Method and system for predicting compound-protein affinity based on protein three-dimensional structure
Through the multimodal end-to-end model, the sequence and structural information of proteins are combined with the deep map convolution network and multi-head attention algorithm, the problem of insufficient utilization of protein three-dimensional structural information in the existing technology is solved, and the accuracy and accuracy of compound-protein affinity prediction is improved.
Patent Information
- Application Number
- CN202210828457.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Priority Date
- 2022-03-29
- Filing Date
- 2022-07-13
- Publication Date
- 2025-08-26
- Estimated Expiration
- 2042-07-13
AI Technical Summary
The prior art is difficult to effectively utilize the three-dimensional structural information of proteins in the prediction of compound-protein affinity, resulting in insufficient prediction accuracy, especially in the structural-free model, the three-dimensional spatial structural information is ignored, and the structure-based methods lack sufficient three-dimensional information applications.
The multimodal end-to-end model is used to combine the sequence information and structural information of the protein, and the compound characteristics are extracted through deep map convolution networks and multi-head attention algorithms, and a co-evolution strategy and affinity learning unit are introduced to optimize the characteristic interaction between proteins and compounds, and affinity prediction is used to use the three-dimensional structural information of the proteins to make affinity prediction.
The accuracy and accuracy of compound-protein affinity prediction are improved, and by integrating the three-dimensional structure and sequence information of proteins, more comprehensive feature extraction and interaction are achieved, improving the performance of the prediction model.
Smart Images

Figure CN116959555B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of drug research and development, and in particular to a method and system for predicting compound-protein affinity based on protein three-dimensional structure. Background Art
[0002] Accurately predicting the binding affinity of compounds to proteins is the main challenge in the virtual drug candidate screening process. Using computational methods to conduct effective high-throughput virtual screening of lead compounds can greatly accelerate the drug development process by reducing R&D time and experimental workload. Especially in recent years, with the development of technology, more and more data sharing projects have been vigorously promoted, and a large number of new biomedical data sources have become accessible. For example, PubChem currently contains more than 111 million compound structures and 271 million bioactivity data points (reference https: / / doi.org / 10.1093 / nar / gkaa971), which greatly increases the predictive potential of compound-protein binding affinity (CPA). However, further improving the accuracy of predicted CPA remains the main challenge facing virtual drug screening technology before it is actually applied. In order to improve the accuracy of CPA prediction, the development of computational methods basically follows two paths: structure-free or structure-based. Although these two methods have helped to improve the accuracy of CPA prediction, both methods also face the challenge of further improving their CPA prediction accuracy.
[0003] Non-structural parameter models consider the compound (drug) as atom-by-atom and the protein (target) as residue-by-residue information. They calculate pairwise atom-residue distances to form a [number of atoms]-[number of residues] pairwise interaction matrix. A compound atom is considered to be in contact (i.e., non-covalently interacting) with the corresponding protein residue below a certain threshold. Each element in the pairwise interaction matrix is a binary value, indicating whether the corresponding atom-residue pair interacts. These non-structural parameter models rely on the sequences of the drug and target to learn interactions from the pairwise matrix, aiming to predict drug-target affinity. For example, multi-layer one-dimensional convolutional neural networks (1D-CNNs) are used to extract features from protein sequences. These features are then fed into a downstream deep learning (DL) model to ultimately learn drug-target affinity. However, these interactions are constantly changing over time and with environmental changes, as protein structure is dynamic, while the acquired pairwise information is only a single state captured by electron microscopy. Therefore, this structure-free parameter model that ignores the binding site information in the three-dimensional spatial structure and only extracts information from a single type of protein feature (considering only sequence features) is likely to have a great limitation in CPA prediction.
[0004] Previous studies have shown that the predicted pairwise non-covalent traction results can be used in structure-free models to further improve their accuracy in CPA prediction. However, simply using the area under the curve (AUC Area Under the Curve) as an evaluation metric to measure the performance of pairwise matrix prediction may be misleading in the interpretation of the results. As an evaluation metric, AUC is the area under the receiver characteristic curve (ROC), and its Y-axis and X-axis are measures of sensitivity, which are used to evaluate the model's ability to predict true positives or true negatives. It draws a graph by considering different thresholds for distinguishing positive and negative values in binary classification. Figure 16 As shown, the compound-protein pair matrix contains a high number of negative values and only a few positive (compound-protein interaction) values. For example, in a dataset with 5 positive pairs and 11,595 negative pairs, the AUC remains high (over 0.95) even when the model predicts an all-negative matrix. The predicted pairwise interaction matrix is an intermediate result generated during model training and is inappropriately fed into the original CPA prediction model. Mathematically, any intermediate result generated during machine learning model training is merely a representation of the input features (i.e., compound sequence, protein sequence, and pairwise interaction matrix). Theoretically, when this representation, along with its upstream features, is fed back into the final reaction prediction (i.e., CPA), the reaction remains a representation of the original input features. In our experiments, we found that feeding the intermediate results along with the original features into an unstructured model did not benefit CPA prediction, indicating that unstructured models cannot improve CPA prediction accuracy without the input of new information. Therefore, models built on this approach may struggle to achieve further accuracy improvements.
[0005] One structure-based approach for predicting CPA is molecular docking. Docking is a bioinformatics modeling process between a target protein fragment and a ligand, which considers potential binding sites and the three-dimensional structure of the protein-drug complex during prediction. However, only a few protein 3D structures are currently available, significantly limiting the application of this approach. To overcome this limitation, an increasing number of approaches based on machine learning (ML) algorithms have been proposed. These methods rely not only on sequence data of the drug and target but also on three-dimensional spatial information (e.g., 3D coordinates, distances between the drug and target in the ligand, and torsion angles between protein residues) to achieve more accurate CPA prediction. Although theoretically attractive, developing a practical paradigm to leverage protein 3D structural information for CPA prediction has been extremely difficult.
[0006] Currently, the main obstacle to applying protein three-dimensional information to computational models is the lack of suitable methods or models that can rationally utilize protein three-dimensional information. Some studies have directly introduced three-dimensional structural information into traditional models, but have not achieved good prediction results. Most existing computational methods for compound-protein affinity prediction also use sequence information to predict affinity values, but the prediction results on some data sets have not significantly improved. Developing a practical multimodal model to rationally utilize protein three-dimensional structural information and sequence information to further improve the accuracy of CPA prediction remains an urgent need. Summary of the Invention
[0007] In order to overcome the defects of the prior art, the purpose of the present invention is to propose a reasonable characterization method for the three-dimensional structure of proteins, and to develop a multimodal end-to-end model that can simultaneously aggregate the sequence information and structural information of proteins. After characterizing and aggregating the structural information and sequence information of proteins, a co-evolutionary attention algorithm is applied to optimize the extraction of the structural features and sequence features of proteins; a graph convolution layer with special residual links is introduced to avoid the over-smoothing problem encountered in deep graph convolution, so as to more comprehensively extract compound features; an interactive attention algorithm is introduced to realize the feature interaction between proteins and compounds in the affinity prediction module, so as to learn the potential interaction information between proteins and compounds and improve the prediction accuracy of affinity.
[0008] The present invention provides a method for predicting compound-protein affinity based on protein three-dimensional structure, which is used to predict compound-protein affinity. Compound features are extracted from compounds, including atomic features and aggregate node features. Protein features containing implicit protein three-dimensional information are extracted from proteins. Compound features and protein features are then combined with affinity prediction algorithms to predict compound-protein affinity. The method specifically includes the following steps:
[0009] S1, compound feature extraction step,
[0010] According to the compound characterization graph, the initial atomic features and aggregate node features are obtained. Using the deep graph convolutional network and multi-head attention mechanism, the atomic features and aggregate node features are cyclically updated to output the final atomic features and aggregate node features of the compound.
[0011] S2, protein feature extraction step;
[0012] Obtain protein sequence information and protein structural features; Based on the protein sequence information and protein structural features, a protein feature aggregation algorithm is used to make the protein embedded sequence features carry protein structural features, thereby obtaining protein embedded sequence features and embedded structural features; By adopting a co-evolution strategy, the embedded sequence features and embedded structural features are cyclically updated, and finally the updated protein sequence features and protein structural features are obtained;
[0013] S3 compound and protein affinity prediction step;
[0014] According to the atomic features and aggregate node features of the compound obtained in step S1 and the protein sequence features in step S2, the affinity learning unit algorithm is used to obtain a predicted affinity value. The larger the affinity value, the greater the probability of affinity between the compound and the protein.
[0015] Preferably, the step S1 is specifically as follows:
[0016] S11, obtain atomic features and aggregation node features based on the compound characterization graph;
[0017] S12, the atomic features are input into the deep graph convolutional network, the output of the deep graph convolutional network and the aggregated node features are added according to the first weight of the compound features, and the updated atomic features are obtained by combining the first gated recurrent function (GRU) and the atomic features;
[0018] S13, the atomic features are input into the multi-head attention representation algorithm, the output of the multi-head attention representation algorithm and the aggregate node features are added according to the second weight of the compound features, and the updated aggregate node features are obtained by combining the second gated recurrent function (GRU) and the aggregate node features;
[0019] S14: After repeating steps S12-S13 K times using the updated atomic features and the updated aggregate node features as the atomic features and aggregate node features, the atomic features and the aggregate node features are output.
[0020] Preferably, the deep graph convolutional network in step S1 has a residual connection loop for aggregating the neighbor information of the atom onto the atom and avoiding the over-smoothing problem caused by deepening the number of network layers. GRU is used to aggregate the information between the front and back layers of the network, and a multi-head attention mechanism characterization algorithm is used to obtain diverse compound features, thereby improving the accuracy of affinity prediction in the affinity prediction step.
[0021] Preferably, the step S2 is specifically as follows:
[0022] S21, obtaining protein sequence information and protein structural features, where the protein structural features at least include a discretized distance matrix;
[0023] S22, the protein feature aggregation algorithm is specifically as follows: the protein sequence information is passed through the first embedding layer (Eembedding layer) to obtain a sequence vector, the discretized distance matrix is passed through the second embedding layer (Eembedding layer) to obtain a discrete distance matrix feature vector, and the discrete distance matrix feature vector is used as the embedded distance matrix after feature embedding;
[0024] S23, the co-evolution strategy is implemented using an N-layer protein encoding algorithm. The protein encoding algorithm of each layer is the same. The specific protein encoding algorithm of one layer is:
[0025] The two results of the embedded distance matrix after row addition and column addition are fused through the gated recurrent unit to obtain the spliced embedded structural information. The spliced embedded structural information then enters the diversity convolution layer to obtain diversified protein structural features, thereby learning protein features with more diverse features.
[0026] The embedded sequence features are sequentially passed through the diversity convolution layer and the common convolution layer to obtain diverse protein sequence features. The diverse protein structure features and diverse protein sequence features are added together through the gating logic, and then sequentially passed through the gated recurrent unit and the convolution layer to obtain a structural feature vector that contains both structural and sequence information.
[0027] The structural feature vector undergoes its own binding transformation (outer sum) to output the protein structural features;
[0028] The updated sequence features are embedded in a diversity convolutional layer to obtain the updated sequence features after diversified embedding. The updated sequence features after diversified embedding are summed with the diversified protein structure features through gated logic to obtain a sequence feature vector that integrates the structure information and sequence information. The diversified protein sequence features are further fused with the sequence feature vector through a gated recurrent unit to output the protein sequence features.
[0029] S24, the new embedding distance matrix output by the protein coding algorithm is the required embedding distance matrix. The obtained embedding distance matrix and embedded sequence features are used as the input of the next layer of protein coding algorithm to continue updating the embedding distance matrix and embedded sequence features until all N layers of protein coding algorithms are completed. At this time, the output embedding distance matrix is the structural feature of the protein, and the embedded sequence feature is the sequence feature of the protein.
[0030] Preferably, the protein structural features in step S21 also include a feature auxiliary matrix; the embedded sequence features obtained in step S22 are further added to the feature auxiliary matrix through gating logic to obtain the embedded updated sequence features. The diversity convolution layer in step S23 divides the input feature vector into four equal parts in the feature dimension, and enters four parallel ordinary convolution layers respectively, and sums the output results. At the same time, the initial features of the sequence information input are aggregated through residual connections to obtain a diversified feature vector.
[0031] Preferably, the characteristic auxiliary matrix is a torsion angle matrix, which is constructed using the sine and cosine values of the dihedral angles φ and ψ between the α-carbon atom of the protein backbone chain and the amino and carboxyl groups.
[0032] Preferably, the discretized distance matrix in step S21 divides the distances between residues in the stable structure of the protein into M equidistant mapping intervals that are statistically substantially in line with the normal distribution, thereby realizing the discretized encoding of the distance matrix, where M is a positive integer.
[0033] Preferably, the step S3 is specifically as follows:
[0034] S31, receiving the atomic features and aggregate node features of the compound from step S1 as compound information and passing them into the affinity learning unit, updating the aggregate node features through a linear layer to obtain updated aggregate node features; updating the compound features through a linear layer, and at the same time obtaining the mean in the atomic dimension, and performing feature splicing with the updated aggregate node features to obtain a comprehensive compound feature;
[0035] S32, receiving the sequence features of the protein from step S2 and passing the sequence features of the protein as protein information into the affinity learning unit, the protein sequence features are updated through the convolution layer, and the average is calculated in the residue dimension to obtain the comprehensive protein features;
[0036] The protein comprehensive features and compound comprehensive features are matrix multiplied (matmul) and the information of the two is fused through a linear layer to obtain the predicted affinity value.
[0037] The present invention also discloses a system for predicting compound-protein affinity based on protein three-dimensional structure, comprising: a compound extractor, a protein extractor and an affinity predictor;
[0038] The compound extractor uses a deep graph convolutional network and a multi-head attention algorithm to update the atomic features and aggregate node features of the input compound through the attention mechanism, and outputs the final atomic features and aggregate node features through a loop; the final compound features and aggregate node features are sent to the affinity predictor;
[0039] The protein extractor obtains embedded sequence features and embedded structural features based on the input protein sequence features and protein structural features through a protein feature aggregation algorithm; and updates the embedded sequence features and embedded structural features by adopting a co-evolution strategy to obtain updated protein sequence features and protein structural features;
[0040] The affinity predictor is used to receive compound features, protein sequence features and aggregation node features, and obtain the predicted affinity value through the affinity learning unit algorithm.
[0041] Preferably, the compound extractor includes a deep graph convolution unit and a multi-head attention representation unit, the deep graph convolution unit is used to implement a deep graph convolution network, and the multi-head attention representation unit is used to implement a multi-head attention algorithm; the protein extractor includes a protein information aggregation unit and a co-evolution update unit, the protein information aggregation unit is used to implement a protein feature aggregation algorithm, and the co-evolution update unit is used to implement a co-evolution strategy; the affinity predictor includes an affinity learning unit.
[0042] Compared with the prior art, the present invention has the following beneficial effects:
[0043] 1. When using a distance matrix to record the distance between protein residues, a discretized distance matrix is used to reduce the dimension of the data representation vector in the word embedding stage. In addition, the torsion angles (dihedral angles) φ and ψ between the α-carbon atom and the amino and carboxyl groups in the protein backbone are used to construct a torsion matrix, which accurately represents the direction of the sequence, fully reflects the three-dimensional structure of the protein, and solves the problem of high-dimensional protein representation.
[0044] 2. Introducing a co-evolutionary strategy to collaboratively update protein sequence and structure information, strengthening the correlation between protein sequence and structure information, and extracting more comprehensive and diverse feature information in the final protein representation vector to characterize the protein;
[0045] 3. A residual structure with initial information of graph nodes is introduced to avoid oversmoothing. On this basis, an aggregation node is introduced to aggregate the global information of the compound. The graph convolution process is supplemented by a gated update unit and a message passing mechanism of graph information to achieve information interaction and update between each atomic node and the aggregation node in the graph structure.
[0046] 4. An affinity learning unit is introduced into the affinity prediction unit to aggregate the features of proteins and compounds to more comprehensively learn the potential interaction information between proteins and compounds, thereby improving the prediction accuracy of affinity. BRIEF DESCRIPTION OF THE DRAWINGS
[0047] Figure 1This is a flow chart of the method for predicting compound-protein affinity based on protein three-dimensional structure;
[0048] Figure 2 It is a flowchart of the specific steps for compound composite feature extraction;
[0049] Figure 3 It is a flowchart of the deep graph convolutional network;.
[0050] Figure 4 It is a flowchart of the specific steps of the deep graph convolutional network;
[0051] Figure 5 It is a schematic diagram of the atomic feature extraction process;
[0052] Figure 6 It is a schematic diagram of the protein feature extraction process;
[0053] Figure 7 This is a flowchart of the specific steps of the protein feature aggregation algorithm;
[0054] Figure 8 This is a flowchart of the specific steps of the single-layer protein encoding algorithm;
[0055] Figure 9 This is a flow chart of the diversity convolutional layer;
[0056] Figure 10 It is a flowchart of the single-layer protein encoding algorithm;
[0057] Figure 11 It is a flowchart of the specific steps of the affinity learning unit algorithm;
[0058] Figure 12 The performance of each model is compared on a dataset (small dataset) after similar clustering of compounds based on molecular fingerprints;
[0059] Figure 13 This is a comparison of the performance of various models based on a data set (small data set) after homology clustering of proteins based on a multiple sequence alignment strategy;
[0060] Figure 14 It is a comparison of the convergence of each model during the training process (model convergence speed and final convergence position);
[0061] Figure 15 This is a graph showing the effect of each performance evaluation index as the number of graph convolution layers increases;
[0062] Figure 16 This is a problem that exists in the performance evaluation of existing non-structural parameter models. DETAILED DESCRIPTION
[0063] To better understand the technical solutions of the present invention, the specific embodiments of the present invention are further described in detail below in conjunction with the accompanying drawings and embodiments. The same reference numerals in the accompanying drawings represent elements with the same or similar functions. Although various aspects of the embodiments are shown in the accompanying drawings, the drawings are not necessarily drawn to scale unless otherwise specified.
[0064] The present invention proposes a method for predicting compound-protein affinity based on protein three-dimensional structure, which is used to predict compound-protein affinity. Compound features are extracted from compounds, including atomic features and aggregate node features. Protein features that implicitly contain protein three-dimensional information are extracted from proteins, including discretized distance matrix and torsion angle matrix. Compound features and protein features are then combined through affinity prediction algorithms to predict compound-protein affinity. In order to implement this method, an end-to-end model based on deep learning and the latest technology is adopted. The model is named Rapidly Evolving Attention and Deep Graph Neural Network (FeatNN). A method for predicting compound-protein affinity based on protein three-dimensional structure implemented by FeatNN is as follows. Figure 1 The specific implementation steps are as follows:
[0065] S1, compound feature extraction step
[0066] The compound feature extraction step is implemented using the compound extractor.
[0067] Compound features include atomic features and bond features. To better characterize compound features, aggregate node features are also introduced. Compound features and aggregate node features are collectively referred to as compound composite features.
[0068] According to the compound characterization graph G = {atom, bond}, the initial atomic features and aggregate node features are obtained. The deep graph convolutional network and multi-head attention algorithm are used to update the atomic features and aggregate node features through the attention mechanism. The final atomic features and aggregate node features are output through a loop. At this time, the atomic features already include the bond features, so the output atomic features can represent the compound features. The specific steps are as follows: Figure 2 shown.
[0069] The invention of this step lies in the special residual connection loop of the deep graph convolutional network, which is used to aggregate the neighbor information of atoms onto the atoms and avoid the over-smoothing problem caused by the deepening of the network layers. GRU is used to aggregate the information between the front and back layers of the network, and a multi-head attention representation algorithm is used to obtain more diverse compound features. Aggregate node features are introduced, and an interactive evolution mechanism is used to update the aggregate node features and atomic features.
[0070] In order to clearly describe the steps of compound composite feature extraction, the following relevant definitions are given first. According to the molecular formula of the compound, it can be converted into a compound representation graph G = {atom, bond}, where atom represents atoms and bond represents bonds. Atom contains atomic features such as element name, aromatic type, vertex degree and atomic valence, and bond contains bond type and shape features. The values of atom and bond can be feature encoded separately through separate hot encoding strategies. According to the compound representation graph G, the features of each atom and bond in the compound can be obtained as follows: For each atom For each key where i = 1, 2, ..., N a ,j=1,2,...,N b , the primitive atomic features are defined as The sum of atomic features is called aggregate node feature, i.e., Considering the graph convolution layer l c and multi-head attention features (k c is the number of multiple heads), atom features or main node features, l c th is defined as and The variable V has k c The head is defined as where l c =1,2,...,l comp ,k c =1, 2, ..., k comp .
[0071] The compound characteristics can be obtained based on the compound graph representation G = {V, E}. More specifically, each node (i.e., atom) v i Initially, a feature vector of length 82 It is represented by the connection of word embeddings representing atomic symbols, representing the atomic symbols, degree, explicit valence, implicit valence and aromaticity of the corresponding atoms. Each edge (i.e. chemical bond) e i E is initially composed of a eigenvector of length 6 It represents the concatenation of word embeddings representing the bond type, including mono, di, tri, aromatic, the bond is conjugated, and the bond is in a ring.
[0072] First, the atomic features of the compound are input into the linear layer (linear) and the deep graph convolutional network respectively. After the atomic features are processed by the linear layer, they are input into the atomic aggregation function, the deep graph convolutional network, the multi-head attention representation algorithm and the first gated recurrent function (GRU). The atomic aggregation function is used to obtain the initial aggregation node features. The aggregation node features and the output results of the deep graph convolutional network are respectively processed by the linear layer and added together, and the first weight of the compound features is obtained through the first activation function (Sigmoid). The aggregation node features and the output results of the multi-head attention representation algorithm are respectively processed by the linear layer and added together, and the compound features are obtained through the first activation function (Sigmoid). The second weight, the aggregate node feature is processed by the linear layer and added to the output result of the deep graph convolutional network. The weight values of the two are the first weight of the compound feature and 1 minus the first weight of the compound feature, respectively. The result of the addition is output to the first gated recurrent function (GRU). The first gated recurrent function outputs the atomic feature, and the atomic feature is updated at this time; the aggregate node feature is processed by the linear layer and added to the output result of the multi-head attention representation algorithm. The weight values of the two are 1 minus the second weight of the compound feature and the second weight of the compound feature, respectively. The result of the addition is output to the second gated recurrent function (GRU). The second gated recurrent function outputs the aggregate node feature, and the aggregate node feature is updated at this time;
[0073] The updated aggregate node features and atomic features are used as input again for looping. The aggregate node features and atomic features obtained after looping K times are the aggregate node features and atomic features that need to be output. K is a positive integer greater than 2, and the specific value is set according to needs.
[0074] Graph convolutional networks (GCNs) can be used in biological research to predict drug-target affinity, but due to depth limitations, this approach often suffers from oversmoothing. GCN performance deteriorates as the number of layers increases, as the representations of nodes in the GCN converge to similar values. Deeper graph convolutional networks employ specialized residual connection loops to aggregate information about the neighbors of atoms onto the atoms themselves, thus avoiding the oversmoothing problem associated with deeper layers. Deeper GCN layers can extract more information.
[0075] The purpose of the deep graph convolutional network in this invention is to aggregate the neighbor information of atoms onto atoms, using the following method: Figure 3 As shown in the figure, for the carbon atom C located in the circle, the features of the surrounding nitrogen atoms N, oxygen atoms O and bromine atoms Br, as well as the bond features between them, will be aggregated into the features of the carbon atom through the deep graph convolutional network. The specific implementation steps are as follows: Figure 4As shown in , the atomic node features and the adjacency matrix aggregate the neighbor node features of each atom through the message passing mechanism and update them on each atom, and are supplemented by a special residual connection loop to prevent the over-smoothing problem encountered when deepening the network layer. Among them, α and θ are the hyper parameters of the residual connection loop. The atomic feature extraction cycle is as follows Figure 5 As shown in the figure, M represents the multi-head attention aggregation operation, and R represents the number of atomic diversity feature loop construction cycles.
[0076] The multi-head attention representation algorithm is used to implement the multi-head attention mechanism, which is specifically implemented as an existing technology. In the present invention, it is used to increase the diversity of the atomic node features of the compound to more comprehensively aggregate the compound information, learn more accurate compound representations, and improve the accuracy of affinity.
[0077] Therefore, at this time, the atomic features of the compound can represent the compound characteristics after repeated processing of deep graph convolutional network and multi-head attention.
[0078] The function of a linear layer is to map an input vector to a specified dimension for subsequent vector or matrix operations. All linear layers mentioned in this application are different linear layers and are set according to actual needs.
[0079] The compound feature extraction step is implemented in a computer using the following algorithm 1. The number of layers of the deep graph convolutional network is set as needed, and the graph convolution is implemented in a computer using the following algorithm 2.
[0080] S2, protein feature extraction step;
[0081] According to the protein sequence features and protein structure features, the embedded sequence features and embedded structure features are obtained through the protein feature aggregation algorithm, and then the updated protein sequence features and protein structure features are obtained through the protein encoding algorithm using the co-evolution strategy, such as Figure 6 The protein feature extraction step is implemented in a computer using the following algorithm 3.
[0082] The inventive point of this step is to utilize the structural features of proteins and to embed the protein structural features into the protein sequence features through the protein feature aggregation algorithm and protein encoding algorithm proposed by the present invention.
[0083] S21, obtaining the sequence and structural characteristics of proteins;
[0084] Protein sequence information can be directly obtained from relevant databases.
[0085] The structural feature is a three-dimensional structural feature, which includes at least a discretized distance matrix and may also include a structural auxiliary feature. In this embodiment, the structural feature includes a discretized distance matrix and a torsion angle matrix as a structural auxiliary feature.
[0086] The discretized distance matrix of a protein is the distance between the residues after the protein is unfolded into a peptide chain, and the discretized distance matrix is a symmetric matrix. Because the distance between residues is usually the Euclidean distance, the specific numerical value is a value within a continuous interval. Since a single numerical value cannot represent the structural characteristics well, it is not easy to understand the machine learning model. Therefore, a discrete distance matrix is proposed to map the distance matrix to a low-dimensional bin with equal width, that is, the continuous interval is divided into multiple equal interval segments, and the numerical values in the same interval segment are set to the same value. This greatly reduces the amount of data in the distance matrix and can better reflect the distance characteristics between residues. Compared with the traditional distance matrix, the distance value change interval is smaller and there is no obvious distinction. It is not convenient for feature extraction when the data is limited. The present invention divides these continuous values with little distinction into equidistant mapping intervals that are statistically basically in line with the normal distribution, thereby realizing the discretized encoding of the distance matrix. In this embodiment, the distance matrix is mapped to 38 low-dimensional bins of equal width between 3.25A and 50.75A (A is the unit of length angstrom, 1 meter = 10 to the power of 10 angstroms), with an additional bin to store any larger distances and a bin to store any smaller distances, for a total of 40 bins.
[0087] The torsion matrix is constructed using existing technology. Considering that the distance information between each residue alone may not accurately describe the three-dimensional structure of a protein, the torsion matrix is used as an auxiliary tool to accurately represent the orientation of the protein sequence. The torsion matrix is constructed using the sine and cosine values of the dihedral angles φ and ψ between the α-carbon atoms of the protein backbone and the amino and carboxyl groups.
[0088] S22, based on protein sequence features and three-dimensional structure features, the embedded sequence features and embedded distance matrix are obtained through protein feature aggregation algorithm, such as Figure 7 As shown, the specific steps are:
[0089] The protein sequence information is embedded in the first embedding layer (Eembedding layer) to obtain sequence features.
[0090] The discrete distance matrix is embedded in words through the second embedding layer to obtain discrete distance matrix features, and the output is a sequence feature with discrete distance matrix information.
[0091] After the twist angle matrix is extracted through the convolution layer, the twist angle feature is obtained; the twist angle feature is output through the linear layer and the first activation function (Sigmoid) is used to obtain the twist angle weight; the twist angle feature and sequence feature are added together through the gated logic with the twist angle weight as the gate weight, and the output is the embedded updated sequence feature.
[0092] The protein feature aggregation algorithm is implemented in a computer using the following Algorithm 5.
[0093] S23, using a co-evolution strategy to obtain updated sequence features and updated structural features for the embedded sequence features and embedded distance matrix;
[0094] The co-evolution strategy is implemented using an N-layer protein coding algorithm. The protein coding algorithm of each layer is the same. The specific steps of the protein coding algorithm of one layer are as follows: Figure 8 shown.
[0095] The two results of the embedded distance matrix after row addition and column addition are respectively fused through the gated recurrent unit to obtain the spliced embedded structural information; the spliced embedded structural information then enters the diversity convolution layer to obtain diversified protein structural features, so as to learn protein features with more diverse features; at the same time, the spliced embedded structural information passes through the linear layer and uses the second activation function (Sigmoid) to obtain the structural information weight.
[0096] The embedded sequence features pass through the diversity convolution layer and the ordinary convolution layer in sequence to obtain diversified protein sequence features; the diversified protein sequence features pass through the linear layer and use the third activation function (Sigmoid) to obtain the sequence information weight.
[0097] The diverse protein structural features and diverse protein sequence features are summed up through the gating unit with the sequence information weight as the gating weight to obtain a structural feature vector that preliminarily aggregates the sequence information and structural information; when adding, the weight of the diverse protein sequence features is the sequence information weight, and the weight of the diverse protein structural features is 1 minus the embedded structural information weight.
[0098] The structural feature vector that initially aggregates sequence information and structural information is updated with diverse protein structural features from the sequence dimension through a gated recurrent unit, and the dimensional mapping is changed through a convolutional layer to obtain a structural feature vector that has both structural information and sequence information. After the structural feature vector specifies the mapping dimension through the convolutional layer, it is itself combined and transformed (outer sum) into a protein structural feature. This embodiment uses the outer sum function to implement this, that is, each time an element of the first input is multiplied by the second input as a row, until all elements of the first input are traversed.
[0099] The updated sequence features after embedding are passed through the diversity convolution layer to obtain the updated sequence features after diversified embedding. The structural information weight is used as the gating weight. The updated sequence features after diversified embedding and the diversified protein structural features are added together through the gating logic to obtain the sequence feature vector that integrates the structural information and sequence information. The updated sequence features after diversified embedding are further fused with the sequence feature vector through the gated recurrent unit to output the protein sequence features. When adding, the weight used for the diversified protein structural features is the sequence information weight, and the weight of the updated sequence features after diversified embedding is 1 minus the sequence information weight.
[0100] In this embodiment, the common convolution layer performs traditional one-dimensional convolution operations; the diversity convolution layer can perform convolution operations in parallel and enhance the diversity of the learned features: the diversity convolution layer first divides the input feature vector into four equal parts in the feature dimension using an equal division strategy, and enters four parallel common convolution layers respectively, and adds the output results. At the same time, the initial features of the sequence information input are aggregated through residual connections, and finally a diversified feature vector is obtained, such as Figure 9 As shown; the equal division strategy is to directly divide the vector into four equal parts from the feature dimension, which needs to satisfy the premise that the feature number can be divided by the number of divisions (in this embodiment, the number of divisions is four). The following algorithm 9 is used in the computer to implement the above-mentioned equal division strategy.
[0101] The diversity convolution layer is implemented in the computer using the following algorithm 7.
[0102] At this point, the distance matrix feature output by the protein encoding algorithm is the distance matrix of the required protein, such as Figure 10 As shown in Figure 2, the distance matrix and sequence features are used as input to the next layer of protein encoding algorithms to continue updating the protein structure and sequence features until all N layers of protein encoding algorithms are completed. At this point, the protein sequence features already imply the protein structure. Therefore, the protein sequence features at this point can represent the protein features.
[0103] Protein encoding is implemented in computer using the following algorithm 4.
[0104] The following reference Figure 11 Explain "S3, Compound and protein affinity prediction steps";
[0105] In the compound and protein affinity prediction step, an affinity learning unit algorithm is introduced to learn the potential features of the interaction between proteins and compounds, and the features are aggregated separately to finally predict the affinity;
[0106] According to step S1, the atomic features and aggregate node features of the compound are received as compound information and input into the affinity learning unit. The aggregate node features are updated through the linear layer to obtain updated aggregate node features. The compound features are updated through the linear layer, and the mean is calculated in the atomic dimension. The features are then concatenated with the updated aggregate node features to obtain the comprehensive compound features.
[0107] According to step S2, the sequence features of the protein are received and passed into the affinity learning unit as protein information. The protein sequence features are updated through the convolution layer and the average is calculated in the residue dimension to obtain the comprehensive protein features.
[0108] The protein comprehensive features and compound comprehensive features are matrix multiplied (matmul) and the information of the two is fused through the linear layer to obtain the predicted affinity value. The basic process is as follows Figure 11 As shown, the following algorithm 8 is used in the computer to implement it.
[0109] The inventive point of this step is to use the protein sequence features that imply protein structural features and the atomic features that represent compound features to fuse them using the affinity learning unit algorithm proposed by the present invention, and combine them with the aggregated node features to finally predict the affinity value.
[0110] The compound and protein affinity prediction steps were implemented in silico using the following algorithm 8.
[0111] The present invention also proposes a compound-protein affinity prediction system based on protein three-dimensional structure, which is used to predict compound-protein affinity, such as Figure 1 As shown, it specifically includes the following modules:
[0112] a first input module, a second input module, a compound extractor, a protein extractor, and an affinity predictor;
[0113] The first input module receives the compound characterization graph and obtains the compound atomic features and aggregation node features according to the compound characterization graph;
[0114] A compound extractor is configured to receive compound atomic features and aggregate node features, obtain updated compound features and aggregate node features based on the input, and send the updated compound features and aggregate node features to an affinity predictor;
[0115] The second input module receives protein sequence features and protein structure features, which include discretized distance matrix and torsion angle matrix, and sends the received three features to the protein extractor;
[0116] The protein extractor is used to receive protein sequence features and protein structure features, update the protein sequence features, and send them to the affinity predictor.
[0117] The affinity predictor is used to receive compound features, protein sequence features and aggregation node features, perform affinity prediction, and provide affinity prediction values for compounds and proteins.
[0118] Furthermore, the compound extractor includes a deep graph convolution unit and a multi-head attention representation unit. The deep graph convolution unit is used to implement a deep graph convolutional network, which aggregates information about the neighbors of atoms onto the atoms themselves. The multi-head attention representation unit is used to extract more diverse compound features to improve the accuracy of affinity prediction.
[0119] The protein extractor includes a protein information aggregation unit and a co-evolutionary update unit. The protein information aggregation unit is used to implement a protein feature aggregation algorithm so that the sequence features of proteins carry protein structural information. The co-evolutionary update unit is used to implement a co-evolutionary strategy so that the extracted protein sequence features can represent the global information of proteins.
[0120] The affinity predictor consists of affinity learning units.
[0121] The method of the present invention is compared with the prior art below, and the model implementing the present invention is named Fast Evolving Attention and Deep Graph Neural Network (FeatNN).
[0122] 1. Build a test dataset
[0123] In the study, two datasets based on PDBbind (version 2020, with a total of 23,496 entries) and BindingDB (version February 6, 2022, with a total of 41,296 entries) were constructed. The PDBbind database provides experimental binding affinity data for a total of 23,496 biomolecular complexes in the PDB. In this example, different measurement values of K values of protein-ligand affinity data (12,628 data) after data cleaning were mainly used. i ,K d and IC 50 , according to the training set: validation set: test set = 7:1:2 as the small data set training sample. BindingDB contains 241268 binding data of 8661 proteins and 1039940 small molecules. In this example, the IC of protein-ligand affinity data (218615 data) after data cleaning is mainly used. 50 The measurement type is based on the ratio of training set: validation set: test set = 7:1:2 as the large data set training samples.
[0124] Check the normality of the correlation data, as extreme data may affect the entire model. The BindingDB dataset has some extremely extreme correlation data, so the outer data (μ-3σ, μ+3σ) are eliminated.
[0125] For the distance matrix and torsion matrix, protein structure data were obtained by searching the RCSB PDB (Protein Data Bank) database using the PDB IDs provided by the two databases. The RSCB PDB files provide sufficient 3D structural data (spatial positions) to extract the distance matrix needed for the model, and the torsion matrix was calculated based on this 3D structural information.
[0126] For SMILES sequences, we obtained data from the RCSB PDB based on the ligand ID.
[0127] 2. Determine the evaluation indicators,
[0128] The following indicators are used to evaluate the prediction performance of the model: R 2 , RMSE (root mean square error), MAE (absolute mean error), MedAE, Pearson correlation coefficient, and Spearman rank correlation coefficient; Among them, R 2 , RMSE (root mean square error), and MAE (mean absolute error) describe the distance between the predicted value and the true value; the Pearson product-moment correlation coefficient and the Spearman rank correlation coefficient describe the correlation between the predicted value and the true value.
[0129] The detailed description of each indicator is as follows:
[0130] R 2 is a dimensionless score describing the effectiveness of the model. It compares the prediction to random guessing based on the mean of the true values;
[0131] RMSE is a standard value of MSE and is often used as a training loss in machine learning research;
[0132] MAE describes the absolute error between the predicted value and the true value;
[0133] MedAE is the median MAE, which describes the median value of the residue;
[0134] The Pearson correlation coefficient describes the linear correlation between two values;
[0135] The Spearman rank correlation coefficient is the rank form of the Pearson correlation coefficient, which describes the correlation between two variables (for example, when one variable increases, the other variable also increases), which is related to the monotonicity of the function;
[0136] 3. Evaluation of FeatNN, BACPI, GraphDTA (GATNet, GAT_GCN, GCNNet, GINConvNet) and MONN.
[0137] Figure 12 To test the clustering of compounds based on IC50 and KIKD, we Figure 12 It can be seen that the FeatNN of the present invention has good performance in three indicators (RMSE, Pearson and R 2 ) are superior to other models.
[0138] Figure 13 To test on new protein data, Figure 13 It can also be seen that the FeatNN of the present invention has good performance in three indicators (RMSE, Pearson and R 2 ) are also better than other models.
[0139] Figure 14 It is a comparison of the convergence of each model during the training process (model convergence speed and final convergence position); it can be seen from the figure that the convergence effect of the invented FeatNN is the best.
[0140] Figure 15 In order to show the advantages of deep graph convolution, the present invention also uses different depths for graph convolution, and it can be seen that the performance evaluation indicators change with the deepening of the number of graph convolution layers.
[0141] Table 1 shows the performance evaluation table of the MONN, GraphDTA (GATGCN, GCNNet, GINConvNet, GATConvNet), BACPI and FeatNN models of the present invention trained on the BindingDB (IC50 measurement standard, a total of 218,615 data);
[0142] Table 1
[0143]
[0144] Tables 2 and 3 show the model performance evaluation of the FeatNN of the present invention and the various models of the prior art after homology similarity clustering of compounds (Table 2) and proteins (Table 3) under different evaluation indicators (KIKD or IC50) trained on the PDBBind small data set (including KIKD and IC50 measurement standards, a total of 12,628 data).
[0145] Table 2
[0146]
[0147] Table 3
[0148]
[0149] In order to further demonstrate that the present invention is feasible, the pseudo code for implementing the main modules of the present invention is provided below, hoping to better understand the method proposed in the present invention.
[0150] First, the main variable definitions in the code are shown in Table 4
[0151] Table 4 Definition of common variables
[0152]
[0153]
[0154] The superscript of the symbol definition is the name, and the subscript is the vector dimension information.
[0155] In addition, N res N represents the number of residues in the input main sequence (pruned during training). tors Indicates the torsion size, N ebd_s represents the embedding size, h represents the hidden size, N atom Indicates the number of atoms, N bond Indicates the number of keys, F atom Indicates atomic characteristics, F bond Denotes the key feature and nbs is the neighbor dimension. Use uppercase operator names when they encapsulate learning parameters, for example, use Linear to represent a linear transformation with a weight matrix W and a bias vector b.
[0156] LayerNorm is used to represent layer normalization that operates on the channel dimension, and the gain and bias of each channel can be learned. We also use uppercase names for random operators, such as those related to dropout. For functions without parameters, we use lowercase operator names. The definitions of operations and variables are shown below. represents multiplication, represents addition, ⊙ represents Hadamard product, a T b represents the dot product of two vectors, [·,·] h Refers to the dimension concatenation operation on the hidden size. Use h to represent the hidden size, k to represent the number of attention heads, and k h represents the size of the attention head, where k h =h / k. softmax(x i ) corresponds to the soft max function, softmax(x i )=exp(x i ) / ∑ iexp(x i ), represents the softmax activation function; σ(·) represents the sigmoid activation function σ(x)=1 / (1+e -x ), tanh(·) corresponds to the tanh activation function tanh(x)=(e x -e -x ) / (e x +e -x ), represents the hyperbolic tangent activation function, gelu represents the Gaussian error linear unit, and f(·) corresponds to the GELU activation function where erf(·) is the Gaussian error function,
[0157] The algorithms of each module are as follows:
[0158] Algorithm 1 Graph Convolution
[0159]
[0160]
[0161]
[0162] con neighhor =concat neighhor (ver neighbor ,edge neighbor )
[0163] neighbor_label=gelu(Linear(con neighbor ))
[0164]
[0165]
[0166] output=theta⊙Linear(support)+(1-theta)⊙support
[0167]
[0168]
[0169] Algorithm 2 Compound Extractor
[0170]
[0171]
[0172] Algorithm 3 Protein feature extraction ProteinExtractor
[0173]
[0174] Algorithm 4 Protein Encoder
[0175]
[0176]
[0177]
[0178] Algorithm 5 Protein Aggregating
[0179]
[0180] seq_embed=Embedding(Seq init )
[0181]
[0182] torsion_embed=CNN Nts→h1 (TorMat init )
[0183] torsion_vector=CNN h1→h (torsion_embed)
[0184] gate=Sigmoid(Linear(torsion_vector))
[0185]
[0186]
[0187]
[0188] Algorithm 6 Single-layer protein encoding EvoUpdating
[0189]
[0190]
[0191]
[0192]
[0193] MixKey = GRU(PairKey1, PairKey2)
[0194] Struct_Features = DivCNN(MixKey)
[0195]
[0196]
[0197] SeqGate = Sigmoid(Linear(Seq2Struct))
[0198] StructGate = Sigmoid(Linear(MixKey))
[0199] Seq2Struct_Vector = SeqGate ⊙ Seq2Struct+(1 - SeqGate)⊙Struct_Features
[0200] Struct_Vector = GRU(Seq2Struct_Vector, Struct_Features)
[0201] Struct2Seq_Mapping = DeepSparseCNN(Struct_Vector)
[0202]
[0203] Seq_Vector = StructGate ⊙ Struect_Features+(1 - StructGate)⊙Seq_Features
[0204]
[0205]
[0206] Algorithm 7 Diversity Convolutional Layer DivCNN
[0207] def DivCNN(x):
[0208] x0 = CNN kernel=1 (x)
[0209] #x(seq, hidden_size)→x k-head (seq, k, head_size)
[0210] #where k is the number of head,and head_size=hidden_size / k
[0211] x k-head =TransposeForScores(x)
[0212]
[0213] return→{x total}
[0214] Algorithm 8 Affinity Prediction AffinityLearning
[0215]
[0216] #Inputs projections
[0217]
[0218]
[0219]
[0220]
[0221]
[0222]
[0223] #Output projection
[0224]
[0225] return→{Affinity Prediction}
[0226] Algorithm 9 TPS function TransposeForScores
[0227] def TransposeForScores({input}):
[0228] #input dimention:(N res / N atom ,h)
[0229] #output dimention:(k,N res / N atom ,ks)
[0230] output=DimentionReshape(input)
[0231] return{output}
[0232] Finally, it should be noted that the embodiments described above are only used to illustrate the technical solutions of the present invention, rather than to limit them. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some or all of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the scope of the technical solutions of the embodiments of the present invention.
Claims
1. A method for predicting compound-protein affinity based on protein three-dimensional structure, for predicting compound-protein affinity, extracting compound features from compounds, including atomic features and aggregate node features, extracting protein features that implicitly contain protein three-dimensional information from proteins, and then using an affinity prediction algorithm to predict compound-protein affinity. The method specifically includes the following steps: S1, compound feature extraction step, According to the compound characterization graph, the initial atomic features and aggregate node features are obtained. Using the deep graph convolutional network and multi-head attention mechanism, the atomic features and aggregate node features are cyclically updated to output the final atomic features and aggregate node features of the compound. S2, protein feature extraction step; Obtain protein sequence information and protein structural features; Based on the protein sequence information and protein structural features, a protein feature aggregation algorithm is used to make the protein embedded sequence features carry protein structural features, thereby obtaining protein embedded sequence features and embedded structural features; By adopting a co-evolution strategy, the embedded sequence features and embedded structural features are cyclically updated, and finally the updated protein sequence features and protein structural features are obtained; S3 compound and protein affinity prediction step; Based on the atomic features and aggregate node features of the compound obtained in step S1 and the protein sequence features in step S2, the affinity learning unit algorithm is used to obtain a predicted affinity value. The larger the affinity value, the greater the probability of binding between the compound and the protein.
2. The method for predicting compound-protein affinity based on protein three-dimensional structure according to claim 1, characterized in that: The step S1 is specifically as follows: S11, obtain atomic features and aggregation node features based on the compound characterization graph; S12, the atomic features are input into the deep graph convolutional network, the output of the deep graph convolutional network and the aggregate node features are added according to the first weight of the compound features, and the updated atomic features are obtained by combining the first gated recurrent function GRU and the atomic features; S13, the atomic features are input into the multi-head attention representation algorithm, the output of the multi-head attention representation algorithm and the aggregate node features are added according to the second weight of the compound features, and the updated aggregate node features are obtained by combining the second gated recurrent function GRU and the aggregate node features; S14: After repeating steps S12-S13 K times using the updated atomic features and the updated aggregate node features as the atomic features and aggregate node features, the atomic features and the aggregate node features are output.
3. The method for predicting compound-protein affinity based on protein three-dimensional structure according to claim 2, characterized in that: The deep graph convolutional network in step S1 has a residual connection loop for aggregating the neighbor information of atoms onto the atoms and avoiding the over-smoothing problem caused by deepening the number of network layers. GRU is used to aggregate the information between the front and back layers of the network, and a multi-head attention mechanism characterization algorithm is used to obtain diverse compound features, thereby improving the accuracy of affinity prediction in the affinity prediction step.
4. The method for predicting compound-protein affinity based on protein three-dimensional structure according to claim 1, characterized in that: The step S2 is specifically as follows: S21, obtaining protein sequence information and protein structural features, where the protein structural features at least include a discretized distance matrix; S22, the protein feature aggregation algorithm is specifically as follows: the protein sequence information is passed through the first embedding layer to obtain a sequence vector, the discretized distance matrix is passed through the second embedding layer to obtain a discrete distance matrix feature vector; the discrete distance matrix feature vector is used as the feature embedding into the updated embedded distance matrix; S23, the co-evolution strategy is implemented using an N-layer protein encoding algorithm. The protein encoding algorithm of each layer is the same. The specific protein encoding algorithm of one layer is: The two results of the embedded distance matrix after row addition and column addition are fused through the gated recurrent unit to obtain the spliced embedded structural information. The spliced embedded structural information then enters the diversity convolution layer to obtain diverse protein structural features. The embedded sequence features are sequentially passed through the diversity convolution layer and the common convolution layer to obtain diverse protein sequence features. The diverse protein structure features and diverse protein sequence features are added together through the gating logic, and then sequentially passed through the gated recurrent unit and the convolution layer to obtain a structural feature vector that contains both structural and sequence information. The structural feature vector undergoes its own binding transformation to output the protein structural features; The updated sequence features are embedded into a diversity convolutional layer to obtain the updated sequence features; The updated sequence features after diversified embedding are summed with the diversified protein structure features through gated logic to obtain a sequence feature vector that integrates the structure information and sequence information. The diversified protein sequence features are then further fused with the sequence feature vector through a gated recurrent unit to output the protein sequence features. S24, the new embedding distance matrix output by the protein coding algorithm is the required embedding distance matrix. The obtained embedding distance matrix and embedded sequence features are used as the input of the next layer of protein coding algorithm to continue updating the embedding distance matrix and embedded sequence features until all N layers of protein coding algorithms are completed. At this time, the output embedding distance matrix is the structural feature of the protein, and the embedded sequence feature is the sequence feature of the protein.
5. The method for predicting compound-protein affinity based on protein three-dimensional structure according to claim 4, characterized in that: The protein structure features in step S21 also include feature auxiliary matrix; The embedded sequence features obtained in step S22 are further added to the feature auxiliary matrix through gated logic to obtain the embedded updated sequence features; The diversity convolution layer in step S23 divides the input feature vector into four equal parts in the feature dimension, and enters four parallel ordinary convolution layers respectively, and adds the output results. At the same time, the initial features of the sequence information input are aggregated through residual connections to obtain a diversified feature vector.
6. The method for predicting compound-protein affinity based on protein three-dimensional structure according to claim 5, characterized in that: The characteristic auxiliary matrix is a torsion angle matrix, which is constructed using the sine and cosine values of the dihedral angles φ and ψ between the α-carbon atom of the protein backbone chain and the amino and carboxyl groups.
7. The method for predicting compound-protein affinity based on protein three-dimensional structure according to claim 4, characterized in that: The discretized distance matrix in step S21 divides the distances between the residues of the protein stable structure into M equally spaced mapping intervals that are statistically substantially in line with the normal distribution, thereby realizing the discretized encoding of the distance matrix, where M is a positive integer.
8. The method for predicting compound-protein affinity based on protein three-dimensional structure according to claim 1, characterized in that: The step S3 is specifically as follows: S31, receiving the atomic features and aggregate node features of the compound from step S1 as compound information and passing them into the affinity learning unit, and updating the aggregate node features through the linear layer to obtain updated aggregate node features; The compound features are updated through the linear layer, and the mean is calculated in the atomic dimension. The features are then concatenated with the updated aggregate node features to obtain the comprehensive compound features. S32, receiving the sequence features of the protein from step S2 and passing the sequence features of the protein as protein information into the affinity learning unit, the protein sequence features are updated through the convolution layer, and the average is calculated in the residue dimension to obtain the comprehensive protein features; The comprehensive features of proteins and compounds are matrix multiplied and their information is fused through a linear layer to obtain the predicted affinity value.
9. A system for predicting compound-protein affinity based on protein three-dimensional structure, characterized in that include: Compound extractors, protein extractors, and affinity predictors; The compound extractor uses a deep graph convolutional network and a multi-head attention algorithm to update the atomic features and aggregate node features of the input compound through the attention mechanism, and outputs the final atomic features and aggregate node features through a loop; the final compound features and aggregate node features are sent to the affinity predictor; The protein extractor obtains embedded sequence features and embedded structural features based on the input protein sequence features and protein structural features through a protein feature aggregation algorithm; and updates the embedded sequence features and embedded structural features by adopting a co-evolution strategy to obtain updated protein sequence features and protein structural features; The affinity predictor is used to receive compound features, protein sequence features and aggregation node features, and obtain the predicted affinity value through the affinity learning unit algorithm.
10. A system for predicting compound-protein affinity based on protein three-dimensional structure according to claim 9, characterized in that include: The compound extractor includes a deep graph convolution unit and a multi-head attention representation unit, wherein the deep graph convolution unit is used to implement a deep graph convolution network, and the multi-head attention representation unit is used to implement a multi-head attention algorithm; The protein extractor includes a protein information aggregation unit and a co-evolution updating unit, wherein the protein information aggregation unit is used to implement a protein feature aggregation algorithm, and the co-evolution updating unit is used to implement a co-evolution strategy; The affinity predictor includes an affinity learning unit.
Citation Information
Patent Citations
Drug molecule recommendation system for regulating and controlling disease targets based on a deep learning, computer equipment and storage medium
CN112562790A
Protein-protein interaction site prediction method based on deep map convolutional network
CN113192559A
