format data to approximate the true protein backbone of the sample protein, including: During training, the noise-adding process of the skeleton generation network adds noise to the real data. By gradually adding Gaussian noise, a series of intermediate states are obtained. Ultimately, it becomes pure noise: ; in It is the noise intensity at each step; The denoising process trains the skeleton generation network to learn how to denoise from noise. By gradually removing noise, the data can be restored. : ; in, and Obtained by neural network prediction; During the noise addition process, the backbone generation network adds noise to the protein backbone structure, and maintains the equivariance to the SE(3) transformation at each step. During the denoising process, the backbone generation network predicts the denoising direction of each step through the equivariance of the SE(3) transformation and gradually restores the backbone structure. During training, an external energy function is introduced as a physical constraint to guide the generation of the skeleton structure: ; in It is the gradient of the energy function with respect to the structure, where λ is the weight; During training, the generation process is controlled in stages; each stage has a different loss function and weights, and the generated skeleton structure is gradually refined by dynamically adjusting the weights. During training, a dynamic alignment strategy is introduced, employing different matching criteria at different stages to provide multi-dimensional structural feedback and guide the skeleton generation network to converge.

6. The protein functional topology design method based on SE(3) isovariant diffusion and ex-situ generation according to claim 4, characterized in that, The following loss function is used to train the non-in situ protein generation prediction framework: For feature extraction networks: Let For the first i The first sample j The actual coordinates of each atom; The first generation of protein prediction framework i The first sample j The coordinates of each atom; This is a mask, where 1 indicates valid and 0 indicates invalid. Using a masked version of RMSE as the loss function: ;in, N It is the sample batch size. A This represents the maximum number of atoms, with the numerator being the sum of the squared errors of all effective atoms and the denominator being the total number of effective atoms. For non-in-situ generated networks, the loss function consists of reconstruction loss and KL divergence loss: The reconstruction loss is used to measure the 3D coordinates of proteins generated by the non-in-situ generation network. and the actual three-dimensional coordinates of proteins x i Differences between them: ; Latent spatial distribution of the output of the non-in-situ generated network encoder constrained by KL divergence loss. Approximately standard normal distribution p(z)=N(0,I) ,in, Here are the encoder parameters, and z is a latent variable. c Represents sequence condition information. x Represents the three-dimensional coordinates of a protein; ; μ This represents the mean. σ Let represent the standard deviation. The KL divergence loss guarantees the regularity of the latent spatial distribution; its analytical expression is: ; For backbone generation networks, the loss functions include translation and rotation fractional loss, main chain atom reconstruction loss, pairing distance matrix loss, energy guidance loss, multi-scale loss, and fragment library guidance loss. Translation fractional loss: set up For true translation fractions, The loss function is: (The score predicted by the model is given by the following formula.) ; in It is a fraction scaling factor. N The number of residues; Rotational fractional loss: Let For the true rotation fraction, The scores predicted by the model are typically parameterized using Lie algebras, with the loss function being: ; Main chain atom reconstruction loss: set up For the first i The residue of the first a The true coordinates of each main chain atom The loss is used to predict the coordinates: ; Pairing distance matrix loss: Let and The actual and predicted numbers are respectively i, j The distance between main chain atom pairs, the loss is: ; in I The indicator function applies the loss only to atom pairs with a distance of less than 6 Å, where M is the effective number of atom pairs; Energy guiding loss: Assume E(x) Let λ be the energy function, λ be the weight, and the loss be: ; Multi-scale loss: Multi-scale loss divides the generation process into a global topology stage, a local folding stage, and an atomic refinement stage. Each stage focuses on a different structural level, and the multi-scale loss is a weighted sum of the losses from each stage. ; , and These represent the global topology loss, local folding loss, and atomic refinement loss, respectively. , and These represent the weights of the global topology loss, local folding loss, and atomic refinement loss, respectively. Fragment library guidance loss: Let To generate fragments, For the retrieved fragment, the loss is: 。 7. A protein functional topology design device based on SE(3) isovariant diffusion and ex-situ generation, characterized in that, The device includes: The feature extraction module is used to input the protein sequence information corresponding to the functional fragment of the target protein into the feature extraction network of the non-in situ generated protein prediction framework; the feature extraction network processes the protein sequence information to obtain the corresponding expression vector. The in-situ generation module is used to input the expression vector into the in-situ generation protein prediction framework's in-situ generation network. The in-situ generation network generates three-dimensional coordinates of a protein with a function similar to the target protein's functional fragment based on the expression vector. The in-situ generation network uses the three-dimensional coordinates of the sample protein as input features and the protein sequence information of the sample protein as conditional information, and learns the probability distribution of the protein's three-dimensional coordinates under specific protein sequence information. The backbone generation module is used to input the three-dimensional coordinates of the protein into the backbone generation network of the non-in-situ protein prediction framework. The backbone generation network generates and predicts the three-dimensional coordinates of the protein and obtains a protein backbone with the target function based on geometric modeling, physical energy guidance, structural hierarchical control and knowledge guidance strategies. An optimization module is used to perform sequence design and structural modeling based on the protein backbone to obtain a protein with the target function.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the protein functional topology design method based on SE(3) isovariant diffusion and non-in-situ generation as described in any one of claims 1-6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the protein functional topology design method based on SE(3) isovariant diffusion and non-in-situ generation as described in any one of claims 1-6.

10. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instruction is executed by the processor, it implements the steps in the protein functional topology design method based on SE(3) isovariant diffusion and in situ generation as described in any one of claims 1-6.

Citation Information

Cited By

  • Drug target association identification method based on spatial atlas and isovariant network

    CN122157758A