A protein mutation stability prediction method based on a low-rank fitting structure inverse folding model

Abstract

The present application belongs to the field of bioinformatics, and particularly relates to a protein mutation stability prediction method based on a low-rank adaptive structure inverse folding model. The present application performs efficient fine-tuning of the parameters of the structure condition inverse folding model through a low-rank adaptive module, and uses an adapter to scale and adjust the contribution of low-rank adaptive update, thereby reducing the amount of training parameters and the consumption of computing resources while retaining the original structural prior information of the model. The present application compares the structure condition probabilities of wild-type residues and mutant residues in the same protein backbone structure and the same mask sequence context, and obtains residue-level contribution results through adjacent residue masking analysis, which can obtain single-point mutation stability prediction scores and provide local structure explanations on which the model prediction is based. The present application does not rely on experimental stability change labels in the target data set for supervised training, can avoid dependence on experimental annotation data, and improves the applicability of the method in new protein families and insufficiently determined protein structures.

Description

Technical Field

[0001] This invention belongs to the field of bioinformatics, specifically relating to a method for predicting protein mutation stability based on a low-rank adaptor reverse folding model. Background Technology

[0002] Protein stability is a crucial property determining the correct folding of proteins, the maintenance of their biological functions, and their engineering potential. Natural proteins are typically in a relatively limited stable state; a single amino acid mutation can alter the protein's internal interaction network, local conformational flexibility, and spatial stacking patterns, leading to reduced protein function, misfolding, aggregation, or inactivation. Therefore, accurately predicting the impact of single-point mutations on protein stability is of great significance for enzyme engineering, disease-related mutation analysis, industrial protein design, and functional protein optimization.

[0003] Currently, changes in protein mutation stability can be measured experimentally, but these methods are typically time-consuming, costly, and unable to cover a large number of candidate mutation combinations. To improve mutation screening efficiency, various computational prediction methods have been proposed, including those based on physical energy functions, statistical potentials, and machine learning. Methods based on physical energy functions usually rely on explicit structural modeling and complex energy term calculations, resulting in high computational costs, and the prediction results are easily affected by parameter settings and structural quality.

[0004] In recent years, data-driven machine learning methods have achieved some success in predicting mutation stability by integrating protein sequence, structure, and evolutionary features. However, these methods typically rely on a large amount of supervised training with experimentally labeled stability data. When there is insufficient labeled data for the target protein family, or when the distribution of the test protein differs significantly from that of the training data, the model is prone to problems such as decreased generalization ability, overfitting, or leakage of homology information.

[0005] With the development of protein structure prediction and protein representation learning, structural conditional reverse folding models can predict the amino acid sequence that matches a given protein backbone structure, thus providing structural prior information for mutation effect analysis. However, the original training objectives of existing structural reverse folding models are mainly sequence recovery or sequence design, rather than directly targeting mutation stability ranking; if full parameter fine-tuning is performed directly, it may lead to high training costs, a large number of parameters, and the destruction of pre-trained structural priors.

[0006] Furthermore, existing methods for predicting mutational stability typically focus more on the final prediction score, while failing to adequately explain the local structural factors upon which the model prediction depends. It is difficult to identify which neighboring residues or local structural environments play a major role in determining mutational stability. Therefore, there is an urgent need for a protein mutational stability prediction method that can utilize protein scaffold structure information, reduce dependence on experimental stability tags, lower model adaptation costs, and simultaneously provide residue-level structural explanations. Summary of the Invention

[0007] Objective: This invention addresses the shortcomings of existing technologies by providing a protein mutation stability prediction method based on a low-rank adaptor structural reverse folding model. This method fully utilizes the protein's three-dimensional scaffold structure information and the sequence-structure compatibility features learned by the structural conditional reverse folding model, enabling rapid prediction of single-point mutation stability changes without relying on experimental stability labels from the target mutation dataset for supervised training. The invention employs a low-rank adaptor module for efficient parameter fine-tuning, reducing the number of training parameters and computational overhead. Simultaneously, an adapter scaling strategy adjusts the relative contribution between the original structural prior and the low-rank adaptor update signal, minimizing structural prior forgetting caused by full-parameter fine-tuning. Furthermore, the invention obtains residue-level contribution results through neighbor residue masking analysis, which can explain the local structural factors upon which the model prediction depends. This method is applicable to scenarios such as protein engineering, enzyme stability modification, disease-related mutation analysis, and functional protein design.

[0008] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0009] The first aspect of this invention provides a method for predicting protein mutation stability based on a low-rank adaptor reverse folding model, comprising the following steps:

[0010] Step 1: Construction of an unsupervised training set of protein structures, which includes the protein backbone structure and the natural amino acid sequence of the protein, but does not include experimental stability change tags.

[0011] Step 2: Construction of protein structure map input. For the protein backbone structure in the protein structure training set, extract the main chain backbone atomic coordinates of each residue, and construct a residue-level spatial nearest neighbor map based on the spatial distance between residues.

[0012] Step 3: Establish a structural conditional inverse folding model, which includes a structural feature extraction module, a node and edge feature embedding module, a sequence embedding module, a structural encoder layer, a sequence conditional decoder layer, a normalization layer, and an amino acid prediction output projection layer. The structural conditional inverse folding model is used to output the conditional probability distribution of amino acids at each residue position based on the input protein backbone structure, residue-level spatial nearest neighbor graph, and masked sequence context.

[0013] Step 4: Introduce a low-rank adaptation module into the structural conditional inverse folding model, specifically: freeze the original parameters in the structural feature extraction module, node and edge feature embedding module, sequence embedding module, structural encoder layer, sequence conditional decoder layer, normalization layer and amino acid prediction output projection layer, and insert the low-rank adaptation module into the linear layers contained in the above modules.

[0014] Step 5: Mask some residues in the protein sequence, and perform unsupervised residue recovery training on the masked residues to optimize the parameters of the low-rank adapter module;

[0015] Step 6: During the inference phase, the relative contribution between the output of the original structural condition inverse folding model and the updated output of the low-rank adaptation module is adjusted by the adapter scaling factor to obtain the final model parameters;

[0016] Step 7: Mutation stability score. For the single-point mutation to be evaluated, the structural conditional probabilities of wild-type residues and mutant residues are calculated in the context of the same protein backbone structure and the same mask sequence, and the difference between their log probabilities is used as the mutation stability prediction score.

[0017] Step 8: Residue-level contribution interpretation. For the mutated residue, select its spatial nearest neighbor residues, mask the amino acid information of the spatial nearest neighbor residues one by one, and repeat step 7 to obtain a new prediction score. Determine the contribution value of the spatial nearest neighbor residues to the central mutated residue based on the score change before and after masking.

[0018] In step 1, the protein structure training set is obtained by sequentially performing structural quality screening, outlier removal, sequence consistency clustering, and redundancy removal on the acquired protein three-dimensional structure data; the threshold for sequence consistency clustering is 30%.

[0019] In some embodiments of the present invention, in step 1, a set of approximately 295,000 protein structure samples is ultimately obtained. , It can be represented as:

[0020] (1)

[0021] in, This represents the protein structure training set used for unsupervised training. Indicates the first The skeletal structure of a protein sample; Indicates the first The natural amino acid sequence corresponding to each protein sample; This indicates the number of protein structure samples in the training set.

[0022] In step 2, residues are used as graph nodes, and the graph is constructed based on the backbone atoms of the main chain. The Euclidean distance between residues determines the spatial proximity relationship between residues. For each residue, the 48 residues with the closest spatial distance are selected as spatial nearest neighbors to obtain a residue-level spatial nearest neighbor map, which is used to represent the local three-dimensional structural environment around the residue when it is a mutation site.

[0023] In some embodiments of the present invention, in step 2, the protein backbone structure... Represented as:

[0024] (2)

[0025] in, Represents the protein backbone structure; Indicates the length of the protein sequence; Indicates the corresponding residue Four main chain backbone atoms; Representing each atom Three-dimensional coordinates, these three-dimensional spatial coordinates can be directly derived from...

[0026] In some embodiments of the present invention, in step 2, the spatial distance between residues is expressed as:

[0027] (3)

[0028] in, Represents residues With residues Spatial distance between them; and Residues respectively and residues of Atomic coordinates.

[0029] In some embodiments of the present invention, in step 3, the masked sequence context is derived from the original sequence in the training samples. .in, This is the natural amino acid sequence of the protein sample; when predicting the first... When considering the conditional probability of amino acids at each residue position, The Middle Replace each residue with a mask marker to obtain The positions of the remaining residues remain unchanged. Therefore, It is the original sequence of the training samples The sequence context obtained after single-position masking is specifically represented as follows:

[0030] For the first The natural amino acid sequence corresponding to each protein sample:

[0031] (4)

[0032] When predicting the first When considering the conditional probability of amino acids at each residue position, The Middle Masking the residue positions yields:

[0033] (5)

[0034] In step 3, the structural conditional reverse folding model is based on the protein backbone structure. Residue-level spatial nearest neighbor graph and the masked sequence context As input, the output is the conditional probability distribution of amino acids at the corresponding residue positions, specifically represented as follows:

[0035] (6)

[0036] in, This represents the conditional probability of amino acids output by the structural conditional inverse folding model. Indicates the first Candidate amino acids at each residue position; Represents the protein backbone structure; Indicated by the spatial distance between residues Constructed residue-level spatial nearest neighbor graph; Indicates the first The sequence context after each residue is masked.

[0037] In some embodiments of the present invention, during the basic model establishment stage in step 3... Indicates the parameters of the inverse folding model under basic structural conditions. After introducing low-rank adaptation and performing adapter scaling in steps 4 to 6, the inference phase... Use final parameters .

[0038] In step 4, the frozen original parameters include the geometric feature linear projection parameters, node and edge feature embedding module parameters, sequence embedding module parameters, structural encoder layer parameters, sequence conditional decoder layer parameters, normalization layer parameters, and amino acid prediction output projection layer parameters in the structural feature extraction module; the low-rank adaptation module inserts layers with linear transformations into each module, specifically including the structural feature linear projection layer, node embedding linear layer, edge embedding linear layer, sequence embedding linear layer, message transformation and feedforward linear layer in the structural encoder layer, message transformation and feedforward linear layer in the sequence conditional decoder layer, and amino acid prediction output projection layer.

[0039] In step 4, for any adapted linear transformation layer, its output satisfies the following relationship:

[0040] (7)

[0041] in, This represents the output of the linear mapping layer; This represents the feature vector passed by the structural conditional inverse folding model; This represents the original weights of the linear layers of the base model that are frozen. yes A specific weight matrix corresponding to this linear transformation; and This represents a trainable low-rank matrix. Denotes the rank of a low-rank matrix. This represents the low-rank adaptation scaling factor.

[0042] In some embodiments of the present invention Take 32, Take 32, so that 1.0; low-rank matrix Initialized using Kaiminguniform, low-rank matrix Initialized to zero, the low-rank adaptive model contains 592,736 trainable parameters.

[0043] In step 5, the masking process includes random residue masking and local spatial residue masking; the random residue masking involves randomly selecting residues to be restored from the effective residue positions; the local spatial residue masking starts from the central residue and follows... Atomic distance is used to select the spatial nearest neighbor residues as the residues to be recovered; the residue masking ratio in the masking process is 10%, and the true amino acid type of the masked residues is predicted during training using the unmasked residues and their structural context.

[0044] In some embodiments of the present invention, the training loss function is expressed as:

[0045] (9)

[0046] in, This indicates unsupervised residue recovery loss; Represents the set of masked residue locations; Indicates the first The actual amino acid type at each residue position; This represents the conditional probability of amino acids output by the structural conditional inverse folding model. Represents the protein backbone structure; Represents a residue-level spatial nearest neighbor graph; Indicates the first The sequence context after each residue is masked.

[0047] In step 6, the final model parameters are:

[0048] (10)

[0049] in, Indicates the parameters of the final structural condition inverse folding model; This represents all the original parameters of the structural conditional inverse folding model without the introduction of low-rank adaptation modules, including the original weights of each module described in step 3. This indicates the parameter updates generated by the low-rank adaptation module; This represents the adapter scaling factor. For any adapted linear layer, yes The components of The corresponding incremental weights are generated by the low-rank matrix.

[0050] In some embodiments of the present invention, for the first At each residue position, after introducing a low-rank adapter and performing adapter scaling, the conditional probability of the candidate amino acid is expressed as follows:

[0051] (11)

[0052] in, This represents the amino acid conditional probability output by the structural conditional inverse folding model after low-rank adaptation and adapter scaling. This represents the final structural condition inverse folding model parameters after low-rank adaptation and adapter scaling; Indicates the first Candidate amino acids at each residue position; Represents the protein backbone structure; Represents a residue-level spatial nearest neighbor graph; Indicates the first The sequence context after each residue is masked.

[0053] In step 7, the single-point mutation to be evaluated is represented as follows:

[0054] (12)

[0055] in, Indicates a single-point mutation. Indicates the location of the mutated residue. Indicates wild-type residues, Indicates mutated residues;

[0056] Mutated residues Replace the masking markers with the structural conditional probabilities of mutant and wild-type residues within the same protein backbone structure and the same masking sequence context. The mutation stability prediction score is expressed as:

[0057] (13)

[0058] in, Indicates mutation The corresponding stability prediction score; This represents the structural condition amino acid probability distribution after low-rank adaptation and adapter scaling.

[0059] In step 8, for the mutated residues ,according to Atomic distance is selected from its 48 spatial nearest neighbor residues, and then each of its spatially nearest neighbor residues is selected one by one. This masks the amino acid information; in the masked residues In the current state, step 7 is repeated to obtain a new predicted score; and the residue-level contribution value is determined based on the score change before and after masking, expressed as:

[0060] (14)

[0061] in, Indicates neighboring residues For the central mutated residue The contribution value of the mutation prediction results. This represents the set of candidate mutant amino acids excluding wild-type residues. Indicates the shielding of neighboring residues The mutation stability prediction score was then recalculated.

[0062] In some embodiments of the present invention, by performing predictions on different datasets and comparing with existing methods, it is shown that the method provided by the present invention has good prediction performance. The present invention does not rely on experimental stability labels of the target dataset, and can utilize unlabeled protein structure data to complete model adaptation, reducing dependence on experimental measurement data. The present invention uses a low-rank adaptation module for efficient parameter fine-tuning, reducing the number of training parameters and computational costs, while also reducing the risk of structural prior forgetting caused by full parameter fine-tuning. The present invention uses a training strategy combining random residue masks and local spatial residue masks, enabling the model to simultaneously learn the overall sequence-structure relationship and the local structural environment near the mutation site. The present invention uses an adapter scaling strategy to balance the original structural prior and the fine-tuned mutation score signal, which is beneficial to improving the model's generalization ability in different protein systems. The present invention uses the log probability difference between wild-type residues and mutated residues under the same structural context for scoring, which can quickly complete single-point mutation stability prediction. The present invention further provides residue-level contribution interpretation results, which helps to analyze the local structural sources of mutation influence and improve the interpretability of prediction results.

[0063] Beneficial effects:

[0064] (1) This invention does not rely on experimental stability change labels in the target dataset for supervised training, but uses unlabeled protein structure data for unsupervised residue recovery training, which avoids dependence on experimental labeled data and improves the applicability of the method in new protein families and proteins whose structures have not been fully determined.

[0065] (2) This invention uses a low-rank adaptation module to efficiently fine-tune the parameters of the structural condition inverse folding model, and uses adapter scaling to adjust the contribution of low-rank adaptation updates during the inference stage. This reduces the number of training parameters and computational resource consumption, while helping to preserve the original structural prior information of the model.

[0066] (3) This invention compares the structural conditional probabilities of wild-type residues and mutant residues in the context of the same protein backbone structure and the same mask sequence, and further obtains residue-level contribution results through neighboring residue masking analysis. It can quickly obtain the single-point mutation stability prediction score and provide the local structural interpretation on which the model prediction is based. Attached Figure Description

[0067] The present invention will be further described in detail below with reference to the accompanying drawings, and the advantages of the present invention in the above and / or other aspects will become clearer.

[0068] Figure 1 This is a flowchart of the protein mutation stability prediction method in an embodiment of the present invention.

[0069] Figure 2This is a scatter plot showing the correlation between predicted values ​​and experimental measurements of the prediction method in this embodiment of the invention on the S669 dataset.

[0070] Figure 3 This is a precision-recall curve of the prediction method in this embodiment of the invention on the S669 dataset.

[0071] Figure 4 This is a plot of the subject operating characteristic curve of the prediction method in this embodiment of the invention on the S669 dataset.

[0072] Figure 5 This is a bar chart comparing the evaluation metrics of the prediction method and existing methods in this embodiment of the invention on the S669 dataset.

[0073] Figure 6 This is a scatter plot showing the correlation between predicted values ​​and experimentally measured values ​​of the prediction method in this embodiment of the invention on the MegaTest dataset.

[0074] Figure 7 This illustrates the prediction performance of the prediction method in different structural environments in the embodiments of the present invention.

[0075] Figure 8 The results are from ablation experiments using the prediction method in this embodiment of the invention.

[0076] Figure 9 The residue-level contribution interpretation results of the prediction method in the embodiments of the present invention. Detailed Implementation

[0077] The present invention will be further described in detail below with reference to specific embodiments, and the advantages of the present invention in the above and / or other aspects will become clearer.

[0078] Unless otherwise specified, the experimental methods described in the following examples are conventional methods; unless otherwise specified, the reagents and materials are commercially available.

[0079] Example 1:

[0080] This embodiment provides a method for predicting protein mutation stability based on a low-rank adaptor reverse folding model. Figure 1 The flowchart for this protein mutation stability prediction method is as follows:

[0081] Step 1: Construction of the protein structure training set for unsupervised training

[0082] To obtain publicly available protein 3D structure data, this embodiment uses the RCSB and BFVD databases. The protein structures are quality-screened, removing samples with severe backbone atom defects, incomplete chain information, or those that cannot be properly resolved. Clustering is performed based on amino acid sequence consistency, with a clustering threshold of 30% to remove highly similar homologous sequences, ensuring data diversity. Ultimately, a set of approximately 295,000 protein structure samples is obtained. . It only contains the protein backbone structure and its natural amino acid sequence, and does not contain experimental stability change tags. The model training process does not require experimentally determined mutation stability values ​​as a supervision signal. It can be represented as:

[0083] (1)

[0084] in, This represents the protein structure training set used for unsupervised training. Indicates the first The skeletal structure of a protein sample; Indicates the first The natural amino acid sequence corresponding to each protein sample; This indicates the number of protein structure samples in the training set.

[0085] Step 2: Constructing the protein structure diagram input

[0086] For unsupervised training of protein structure training sets Each protein backbone structure Extraction length is Each residue in the sequence The protein backbone structure is represented by the coordinates of four backbone atoms as follows:

[0087] (2)

[0088] in, Represents the protein backbone structure; This indicates the length of the protein sequence; 4 indicates the length of each residue. Four main chain backbone atoms; 3 indicates the number of atoms in each atom. Three-dimensional coordinates, these three-dimensional spatial coordinates can be directly derived from...

[0089] Furthermore, each residue is treated as a graph node in a residue-level spatial nearest neighbor graph, according to... The Euclidean distance between atoms determines the spatial proximity relationship between residues, thus constructing a residue-level spatial nearest neighbor graph. The spatial distance between residues is represented as:

[0090] (3)

[0091] in, Represents residues With residues Spatial distance between them; and Residues respectively and residues of Atomic coordinates.

[0092] For each residue, the 48 spatially nearest residues are selected as spatial neighbors, resulting in a residue-level spatial neighbor graph. , is used to represent the local three-dimensional structural environment surrounding the residue when it is a mutation site.

[0093] Step 3: Establishment of the structural condition inverse folding model

[0094] The structural conditional inverse folding model includes a structural feature extraction module, a node and edge feature embedding module, a sequence embedding module, a structural encoder layer, a sequence conditional decoder layer, a normalization layer, and an amino acid prediction output projection layer.

[0095] The structural feature extraction module is used to extract protein backbone structure. Extracting the spatial distance between residues Relative direction, local spatial orientation, and main chain geometric relationships, and based on Determine the spatial nearest neighbor graph of residues The node and edge feature embedding module is used to map residue node features, residue edge features, and geometric features into latent space representations; the sequence embedding module is used to map masked sequence contexts into residue sequence representations.

[0096] Nearest neighbor graph of structural encoder layer at the residue level Message passing and neighborhood aggregation are performed on the first, enabling the second... Each residue node receives the geometric constraints and structural context of its spatial nearest residues to obtain a structural conditional representation; the sequence conditional decoder layer further integrates the mask sequence context and the structural conditional representation, and after the feature distribution is stabilized by the normalization layer, the amino acid prediction output projection layer maps it to the conditional probability distribution of 20 standard amino acids.

[0097] Therefore, the model starts from the protein backbone structure. To begin, first determine the distance between the residues. Constructing a spatial nearest neighbor graph Then , The mask sequence and context are jointly encoded into a residue-level representation, and finally the first residue is output. The conditional probability distribution of amino acids corresponding to each residue position. The mask sequence context is derived from the original sequence in the training samples. Soon The Middle The residues were replaced with mask markers, while the positions of the remaining residues remained unchanged.

[0098] Step 3 is specifically represented as follows:

[0099] For the first The natural amino acid sequence corresponding to each protein sample:

[0100] (4)

[0101] When predicting the first When considering the conditional probability of amino acids at each residue position, The Middle Masking the residue positions yields:

[0102] (5)

[0103] At this point, for the first The candidate amino acids at each residue position. The conditional probability is expressed as:

[0104] (6)

[0105] in, This represents the conditional probability of amino acids output by the structural conditional inverse folding model. Indicates the first Candidate amino acids at each residue position; Represents the protein backbone structure; Indicated by the spatial distance between residues Constructed residue-level spatial nearest neighbor graph; Indicates the first The sequence context after each residue is masked.

[0106] Specifically, in the basic model establishment stage of step 3, Indicates the parameters of the inverse folding model under basic structural conditions. After introducing low-rank adaptation and performing adapter scaling in steps 4 to 6, the inference phase... Use final parameters .

[0107] Step 4: Introduction of the low-rank adaptation module

[0108] To reduce the number of training parameters and retain the existing structural priors in the base model, all original parameters of the pre-trained structural conditional inverse folding model are first frozen. The original parameters include geometric feature linear projection parameters from the structural feature extraction module, node and edge feature embedding module parameters, sequence embedding module parameters, structural encoder layer parameters, sequence conditional decoder layer parameters, normalization layer parameters, and amino acid prediction output projection layer parameters. Subsequently, low-rank adaptation modules are introduced into the layers with linear transformations in the above modules, specifically including a structural feature linear projection layer, a node embedding linear layer, an edge embedding linear layer, a sequence embedding linear layer, a message transformation and feedforward linear layer within the structural encoder layer, a message transformation and feedforward linear layer within the sequence conditional decoder layer, and an amino acid prediction output projection layer. For the normalization layer, only its original scale and translation parameters are frozen; no additional low-rank adaptation modules are inserted. For any adapted linear transformation, its output is expressed as:

[0109] (7)

[0110] in, This represents the output of the linear mapping layer; This represents the feature vector passed by the structural conditional inverse folding model; This represents the original weights of the linear layers of the base model that are frozen. yes A specific weight matrix corresponding to this linear transformation; and This represents a trainable low-rank matrix. Denotes the rank of a low-rank matrix. This represents the low-rank adaptation scaling factor.

[0111] Among the specific parameters, Take 32, Take 32, so that 1.0; low-rank matrix Initialized using Kaiming uniform, low-rank matrix Initialized to zero, the low-rank adaptive model contains 592,736 trainable parameters.

[0112] Step 5: Unsupervised residue recovery training

[0113] A training method combining random residue masking and local spatial residue masking is employed. The low-rank adaptor module is trained by predicting the true amino acid type of the masked residue using the structural context surrounding it, enabling it to recognize the influence of the local structural environment on amino acid stability. Random residue masking randomly selects residues to be restored from effective residue positions; local spatial residue masking selects neighboring residues as residues to be restored, starting from the central residue and based on spatial distance. During training, 10% of the effective residues are selected as mask positions, specifically as follows:

[0114] (8)

[0115] in, This indicates the proportion of residues in each protein sequence that are selected as mask positions.

[0116] The model predicts the true amino acid type of masked residues based on unmasked residues, protein backbone structure, and residue nearest neighbor relationships. The training loss function is expressed as:

[0117] (9)

[0118] in, This indicates unsupervised residue recovery loss; Represents the set of masked residue locations; Indicates the first The actual amino acid type at each residue position; This represents the conditional probability of amino acids output by the structural conditional inverse folding model. Represents the protein backbone structure; Represents a residue-level spatial nearest neighbor graph; Indicates the first The sequence context after each residue is masked.

[0119] During training, a standard deviation of 0.05 was added to the skeleton atomic coordinates. Gaussian noise; the optimizer uses AdamW with an initial learning rate of 3×10⁻⁶. -4 The weight decays to 1×10 -3 The minimum learning rate is 1×10. -6 The equivalent batch size is 32, and mixed precision training and gradient pruning are used.

[0120] Step 6: Adapter Scaling Adjustment

[0121] During the inference phase, a scaling factor is applied to the parameter updates generated by the low-rank adaptation module to adjust the relative contribution between the original structural conditional model output and the low-rank adaptation update output. The final structural conditional inverse folding model parameters used for mutation stability prediction are expressed as follows:

[0122] (10)

[0123] in, Indicates the parameters of the final structural condition inverse folding model; This represents all the original parameters of the structural conditional inverse folding model without the introduction of the low-rank adaptor module; This indicates the parameter updates generated by the low-rank adaptation module. This represents the adapter scaling factor. A value of 0 indicates that only the basic model is used. A value of 1 indicates that a low-rank adaptation update is fully adopted; in this embodiment... A value of 0.56 is used to balance the original structural prior and the mutation score signal after low-rank adaptation.

[0124] At this point, for the first At each residue position, after introducing a low-rank adaptor module, the candidate amino acids... The conditional probability is expressed as:

[0125] (11)

[0126] in, This represents the amino acid conditional probability output by the structural conditional inverse folding model after low-rank adaptation and adapter scaling. This represents the final structural condition inverse folding model parameters after low-rank adaptation and adapter scaling; Indicates the first Candidate amino acids at each residue position; Represents the protein backbone structure; Represents a residue-level spatial nearest neighbor graph; Indicates the first The sequence context after each residue is masked.

[0127] Step 7: Calculate the mutation stability prediction score

[0128] For the single-point mutation to be evaluated, the structural conditional probabilities of wild-type residues and mutated residues are calculated separately within the same protein backbone structure and the same mask sequence context, and the difference between their log probabilities is used as the mutation stability prediction score. The single-point mutation to be evaluated is represented as follows:

[0129] (12)

[0130] in, Indicates a single-point mutation. Indicates the location of the mutated residue. Indicates wild-type residues, This indicates a mutated residue.

[0131] Mutated residues Replace the masking markers with the structural conditional probabilities of mutant and wild-type residues within the same protein backbone structure and the same masking sequence context. The mutation stability prediction score is expressed as:

[0132] (13)

[0133] in, Indicates mutation The corresponding mutation stability prediction score; This represents the structural condition amino acid probability distribution after low-rank adaptation and adapter scaling. This score compares mutant and wild-type residues within the same structural context, reducing scoring bias caused by different input backgrounds.

[0134] Step 8: Explanation of Residue-Level Contributions

[0135] For mutated residues Firstly, according to Atomic distance is selected from its 48 spatial nearest neighbor residues, and then each of its spatially nearest neighbor residues is selected one by one. This masks the amino acid information; in the masked residues In the current state, step 7 is repeated to obtain a new predicted score; and the residue-level contribution value is determined based on the score change before and after masking, expressed as:

[0136] (14)

[0137] in, Indicates neighboring residues For the central mutated residue The contribution value of the mutation prediction results. This represents the set of candidate mutant amino acids excluding wild-type residues. Indicates the shielding of neighboring residues The predicted mutation stability score was then recalculated. If the predicted score changes significantly after masking a neighboring residue, it indicates that the neighboring residue makes a significant contribution to the model's assessment of mutation stability.

[0138] The overall prediction performance of the prediction method (StabIF) provided in this embodiment is characterized on different benchmark datasets.

[0139] Evaluation metrics include Spearman's rank correlation coefficient, Pearson's linear correlation coefficient, accuracy, area under the receiver operating characteristic curve (AUROC), and area under the precision-recall curve (AUPRC).

[0140] The Spearman rank correlation coefficient is expressed as:

[0141] (15)

[0142] The Pearson linear correlation coefficient is expressed as:

[0143] (16)

[0144] in, This represents the model's predicted score vector. Represents the vector of experimental stability changes. This indicates the number of mutations to be evaluated.

[0145] First, the correlation between the prediction scores obtained by the prediction method and the experimental stability changes is characterized on the S669 and MegaTest datasets. Figure 2 This is a scatter plot showing the correlation between predicted and experimentally measured values ​​of the method of this invention on the S669 dataset. Figure 2 As can be seen, the Spearman rank correlation coefficient of the method of the present invention is 0.490 and the Pearson linear correlation coefficient is 0.445 on the S669 dataset, indicating that the method of the present invention can reflect the trend of protein mutation stability changes well and has consistent predictive ability on different test datasets.

[0146] Figure 3 This is a precision-recall curve of the method of the present invention on the S669 dataset, used to evaluate the detection capability of the method of the present invention in the task of identifying stable mutation candidates. Figure 3 As can be seen, the area under the curve is 0.88, indicating that the method of the present invention can maintain a high precision while maintaining a high recall rate, and can effectively identify mutation samples with stable change characteristics.

[0147] Figure 4 This is the receiver operating characteristic (ROC) curve of the method of this invention on the S669 dataset, used to evaluate the ability of the method to distinguish between stability-enhancing and stability-reducing mutations. The area under the curve is 0.72, which is higher than the random classification baseline, indicating that the method of this invention has good binary classification ability and can effectively distinguish mutations with different directions of stability change.

[0148] Figure 5This is a bar chart comparing the evaluation metrics of the method of this invention with existing methods on the S669 dataset. Existing methods include I-Mutant3.0, DUET, MAESTRO, PopMusic, ThermoNet, Pythia, and DynaMut. Evaluation metrics include Spearman rank correlation coefficient, Pearson linear correlation coefficient, accuracy, area under the receiver operating characteristic (ROC) curve, area under the precision-recall curve, and overall mean. Figure 5 As can be seen, the method of the present invention achieves superior results in multiple indicators, including an accuracy of 0.75, an area under the receiver operating characteristic curve of 0.72, an area under the precision-recall curve of 0.88, and a comprehensive average of 0.66. This indicates that the method of the present invention has good comprehensive performance in terms of relevance evaluation, classification ability, and overall prediction performance.

[0149] Figure 6 This is a scatter plot showing the correlation between predicted and experimentally measured values ​​of the method of this invention on the MegaTest dataset. Figure 6 It can be seen that there is also a significant positive correlation between the predicted values ​​and the experimental values. The Spearman rank correlation coefficient is 0.645 and the Pearson linear correlation coefficient is 0.637, indicating that the method of the present invention still has good stability, prediction consistency and generalization ability on independent test datasets.

[0150] Table 1 compares the performance of the proposed method (StabIF) with FoldX, Rosetta, MIF, ABACUS2, ESM-IF, ESM2-t33, and ProteinMPNN on the S2648 dataset. The comparison shows that the proposed method achieves the best results in Spearman rank correlation coefficient, accuracy, and AUPRC index, indicating that it has better overall performance in ranking mutation effects and distinguishing stability effects.

[0151] Table 1. Performance comparison of different methods on the S2648 dataset.

[0152]

[0153] As shown in Table 1, the method of this invention achieves a Spearman rank correlation coefficient of 0.60, a Pearson linear correlation coefficient of 0.58, an accuracy of 0.81, and an AUPRC of 0.53 on the S2648 dataset. The Spearman rank correlation coefficient is higher than that of FoldX, Rosetta, MIF, ABACUS2, ESM-IF, ESM2-t33, and ProteinMPNN; the accuracy and AUPRC are also the highest, indicating that this method can effectively maintain the order of mutation effects and has good ability to distinguish stable effects.

[0154] Figure 7 This paper demonstrates the predictive performance of the method of this invention in different structural environments. Mutation sites in the SSYM dataset were categorized into α-helices, β-sheets, random coils, surface residues, buried residues, and all residues based on secondary structure and solvent accessibility. The results show that the Spearman rank correlation coefficients of the method of this invention are 0.723 and 0.722 in the β-sheet and α-helical regions, respectively; the Spearman rank correlation coefficient for buried residues is 0.643, close to the 0.640 for all residues; and the Spearman rank correlation coefficient for random coil regions is 0.434. These results indicate that the method has good modeling ability for locally constrained environments.

[0155] Figure 8 The ablation experiments of the method of this invention are presented. The ablation experiments compared a base model without low-rank adaptation fine-tuning, a model with low-rank adaptation but without adapter scaling, and a complete model with both low-rank adaptation and adapter scaling. All three models used the same data processing flow, mutation scoring function, and evaluation metrics, and were evaluated on the S669, S2648, MegaTest, SSYM-dir, and SSYM-inv datasets, respectively. Here, SSYM-dir represents the symmetric mutation subset calculated according to the original mutation direction, and SSYM-inv represents the symmetric mutation subset calculated after reversing the corresponding mutation direction, used to verify the model's predictive consistency regarding mutation direction and its reverse stability changes. Figure 8 The results show that low-rank adaptation can enhance the model's ability to capture changes in mutation stability, and adapter scaling can alleviate the problem of excessive deviation of the model from the basic prior after low-rank adaptation, so that the complete model can achieve more balanced performance on multiple benchmark datasets.

[0156] Figure 9 The residue-level contribution interpretation results of the method of the present invention are presented. Figure 9 Figure A in the figure shows the spatial relationship between the F52 site and its surrounding key neighboring residues in the GB1 protein; Figure 9The B diagram in the image shows the connectivity between F52 and its surrounding residues in the form of a residue nearest neighbor diagram. Figure 9 A and B in the figure show that the StabIF model takes into account F52 and its surrounding local environment, which is consistent with known biological facts, proving that the method is accurate in identifying local key structures. Figure 9 Figure C in the diagram shows the spatial distribution of the D186 site and surrounding engineered mutation sites in the PETase / DuraPETase structure; Figure 9 The D-plot in the present invention further quantifies the contribution relationship between D186 and multiple mutation-related residues, showing that these sites have different degrees of influence on the model stability score; Figure 9 The E-plot in the figure compares the average contribution of the mutation site with other neighboring residues, showing that the engineered mutation site exhibits a higher model contribution in DuraPETase, proving that the method of the present invention can effectively filter background noise and capture key structural regions related to stability improvement.

[0157] This invention provides a method for predicting protein mutation stability based on a low-rank adaptive structure backfolding model. Many methods and approaches exist for implementing this technical solution; the above description is merely a preferred embodiment. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of this invention, and these improvements and modifications should also be considered within the scope of protection of this invention. All components not explicitly stated in this embodiment can be implemented using existing technologies.

Claims

1. A method for predicting protein mutation stability based on a low-rank adaptor reverse folding model, characterized in that, Includes the following steps: Step 1: Construction of an unsupervised training set of protein structures, which includes the protein backbone structure and the natural amino acid sequence of the protein, but does not include experimental stability change tags. Step 2: Construction of protein structure map input. For the protein backbone structure in the protein structure training set, extract the main chain backbone atomic coordinates of each residue, and construct a residue-level spatial nearest neighbor map based on the spatial distance between residues. Step 3: Establish a structural conditional inverse folding model, which includes a structural feature extraction module, a node and edge feature embedding module, a sequence embedding module, a structural encoder layer, a sequence conditional decoder layer, a normalization layer, and an amino acid prediction output projection layer. The structural conditional inverse folding model is used to output the conditional probability distribution of amino acids at each residue position based on the input protein backbone structure, residue-level spatial nearest neighbor graph, and masked sequence context. Step 4: Introduce a low-rank adaptation module into the structural conditional inverse folding model, specifically: freeze the original parameters in the structural feature extraction module, node and edge feature embedding module, sequence embedding module, structural encoder layer, sequence conditional decoder layer, normalization layer and amino acid prediction output projection layer, and insert the low-rank adaptation module into the linear layers contained in the above modules. Step 5: Mask some residues in the protein sequence, and perform unsupervised residue recovery training on the masked residues to optimize the parameters of the low-rank adapter module; Step 6: During the inference phase, the relative contribution between the output of the original structural condition inverse folding model and the updated output of the low-rank adaptation module is adjusted by the adapter scaling factor to obtain the final model parameters; Step 7: Mutation stability score. For the single-point mutation to be evaluated, the structural conditional probabilities of wild-type residues and mutant residues are calculated in the context of the same protein backbone structure and the same mask sequence, and the log probability difference between the two is used as the mutation stability prediction score. Step 8: Residue-level contribution interpretation. For the mutated residue, select its spatial nearest neighbor residues, mask the amino acid information of the spatial nearest neighbor residues one by one, and repeat step 7 to obtain a new prediction score. Determine the contribution value of the spatial nearest neighbor residues to the central mutated residue based on the score change before and after masking.

2. The protein mutation stability prediction method according to claim 1, characterized in that, In step 1, the protein structure training set is obtained by sequentially performing structural quality screening, outlier removal, sequence consistency clustering, and redundancy removal on the acquired protein three-dimensional structure data; the threshold for sequence consistency clustering is 30%.

3. The method for predicting protein mutation stability according to claim 1, characterized in that, In step 2, residues are used as graph nodes, and the main chain backbone atoms are used as the basis for the graph. The Euclidean distance between residues determines the spatial proximity relationship between residues. For each residue, the 48 residues with the closest spatial distance are selected as spatial nearest neighbors to obtain a residue-level spatial nearest neighbor map, which is used to represent the local three-dimensional structural environment around the residue when it is a mutation site.

4. The protein mutation stability prediction method according to claim 1, characterized in that, In step 3, the structural conditional inversion model includes a structural feature extraction module, a node and edge feature embedding module, a sequence embedding module, a structural encoder layer, a sequence conditional decoder layer, a normalization layer, and an amino acid prediction output projection layer. The structural conditional inversion model takes the protein backbone structure, the residue-level spatial nearest neighbor graph, and the masked sequence context as input, and outputs the conditional probability distribution of amino acids at corresponding residue positions, specifically as follows: (6) in, This represents the conditional probability of amino acids output by the structural conditional inverse folding model. Indicates the first Candidate amino acids at each residue position; Represents the protein backbone structure; Indicated by the spatial distance between residues Constructed residue-level spatial nearest neighbor graph; Indicates the first The sequence context after masking each residue.

5. The method for predicting protein mutation stability according to claim 1, characterized in that, In step 4, the frozen original parameters include the geometric feature linear projection parameters in the structural feature extraction module, the node and edge feature embedding module parameters, the sequence embedding module parameters, the structural encoder layer parameters, the sequence conditional decoder layer parameters, the normalization layer parameters, and the amino acid prediction output projection layer parameters. The low-rank adaptation module inserts layers with linear transformations into each module, specifically including the structural feature linear projection layer, the node embedding linear layer, the edge embedding linear layer, the sequence embedding linear layer, the message transformation and feedforward linear layer in the structural encoder layer, the message transformation and feedforward linear layer in the sequence conditional decoder layer, and the amino acid prediction output projection layer.

6. The protein mutation stability prediction method according to claim 5, characterized in that, In step 4, for any layer that has undergone the adapted linear transformation, its output satisfies the following relationship: (7) in, This represents the output of the linear mapping layer; This represents the feature vector passed by the structural conditional inverse folding model; This represents the original weights of the linear layers of the base model that are frozen. yes A specific weight matrix corresponding to this linear transformation; and This represents a trainable low-rank matrix. Denotes the rank of a low-rank matrix. This represents the low-rank adaptation scaling factor.

7. The method for predicting protein mutation stability according to claim 1, characterized in that, In step 5, the masking process includes random residue masking and local spatial residue masking; the random residue masking involves randomly selecting residues to be restored from the effective residue positions; the local spatial residue masking starts from the central residue and follows... Atomic distance is used to select the spatial nearest neighbor residues as the residues to be recovered; the residue masking ratio in the masking process is 10%, and the true amino acid type of the masked residues is predicted during training using the unmasked residues and their structural context.

8. The method for predicting protein mutation stability according to claim 1, characterized in that, In step 6, the final model parameters are: (10) in, Indicates the parameters of the final structural condition inverse folding model; This represents the original parameters of the structural conditional inverse folding model without the introduction of LoRA; This indicates the parameter updates generated by the low-rank adaptation module. This represents the adapter scaling factor.

9. The method for predicting protein mutation stability according to claim 1, characterized in that, In step 7, the single-point mutation to be evaluated is represented in the following form: (12) in, Indicates a single-point mutation. Indicates the location of the mutated residue. Indicates wild-type residues, Indicates mutated residues; Mutated residues Replace the masking markers with the structural conditional probabilities of mutant and wild-type residues within the same protein backbone structure and the same masking sequence context. The mutation stability prediction score is expressed as: (13) in, Indicates mutation The corresponding stability prediction score; This represents the structural condition amino acid probability distribution after low-rank adaptation and adapter scaling.

10. The method for predicting protein mutation stability according to claim 1, characterized in that, In step 8, for the mutated residues ,according to Atomic distance is selected from its 48 spatial nearest neighbor residues, and then each of its spatially nearest neighbor residues is selected one by one. This masks the amino acid information; in the masked residues In the current state, step 7 is repeated to obtain a new predicted score; and the residue-level contribution value is determined based on the score change before and after masking, expressed as: (14) in, Indicates neighboring residues For the central mutated residue The contribution value of the mutation prediction results. This represents the set of candidate mutant amino acids excluding wild-type residues. Indicates the shielding of neighboring residues The mutation stability prediction score was then recalculated.

Images