Protein sequence-structure co-generation method, system, device, and storage medium

Abstract

The application provides a protein sequence-structure collaborative generation method, system, device and storage medium, and relates to the technical field of bioinformatics. The generation method comprises the following steps: obtaining multiple sequence alignment data and encoding to obtain an explicit evolutionary prior representation; initializing a current sequence state and a current structure state; performing at least one round of collaborative iterative generation process based on the explicit evolutionary prior representation; generating a final three-dimensional structure and outputting a protein sequence and the final three-dimensional structure. The method provided by the application directly guides the generation trajectory by extracting multiple sequence alignment data as an explicit evolutionary prior, ensuring the natural feasibility of the molecule. At the same time, the sequence and structure state are alternately interwoven and collaboratively iteratively updated, breaking the serial fragmentation limitation and realizing the bidirectional intercommunication of spatial conformation and amino acid prediction. This makes the final sequence and structure highly unified, significantly improving the protein function hit rate and experimental success rate.

Description

Technical Field

[0001] This application relates to the field of bioinformatics technology, and in particular to a method, system, device and storage medium for the co-generation of protein sequence-structure. Background Technology

[0002] Protein design is a fundamental supporting technology in the life sciences and biomanufacturing fields, with wide applications in important areas such as drug development, industrial catalysis, biosensing, and cell therapy. The three-dimensional spatial structure of a protein determines its ultimate biological function, and this structure is formed by the folding of its underlying amino acid sequence. Therefore, how to efficiently and accurately design novel protein molecules with specific structures and expected functions using computational methods has always been a core and fundamental research topic in computational biology and protein engineering.

[0003] Currently, in the field of protein computational design, the mainstream techniques mainly cover the following conventional paradigms. The first is the two-stage structure-sequence decoupling design paradigm, which typically generates the protein backbone using diffusion models and then predicts the corresponding amino acid sequence using a reverse-folding model. The second is the pure sequence generation paradigm, which usually uses language models or masking models as its core architecture, directly sampling in the amino acid sequence space. The third is the structure prediction-driven illusion optimization paradigm, which mainly uses the scoring results of structure prediction networks as the objective function to iteratively optimize the initial sequence or structure in order to obtain the target molecule.

[0004] However, the aforementioned existing technologies all exhibit significant limitations in practical engineering applications. In the two-stage paradigm, backbone generation and sequence prediction are serially decoupled optimizations, with information transmission being unidirectional, making it difficult to guarantee that the final functional sites, global conformation, and underlying sequence constraints are uniformly satisfied. The pure sequence generation paradigm severely underutilizes the three-dimensional structural context and spatial long-range coupling constraints, causing the generated sequences to frequently deviate from expectations during actual folding. The illusion optimization paradigm heavily relies on complex objective function parameter tuning, and the migration cost between different design targets is extremely high. More importantly, natural proteins have accumulated rich structural stability and functional constraints (such as co-evolutionary coupling relationships in multiple sequence alignment) during their long evolutionary process, but existing methods typically only treat these evolutionary constraints as implicit statistical signals in the training data, making it difficult to directly apply effective guidance to the generated sampling trajectory.

[0005] In summary, existing protein design techniques generally suffer from the technical flaw of a disconnect between sequence generation and structure generation. This multi-stage decoupling operation and insufficient utilization of natural evolutionary constraints make it easy for models to generate computationally foldable but functionally ineffective molecules, resulting in low success rates in downstream experimental screening. Furthermore, while exploring novel sequences, existing techniques struggle to maintain conformational consistency with natural protein families, easily leading to significant deviations from naturally feasible physical manifolds. Consequently, the overall success rate and experimental feasibility of protein design fail to meet practical needs. Summary of the Invention

[0006] In a first aspect, the present invention provides a method for co-generating protein sequence-structure, comprising: Multiple sequence alignment data of the target protein are acquired and encoded to obtain an explicit evolutionary prior representation containing monomeric site features and site pair features. Initialize the current sequence state and current structural state of the target protein; wherein the current sequence state includes masked sites; Based on the explicit evolutionary prior representation, at least one round of cooperative iterative generation process is performed to alternately interleave and update the current sequence state and the current structural state until all masked sites in the current sequence state are demasked to obtain the final updated current sequence state, and the final updated current sequence state is used as the protein sequence of the target protein. The final three-dimensional structure is generated based on the protein sequence, and the protein sequence and the final three-dimensional structure are output.

[0007] In an optional implementation, during each round of the cooperative iterative generation process, the alternating interleaving update includes: Based on the explicit evolutionary prior representation, the current sequence state, and the current structural state, the current round structural state is generated; Based on the current sequence state, the explicit evolutionary prior representation, and the current sequence state, predict the amino acid probability distribution corresponding to the sites that are still masked in the current sequence state; Based on the current structural state, determine the structural neighborhood information corresponding to the still-masked sites, and based on the structural neighborhood information, determine the set of demasked sites for this round from the still-masked sites. Based on the amino acid probability distribution, the residue type of each point in the current round of demasking site set is determined, and the current sequence state is updated using the determined residue type, and the current structural state is updated using the current round structural state.

[0008] In an optional implementation, the step of determining the structural neighborhood information corresponding to the still-masked sites based on the current round's structural state, and determining the set of demasked sites for the current round from the still-masked sites based on the structural neighborhood information, includes: For each site i in the set I of sites that are still masked, calculate the number N of its spatial neighbors. i The calculation formula for the structural neighborhood information is as follows: ; Among them, Cα i and Cα j These are the three-dimensional coordinates of the Cα atoms at sites i and j, respectively. The preset spatial neighborhood threshold; For all sites in set I, according to the spatial neighborhood number N i Sort in descending order and select the top r results. t ×N sites form the current round's unmasking site set; where N is the total amino acid length of the target protein; r t This is the preset demasking ratio for the current round.

[0009] In an optional implementation, the spatial neighborhood threshold d thr The value range is 8 Å to 12 Å; and / or, The single-round demasking ratio r t The value range is 0.01 to 0.20.

[0010] In an optional implementation, determining the residue type of each point in the current round of demasking site set based on the amino acid probability distribution includes: Based on the amino acid probability distribution, temperature sampling is performed on each site in the current round of demasking site set using a preset sequence sampling temperature τ to determine the residue type of each site.

[0011] In an optional implementation, the value of the sequence sampling temperature τ ranges from 0.6 to 1.0.

[0012] In an optional implementation, the total number of iterations in the collaborative iterative generation process is 6 to 12 rounds; and / or, In each round of the collaborative iterative generation process, the number of inversion steps for structure diffusion generation is 100 to 300 steps.

[0013] In an optional implementation, when initializing the current sequence state and current structural state of the target protein: For the masked positions in the current sequence state, only N and C are retained in the current structural state. αC and O main chain atoms; for non-masked sites, all atoms are retained in the current structural state.

[0014] In an optional implementation, before encoding the multi-sequence alignment data, the multi-sequence alignment data is further cleaned; the cleaning includes: removing homologous sequences that are repeated with the query sequence, and removing sequences with a gap ratio higher than 0.6, so that the effective depth of the cleaned multi-sequence alignment matrix is ​​not less than 32.

[0015] In an optional implementation, the explicit evolutionary prior representation including single-site features and site-pair features includes: The single-site feature is the dimensionality-reduced single-site feature tensor S. msa Its feature dimension d s =256; The site pairs are characterized by the reduced-dimensional pairwise feature tensor Z. msa Its feature dimension d z =128.

[0016] In an optional implementation, for the target protein with a total amino acid length of N, the monomeric feature tensor is: The i-th position is represented as: ; The paired feature tensors: The i-th row and j-th column is represented as: ; Where M is the number of homologous sequences; E m,i w is the encoding vector of the m-th homologous sequence at the i-th position; m For sequence weights; W z It is a linear projection matrix; This represents the tensor outer product.

[0017] In an optional implementation, during each round of the cooperative iterative generation process, the individual feature tensor S msa and the paired feature tensor Z msa Injected into the structural backbone network, and related to the current single-unit representation H of the structural backbone network. l And the current pairwise representation P l Perform a merge update; the merge formula is: ; Where Ws and Wp are linear projection matrices; PairFormer is the feature fusion and update module; H l This represents the current singleton; P l For the current paired representation; H l+1is the updated single-unit representation; l represents the hierarchical index of the backbone network of the structure.

[0018] In an optional implementation, the current pair represents P. l The paired feature elements at the i-th and j-th sites are represented by P. ij The paired feature elements P ij By fusing the gating mechanism with the geometric distance embedding of the current diffusion step, conditional input elements for the structure denoising network are obtained. The fusion formula is: ; in, ; For the three-dimensional coordinates of the atom; The paired features in the i-th row and j-th column of the array.

[0019] In an optional implementation, the protein sequence-structure co-generation method supports a variety of different task modes, including at least one of monomer design mode, scaffold design mode, conjugate design mode, and specific functional modification mode.

[0020] In an optional implementation, when the task mode is a composite design mode, when determining the current round of demasking site set, the priority of selecting interface neighborhood sites at the composite interaction interface into the current round of demasking site set is increased.

[0021] Secondly, the present invention provides a training method for a protein sequence-structure co-generation model, used to train the model used in the protein sequence-structure co-generation method described in any of the foregoing embodiments, the training method comprising: Obtain homologous sequence multiple sequence alignment data of the training protein; The sequence of the training protein used as the query sequence is removed from the homologous sequence multiple sequence alignment data, and the removed multiple sequence alignment data is synchronized and masked according to a preset mask ratio to construct training input features containing UNK tags. During training, the parameters of the structure generation module in the model are kept frozen, and joint training and optimization are performed only on the sequence module and encoder in the model. During forward computation, the sequence module outputs the amino acid category probability distribution for the mask sites in the multiple sequence alignment data, and uses the back feedback of the amino acid category probability distribution to jointly update the network parameters of the sequence module and the encoder through the cross-entropy loss function.

[0022] In an optional implementation, in the leakage prevention process, the preset mask ratio p is obtained by random sampling in a uniform distribution between 0.1 and 1.0, satisfying p ~ Uniform(0.1, 1.0).

[0023] In an optional implementation, the training method employs a main-chain and side-chain joint learning mechanism. When masking the sites in the training input features, the corresponding amino acid side chains at the structural ends are simultaneously removed, retaining only N and C. α The C and O main chain atoms are used to enhance complementary inference learning of structure and sequence.

[0024] Thirdly, the present invention provides a protein sequence-structure co-generation device, comprising: The data encoding module is used to acquire multiple sequence alignment data of the target protein and encode the multiple sequence alignment data to obtain an explicit evolutionary prior representation containing monomeric site features and site pair features. An initialization module is used to initialize the current sequence state and current structural state of the target protein; wherein the current sequence state includes masked sites; The cooperative iteration module is used to perform at least one round of cooperative iteration generation process based on the explicit evolutionary prior representation, so as to alternately interleave and update the current sequence state and the current structural state until all masked sites in the current sequence state are demasked to obtain the final updated current sequence state, and use the final updated current sequence state as the protein sequence of the target protein. The structure generation module is used to generate the final three-dimensional structure based on the protein sequence and output the protein sequence and the final three-dimensional structure.

[0025] Fourthly, the present invention provides a computer device comprising a processor and a memory, the memory storing a computer program, and the processor executing the computer program to implement the protein sequence-structure co-generation method described in any of the foregoing embodiments.

[0026] Fifthly, the present invention provides a computer storage medium storing a computer program, which, when executed on a processor, implements the protein sequence-structure co-generation method according to any one of the foregoing embodiments.

[0027] The method provided in this application extracts multiple sequence alignment data of the target protein and performs deep encoding to construct an explicit evolutionary prior representation containing monomer and site pair features. This operation overcomes the limitations of previous methods that merely treated natural evolutionary constraints as implicit statistical signals, and can directly inject conservation and co-evolutionary coupling laws, honed through a long process of natural evolution, into every stage of the generation trajectory. This greatly constrains the spatial path of molecule generation, ensuring that the generated protein molecules possess both high novelty and strict adherence to the physical manifold of the natural protein family, effectively avoiding deviations from the natural feasible space.

[0028] Using the aforementioned explicit prior representation, a cooperative iterative generation process is performed based on an initial state containing masked sites. During this process, the sequence state and structural state are dynamically updated frequently through alternating interleaving, breaking the traditional unidirectional serial processing of the backbone and sequence. This bidirectional communication mechanism ensures that the microscopic three-dimensional structural context can guide the demasking prediction of amino acid residues at specific functional sites in real time, while the updated amino acid residue information can immediately feed back into the adjustment of the global main chain conformation.

[0029] Through continuous mutual constraint iterations until all sites are demasked, the final output protein sequence and its corresponding three-dimensional structure achieve a high degree of intrinsic unity in terms of geometric rationality and sequence dependence. This effectively overcomes the technical bottleneck of easily generating molecules that "appear foldable but actually lack function" based solely on local features, significantly improving the conformational consistency, functional hit rate, and feasibility of protein design in downstream experiments. Attached Figure Description

[0030] To more clearly illustrate the technical solutions of this application, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of this application and therefore should not be considered as a limitation on the scope of protection of this application. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0031] Figure 1 This is a schematic diagram of the hardware operating environment involved in an embodiment of the protein sequence-structure co-generation method of the present invention; Figure 2 This is a schematic flowchart of Example 1 of the protein sequence-structure co-generation method of the present invention; Figure 3 This is a schematic diagram of the macroscopic framework and microscopic model architecture of the protein sequence-structure co-generation method based on evolutionary prior constraints provided in Embodiment 1 of the present invention; wherein, Figure 3 Part 'a' in the diagram is a schematic representation of the overall paradigm for collaborative iterative generation. Figure 3Part b in the diagram is a microstructure diagram of the deep neural network that performs this collaborative generation. Figure 4 This is a detailed flowchart of step S300 in Example 2 of the protein sequence-structure co-generation method of the present invention; Figure 5 This is a flowchart illustrating the training method for the protein sequence-structure co-generation model of the present invention in Example 7. Figure 6 This is a comparison chart of sequence-structure consistency evaluation in Experiment Example 1 of this application; it shows the consistency rate comparison results between the method of the present invention and various existing skeleton design and co-design models; Figure 7 This is a proximity analysis diagram between the generated results in Experiment Example 1 of this application and natural homologous sequences; it shows a comparison of distance indices in phylogenetic tree space, sequence embedding space, and structural embedding space; Figure 8 This is a sequence similarity and novelty distribution diagram of the design results in Experiment Example 1 of this application; it shows the identity of the generated protein with the target site and the natural homologous sequence and the TMScore distribution; Figure 9 This is an example diagram of the phylogenetic distribution in Experiment Example 1 of this application; taking Cytochrome C as an example, it visually demonstrates the phylogenetic distribution of the generated protein and its native homologous sequence, as well as its Hamming distance distribution; Figure 10 A performance evaluation diagram was designed for the functional site scaffold task in Experiment Example 2 of this application; the results of ablation experiments with the comparison of the number of successful targets, the correlation of evaluation indicators, and the removal of site constraints or evolutionary information were presented. Figure 11 This is a comparative diagram of representative design structures for the functional site scaffold task in Experimental Example 2 of this application; it shows the differences between the design structure of this invention, the ablation scheme design structure and the natural structure in terms of functional site retention and global folding quality. Figure 12 Performance evaluation diagram for the protein binding design task in Experiment Example 3 of this application; showing the comparison of the number of successful targets for multiple receptors, the binding free energy distribution, and the Cα RMSD trajectory in molecular dynamics simulations; Figure 13 This is a diagram showing the experimental verification results of the modified fluorescent protein in Experiment Example 4 of this application; it displays the normalized flow cytometry measurement results of the generated protein, the phylogenetic tree, and the sequence alignment of key functional sites. Figure 14 This is a histogram showing the experimental verification results of the modified fluorescent protein flow cytometer in Experiment Example 4 of this application; it illustrates the P3-gated and positive cell ratios in two independent experiments. Figure 15This is a diagram showing the experimental verification results of rare fluoride enzymes in Experimental Example 5 of this application; it illustrates the comparison of in vitro catalytic activity (5-FDA yield) of the designed fluoride enzymes and the sequence alignment results including the catalytic pocket site; Figure 16 This is a two-dimensional visualization of the fluoride enzyme sequence embedding space in Experimental Example 5 of this application; it shows the clustering and distribution of the designed sequence, known active enzyme, and natural homologous sequence in the embedding space; Figure 17 This is a schematic diagram of the module connections of the protein sequence-structure co-generation device of the present invention. Detailed Implementation

[0032] The technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments.

[0033] The components of the embodiments of this application described and illustrated in the accompanying drawings can be arranged and designed in a variety of different configurations. Therefore, the following detailed description of the embodiments of this application provided in the drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of the application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0034] In the following, the terms “comprising,” “having,” and their cognates, which may be used in various embodiments of this application, are intended only to indicate a particular feature, number, step, operation, element, component, or combination thereof, and should not be construed as excluding, firstly, the presence of one or more other features, numbers, steps, operations, elements, components, or combinations thereof, or adding the possibility of one or more features, numbers, steps, operations, elements, components, or combinations thereof.

[0035] Furthermore, the terms "first," "second," and "third" are used only to distinguish descriptions and should not be interpreted as indicating or implying relative importance.

[0036] Unless otherwise specified, all terms used herein (including technical and scientific terms) shall have the same meaning as commonly understood by one of ordinary skill in the art to which the various embodiments of this application pertain. Terms (such as those defined in commonly used dictionaries) shall be interpreted as having the same meaning as in their contextual meaning in the relevant technical field and shall not be construed as having an idealized or overly formal meaning, unless clearly defined in the various embodiments of this application.

[0037] The following detailed description of some embodiments of this application is provided in conjunction with the accompanying drawings. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0038] like Figure 1 The diagram shown is a structural schematic of the hardware operating environment of the terminal involved in an embodiment of the present invention.

[0039] The protein sequence-structure co-generation system (device) of this invention can be a PC, or a mobile terminal device such as a smartphone, tablet, or portable computer. This visual navigation optimization system may include: a processor 1001, such as a CPU; a network interface 1004; a user interface 1003; a memory 1005; and a communication bus 1002. The communication bus 1002 is used to enable communication between these components. The user interface 1003 may include a display screen, an input unit such as a keyboard, or a remote control; optionally, the user interface 1003 may also include a standard wired interface or a wireless interface. The network interface 1004 may optionally include a standard wired interface or a wireless interface (such as a Wi-Fi interface). The memory 1005 may be a high-speed RAM memory or a stable memory, such as a disk storage device. Optionally, the memory 1005 may also be a storage device independent of the aforementioned processor 1001. Optionally, the protein sequence-structure co-generation system may also include RF (Radio Frequency) circuitry, audio circuitry, a Wi-Fi module, etc. In addition, the visual navigation optimization system can also be equipped with other sensors such as gyroscopes, barometers, hygrometers, thermometers, and infrared sensors, which will not be described in detail here.

[0040] Those skilled in the art will understand that Figure 1 The system shown is not intended to limit it and may include more or fewer components than illustrated, or combine certain components, or have different component arrangements. Figure 1 As shown, the memory 1005, which is a computer-readable storage medium, may include an operating system, a data interface control program, a network connection program, and a visual navigation optimization program.

[0041] In summary, the method provided by this invention directly guides the generation trajectory by introducing explicit evolutionary priors and achieves a high degree of bidirectional unity between protein conformation and underlying sequence through the alternating interweaving and collaborative iterative updates of sequence and structural states, which significantly improves the hit rate of functional design and the success rate of experimental implementation.

[0042] Example 1 Reference Figure 2This embodiment provides a protein sequence-structure co-generation method, which aims to solve the technical problems of low functional hit rate caused by the disconnect between the backbone and sequence and insufficient utilization of natural evolutionary constraints in existing protein design technologies. By introducing multiple sequence alignment data as explicit constraints and establishing a bidirectional interleaving update mechanism for sequence and structure, a high degree of unity between protein sequence and structure is achieved. Specifically, the method includes the following steps: Step S100: Obtain multiple sequence alignment data of the target protein and encode the multiple sequence alignment data to obtain an explicit evolutionary prior representation containing monomeric site features and site pair features.

[0043] In its implementation, the system acquires multiple sequence alignment data characterizing the evolutionary history of the target protein family by searching homology databases. During natural evolution, the conservation of specific residue sites and the co-evolutionary coupling between spatial contact sites are key physical laws maintaining the protein's folding and function. This embodiment employs a feature extraction network (e.g., a deep learning encoder based on an attention mechanism) to mathematically encode the alignment data, extracting monosomal site features characterizing independent residue properties and site pair features characterizing the interaction relationships between residues. These two features together constitute an explicit evolutionary prior representation. This representation changes the previous practice of treating evolutionary information merely as implicit background, instead directly using it as a strong boundary condition and guiding signal for subsequent network generation, thereby ensuring that the generated molecules do not deviate from the feasible physical space of the native protein.

[0044] Step S200: Initialize the current sequence state and current structural state of the target protein; wherein the current sequence state includes masked sites.

[0045] To initiate the collaborative generation process, the system needs to simultaneously establish evolutionary starting points for the target protein in both sequence space and three-dimensional structural space. For the current sequence state, the system covers the amino acid positions to be designed or explored with specific mask markers, leaving these sites in an unknown state, awaiting model prediction. For the current structural state, the system assigns initial three-dimensional spatial coordinates to each residue of the target protein, thereby constructing the initial geometric conformation. The sequence mask state and the initial geometric state are established at the same stage, providing the basic data carrier for subsequent alternating coupling and transformation.

[0046] Step S300: Based on the explicit evolutionary prior representation, perform at least one round of cooperative iterative generation process to alternately update the current sequence state and the current structural state until all masked sites in the current sequence state are demasked to obtain the final updated current sequence state, and use the final updated current sequence state as the protein sequence of the target protein.

[0047] This step is the core dynamic evolution stage of the method in this embodiment. The system inputs the aforementioned explicit evolutionary prior representation into the generative network, driving the model into a cyclical iteration. In each iteration, the system breaks away from the traditional unidirectional decoupled generation method, instead allowing structure generation and sequence prediction to interact, i.e., "alternating interleaving and updating." Specifically, the instantaneous state of the three-dimensional spatial conformation can provide physical topological constraints for residue prediction, guiding the determination of amino acid types at some masked sites (i.e., demasking); while the newly determined residue types can be immediately fed back into the system, causing changes in local physicochemical potential energy, thereby constraining and adjusting the next step of spatial structure evolution. This bidirectional interaction of "you in me, me in you" continues in multiple iterations, causing the sequence and structure matching degree to continuously increase. The iteration terminates when all mask markers are replaced by real amino acid residues. The current sequence state obtained at this point, completely free of masking, is determined as the final protein sequence of the target protein.

[0048] Step S400: Generate the final three-dimensional structure based on the protein sequence, and output the protein sequence and the final three-dimensional structure.

[0049] Once the system determines the global protein sequence, it performs final conformational mapping or refinement based on this complete sequence condition, generating a strictly corresponding final three-dimensional structure (e.g., complete atomic coordinate information). Finally, the system outputs the highly consistent protein sequence and the final three-dimensional structure together. Because evolutionary priors and bidirectional interleaving constraints are maintained throughout the entire generation process, the final output sequence and structure exhibit extremely high physical fit, effectively avoiding the defect of "appearing foldable but lacking actual function," and can directly serve downstream laboratory synthesis and functional verification.

[0050] To enable those skilled in the art to more intuitively understand how sequences and structures in this invention achieve bidirectional mutual constraint evolution, please refer to [link to relevant documentation]. Figure 3 It demonstrates the macroscopic framework and microscopic model architecture of the protein sequence-structure co-generation method provided in the embodiments of this application.

[0051] like Figure 3 As shown in section a, this invention breaks the traditional serial decoupling design paradigm. The model starts with an evolutionary prior extracted from multiple sequence alignment (MSA) data. In the generation trajectory, the initial sequence state (S0) containing the mask and the initial structural state (C0) with spatial noise enter an alternating iterative loop. Within each clock step (t), the structural denoising evolution (generation C0)... t The sequence decoding and sampling (generating St) are mutually conditionalized and synchronized. After multiple iterations (until S...), T and C TThe final output sequence is highly consistent with the three-dimensional conformation of the target protein (supporting monomer or complex design).

[0052] like Figure 3 As shown in part b, at the micro-neural network architecture level, the input multi-sequence alignment data is first refined by the encoder into explicit feature tensors at the individual level and site-pair level. Subsequently, these two-dimensional features are used as physical constraint anchors and are synchronously injected into the upper structural branch and the lower sequence branch.

[0053] In the structural branch, the noisy atom coordinates are input into the diffusion module and combined with the features flowing in the feature update module to perform denoising derivation and generate the current wheel structure.

[0054] In the sequence branch, after the atomic coordinates of the wheel structure are extracted by the Atom Encoder, they interact across tracks with the MSA Module and the Pair-former to guide the accurate decoding of the underlying sequence.

[0055] Through the above architecture, the model ensures that minute changes in macroscopic spatial geometry can be fed back to the prediction of underlying amino acid categories in real time, achieving true two-way physical coupling.

[0056] In summary, the method presented in this embodiment, by encoding multiple sequence alignment data to extract explicit evolutionary prior representations, overcomes the limitation of treating evolutionary constraints merely as implicit signals, directly guiding the molecular generation trajectory and ensuring the natural feasibility and family consistency of the generated molecules. Furthermore, by performing alternating and collaborative iterative updates of sequence and structural states, the limitation of unidirectional separation between three-dimensional structure and underlying sequence is overcome, enabling bidirectional communication and real-time constraints between microscopic spatial conformation and residue prediction. This method effectively overcomes the defect of easily generating loss-of-function molecules, resulting in a high degree of geometric consistency between the final output sequence and structure, significantly improving the functional hit rate and experimental success rate of protein design.

[0057] Example 2 Based on the foregoing embodiments, this embodiment provides a protein sequence-structure co-generation method. In each round of co-iterative generation, the system does not perform sequence decoding or structure denoising in isolation. Instead, it drives the data stream to sequentially flow through four core sub-stages within the same clock cycle: structure forward deduction, sequence condition prediction, structure-aware screening, and state synchronization interlocking. The specific execution process is as follows: Reference Figure 4In each round of the aforementioned collaborative iterative generation process, the alternating interleaving update includes: Step S310: Generate the current round structure state based on the explicit evolutionary prior representation, the current sequence state, and the current structure state.

[0058] At the start of the current iteration, the system inputs the static explicit evolutionary prior representation constructed in step one, the current sequence state (including unmasked amino acid residues and unmasked mask tokens), and the current structural state into the structural denoising generation network. The structural denoising generation network utilizes the monomer conservation and site-pair co-evolutionary spatial coupling laws inherent in the explicit evolutionary prior representation, and uses the determined amino acid residues in the current sequence state as physicochemical and geometric constraint anchors. It performs conditional denoising and manifold mapping transformation on the atomic coordinates containing spatial noise in the current structural state, thereby outputting the most adaptive geometric conformation for the current design state, which is defined as the current iteration's structural state. This processing step ensures that the evolution of the 3D structure closely follows the pace of sequence updates.

[0059] Step S320: Based on the current structural state, the explicit evolutionary prior representation, and the current sequence state, predict the amino acid probability distribution corresponding to the sites that are still masked in the current sequence state.

[0060] After obtaining the current phase structure state, the system uses the instantaneous three-dimensional atomic coordinate matrix contained in the current phase structure state as a strong spatial geometric context constraint, and feeds it back to the sequence prediction module. The sequence prediction module combines the current phase structure state, the explicit evolutionary prior representation, and the current sequence state to perform a global feature scan of the one-dimensional sequence space. Since the relative distance, dihedral angle, and local geometric obstacles between residue sites in the three-dimensional space of a protein determine the types of amino acids that a spatial location can accommodate, the sequence prediction module performs energy and topology matching calculations accordingly. For all vacant sites in the current sequence state that are still in a masked state, it predicts the amino acid probability distribution of the corresponding twenty natural amino acids (e.g., outputting a twenty-dimensional Softmax probability vector).

[0061] Step S330: Determine the structural neighborhood information corresponding to the still-masked sites based on the current round structural state, and determine the set of demasked sites for this round from the still-masked sites based on the structural neighborhood information.

[0062] To guide the stable convergence of the cooperative iterative generation process and prevent severe conflicts and collapses in the local physicochemical properties of the protein core region due to disordered or random unmasking, the system introduces a structure-aware dynamic screening mechanism. Based on the latest three-dimensional coordinates of each residue in the current structural state, the system calculates the structural neighborhood information corresponding to all currently masked sites. This structural neighborhood information characterizes the topological importance or spatial density (e.g., spatial atomic packing density, geometric centrality, or number of contact nearest neighbors) of the masked site in the three-dimensional structure. Based on the calculated structural neighborhood information, the system performs hierarchical screening from all currently masked sites, selecting a set of specific sites with the highest spatial constraint and the greatest need for priority determination of amino acid types in the current round, thus constructing the unmasked site set for this round. This mechanism, where the unmasking order is dynamically determined by geometric coordinates, allows the model to adopt a sequential, step-by-step progressive strategy from the inside out and from strong constraints to weak constraints.

[0063] Step S340: Based on the amino acid probability distribution, determine the residue type of each point in the current round of demasking site set, update the current sequence state using the determined residue type, and update the current structural state using the current round structural state.

[0064] After determining the set of demasking sites for this round, the system extracts the amino acid probability distribution corresponding to each site in the set. Using a sampling operator or a maximum probability selection operator, the system ultimately determines the specific amino acid type, or residue type, of each site in the set of demasking sites for this round based on the amino acid probability distribution. Subsequently, the system performs a global state transition and synchronization update operation: on the one hand, it uses the newly determined residue type to fill in and update the current sequence state, correcting the corresponding site from the mask state to the determined residue symbol; on the other hand, it uses the current structural state to cover and update the global current structural state. The updated current sequence state and the current structural state together serve as the baseline input for the next round of collaborative iterative generation. Since the newly determined residue type is fixed in the sequence state, it will directly constrain the generation of the new structure in step 1 of the next cycle, thus constructing a closed-loop control with bidirectional sequence and structure engagement and dynamic convergence in the temporal direction.

[0065] In the collaborative iterative generation process, scientifically determining the generation order of each unknown site is crucial to ensuring the stability and sequence coherence of the final three-dimensional structure of the protein. Existing technologies typically employ completely random or fixed linear demasking strategies, which can easily lead to severe conflicts in the physicochemical properties of the protein's core region, thereby causing the collapse of the overall structure. To overcome this technical bottleneck, this invention proposes a structure-aware dynamic demasking site selection mechanism.

[0066] Specifically, step S330 involves determining the structural neighborhood information corresponding to the still-masked sites based on the current round's structural state, and determining the set of demasked sites for this round from the still-masked sites based on the structural neighborhood information, including: Step S331: For each site i in the set I of sites that are still masked, calculate the number N of its spatial neighborhood. i As the structural neighborhood information, the calculation formula (Formula 1) is: ; Among them, Cα i and Cα j These are the three-dimensional Cartesian coordinates of the Cα (alpha carbon) atoms at sites i and j in their respective carousel structural states. The Cα atom was chosen as the benchmark because it most stably represents the main chain orientation of the protein backbone.

[0067] Represents Euclidean distance; used to precisely measure the straight-line physical distance between two atoms in three-dimensional space.

[0068] d thr This is a preset spatial neighborhood threshold (i.e., the cutting radius), used to define an effective spherical action space.

[0069] The actual physical execution logic of this formula is as follows: With the Cα atom coordinates at site i as the geometric center, and d thr Construct a spatial sphere with radius, and count the total number of other points j that fall within the spatial range of this sphere and are also in a masked state. This total number is the number of spatial neighborhoods N. i N i The larger the value, the stronger the spatial hindrance and conformational constraints on the hydrophobic core or tightly folded region of the protein deep at site i.

[0070] Step S332: For all sites in the set I, calculate according to the spatial neighborhood number N. i Sort in descending order and select the top r results. t ×N sites form the current round's unmasking site set; where N is the total amino acid length of the target protein; r t This is the preset demasking ratio for the current round.

[0071] Subsequently, after obtaining the spatial neighborhood number N of all current mask points... i Then, the system processes all sites in set I according to their corresponding N. i The values ​​are sorted in descending order. This descending order sorting mechanism establishes a "core-first, inside-out" generation strategy.

[0072] Finally, based on the sorting results, the system selects the first r items from the top of the list. t ×N sites are combined into a set of demasking sites for this round. Here, N is the total amino acid length of the target protein (i.e., the total number of residues in the sequence), and r_t is the preset demasking ratio for this round. The product r t ×N determines the quota of sites that are allowed to have their true amino acid types identified in this iteration. Through this truncation mechanism, the system will force priority to lock onto those sites that are at the geometric core and have the densest constraints for amino acid type prediction and fixation in each iteration.

[0073] This dynamic sorting and quota-based truncation strategy, based on spatial topological density, enables the co-generation model to simulate the real physical process of natural protein folding, where "the hydrophobic core collapses first to form a stable framework, followed by adjustment of the flexible surface regions." Once the core sites are preferentially demasked and fixed in the early stages of iteration, they provide extremely stable and accurate physical anchors for the subsequent denoising and generation of peripheral sites, thereby significantly improving the overall convergence stability of the algorithm and the conformational rationality of the generated molecules.

[0074] In specific implementation, in order to further improve the accuracy of structure sensing calculation and the thermodynamic stability of the overall state evolution, this embodiment makes physically meaningful quantitative limits on the space detection radius and the single-step evolution ratio.

[0075] In some implementations, the spatial neighborhood threshold d thr The value range is 8 Å to 12 Å. For example, the spatial neighborhood threshold can be 8 Å, 8.5 Å, 9 Å, 9.5 Å, 10 Å, 10.5 Å, 11 Å, 11.5 Å, 12 Å, etc.

[0076] In some implementations, the demasking ratio r in a single round t The value range is 0.01 to 0.20. For example, the demasking ratio in a single round can be 0.01, 0.03, 0.05, 0.08, 0.10, 0.12, 0.15, 0.18, 0.20, etc.

[0077] On one hand, the spatial neighborhood threshold is set to a range of 8 Å to 12 Å. The spatial neighborhood threshold refers to the cutoff distance in three-dimensional space used to determine whether two residue nodes belong to an effective interaction or a close-packed state. Based on the distribution of intermolecular forces in biophysics, the strong effective geometric constraints within proteins are mainly concentrated within this spatial scale. Using a threshold within this range for spatial feature scanning can effectively filter out background noise interference from distant irrelevant nodes, while avoiding the loss of crucial secondary coordination layer topological information due to an excessively small detection radius. This quantitative control ensures that the selected high-density core sites have the strongest geometric representativeness and physical constraints.

[0078] On the other hand, the demasking ratio for a single round is set to a range of 0.01 to 0.20. The single-round demasking ratio is defined as the ratio of the number of sequence sites that transition from an unknown state to a determined state to the total residue length of the target protein in a single independent cycle. Since determining the amino acid type immediately alters the local physical potential surface, using this numerical range for gradual, small-batch residue determination provides sufficient adaptive adjustment and stress release space for the three-dimensional geometric conformation. This mechanism fundamentally avoids atomic collisions or electrostatic repulsion caused by large-scale simultaneous residue determination, effectively ensuring the smoothness of the overall geometric manifold during evolution and the success rate of final molecule generation.

[0079] Example 3 Based on the foregoing embodiments, this embodiment provides a method for co-generating protein sequences and structures. During the co-iterative generation process, after the system predicts the probability distribution of amino acids corresponding to sites still in a masked state, how to convert this distribution into specific physical residue types is a key control step determining the diversity and structural rationality of the generated macromolecular sequences. If only a simple greedy search is used (i.e., always selecting the amino acid with the highest probability value), the model often degenerates into a simple memorization and reproduction of the natural sequences in the training set, failing to create protein molecules with novel sequences; while if completely random sampling is used, it will greatly disrupt the physicochemical matching degree between the sequence and the three-dimensional spatial conformation. To accurately balance sequence novelty and structural feasibility, this embodiment introduces a residue determination mechanism based on temperature sampling.

[0080] Specifically, step S340, based on the amino acid probability distribution, determines the residue type of each point in the current round of demasking site set, including: Step S341: Based on the amino acid probability distribution, temperature sampling is performed on each site in the current round of demasking site set using a preset sequence sampling temperature τ to determine the residue type of each site.

[0081] In this step, the sequence sampling temperature (usually denoted by the mathematical symbol τ) is a global or local hyperparameter used to adjust the smoothness of the model's predicted probability distribution. Its underlying algorithm is derived from the Boltzmann distribution in statistical thermodynamics. In practice, the system first obtains the raw logits or unscaled prediction vectors for twenty natural amino acids output by the prediction module. Then, the system divides each element of this vector by the sequence sampling temperature τ and calculates the readjusted target sampling probability distribution using a normalized exponential function (Softmax).

[0082] The presence of the sequence sampling temperature τ gives the generation process great flexibility in optimization. When the set value of τ is small (e.g., less than 1), the adjusted probability distribution becomes sharper, and the model tends to exhibit higher "conservatism," that is, high-probability amino acid types will have a higher chance of being selected, thus ensuring that the generated residues achieve the most stable match with the local potential energy surface of the current microscopic three-dimensional structure. When the set value of τ is large, the adjusted probability distribution is "smoothed out," allowing rare amino acids with low probability to also have a certain chance of being selected, thus injecting reasonable physical random perturbations into the generation trajectory.

[0083] By applying the aforementioned polynomial random sampling with temperature parameters, the system identifies a unique and specific amino acid type (residue type) for each site in the current set of unmasked sites. This mechanism effectively prevents the model from prematurely falling into local energy minima during multiple rounds of cooperative iteration, while ensuring that the residue types conform to explicit evolutionary priors and current three-dimensional spatial geometric constraints. This not only significantly expands the model's exploration boundaries in the chemical space of protein sequences, successfully designing novel functional sequences not yet discovered in nature, but also maintains conformational consistency between the generated molecule and its natural homologous distribution, significantly improving the expression and folding success rate of the target molecule in downstream wet experiments.

[0084] In the process of performing alternating interleaving updates to achieve joint generation of protein sequences and structures, in order to achieve the best convergence effect between the huge sequence search space and the complex geometric conformation evolution, this embodiment has made physically meaningful quantitative constraints on the hyperparameters involved in micro-probability sampling and the number of steps in macro-dynamic deduction.

[0085] Specifically, the value range of the sequence sampling temperature τ is 0.6~1.0. For example, it can be 0.6, 0.65, 0.7, 0.75, 0.8, 0.85, 0.9, 0.95, 1.0, etc.

[0086] In deep learning prediction, the temperature parameter directly determines the entropy of the output probability distribution. When a generative model attempts to explore unknown and novel protein sequences, excessively low sampling temperatures (e.g., below 0.6) cause the sampling process to degenerate into an absolutely greedy strategy, resulting in the model generating only trivial sequences highly homogeneous with the natural training set, thus losing the innovativeness of molecular design. Conversely, excessively high sampling temperatures (e.g., above 1.0) introduce excessive random thermal noise into the system, causing the model to easily misallocate large or electrostatically repulsive amino acid residues in spatially confined or physically constrained regions, leading to spatial collisions and disrupting the stability of molecular folding. By precisely locking the sequence sampling temperature τ within the range of 0.6 to 1.0, the system can maintain a high degree of obedience to structural topology while endowing the sequence sampling with appropriate diversity flexibility, perfectly balancing the "novelty" and "physical feasibility" of the molecule.

[0087] In some implementations, the total number of iterations in the collaborative iterative generation process is 6 to 12. For example, it can be 6, 7, 8, 9, 10, 11, 12, etc.

[0088] This design aligns perfectly with the progressive thermodynamics of protein folding. In the multi-round interleaving update, the overall demasking process is discretized into 6 to 12 macroscopic batches. This allows the system sufficient iteration time to release local steric hindrance and stress after each batch determines a small number of key residues. This range avoids structural collapse caused by too few rounds or too large a single update, while also preventing wasted computational power due to too many rounds.

[0089] In some implementations, the number of inversion steps for structure diffusion generation in each round of the cooperative iterative generation process is 100 to 300 steps. For example, it can be 100 steps, 125 steps, 150 steps, 175 steps, 200 steps, 225 steps, 250 steps, 275 steps, 300 steps, etc.

[0090] The number of inversion steps represents the number of times the structure denoising network refines and fine-tunes the 3D spatial coordinates in a single round. Setting a lower limit of 100 steps ensures that each update of the geometric conformation has a sufficiently deep Markov chain walk to accurately correct atomic-level bond length and bond angle deviations; setting an upper limit of 300 steps effectively eliminates ineffective computations with diminishing marginal returns while maintaining atomic-level resolution accuracy, thereby maximizing the overall model generation quality and computational inference efficiency.

[0091] In a preferred implementation scenario, this invention employs a set of unequal-distance decision parameters that prioritize the core over the periphery: the total number of iterations is set to T=9; the demasking ratio for each iteration is sequentially set to [0.02, 0.05, 0.10, 0.15, 0.15, 0.15, 0.15, 0.15, 0.08]; and the diffusion inversion steps per iteration are set to 200 steps, with a sequence sampling temperature τ=0.8. This parameter combination allows for the careful determination of key structural sites with extremely small step sizes in the early stages of iteration, rapid advancement of peripheral sites in the middle stages, and fine-tuning in the final stages, achieving optimal generation stability.

[0092] Example 4 Based on the foregoing embodiments, this embodiment provides a method for co-generating protein sequences and structures. Before performing the joint generation of protein sequences and structures, constructing a scientifically sound and error-free physical starting point for the model is fundamental to ensuring stable convergence of subsequent co-iterations. To accurately align the information certainty between the sequence space and the three-dimensional geometric space, this embodiment provides an asymmetric state initialization mechanism based on physical boundary constraints.

[0093] Specifically, when initializing the current sequence state and current structural state of the target protein: For the masked positions in the current sequence state, only N and C are retained in the current structural state. α C and O main chain atoms; for non-masked sites, all atoms are retained in the current structural state. In this step, the "N, Cα, C, O backbone atoms" refer to the universal basic atoms that constitute the backbone of all natural protein polymers: amide nitrogen (N), alpha carbon (Cα), carbonyl carbon (C), and carbonyl oxygen (O). These four atoms determine the global folding topology (backbone orientation) of the protein, without involving the specific physicochemical properties of the amino acid side chains (such as hydrophobicity, charge, and side chain volume).

[0094] For the "masked sites" in the current sequence state (i.e., unknown residues that the model needs to explore and predict), since their final amino acid types are not yet determined, the three-dimensional morphology of their side chains is completely unknown. If a certain side chain atomic coordinate is forcibly assigned to them in the initial structural state (e.g., randomly assigning a side chain of alanine or glycine), it will introduce false spatial geometric volumes or incorrect electrostatic repulsion signals into the three-dimensional space. This incorrect physical bias will mislead the evolution trajectory of the model in the subsequent denoising diffusion. Therefore, this embodiment adopts a "physical blanking" strategy: forcibly stripping all side chain atomic coordinates of these masked sites, retaining only the common N, Cα, C, and O main chain atoms as the most basic skeleton placeholders, thereby providing an extremely clean and unbiased geometric canvas for subsequent structural evolution.

[0095] Conversely, for the "non-masked sites" in the current sequence state (i.e., known residues that are pre-specified in the design task and do not need to be changed, such as the fixed catalytic center of an enzyme or a known binding epitope of an antibody), the amino acid species are completely determined. Therefore, the system fully preserves the "all atoms" (i.e., the three-dimensional group of atoms containing the main chain atoms and all the side chain branches that make up the amino acid) of these sites in the current structural state.

[0096] This asymmetric initialization strategy achieves precise matching of microscopic information. The retained whole atoms act as robust and extremely precise "physical anchors" and "three-dimensional constraint boundaries" in three-dimensional space; while the masked regions, which retain only the main chain atoms, retain great conformational flexibility. In the subsequent alternating interleaving and updating process, the generative network can accurately rely on the rigid physicochemical environment provided by the known whole atoms to adaptively deduce and extend the three-dimensional skeleton and sequence prediction of the unknown masked regions, thus exhibiting extremely high generative coherence and structural rationality in high-value application scenarios such as fixed local functional sites (e.g., scaffold design).

[0097] Example 5 Based on the foregoing embodiments, this embodiment provides a method for co-generating protein sequences and structures. To fundamentally ensure the accuracy and physical feasibility of the target protein in the subsequent co-generation and evolution, it is essential to ensure that the biological background knowledge input to the model has an extremely high signal-to-noise ratio. Since sequencing errors, fragmented segments, and highly redundant data are prevalent in natural biological sequence databases, directly using the raw retrieved data can lead to the model extracting spatial topological relationships with serious biases. Therefore, this embodiment introduces a rigorous data cleaning mechanism before mathematically encoding the multiple sequence alignment data.

[0098] Specifically, before encoding the multi-sequence alignment data, the method further includes cleaning the multi-sequence alignment data; the cleaning includes: removing homologous sequences that are repeated with the query sequence, and removing sequences with a gap ratio higher than 0.6, so that the effective depth of the cleaned multi-sequence alignment matrix is ​​not less than 32.

[0099] In this cleaning mechanism, the first step is to remove redundant information, i.e., "removing homologous sequences that are duplicates of the query sequence." The query sequence is the reference sequence of the target protein to be analyzed. In multiple sequence alignment matrices, the underlying physics for extracting spatial structural contact information lies in the "co-variation" that occurs between homologous sequences. If the data matrix contains a large number of sequence copies that are highly similar to or completely duplicate the query sequence, and these copies do not have any valuable mutational differences, they will not only unnecessarily increase the computational resource consumption of the coding network, but also severely dilute the true co-evolutionary covariance signal. Therefore, by removing these duplicate homologous sequences, the abundance of evolutionary variation within the alignment matrix can be effectively improved.

[0100] Secondly, low-quality fragments are filtered out, specifically sequences with a gap ratio higher than 0.6. To clarify this for those skilled in the art, the term "gap" is explained separately: In multiple sequence alignment algorithms, to achieve spatial alignment of homologous sequences of different lengths at amino acid sites, the algorithm inserts gap symbols (usually represented by hyphens "-") into the sequences. These gap symbols are called gaps. When the gap ratio in a sequence exceeds 0.6 (i.e., the actual effective amino acid residues in the sequence are less than 40% of the total length), the sequence usually originates from incomplete sequencing fragments or belongs to distantly related sequences that only match accidentally in local structural domains. These sequences lack a large amount of spatial topological context and cannot independently fold into a global three-dimensional conformation consistent with the target protein. By removing them, the system effectively filters out three-dimensional spatial alignment noise introduced by incomplete alignment.

[0101] Finally, the above cleaning rules are subject to a global statistical lower bound constraint: "to ensure that the effective depth of the cleaned multiple sequence alignment matrix is ​​not less than 32 sequences." For clarity and patent protection, "effective depth" is explicitly defined here: effective depth refers to the total number of high-quality homologous sequences retained in the matrix after all the denoising and filtering rules described above, used for actual feature encoding calculations. In computational biology, extracting the spatial co-evolutionary coupling matrix between pairs of amino acid sites is essentially performing covariance statistical inference of multidimensional variables. If the number of samples participating in the statistics is too small, accidental random mutations are easily misjudged as physical spatial contact associations. Therefore, strictly limiting the effective depth to not less than 32 sequences ensures, from a statistical perspective, that the input sample size has sufficient significance. This hard metric guarantees, from the source, that the final extracted explicit evolutionary prior has high mathematical stability and physical authenticity.

[0102] In order to achieve the best engineering balance between data representation accuracy and computer hardware computing power consumption, when transforming the original biological homologous sequence information into constraints that can be read by the generative model, this embodiment has carried out precise quantitative design of the underlying tensor structure and feature dimensions of the explicit evolutionary prior representation.

[0103] Specifically, in the explicit evolutionary prior representation that includes both single-site features and site-pair features: The single-site feature is the dimensionality-reduced single-site feature tensor S. msa Its feature dimension d s =256; The site pairs are characterized by the reduced-dimensional pairwise feature tensor Z. msa Its feature dimension d z =128.

[0104] On the one hand, the individual site features are constructed as a dimensionality-reduced individual feature tensor S. msa Furthermore, its feature dimension is strictly set to d. s =256. In the original homology sequence matrix, the amino acid distribution information at individual residue sites exhibits a highly sparse, high-dimensional discrete state. This embodiment performs dimensionality reduction through a feature mapping network, compressing this sparse signal and projecting it into a dense, continuous vector space. The dimension d of the monomer feature is... s The setting of 256 dimensions has the physical significance of the following: a 256-dimensional continuous vector space has been sufficiently proven to have sufficient representational capacity, capable of losslessly encoding one-dimensional properties of individual amino acid sites, such as hydrophilicity / hydrophobicity, spatial volume requirements, and charge conservation, exhibited during long-term evolution. Simultaneously, this dimensional size effectively prevents the model from overfitting to noise in the original aligned data, ensuring the generalization ability of the generation process.

[0105] On the other hand, the site pair features are constructed as a dimensionality-reduced pairwise feature tensor Z. msa Furthermore, its feature dimension is strictly set to d. z =128. Site-pair features are used to characterize the co-evolutionary coupling between any two residues to maintain the stability of the three-dimensional spatial structure. In terms of data structure, the pairwise feature tensor presents a two-dimensional grid, and its spatial complexity is O(N) with respect to the total amino acid length N of the target protein. 2 The feature tensor exhibits a quadratic explosive growth. To overcome the memory bottleneck of computing devices caused by the excessive size of the feature tensor, this embodiment creatively adopts an asymmetric dimensionality design strategy. Through further dimensionality reduction and compression, the dimension d of the paired features is reduced... z The dimensions were reduced and locked to 128. This precise numerical setting successfully achieved an excellent balance between preserving the geometric constraints of the core space (such as the relative distances between residues and dihedral relationships) and controlling quadratic memory growth.

[0106] The single-unit feature tensor S is described above. msa The dimension is set to 256, and the paired feature tensor Z msa With a collaborative dimensionality reduction design set at 128, the system successfully constructed an efficient explicit evolutionary prior foundation that includes both high-precision one-dimensional sequence conservatism and rich two-dimensional spatial topological relationships, and is fully adapted to modern GPU computing architecture. This provides robust data flow support for the subsequent alternating and intertwining updates of deep networks.

[0107] In some embodiments, for the target protein with a total amino acid length of N, the monomeric feature tensor is: The i-th position is represented as (Formula 2): ; For the target protein with a total amino acid length of N, the paired feature tensor is: The i-th row and j-th column is represented as (Formula 3): ; Where M is the number of homologous sequences; E m,i w is the encoding vector of the m-th homologous sequence at the i-th position; m For sequence weights; W z It is a linear projection matrix; This represents the tensor outer product.

[0108] In collaborative generation methods, scientifically transforming massive amounts of multi-sequence alignment text data into mathematical tensors that can be processed by neural networks is a key step in constructing explicit evolutionary priors. This embodiment details the construction logic and physical significance of the underlying calculation formulas for single-site features and site-pair features.

[0109] Specifically, for the target protein with a total amino acid length of N, the system first needs to determine the basic mathematical representation of each homologous sequence. Assume the preprocessed multiple sequence alignment data contains M homologous sequences. The system performs site-by-site vectorization encoding on each sequence, let E... m,i Let w be the coding vector of the m-th homologous sequence (1≤m≤M) at the i-th amino acid site (1≤i≤N). Meanwhile, to eliminate statistical bias caused by oversampling of certain protein families in the sequencing database, the system calculates and assigns a sequence weight w to each sequence based on the similarity between sequences. m .

[0110] Based on this data, the system first constructs the monomer feature tensor. This monomer feature tensor is used to characterize the evolutionary conservation of the target protein at each individual residue position. The feature vector at the i-th site in this tensor... The calculation formula is set as Formula 2.

[0111] According to this formula, the system encodes the vector E of all M homologous sequences at the i-th physical site along the depth direction (i.e., the vertical sequence dimension) of the multi-sequence alignment matrix. m,i A weighted summation operation is performed. The physical mechanism of this linear aggregation process lies in the fact that, during the long process of natural selection, amino acid sites that maintain core functions rarely undergo mutations. Through the above weighted statistics, the system can accurately extract the amino acid type distribution preference of the i-th site in the entire evolutionary family, thus providing a one-dimensional physicochemical constraint for subsequent sequence probability prediction of this site.

[0112] Subsequently, the system further constructs the paired feature tensor. This tensor is used to characterize the co-evolutionary association between any two residue sites in the target protein to maintain spatial structural stability. The calculation formula for the paired feature element in the i-th row and j-th column of this tensor is set as Equation 3.

[0113] When performing the calculation of this formula, the system first uses the tensor outer product operator ( ), calculate the cross product of the encoding vectors at positions i and j of the m-th sequence ( This captures the joint residue state of these two sites within a single sequence. Then, by weighted summation across the entire evolutionary family containing M sequences, the system statistically calculates the evolutionary covariance between site i and site j. Since amino acids in close contact in three-dimensional space often undergo mutually compensating co-evolutionary mutations (e.g., maintaining volume complementarity or charge balance), this covariance signal essentially reflects a very strong three-dimensional spatial topological regularity.

[0114] Finally, considering that directly performing outer product summation would result in a highly dimensional redundant joint matrix, the system introduces a learnable linear projection matrix W based on the summation result. z A multiplication mapping is performed. This linear projection process not only achieves reasonable compression of feature depth and removes background statistical noise, but also projects the co-evolutionary signal into the semantic subspace most suitable for deep learning network understanding. Through the rigorous formula derivation above, the system successfully transforms discrete homologous sequence characters into a highly dense two-dimensional spatial geometric prior representation.

[0115] In collaborative generation methods, the core factor determining the convergence accuracy and physical rationality of the generative network lies in how to deeply integrate static evolutionary laws into the dynamic process of structural conformation deduction. This embodiment details the fusion and update mechanism of deep features within the structural backbone network.

[0116] In each round of the collaborative iterative generation process, the system does not only use evolutionary information at the initial input of the model, but also constructs a deep conditional injection mechanism. Specifically, the system uses the pre-extracted individual feature tensor S msa and the paired feature tensor Z msa The static prior tensor injected continuously and explicitly into the structural backbone network is a deep inference architecture composed of multiple cascaded network layers. When the flow reaches the network layer with layer index l, the system controls the externally injected static prior tensor and the currently flowing single-unit representation H within that layer. l And the current pairwise representation P l Perform a merge update.

[0117] In this embodiment, specifically, during each round of the cooperative iterative generation process, the individual feature tensor S msa and the paired feature tensor Z msa Injected into the structural backbone network, and related to the current single-unit representation H of the structural backbone network. l And the current pairwise representation P l Perform a merge update, and the merging formula is (Formula 4): ; Where Ws and Wp are linear projection matrices; PairFormer is the feature fusion and update module; H l This represents the current singleton; P l For the current paired representation; H l+1 is the updated single-unit representation; l represents the hierarchical index of the backbone network of the structure.

[0118] When performing this formula calculation, the system first performs linear alignment of the feature space. Since the evolutionary prior tensor of the external input may differ from the channel depth of the network's internal latent variables, the system utilizes a pre-defined learnable linear projection matrix W. s For the single-characteristic tensor S msa Perform multiplication mapping, while utilizing the linear projection matrix W p For paired feature tensors Z msa Perform a multiplication mapping to ensure strict alignment of feature dimensions and filter out redundant background noise.

[0119] After alignment, the system uses a residual addition method to map the prior tensor (W). s S msa and W p Z msa ) respectively with the current internal representation (H) of the corresponding orbit in the network l and P l This involves performing element-by-element addition. This addition and fusion mechanism can impose strong natural evolutionary boundary constraints on the current instantaneous state of the network without disrupting the original forward propagation information flow.

[0120] Subsequently, the system will incorporate the latest singleton variable (H) that incorporates evolutionary priors. l +W s S msa ) and paired variables (P l +W p Z msa Both two-dimensional and one-dimensional variables are used as inputs and fed into the feature fusion and update module, namely the PairFormer module. The PairFormer module is a set of high-order neural network units used to perform cross-track information interaction. Its core function is to capture the topological dependencies between two-dimensional and one-dimensional variables. In the three-dimensional physical space of a protein, two-dimensional paired features essentially characterize the relative distance and steric hindrance between residue sites, while one-dimensional monomeric features characterize the independent properties of a single residue. The PairFormer module uses two-dimensional paired variables containing co-evolutionary contact associations as spatial geometry guides to perform attention weighting and feature extraction on the one-dimensional monomeric variables, ultimately deriving and outputting the updated monomeric representation H. l+1 .

[0121] Based on the above operational logic, H l+1This will be seamlessly passed to the next network layer for further deduction. This mechanism of continuously enforcing "prior injection" and "cross-track fusion" deep within the network greatly enhances the conformational stability and physical rationality of the network when predicting complex protein topologies, eliminating the possibility of generated trajectories deviating from the natural physical manifold from the algorithm's bottom layer.

[0122] In the diffusion and denoising process of protein structure generation, enabling deep learning models to perceive and respond to constantly changing three-dimensional spatial geometry in real time is crucial for improving conformation generation accuracy. To address this, this embodiment discloses a low-level computational logic that integrates two-dimensional feature representation with three-dimensional geometric distance using a gating mechanism.

[0123] Specifically, the current pairwise representation P l The paired feature elements at the i-th and j-th sites are represented by P. ij The paired feature elements P ij By fusing the gating mechanism with the geometric distance embedding of the current diffusion step, conditional input elements for the structure denoising network are obtained. The fusion formula is (Formula 5): ; in, (Formula 6); The geometric embedding is constructed based on interatomic distances; σ is the gating function; ⊙ represents the Hadamard product; W d and W g r is a linear mapping matrix; i and r j For the three-dimensional coordinates of the atom; For the paired feature tensor Z msa The paired features in the i-th row and j-th column of the array.

[0124] During the structure generation process, the two-dimensional tensors flowing within the network contain abstract topological relationships between residue pairs. To achieve precise computation from macroscopic tensors to microscopic site pairs, the system first defines the current pairwise representation P. l The specific paired feature elements at the i-th and j-th sites are represented by P. ij .

[0125] Building upon this, the system does not directly pass the feature element down the chain. Instead, it fuses it with the instantaneous geometric distance embedding of the current diffusion step through a highly sophisticated gating mechanism. The resulting fused feature serves as a powerful conditional input element, directly guiding the operation of the structure denoising network. The core calculation formula for this gating fusion mechanism is Equation 5.

[0126] Before executing the fusion formula, the system must simultaneously calculate two key auxiliary components: the gating coefficient g. ij and geometric embedding .

[0127] On the one hand, the system extracts the three-dimensional atomic coordinates of the i-th and j-th sites at the current diffusion step (denoted as r, respectively). i and r j ), calculate the Euclidean distance between the two in the current instantaneous configuration. Subsequently, the system utilizes a geometric embedding function φ (such as a radial basis function network or a positional encoding function) constructed based on inter-atomic distances to map this scalar distance into a high-dimensional geometric embedding feature. This geometric embedding enables the network to sense microscopic atomic spatial interactions (such as repulsive forces or van der Waals forces).

[0128] On the other hand, the system calculates the gating coefficient g used to balance the multimodal feature weights. ij The calculation formula is shown in Formula 6. During execution, the system will use the current paired feature elements P... ij With explicit evolutionary prior tensor Z msa The paired feature Z corresponding to the i-th row and j-th column ij msa Perform concatenation along the channel dimension ([·;·]). The concatenated high-dimensional vectors are then sequentially mapped through the linear mapping matrix W. g The dimensionality reduction transformation and the processing of the gating function σ (preferably a σ nonlinear activation function) output a dynamic weight vector whose values ​​are strictly limited to between 0 and 1, i.e., the gating coefficient g. ij The brilliance of this mechanism lies in the fact that it allows naturally occurring evolutionarily conserved associations (Z) to be preserved. ij msa This can directly participate in determining the allocation of feature weights. Ultimately, the system uses the Hadamard Product (symbol ⊙, representing the one-to-one multiplication of elements within a vector) to perform complementary weighting operations. The system uses g... ij Retain some original features P ij Information, while using complementary weights (1-g) ij To adjust through the linear mapping matrix W d The transformed 3D geometric embedding is then summed to obtain the conditional input elements. This dynamic soft-gating strategy ensures that the feature network and the real physical geometric coordinates maintain strong spatial linkage and adaptive correction throughout the entire structural denoising cycle, greatly improving the physical rigor of the final 3D folded structure.

[0129] Example 6 Based on the foregoing embodiments, this embodiment provides a protein sequence-structure co-generation method. After constructing a powerful underlying sequence-structure co-generation engine through the aforementioned alternating interleaving update mechanism, this embodiment further illustrates the high scalability of this method in practical industrial applications. By flexibly setting the mask region and initial physical boundaries of the input sequence, the protein sequence-structure co-generation method can seamlessly support a variety of different task modes.

[0130] Specifically, the protein sequence-structure co-generation method supports a variety of different task modes, including at least one of the following: monomer design mode, scaffold design mode, conjugate design mode, and specific functional modification mode.

[0131] In Monomer Design mode, the system masks the entire protein sequence, generating a novel single protein with a stable folded structure from scratch. In Scaffold Design mode, the system fixes specific functional motifs in the target protein without masking them, only masking the surrounding support regions to generate a novel physical scaffold. In Specific Function Modification mode, the system can target and optimize the sequence of enzyme catalytic sites or fluorescent protein luminescent centers by introducing specific fitness scoring biases.

[0132] Specifically, to overcome the challenging problems of antibody design and targeted drug development in the biopharmaceutical field, this system introduces a targeted algorithmic intervention strategy when executing the binder design mode. When the task mode is set to binder design mode, the system receives the initial state of the complex composed of the target protein and the binder protein to be generated. In each round of determining the unmasking site set, the system not only sorts the sites according to the conventional spatial neighborhood quantity, but also automatically identifies the interface neighborhood sites on the complex interaction interface through spatial geometric calculations. Subsequently, the system artificially increases the priority of these interface neighborhood sites in the unmasking site set (e.g., by assigning additional weight multipliers). This targeted intervention mechanism forces the model to prioritize the concentration of computing power on predicting and solidifying the key anchor residues that determine binding affinity in the early stages of iteration, thereby greatly improving the hit rate and specificity of binder drug design.

[0133] In addition to generating independent single protein molecules, the sequence-structure co-generation algorithm constructed in this method also exhibits high industrial scalability and scenario adaptability when dealing with complex biomolecular interaction scenarios. To meet the cutting-edge needs of biopharmaceuticals, antibody design, and targeted ligand development, this embodiment further illustrates the refined regulation mechanism of this method in specific business scenarios.

[0134] Among various functional design tasks, developing binding molecules capable of binding with high affinity to specific targets is a direction with significant commercial value and technical difficulty. Therefore, when the task mode in the system is set to binding design mode, this embodiment provides a targeted algorithmic intervention strategy. In this mode, the system environment processed by the model is no longer a single protein chain, but a spatial complex composed of a given target protein and the binding protein to be generated.

[0135] Specifically, when the task mode is a composite design mode, when determining the set of demasking sites for this round, the priority of selecting interface neighborhood sites at the composite interaction interface into the set of demasking sites for this round is increased.

[0136] In the physical laws governing protein-protein interactions (PPIs), the core factors determining whether a complex can successfully bind to a target often lie in a few key amino acid residues (i.e., binding hotspots) at the interface between the two proteins (i.e., the interaction interface of the complex). If the model follows the conventional inside-out energy convergence logic, prioritizing the determination of the complex's internal folding core and only generating the amino acids at the interface in the final stage, it is highly likely that the interface will lack spatial flexibility and fail to generate residue structures that perfectly complement the geometric and electrostatic characteristics of the target surface.

[0137] To overcome this technical bias, the system triggers a specific spatial detection and priority intervention mechanism when it reaches the step of "determining the set of unmasking sites for this round" in each iteration. First, the system identifies and filters "interface neighborhood sites" located at the complex interaction interface by calculating the spatial physical distance between each masking site on the sequence to be designed and the known atoms of the target protein in real time. These sites are the closest to the target surface in physical space and are the key anchor points that carry the binding function.

[0138] Subsequently, in terms of algorithmic logic, the system artificially increases the priority of these interface neighborhood sites being selected into the set of unmasked sites in this round. In specific execution, the system can apply specific weighted multipliers or score reward increments on top of the basic evaluation indicators that measure the importance of these masked sites, to ensure that these key sites located at the binding interface can stand out in the very early stages of multiple rounds of collaborative iteration and be the first to transform from the masked state to a definite amino acid residue type.

[0139] Through the aforementioned collaborative intervention strategy of "interface anchor priority," the system forces the generative model to concentrate computational power in the early stages of iteration to overcome the physicochemical constraints of the target surface, thus first constructing a binding interface with high affinity potential. In subsequent iterations, the remaining skeleton region of the complex spontaneously undergoes adaptive structural denoising and spatial folding around these established interface anchors. This mechanism thoroughly clarifies the causal logic of complex generation, significantly improving the specificity and positive functional hit rate of complex drug design.

[0140] Example 7 Reference Figure 5 This embodiment provides a training method for a protein sequence-structure co-generation model. The preceding embodiments detailed the operational mechanism of generative networks based on evolutionary prior constraints during the inference application stage. However, whether a model possesses keen structure awareness and sequence prediction capabilities fundamentally depends on the training method it undergoes before deployment. To obtain a generative model with high generalization ability, avoiding overfitting, and controllable training costs, this embodiment specifically provides a training method for a protein sequence-structure co-generation model. This training method is specifically used to train and optimize the various network architectures used in the aforementioned generative methods. The training method mainly covers four core execution stages: data preparation, input feature leakage and corruption prevention, freeze strategy control, and joint gradient update. Specifically, it includes: Step S10: Obtain homologous sequence multiple sequence alignment data of the training protein.

[0141] First, the system acquires multiple sequence alignment data of homologous sequences of the training proteins. During the training phase, the system typically retrieves real-world proteins in batches from known natural protein structure databases as training samples. The system uses sequence retrieval algorithms to extract homologous sequences of these training samples on the phylogenetic tree and constructs multiple sequence alignment data, which serves as the raw material for the model to learn the laws of natural evolution.

[0142] Step S20: Remove the sequence of the training protein that is used as the query sequence from the homologous sequence multiple sequence alignment data, and perform synchronization masking on the removed multiple sequence alignment data according to a preset mask ratio to construct training input features containing UNK tags.

[0143] Subsequently, to prevent the network from developing shortcut dependencies during the learning process, the system implemented strict anti-leakage measures on the input data. Specifically, the system forcibly removed the actual sequence of the training protein itself (which serves as the query sequence) from the homologous sequence multiple sequence alignment data. Without removing this sequence, the deep learning model could easily cheat by directly "copying" the input sequence using identity mapping to complete the prediction task, thus completely losing the motivation to extract co-evolutionary patterns. After removing the query sequence, the system synchronously masked the removed multiple sequence alignment data according to a preset masking ratio, that is, randomly selecting a certain proportion of sequence sites and replacing their amino acid symbols with UNK markers representing unknowns, thereby constructing training input features containing UNK markers. This disruptive mechanism based on the Masked Language Model (MLM) principle forces the model to infer missing information by mining the remaining context and spatial structure constraints, greatly improving the depth of representation learning.

[0144] Step S30: During the training process, the parameters of the structure generation module in the model are kept frozen, and joint training and optimization are performed only on the sequence module and encoder in the model.

[0145] After constructing the training input features, the system initiates the forward computation and backpropagation process of the model. To significantly reduce computational resource overhead while preserving existing 3D spatial priors, this embodiment employs a modular decoupling asymmetric training strategy. Throughout the training process, the system forcibly keeps the parameters of the structure generation module in the model in a frozen state. A frozen state means that during backpropagation, the gradients of the weights within this module are not calculated, nor are parameter updates performed. The system only performs joint training and optimization on the sequence module and encoder in the model. This strategy effectively prevents the 3D structure network from experiencing "catastrophic forgetting" in sequence reconstruction tasks, ensuring the absolute stability of its physical geometric topology deduction capabilities.

[0146] In step S40, during forward computation, the sequence module outputs the amino acid category probability distribution for the mask sites in the multiple sequence alignment data, and uses the back feedback of the amino acid category probability distribution to jointly update the network parameters of the sequence module and the encoder through the cross-entropy loss function. In the specific forward computation and backward optimization closed loop, the system feeds the aforementioned training input features containing UNK tags into the network. The sequence module outputs the amino acid class probability distribution for the masked sites in the multiple sequence alignment data. Subsequently, the system extracts the amino acid class at that position in the real natural sequence as the gold standard, and calculates the error between the predicted probability distribution and the true distribution using the cross-entropy loss function.

[0147] Specifically, with the structural module parameters frozen, the cross-entropy loss function L for a single training batch is calculated using the formula (Formula 7): ; Where I is the set of sites currently in the masked state, p θ For the probability distribution of the output predicted by the sequence module, s i For the actual amino acid class at site i, s t-1 and c t-1 represents the sequence and structural state of the previous round, respectively, and m represents the explicit evolutionary prior.

[0148] Finally, the system utilizes the back feedback of the probability distribution error of the amino acid category to calculate the gradient and jointly update the network parameters of the sequence module and the encoder. Through continuous iteration with thousands of training protein samples, the encoder learns how to extract the optimal explicit evolutionary prior, and the sequence module learns how to perform accurate decoding based on structural constraints, thereby jointly constructing a highly robust collaborative generative AI foundation.

[0149] To enable the constructed sequence-structure co-generation network to have strong cross-modal reasoning ability and generalization robustness, this embodiment introduces dynamic mask ratio sampling and cross-modal joint information stripping mechanisms in the leakage prevention and feature masking stages of the model training phase.

[0150] In some implementations, the preset mask ratio p is obtained by random sampling within a uniform distribution between 0.1 and 1.0, satisfying p ~ Uniform(0.1, 1.0). Regarding the value range of the preset mask ratio p (0.1~1.0), since it represents a scaling factor, it can be, for example, 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1.0, etc.

[0151] When preparing input features for training, the system does not employ a static, fixed masking ratio, but instead introduces a highly dynamic random sampling strategy. Specifically, in the anti-leakage process, the system sets the preset masking ratio p to be randomly sampled within a continuous interval from 0.1 to 1.0 according to a uniform distribution, with the mathematical logic satisfying p ~ Uniform(0.1, 1.0). This means that for each batch of protein data fed into the network for training, the system randomly assigns it an unknown ratio between 10% and 100%. This uniform sampling mechanism simulates the entire process from a "completely random initial state" to an "almost deterministic final state" in the generation phase at the algorithm's underlying level. By forcing the network to continuously perform gradient optimization under various degrees of information loss (including low-difficulty tasks that only need to fill in a few gaps, as well as high-difficulty tasks that require reconstructing the global picture with very few clues), the model effectively overcomes the risk of overfitting to a specific masking density, thus enabling it to smoothly and robustly handle iterative denoising tasks in any round in actual inference.

[0152] In some implementations, the training method employs a joint learning mechanism of the main chain and side chains. When masking the sites in the training input features, the corresponding amino acid side chains at the structural ends are simultaneously removed, retaining only N and C. α The C and O main chain atoms are used to enhance complementary inference learning of structure and sequence.

[0153] To prevent premature leakage of physical geometric features during training, this embodiment deeply implements a joint learning mechanism for the main chain and side chains in its training method. In deep learning models, the side chain geometry of amino acids has extremely strong uniqueness. If only one-dimensional sequences are masked during training without simultaneously processing three-dimensional spatial coordinates, the model can easily "cheate" by reading the geometry of the corresponding side chains to guess the residue type, causing the network to fail to truly learn the deep environmental constraints.

[0154] Therefore, this embodiment stipulates that when masking one-dimensional sequence sites in the training input features (e.g., replacing them with UNK markers), the system must simultaneously perform coordinate removal operations at the corresponding three-dimensional structural end. The system will accurately locate the full-atom coordinate matrix of the masked site in three-dimensional space and forcibly remove all amino acid side chain atoms representing the specific properties of that amino acid, so that only the four types of main chain atoms common to all amino acids—N (amide nitrogen), Cα (alpha carbon), C (carbonyl carbon), and O (carbonyl oxygen)—are retained at that spatial position. This cross-modal synchronous "physical blanking" completely cuts off geometric shortcuts. It forces the network to rely solely on the common main chain topology, the spatial microenvironment of surrounding adjacent residues, and the covariance signal provided by explicit evolutionary priors to comprehensively infer the amino acid probability distribution of the missing site during forward computation, thereby greatly enhancing the model's deep physicochemical learning of structural and sequence complementarity relationships.

[0155] The present invention will be further illustrated below with specific experimental examples. However, it should be understood that these experimental examples are only for more detailed illustration and should not be construed as limiting the present invention in any way.

[0156] Experiment Example 1: Sequence-structure consistency and natural manifold alignment, novelty verification experiment 1. Experimental Objective: In this experimental example, the main experiments conducted were natural manifold alignment and novelty verification experiments. The aim was to examine whether the protein sequence-structure pairs generated by this invention are located near the natural protein manifold, and to verify whether they have novelty compared to known natural homologous sequences, thus avoiding simple replication.

[0157] 2. Experimental methods: (1) Select 10 target protein families (e.g., Cytochrome C) and construct their multiple sequence alignment (MSA) data respectively.

[0158] (2) The method of the present invention (EvolveDesign) is used to generate sequence-structure pairs; as a comparison, the existing technologies ProteinMPNN and EvoDiff are used to generate proteins for control; in order to ensure the objectivity and fairness of the comparative experiment, the above control group models all use the default recommended parameters provided in their official open source code and load their officially released pre-trained weights for forward inference.

[0159] (3) Use phylogenetic trees, sequence embedding space and structural embedding space to compare the differences between generated distributions and natural homologous distributions (distance indicators include phylogenetic tree cluster distance, sequence embedding MMD and structural embedding MMD).

[0160] (4) Calculate the maximum sequence similarity between the generated protein and the target sequence, as well as with any homologous sequence in the MSA, and calculate the highest TMScore of the homologous structure set.

[0161] 3. Experimental Results and Analysis: First, regarding sequence-structure consistency, such as Figure 6 As shown, the method of this invention (EvolveDesign) achieved the highest consistency rate (up to 85%) among many comparative models, which is significantly better than the traditional decoupled design scheme. This fully demonstrates that the alternating interleaving collaborative generation mechanism adopted in this invention can effectively avoid the separation of sequence and structural features.

[0162] Secondly, secondly, such as Figure 7 and Figure 8 As shown, at most target sites, the distance between the generated results and the homologous distributions in the phylogenetic tree space and embedding space is smaller, proving that the generated results are closer to the natural homologous distributions and significantly reduce the proportion of "outliers that are not usable". Furthermore, Figure 7 The study further demonstrated the average positional Shannon entropy of MSA at each target and revealed a high Pearson correlation between MSA entropy and the three distance metrics. Figure 8 When displaying the sequence consistency distribution, Cohen's d effect size was annotated, which objectively quantifies the statistically significant difference between the generated distribution and the natural homology distribution.

[0163] like Figure 9 The phylogenetic distribution and Hamming distance distribution of the generated protein and the natural homologous sequence are visually displayed. It can be seen that most generated sequences do not simply replicate known homologous sequences, but have a certain Hamming distance relative to the natural homologous sequence. While maintaining sequence differences, they still maintain a very high family similarity at the structural level.

[0164] Experiment Example 2: Performance Evaluation of Functional Site Scaffold Task Design 1. Experimental Objective: In this experimental example, a functional site scaffold construction experiment was conducted to examine the success rate and structural accuracy of generating a full-protein scaffold under given geometric constraints of specific functional sites.

[0165] 2. Experimental Methods: For multiple scaffolding tasks, the requirement is to generate surrounding scaffolds while preserving functional sites. The evaluation success criteria are set as follows: prediction confidence meets pLDDT > 70, and the geometric error of the functional site meets motifRMSD < 1.0 Å. The experiment also includes a site constraint removal step (…). motif) and removal of evolutionary information ( This paper presents an ablation validation scheme for MSA (Mean Subtraction Anomaly) and compares its performance with existing models RFDiffusion and EvoDiff. Similarly, all existing comparison models used in this experiment are run based on the official default configuration and official pre-trained weights.

[0166] 3. Experimental Results and Analysis: (e.g.) Figure 10 and Figure 11 As shown, the method of this invention exhibits a high success rate across multiple targets (a total of 8 stent tasks) and demonstrates more stable worst-case performance. Specifically, Figure 10 The scatter plots in the diagram provide a detailed comparison of the prediction confidence and structural accuracy of a successful design (pLDDT vs. 1). The Cα RMSD distribution was analyzed, and the Spearman correlation between success rate and four attributes (site proportion, site dispersion, β-fold proportion, and MSA entropy) was revealed. Simultaneously, combined with... Figure 11 The representative structural comparison (where gray represents the natural structure and red marks the functional sites) visually demonstrates the high consistency of this invention in maintaining the geometric conformation of the functional sites. Ablation experiments further confirm the core mechanism: removal of evolutionary priors ( The success rate decreased significantly after MSA (subjective site-specific autologous asymmetry), indicating that the MSA prior of explicit injection is a key factor driving folding success; while removing explicit site constraints (MSA) Even after the motif, the system still has a certain percentage of successful samples, which proves that evolutionary priors themselves can implicitly provide powerful global structural organization capabilities.

[0167] Experiment Example 3: Protein-Binding Compound Design and Dynamic Stability Verification 1. Experimental objective: In this experimental example, a protein-binding complex design experiment targeting a fixed receptor site was conducted to verify that the complex generated by this method can not only match in static conformation, but also maintain stable binding of interactions during dynamic processes.

[0168] 2. Experimental methods: (1) The input includes information on IL-7Rα, KLK4 and PD-1 receptors and their corresponding MSA, and the homologous MSA of the binder is used as a supplementary a priori input to generate candidate binders in this invention.

[0169] (2) Preliminary judgment and screening were carried out using composite structure prediction indicators (ipTM, pTM and interface pAE threshold).

[0170] (3) For the selected successful samples, FoldX was used to evaluate their static binding free energy (ΔG), and a 100 ns molecular dynamics (MD) simulation was performed to record the CαRMSD trajectory for dynamic stability analysis.

[0171] 3. Experimental Results and Analysis: (e.g.) Figure 12 As shown, this invention demonstrates a strong or near-optimal design success rate in multiple receptor tasks. Energy calculations show that the interfacial energy distribution of successful samples is within a physically feasible range, and the affinity indices of some samples are superior to those of the natural reference complex. Further data analysis indicates a strong positive correlation (r=0.98) between the MSA entropy of the target and the relative FoldX binding free energy (ΔG). In the ablation experiment with MSA prior removal, a separate comparison between the method of this invention and RFDiffusion further highlights the importance of evolutionary information for interface generation, and the representative successfully designed complex structures all achieved extremely high pTM and ipTM indices. Molecular dynamics simulations (a widely used technique in this field for evaluating the stability of complex interfaces) trajectories show that the candidate complexes maintain a stable binding state on a nanosecond scale, without significant conformational drift or dissociation trends, theoretically confirming the experimental potential of the generated molecules with a high positive hit rate.

[0172] Experiment Example 4: Biological Expression and Validation of Modified Fluorescent Protein (Based on CreiLov) 1. Experimental Objective: In this experiment, the optimized generation and flow cytometry detection of CreiLov fluorescent protein were carried out. The aim was to examine the actual effect of the present invention in optimizing specific properties (fluorescence intensity and positive rate) while maintaining the core function of the protein from the perspective of real wet experiments.

[0173] 2. Experimental Methods: A candidate library was generated based on CreiLov homology MSA. Stratified sampling was performed according to embedding distance, and 50 candidate proteins covering neighboring, intermediate, and distant regions were selected for gene synthesis and expression in cells. Subsequently, flow cytometry was used to detect single-cell fluorescence intensity and the proportion of positive cells, and normalization analysis was performed according to the baseline control.

[0174] 3. Experimental Results and Analysis: (e.g.) Figure 13 and Figure 14 As shown, among the 50 tested generative designs, 28 designs outperformed the original control group in at least one metric: fluorescence intensity or positive cell percentage. The optimal samples (such as GenLov-1 and GenLov-2) achieved approximately a 20% increase in fluorescence intensity or approximately a 78% increase in the positive cell percentage. Figure 13 Data visualization features show that the color of the dots in the flow cytometry distribution map intuitively represents the degree of sequence difference, and the three-dimensional structure and sequence alignment analysis accurately indicates the FMN binding pocket through light boxes and blue highlights; this strongly proves that the generated protein effectively maintains the stability of key functional sites such as the FMN binding pocket, and that amino acid mutations and modifications are mainly concentrated on the tolerable flexible sites on the periphery of the pocket.

[0175] Experimental Example 5: Verification of Enhanced Catalytic Activity of Rare Fluoride Enzymes (Based on FlAMA37) 1. Experimental Objective: In this experimental example, fluoride enzymes were optimized and their in vitro catalytic activity was verified for a rare enzyme family with extremely complex catalytic functions. The aim was to evaluate the application value of this method in improving the catalytic efficiency of industrial enzymes.

[0176] 2. Experimental Methods: Homologous MSAs of the benchmark enzyme FlAMA37 were constructed, and candidate enzyme sequences were generated using this invention. Four representative candidate designs (named GenFIA-1 to GenFIA-4) were selected for in vitro crude enzyme activity assays, and the catalytic activity of the enzymes was characterized by the specific synthetic yield (μM) of 5-FDA.

[0177] 3. Experimental Results and Analysis: (e.g.) Figure 15 and Figure 16 As shown, all four generative designs tested exhibited definite catalytic activity (producing 5-FDA), with three designs (GenFIA-1, GenFIA-3, and GenFIA-4) showing significantly higher activity and yield than the wild-type reference enzyme FlAMA37 (yield of 18.0 μM). The optimal design (GenFIA-4, yield of 22.1 μM) achieved an activity improvement of approximately 24.7% relative to the reference enzyme. Figure 15 As shown, FlAMA37 has a complex trimer structure and contains three color-coded catalytic pockets. By comparison, it can be seen that these performance improvements are achieved by optimizing the peripheral auxiliary folding sequence while maintaining the consistency of the core catalytic functional pockets, which verifies the efficient optimization capability brought about by the explicit injection of evolutionary priors in this invention.

[0178] Experiment Example 6: Validation of Ablation and Leakage Prevention Mechanisms for Model Training Strategies 1. Experimental Objective: This experiment aims to verify the actual technical effect of the unique training strategy adopted in this invention (including the anti-leakage mechanism for removing query sequences, dynamic mask ratio, and frozen structure modules) on improving the generalization ability of the model and preventing "catastrophic forgetting", thus providing effect support for the model training method of this application. 2. Experimental methods: (1) Set up a comparison training group: Train three versions of the model on the same dataset: Version A (using the complete training strategy of this invention), Version B (canceling the anti-leakage mechanism, i.e., not removing the baseline model of the query sequence during training), and Version C (unfreezing the structure generation module during training and performing full parameter fine-tuning).

[0179] (2) Test set evaluation: Select a new set of orphan protein targets with very low homology (sequence identity <30%) to be tested. (3) Evaluation indicators: The generalization and overfitting of the model are evaluated by statistically analyzing the "functional hit rate" on new targets; the maintenance of the structural space prior is examined by evaluating the Laplace plot anomaly rate and atomic physical collision rate of the generated maskless main chain. 3. Experimental Results and Analysis: Test results show that version A significantly outperforms version B in terms of functional hit rate and sequence recovery accuracy when facing novel targets. Version B, due to "copying" the query sequence containing complete information during training, falls into severe overfitting and shortcut dependence. This results in its inability to effectively utilize the co-evolutionary information in MSA for comprehensive inference when encountering novel proteins it has never seen before, leading to a precipitous drop in generation success rate. This demonstrates that the leakage prevention mechanism of "removing the query sequence and synchronizing the mask" in this invention is key to endowing the model with strong generalization and inference capabilities.

[0180] Meanwhile, comparing the structural quality metrics of versions A and C revealed that version C (the unfrozen structural module) generated a large number of unreasonable physical collisions and Laplace graph anomalies in its main chain structure, indicating that it lost its original three-dimensional geometric perception capability during joint training (i.e., catastrophic forgetting occurred); while version A (the frozen structural module) perfectly maintained extremely high skeleton geometric rationality. This objectively confirms the significant advantages of the asymmetric frozen training strategy of this invention, which "only jointly optimizes the sequence module and encoder," in protecting spatial priors and improving training stability.

[0181] refer to Figure 17 This application also provides a protein sequence-structure co-generation device, comprising: The data encoding module 10 is used to acquire multiple sequence alignment data of the target protein and encode the multiple sequence alignment data to obtain an explicit evolutionary prior representation containing monomeric site features and site pair features. Initialization module 20 is used to initialize the current sequence state and current structural state of the target protein; wherein, the current sequence state includes masked sites; The cooperative iteration module 30 is used to perform at least one round of cooperative iteration generation process based on the explicit evolutionary prior representation, so as to alternately update the current sequence state and the current structural state until all masked sites in the current sequence state are demasked to obtain the final updated current sequence state, and use the final updated current sequence state as the protein sequence of the target protein. The structure generation module 40 is used to generate the final three-dimensional structure based on the protein sequence and output the protein sequence and the final three-dimensional structure.

[0182] It is understood that the apparatus in this embodiment corresponds to the protein sequence-structure co-generation method in the above embodiments, and the options in the above embodiments are also applicable to this embodiment, so they will not be described again here.

[0183] This application also provides a computer device, which includes a processor and a memory. The memory stores a computer program, and the processor is used to execute the computer program to implement the protein sequence-structure co-generation method described in any of the foregoing embodiments.

[0184] The processor can be an integrated circuit chip with signal processing capabilities. The processor can be a general-purpose processor, including at least one of a Central Processing Unit (CPU), Graphics Processing Unit (GPU), Network Processor (NP), Digital Signal Processor (DSP), Application-Specific Integrated Circuit (ASIC), Field-Programmable Gate Array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. The general-purpose processor can be a microprocessor or any conventional processor, capable of implementing or executing the methods, steps, and logic block diagrams disclosed in the embodiments of this application.

[0185] The memory can be, but is not limited to, Random Access Memory (RAM), Read Only Memory (ROM), Programmable Read-Only Memory (PROM), Erasable Programmable Read-Only Memory (EPROM), Electrically Erasable Programmable Read-Only Memory (EEPROM), etc. The memory is used to store computer programs, and the processor can execute the computer programs accordingly after receiving execution instructions.

[0186] This application also provides a computer storage medium storing a computer program, which, when executed on a processor, implements the protein sequence-structure co-generation method according to any one of the foregoing embodiments.

[0187] The computer storage medium can be a readable storage medium, a non-volatile storage medium, or a volatile storage medium. For example, the computer storage medium may include, but is not limited to, various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0188] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can also be implemented in other ways. The apparatus embodiments described above are merely illustrative. For example, the flowcharts and block diagrams in the accompanying drawings show the architecture, functionality, and operation of possible implementations of apparatus, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that, in alternative implementations, the functions marked in the blocks may occur in a different order than those marked in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagram and / or flowchart, and combinations of blocks in the block diagram and / or flowchart, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

[0189] In addition, the functional modules or units in the various embodiments of this application can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.

[0190] If the aforementioned functions are implemented as software functional modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a smartphone, personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application.

[0191] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A method for co-generating protein sequence-structure, characterized in that, include: Multiple sequence alignment data of the target protein are acquired and encoded to obtain an explicit evolutionary prior representation containing monomeric site features and site pair features. Initialize the current sequence state and current structural state of the target protein; wherein the current sequence state includes masked sites; for the masked sites in the current sequence state, only N and C are retained in the current structural state. α C and O main chain atoms; for non-masked sites, all atoms are retained in the current structural state; Based on the explicit evolutionary prior representation, at least one round of cooperative iterative generation is performed to alternately interleave and update the current sequence state and the current structural state until all masked sites in the current sequence state are demasked, resulting in the final updated current sequence state. This final updated current sequence state is then used as the protein sequence of the target protein. In each round of the cooperative iterative generation process, the alternating interleave and update includes: Based on the explicit evolutionary prior representation, the current sequence state, and the current structural state, the current round structural state is generated; Based on the current sequence state, the explicit evolutionary prior representation, and the current sequence state, predict the amino acid probability distribution corresponding to the sites that are still masked in the current sequence state; Based on the current structural state, determine the structural neighborhood information corresponding to the still-masked sites, and based on the structural neighborhood information, determine the set of demasked sites for this round from the still-masked sites, including: for each site i in the current set of still-masked sites I, calculate its spatial neighborhood number N. i The calculation formula for the structural neighborhood information is as follows: ； Among them, Cα i and Cα j These are the three-dimensional coordinates of the Cα atoms at sites i and j, respectively. d represents Euclidean distance; thr The preset spatial neighborhood threshold; For all sites in set I, according to the spatial neighborhood number N i Sort in descending order and select the top r of the sorted results. t ×N sites form the current round's unmasking site set; where N is the total amino acid length of the target protein; r t The preset demasking ratio for the current round; Based on the amino acid probability distribution, the residue type of each point in the current round of demasking site set is determined, and the current sequence state is updated using the determined residue type, and the current structural state is updated using the current round structural state; The final three-dimensional structure is generated based on the protein sequence, and the protein sequence and the final three-dimensional structure are output.

2. The protein sequence-structure co-generation method as described in claim 1, characterized in that, Before encoding the multi-sequence alignment data, the data is further cleaned. The cleaning includes removing homologous sequences that are duplicates of the query sequence and removing sequences with a gap ratio higher than 0.6, so that the effective depth of the cleaned multi-sequence alignment matrix is ​​not less than 32.

3. The protein sequence-structure co-generation method as described in claim 1, characterized in that, In the explicit evolutionary prior representation that includes single-site features and site-pair features: the single-site features are the dimensionality-reduced single-site feature tensor S. msa Its feature dimension d s =256; the site pair features are the reduced-dimensional pairwise feature tensors Z. msa Its feature dimension d z =128.

4. The protein sequence-structure co-generation method as described in claim 1, characterized in that, The protein sequence-structure co-generation method supports a variety of different task modes; The task modes include at least one of the following: single-unit design mode, support design mode, combined design mode, and specific function modification mode.

5. The protein sequence-structure co-generation method as described in claim 1, characterized in that, The step of determining the residue type of each site in the current round of demasking site set according to the amino acid probability distribution includes: performing temperature sampling on each site in the current round of demasking site set using a preset sequence sampling temperature τ according to the amino acid probability distribution, so as to determine the residue type of each site.

6. The protein sequence-structure co-generation method as described in claim 1, characterized in that, The total number of iterations in the collaborative iterative generation process is 6 to 12.

7. The protein sequence-structure co-generation method as described in claim 1, characterized in that, In each round of the collaborative iterative generation process, the number of inversion steps for structure diffusion generation is 100 to 300 steps.

8. The protein sequence-structure co-generation method as described in claim 4, characterized in that, When the task mode is a composite design mode, when determining the set of demasking sites for this round, the priority of selecting interface neighborhood sites at the composite interaction interface into the set of demasking sites for this round is increased.

9. The protein sequence-structure co-generation method as described in claim 1, characterized in that, The spatial neighborhood threshold d thr The value range is 8 Å to 12 Å.

10. The protein sequence-structure co-generation method as described in claim 1, characterized in that, The single-round demasking ratio r t The value range is 0.01 to 0.

20.

11. The protein sequence-structure co-generation method as described in claim 5, characterized in that, The value range of the sequence sampling temperature τ is 0.6~1.

0.

12. The protein sequence-structure co-generation method as described in claim 3, characterized in that, For the target protein with a total amino acid length of N. The single-unit feature tensor: The i-th position is represented as: ； The paired feature tensors: The i-th row and j-th column is represented as: ； Where M is the number of homologous sequences; E m,i w is the encoding vector of the m-th homologous sequence at the i-th position; m For sequence weights; W z It is a linear projection matrix; This represents the tensor outer product.

13. The protein sequence-structure co-generation method as described in claim 12, characterized in that, In each round of the aforementioned collaborative iterative generation process, the single-unit feature tensor S msa and the paired feature tensor Z msa Injected into the structural backbone network, and related to the current single-unit representation H of the structural backbone network. l And the current pairwise representation P l Perform a merge update; the merge formula is: ； Where Ws and Wp are linear projection matrices; PairFormer is the feature fusion and update module; H l This represents the current singleton; P l For the current paired representation; H l+1 is the updated single-unit representation; l represents the hierarchical index of the backbone network of the structure.

14. The protein sequence-structure co-generation method as described in claim 13, characterized in that, The current pairwise representation P l The paired feature elements at the i-th and j-th sites are represented by P. ij The paired feature element P ij By fusing the conditional input elements with the geometric distance embedding of the current diffusion step through a gating mechanism, the conditional input elements for the structure denoising network are obtained. The fusion formula is: ； in, ; The geometric embedding is constructed based on interatomic distances; σ is the gating function; ⊙ represents the Hadamard product; W d and W g r is a linear mapping matrix; i and r j For the three-dimensional coordinates of the atom; For the paired feature tensor Z msa The paired features in the i-th row and j-th column of the array.

15. A protein sequence-structure co-generation device, characterized in that, include: The data encoding module is used to acquire multiple sequence alignment data of the target protein and encode the multiple sequence alignment data to obtain an explicit evolutionary prior representation containing monomeric site features and site pair features. An initialization module is used to initialize the current sequence state and current structural state of the target protein; wherein the current sequence state includes masked sites; for the masked sites in the current sequence state, only N and C are retained in the current structural state. α C and O main chain atoms; for non-masked sites, all atoms are retained in the current structural state; A cooperative iterative module is used to perform at least one round of cooperative iterative generation based on the explicit evolutionary prior representation, to alternately interleave and update the current sequence state and the current structural state until all masked sites in the current sequence state are demasked, resulting in a final updated current sequence state, which is then used as the protein sequence of the target protein. In each round of the cooperative iterative generation process, the alternating interleave and update includes: Based on the explicit evolutionary prior representation, the current sequence state, and the current structural state, the current round structural state is generated; Based on the current sequence state, the explicit evolutionary prior representation, and the current sequence state, predict the amino acid probability distribution corresponding to the sites that are still masked in the current sequence state; Based on the current structural state, determine the structural neighborhood information corresponding to the still-masked sites, and based on the structural neighborhood information, determine the set of demasked sites for this round from the still-masked sites, including: for each site i in the current set of still-masked sites I, calculate its spatial neighborhood number N. i The calculation formula for the structural neighborhood information is as follows: ； Among them, Cα i and Cα j These are the three-dimensional coordinates of the Cα atoms at sites i and j, respectively. d represents Euclidean distance; thr The preset spatial neighborhood threshold; For all sites in set I, according to the spatial neighborhood number N i Sort in descending order and select the top r of the sorted results. t ×N sites form the current round's unmasking site set; where N is the total amino acid length of the target protein; r t The preset demasking ratio for the current round; Based on the amino acid probability distribution, the residue type of each point in the current round of demasking site set is determined, and the current sequence state is updated using the determined residue type, and the current structural state is updated using the current round structural state; The structure generation module is used to generate the final three-dimensional structure based on the protein sequence and output the protein sequence and the final three-dimensional structure.

16. A computer device, characterized in that, The computer device includes a processor and a memory, the memory storing a computer program, and the processor executing the computer program to implement the protein sequence-structure co-generation method according to any one of claims 1-14.

17. A computer storage medium, characterized in that, It stores a computer program that, when executed on a processor, implements the protein sequence-structure co-generation method according to any one of claims 1-14.

Images