CN116189759A - Virtual screening method and application of group induction lead compound - Google Patents

Abstract

The invention discloses a virtual screening method of a group induction lead compound, which mainly comprises the following steps: the input molecular compound structure constructs a molecular adjacency matrix through pretreatment, and is sent into a GNN1 network to generate compound characteristics; the input protein sequence extracts the protein amino acid composition and dipeptide frequency composition of the protein sequence to form a preliminary protein feature vector, and the preliminary protein feature vector is sent into a cross network to generate cross fusion features; simultaneously, generating a corresponding contact diagram by the protein sequence, and then sending the contact diagram into a GNN2 network to generate protein sequence characteristics; finally, three characteristic combinations are sent to the full-connection layer for prediction to obtain an affinity value. The invention can be used for discovering new compounds with quorum sensing activity and providing new thinking and means for controlling and preventing bacteria such as bacterial wilt and the like; meanwhile, the method can be used for efficiently screening out the compounds combined with the PhcA and PhcR proteins, so that the compounds with quorum sensing activity can be found.

Description

A virtual screening method and application of quorum sensing lead compounds

technical field

The invention relates to the field of medicinal chemistry, in particular to a targeted virtual screening method and application of quorum sensing lead compounds of R. solanacearum.

Background technique

Ralstonia solanacearum is one of the most destructive soil-borne pathogens in the world. The pathogen is widely distributed in tropical, subtropical and temperate climate regions around the world, and gradually spreads to high-latitude and high-altitude regions. The virulence behaviors related to soil-borne R. solanacearum invasion into crop rhizosphere are regulated by quorum sensing. Ralstonia solanacearum has two quorum sensing systems: AHL system and trihydroxypalmitic acid methyl ester (3-OH PAME) system, and the AHL system does not affect the virulence. R. solanacearum globally regulates metabolism and virulence behavior through the 3-OH PAME quorum sensing system, coordinates the assembly of various secretion systems, and directs the sequential expression and secretion of various virulence factors, thus successfully completing the rhizosphere invasion process. The system consists of PhcBSR synthesis components and regulatory factor PhcA. Among them, PhcB is responsible for synthesizing the signal molecule 3-OH PAME, and PhcS is responsible for receiving the sensory signal molecule. When the concentration of 3-OH PAME exceeds a certain threshold, PhcS activates PhcR, thereby releasing the inhibition of PhcR on PhcA. PhcA not only regulates the primary metabolism and AHL quorum sensing system of R. solanacearum, but also regulates the motility, siderophore, biofilm, cell wall degrading enzymes, type III virulence factors and exopolysaccharides of R. solanacearum, which destroy the plant immune system, etc. Virulence behavior closely related to the rhizosphere invasion process. Some studies have tried to block the rhizosphere invasion process of soil-borne R. solanacearum by degrading quorum-sensing molecules, but the blocking effect is not satisfactory.

Currently, virtual screening is a very common strategy in computer-aided drug design and has been widely used. Drug-target affinity (DTA) prediction is an important step in virtual screening, which can quickly match targets and drugs and speed up the drug development process. DTA prediction provides information on the binding strength of a drug to a target protein and can be used to show whether a small molecule is bound to a protein. For proteins with known structure and site information, we can use molecular simulation and molecular docking to perform detailed simulations to obtain more accurate results, which is called structure-based virtual screening. However, there are still many proteins without structural information. Even with homology models, it is still difficult to obtain structural information for many proteins. Therefore, it is an urgent problem to use sequences (sequence-based virtual screening) to predict the binding affinity of proteins and drug molecules, which is also the focus of the present invention.

Virtual screening based on molecular docking has become the core technology of computer-aided compound design and is widely used in the targeted development of new compounds. Therefore, using virtual screening of quorum sensing lead compounds of R. solanacearum to interfere with quorum sensing may be one of the important ways to control soil-borne bacterial wilt.

Contents of the invention

The object of the present invention is to provide a targeted virtual screening method for quorum sensing lead compounds of R. solanacearum, which can automatically search for the best graph structure based on reinforcement learning, and improve the performance of the virtual screening model for quorum sensing lead compounds of the graph network , and it is a screening method for quorum sensing lead compounds based on the protein structure of PhcA and PhcR, which can be used to discover new compounds with quorum sensing activity, and provide new ideas and means for the control and prevention of bacteria such as Ralstonia solanacearum.

The technical scheme that the present invention takes is as follows:

A virtual screening method for quorum sensing lead compounds, which sends molecular compound structure and protein sequence as input to the preprocessing module to extract preliminary features, and then sends it to the prediction model network, wherein the structure and parameters of the prediction model are passed through the LSTM controller Training generation; the specific process is as follows:

The input molecular compound structure is preprocessed to construct a molecular adjacency matrix, which is sent to the GNN1 network to generate compound characteristics;

From the input protein sequence, extract its protein amino acid composition and dipeptide frequency to form a preliminary protein feature vector, send it to the cross network to generate cross fusion features; at the same time, generate a corresponding contact map for the protein sequence, and then send it to the GNN2 network to generate a protein sequence feature;

Finally, the combination of the three features is sent to the fully connected layer to predict the affinity value.

Further, let the constructed molecular adjacency matrix be X ₁ , the element value of the adjacent atomic matrix on the molecular structure diagram is 1, and the non-adjacent element is 0, and the size of the molecular adjacency matrix is (n*n), where n is The number of nodes, i.e. the number of all atoms.

Further, use the Pconsc4 software to process the protein sequence, and output the probability matrix of residuals versus contact or not, the size of which is m*m, retain values greater than 0.5 in the matrix, and set other values to 0, and the filtered matrix is the protein contact map X ₂ , where m is the number of residuals.

Further, the amino acid composition is the frequency of occurrence of each of the 20 amino acids constituting the sequence, and the frequency of the dipeptide is the frequency of occurrence of an amino acid pair composed of any two amino acids.

Further, the prediction model network is composed of GNN1 network, GNN2 network, and crossover network connected in parallel, which are combined and sent to a splicing layer and DROPOUT layer, and then connected to two fully connected layers.

Further, the cross network is composed of 5 cross layers connected in series, and finally connected with a 128-dimensional fully connected layer, the output of the fully connected layer is f ₃ , and each cross layer has the following formula:

C _l+1 ＝X ₀ C ^T _l W _c,l +b _c,l +C _l

Where: l=1,2,...,5, C _l and C _l+1 are the outputs of the l-th layer and l+1-th layer cross layer respectively, C ₀ is the combination of amino acid composition and dipeptide frequency X ₃ , W _c,l and b _c,l are the connection parameters between the two layers; all variables in the above formula are column vectors. The output of each layer is the output of the previous layer plus the feature cross.

Further, the molecular adjacency matrix and the protein contact map are respectively input into two different GNN1 and GNN2 networks, each network is composed of three GNN layers, and the output features of the two GNN networks are f ₁ and f ₂ , and then Add the cross-fusion feature, and after splicing, it will be f ₁ +f ₂ +f ₃ , to obtain the overall features of the corresponding small molecule-protein pair for prediction; then send it to a fully connected layer with an output dimension of 128, and then send it to the first Two fully connected layers, the output dimension is 1, which is the affinity value predicted by the network.

Furthermore, the LSTM controller is used to optimize the virtual screening network model. Model optimization is to use reinforcement learning in the determined parameter space to obtain the optimal structure parameters of two GNNs and the parameters of other neurons in the entire network; the GNN structure M needs to determine how many parameters: sampling function (S), correlation metric function (Att), aggregation function (Agg), multi-head attention amount (K), output hidden embedding (Dim), and activation function (Act).

Furthermore, the optimization consists of two steps. First, the LSTM predicts the corresponding operations of S, Att, Agg, Act, K, and Dim of a GNN1. Each prediction is performed by the softmax classifier of the LSTM, and then the predicted value is input to the next At one time point, the next parameter prediction is obtained; when the layer number of GNN Layer reaches 3, the LSTM controller completes the generation of the architecture; repeat this process to generate the parameters of GNN2; build and train the entire prediction network, and obtain the GNN network and other networks Layer weight parameters; then optimize the parameters of LSTM with reinforcement learning based on the accuracy obtained after network training to obtain an optimized controller model; the two steps are alternately executed for a certain number of steps to obtain the final screening network model.

The above method can be applied in the virtual screening of quorum sensing lead compounds of R. solanacearum.

The beneficial effects of the present invention are as follows: firstly, the method extracts molecular adjacency matrix, protein contact map, and protein sequence cross-fusion features to form multi-dimensional features, which can better reflect the characteristics of molecules and proteins; secondly, the method uses reinforcement learning to optimize the model structure, It avoids selecting model parameters by experience or a lot of manual work in the past. The perfection and promotion of this method can effectively carry out virtual screening of lead compounds, which has broad prospects and extraordinary significance.

Description of drawings

Fig. 1 is a virtual screening model figure of the present invention;

Figure 2 is a molecular graph representation.

Detailed ways

The present invention will be described in detail below in conjunction with the accompanying drawings.

As shown in Fig. 1, a kind of virtual screening method of quorum sensing lead compound, this method is the screening method of the quorum sensing lead compound based on PhcA and PhcR protein structure, it is through the computer aided virtual screening technology based on PhcA and PhcR protein structure, molecular A screening method for quorum sensing lead compounds of R. solanacearum established by Biological Technology. The realization of the method is through the virtual screening model. The input of the virtual screening model is the molecular compound structure and protein sequence, which are then sent to the preprocessing module to extract preliminary features, and then sent to the prediction model network. The structure and parameters of the prediction model are passed through the LSTM controller. training generated. Basic process: input the molecular formula of the compound, construct the molecular adjacency matrix, and send it to the GNN1 sub-network to generate the compound characteristics. Input the protein sequence, extract its protein amino acid composition and dipeptide frequency to form a preliminary protein feature vector, send it to the cross network to generate cross fusion features; at the same time, generate the corresponding contact map of the protein sequence, and then send it to the GNN2 network to generate protein sequence features . Finally, the combination of the three features is sent to the fully connected layer to predict the affinity value, and the output value is 0 without affinity and 1 with affinity. The following introduces the construction of molecular adjacency matrix, protein contact map, generation of protein sequence cross-fusion features, GNN structure optimization, and optimization of other parameters of the prediction network are realized by the LSTM controller.

1. Data preprocessing

(1) Construct molecular adjacency matrix

Molecules are represented in the dataset, in SMILES format. The molecular graph is constructed according to the drug SMILES string, which uses atoms as nodes and bonds as edges. The molecular graph structure construction process representation is shown in Figure 2. Let the molecular adjacency matrix obtained by construction be X ₁ , the element value of the adjacent atomic matrix is 1, and the value of non-adjacent is 0, and the size is (n*n), where n is the number of nodes in the graph, that is, the number of all atoms.

(2) Protein contact map

The protein contact diagram is a graphical representation method for describing the interaction between proteins, which shows the contacts and interactions between proteins, and is used to describe protein structure and function. Use Pconsc4 software to process the protein sequence, output the probability matrix of residuals versus contact or not, the size is m*m, keep the values greater than 0.5 in the matrix, and set other values to 0, and the filtered matrix is the protein contact map X ₂ .

(3) Protein composition characteristics

The protein composition feature is the combination X ₃ of amino acid composition and dipeptide frequency, and the size is 420 dimensions. Amino acid composition is the frequency of occurrence of each of the 20 amino acids that make up the sequence. The frequency of a dipeptide is the frequency of any pair of amino acids composed of two amino acids. There are 20 amino acids that make up a protein sequence, and there are 400 dipeptides.

2. Prediction model architecture

(1) Cross network

The input of the cross network is X ₃ , the network is composed of 5 cross layers in series, and finally connected with a 128-dimensional fully connected layer, the output of the fully connected layer is f ₃ , and each cross layer has the following formula:

C _l+1 ＝X ₀ C ^T _l W _c,l +b _c,l +C _l

Where: l=1,2,...,5. C _l and C _l+1 are the outputs of the l-th layer and the l+1-th layer crosslayer respectively, C ₀ is X ₃ , W _c,l and b _c,l are the connection parameters between these two layers. All variables in the above formula are column vectors. The output of each layer is the output of the previous layer plus the feature cross.

(2) The overall structure of the affinity prediction network

The prediction model consists of two GNN networks (GNN1, GNN2) and cross-connected networks in parallel, which are combined and sent to a splicing layer and DROPOUT layer, and then connected to two fully connected layers. Molecular adjacency matrices and protein contact graphs of drug molecules and proteins are input into two different GNN1, GNN2 networks. Each network consists of 3 GNN layers. The output features of the two GNN networks are f ₁ , f ₂ , plus the cross-fusion feature, which is f ₁ +f ₂ +f ₃ after splicing, and the overall features of the corresponding small molecule-protein pairs for prediction are obtained. Then it is sent to a fully connected layer with an output dimension of 128, and then sent to the second fully connected layer with an output dimension of 1, which is the affinity value predicted by the network. The specific structural parameters of the GNN layer and other network layers are obtained from the following network model optimization training process.

(3) LSTM controller realizes virtual screening network model optimization

Model optimization is to use reinforcement learning in a determined parameter space to obtain the optimal structural parameters of two GNNs and the parameters of other neurons in the entire network. The GNN structure M needs to determine several parameters: sampling function (S), correlation measurement function (Att), aggregation function (Agg), multi-head attention number (K), output hidden embedding (Dim) and activation function (Act).

The specific description and the corresponding parameter values of each parameter are as follows:

1. Output the hidden embedding (Dim). Dim is the output dimension of each layer of GNN, which is an integer value.

2. Sampling function (S). For each layer of GNN, a sampling function is required. Sampling is used in graph neural networks to select receptive fields for a given target node. Three methods are used in this method, a. fixed neighbor number sampling method, b. importance sampling method, c. first-order neighbor sorting method.

3. Correlation metric function (Att) and the number K of multi-head attention. For each layer of GNN, we choose an Att method and the number K of multi-head attention. Att can choose two measurement functions: GAT and GCN, which correspond to the two network structures of GAT and GCN respectively. The GCN network includes 2 graph convolution layers, 1 ReLU activation function layer and 1 Dropout layer. The GAT network includes K graph multi-head attention layers, a Softmax activation function layer and a Dropout layer. Among them, GAT assigns neighborhood importance by using attention layers, and GCN assigns neighborhood importance according to the degree of nodes.

4. Aggregation function (Agg). For each layer of GNN, Agg aggregation needs to be used. Optional aggregation functions Agg are: a. sum aggregator, b. mean aggregator, c. pool aggregator.

5. Activation function (Act). For each layer of GNN, the Act activation function needs to be used. The optional activation functions Act are: a.ReLU, b.LeakyReLU, c.ELU, d.Linear, e.Softmax. Increase the nonlinear fitting ability of the graph network and improve the expression ability of the model.

Network model optimization uses an LSTM controller neural network to train the network, which consists of two steps. First, the LSTM predicts the corresponding operation of [S, Att, Agg, Act, K, Dim] of a GNN1. Each prediction is determined by the softmax of the LSTM. The classifier executes, then feeds that prediction to the next time point to get the next parameter prediction. When the layer number 3 of GNN Layer is reached, the LSTM controller completes the generation of an architecture; repeat this process to generate the parameters of GNN2. Construct and train the entire prediction network to obtain the weight parameters of the GNN network and other network layers. Then, based on the accuracy rate obtained after network training, the parameters of LSTM are optimized by reinforcement learning to obtain an optimized controller model. The two steps are executed alternately for a certain number of steps to obtain the final screening network model.

The training of the model is further explained below:

Model training uses the public KIBA dataset. The data set includes a total of 229 unique proteins and 2111 unique drugs, and there are 118254 affinity pairs between proteins and drug molecules. In the training method, the data set is divided into training set, verification set and test set according to the ratio of 80%:10%:10%.

Training parameters: The controller is an LSTM network with 100 hidden units. It is trained using the ADAM optimizer with a learning rate of 0.0035. The controller samples the graph network structure, generates sub-models, and trains for 200 epochs. During training, L2 regularization with λ = 0.0005 is applied. Furthermore, dropout with p = 0.5 is applied to the inputs of both layers as well as the normalized attention layer.

LSTM consists of three layers: input layer, hidden layer and output layer, the input dimension is 6*1 dimension, the hidden layer neurons are 100, and the time step is 10.

The LSTM network parameters are set as follows:

Layer1: LSTM (input_size=8, hidden_size=100, num_layers=2)

Layer2: Dropout (p=0.5)

Layer3: Linear (hidden_size=500, n_class=1)

Among them, input_size represents the input data dimension; hidden_size represents the output dimension; num_layers represents the stacked layers of LSTM, the default is 1; n_class represents the output dimension of the LSTM network, and 1 represents the output regression value.

After the controller is trained 1000 times, we let the controller output the best model from the 200 sampled GNNs. The results show that this optimization method can design the best performance model of the original virtual screening model.

Experimental results

After the controller is trained 1000 times, we let the controller output the best model from the 200 sampled GNNs. The results show that this optimization method can design the best performance model of the original virtual screening model.

The structure of the best virtual screening model after optimization and screening is as follows:

GNN1 structure:

Layer1: molecular graph attention layer GAT1 (input dimension = (n*n), output dimension = 128, number of hidden layer units = 128, number of head attention = 4, activation function = elu(), aggregation function = sum ()).

Layer2: Molecular graph convolutional layer GCN2 (input dimension=128, output dimension=256, number of hidden layer units=256, number of head attention=4, activation function=relu(), aggregation function=max()).

Layer3: molecular graph attention layer GAT3 (input dimension=256, output dimension=128, number of hidden layer units=128, number of head attention=8, activation function=elu(), aggregation function=avg()).

GNN2 structure:

Layer1: Protein graph convolutional layer GCN4 (input dimension = (m*m), output dimension = 64, number of hidden layer units = 64, number of head attention = 16, activation function = relu(), aggregation function = max ()).

Layer2: Protein graph attention layer GCN5 (input dimension=64, output dimension=256, number of hidden layer units=256, number of head attention=4, activation function=elu(), aggregation function=pooling()).

Layer3: Protein graph convolutional layer GCN6 (input dimension=256, output dimension=256, number of hidden layer units=256, number of head attention=16, activation function=relu(), aggregation function=max()).

Cross network: input dimension=420, output dimension=128, number of layers n=5

Feature splicing layer: Concat1 (input dimension=(128,256,128), output dimension=512).

Dropout layer: (512, 512), ratio p=0.5.

Fully connected layer: Linear(512,128).

Fully connected layer: Linear(128,1).

After 300 iterations, the network reached a better state, and the corresponding parameters were saved for virtual screening of quorum sensing lead compounds of R. solanacearum.

In a word, the present invention provides a screening method for quorum sensing lead compounds based on the protein structure of PhcA and PhcR, which can be used to discover new compounds with quorum sensing activity, and provide new ideas and ideas for the control and prevention of bacteria such as Ralstonia solanacearum. means. The method combines computer-aided virtual screening technology and molecular biology technology to efficiently screen compounds that bind to PhcA and PhcR proteins, thereby discovering compounds with quorum sensing activity. The method has the advantages of simple operation, high efficiency and rapidity, and high screening accuracy, and can be widely used in fields such as biomedicine and agriculture.

It is worth noting that PhcA and PhcR in the present invention are two key proteins in Ralstonia solanacearum, and their structure and function play an important role in quorum sensing of bacteria. However, in other bacteria, there may be different quorum-sensing proteins, so it is necessary to screen and study according to different bacterial species in order to find quorum-sensing lead compounds suitable for different bacteria.

The basic principles, main features and advantages of the present invention have been shown and described above. Those of ordinary skill in the art should understand that the above-mentioned embodiments do not limit the protection scope of the present invention in any form, and all technical solutions obtained by means of equivalent replacement or the like fall within the protection scope of the present invention.

The parts not involved in the present invention are the same as the prior art or can be realized by adopting the prior art.

Claims (10)

1. A method for virtual screening of quorum sensing lead compounds, characterized in that molecular compound structures and protein sequences are sent to the preprocessing module as input to extract preliminary features, and then sent to the prediction model network, wherein the structure of the prediction model and The parameters are generated through LSTM controller training; the specific process is as follows: The input molecular compound structure is preprocessed to construct a molecular adjacency matrix, which is sent to the GNN1 network to generate compound features; From the input protein sequence, extract its protein amino acid composition and dipeptide frequency to form a preliminary protein feature vector, send it to the cross network to generate cross fusion features; at the same time, generate a corresponding contact map for the protein sequence, and then send it to the GNN2 network to generate a protein sequence feature; Finally, the combination of the three features is sent to the fully connected layer to predict the affinity value. 2. the virtual screening method of a kind of quorum sensing lead compound according to claim 1, it is characterized in that, suppose the molecular adjacency matrix that builds to obtain is X ₁ , the element value of the adjacent atomic matrix on the molecular structure diagram is 1, not Adjacency is 0, and the molecular adjacency matrix size is (n*n), where n is the number of nodes in the structure graph, that is, the number of all atoms. 3. the virtual screening method of a kind of quorum sensing lead compound according to claim 1, it is characterized in that, use Pconsc4 software to process protein sequence, output the probability matrix of residual error pair whether to contact, size is m*m, keep in the matrix If the value is greater than 0.5, other values are set to 0, and the filtered matrix is the protein contact map X ₂ , where m is the number of residuals. 4. the virtual screening method of a kind of quorum sensing lead compound according to claim 1, is characterized in that, amino acid composition is the frequency that 20 kinds of amino acids that constitute sequence occur respectively, and the frequency of dipeptide is the amino acid that any two amino acids form to the frequency of occurrence. 5. the virtual screening method of a kind of quorum sensing lead compound according to claim 1, is characterized in that, described predictive model network is made up of GNN1 network, GNN2 network, cross network parallel connection, sends into a splicing layer and DROPOUT after merging layer, followed by two fully connected layers. 6. the virtual screening method of a kind of quorum sensing lead compound according to claim 1, is characterized in that, described intersection network is made up of 5 intersection layers in series, connects a 128-dimensional fully connected layer at last, fully connected layer output is f ₃ , each intersection layer has the following formula: C _l+1 ＝X ₀ C ^T _l W _c,l +b _c,l +C _l Where: l=1,2,...,5, C _l and C _l+1 are the outputs of the l-th layer and l+1-th layer cross layer respectively, C ₀ is the combination of amino acid composition and dipeptide frequency X ₃ , W _c,l and b _c,l are the connection parameters between the two layers; all variables in the above formula are column vectors. The output of each layer is the output of the previous layer plus the feature cross. 7. according to the virtual screening method of a kind of quorum sensing lead compound described in claim 1 or 5 or 6, it is characterized in that, described molecular adjacency matrix, protein contact map are respectively input into two different GNN1 and GNN2 networks, Each network is composed of 3 GNN layers. The output features of the two GNN networks are f ₁ , f ₂ , plus the cross fusion feature, which is f ₁ +f ₂ +f ₃ after splicing, and the corresponding small The overall characteristics of the molecule-protein pair; then sent to a fully connected layer with an output dimension of 128, and then sent to a second fully connected layer with an output dimension of 1, which is the affinity value predicted by the network. 8. the virtual screening method of a kind of quorum sensing lead compound according to claim 1, is characterized in that, realizes virtual screening network model optimization by LSTM controller, and model optimization is exactly to use reinforcement learning to obtain two GNNs in the determined parameter space The optimal structure parameters and the parameters of other neurons in the entire network; GNN structure M needs to determine several parameters: sampling function (S), correlation measurement function (Att), aggregation function (Agg), multi-head attention number (K) , the output hidden embedding (Dim) and activation function (Act). 9. the virtual screening method of a kind of quorum sensing lead compound according to claim 8, is characterized in that, optimization is made up of two steps, first LSTM predicts the corresponding operation of S, Att, Agg, Act, K, Dim of a GNN1, Each prediction is performed by the softmax classifier of LSTM, and then the predicted value is input to the next time point to obtain the next parameter prediction; when the number of layers of GNN Layer reaches 3, the LSTM controller completes the generation of an architecture; repeat This process generates the parameters of GNN2; builds and trains the entire prediction network to obtain the weight parameters of the GNN network and other network layers; then optimizes the parameters of the LSTM based on the accuracy rate obtained after network training to obtain an optimized controller model; two The steps are executed alternately for a certain number of steps to obtain the final screening network model. 10. The application of a virtual screening method for quorum sensing lead compounds in Ralstonia solanacearum.

Images