A protein-protein interaction modulator prediction method based on hierarchical graph modeling

Abstract

The application discloses a protein-protein interaction modulator (PPIM) prediction method based on hierarchical graph modeling, and belongs to the cross technical field of computers and drug screening.The application solves the problems of low accuracy, low efficiency and high complexity of traditional PPIM prediction methods.The application uses a graph attention network to learn the features of a given chemical structure at the atomic and functional group levels, overcomes the limitation that traditional machine learning models need to rely on manual screening and extraction of molecular features, and reduces the complexity of the prediction method.The application uses a hypergraph attention network to aggregate and convert two levels of molecular structure information, and overcomes the problem that existing methods cannot capture hierarchical information.The application can learn the effective representation of PPIM by capturing multi-scale substructure patterns, so as to improve the accuracy and efficiency of PPIM prediction and guarantee the effectiveness of PPIM prediction.The method can be applied to the prediction of protein-protein interaction modulators.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of computer science and drug screening, and specifically relates to a method for predicting protein interaction regulators. Background Technology

[0002] Protein-protein interactions (PPIs) play a central role in cellular processes and biochemical events, including signal transduction, cell adhesion, cell proliferation, gene expression, and immune regulation. Studies have shown that PPIs are involved in various disease pathways, such as multiple stages of cancer development, cardiovascular disease, neurodegenerative diseases, and infectious diseases. Therefore, PPIs are considered novel and highly attractive drug targets. Modulating or blocking protein-protein interactions using inhibitors is a promising therapeutic intervention strategy. Several small-molecule PPI inhibitors have already been identified and tested clinically, particularly in oncology.

[0003] The critical importance of PPIs makes them a highly promising therapeutic target, and their modulation is considered a novel and promising drug discovery strategy. Therefore, extensive research has been conducted over the past two decades to explore the design and development of PPI modulators (PPIMs). Currently, an increasing number of small molecule PPIMs are entering the market or clinical trials. For example, the BCL2 inhibitor venetoclax (ABT-199) has been approved by the FDA for the treatment of chronic lymphocytic leukemia; the LFA-1 / ICAM-1 inhibitor lifitegrast (SAR 1118) has been approved for the treatment of dry eye syndrome. Generally, small molecule PPIMs have different mechanisms of action and can be divided into two categories: inhibitory and stabilizing. Most known modulators belong to the inhibitor class.

[0004] Despite significant successes, the screening and optimization of small molecule PPIMs remains fraught with both opportunities and challenges. On the one hand, the general characteristics of PPI interfaces—large, flat, and featureless—make existing compound libraries used for screening conventional targets unsuitable for PPIM screening. On the other hand, with the rapid development of chemical biology and structural biology, specialized PPIM databases have emerged, including TIMBAL, ippI-DB, and 2P2I-DB. These provide the foundation and evidence for computational screening methods, especially those based on machine learning. Supported by these databases, several studies have been conducted to identify novel and effective small molecule PPIMs by analyzing key features between compounds and chemical properties. Early studies utilized chemical descriptors to construct general decision tree models to determine whether compounds are potential PPI inhibitors. Recent research focuses on modeling and analyzing different PPI families. Gupta et al. proposed a two-stage PPIM classification method based on Morgan fingerprints. In the first stage, potential PPI inhibitors were identified from a large compound library; in the second stage, specific predictors were built for 11 different PPI families. Recently, Gao et al. proposed an ensemble learning framework based on tree models, designed to screen effective modulators for different PPI families.

[0005] While existing methods have contributed to the discovery and identification of PPIMs, several limitations remain. These traditional machine learning-based approaches rely on the manual extraction and selection of molecular descriptors or fingerprint features, requiring significant expertise and human intervention. Furthermore, different PPI targets exhibit varying sensitivities and affinities for regulators due to their different roles in physiological processes. Therefore, specific features need to be designed for regulators targeting each PPI family, further increasing the complexity of developing predictive models.

[0006] In recent years, deep learning has made significant progress in various fields such as image recognition, speech recognition, natural language processing, and drug discovery. The ability of deep learning models to automatically extract molecular features allows them to overcome the limitations of conventional machine learning methods. Molecular compounds can be naturally represented as graph structures, where atoms correspond to nodes and chemical bonds to edges. Graph Neural Networks (GNNs) are networks specifically designed for processing graph data, which can better capture the similarities and differences between molecules by utilizing the structural information of molecular graphs. Representative GNNs, including Graph Convolutional Networks (GCNs), Graph Attention Networks (GATs), and Graph Isomorphism Networks (GINs), have achieved impressive performance in various tasks such as molecular property prediction and molecular generation. However, these GNNs, which operate on paired graphs (atomic-level graphs), often neglect hierarchical information of molecules and lack domain knowledge.

[0007] In summary, traditional machine learning-based methods rely on manual design of features for regulators targeting each PPI family, resulting in high complexity and low prediction efficiency in developing predictive models. Many deep learning methods fail to capture molecular-level information, leading to low accuracy in PPIM prediction. Therefore, it is essential to propose a new method for predicting protein-protein interaction regulators. Summary of the Invention

[0008] The purpose of this invention is to address the problems of low accuracy, low efficiency, and high complexity of traditional PPIM prediction methods, and to propose a protein interaction regulator prediction method based on hierarchical graph modeling.

[0009] The technical solution adopted by the present invention to solve the above-mentioned technical problems is as follows:

[0010] A method for predicting protein interaction regulators based on hierarchical graph modeling, the method specifically includes the following steps:

[0011] Step 1: Construct a PPI regulator identification dataset and a PPI regulator efficacy prediction dataset for each PPI family. Divide the constructed PPI regulator identification dataset into three parts: training set, validation set, and test set. Then divide the constructed PPI regulator efficacy prediction dataset into three parts: training set, validation set, and test set.

[0012] Step 2: Generate atomic-level and functional group-level maps for each compound in the PPI regulator identification dataset targeting each PPI family; then generate atomic-level and functional group-level maps for each PPI regulator in the PPI regulator efficacy prediction dataset targeting each PPI family.

[0013] Step 3: Construct hierarchical hypergraph attention classification models with identical structures for each PPI family. Then, use the PPI modifiers of each PPI family to identify the atomic-level and functional group-level graphs of the training set to train the hierarchical hypergraph attention classification models. Finally, use the validation set and test set to validate and test the trained hierarchical hypergraph attention classification models.

[0014] Step 4: Construct hierarchical hypergraph attention regression models with identical structures for each PPI family. Then, use the atomic-level and functional group-level graphs of the PPI modulator efficacy prediction training set for each PPI family to train the hierarchical hypergraph attention regression models. Finally, use the validation set and test set to validate and test the trained hierarchical hypergraph attention regression models.

[0015] Step 5: For each compound to be identified, input it into a pre-trained hierarchical hypergraph attention classification model to determine whether it is a regulator of a certain PPI family. If the compound is a regulator of a certain PPI family and its regulatory efficacy needs to be predicted, then input the compound into the hierarchical hypergraph attention regression model of the corresponding PPI family to obtain the regulatory efficacy prediction result. If the compound is a regulator of a certain PPI family and regulatory efficacy does not need to be predicted, then the process ends. If the compound is not a regulator of a certain PPI family, then the process ends.

[0016] If a compound is known to be a PPI regulator of a certain PPI family and the regulatory efficacy of the PPI regulator needs to be predicted, the PPI regulator can be directly input into the hierarchical hypergraph attention regression model of the corresponding PPI family to obtain the regulatory efficacy prediction result.

[0017] The beneficial effects of this invention are:

[0018] This invention utilizes graph attention networks to learn the features of a given chemical structure at the atomic and functional group levels, overcoming the limitations of traditional machine learning models that rely on manual screening and extraction of molecular features, thus reducing the complexity of prediction methods. Furthermore, it leverages hypergraph attention networks to aggregate and transform molecular structural information at two levels, overcoming the problem of existing methods failing to capture hierarchical information. By capturing multi-scale substructure patterns, this invention can learn an effective representation of PPIM, thereby improving the accuracy and efficiency of PPIM prediction and ensuring its effectiveness. Attached Figure Description

[0019] Figure 1 This is a flowchart of a protein interaction regulator prediction method based on hierarchical graph modeling according to the present invention;

[0020] Figure 2 This is a schematic diagram of the structure of a hierarchical hypergraph attention network. Detailed Implementation

[0021] Specific Implementation Method 1: Combination Figure 1 This embodiment describes a method for predicting protein interaction regulators based on hierarchical graph modeling. The method specifically includes the following steps:

[0022] Step 1: Construct a PPI regulator identification (classification) dataset and a PPI regulator efficacy prediction (regression) dataset for each PPI family. Divide the constructed PPI regulator identification dataset into three parts: training set, validation set, and test set. Then divide the constructed PPI regulator efficacy prediction dataset into three parts: training set, validation set, and test set.

[0023] That is, for each PPI family, we can obtain its PPI regulator identification training set, PPI regulator identification validation set, and PPI regulator identification test set, as well as its PPI regulator efficacy prediction training set, PPI regulator efficacy prediction validation set, and PPI regulator efficacy prediction test set; the ratio of the training set, validation set, and test set is 8:1:1.

[0024] Step 2: Generate atomic-level and functional group-level maps for each compound in the PPI regulator identification dataset targeting each PPI family; then generate atomic-level and functional group-level maps for each PPI regulator in the PPI regulator efficacy prediction dataset targeting each PPI family.

[0025] Step 3: Construct hierarchical hypergraph attention classification models with identical structures for each PPI family. Then, use the PPI modifiers of each PPI family to identify the atomic-level and functional group-level graphs of the training set to train the hierarchical hypergraph attention classification models. Finally, use the validation set and test set to validate and test the trained hierarchical hypergraph attention classification models.

[0026] Step 4: Construct hierarchical hypergraph attention regression models with identical structures for each PPI family. Then, use the atomic-level and functional group-level graphs of the PPI modulator efficacy prediction training set for each PPI family to train the hierarchical hypergraph attention regression models. Finally, use the validation set and test set to validate and test the trained hierarchical hypergraph attention regression models.

[0027] Step 5: For the compound to be identified, input the compound into each of the trained hierarchical hypergraph attention classification models to determine whether the compound belongs to a certain PPI family. If the compound belongs to a certain PPI family and its regulatory efficacy needs to be predicted (here, "its" refers to the compound to be identified), then input the compound into the hierarchical hypergraph attention regression model of the corresponding PPI family to obtain the regulatory efficacy prediction result. If the compound belongs to a certain PPI family and regulatory efficacy does not need to be predicted, then the process ends. If the compound does not belong to a certain PPI family, then the process ends.

[0028] If a compound is known to be a PPI regulator of a certain PPI family and the regulatory efficacy of the PPI regulator needs to be predicted, the PPI regulator can be directly input into the hierarchical hypergraph attention regression model of the corresponding PPI family to obtain the regulatory efficacy prediction result.

[0029] Specific Implementation Method Two: This implementation method differs from Specific Implementation Method One in that, in step one, a PPI modifier identification dataset is constructed for each PPI family, and the constructed dataset is divided into three parts: a training set, a validation set, and a test set; the specific process is as follows:

[0030] Perform steps 1 through 4 separately for each PPI family:

[0031] Step 1: Collect experimentally characterized small molecule modulators targeting the current PPI family from the iPPI-DB and 2P2I-DB v2 databases;

[0032] Step 2: Use Morgan fingerprint and Butina clustering algorithms to cluster the regulators collected in Step 1. After obtaining the clustering results, select one regulator from each category and use the selected regulators to form positive samples in the PPI regulator identification dataset that target the current PPI family.

[0033] Step 3: Obtain a number of negative samples equal to the number of positive samples, and use the positive and negative samples together to form a PPI modifier identification dataset targeting the current PPI family.

[0034] Of these, half of the negative samples came from positive samples targeting other PPI families, while the other half of the negative samples were compounds randomly selected from the ChEMBL database.

[0035] Step 4: Divide the PPI regulator identification dataset targeting the current PPI family into three parts: training set, validation set, and test set.

[0036] The other steps and parameters are the same as in Specific Implementation Method 1.

[0037] In steps one and two, the radius of the Morgan fingerprint is 2, and the Tanimoto coefficient similarity cutoff value of the Butina clustering algorithm is 0.8. Clustering is performed, and a modifier is selected in each cluster to remove duplicate and redundant compounds. In step one and three, each negative sample selected from the ChEMBL database meets the following criteria: targeting a single protein, and the compound has a reasonable half-maximum inhibitory concentration (IC50) value.

[0038] Specific Implementation Method Three: This implementation method differs from Specific Implementation Method One or Two in that, in step one, a PPI modulator efficacy prediction dataset is constructed for each PPI family, and the constructed PPI modulator efficacy prediction dataset is divided into three parts: a training set, a validation set, and a test set; the specific process is as follows:

[0039] Perform steps (1) through (3) for each PPI family separately:

[0040] Step (1): Collect experimentally characterized small molecule modulators targeting the current PPI family from the iPPI-DB database;

[0041] Step (2): The regulators collected in step (1) are clustered using Morgan fingerprint and Butina clustering algorithms. After obtaining the clustering results, one regulator is selected from each category and a label is added to each selected regulator. The added label is the activity value of the corresponding regulator (pIC50, pKi, etc.).

[0042] A dataset predicting the effectiveness of PPI modifiers targeting the current PPI family was constructed using the selected modifiers and added labels.

[0043] Step (3): Divide the PPI regulator efficacy prediction dataset targeting the current PPI family into three parts: training set, validation set, and test set.

[0044] Other steps and parameters are the same as in specific implementation method one or two.

[0045] Specific Implementation Method Four: This implementation method differs from Specific Implementation Methods One to Three in that, in step two, atomic-level and functional group-level maps are generated for each compound in the PPI modifier identification dataset targeting each PPI family; the specific process is as follows:

[0046] Define the atomic level diagram as G atom =(V atom E atom ), where V atom It is a set of nodes in an atomic-level diagram, where each atom in the current compound is treated as a node, E atom It is the set of edges in the atomic-level graph, where each bond in the current compound is treated as an edge;

[0047] Define the functional group level diagram as G fg =(V fg E fg ), where V fg It is a set of nodes in the functional group hierarchy graph, where each functional group in the current compound is treated as a node, E fg It is the set of edges of the functional group level graph. If two functional groups have adjacent atoms in the atomic level graph, then there is an edge between the two functional groups.

[0048] The other steps and parameters are the same as those in one of the specific implementation methods one to three.

[0049] Specific Implementation Method Five: Combination Figure 2This embodiment is described below. The difference between this embodiment and one of the specific embodiments one through four is that the training of the hierarchical hypergraph attention classification model using the PPI modifier identification training set of each PPI family is as follows:

[0050] Step S1: For any PPI family, generate a hypergraph for each compound in the PPI regulator identification training set of that PPI family (since the PPI regulator identification training set includes positive and negative samples, each positive and negative sample is referred to as a compound here), and define the hypergraph as G. h =(V h E h ), where V h It is the set of nodes in the hypergraph, where each atom in the current compound is treated as a node, E h It is the set of hyperedges in the hypergraph, where each functional group in the current compound is treated as a separate hyperedge;

[0051] Using the association matrix To model the node relationships in the hypergraph, if the i-th atom is in the j-th functional group, then H i,j =1, otherwise H i,j =0, H i,j It is the element in the i-th row and j-th column of the correlation matrix;

[0052] Step S2: For any compound in the PPI regulator identification training set of this PPI family, use the atomic-level map and functional group-level map of the compound as inputs to two GATs (Graph Attention Networks) in the first-level graph attention unit, respectively, to obtain the atomic features X′ output by the two GATs. atom and functional group characteristics X′ fg ;

[0053] Then, the elements in the correlation matrix are updated based on the atomic and functional group features output by the two GATs to obtain the updated correlation matrix.

[0054] Step S3: Use the atomic features and functional group features output by the two GATs as input to the hypergraph attention network in the first-level graph attention unit, and obtain the functional group-level features containing atomic information and the atomic-level features containing functional group information based on the updated correlation matrix.

[0055] Atomic-level features X″ containing functional group information atom Atomic features X′ output by GAT atom By combining residual joins, the join result is obtained. Functional group-level features X″ containing atomic information fg and the functional group feature X′ output by GAT fgBy combining residual joins, the join result is obtained.

[0056] Step S4, and As input to the second-level graph attention unit, the working process within the second-level graph attention unit is the same as the working process within the first-level graph attention unit;

[0057] This process continues until the output of the Lth level graph attention unit is obtained. and Then Obtained through GAT Will Obtained through GAT

[0058] Using summation pooling, average pooling, and max pooling operators Perform global feature extraction and concatenate the feature extraction results to obtain the concatenation result. Using summation pooling, average pooling, and max pooling operators Perform global feature extraction and concatenate the feature extraction results to obtain the concatenation result.

[0059] Step S5: Connect the results Adjust to size C n ×C m Feature map (C) n ×C m The value is equal to The input consists of two parallel channels (of dimension n). One channel comprises n convolutional blocks (each convolutional block includes a convolutional layer, a BatchNorm layer, and a ReLU activation function layer), and the other channel comprises a convolutional layer with a kernel size of 1×1. The outputs of the two channels are residually concatenated, and then the residual concatenation result is passed through a ReLU activation function layer. The atomic-level global features output by the ReLU activation function layer are denoted as...

[0060] Connect results Adjust to size C n ×C m The feature map is input to two parallel channels. One channel consists of n convolutional blocks (each convolutional block includes a convolutional layer, a BatchNorm layer, and a ReLU activation function layer), and the other channel consists of a convolutional layer with a kernel size of 1×1. The outputs of the two channels are residually concatenated, and then the residual concatenation result is passed through a ReLU activation function layer. The functional group-level global features output by the ReLU activation function layer are denoted as...

[0061] Step S6, and As input to the multilayer perceptron, the compound identification result is output by the multilayer perceptron.

[0062]

[0063] Among them, MLP is a two-layer MLP with ReLU activation function, FC is a fully connected layer, Flatten is the flattening operator, and ⊙ is the Hadamard product;

[0064] Step S7: Perform steps S2 to S6 for each compound in the PPI regulator identification training set for the PPI family. Calculate the loss function based on the identification results of each compound, and adjust the parameters of the hierarchical hypergraph attention classification model based on the loss function.

[0065] The other steps and parameters are the same as those in one of the specific implementation methods one to four.

[0066] In this invention, the hierarchical hypergraph attention regression model has the same structure as the hierarchical hypergraph attention classification model. Similarly, the hierarchical hypergraph attention regression model is trained using the PPI modulator efficacy prediction training set.

[0067] Specific Implementation Method Six: This implementation method differs from Specific Implementation Methods One to Five in that the functional groups in the current compound are defined as follows:

[0068] (1) Extract the set of atoms within the smallest ring in a compound molecule as a functional group;

[0069] (2) Find all the central atoms (including carbon, nitrogen, oxygen, phosphorus and sulfur) in the compound molecule that are not in the smallest ring. For any central atom, combine the central atom and the environmental atoms connected to the central atom through single bonds, double bonds and triple bonds to form a functional group. The environmental atoms are determined according to the prior knowledge of the functional group. That is, a functional group is obtained for each central atom.

[0070] (3) Find chemical bonds not covered in (1) and (2) and define the atoms connected by such chemical bonds as a functional group.

[0071] The other steps and parameters are the same as those in one of the specific implementation methods one to five.

[0072] Using the strategy of this implementation, the atomic-level graph is divided into various functional groups, which serve as nodes in the functional group-level graph; if two functional groups have connected atoms in the atomic-level graph, an edge is set in the functional group-level graph to reflect their interaction relationship.

[0073] Specific Implementation Method Seven: This implementation method differs from Specific Implementation Methods One through Six in that the elements in the correlation matrix are updated based on the atomic features and functional group features output by the two GATs, resulting in an updated correlation matrix:

[0074]

[0075] Among them, s j It is a feature of the j-th functional group (based on the functional group feature X′) fg (obtained), z i It is a characteristic of the i-th atom in the j-th functional group (based on atomic characteristic X′). atom (obtained), W p and W q It is a weight matrix, m is the weight vector, and the superscript T represents the transpose. It is the set of atoms in the j-th functional group, hv k It is the k-th atom in the j-th functional group, z k It is a characteristic of the k-th atom in the j-th functional group, H i ′ ,j It is the element in the i-th row and j-th column of the updated correlation matrix H′.

[0076] The other steps and parameters are the same as those in one of the specific implementation methods one to six.

[0077] In a molecule, each functional group can contain multiple atoms, and each atom can belong to more than one functional group; however, the importance of different atoms (functional groups) to their related functional groups (atoms) varies due to factors such as their position and electronegativity; therefore, in the process of clustering node (atom) and hyperedge (functional group) features, an attention mechanism is applied to the correlation matrix H to distinguish the contributions of atoms and functional groups; the goal of the attention mechanism is to learn a dynamic correlation matrix H, which measures the degree of correlation between atoms and functional groups through non-binary values.

[0078] Graph attention networks are used to extract molecular features from atomic-level and functional group-level graphs, respectively. Then, the concept of attention mechanism is used to assign importance weights to the neighbors of nodes during information propagation, so that the network can focus on the most relevant nodes and their interactions, thereby extracting more meaningful features.

[0079] Specific Implementation Method Eight: This implementation method differs from Specific Implementation Methods One to Seven in that the specific process of step S3 is as follows:

[0080] X″ fg =σ(B -1 H′ T X′ atom Θ atom )

[0081] Where σ is the LeakyReLU activation function, It is the parameter matrix that needs to be trained. It is the real number field, and M′×M″ is the matrix Θ. atom Dimensions It is the hypermarginality matrix, |E h | represents the number of superedges in the superedge set, X″ fg It is a functional group-level feature that contains atomic information;

[0082] X″ atom =σ(D -1 H′X′ fg Θ fg )

[0083] in, It is the parameter matrix that needs to be trained, X″ atom It is an atomic-level feature that contains information about functional groups. It is the node degree matrix in the hypergraph, |V h | represents the number of nodes in the node set of the hypergraph, and the superscript -1 represents the inverse of the matrix.

[0084] The other steps and parameters are the same as those in any of the specific implementation methods one to seven.

[0085] For the l-th level graph attention unit, X″ will be calculated. fg The process is denoted as Calculate X″ atom The process is denoted as Calculate X′ atom The process is denoted as Calculate X′ fg The process is denoted as The process of residual join is denoted as:

[0086]

[0087]

[0088] Two graph attention networks operate on the atomic-level graph and the functional group-level graph, respectively, to update node features. Then, based on the constructed hypergraph, atomic convolution (AtomConv) aggregates the atomic features output by the graph attention network to the functional group, and functional group convolution (FgConv) aggregates the functional group features output by the functional group to the atom. Finally, the output features of the graph attention network and the hypergraph attention network are combined through residual connections.

[0089] Specific Implementation Method Nine: This implementation method differs from Specific Implementation Methods One through Eight in that the connection result... for:

[0090]

[0091] Where || represents a channel-level connection, It is the result of the summation pooling operation. This is the result of the average pooling operation. It is the result of the max pooling operation;

[0092] atomic-level global features for:

[0093]

[0094] in, It is the output of the nth convolutional block. yes The output after passing through a 1×1 convolutional layer.

[0095] The other steps and parameters are the same as those in one of the specific implementation methods one to eight.

[0096] To preserve more global information, summative pooling, average pooling, and max pooling operators are used to encode local features into graph-level features. Stacking multiple convolutional blocks can extract higher-level features, and residual connections with 1×1 convolutional layers are introduced to accelerate the training process. Similarly, the method in this implementation can obtain functional group-level global features.

[0097] Specific Implementation Method Ten: This implementation method differs from Specific Implementation Methods One through Nine in that the loss function used in the hierarchical hypergraph attention classification model is... for:

[0098]

[0099] Where Γ is the number of compounds in the PPI regulator identification training set. y is the identification result of the i-th compound. i is the actual label of the i-th compound, and σ′ is the sigmoid function;

[0100] The loss function used in the hierarchical hypergraph attention regression model is... for:

[0101]

[0102] The other steps and parameters are the same as those in any of the specific implementation methods one to nine.

[0103] For classification tasks, this invention uses Matthew correlation coefficient (MCC), F1 score, area under the receiver operating curve (AUROC), and accuracy (ACC) as evaluation metrics; for regression tasks, this invention uses mean absolute error (MAE), root mean square error (RMSE), Pearson correlation coefficient (ρ), Kendall correlation coefficient (τ), and Spearman correlation coefficient (r). s () as an evaluation indicator.

[0104] This invention selects the state-of-the-art PPI modifier prediction method SELPPI as a reference method, and the experimental results on the classification task are shown in Table 1. It can be seen that, in terms of various evaluation metrics, this invention achieves good performance on all datasets. In particular, on the Bromodomain / Histone, LEDGF / IN, and MDM2-Like / P53 datasets, the accuracy of this invention is improved by 4.5%-4.9%, 5.6%-8.9%, and 1.7%-6.3%, respectively, demonstrating the powerful ability of this invention to identify PPIMs.

[0105] Table 1

[0106]

[0107] The experimental results on the regression task are shown in Table 2. It can be seen that, compared with SELPPI, this invention demonstrates comparable performance on almost all datasets. Specifically, it performs best on the Bromodomain / Histone dataset, with the three correlation coefficients being 0.839, 0.736, and 0.849, respectively, while the RMSE and MAE values ​​are only 0.610 and 0.449, respectively. Furthermore, this invention achieves the highest Pearson correlation coefficient values ​​on all datasets, indicating a strong linear correlation between the predicted and true labels.

[0108] Table 2

[0109]

[0110]

[0111] The above examples of the present invention are merely illustrative of the computational model and process of the present invention, and are not intended to limit the implementation of the present invention. Those skilled in the art will recognize that other variations or modifications can be made based on the above description. It is impossible to exhaustively list all possible implementations here. Any obvious variations or modifications derived from the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A method for predicting protein interaction regulators based on hierarchical graph modeling, characterized in that, The method specifically includes the following steps: Step 1: Construct a PPI regulator identification dataset and a PPI regulator efficacy prediction dataset for each PPI family. Divide the constructed PPI regulator identification dataset into three parts: training set, validation set, and test set. Then divide the constructed PPI regulator efficacy prediction dataset into three parts: training set, validation set, and test set. Step 2: Generate atomic-level and functional group-level maps for each compound in the PPI regulator identification dataset targeting each PPI family; then generate atomic-level and functional group-level maps for each PPI regulator in the PPI regulator efficacy prediction dataset targeting each PPI family. Step 3: Construct hierarchical hypergraph attention classification models with identical structures for each PPI family. Then, use the PPI modifiers of each PPI family to identify the atomic-level and functional group-level graphs of the training set to train the hierarchical hypergraph attention classification models. Finally, use the validation set and test set to validate and test the trained hierarchical hypergraph attention classification models. Step 4: Construct hierarchical hypergraph attention regression models with identical structures for each PPI family. Then, use the atomic-level and functional group-level graphs of the PPI modulator efficacy prediction training set for each PPI family to train the hierarchical hypergraph attention regression models. Finally, use the validation set and test set to validate and test the trained hierarchical hypergraph attention regression models. Step 5: For each compound to be identified, input it into a pre-trained hierarchical hypergraph attention classification model to determine whether it is a regulator of a certain PPI family. If the compound is a regulator of a certain PPI family and its regulatory efficacy needs to be predicted, then input the compound into the hierarchical hypergraph attention regression model of the corresponding PPI family to obtain the regulatory efficacy prediction result. If the compound is a regulator of a certain PPI family and regulatory efficacy does not need to be predicted, then the process ends. If the compound is not a regulator of a certain PPI family, then the process ends. If a compound is known to be a PPI regulator of a certain PPI family and the regulatory efficacy of the PPI regulator needs to be predicted, the PPI regulator can be directly input into the hierarchical hypergraph attention regression model of the corresponding PPI family to obtain the regulatory efficacy prediction result.

2. The method for predicting protein interaction regulators based on hierarchical graph modeling according to claim 1, characterized in that, In step one, a PPI modifier identification dataset is constructed for each PPI family, and the constructed dataset is divided into three parts: a training set, a validation set, and a test set; the specific process is as follows: Perform steps 1 through 4 separately for each PPI family: Step 1: Collect modulators targeting the current PPI family from the database; Step 2: Use Morgan fingerprinting and Butina clustering algorithms to cluster the regulators collected in Step 1. After obtaining the clustering results, select one regulator from each category and use the selected regulators to form positive samples in the PPI regulator identification dataset that target the current PPI family. Step 3: Obtain a number of negative samples equal to the number of positive samples, and use the positive and negative samples together to form a PPI modifier identification dataset targeting the current PPI family. Of these, half of the negative samples came from positive samples targeting other PPI families, while the other half of the negative samples were compounds randomly selected from the ChEMBL database. Step 4: Divide the PPI regulator identification dataset targeting the current PPI family into three parts: training set, validation set, and test set.

3. The method for predicting protein interaction regulators based on hierarchical graph modeling according to claim 2, characterized in that, In step one, a PPI modulator efficacy prediction dataset is constructed for each PPI family, and the constructed PPI modulator efficacy prediction dataset is divided into three parts: a training set, a validation set, and a test set; the specific process is as follows: Perform steps (1) through (3) for each PPI family separately: Step (1): Collect modulators targeting the current PPI family from the database; Step (2): The regulators collected in step (1) are clustered using Morgan fingerprint and Butina clustering algorithms. After obtaining the clustering results, one regulator is selected from each category and a label is added to each selected regulator. The added label is the activity value of the corresponding regulator. A dataset predicting the effectiveness of PPI modifiers targeting the current PPI family was constructed using the selected modifiers and added labels. Step (3): Divide the PPI regulator efficacy prediction dataset targeting the current PPI family into three parts: training set, validation set, and test set.

4. The method for predicting protein interaction regulators based on hierarchical graph modeling according to claim 3, characterized in that, In step two, atomic-level and functional group-level maps are generated for each compound in the dataset identifying PPI modulators targeting various PPI families; the specific process is as follows: Define the atomic level diagram as G atom =(V atom E atom ), where V atom It is a set of nodes in an atomic-level diagram, where each atom in the current compound is treated as a node, E atom It is the set of edges in the atomic-level graph, where each bond in the current compound is treated as an edge; Define the functional group level diagram as G fg =(V fg E fg ), where V fg It is a set of nodes in the functional group hierarchy graph, where each functional group in the current compound is treated as a node, E fg It is the set of edges of the functional group level graph. If two functional groups have adjacent atoms in the atomic level graph, then there is an edge between the two functional groups.

5. The method for predicting protein interaction regulators based on hierarchical graph modeling according to claim 4, characterized in that, The training of the hierarchical hypergraph attention classification model using the PPI modifier identification training set of each PPI family is specifically as follows: Step S1: For any PPI family, generate a hypergraph for each compound in the PPI regulator identification training set for that PPI family, and define the hypergraph as G. h =(V h E h ), where V h It is the set of nodes in the hypergraph, where each atom in the current compound is treated as a node, E h It is the set of hyperedges in the hypergraph, where each functional group in the current compound is treated as a separate hyperedge; Using the association matrix To model the node relationships in the hypergraph, if the i-th atom is in the j-th functional group, then H i,j =1, otherwise H i,j =0, H i,j It is the element in the i-th row and j-th column of the correlation matrix; Step S2: For any compound in the PPI modulator identification training set of this PPI family, use the atomic-level map and functional group-level map of the compound as inputs to the two GATs in the first-level graph attention unit, and obtain the atomic features X′ output by the two GATs. atom and functional group characteristics X′ fg ; Then, the elements in the correlation matrix are updated based on the atomic and functional group features output by the two GATs to obtain the updated correlation matrix. Step S3: Use the atomic features and functional group features output by the two GATs as input to the hypergraph attention network in the first-level graph attention unit, and obtain the functional group-level features containing atomic information and the atomic-level features containing functional group information based on the updated correlation matrix. Atomic-level features X″ containing functional group information atom Atomic features X′ output by GAT atom By combining residual joins, the join result is obtained. Functional group-level features X″ containing atomic information fg and the functional group feature X′ output by GAT fg By combining residual joins, the join result is obtained. Step S4, and As input to the second-level graph attention unit, the working process within the second-level graph attention unit is the same as the working process within the first-level graph attention unit; This process continues until the output of the Lth level graph attention unit is obtained. and Then Obtained through GAT Will Obtained through GAT Using summation pooling, average pooling, and max pooling operators Perform global feature extraction and concatenate the feature extraction results to obtain the concatenation result. Using summation pooling, average pooling, and max pooling operators Perform global feature extraction and concatenate the feature extraction results to obtain the concatenation result. Step S5: Connect the results The input consists of two parallel channels, one of which contains n convolutional blocks, and the other contains a convolutional layer with a kernel size of 1×1. The outputs of the two channels are residually concatenated, and then the result is passed through a ReLU activation function layer. The atomic-level global features output by the ReLU activation function layer are denoted as... Connect results The input consists of two parallel channels, one of which contains n convolutional blocks, and the other contains a convolutional layer with a kernel size of 1×1. The outputs of the two channels are residually concatenated, and then the result is passed through a ReLU activation function layer. The functional group-level global features output by the ReLU activation function layer are denoted as... Step S6, and As input to the multilayer perceptron, the compound identification result is output by the multilayer perceptron. Among them, MLP is a two-layer MLP with ReLU activation function, FC is a fully connected layer, Flatten is the flattening operator, and ⊙ is the Hadamard product; Step S7: Perform steps S2 to S6 for each compound in the PPI regulator identification training set for the PPI family. Calculate the loss function based on the identification results of each compound, and adjust the parameters of the hierarchical hypergraph attention classification model based on the loss function.

6. The method for predicting protein interaction regulators based on hierarchical graph modeling according to claim 5, characterized in that, The functional groups in the current compound are defined as follows: (1) Extract the set of atoms within the smallest ring in a compound molecule as a functional group; (2) Find all the central atoms in the compound molecule that are not in the smallest ring. For any central atom, combine the central atom and the environmental atoms connected to the central atom by single, double and triple bonds to form a functional group. The environmental atoms are determined according to the prior knowledge of the functional group, that is, a functional group is obtained for each central atom. (3) Find chemical bonds not covered in (1) and (2) and define the atoms connected by such chemical bonds as a functional group.

7. The method for predicting protein interaction regulators based on hierarchical graph modeling according to claim 6, characterized in that, The elements in the correlation matrix are updated based on the atomic and functional group features output by the two GATs to obtain the updated correlation matrix: Among them, s j It is a feature of the j-th functional group, z i It is a characteristic of the i-th atom in the j-th functional group, W p and W q It is a weight matrix, m is the weight vector, and the superscript T represents the transpose. It is the set of atoms in the j-th functional group, hv k It is the k-th atom in the j-th functional group, z k It is a characteristic of the k-th atom in the j-th functional group, H′ i,j It is the element in the i-th row and j-th column of the updated correlation matrix H′.

8. The method for predicting protein interaction regulators based on hierarchical graph modeling according to claim 7, characterized in that, The specific process of step S3 is as follows: X″ fg =σ(B -1 H′ T X′ atom I atom ) Where σ is the LeakyReLU activation function, It is the parameter matrix that needs to be trained. It is the real number field, and M′×M″ is the matrix Θ. atom Dimensions It is the hypermarginality matrix, |E h | represents the number of superedges in the superedge set, X″ fg It is a functional group-level feature that contains atomic information; X″ atom =σ(D -1 H′X′ fg I fg ) in, It is the parameter matrix that needs to be trained, X″ atom It is an atomic-level feature that contains information about functional groups. It is the node degree matrix in the hypergraph, |V h | represents the number of nodes in the node set of the hypergraph, and the superscript -1 represents the inverse of the matrix.

9. The method for predicting protein interaction regulators based on hierarchical graph modeling according to claim 8, characterized in that, The connection result for: Where || represents a channel-level connection, It is the result of the summation pooling operation. This is the result of the average pooling operation. It is the result of the max pooling operation; atomic-level global features for: in, It is the output of the nth convolutional block. yes The output after passing through a 1×1 convolutional layer.

10. The method for predicting protein interaction regulators based on hierarchical graph modeling according to claim 9, characterized in that, The loss function used in the hierarchical hypergraph attention classification model is... for: Where Γ is the number of compounds in the PPI regulator identification training set. y is the identification result of the i-th compound. i is the actual label of the i-th compound, and σ′ is the sigmoid function; The loss function used in the hierarchical hypergraph attention regression model is... for:

Images