Protein ddg prediction model optimization method, protein ddg prediction method, and related apparatus
By optimizing variable quantum circuits and protein three-dimensional structure data, the accuracy and efficiency of the protein DDG prediction model were improved, solving the problem of insufficient accuracy of existing models and achieving more accurate DDG prediction.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- TURINGQ CO LTD
- Filing Date
- 2025-06-18
- Publication Date
- 2026-06-02
AI Technical Summary
The accuracy of existing protein DDG prediction models needs to be improved.
A variable quantum circuit-based approach is adopted, which combines protein three-dimensional structure data and optimizes the encoder and decoder to improve the accuracy of the prediction model. This includes preprocessing the training data and extracting and combining feature vectors, and obtaining the DDG value by measuring the entangled state of qubits.
It improves the accuracy and computational efficiency of protein DDG prediction, and can more accurately capture the impact of protein structural changes before and after mutation on free energy.
Smart Images
Figure CN120748503B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of bioinformatics data processing technology, and more specifically, to a method for optimizing a protein DDG prediction model, a protein DDG prediction method, and related apparatus. Background Technology
[0002] Protein-protein complex DDG (Delta Delta G) (hereinafter referred to as protein DDG) is a measure of the impact of mutations on protein stability or interactions, representing the difference in free energy change caused by the mutation. Specifically, it is the difference between the free energy change (ΔGmt) of the mutant protein and the free energy change (ΔGwt) of the wild-type (unmutated) protein. In related technologies, predictive models are generally used to predict protein DDG.
[0003] However, the inventors' research revealed that the accuracy of existing prediction models for the protein DDG needs to be improved. Summary of the Invention
[0004] The objectives of this invention include, for example, providing a method for optimizing a protein DDG prediction model, a method for predicting protein DDG, and related apparatus, which can at least partially solve the aforementioned technical problems.
[0005] The embodiments of the present invention can be implemented as follows:
[0006] In a first aspect, embodiments of the present invention provide a method for optimizing a protein DDG prediction model, the method comprising:
[0007] Acquire preprocessed training data, which includes wild-type protein three-dimensional structure data, mutant protein three-dimensional structure data corresponding to the wild-type protein three-dimensional structure data, and the first DDG value of the mutation;
[0008] The wild-type protein three-dimensional structure data and the mutant protein three-dimensional structure data are respectively input into the optimized encoder of the protein DDG initial prediction model to obtain the first feature vector corresponding to the wild-type protein three-dimensional structure data and the second feature vector corresponding to the mutant protein three-dimensional structure data.
[0009] Based on the variable quantum circuit, the second DDG value is obtained according to the first feature vector and the second feature vector;
[0010] The protein DDG prediction model is iteratively optimized based on the first DDG value and the second DDG value to obtain the protein DDG prediction model.
[0011] Optionally, the step of inputting the wild-type protein three-dimensional structure data and the mutant protein three-dimensional structure data into the optimized encoder of the protein DDG initial prediction model to obtain the first feature vector corresponding to the wild-type protein three-dimensional structure data and the second feature vector corresponding to the mutant protein three-dimensional structure data includes:
[0012] Microenvironmental data of the three-dimensional structure data of the wild-type protein and the three-dimensional structure data of the mutant protein were extracted respectively.
[0013] The optimized encoder is used to vectorize the microenvironmental data of the wild-type protein three-dimensional structure data and the mutant protein three-dimensional structure data respectively to obtain the corresponding prompt codes.
[0014] Based on the cue code adapter, the cue codes corresponding to the three-dimensional structure data of the wild-type protein are combined to obtain the first feature vector;
[0015] Based on the cue code adapter, the cue codes corresponding to the three-dimensional structural data of the mutant protein are combined to obtain the second feature vector.
[0016] Optionally, the step of obtaining the second DDG value based on the variable quantum circuit, according to the first feature vector and the second feature vector, includes:
[0017] Calculate the difference vector between the first feature vector and the second feature vector;
[0018] Map the difference vector onto the quantum state of the qubit;
[0019] Parameterized quantum circuits are constructed based on rotating gates, and entangled states are established between the qubits through controlled gates.
[0020] The expected value of the quantum state is measured, and the quantum state is converted into a numerical value to obtain the second DDG value.
[0021] Optionally, the method further includes a step of optimizing the encoder, comprising:
[0022] Obtain the three-dimensional structure data of multiple wild-type proteins from the training data;
[0023] For each wild-type protein's three-dimensional structure data, the amino acid residue types, angular statistical feature data, and three-dimensional spatial coordinates of the wild-type protein's three-dimensional structure data are extracted;
[0024] Hierarchical masking is performed on the amino acid residue types, the angular statistical feature data, and the three-dimensional spatial coordinates, respectively.
[0025] Based on the encoder and decoder, the amino acid residue types, angular statistical feature data and three-dimensional spatial coordinates processed by the hierarchical masking are restored and reconstructed to obtain the restored three-dimensional structure of wild-type protein.
[0026] The encoder is iteratively optimized based on the restored wild-type protein 3D structure data and the wild-type protein 3D structure data to obtain the optimized encoder.
[0027] Optionally, based on the encoder and decoder, the amino acid residue types, angular statistical feature data, and three-dimensional spatial coordinates processed by the hierarchical masking are restored and reconstructed to obtain the restored three-dimensional structure data of the wild-type protein, including:
[0028] Based on the encoder, the amino acid residue types, angular statistical feature data and three-dimensional spatial coordinates processed by the hierarchical mask are encoded and fused to obtain the corresponding hidden representation vector;
[0029] The hidden representation vector is vector quantized using a preset hint codebook to obtain the hint code;
[0030] The decoder is used to reconstruct and restore the prompt code, thereby obtaining the restored three-dimensional structural data of the wild-type protein.
[0031] Optionally, the encoder performs encoding and fusion on the amino acid residue types, angular statistical feature data, and three-dimensional spatial coordinates processed by the hierarchical masking to obtain the corresponding hidden representation vector, including:
[0032] Calculate the attention scores for the amino acid residue type, the angular statistical feature data, and the three-dimensional spatial coordinates, respectively;
[0033] Convert the attention scores into weights;
[0034] Based on the weights, the amino acid residue types, the angular statistical feature data, and the three-dimensional spatial coordinates are fused to obtain the hidden representation vector.
[0035] Optionally, the method further includes a step of preprocessing the training data, including:
[0036] Obtain multiple raw protein three-dimensional structure data;
[0037] For each of the original protein three-dimensional structure data, the loop region of the original protein three-dimensional structure data is filled with blanks to obtain the wild-type protein three-dimensional structure data;
[0038] The amino acid structure of the wild-type protein is replaced according to the preset mutation information to obtain the three-dimensional structure data of the mutant protein.
[0039] Secondly, embodiments of the present invention provide a method for predicting protein DDG, the method comprising:
[0040] Receive first or second protein information to be predicted from the user terminal, wherein the first protein information to be predicted includes the original three-dimensional structure of the protein and mutation information, and the second protein information to be predicted includes the original three-dimensional structure of the protein and the corresponding three-dimensional structure of the mutant protein.
[0041] Input the first protein information to be predicted or the second protein information to be predicted into the protein DDG prediction model described above.
[0042] Receive the DDG prediction value corresponding to the first or second protein information to be predicted.
[0043] Thirdly, embodiments of the present invention provide a protein DDG prediction model optimization device, comprising:
[0044] The training data acquisition unit is used to acquire preprocessed training data, which includes wild-type protein three-dimensional structure data, mutant protein three-dimensional structure data corresponding to the wild-type protein three-dimensional structure data, and the first DDG value of the mutation.
[0045] The feature vector acquisition unit is used to input the three-dimensional structure data of the wild-type protein and the three-dimensional structure data of the mutant protein into the optimized encoder in the protein DDG initial prediction model, respectively, to obtain the first feature vector corresponding to the three-dimensional structure data of the wild-type protein and the second feature vector corresponding to the three-dimensional structure data of the mutant protein.
[0046] The second DDG value determination unit is used to obtain the second DDG value based on the variable quantum line, according to the first feature vector and the second feature vector.
[0047] The model optimization unit is used to iteratively optimize the initial protein DDG prediction model based on the first DDG value and the second DDG value to obtain the protein DDG prediction model.
[0048] Fourthly, embodiments of the present invention provide an electronic device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of any of the methods described above.
[0049] Fifthly, embodiments of the present invention provide a computer-readable storage medium, the computer-readable storage medium including a computer program, wherein the computer program, when executed, controls a server where the computer-readable storage medium is located to implement the steps of any of the methods described above.
[0050] The beneficial effects of the embodiments of the present invention include, for example:
[0051] The three-dimensional structural data of proteins are used to construct and optimize protein DDG prediction models. By directly analyzing the spatial structure of proteins before and after mutation, the influence of protein structural changes on free energy can be captured more accurately, thereby improving the accuracy of DDG prediction. Attached Figure Description
[0052] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.
[0053] Figure 1 A block diagram illustrating an electronic device according to an embodiment of the present invention;
[0054] Figure 2 A flowchart illustrating the steps of a protein DDG prediction model optimization method provided in this embodiment of the invention;
[0055] Figure 3 A flowchart illustrating the steps of a protein DDG prediction method provided in this embodiment of the invention;
[0056] Figure 4 This is an architectural diagram of a protein DDG prediction model optimization device provided in an embodiment of the present invention.
[0057] Icons: 100 - Electronic device; 110 - Memory; 120 - Processor; 130 - Communication module; 300 - Protein DDG prediction model optimization device; 301 - Training data acquisition unit; 302 - Feature vector acquisition unit; 303 - Second DDG value determination unit; 304 - Model optimization unit. Detailed Implementation
[0058] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.
[0059] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.
[0060] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.
[0061] Furthermore, the terms "first" and "second" are used only to distinguish descriptions and should not be interpreted as indicating or implying relative importance.
[0062] It should be noted that, where there is no conflict, the features in the embodiments of the present invention can be combined with each other.
[0063] Currently, there are three main types of models in the field of protein DDG prediction: primary sequence-based models, energy-based models, and pre-trained models.
[0064] Models based on primary sequences predict protein DDG entirely based on the protein's primary sequence, such as the ESM (Evoformer Sequence Model), PSSM (Position-Specific Scoring Matrix), and MSA (Multi-Head Self-Attention) Transformer model.
[0065] This method uses limited data, and the data acquisition and algorithm implementation are relatively simple. However, because the primary sequence contains too little information, it cannot adequately represent the overall information of the protein, and therefore often fails to yield satisfactory prediction results.
[0066] Energy-based models typically utilize empirical energy functions, combined with extensive statistical and physical principles. Examples include the Rosetta model and the FoldX model. Empirical energy functions include a series of scoring terms, such as van der Waals interactions, charge interactions, hydrogen bonds, and solvation effects. Based on experimental data of known protein structures and mutations, these models are optimized to accurately predict the impact of mutations.
[0067] These methods, which combine statistical and physical principles, offer good accuracy when parameters are set correctly. However, the Rosetta model suffers from an excessive number of parameters that require fine-tuning to obtain accurate results.
[0068] Pre-trained models learn protein representations through pre-training tasks. These tasks include Protein Language Models (PLMs) and partial structure reconstruction. PLAs utilize large-scale protein sequence data for unsupervised learning, aiming to capture amino acid dependencies and local and global structural information from the sequences. Partial structure reconstruction leverages known partial protein structure data, learning protein structural representations through pre-training tasks such as "masked partial structure reconstruction" or "self-supervised sequence-to-structure prediction." The weights obtained from these training tasks are then used as a feature extractor (encoder) to map the input protein structure into a high-dimensional representation space. This representation can then serve as input to a predictive DDG model.
[0069] These methods possess rich representational capabilities and the ability to capture complex dependencies. However, designing and training the models is challenging. Efficiently and accurately extracting features to achieve both high accuracy and strong generalization requires repeated adjustments and optimizations.
[0070] Based on the above, embodiments of the present invention provide a method for optimizing a protein DDG prediction model, a method for predicting protein DDG, and a related apparatus, which can effectively alleviate the above-mentioned technical problems.
[0071] Please refer to Figure 1 This is a block diagram of an electronic device 100 provided in this application. The electronic device 100 can be a data processing device, and this embodiment does not limit this. The electronic device 100 includes a memory 110, a processor 120, and a communication module 130. The memory 110, processor 120, and communication module 130 are electrically connected directly or indirectly to achieve data transmission or interaction. For example, these components can be electrically connected to each other through one or more communication buses or signal lines.
[0072] The memory 110 is used to store programs or data. The memory 110 may be, but is not limited to, random access memory (RAM), read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), etc.
[0073] The processor 120 is used to read / write data or programs stored in memory and to perform corresponding functions.
[0074] The communication module 130 is used to establish a communication connection between the server and other communication terminals through the network, and to send and receive data through the network.
[0075] It should be understood that, Figure 1 The structure shown is only a schematic diagram of the electronic device 100. The electronic device 100 may also include components that are larger than... Figure 1 The more or fewer components shown, or having the same Figure 1 The different configurations shown. Figure 1 The components shown can be implemented using hardware, software, or a combination thereof. This electronic device 100 can be integrated into other devices or configured as a standalone device.
[0076] Corresponding to electronic device 100, this embodiment of the invention provides a method for optimizing a protein DDG prediction model. The method includes, for example: Figure 2 The following steps are shown:
[0077] Step S110: Obtain preprocessed training data, which includes wild-type protein three-dimensional structure data, mutant protein three-dimensional structure data corresponding to the wild-type protein three-dimensional structure data, and the first DDG value of the mutation.
[0078] Step S120: Input the wild-type protein three-dimensional structure data and the mutant protein three-dimensional structure data into the optimized encoder of the protein DDG initial prediction model to obtain the first feature vector corresponding to the wild-type protein three-dimensional structure data and the second feature vector corresponding to the mutant protein three-dimensional structure data.
[0079] Step S130: Based on the variable quantum circuit, obtain the second DDG value according to the first feature vector and the second feature vector.
[0080] Step S140: Iteratively optimize the initial protein DDG prediction model based on the first DDG value and the second DDG value to obtain the protein DDG prediction model.
[0081] In step S110, preprocessed training data is obtained, including wild-type protein three-dimensional structure data, mutant protein three-dimensional structure data corresponding to the wild-type protein three-dimensional structure data, and the first DDG value of the mutation.
[0082] Training data can be data used to train a model to optimize a protein DDG prediction model. In this embodiment of the invention, training data can be preprocessed raw protein three-dimensional structure data (i.e., wild-type protein three-dimensional structure data), mutant protein three-dimensional structure data generated by mutating wild-type protein three-dimensional structure data, and the true DDG value (i.e., the first DDG value) of wild-type protein three-dimensional structure data mutated into mutant protein three-dimensional structure data.
[0083] Preprocessing of training data can involve cleaning the training data and removing unreasonable wild-type protein 3D structure data, mutant protein 3D structure data, or first DDG values.
[0084] As an optional implementation, to effectively increase the amount of training data and alleviate the problems of data skewness and sample scarcity, the DDG value from wild-type protein 3D structure data to mutant protein 3D structure data can be reversed and treated as the -DDG value from mutant protein 3D structure data to wild-type protein 3D structure data. This results in twice the amount of training data as the original training data.
[0085] Optionally, the method further includes a step of preprocessing the training data, including: acquiring multiple raw protein three-dimensional structure data.
[0086] For each of the original protein three-dimensional structure data, the loop region of the original protein three-dimensional structure data is filled with blanks to obtain the wild-type protein three-dimensional structure data.
[0087] The amino acid structure of the wild-type protein is replaced according to the preset mutation information to obtain the three-dimensional structure data of the mutant protein.
[0088] In another method of preprocessing training data, multiple unprocessed raw protein 3D structure data can be obtained first. For each raw protein 3D structure data, the loop region of the raw protein 3D structure data is filled with blanks using software tools (such as ChimeraX, Modeller, and Schrodinger) to correct potential geometric and energy unstable regions, thus obtaining wild-type protein 3D structure data.
[0089] Then, based on pre-defined mutation information (such as the location of amino acid mutations and the type of amino acids after mutation), amino acid replacements are performed on the wild-type protein's 3D structure data. For example, amino acids in the MPAST sequence are modified to MPASL to obtain the mutant protein's 3D structure data, ensuring the rationality and accuracy of the protein structure. Finally, the wild-type protein's 3D structure data and the mutant protein's 3D structure data are used as preprocessed training data.
[0090] In step S120, the three-dimensional structure data of the wild-type protein and the three-dimensional structure data of the mutant protein are respectively input into the optimized encoder of the protein DDG initial prediction model to obtain the first feature vector corresponding to the three-dimensional structure data of the wild-type protein and the second feature vector corresponding to the three-dimensional structure data of the mutant protein.
[0091] The encoder in the initial prediction model for protein DDG can be a module used to encode the three-dimensional structure data of the protein into feature vectors. In this embodiment of the invention, the encoder can be optimized first to make the feature vectors encoded by the encoder more accurate in reconstructing the three-dimensional structure data of the protein.
[0092] After obtaining the three-dimensional structure data of wild-type protein and mutant protein, they are respectively input into the optimized encoder. After encoding processing by the encoder, the first feature vector corresponding to the three-dimensional structure data of wild-type protein and the second feature vector corresponding to the three-dimensional structure data of mutant protein are obtained.
[0093] Optionally, the method further includes a step of optimizing the encoder, comprising:
[0094] Obtain the three-dimensional structure data of multiple wild-type proteins from the training data.
[0095] For each wild-type protein's three-dimensional structure data, the amino acid residue types, angular statistical feature data, and three-dimensional spatial coordinates of the wild-type protein's three-dimensional structure data are extracted.
[0096] The amino acid residue types, the angular statistical feature data, and the three-dimensional spatial coordinates are respectively subjected to hierarchical masking processing.
[0097] Based on the encoder and decoder, the amino acid residue types, angular statistical feature data and three-dimensional spatial coordinates processed by the hierarchical masking are restored and reconstructed to obtain the restored three-dimensional structure of the wild-type protein.
[0098] The encoder is iteratively optimized based on the restored wild-type protein 3D structure data and the wild-type protein 3D structure data to obtain the optimized encoder.
[0099] In one alternative implementation, the encoder can be optimized in the following way:
[0100] Multiple wild-type protein 3D structure data were randomly obtained from the training data. For each wild-type protein 3D structure data, the amino acid residue type (one-dimensional feature), angular statistical feature data (two-dimensional feature), and three-dimensional spatial coordinates (three-dimensional feature) were extracted from the 3D structure data.
[0101] Hierarchical masking was performed on amino acid residue types, angular statistical feature data, and 3D spatial coordinates. The encoder then encoded the hierarchically masked amino acid residue types, angular statistical feature data, and 3D spatial coordinates into feature vectors. The decoder then reconstructed the 3D structure of the wild-type protein using these feature vectors. The encoder was optimized by comparing the reconstructed wild-type protein 3D structure data with the wild-type protein 3D structure data, resulting in an optimized encoder.
[0102] Optionally, based on the encoder and decoder, the amino acid residue types, angular statistical feature data, and three-dimensional spatial coordinates processed by the hierarchical masking are restored and reconstructed to obtain the restored three-dimensional structure data of the wild-type protein, including:
[0103] Based on the encoder, the amino acid residue types, angular statistical feature data and three-dimensional spatial coordinates processed by the hierarchical mask are encoded and fused to obtain the corresponding hidden representation vector.
[0104] The hidden representation vector is vector quantized using a preset prompt codebook to obtain the prompt code.
[0105] The decoder is used to reconstruct and restore the prompt code, thereby obtaining the restored three-dimensional structural data of the wild-type protein.
[0106] As an optional implementation, specifically, an encoder can encode and fuse the amino acid residue types, angular statistical feature data, and three-dimensional spatial coordinates after hierarchical masking to obtain the corresponding hidden representation vector. A preset cue codebook is then used to perform vector quantization on the hidden representation vector to obtain a cue code. Finally, a decoder reconstructs and restores the cue code to obtain the restored three-dimensional structure data of the wild-type protein.
[0107] For example, cross-entropy is used to predict the residue types of the masked residue set; the difference between the predicted angle and the true angle is calculated using the MSE loss function; and Huber loss is used to predict the difference between the noisy structure and the true structure through Gaussian noise estimation. Finally, the cue code is reconstructed using a decoder to obtain the restored three-dimensional structure data of the wild-type protein.
[0108] Optionally, the encoder performs encoding and fusion on the amino acid residue types, angular statistical feature data, and three-dimensional spatial coordinates processed by the hierarchical masking to obtain the corresponding hidden representation vector, including:
[0109] Attention scores are calculated for the amino acid residue type, the angular statistical feature data, and the three-dimensional spatial coordinates, respectively. Each attention score is then converted into a weight.
[0110] Based on the weights, the amino acid residue types, the angular statistical feature data, and the three-dimensional spatial coordinates are fused to obtain the hidden representation vector.
[0111] In an alternative implementation, the hidden representation vector can be obtained using Graph Attention Networks (GATs).
[0112] First, the attention scores for amino acid residue types, angular statistical feature data, and three-dimensional spatial coordinates are calculated separately, using the following formulas:
[0113]
[0114]
[0115] L pair =WZ
[0116]
[0117] Among them, L node L represents the attention score between nodes. pair For the correlation between node pairs, L spatialThe similarity of spatial location corresponds to amino acid residue type, angular statistical feature data, and three-dimensional spatial coordinates, respectively.
[0118] After obtaining the attention scores based on amino acid residue types, angular statistical features, and three-dimensional spatial coordinates, the attention scores are converted into weights using the following formula:
[0119] α = softmax(L)
[0120] Where α is the weight.
[0121] Then, feature fusion is performed on the amino acid residue type, the angular statistical feature data, and the three-dimensional spatial coordinates based on the weights, as shown in the following formula:
[0122] F agg =[F node ;F pair ;F spatial ]
[0123]
[0124]
[0125]
[0126] Finally, the node features are updated to obtain the hidden representation vector, as shown in the following formula:
[0127] feats=LayerNorm(x+MLP(F agg ))
[0128] Where W represents the learnable weight matrix, Qn, Kn, and Vn are the projections of the query, key, and value of a node, Qp, Kp, and Vp are the projections of the query, key, and value of a point, dn is the node feature dimension, P is the number of spatial points, γ is the scaling factor for spatial attention, and α... ij Let represent the attention weight of node i to node j, norm() is the normalization function, LayerNorm is the layer normalization, and MLP is the multilayer perceptron mechanism.
[0129] Optionally, the step of inputting the wild-type protein three-dimensional structure data and the mutant protein three-dimensional structure data into the optimized encoder of the protein DDG initial prediction model to obtain the first feature vector corresponding to the wild-type protein three-dimensional structure data and the second feature vector corresponding to the mutant protein three-dimensional structure data includes:
[0130] Microenvironmental data of the three-dimensional structure data of the wild-type protein and the three-dimensional structure data of the mutant protein were extracted respectively.
[0131] The optimized encoder is used to vectorize the microenvironmental data of the wild-type protein three-dimensional structure data and the mutant protein three-dimensional structure data, respectively, to obtain the corresponding prompt codes.
[0132] Based on the cue code adapter, the cue codes corresponding to the three-dimensional structure data of the wild-type protein are combined to obtain the first feature vector.
[0133] Based on the cue code adapter, the cue codes corresponding to the three-dimensional structural data of the mutant protein are combined to obtain the second feature vector.
[0134] The microenvironment refers to the environment within or around a specific region of a protein molecule. The structure and chemical properties of this region can significantly affect the protein's structure, function, and interactions with other molecules. In this embodiment of the invention, it can refer to residue type, angular statistics, and local conformational changes. As an optional implementation, microenvironment data can be extracted from the three-dimensional structure data of wild-type and mutant proteins, respectively. Each microenvironment data is then encoded using an optimized encoder to obtain corresponding cue codes. A cue code adapter is then used to combine the cue codes corresponding to the three-dimensional structure data of the wild-type protein to obtain a first feature vector; and to combine the cue codes corresponding to the three-dimensional structure data of the mutant protein to obtain a second feature vector.
[0135] In step S130, based on the variable quantum circuit, the second DDG value is obtained according to the first feature vector and the second feature vector.
[0136] The Variational Quantum Classifier (VQC) can improve prediction performance and computational efficiency, further enhancing the prediction accuracy of DDG. Therefore, the first and second eigenvectors can be processed and fed into the VQC to obtain the DDG prediction value (i.e., the second DDG value).
[0137] Optionally, the step of obtaining the second DDG value based on the variable quantum circuit, according to the first feature vector and the second feature vector, includes:
[0138] Calculate the difference vector between the first feature vector and the second feature vector.
[0139] The difference vector is mapped onto the quantum state of the qubit.
[0140] A parameterized quantum circuit is constructed based on a rotating gate, and an entangled state is established between the qubits through a controlled gate.
[0141] The expected value of the quantum state is measured, and the quantum state is converted into a numerical value to obtain the second DDG value.
[0142] In one alternative implementation, the difference vector between the first feature vector and the second feature vector can be calculated first. For example, the first feature vector and the second feature vector can be represented as difference vector 1 and difference vector 2, where difference vector 1 represents the difference between the residue type after mutation and the residue type before mutation; and difference vector 2 represents the difference in change of the mutant relative to the wild type.
[0143] The difference vector is mapped onto the quantum state of the qubit. A parameterized quantum circuit is constructed using rotation gates (such as RX, RY, and RZ gates) with rotation angles that can be trained parameters. Entangled states are then established between the qubits through controlled gates. Finally, the expected value of the quantum state is measured, and the quantum state is converted into a numerical value to obtain the second DDG value.
[0144] In step S140, the protein DDG initial prediction model is iteratively optimized based on the first DDG value and the second DDG value to obtain the protein DDG prediction model.
[0145] After obtaining the second DDG value, the loss is calculated by comparing it with the first DDG value obtained from the training data. The initial protein DDG prediction model is optimized by minimizing the mean squared error loss between the second DDG value and the first DDG value, and finally the protein DDG prediction model is obtained.
[0146] Based on the same inventive concept, such as Figure 3 As shown in the embodiments of this invention, a method for predicting protein DDG is provided, the method comprising:
[0147] Step S210: Receive the first or second protein information to be predicted input from the user terminal, wherein the first protein information to be predicted includes the original three-dimensional structure of the protein and mutation information, and the second protein information to be predicted includes the original three-dimensional structure of the protein and the corresponding three-dimensional structure of the mutant protein.
[0148] Step S220: Input the first protein information to be predicted or the second protein information to be predicted into the protein DDG prediction model described above.
[0149] Step S230: Receive the DDG prediction value corresponding to the first protein information to be predicted or the second protein information to be predicted.
[0150] The original protein 3D structure can be a PDB file, and the mutation information can include the mutation site, the chain at the mutation site, and the amino acid types before and after the mutation. If the user inputs the information of the first protein to be predicted, the PDB file of the original protein 3D structure and the corresponding mutation information are input into the protein DDG prediction model to obtain the corresponding DDG prediction value. It should be noted that if the first protein to be predicted is input, cases where the sequence length has changed due to mutation are not accepted.
[0151] The 3D structure of the mutant protein corresponding to the original 3D protein structure can also be a PDB file. If the user inputs information about a second protein to be predicted, the PDB files of the original protein's 3D structure and the corresponding PDB files of the mutant protein's 3D structure are input into the protein DDG prediction model to obtain the corresponding DDG prediction value. If information about a second protein to be predicted is input, the case where the sequence length changes due to mutation is accepted.
[0152] Based on the same inventive concept, such as Figure 4 As shown in the embodiment of this invention, a protein DDG prediction model optimization device 300 is provided, comprising:
[0153] The training data acquisition unit 301 is used to acquire preprocessed training data, which includes wild-type protein three-dimensional structure data, mutant protein three-dimensional structure data corresponding to the wild-type protein three-dimensional structure data, and the first DDG value of the mutation.
[0154] The feature vector acquisition unit 302 is used to input the wild-type protein three-dimensional structure data and the mutant protein three-dimensional structure data into the optimized encoder in the protein DDG initial prediction model, respectively, to obtain the first feature vector corresponding to the wild-type protein three-dimensional structure data and the second feature vector corresponding to the mutant protein three-dimensional structure data.
[0155] The second DDG value determination unit 303 is used to obtain a second DDG value based on the variable quantum circuit, according to the first feature vector and the second feature vector.
[0156] The model optimization unit 304 is used to iteratively optimize the initial protein DDG prediction model based on the first DDG value and the second DDG value to obtain the protein DDG prediction model.
[0157] Regarding the protein DDG prediction model optimization device 300 described above, the specific functions of each unit have been described in detail in the embodiments of the protein DDG prediction model optimization method provided in this specification, and will not be elaborated here.
[0158] Based on the same inventive concept, embodiments of this invention provide a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of any of the methods in the aforementioned protein DDG prediction model optimization method.
[0159] The present invention has at least the following beneficial effects:
[0160] 1. The three-dimensional structural data of proteins are used to construct and optimize the protein DDG prediction model. By directly analyzing the spatial structure of proteins before and after mutation, the influence of protein structural changes on free energy can be captured more accurately, thereby improving the accuracy of DDG prediction.
[0161] 2. Introducing variable quantum circuits improves prediction performance and computational efficiency, further enhancing the prediction accuracy of DDG.
[0162] 3. Before optimizing the model, the encoder is optimized based on the graph attention network to make the encoder output more accurate.
[0163] In the several embodiments provided by this invention, it should be understood that the disclosed apparatus and methods can also be implemented in other ways. The apparatus embodiments described above are merely illustrative; for example, the flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of apparatus, methods, and computer program products according to various embodiments of the invention. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than those marked in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram and / or flowchart, and combinations of blocks in block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.
[0164] In addition, the functional modules in the various embodiments of the present invention can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.
[0165] If the aforementioned functions are implemented as software functional modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, essentially, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.
[0166] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.
Claims
1. A method for optimizing a protein DDG prediction model, characterized in that, The method includes: Acquire preprocessed training data, which includes wild-type protein three-dimensional structure data, mutant protein three-dimensional structure data corresponding to the wild-type protein three-dimensional structure data, and the first DDG value of the mutation; The wild-type protein three-dimensional structure data and the mutant protein three-dimensional structure data are respectively input into the optimized encoder of the protein DDG initial prediction model to obtain the first feature vector corresponding to the wild-type protein three-dimensional structure data and the second feature vector corresponding to the mutant protein three-dimensional structure data. Based on a variable quantum circuit, the second DDG value is obtained according to the first eigenvector and the second eigenvector, including: calculating the difference vector between the first eigenvector and the second eigenvector; mapping the difference vector to the quantum state of the qubit; constructing a parameterized quantum circuit based on a rotating gate, and establishing an entangled state between the qubits through a controlled gate; measuring the expected value of the quantum state, and converting the quantum state into a numerical value to obtain the second DDG value. The protein DDG prediction model is iteratively optimized based on the first DDG value and the second DDG value to obtain the protein DDG prediction model.
2. The protein DDG prediction model optimization method as described in claim 1, characterized in that, The step of inputting the wild-type protein's three-dimensional structure data and the mutant protein's three-dimensional structure data into the optimized encoder of the protein DDG initial prediction model to obtain the first feature vector corresponding to the wild-type protein's three-dimensional structure data and the second feature vector corresponding to the mutant protein's three-dimensional structure data includes: Microenvironmental data of the three-dimensional structure data of the wild-type protein and the three-dimensional structure data of the mutant protein were extracted respectively. The optimized encoder is used to vectorize the microenvironmental data of the wild-type protein three-dimensional structure data and the mutant protein three-dimensional structure data respectively to obtain the corresponding prompt codes. Based on the cue code adapter, the cue codes corresponding to the three-dimensional structure data of the wild-type protein are combined to obtain the first feature vector; Based on the cue code adapter, the cue codes corresponding to the three-dimensional structural data of the mutant protein are combined to obtain the second feature vector.
3. The protein DDG prediction model optimization method as described in claim 1, characterized in that, The method also includes a step of optimizing the encoder, including: Obtain the three-dimensional structure data of multiple wild-type proteins from the training data; For each wild-type protein's three-dimensional structure data, the amino acid residue types, angular statistical feature data, and three-dimensional spatial coordinates of the wild-type protein's three-dimensional structure data are extracted; Hierarchical masking is performed on the amino acid residue types, the angular statistical feature data, and the three-dimensional spatial coordinates, respectively. Based on the encoder and decoder, the amino acid residue types, angular statistical feature data and three-dimensional spatial coordinates processed by the hierarchical masking are restored and reconstructed to obtain the restored three-dimensional structure of wild-type protein. The encoder is iteratively optimized based on the restored wild-type protein 3D structure data and the wild-type protein 3D structure data to obtain the optimized encoder.
4. The protein DDG prediction model optimization method as described in claim 3, characterized in that, Based on the encoder and decoder, the amino acid residue types, angular statistical feature data, and three-dimensional spatial coordinates processed by the hierarchical masking are restored and reconstructed to obtain the restored three-dimensional structure data of the wild-type protein, including: Based on the encoder, the amino acid residue types, angular statistical feature data and three-dimensional spatial coordinates processed by the hierarchical mask are encoded and fused to obtain the corresponding hidden representation vector; The hidden representation vector is vector quantized using a preset hint codebook to obtain the hint code; The decoder is used to reconstruct and restore the prompt code, thereby obtaining the restored three-dimensional structural data of the wild-type protein.
5. The protein DDG prediction model optimization method as described in claim 4, characterized in that, The encoder is used to encode and fuse the amino acid residue types, angular statistical feature data, and three-dimensional spatial coordinates processed by the hierarchical masking to obtain the corresponding hidden representation vector, including: Calculate the attention scores for the amino acid residue type, the angular statistical feature data, and the three-dimensional spatial coordinates, respectively; Convert the attention scores into weights; Based on the weights, the amino acid residue types, the angular statistical feature data, and the three-dimensional spatial coordinates are fused to obtain the hidden representation vector.
6. The method for optimizing the protein DDG prediction model as described in claim 1, characterized in that, The method further includes a step of preprocessing the training data, including: Obtain multiple raw protein three-dimensional structure data; For each of the original protein three-dimensional structure data, the loop region of the original protein three-dimensional structure data is filled with blanks to obtain the wild-type protein three-dimensional structure data; The amino acid structure of the wild-type protein is replaced according to the preset mutation information to obtain the three-dimensional structure data of the mutant protein.
7. A method for predicting DDG in proteins, characterized in that, The method includes: Receive first or second protein information to be predicted from the user terminal, wherein the first protein information to be predicted includes the original three-dimensional structure of the protein and mutation information, and the second protein information to be predicted includes the original three-dimensional structure of the protein and the corresponding three-dimensional structure of the mutant protein. Input the first or second protein information to be predicted into the protein DDG prediction model according to any one of claims 1 to 6; Receive the DDG prediction value corresponding to the first or second protein information to be predicted.
8. A device for optimizing a protein DDG prediction model, characterized in that, include: The training data acquisition unit is used to acquire preprocessed training data, which includes wild-type protein three-dimensional structure data, mutant protein three-dimensional structure data corresponding to the wild-type protein three-dimensional structure data, and the first DDG value of the mutation. The feature vector acquisition unit is used to input the three-dimensional structure data of the wild-type protein and the three-dimensional structure data of the mutant protein into the optimized encoder in the protein DDG initial prediction model, respectively, to obtain the first feature vector corresponding to the three-dimensional structure data of the wild-type protein and the second feature vector corresponding to the three-dimensional structure data of the mutant protein. The second DDG value determination unit is used to obtain a second DDG value based on a variable quantum circuit, according to the first eigenvector and the second eigenvector, including: calculating the difference vector between the first eigenvector and the second eigenvector; mapping the difference vector to the quantum state of the qubit; constructing a parameterized quantum circuit based on a rotating gate, and establishing an entangled state between the qubits through a controlled gate; measuring the expected value of the quantum state, and converting the quantum state into a numerical value to obtain the second DDG value. The model optimization unit is used to iteratively optimize the initial protein DDG prediction model based on the first DDG value and the second DDG value to obtain the protein DDG prediction model.
9. An electronic device, characterized in that, include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, implements the steps of the method according to any one of claims 1 to 6.
10. A computer-readable storage medium, characterized in that, The computer-readable storage medium includes a computer program, which, when executed, controls the server where the computer-readable storage medium is located to implement the steps of the method according to any one of claims 1 to 6.