Biomacromolecule sequence structure prediction method

Abstract

The invention discloses a biomacromolecule sequence structure prediction method, and relates to the technical field of computers. The attention weights corresponding to different attention types of the processed target information are broadcasted in different broadcast range modes by using a multi-scale attention broadcast mechanism, and the attention weights in different time steps in the broadcast range are repeatedly used, so that the same attention weight is adopted in the steps of the broadcast range, the calculation amount is reduced, and the efficiency is improved. The efficiency performance of the model is improved, and the perception ability of the model to amino acid position information in a protein sequence can be effectively improved, so that the model can fully consider the position relationship between amino acids when calculating the attention weight, and the model can be helped to better identify the interaction of long-chain amino acids; the structure prediction error caused by the missing of the position information is reduced, and the prediction precision is improved.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to a method for predicting the sequence structure of biological macromolecules. Background Technology

[0002] Traditional protein structure determination methods, such as X-ray crystallography and nuclear magnetic resonance spectroscopy, while providing high-precision structural information, suffer from drawbacks such as long experimental cycles, high costs, and complex operations, making them unsuitable for large-scale protein structure determination. For large-scale protein structure determination, the artificial intelligence (AI) software AlphaFold3, developed using related technologies, has successfully predicted the structure and interactions of almost all biological macromolecules (proteins, DNA, RNA, ligands, etc.) with unprecedented accuracy. This ability to accurately predict and analyze the complex interactions between proteins and other molecules using computers can bring new insights to fields such as disease pathways, genomics, therapeutic targets, protein engineering, and synthetic biology. However, the original diffusion model used in AlphaFold3 suffers from a decrease in prediction accuracy as amino acid length increases during iterative denoising. This is because traditional attention mechanisms struggle to maintain long-chain amino acid interactions. When calculating attention weights, traditional attention mechanisms fail to accurately capture the dependencies between long-chain amino acids as sequence length increases, leading to a decrease in the overall structure prediction accuracy for long protein sequences. Summary of the Invention

[0003] This application provides a method for predicting the sequence structure of biological macromolecules, which at least solves the problem in related technologies where diffusion models, during iterative denoising, struggle to accurately capture the dependencies between long-chain amino acids when calculating attention weights as the amino acid length increases, resulting in decreased prediction accuracy.

[0004] This application provides a method for predicting the sequence structure of biological macromolecules, including: The geometric position information of amino acids in the sequence of biological macromolecules is encoded to determine the dynamic geometric position encoding of the relative spatial relationship between multiple residue pairs. The weight of the dynamic geometric position encoding is adjusted by a gating signal to generate a gated geometric position encoding. The geometric position encoding is combined with the attention mechanism to determine the attention weights that adjust the positional relationships between amino acids, and the broadcast range of the attention weights is adjusted according to the attention type of the target information being processed. The residue positions are corrected based on the geometric position encoding, the attention weight, and the broadcast range of the attention weight, and the geometric position encoding is updated according to the corrected residue positions; The predicted biomolecular sequence structure was determined based on the updated geometric position coding.

[0005] This application introduces a geometric position encoding gating mechanism during the diffusion process and utilizes a multi-scale attention broadcasting (PAB) mechanism to broadcast attention weights corresponding to different attention types of the processed target information in different broadcast ranges. Attention weights are reused in different time steps within the broadcast range, achieving the use of the same attention weights within the steps of the broadcast range. This reduces the amount of computation, improves model efficiency and performance, and effectively enhances the model's ability to perceive amino acid position information in protein sequences. This allows the model to fully consider the positional relationships between amino acids when calculating attention weights, helping the model to better identify long-chain amino acid interactions, reducing structural prediction errors caused by missing positional information, and improving prediction accuracy. Attached Figure Description

[0006] To more clearly illustrate the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0007] Figure 1 This is a schematic diagram illustrating the principle that the original diffusion model used in AlphaFold3 has difficulty maintaining long-chain amino acid interactions. Figure 2 This is a diagram illustrating the application environment of the biomacromolecule sequence structure prediction method in one embodiment of this application; Figure 3 This is a logical diagram of the attention weights corresponding to each attention type in a broadcast process that incorporates a geometric position encoding gating mechanism during the diffusion process, as described in one embodiment of this application. Figure 4 This is a flowchart illustrating a method for predicting the sequence structure of biological macromolecules in one embodiment of this application. Figure 5 This is a structural block diagram of a biomacromolecule sequence structure prediction device in one embodiment of this application; Figure 6 This is an internal structural diagram of a computer device in one embodiment of this application. Detailed Implementation

[0008] The following will be combined with the accompanying drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of them. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0009] It should be noted that, in the description of this application, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. The terms "first," "second," etc., in this application are used to distinguish similar objects and are not used to describe a specific order or sequence.

[0010] To enable those skilled in the art to better understand the present application, the present application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0011] Every plant, animal, and human cell in nature contains billions of molecular machines. These machines are composed of proteins, DNA, RNA, and other ligand molecules. It is these tiny machines made up of biological macromolecules that sustain and perpetuate life. Essentially, life is built on molecular-level structural support and the interactions between molecules. Proteins are the main carriers of life activities, and their functions are closely related to their three-dimensional structure. Understanding the three-dimensional structure of proteins helps to reveal their mechanisms of action within organisms, such as the catalytic function of enzymes and the immune recognition function of antibodies. Traditional methods for determining protein structure, such as X-ray crystallography and nuclear magnetic resonance spectroscopy, while providing high-precision structural information, suffer from drawbacks such as long experimental cycles, high costs, and complex operations, making them unsuitable for large-scale protein structure determination.

[0012] On May 9, 2024, researchers from DeepMind and Isomorphic Labs developed a new artificial intelligence (AI) software called AlphaFold3, which successfully predicted the structure and interactions of almost all biological macromolecules (proteins, DNA, RNA, ligands, etc.) with unprecedented accuracy.

[0013] This ability to precisely predict and analyze the complex interactions between proteins and other molecules using computers can help bring new insights into fields such as disease pathways, genomics, therapeutic targets, protein engineering, and synthetic biology. More importantly, AlphaFold3 opens up exciting possibilities for drug development and has the potential to revolutionize traditional drug discovery models.

[0014] The original diffusion model used in AlphaFold3 suffers from a decrease in prediction accuracy as the amino acid length increases during iterative denoising, as shown in the attached figure. Figure 1As shown. This is because traditional attention mechanisms struggle to maintain long-chain amino acid interactions. In long-sequence proteins, interactions between amino acids include not only short-range interactions but also long-range interactions. These long-range interactions are crucial for maintaining the overall structural stability and function of the protein. When calculating attention weights, traditional attention mechanisms struggle to accurately capture the dependencies between long-chain amino acids as sequence length increases, leading to a decrease in the accuracy of overall structure prediction for long-sequence proteins.

[0015] To address the issue of decreased prediction accuracy of the AlphaFold3 model for long-chain amino acids during iterative denoising, this application provides a method for predicting the sequence structure of biomolecules, which can be applied to applications such as... Figure 1 In the application environment shown, terminal 102 and server 104 communicate via a network. Terminal 102 can input biomolecule sequences, ligands, and covalent bond information into server 104. Server 104 processes the input biomolecule sequences, ligands, and covalent bond information, predicts the biomolecule sequence structure, and outputs the result. Terminal 102 can be, but is not limited to, various personal computers, laptops, smartphones, tablets, and portable wearable devices. Server 104 can be a standalone server or a server cluster consisting of multiple servers.

[0016] like Figure 3 As shown, this application introduces a geometric position encoding gating mechanism during the diffusion process and utilizes a multi-scale attention broadcasting (PAB) mechanism to employ different broadcast ranges for calculating attention weights with different frequencies of change. The broadcast range refers to the reuse of attention weights across different time steps, ensuring the same attention weights are used within each broadcast range step. This mechanism not only effectively enhances the model's ability to perceive amino acid position information in protein sequences, thereby improving the accuracy of structure prediction, but also reduces computational cost and improves model efficiency and performance. The geometric position encoding gating mechanism encodes the geometric position information of amino acids and combines it with the attention mechanism, enabling the model to fully consider the positional relationships between amino acids when calculating attention weights. This mechanism helps the model better identify long-chain amino acid interactions and reduces structure prediction errors caused by missing positional information.

[0017] Among them, the multi-scale attention broadcasting mechanism (PAB) accelerates the process by using the same attention weights across steps within the broadcast range, reducing redundant attention computations. Attention differences across different time steps exhibit a U-shaped pattern, with minimal differences in intermediate steps; different types of attention show varying stability and variability. Based on this, PAB proposes a pyramid-shaped attention broadcasting mechanism, setting different broadcast ranges according to the stability and variability of different attention types to avoid redundant attention computations.

[0018] like Figure 4 As shown, embodiments of this application provide a method for predicting the sequence structure of biological macromolecules, including the following steps: Step S1: Encode the geometric position information of amino acids in the biomacromolecule sequence, determine the dynamic geometric position encoding of the relative spatial relationship between multiple residue pairs, and generate the gated geometric position encoding by adjusting the weight of the dynamic geometric position encoding through a gating signal. Step S2: Combine geometric position encoding with attention mechanism to determine attention weights that adjust the positional relationship between amino acids, and adjust the broadcast range of attention weights according to the attention type of the target information being processed; Step S3: Correct the residue position based on the geometric position encoding, attention weight, and the broadcast range of the attention weight, and update the geometric position encoding according to the corrected residue position; Step S4: Determine the predicted biomacromolecule sequence structure based on the updated geometric position encoding.

[0019] This application introduces a geometric position encoding gating mechanism during the diffusion process and utilizes a multi-scale attention broadcasting (PAB) mechanism to broadcast attention weights corresponding to different attention types of the target information being processed in different broadcast ranges. Attention weights are reused in different time steps within the broadcast range, achieving the use of the same attention weights within the steps of the broadcast range. This reduces the amount of computation, improves model efficiency and performance, and effectively enhances the model's ability to perceive amino acid position information in protein sequences. This allows the model to fully consider the positional relationships between amino acids when calculating attention weights, helping the model to better identify long-chain amino acid interactions, reducing structural prediction errors caused by missing positional information, and improving prediction accuracy.

[0020] Furthermore, this application can broadcast the attention weights corresponding to different attention types of the target information being processed using different broadcast range methods, which can fully consider the interactions of long-chain amino acids and accurately calculate the attention weights for long-chain amino acids. Moreover, the calculation methods for attention weights corresponding to different attention types are different, which effectively improves the computational resources for attention weights for long-chain amino acids and improves the overall prediction accuracy and efficiency of biological macromolecular sequence structures.

[0021] By integrating spatial geometric information with sequence interactions, and progressively refining the structure through a diffusion model, accurate prediction of the sequence structure of biological macromolecules can be achieved.

[0022] In this embodiment, the geometric position information of amino acids in a biomacromolecule sequence is encoded to determine the dynamic geometric position encoding of the relative spatial relationship between multiple residue pairs. The gating signal is used to adjust the weights of the dynamic geometric position encoding to generate a gated geometric position encoding, including: It receives input biomolecule sequences, ligands, and covalent bond information, and generates initial residue feature representations of the biomolecule sequences. The starting three-dimensional coordinates of the generated biomacromolecule sequence serve as the starting point for dynamic geometric position coding; Based on the initial three-dimensional coordinate points, the three-dimensional coordinates of the residues are mapped to high-dimensional features through linear transformation. Combined with the noise perturbation process of the diffusion model, an encoding vector characterizing the relative spatial relationship between residue pairs is generated as the dynamic geometric position encoding between multiple residue pairs. Generate one or more gating signals, which are used to adjust the weights of the dynamic geometric position encoding; The gating signal is applied to the dynamic geometric position code to generate a gating-adjusted geometric position code.

[0023] The conformer generation module generates initial three-dimensional coordinates for each residue. The generation of the gating signal aims to suppress noise interference during diffusion denoising and enhance the spatial relationships between key residues. The gating signal is used to adjust the weights of the dynamic geometric position encoding to selectively control its influence on subsequent feature updates or attention mechanisms; wherein, the geometric position encoding includes, but is not limited to: discretized encoding of the relative distance between the central atoms (e.g., Cα atoms) of residue pairs; encoding of the relative orientation of the local coordinate systems between residue pairs, for example, represented by rotation matrices or quaternions; and encoding of the relative angles or torsion angles (e.g., dihedral angles) formed by the key atoms (e.g., Cα, C, N atoms) between residue pairs.

[0024] The process involves receiving initial input, generating an initial structural representation, dynamically calculating the geometric relationships between residues using diffuse noise, and adjusting the weights of this geometric information through a gating mechanism to provide accurate and controlled geometric information for subsequent attention mechanisms.

[0025] In this embodiment, combining geometric position encoding with an attention mechanism to determine the attention weights for adjusting the positional relationships between amino acids includes: Geometric position encoding is injected as a logarithmic bias term into the attention score calculation of the attention mechanism, and the attention score between amino acid residue pairs is used as the attention weight to adjust the positional relationship between amino acids.

[0026] In this method, by injecting geometric encoding as a logarithmic bias term into the attention score calculation, the attention between amino acid residue pairs is directly adjusted, thereby enhancing the model's ability to perceive the interactions between neighboring atoms / residues in local space.

[0027] In this embodiment, adjusting the broadcast range of attention weights according to the attention type of the processed target information includes: Determine the frequency of change or rate of change of the target information being processed during the iterative denoising process; Determine the attention type corresponding to the target information based on the frequency of change or the rate of change of the impact of the target information, and set the broadcast ratio of the attention type corresponding to the target information. The broadcast range is controlled based on the broadcast ratio, which determines the attention weight corresponding to the target information.

[0028] Specifically, by evaluating the frequency of change or rate of change of the target information processed during iterative denoising, the attention type and broadcast ratio are adaptively determined, thereby optimizing the information propagation efficiency and ensuring that different types of information receive a reasonable range of attention in the model.

[0029] In this embodiment, the attention type corresponding to the target information is determined based on the frequency of change or the rate of change of influence of the target information, and the broadcast ratio of the attention type corresponding to the target information is set as follows: When the target information is to process atomic-level local structural information and capture instantaneous changes in hydrogen bonds or van der Waals forces, the target information is classified as spatial attention feature information. When the target information is sequence evolution information, the target information is classified as temporal attention feature information; When the target information is the ligand-protein binding pattern in the global structure generation stage, the target information is classified as cross-modal attention feature information. The broadcast ratio of spatial attention features is set to be less than that of temporal attention features, and the broadcast ratio of temporal attention features is set to be less than that of cross-modal attention features.

[0030] Based on the nature of the information being processed (atomic-level local structure, sequence evolution, cross-modal combination), attention is categorized into three types: spatial, temporal, and cross-modal. An increasing broadcast ratio is set to match the different information types' needs for propagation range.

[0031] In this embodiment, controlling the broadcast range of the attention weight corresponding to the target information based on the broadcast ratio includes: The total number of diffusion steps when setting the attention weights corresponding to spatial attention features for a single broadcast is less than the total number of diffusion steps when setting the attention weights corresponding to temporal attention features for a single broadcast. The total number of diffusion steps when setting the attention weights corresponding to temporal attention features for a single broadcast is less than the total number of diffusion steps when setting the attention weights corresponding to cross-modal attention features for a single broadcast. In response to a single broadcast of the target attention weights, determine at each time step within the total number of diffusion steps whether the current time step is within the broadcast range; In response to the current time step being within the broadcast range, control the reuse of the target attention weight of the previous time step in the next time step; In response to the current time step not being within the broadcast range, the attention weight for the next time step is recalculated according to the attention mechanism, and the attention weight for the next time step is saved for broadcasting within the total number of diffusion steps of the next single broadcast.

[0032] Specifically, by setting the total number of diffusion steps for a single broadcast for different attention types, and determining at each time step whether it is within the broadcast range, the system decides whether to reuse existing weights or recalculate them, in order to balance computational efficiency and accuracy and avoid unnecessary duplicate calculations.

[0033] like Figure 3 As shown, the time steps for broadcasting and propagating the attention weights corresponding to spatial attention features, temporal attention features, and cross-modal attention features include: Let the current time step be X. t Then, the remaining time steps within the total number of diffusion steps in a single broadcast are X. t-1 X t-2 X t-3 , ...X t-n Where tn=1, and n is an integer; Following the amino acid sequence order that generates the biological macromolecular sequence structure, control the spatial attention feature information from the current time step X. t The corresponding attention weights are broadcast sequentially to control the temporal attention feature information from the current time step X. t The corresponding attention weights are broadcast sequentially to control the cross-modal attention feature information from the current time step X. t The corresponding attention weights are broadcast sequentially.

[0034] Because the total number of diffusion steps for a single broadcast of attention weights corresponding to spatial attention features is less than the total number of diffusion steps for a single broadcast of attention weights corresponding to temporal attention features, and the total number of diffusion steps for a single broadcast of attention weights corresponding to temporal attention features is less than the total number of diffusion steps for a single broadcast of attention weights corresponding to cross-modal attention features, Figure 3 The example illustrates how the attention weights corresponding to spatial attention feature information change from the current time step X. t Broadcast to time step X t-1 The attention weights corresponding to the time attention feature information change from the current time step X. t Broadcast to time step X t-2 The attention weights corresponding to cross-modal attention feature information change from the current time step X. t Broadcast to time step X t-3 .

[0035] In this embodiment, controlling the broadcast range of the attention weight corresponding to the target information based on the broadcast ratio further includes: During the broadcasting of attention weights corresponding to target information, the computational resource load rate is obtained. Adjust the broadcast ratio of spatial attention feature information, temporal attention feature information, and cross-modal attention feature information based on the computing resource load rate to adjust the broadcast range.

[0036] Specifically, when broadcasting complex target information, the broadcast frequency of corresponding attention feature information is automatically increased; while for simple target information, the broadcast ratio is reduced, thereby improving efficiency while ensuring quality.

[0037] In this embodiment, when broadcasting spatial attention feature information, temporal attention feature information, and cross-modal attention feature information, the method further includes: Obtain spatial attention feature information, temporal attention feature information, and cross-modal attention feature information for broadcasting at the current time step; Set the spatial attention feature information, temporal attention feature information, and cross-modal attention feature information for broadcasting to either group broadcast mode or hierarchical broadcast mode.

[0038] In addition to a single broadcast mechanism, we are exploring different types of broadcast mechanisms, such as group broadcast mode and hierarchical broadcast mode, to uncover greater acceleration potential.

[0039] In this embodiment, determining whether the current time step is within the broadcast range at each time step within the total number of diffusion steps in response to a single broadcast of the target attention weight includes: At the current time step, obtain the current diffusion step of the target attention weight, obtain the total number of single diffusion steps for broadcasting the target attention weight, and determine the real-time diffusion step position based on the current diffusion step and the total number of single diffusion steps. Calculate the U-shaped stability curve based on the real-time diffusion step location; If the following conditions are met: the next diffusion step is within the total number of diffusion steps in a single run, the U-shaped stability curve is greater than the first threshold, and the attention stability is greater than the attention stability threshold, then the current time step is determined to be within the broadcast range; otherwise, the current time step is determined not to be within the broadcast range.

[0040] Among them, a method for calculating the U-shaped stability curve based on the real-time diffusion step position is introduced, and combined with the attention stability threshold for comprehensive judgment, so as to make a smarter and more robust decision on whether to reuse attention weights, ensuring that broadcasting is carried out in the stage where attention weights are stable.

[0041] In this embodiment, determining the real-time diffusion step location based on the current diffusion step and the total number of diffusion steps in a single run includes: Obtain the current diffusion step and the total number of diffusion steps in a single run; The location of the real-time diffusion step is determined based on the quotient of the current diffusion step and the total number of diffusion steps in a single run.

[0042] Specifically, the real-time diffusion step position is calculated by normalized_step = step / total_steps, where normalized_step is the real-time diffusion step position, step is the current diffusion step, and total_steps is the total number of diffusion steps in a single run.

[0043] Standardizing the current diffusion progress provides a unified input for the subsequent calculation of the U-shaped stability curve.

[0044] In this embodiment, calculating the U-shaped stability curve based on the location of the real-time diffusion step includes: Obtain the absolute value of the difference between the location of the real-time diffusion step and the first threshold; Calculate the product of the absolute value and the second threshold; The U-shaped stability curve is determined based on the difference between the third threshold and the product.

[0045] Specifically, the U-shaped stability curve is calculated using stability_curve = 1.0-2.0 * abs(normalized_step-0.5), where stability_curve is the U-shaped stability curve and abs() is the absolute value function.

[0046] The first threshold is preferably 0.5, the second threshold is preferably 2.0, and the third threshold is preferably 1.0. These values ​​are used to quantitatively evaluate the stability of the attention weights in the current diffusion step, serving as a key basis for determining whether to broadcast.

[0047] In this embodiment, controlling the broadcast range of the attention weight corresponding to the target information based on the broadcast ratio includes: Obtain the total number of diffusion steps in a single run and the broadcast ratio; The first product is obtained based on the total number of diffusion steps in a single run and the broadcast ratio; The first product is rounded down to determine the broadcast range of the attention weight corresponding to the target information.

[0048] Specifically, the broadcast range of the attention weights corresponding to the target information is calculated by broadcast_range = int(total_steps*broadcast_ratio), where broadcast_range is the broadcast range, int() is the rounding function, total_steps is the total number of diffusion steps in a single broadcast, and broadcast_ratio is the broadcast ratio.

[0049] Among them, the effective propagation distance of attention weight in the diffusion process is quantitatively determined based on the total number of diffusion steps and the broadcast ratio.

[0050] In this embodiment, setting the broadcast ratio of the attention type corresponding to the target information further includes: Obtain the target attention type corresponding to the target information, and obtain the confidence score for broadcasting the attention weights corresponding to the target information when the target attention type is selected; If the confidence score is less than the preset confidence score threshold, the broadcast ratio of the target attention type corresponding to the target information is increased.

[0051] Specifically, when the broadcast confidence score of the target attention weight is lower than the preset confidence score threshold (preferably 0.7), its broadcast ratio is increased to improve the information update frequency, thereby prompting the model to re-evaluate and update more frequently when the model is uncertain about the current attention weight.

[0052] In this embodiment, the residue position is corrected based on the geometric position encoding, attention weight, and the broadcast range of the attention weight. Updating the geometric position encoding according to the corrected residue position includes: During the iterative denoising process, the initial three-dimensional coordinates of the corresponding amino acid sequence atoms in the geometric position encoding are obtained; At each iteration, the initial three-dimensional coordinates and amino acid sequence pairing information are input into the diffusion model. The diffusion model combines geometric position encoding and attention weights to identify and predict noise vectors. The noise vector is subtracted from the initial 3D coordinates to correct the residue positions, and the geometric position encoding is updated based on the corrected 3D coordinates.

[0053] In this process, new noise is introduced in each iteration to explore various biomolecular sequence structures until the target three-dimensional coordinates of the corresponding amino acid sequence atoms are obtained. By continuously removing existing noise and introducing new noise through an iterative denoising process, the diversity of biomolecular sequence structures can be increased, the optimal structure can be identified, the target three-dimensional coordinates of the corresponding amino acid sequence can be obtained, and the accuracy of structure prediction can be improved.

[0054] In this embodiment, determining the predicted biomacromolecule sequence structure based on the updated geometric position encoding includes: By using a distance-constrained network, the target's three-dimensional coordinates are optimized based on the predicted distance matrix to conform to the molecular force field rules, thus obtaining the optimized target's three-dimensional coordinates. The predicted biomolecular sequence structure is generated based on the optimized target three-dimensional coordinates.

[0055] In this method, by introducing a distance constraint network, the three-dimensional coordinates of the target are optimized based on the predicted distance matrix to ensure that the predicted structure conforms to the rules of molecular mechanics and physical constraints, thereby improving the quality and biological rationality of the predicted structure.

[0056] In the aforementioned method for predicting the sequence structure of biological macromolecules, this application introduces a geometric position encoding gating mechanism during the diffusion process and utilizes a multi-scale attention broadcasting (PAB) mechanism to broadcast attention weights corresponding to different attention types of the processed target information in different broadcast ranges. Attention weights are reused in different time steps within the broadcast range, achieving the use of the same attention weights within the steps of the broadcast range. This reduces the amount of computation, improves model efficiency and performance, and effectively enhances the model's ability to perceive amino acid position information in protein sequences. This allows the model to fully consider the positional relationships between amino acids when calculating attention weights, helping the model to better identify long-chain amino acid interactions, reducing structure prediction errors caused by missing positional information, and improving prediction accuracy.

[0057] This application implements a two-level architecture for the diffusion module of the AlphaFold3 model: first processing atomic coordinates, then processing token representations, and finally regressing to atomic coordinates. PAB can establish cross-level attention bridges during this process. For example, at the atomic level, local attention captures chemical details such as bond lengths and bond angles; at the token level, global attention integrates sequence evolution information and broadcasts stable attention results (such as interaction patterns of conserved domains) to the atomic level to guide local structure optimization. This cross-level interaction enhances the synergy of features at different levels of abstraction, improving the coherence of structure prediction. Experiments show that the AlphaFold3 model with PAB improves the predicted residue pLDDT score by approximately 5 to 10 points.

[0058] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods according to the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method.

[0059] In one embodiment, such as Figure 5 As shown, a biomacromolecule sequence structure prediction device 10 is provided, including: a geometric position encoding module 1, a broadcast range management module 2, a correction module 3, and a biomacromolecule generation module 4.

[0060] The geometric position encoding module 1 is used to encode the geometric position information of amino acids in the sequence of biological macromolecules, determine the dynamic geometric position encoding of the relative spatial relationship between multiple residue pairs, and generate the gated geometric position encoding by adjusting the weight of the dynamic geometric position encoding through a gating signal.

[0061] The broadcast range management module 2 is used to combine geometric position encoding with attention mechanisms to determine the attention weights that adjust the positional relationships between amino acids, and to adjust the broadcast range of the attention weights according to the attention type of the target information being processed.

[0062] The correction module 3 is used to correct residue positions based on geometric position encoding, attention weight, and the broadcast range of attention weight, and to update the geometric position encoding according to the corrected residue positions.

[0063] The biomacromolecule generation module 4 is used to determine the predicted biomacromolecule sequence structure based on the updated geometric position encoding.

[0064] In this embodiment, the geometric position information of amino acids in a biomacromolecule sequence is encoded to determine the dynamic geometric position encoding of the relative spatial relationship between multiple residue pairs. The gating signal is used to adjust the weights of the dynamic geometric position encoding to generate a gated geometric position encoding, including: It receives input biomolecule sequences, ligands, and covalent bond information, and generates initial residue feature representations of the biomolecule sequences. The starting three-dimensional coordinates of the generated biomacromolecule sequence serve as the starting point for dynamic geometric position coding; Based on the initial three-dimensional coordinate points, the three-dimensional coordinates of the residues are mapped to high-dimensional features through linear transformation. Combined with the noise perturbation process of the diffusion model, an encoding vector characterizing the relative spatial relationship between residue pairs is generated as the dynamic geometric position encoding between multiple residue pairs. Generate one or more gating signals, which are used to adjust the weights of the dynamic geometric position encoding; The gating signal is applied to the dynamic geometric position code to generate a gating-adjusted geometric position code.

[0065] In this embodiment, combining geometric position encoding with an attention mechanism to determine the attention weights for adjusting the positional relationships between amino acids includes: Geometric position encoding is injected as a logarithmic bias term into the attention score calculation of the attention mechanism, and the attention score between amino acid residue pairs is used as the attention weight to adjust the positional relationship between amino acids.

[0066] In this embodiment, adjusting the broadcast range of attention weights according to the attention type of the processed target information includes: Determine the frequency of change or rate of change of the target information being processed during the iterative denoising process; Determine the attention type corresponding to the target information based on the frequency of change or the rate of change of the impact of the target information, and set the broadcast ratio of the attention type corresponding to the target information. The broadcast range is controlled based on the broadcast ratio, which determines the attention weight corresponding to the target information.

[0067] In this embodiment, the attention type corresponding to the target information is determined based on the frequency of change or the rate of change of influence of the target information, and the broadcast ratio of the attention type corresponding to the target information is set as follows: When the target information is to process atomic-level local structural information and capture instantaneous changes in hydrogen bonds or van der Waals forces, the target information is classified as spatial attention feature information. When the target information is sequence evolution information, the target information is classified as temporal attention feature information; When the target information is the ligand-protein binding pattern in the global structure generation stage, the target information is classified as cross-modal attention feature information. The broadcast ratio of spatial attention features is set to be less than that of temporal attention features, and the broadcast ratio of temporal attention features is set to be less than that of cross-modal attention features.

[0068] In this embodiment, controlling the broadcast range of the attention weight corresponding to the target information based on the broadcast ratio includes: The total number of diffusion steps when setting the attention weights corresponding to spatial attention features for a single broadcast is less than the total number of diffusion steps when setting the attention weights corresponding to temporal attention features for a single broadcast. The total number of diffusion steps when setting the attention weights corresponding to temporal attention features for a single broadcast is less than the total number of diffusion steps when setting the attention weights corresponding to cross-modal attention features for a single broadcast. In response to a single broadcast of the target attention weights, determine at each time step within the total number of diffusion steps whether the current time step is within the broadcast range; In response to the current time step being within the broadcast range, control the reuse of the target attention weight of the previous time step in the next time step; In response to the current time step not being within the broadcast range, the attention weight for the next time step is recalculated according to the attention mechanism, and the attention weight for the next time step is saved for broadcasting within the total number of diffusion steps of the next single broadcast.

[0069] In this embodiment, determining whether the current time step is within the broadcast range at each time step within the total number of diffusion steps in response to a single broadcast of the target attention weight includes: At the current time step, obtain the current diffusion step of the target attention weight, obtain the total number of single diffusion steps for broadcasting the target attention weight, and determine the real-time diffusion step position based on the current diffusion step and the total number of single diffusion steps. Calculate the U-shaped stability curve based on the real-time diffusion step location; If the following conditions are met: the next diffusion step is within the total number of diffusion steps in a single run, the U-shaped stability curve is greater than the first threshold, and the attention stability is greater than the attention stability threshold, then the current time step is determined to be within the broadcast range; otherwise, the current time step is determined not to be within the broadcast range.

[0070] In this embodiment, determining the real-time diffusion step location based on the current diffusion step and the total number of diffusion steps in a single run includes: Obtain the current diffusion step and the total number of diffusion steps in a single run; The location of the real-time diffusion step is determined based on the quotient of the current diffusion step and the total number of diffusion steps in a single run.

[0071] In this embodiment, calculating the U-shaped stability curve based on the location of the real-time diffusion step includes: Obtain the absolute value of the difference between the location of the real-time diffusion step and the first threshold; Calculate the product of the absolute value and the second threshold; The U-shaped stability curve is determined based on the difference between the third threshold and the product.

[0072] In this embodiment, controlling the broadcast range of the attention weight corresponding to the target information based on the broadcast ratio includes: Obtain the total number of diffusion steps in a single run and the broadcast ratio; The first product is obtained based on the total number of diffusion steps in a single run and the broadcast ratio; The first product is rounded down to determine the broadcast range of the attention weight corresponding to the target information.

[0073] In this embodiment, setting the broadcast ratio of the attention type corresponding to the target information further includes: Obtain the target attention type corresponding to the target information, and obtain the confidence score for broadcasting the attention weights corresponding to the target information when the target attention type is selected; If the confidence score is less than the preset confidence score threshold, the broadcast ratio of the target attention type corresponding to the target information is increased.

[0074] In this embodiment, the residue position is corrected based on the geometric position encoding, attention weight, and the broadcast range of the attention weight. Updating the geometric position encoding according to the corrected residue position includes: During the iterative denoising process, the initial three-dimensional coordinates of the corresponding amino acid sequence atoms in the geometric position encoding are obtained; At each iteration, the initial three-dimensional coordinates and amino acid sequence pairing information are input into the diffusion model. The diffusion model combines geometric position encoding and attention weights to identify and predict noise vectors. The noise vector is subtracted from the initial 3D coordinates to correct the residue positions, and the geometric position encoding is updated based on the corrected 3D coordinates.

[0075] In this embodiment, determining the predicted biomacromolecule sequence structure based on the updated geometric position encoding includes: By using a distance-constrained network, the target's three-dimensional coordinates are optimized based on the predicted distance matrix to conform to the molecular force field rules, thus obtaining the optimized target's three-dimensional coordinates. The predicted biomolecular sequence structure is generated based on the optimized target three-dimensional coordinates.

[0076] In the aforementioned biomacromolecule sequence structure prediction device, this application introduces a geometric position encoding gating mechanism during the diffusion process and utilizes a multi-scale attention broadcasting (PAB) mechanism to broadcast attention weights corresponding to different attention types of the processed target information in different broadcast ranges. Attention weights in different time steps within the broadcast range are reused, achieving the use of the same attention weights within the steps of the broadcast range. This reduces the computational load, improves model efficiency and performance, and effectively enhances the model's ability to perceive amino acid position information in protein sequences. This allows the model to fully consider the positional relationships between amino acids when calculating attention weights, helping the model to better identify long-chain amino acid interactions, reducing structure prediction errors caused by missing position information, and improving prediction accuracy.

[0077] For a description of the features in the embodiments corresponding to the biomacromolecule sequence structure prediction device, please refer to the relevant descriptions in the embodiments corresponding to the biomacromolecule sequence structure prediction method, which will not be repeated here.

[0078] Embodiments of this application also provide an electronic device, including a memory and a processor, wherein the memory stores a computer program and the processor is configured to run the computer program to perform the steps in any of the above embodiments of the biomacromolecule sequence structure prediction method.

[0079] In one embodiment, the electronic device may be a server, and its internal structure diagram may be as follows: Figure 6 As shown, this electronic device includes a processor, memory, network interface, and database connected via a system bus. The processor provides computational and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores the operating system, computer programs, and database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database stores biomolecular sequence structure prediction data. The network interface communicates with external terminals via a network connection. When the computer program is executed by the processor, it implements a biomolecular sequence structure prediction method.

[0080] Embodiments of this application also provide a computer-readable storage medium storing a computer program, wherein the computer program is configured to execute the steps in any of the above embodiments of the biomacromolecule sequence structure prediction method when run: The geometric position information of amino acids in the sequence of biological macromolecules is encoded to determine the dynamic geometric position encoding of the relative spatial relationship between multiple residue pairs. The weight of the dynamic geometric position encoding is adjusted by a gating signal to generate a gated geometric position encoding. By combining geometric position encoding with attention mechanisms, attention weights are determined to adjust the positional relationships between amino acids, and the broadcast range of attention weights is adjusted according to the attention type of the target information being processed. The residue positions are corrected based on the geometric position encoding, attention weights, and the broadcast range of the attention weights, and the geometric position encoding is updated based on the corrected residue positions. The predicted biomolecular sequence structure was determined based on the updated geometric position coding.

[0081] In one exemplary embodiment, the aforementioned computer-readable storage medium may include, but is not limited to, various media capable of storing computer programs, such as a USB flash drive, read-only memory (ROM), random access memory (RAM), portable hard disk, magnetic disk, or optical disk.

[0082] Embodiments of this application also provide a computer program product, which includes a computer program that, when executed by a processor, implements the steps in any of the above embodiments of the method for predicting the sequence structure of biological macromolecules: The geometric position information of amino acids in the sequence of biological macromolecules is encoded to determine the dynamic geometric position encoding of the relative spatial relationship between multiple residue pairs. The weight of the dynamic geometric position encoding is adjusted by a gating signal to generate a gated geometric position encoding. By combining geometric position encoding with attention mechanisms, attention weights are determined to adjust the positional relationships between amino acids, and the broadcast range of attention weights is adjusted according to the attention type of the target information being processed. The residue positions are corrected based on the geometric position encoding, attention weights, and the broadcast range of the attention weights, and the geometric position encoding is updated based on the corrected residue positions. The predicted biomolecular sequence structure was determined based on the updated geometric position coding.

[0083] Embodiments of this application also provide another computer program product, including a non-volatile computer-readable storage medium storing a computer program. When executed by a processor, the computer program implements the steps in any of the above embodiments of the biomacromolecule sequence structure prediction method. The geometric position information of amino acids in the sequence of biological macromolecules is encoded to determine the dynamic geometric position encoding of the relative spatial relationship between multiple residue pairs. The weight of the dynamic geometric position encoding is adjusted by a gating signal to generate a gated geometric position encoding. By combining geometric position encoding with attention mechanisms, attention weights are determined to adjust the positional relationships between amino acids, and the broadcast range of attention weights is adjusted according to the attention type of the target information being processed. The residue positions are corrected based on the geometric position encoding, attention weights, and the broadcast range of the attention weights, and the geometric position encoding is updated based on the corrected residue positions. The predicted biomolecular sequence structure was determined based on the updated geometric position coding.

[0084] Those skilled in the art will further recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0085] The above provides a detailed description of a method and electronic device for predicting the sequence structure of biological macromolecules provided in this application. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only intended to help understand the method and its core ideas. It should be noted that those skilled in the art can make various improvements and modifications to this application without departing from its principles, and these improvements and modifications also fall within the protection scope of this application.

Claims

1. A method for predicting the sequence structure of biological macromolecules, characterized in that, include: The geometric position information of amino acids in the sequence of biological macromolecules is encoded to determine the dynamic geometric position encoding of the relative spatial relationship between multiple residue pairs. The weight of the dynamic geometric position encoding is adjusted by a gating signal to generate a gated geometric position encoding. The geometric position encoding is combined with the attention mechanism to determine the attention weights that adjust the positional relationships between amino acids, and the broadcast range of the attention weights is adjusted according to the attention type of the target information being processed. The residue positions are corrected based on the geometric position encoding, the attention weight, and the broadcast range of the attention weight, and the geometric position encoding is updated according to the corrected residue positions; The predicted biomolecular sequence structure was determined based on the updated geometric position coding.

2. The method for predicting the sequence structure of biological macromolecules according to claim 1, characterized in that, The process of encoding the geometric position information of amino acids in a biomacromolecule sequence, determining the dynamic geometric position encoding of the relative spatial relationships between multiple residue pairs, and generating a gated geometric position encoding by adjusting the weights of the dynamic geometric position encoding through a gating signal includes: Receive input biomacromolecule sequences, ligands, and covalent bond information, and generate an initial residue feature representation of the biomacromolecule sequence; The starting three-dimensional coordinates of the biomacromolecule sequence are generated as the starting point of the dynamic geometric position encoding; Based on the initial three-dimensional coordinate point, the three-dimensional coordinates of the residues are mapped to high-dimensional features through linear transformation. Combined with the noise perturbation process of the diffusion model, an encoding vector characterizing the relative spatial relationship between residue pairs is generated as the dynamic geometric position encoding between multiple residue pairs. Generate one or more gating signals, which are used to adjust the weights of the dynamic geometric position encoding; The gating signal is applied to the dynamic geometric position code to generate a gating-adjusted geometric position code.

3. The method for predicting the sequence structure of biological macromolecules according to claim 1, characterized in that, The step of combining the geometric position encoding with the attention mechanism to determine the attention weights for adjusting the positional relationships between amino acids includes: The geometric position encoding is injected as a logarithmic bias term into the attention score calculation of the attention mechanism, and the attention score between amino acid residue pairs is used as the attention weight to adjust the positional relationship between amino acids.

4. The method for predicting the sequence structure of biological macromolecules according to claim 1, characterized in that, The broadcast range for adjusting the attention weight based on the attention type of the processed target information includes: Determine the frequency of change or rate of change of the target information being processed during the iterative denoising process; Determine the attention type corresponding to the target information based on the frequency of change or the rate of change of influence of the target information, and set the broadcast ratio of the attention type corresponding to the target information. The broadcast range of the attention weight corresponding to the target information is controlled according to the broadcast ratio.

5. The method for predicting the sequence structure of biological macromolecules according to claim 4, characterized in that, The step of determining the attention type corresponding to the target information based on the change frequency or influence change rate of the target information, and setting the broadcast ratio of the attention type corresponding to the target information, includes: When the target information is to process atomic-level local structural information and capture instantaneous changes in hydrogen bonds or van der Waals forces, the target information is classified as spatial attention feature information. When the target information is processing sequence evolution information, the target information is classified as temporal attention feature information; When the target information is a ligand-protein binding mode in the global structure generation stage, the target information is classified as cross-modal attention feature information; The broadcast ratio of the spatial attention feature information is set to be less than the broadcast ratio of the temporal attention feature information, and the broadcast ratio of the temporal attention feature information is less than the broadcast ratio of the cross-modal attention feature information.

6. The method for predicting the sequence structure of biological macromolecules according to claim 5, characterized in that, The broadcast range for controlling the attention weight corresponding to the target information based on the broadcast ratio includes: The total number of diffusion steps when the attention weight corresponding to the spatial attention feature information is broadcast once is less than the total number of diffusion steps when the attention weight corresponding to the temporal attention feature information is broadcast once, and the total number of diffusion steps when the attention weight corresponding to the temporal attention feature information is broadcast once is less than the total number of diffusion steps when the attention weight corresponding to the cross-modal attention feature information is broadcast once. In response to a single broadcast of the target attention weights, determine at each time step within the total number of diffusion steps whether the current time step is within the broadcast range; In response to the current time step being within the broadcast range, control the reuse of the target attention weight of the previous time step in the next time step; In response to the current time step not being within the broadcast range, the attention weight for the next time step is recalculated according to the attention mechanism, and the attention weight for the next time step is saved for broadcasting within the total number of diffusion steps of the next single broadcast.

7. The method for predicting the sequence structure of biological macromolecules according to claim 6, characterized in that, The response to determining whether the current time step is within the broadcast range at each time step within the total number of diffusion steps during a single broadcast of the target attention weights includes: At the current time step, obtain the current diffusion step of the target attention weight, obtain the total number of single diffusion steps for the target attention weight to be broadcast, and determine the real-time diffusion step position based on the current diffusion step and the total number of single diffusion steps. Calculate the U-shaped stability curve based on the location of the real-time diffusion step; If the following conditions are met: the next diffusion step is within the total number of diffusion steps in a single run, the U-shaped stability curve is greater than the first threshold, and the attention stability is greater than the attention stability threshold, then the current time step is determined to be within the broadcast range; otherwise, the current time step is determined not to be within the broadcast range.

8. The method for predicting the sequence structure of biological macromolecules according to claim 7, characterized in that, Determining the real-time diffusion step location based on the current diffusion step and the total number of single diffusion steps includes: Obtain the current diffusion step and the total number of diffusion steps in a single run; The real-time diffusion step position is determined based on the quotient of the current diffusion step and the total number of diffusion steps in a single run; The calculation of the U-shaped stability curve based on the location of the real-time diffusion step includes: Obtain the absolute value of the difference between the location of the real-time diffusion step and the first threshold; Calculate the product of the absolute value and the second threshold; The U-shaped stability curve is determined based on the difference between the third threshold and the product. The step of controlling the broadcast range of the attention weight corresponding to the target information according to the broadcast ratio includes: Obtain the total number of diffusion steps in a single run and the broadcast ratio; The first product is obtained based on the total number of diffusion steps in a single run and the broadcast ratio; The first product is rounded down to determine the broadcast range of the attention weight corresponding to the target information.

9. The method for predicting the sequence structure of biological macromolecules according to claim 4, characterized in that, The setting of the broadcast ratio for the attention type corresponding to the target information also includes: Obtain the target attention type corresponding to the target information, and obtain the confidence score for broadcasting the attention weight corresponding to the target information when the target attention type is specified. In response to the confidence score being less than a preset confidence score threshold, the broadcast ratio of the target attention type corresponding to the target information is increased.

10. The method for predicting the sequence structure of biological macromolecules according to claim 1, characterized in that, The step of correcting the residue position based on the geometric position encoding, the attention weight, and the broadcast range of the attention weight, and updating the geometric position encoding according to the corrected residue position, includes: During the iterative denoising process, the initial three-dimensional coordinates of the corresponding amino acid sequence atoms in the geometric position encoding are obtained; In each iteration, the initial three-dimensional coordinates and amino acid sequence pairing information are input into the diffusion model, which keeps the broadcast range of the attention weights unchanged. The model combines the geometric position encoding and the attention weights to identify and predict noise vectors. The noise vector is subtracted from the initial three-dimensional coordinates to correct the residue position, and the geometric position code is updated based on the corrected three-dimensional coordinates.

Images