A method for multi-scale quality assessment of protein complex structure model

Abstract

The application discloses a protein complex structure model multi-scale quality evaluation method, and belongs to the technical field of bioinformatics and artificial intelligence assisted protein structure analysis. First, three-dimensional atomic coordinates and amino acid sequence information of a protein complex structure model to be evaluated are acquired. Then, multi-scale structure features are extracted from four scales of atoms, residues, interfaces and the whole. Further, a unified representation space is constructed through feature alignment, scale normalization and unified mapping, and bidirectional cross-linking and information coupling between different scale features are realized based on a multi-head attention mechanism. Then, the representation is input into a hierarchical perception deep neural network, and through multi-task learning and a multi-head output mechanism, atomic scale, residue scale, interface scale and global scale are cooperatively evaluated. Finally, atomic scores, residue scores, interface region scores and global structure quality scores are output. The application improves the reliability of protein complex structure model quality evaluation results.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of bioinformatics and deep learning, and specifically relates to a multi-scale quality assessment method for protein complex structure models. Background Technology

[0002] Proteins are the fundamental functional carriers of life activities, and their three-dimensional spatial structure directly determines their biological functions. With the continuous development of bioinformatics and deep learning technologies, the focus of protein structure research has shifted from single protein chains to protein complexes. Especially in the AlphaFold era, various protein complex structure prediction methods have emerged, enabling the rapid generation of numerous protein complex structure models. However, due to the complex inter-chain interactions, diverse spatial organization, and significant interfacial coupling relationships of protein complexes, existing predicted structure models still exhibit differences in atomic arrangement rationality, residue folding accuracy, inter-chain interfacial interactions, and overall topological consistency. Therefore, accurate and comprehensive quality assessment of protein complex structure models has become a crucial step in bioinformatics research.

[0003] Currently, quality assessment methods for protein complex structure models are mainly divided into two categories: self-assessment methods and third-party model quality assessment methods. Self-assessment methods are typically integrated within the structure prediction model, using intermediate features, attention information, or structure confidence information generated during the prediction process to assess the quality of the generated structure. These methods can directly combine information from within the prediction model for structure quality estimation; however, their assessment results usually depend on the corresponding structure prediction model itself, and are only applicable to structures generated by the model itself, limiting their generality and failing to meet the need for unified assessment of structures from multiple sources. Third-party model quality assessment methods do not depend on a specific prediction model, offering greater generality and objectivity; however, existing methods generally suffer from the limitation of single-scale modeling. Specifically, there is a lack of information exchange between the rationality of local atomic arrangement and the correctness of residue-level folding, a lack of synergistic constraints between the accuracy of residue structure and the stability of inter-chain interfaces, and a lack of bidirectional information exchange mechanism between the stability of interface interactions and the consistency of global topology. This makes it difficult for existing methods to simultaneously take into account the rationality of local atomic arrangement, the stability of inter-chain interface interactions, and the consistency of global topology. The technical root cause is the fragmentation of cross-scale information and the lack of bidirectional interaction mechanism, which makes it impossible to achieve synergistic constraints and fusion of multi-scale features. Ultimately, this manifests as defects such as coarse evaluation granularity and insufficient stability and reliability of results.

[0004] Therefore, there is an urgent need for a multi-scale joint evaluation method for protein complex structure models covering four scales: atoms, residues, interfaces, and the global scale. This method would address the problems of fragmented cross-scale information and insufficient evaluation accuracy in existing methods, and provide a reliable basis for protein complex structure screening, optimization, and downstream research. Summary of the Invention

[0005] To overcome the shortcomings of existing technologies, this invention provides a multi-scale quality assessment method for protein complex structure models. The core objective is to achieve bidirectional coupled assessment of four scales—atoms, residues, interfaces, and the global scale—for the first time, thereby enhancing the correlation and expression ability between structural information at different scales, improving the reliability of protein complex structure model quality assessment results, and providing a reliable basis for structural quality assessment in subsequent bioinformatics research.

[0006] To achieve the above-mentioned objectives, the present invention adopts the following technical solution: A multi-scale quality assessment method for protein complex structural models is characterized by: firstly, acquiring the three-dimensional atomic coordinates and amino acid sequence information of the protein complex structural model to be assessed; then, extracting multi-scale structural features from four scales: atoms, residues, interfaces, and the global scale; further constructing a unified representation space through feature alignment, scale normalization, and unified mapping; and achieving bidirectional cross-correlation and information coupling between features at different scales based on a multi-head attention mechanism to form a cross-scale coupled representation and hierarchical interactive representation; then, inputting the representation into a hierarchical perceptual deep neural network, and performing collaborative quality assessment at the atomic, residue, interface, and global scales through a multi-task learning and multi-head output mechanism; and finally outputting a per-atom score, a per-residue score, a per-interface region score, and a global structural quality score.

[0007] Furthermore, the multi-scale quality assessment method for the protein complex structure model includes the following steps: Step S1: Obtain the structural model of the protein complex to be evaluated, wherein the structural model of the protein complex includes the three-dimensional atomic coordinates and amino acid sequence information of at least two interacting protein chains; Step S2: Standardize the protein complex structure model to eliminate data interference, and then extract multi-scale structural features from four dimensions: atomic scale, residue scale, interface scale, and global scale, respectively, including covering sequence, atomic geometry, residue-level structure, interface interaction, and global topology. Step S3: Perform feature preprocessing on the extracted multi-scale structural features, construct cross-scale coupled representation and hierarchical interactive representation of multi-scale structural features of protein complexes, realize bidirectional cross-association and bidirectional information flow between the four scales of atoms-residues-interfaces-global based on the cross-scale attention weight matrix, and realize bidirectional feature interaction mapping and information fusion between each scale. Step S4: Input the cross-scale coupled representation and hierarchical interactive representation into the hierarchical perceptual deep neural network trained by multiple tasks. Through the multi-head output mechanism of the network, multi-scale feature representation learning is performed on the atomic scale, residue scale, interface scale and global scale respectively. Step S5: Output multi-scale quality scores: atomic-level atom-by-atom score vector, residue-by-residue score vector, interface-level interface region score set, and global structural quality scalar score. The scores are normalized to the 0-1 range, and the higher the value, the better the structural quality.

[0008] Compared with the prior art, the beneficial effects of the present invention are: 1. This invention innovatively extracts multi-scale structural features from four dimensions: atomic scale, residue scale, interface scale, and global scale. It covers sequence information, atomic geometric information, residue structural information, interface interaction information, and global topological information. This enables the extracted features to comprehensively and accurately describe protein complexes from different structural levels, improves the completeness of structural information expression, effectively solves the problem of missing information in structural characterization by single-scale features, and provides a sufficient and reliable feature basis for subsequent multi-scale quality assessment.

[0009] 2. This invention proposes for the first time a four-scale, all-dimensional, bidirectional coupling architecture involving atoms, residues, interfaces, and the global scale. This differs from conventional solutions that rely on single-scale modeling, simple feature splicing, or unidirectional information transfer. It possesses three core innovative features: First, it dynamically quantifies the dependence strength of features at different scales through a cross-scale attention weight matrix, precisely controlling the contribution ratio of information fusion at each scale. Second, it achieves bidirectional cross-correlation and bidirectional information flow between each pair of scales, rather than unidirectional feature transfer. Third, it employs a multi-task joint training framework, constraining the evaluation at each scale through independent loss functions, thus achieving collaborative optimization of multi-scale quality. This design fundamentally solves the core pain points of existing technologies—the fragmentation of cross-scale information and the lack of bidirectional interaction mechanisms. It is not a conventional improvement based on existing technology by those skilled in the art, effectively enhancing the expressive power of multi-scale structural features and the stability of evaluation results.

[0010] 3. This invention employs a hierarchical perceptual deep neural network that can provide finer-grained structural quality assessment scores. Through a multi-task learning framework, combined with a multi-head output mechanism, each output branch corresponds to an independent loss function and is jointly optimized to achieve synchronous output of atomic-level atom-by-atom scoring vectors, residue-level residue-by-residue scoring vectors, interface-level interface residue pairs or interface region scoring sets, and global structural single scalar scores.

[0011] 4. The multi-scale quality scores output by this invention are all normalized to the 0–1 range. The atomic, residue, and interface levels are output in element-wise form, while the global level is in scalar form, which makes the quality assessment results of different structural models at different scales highly comparable. At the same time, various scores can be organized into structured output results for subsequent analysis, which significantly improves the consistency of the assessment results and the convenience of application, and provides a reliable reference for protein complex structure screening, structure optimization, and subsequent bioinformatics analysis. Attached Figure Description

[0012] Figure 1 This is a flowchart illustrating the overall process of the multi-scale quality assessment method for protein complex structure models according to an embodiment of the present invention.

[0013] Figure 2 This is a schematic diagram of the cross-scale feature coupling and hierarchical interaction structure according to an embodiment of the present invention. Detailed Implementation

[0014] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.

[0015] Reference Figure 1 and Figure 2 A multi-scale quality assessment method for protein complex structure models includes the following steps: Step S1: Obtain the structural model of the protein complex to be evaluated. The structural model of the protein complex includes the three-dimensional atomic coordinates and amino acid sequence information of at least two interacting protein chains. Specifically, it is obtained by downloading from the PDB database or importing the local structural file to ensure that the model contains complete atomic coordinates (X, Y, Z axes) and amino acid sequence information without missing any key information.

[0016] Step S2: The protein complex structure model is standardized and preprocessed to eliminate data interference. Then, multi-scale structural features are extracted from four dimensions: atomic scale, residue scale, interface scale, and global scale, including covering sequence, atomic geometry, residue-level structure, interface interaction, and global topology. The standardized preprocessing sequentially performs atomic repetition information removal, chain identifier unification, and three-dimensional coordinate normalization. The feature extraction is consistent with the five major feature categories, without adding any additional feature types. In this embodiment, the standardization preprocessing includes atomic repetition information removal, chain identifier unification, and three-dimensional coordinate normalization, used to eliminate data bias from structural models from different sources; the multi-scale structural features are divided into five categories, as follows: Sequence features are obtained from the amino acid sequence of the protein complex, including amino acid type, sequence position, BLOSUM62 substitution matrix, and multiple sequence alignment features; Atomic-level features are obtained from the three-dimensional structural information of the protein complex atoms, including atomic coordinate features, atom type, atomic charge, atomic validity mask, and atomic spatial geometric features based on the local coordinate system; the atomic spatial geometric features include interatomic distance, included angle, and dihedral angle information; Residue-level features are obtained from the spatial structure and physicochemical properties of protein complex residues, including residue distance diagrams, residue orientation diagrams, dihedral angles of the main chain, bond lengths and bond angles, solvent-accessible surface area, amino acid physicochemical properties, secondary structure, and Rosetta monomer energy characteristics. Interface-level features are obtained from the cross-chain residue contact relationships between different protein chains, including cross-chain residue contact maps, cross-chain distance matrices, cross-chain relative orientations, interface contact probabilities, and Rosetta binary energy terms. The interface region is defined based on the CA atomic distance or the shortest heavy atom distance being less than a set threshold (12 Å) to define the set of interface residue pairs. Global topological features are obtained from the overall spatial structure information of the protein complex, including the overall residue contact map, interchain spatial distribution, overall contact density, structural compactness, voxelization, and structural template spectrum features.

[0017] Step S3: Perform feature preprocessing on the extracted multi-scale structural features, construct cross-scale coupled representation and hierarchical interactive representation of multi-scale structural features of protein complexes, realize bidirectional cross-association and bidirectional information flow between the four scales of atoms-residues-interfaces-global based on the cross-scale attention weight matrix, and realize bidirectional feature interaction mapping and information fusion between each scale; feature preprocessing includes feature alignment, scale normalization and unified mapping, and cross-scale coupling is constructed based on the multi-head attention mechanism to clarify the bidirectional coupling relationship between the four scales; In this embodiment, the feature preprocessing includes feature alignment, scale normalization, and unified mapping preprocessing, transforming features of different dimensions and scales into a unified representation space to form a unified hierarchical feature representation. The construction of the cross-scale coupled representation and hierarchical interactive representation achieves bidirectional deep interactive fusion between atomic scale, residue scale, interface scale, and global scale. Specifically, the coupling relationships are as follows: atomic scale features are bidirectionally coupled with residue scale features and global scale features, residue scale features are bidirectionally coupled with interface scale features and global scale features, and interface scale features are bidirectionally coupled with global scale features. The coupling between each scale is based on a multi-head attention mechanism. Currently, by linearly projecting features at different scales, query vector key vectors and numerical vectors of specific dimensions are obtained. Feature association weights are calculated based on a cross-scale attention weight matrix. This weight matrix is ​​used to quantify the strength of dependencies between features at different scales and serves as a coefficient matrix for cross-scale information propagation and feature reweighting. It is used to control the information contribution ratio of features at different scales in the fusion process. Then, different attention heads are used to characterize the spatial dependencies, structural associations, and hierarchical constraints between features at different scales. Finally, the outputs of each attention head are spliced ​​or weighted and fused to obtain cross-scale coupled representations and hierarchical interactive representations, which are used to support subsequent multi-scale structural quality assessment tasks.

[0018] Step S4: Input the cross-scale coupled representation and hierarchical interactive representation into the hierarchical perceptual deep neural network trained by multiple tasks. Through the multi-head output mechanism of the network, multi-scale feature representation learning is performed on the atomic scale, residue scale, interface scale and global scale respectively to achieve accurate evaluation of the structural quality at each scale.

[0019] The hierarchical perceptual deep neural network adopts a multi-task learning architecture to adapt to multi-scale quality assessment requirements. The network includes five functional modules, as follows: The atomic-scale evaluation module is constructed using a graph neural network based on an attention mechanism, and takes atomic-level geometric features and corresponding cross-scale coupling features as input. The residue-scale assessment module adopts an architecture combining a multilayer perceptron (MLP) and a bidirectional long short-term memory (BiLSTM) network, and inputs residue-level structural features and corresponding cross-scale coupling features. The interface scale evaluation module is constructed using a coding structure based on a self-attention mechanism, and the input interface interaction features and corresponding cross-scale coupling features are used. Global topology evaluation module: It adopts a combination architecture of convolutional neural network (CNN) and graph attention mechanism (GAT), inputting global topological features and cross-scale coupling features; extracting global spatial features through CNN, and combining GAT to capture the relationship between chain distribution and contact density, thus characterizing the overall topological relationship of the complex. The multi-task multi-head output module adopts a parallel fully connected layer structure, which corresponds to the evaluation output at four scales: the atomic scale output is a per-atom score vector, the residue scale output is a per-residue score vector, the interface scale output is a per-interface residue pair or interface region score set, and the global scale output is a single scalar score. Each output branch corresponds to an independent loss function and is jointly optimized. The collaborative optimization learning of structural quality is achieved through multi-scale consistency constraints.

[0020] The atomic-scale evaluation module in this embodiment is constructed using a graph neural network (GAT) based on an attention mechanism to achieve accurate evaluation of the rationality of atomic-level structures. It contains two graph attention layers: the first layer has an input dimension of 128 (the dimension of the spliced ​​atomic-level features and cross-scale coupled representations), an output dimension of 64, 8 attention heads, and the activation function LeakyReLU (negative slope 0.2); the second layer has an input dimension of 64, an output dimension of 32, 4 attention heads, and the activation function LeakyReLU (negative slope 0.2); a Dropout layer (probability 0.3) is added after each layer to prevent overfitting; finally, a linear layer maps the output dimension to 1, used to output the structural quality score of a single atom. The residue-scale evaluation module in this embodiment employs an architecture combining a multilayer perceptron (MLP) and a bidirectional long short-term memory (BiLSTM) network to capture temporal correlation features of residue-level structures and achieve residue-by-residue quality scoring. The specific structure is as follows: Input layer (128 dimensions, representing the concatenation of residue-level features and cross-scale coupling representations) → First MLP layer (64 output dimensions, ReLU activation function) → Dropout layer (probability 0.3) → Second MLP layer (32 output dimensions, ReLU activation function) → BiLSTM layer (64 hidden units, 2 layers, bidirectional propagation, Dropout probability 0.3) → Linear layer (1 output dimension). The length of the input sequence to the BiLSTM layer corresponds to the number of residues in the protein complex structure model to be evaluated. The output at each time step corresponds to the intermediate features of a single residue, and the linear layer ultimately outputs the structural quality score of each residue. The interface scale evaluation module in this embodiment employs a self-attention mechanism-based encoding structure to mine inter-chain interface interaction features and achieve interface region quality evaluation. Specifically, it includes a 4-layer self-attention encoder. Each self-attention encoder layer has 8 attention heads, an input dimension of 128 (the dimension obtained by concatenating interface-level features with cross-scale coupling representations), a hidden layer dimension of 64, a feedforward network dimension of 256, and uses ReLU as the activation function. To ensure model stability, each self-attention encoder layer includes a normalization layer and a Dropout layer (probability 0.3). Finally, a single linear layer (input dimension 64, output dimension 1) outputs the structural quality score of a single interface region. The global topology evaluation module in this embodiment employs an architecture combining a Convolutional Neural Network (CNN) and a Graph Attention (GAT) mechanism to extract the overall spatial topological features of the protein complex, achieving global structural quality evaluation. The specific structure is as follows: Input layer (dimension 256, which is the dimension resulting from the concatenation of global topological features and cross-scale coupled representations) → First CNN layer (3×3 kernel size, stride 1, 32 output channels, ReLU activation function) → Max pooling layer (2×2 kernel size, stride 2) → Second CNN layer (3×3 kernel size, stride 1, 64 output channels, ReLU activation function) → Max pooling layer (2×2 kernel size, stride 2) → The third CNN layer (3×3 kernel size, stride 1, 128 output channels, ReLU activation function) → Global Average Pooling layer (128 output dimensions) → GAT layer (4 attention heads, 128 input dimensions, 64 output dimensions, LeakyReLU activation function) → Linear layer (1 output dimension); this module ultimately outputs a global structural quality scalar score, used to characterize the overall structural quality of the entire protein complex; The multi-task, multi-head output module in this embodiment employs a parallel fully connected layer structure, corresponding to evaluation outputs at four scales. Specifically, the atomic-scale output layer has an input dimension of 32, and its output dimension matches the total number of atoms; the residue-scale output layer has an input dimension of 64, and its output dimension matches the total number of residues; the interface-scale output layer has an input dimension of 64, and its output dimension matches the number of interfaces; and the global-scale output layer has an input dimension of 64 and an output dimension of 1. Each output branch corresponds to an independent mean squared error (MSE) loss function. The multi-objective loss function is obtained by weighted summation of the four loss functions, with the following weights for each loss term: atomic-level loss weight 0.2, residue-level loss weight 0.3, interface-level loss weight 0.3, and global-level loss weight 0.2. During model training, an adaptive moment estimation (Adam) optimizer is used, with parameters set as follows: learning rate 1e-4, β1=0.9, β2=0.999, and weight decay 1e-5. An early stopping strategy is also implemented, terminating training when the validation loss no longer decreases for 10 consecutive epochs to avoid overfitting.

[0021] The training process of the hierarchical perceptual deep neural network includes: four independent quality assessment objectives corresponding to the atomic scale, residue scale, interface scale, and global structure, with each objective corresponding to a dedicated supervision label; training the hierarchical perceptual deep neural network using a multi-objective loss function, which is obtained by weighted summation of atomic-level quality score loss, residue-level quality score loss, interface-level quality score loss, and global structure quality score loss. The weights of each loss term are dynamically adjusted based on the assessment difficulty of different scale tasks, sample distribution characteristics, and the convergence speed and gradient change trend of the loss function at each scale on the validation set; during training, an adaptive moment estimation optimizer (Adam) is used to iteratively update the network parameters, and the multi-objective loss function is minimized through backpropagation algorithm to achieve joint learning of multi-scale quality scores; simultaneously, a validation set is set up, and an early stopping strategy is adopted based on the multi-objective loss function values ​​on the validation set. Training is terminated when the validation loss no longer decreases for 10 consecutive epochs to avoid model overfitting and improve model generalization ability.

[0022] The method for constructing the training and validation sets includes: screening protein structures from the Protein Data Bank (PDB) database, with screening criteria including a structural resolution of less than or equal to 2.5 Å and a residue length between 50 and 400; using the MMseqs2 tool to perform sequence redundancy removal on the screened protein structures, with a sequence consistency threshold set to 40% and a bidirectional sequence coverage threshold set to 80%, to obtain a set of non-redundant protein structures; based on the set of non-redundant protein structures, constructing multiple structural perturbation samples as training inputs by comparing modeling methods, structural perturbation methods, and deep learning-based structure generation methods; and dividing the dataset into training and validation sets in an 8:2 ratio.

[0023] Step S5: Output the multi-scale quality score results of the protein complex structure model, including: atomic-level atom-by-atom quality score vector, residue-level residue-by-residue quality score vector, interface region score set, and global structural quality single score value; the atomic-level, residue-level, and interface-level scores are all output in element-by-element form, and each score value is normalized to the 0-1 interval. The global structural quality score is in scalar form and normalized to the 0-1 interval. The higher the value, the better the structural quality.

[0024] In this embodiment, the multi-scale quality scoring results are output synchronously through a multi-task multi-head output module, and include four types of scores, as follows: The atomic-level quality score is output by the atomic-scale assessment module and is an atom-by-atom score vector, representing the structural rationality score of each atom. The residue-level quality score is output by the residue-scale assessment module and is a residue-by-residue score vector, representing the structural quality score of each amino acid residue. The interface region quality score is output by the interface scale assessment module. It is a set of interface region scores calculated based on the set of interface residue pairs or interface contact clusters, representing the structural quality score of each interface contact region. The global structural quality score is output by the global topology evaluation module and is a single scalar score value, representing the overall quality level of the entire protein complex structure. Each element-wise score is normalized to the 0-1 range, and the global score is also normalized to the 0-1 range.

[0025] In one specific embodiment, the above steps are performed using the following specific parameters. Taking the protein complex structure model 8bwl_rank0.pdb as an example, this complex has four chains: A, B, C, and D. Chain A contains 69 amino acid residues, chain B contains 121 amino acid residues, chain C contains 69 amino acid residues, and chain D contains 121 amino acid residues. The total length of the residue sequence is 380, and the total number of atoms is 2932. The specific implementation process is as follows: Step S1: Obtain the structural model of the protein complex to be evaluated, 8bwl_rank0.pdb. Open the model using PyMOL software to confirm that the model contains four interacting protein chains, A, B, C, and D. Extract the three-dimensional atomic coordinates of the model (X, Y, and Z axes with 3 decimal places) and complete amino acid sequence information, ensuring no missing atomic coordinates, no sequence disorder, and no missing atoms. The process is as follows: Step 1.1: Use the PDBParser module of the Biopython library in Python to read the 8bwl_rank0.pdb file, parse it to obtain the residue list, atom list and corresponding coordinate information of each chain, and store it as a NumPy array (data type float32) for subsequent preprocessing operations; Step 1.2: Verify the validity of the parsed coordinate data and the number of atoms, remove abnormal atoms with coordinates of NaN or infinity, and confirm that the total number of atoms after parsing is 2932, ensuring the integrity and accuracy of the model data.

[0026] Step S2: Perform standardization preprocessing on the 8bwl_rank0.pdb model. The process is as follows: Step 2.1: Removal of duplicate atomic information. By comparing the name, coordinates and residues to which the atoms belong, duplicate atoms are removed (a total of 6 duplicate atoms are removed). After removal, the total number of atoms is 2926, ensuring that the number of atoms is accurate. Step 2.2: Unify the chain identifiers. Unify the identifiers of the four chains as "A", "B", "C", and "D" respectively to avoid confusion in the chain identifiers; Step 2.3: Three-dimensional coordinate normalization processing. The Z-score normalization method is used to convert the X, Y, and Z coordinates of all atoms into normalized coordinates with a mean of 0 and a standard deviation of 1, thereby eliminating the interference of different coordinate magnitudes. After preprocessing, multi-scale structural features are extracted from four dimensions: atomic scale, residue scale, interface scale, and global scale. The feature types are consistent with the five categories of features described in the invention, without additional refinement, to ensure that the feature extraction process is completely consistent with the invention.

[0027] Step S3: Perform feature preprocessing on the extracted multi-scale structural features. The process is as follows: Step 3.1: Feature alignment. Align features of different dimensions to the same length (2048 dimensions), and use zero-padding to supplement insufficient dimensions. Step 3.2: Scale normalization. The Min-Max normalization method is used to normalize all feature values ​​to the [0,1] interval. Step 3.3: Unified mapping, through one linear layer (input dimension adjusted according to feature type, output dimension 128), features of different scales are mapped to a unified representation space; Step 3.4: Subsequently, a cross-scale coupled representation and hierarchical interaction representation are constructed. An 8-head attention mechanism is used to linearly project the features at the four scales (query vector, key vector, and value vector are all 64-dimensional). The cross-scale attention weight matrix (64×64-dimensional) is calculated to quantify the strength of the dependency relationship between features at different scales. The outputs of the 8 attention heads are concatenated (each attention head output is 64-dimensional, and the concatenated result is 512-dimensional), and layer normalization is applied to obtain the cross-scale coupled representation and hierarchical interaction representation. The above bidirectional coupling and hierarchical interaction logic can be intuitively seen in [reference needed]. Figure 2 (Schematic diagram of cross-scale feature coupling and hierarchical interaction structure). Figure 2 It presents the dependencies between different scales, attention interaction paths, and information fusion methods.

[0028] Step S4: Input the cross-scale coupled representation and hierarchical interactive representation into the hierarchical perceptual deep neural network jointly trained by multiple tasks. The execution process is as follows: Step 4.1: The atomic scale evaluation module processes atomic-level features and cross-scale coupling features, extracts atomic-level feature correlation information through two GAT layers, and finally outputs the quality score of 2926 atoms. Step 4.2: The residue-scale assessment module processes residue-level features and cross-scale coupling features. Through an architecture combining MLP and BiLSTM, it extracts the correlation information of residue-level structural features and finally outputs the quality score of 380 residues. Step 4.3: The interface scale evaluation module processes the interface-level features and cross-scale coupling features. It extracts the interface interaction features through a 4-layer self-attention encoder and obtains 5 interfaces (AB, AC, AD, BC, CD) based on the CA atomic distance of 8 Å as the threshold. The quality scores of the 5 interfaces are then output. Step 4.4: The global topology evaluation module processes the global topology features and cross-scale coupling features. Through the architecture combining CNN and GAT, it extracts the spatial topology association information of the global structure and outputs a global structure quality scalar score. Step 4.5: Through the multi-task multi-head output module, the scoring results of the four scales are output synchronously, and each score is normalized to the [0,1] interval to ensure that the scoring output is consistent with the model parameters and preprocessing results.

[0029] Step S5: Output the multi-scale quality score results of the 8bwl_rank0.pdb model, as follows: Atom-level mass rating vector: a total of 2926 rating values, all normalized to the [0,1] interval, as shown in the following examples: the α-carbon atom (CA) score of the 10th residue of chain A is 0.89, the β-carbon atom (CB) score is 0.87, and the oxygen atom (O) score is 0.88; the CA atom score of the 25th residue of chain B is 0.91, the CB atom score is 0.89, and the O atom score is 0.86; the scores of the remaining atoms are all distributed between 0.75 and 0.93. The higher the value, the more reasonable the atom arrangement. The number of scores corresponds exactly to the total number of 2926 atoms after preprocessing.

[0030] Residue-level quality score vector: A total of 380 score values, all normalized to the [0,1] interval. Examples are shown below (selecting some residues from each chain): Chain A: Residues 1-10 scores are 0.85, 0.88, 0.91, 0.86, 0.89, 0.87, 0.90, 0.88, 0.86, 0.87; Chain B: Residues 1-10 scores are 0.87, 0.89, 0.84, 0.90, 0.88, 0.86, 0.89, 0.91, 0.87, 0.88; Chain C: Residues 1-10 scores are... The scores for the first 10 residues of chain D are 0.86, 0.88, 0.85, 0.89, 0.87, 0.88, 0.86, 0.89, 0.90, and 0.87, respectively. The scores for the remaining residues are distributed between 0.82 and 0.92. The higher the value, the more accurate the residue folding. The number of scores corresponds exactly to the total length of 380 residues and matches the number of residues in chains A, B, C, and D.

[0031] Interface-level quality score set for each interface region: Based on the 8 Å threshold for CA atomic distance, five interfaces were obtained, with specific scores as follows: Interface AB (contact region between A and B chains) score 0.88, Interface AC (contact region between A and C chains) score 0.90, Interface AD ​​(contact region between A and D chains) score 0.87, Interface BC (contact region between B and C chains) score 0.86, and Interface CD (contact region between C and D chains) score 0.89. All interface scores are distributed between 0.85 and 0.90, with higher values ​​indicating more stable inter-chain interface interactions.

[0032] The global structural quality single score is 0.89 (normalized to the [0,1] interval), indicating that the overall quality of the 8bwl_rank0.pdb protein complex structure model is good and the global topological consistency is high.

[0033] Although specific embodiments of the present invention have been described in detail above with reference to the accompanying drawings, this does not constitute a limitation on the scope of protection of the present invention. Those skilled in the art will readily conceive of other embodiments of the present invention upon considering the specification and practicing the disclosure of the present invention. The present invention is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein. The specification and embodiments are considered exemplary only, and any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A multi-scale quality assessment method for protein complex structural models, characterized in that, First, the three-dimensional atomic coordinates and amino acid sequence information of the protein complex structure model to be evaluated are obtained. Then, multi-scale structural features are extracted from four scales: atoms, residues, interfaces, and the global scale. Furthermore, a unified representation space is constructed through feature alignment, scale normalization, and unified mapping. Based on the multi-head attention mechanism, bidirectional cross-correlation and information coupling between features at different scales are realized, forming a cross-scale coupled representation and hierarchical interactive representation. Then, the representation is input into a hierarchical perceptual deep neural network. Through multi-task learning and multi-head output mechanisms, collaborative quality evaluation is performed at the atomic scale, residue scale, interface scale, and global scale. Finally, the atom-by-atom score, residue-by-residue score, interface-by-interface region score, and global structural quality score are output.

2. The multi-scale quality assessment method for protein complex structure models as described in claim 1, characterized in that, The method includes the following steps: Step S1: Obtain the structural model of the protein complex to be evaluated, wherein the structural model of the protein complex includes the three-dimensional atomic coordinates and amino acid sequence information of at least two interacting protein chains; Step S2: The protein complex structure model is standardized and preprocessed, and then multi-scale structural features are extracted from four dimensions: atomic scale, residue scale, interface scale, and global scale, respectively, including covering sequence, atomic geometry, residue-level structure, interface interaction, and global topology. Step S3: Perform feature preprocessing on the extracted multi-scale structural features, construct cross-scale coupled representation and hierarchical interactive representation of multi-scale structural features of protein complexes, realize bidirectional cross-association and bidirectional information flow between the four scales of atoms-residues-interfaces-global based on the cross-scale attention weight matrix, and realize bidirectional feature interaction mapping and information fusion between each scale. Step S4: Input the cross-scale coupled representation and hierarchical interactive representation into the hierarchical perceptual deep neural network trained by multiple tasks. Through the multi-head output mechanism of the network, multi-scale feature representation learning is performed on the atomic scale, residue scale, interface scale and global scale respectively. Step S5: Output multi-scale quality scores normalized to the 0-1 range, including atomic-level atom-by-atom score vectors, residue-level residue-by-residue score vectors, interface-level interface region score sets, and global structural quality scalar scores.

3. The multi-scale quality assessment method for protein complex structure models according to claim 2, characterized in that, In step S2, the standardization preprocessing includes atomic repetition information removal, chain identifier unification, and three-dimensional coordinate normalization, used to eliminate data bias from structural models from different sources; the multi-scale structural features are divided into five categories, as follows: Sequence features are obtained from the amino acid sequence of the protein complex, including amino acid type, sequence position, BLOSUM62 substitution matrix, and multiple sequence alignment features; Atomic-level features are obtained from the three-dimensional structural information of the protein complex atoms, including atomic coordinate features, atom type, atomic charge, atomic validity mask, and atomic spatial geometric features based on the local coordinate system. The atomic spatial geometric features include information on interatomic distances, included angles, and dihedral angles. Residue-level features are obtained from the spatial structure and physicochemical properties of protein complex residues, including residue distance diagrams, residue orientation diagrams, dihedral angles of the main chain, bond lengths and bond angles, solvent-accessible surface area, amino acid physicochemical properties, secondary structure, and Rosetta monomer energy characteristics. Interface-level features are obtained from the contact relationships of cross-chain residues between different protein chains, including cross-chain residue contact maps, cross-chain distance matrices, cross-chain relative orientations, interface contact probabilities, and Rosetta binary energy terms. The interface region is defined as a set of interface residue pairs based on the CA atomic distance or the shortest heavy atom distance being less than a set threshold. Global topological features are obtained from the overall spatial structure information of the protein complex, including the overall residue contact map, interchain spatial distribution, overall contact density, structural compactness, voxelization, and structural template spectrum features.

4. The multi-scale quality assessment method for protein complex structure models according to claim 2, characterized in that, In step S3, the feature preprocessing includes feature alignment, scale normalization, and unified mapping preprocessing, transforming features of different dimensions and scales into a unified representation space to form a unified hierarchical feature representation. The construction of the cross-scale coupled representation and hierarchical interactive representation achieves bidirectional deep interactive fusion between atomic scale, residue scale, interface scale, and global scale. Specifically, the coupling relationships are as follows: atomic scale features are bidirectionally coupled with residue scale features and global scale features, residue scale features are bidirectionally coupled with interface scale features and global scale features, and interface scale features are bidirectionally coupled with global scale features. The coupling between each scale is based on a multi-head attention mechanism. The method achieves this by linearly projecting features at different scales to obtain query vectors, key vectors, and numerical vectors. It then calculates feature association weights based on the corresponding cross-scale attention weight matrix. This weight matrix quantifies the strength of dependencies between features at different scales and serves as a coefficient matrix for cross-scale information propagation and feature reweighting, controlling the information contribution ratio of features at different scales during the fusion process. Different attention heads are then used to characterize the spatial dependencies, structural relationships, and hierarchical constraints between features at different scales. Finally, the outputs of each attention head are concatenated or weighted to obtain a cross-scale coupled representation and hierarchical interactive representation, supporting subsequent multi-scale structural quality assessment tasks.

5. The multi-scale quality assessment method for protein complex structure models according to claim 2, characterized in that, In step S4, the hierarchical perceptual deep neural network adopts a multi-task learning architecture to adapt to multi-scale quality assessment requirements. The hierarchical perceptual deep neural network includes: The atomic-scale evaluation module is used to construct an attention-based graph neural network, taking atomic-level geometric features and corresponding cross-scale coupling features as input. The residue-scale evaluation module is used to take into account residue-level structural features and corresponding cross-scale coupling features using an architecture that combines a multilayer perceptron and a bidirectional long short-term memory network. The interface scale evaluation module is used to construct an input interface interaction feature and the corresponding cross-scale coupling feature using a self-attention mechanism-based encoding structure. The global topology evaluation module is used to adopt an architecture that combines convolutional neural networks and graph attention mechanisms. It takes global topological features and corresponding cross-scale coupling features as input, and based on the global topological features and the cross-scale coupling information between these features and atomic-scale, residue-scale, and interface-scale features, it characterizes the topological association of the overall spatial structure of the protein complex. It extracts the spatial features of the global structure through convolutional neural networks and captures the relationship between interchain spatial distribution and overall contact density by combining graph attention mechanisms. The multi-task multi-head output module adopts a parallel fully connected layer structure, which corresponds to the evaluation output at four scales: the atomic scale output is a per-atom score vector, the residue scale output is a per-residue score vector, the interface scale output is a per-interface residue pair or interface region score set, and the global scale output is a single scalar score. Each output branch corresponds to an independent loss function and is jointly optimized. The collaborative optimization learning of structural quality is achieved through multi-scale consistency constraints.

6. The multi-scale quality assessment method for protein complex structure models according to claim 5, characterized in that, In step S5, the multi-scale quality scoring results are output synchronously through a multi-task multi-head output module, including atomic-level quality scores, residue-level quality scores, interface region quality scores, and global structural quality scores. The atomic-level quality scores are output by the atomic-scale evaluation module and are per-atom scoring vectors, representing the structural rationality score of each atom. The residue-level quality scores are output by the residue-scale evaluation module and are per-residue scoring vectors, representing the structural quality score of each amino acid residue. The interface region quality scores are output by the interface-scale evaluation module and are per-interface region scoring sets calculated based on the set of interface residue pairs or interface contact clusters, representing the structural quality score of each interface contact region. The global structural quality score is output by the global topology evaluation module and is a single scalar score value, representing the overall quality level of the entire protein complex structure. Each element-level score is normalized to the 0-1 range, and the global score is also normalized to the 0-1 range.

7. The multi-scale quality assessment method for protein complex structure models according to claim 5, characterized in that, The training process of the hierarchical perceptual deep neural network includes: four independent quality assessment objectives corresponding to the atomic scale, residue scale, interface scale, and global structure, with each objective corresponding to a dedicated supervision label; training the hierarchical perceptual deep neural network using a multi-objective loss function, which is obtained by weighted summation of atomic-level quality score loss, residue-level quality score loss, interface-level quality score loss, and global structure quality score loss. The weights of each loss term are dynamically adjusted based on the assessment difficulty of different scale tasks, sample distribution characteristics, and the convergence speed and gradient change trend of the loss function at each scale on the validation set; during training, an adaptive moment estimation optimizer is used to iteratively update the network parameters, and the multi-objective loss function is minimized through backpropagation algorithm to achieve joint learning of multi-scale quality scores; simultaneously, a validation set is set up, and an early stopping strategy is adopted based on the multi-objective loss function values ​​on the validation set, terminating training when the validation loss no longer decreases for a set number of consecutive epochs.

8. The multi-scale quality assessment method for protein complex structure models according to claim 7, characterized in that, The method for constructing the training and validation sets includes: screening protein structures from the PDB database, with screening criteria including structural resolution less than or equal to a preset threshold and residue length within a set range; using the MMseqs2 tool to perform sequence redundancy removal on the screened protein structures to obtain a set of non-redundant protein structures; based on the set of non-redundant protein structures, constructing multiple structural perturbation samples as training inputs by comparing modeling methods, structural perturbation methods, and deep learning-based structure generation methods; and dividing the dataset into training and validation sets according to a ratio.

Images