A method for predicting compound blood-brain barrier permeability based on ternary mixed-level fusion convolutional neural network

Abstract

The present invention proposes a compound blood-brain barrier permeability prediction method based on a ternary mixed-level fusion convolutional neural network, and the specific implementation steps are as follows: (1) for the acquired compound, extract the SMILES expression and molecular structure diagram of the compound; (2) perform data preprocessing operations: extract MGF, MFF and MDF, obtain P-MGF, P-MFF, P-MDF respectively through feature screening, and normalize P-MGF, P-MFF, and P-MDF in turn to obtain,,; (3) construct a ternary mixed-level fusion convolutional neural network, design a ternary module to process,, three feature sets, obtain the processing results through the splicing layer fusion, and use the fully connected layer and the output layer for processing in turn; (4) input,,, into the ternary mixed-level fusion convolutional neural network, adjust the parameters to obtain the best prediction network; (5) apply the best prediction network to the prediction of the compound sample to be tested.

Description

Technical Field

[0001] The present invention belongs to the field of biological information processing, and further relates to the prediction of blood-brain barrier permeability of compounds based on ternary mixed-level fusion convolutional neural network in the field of deep learning technology. The present invention can be used to predict the blood-brain barrier permeability of compounds under development. Background Art

[0002] In the development of central nervous system (CNS) compounds, the permeability of the blood-brain barrier (BBB) ​​is an important indicator. As a key barrier to protect the brain and its surrounding neurons, the BBB can selectively screen and restrict the entry and exit of substances inside the brain, prevent harmful substances, pathogens or foreign substances from entering the brain, and maintain the stability of the internal environment of brain cells. However, the BBB also prevents most compounds from entering the brain, which poses a challenge to the treatment of brain diseases. In the traditional development of central nervous system compounds, clinical trials are required to evaluate whether the compounds can penetrate the BBB and have a therapeutic effect on the brain. However, this process requires a lot of capital investment and a long development cycle. Therefore, being able to effectively predict the BBB permeability of compounds before clinical trials can help R&D personnel adjust the experimental plan or optimize the structure and properties of the compound in a timely manner, thereby reducing R&D costs and improving R&D efficiency. At present, traditional BBB permeability prediction models mainly include logistic regression (LR), support vector machine (SVM), etc. However, the prediction methods of these models only focus on the characteristics of SMILES expressions and face the problem of data imbalance: there are more compounds that can permeate the BBB and fewer compounds that cannot permeate the BBB in the dataset, which causes the model to be more biased towards predicting the majority class and has a lower prediction accuracy for the minority class, making it difficult for these traditional models to meet the needs of actual compound research and development.

[0003] The Chinese patent document with publication number CN114360660A discloses a BBB permeability prediction method based on a multi-layer perceptron machine learning model. The invention method uses a multi-layer perceptron to learn specific molecular features, does not focus on the features of the molecular structure graph, and selects a single prediction model.

[0004] The Chinese patent document with publication number CN112802561A discloses a BBB permeability prediction method based on ensemble learning. The invention uses machine learning and ensemble methods to learn molecular fingerprint features, and takes the best learner to build the final prediction model. The model only focuses on molecular fingerprint features, and there is still room for optimization in model selection. Summary of the invention

[0005] The purpose of the present invention is to address the problems existing in the above-mentioned background technology: (1) only focusing on SMILES expression information while ignoring molecular structure diagram information; (2) the prediction method is single and only uses machine learning methods; (3) the accuracy of predicting compounds that cannot penetrate the BBB is too low. A compound blood-brain barrier permeability prediction method based on a ternary mixed-level fusion convolutional neural network is proposed. The method combines molecular structure diagrams, molecular fingerprints, and molecular descriptors to accurately predict whether a compound can penetrate the BBB and classify it, which is helpful for the diversified construction of BBB prediction models and improves the construction efficiency and prediction performance of BBB prediction models.

[0006] The present invention also provides a compound blood-brain barrier permeability prediction device based on a ternary mixed-level fusion convolutional neural network. After the SMILES expression and molecular structure diagram of the compound to be predicted are input into the system, the system can predict and output the BBB permeability of the compound with the help of the prediction model, which can be used to assist in compound research and development.

[0007] Terminology explanation:

[0008] BBB label: A label used to indicate whether a compound can penetrate the BBB. If a compound can penetrate the BBB, the BBB label of the compound is recorded as BBB+, otherwise it is recorded as BBB-.

[0009] Simplified molecular input line entry specification: Simplified molecular input line entry specification (SMILES) is a specification that uses ASCII strings to clearly describe molecular structures.

[0010] SMILES expression: Simplified molecular linear input specification expression. SMILES expression is a simplified molecular linear input specification expression, commonly used to represent chemical structures. It consists of a series of atomic symbols and connectors, which can accurately describe the structure of the molecule for computer processing and storage. SMILES expression can be imported by most molecular editing software and converted into two-dimensional graphics or three-dimensional models of molecules. Based on the main principle of cheminformatics that similar molecules have similar properties, it can be used to predict biochemical properties. By comparing the SMILES expressions of molecules, similar molecular structures can be quickly identified, and then it can be inferred that they may have similar properties.

[0011] Molecular structure graph: A graphical structure representing chemical molecules, where atoms represent nodes and chemical bonds represent edges. These atoms and chemical bonds can have various attributes, such as the type of atom, electronegativity, number of valence electrons, hybridization state, and type of chemical bond. Encoding these attributes into vector form, i.e., feature vectors, enables machine learning models to effectively process and analyze the structural information of molecules, and then perform tasks such as compound design and molecular property prediction.

[0012] Compound sample: Each compound in the compound dataset and its various properties, including but not limited to the SMILES and BBB labels of the compound.

[0013] Test compounds: refer to compound samples that need to be predicted for blood-brain barrier permeability through the model. These compounds may be known, under development, or need to evaluate their compound properties. By inputting the feature vector of the test compound into the trained model, the prediction result, that is, the blood-brain barrier permeability classification or probability of the compound, can be obtained.

[0014] Positive sample: a compound sample that can cross the BBB, with the BBB label as BBB+, also called BBB+ sample.

[0015] Negative sample: a compound sample that cannot cross the BBB, labeled as BBB-, also called BBB-sample.

[0016] ROC curve: Receiver operating characteristic curve, is a curve that reflects the relationship between sensitivity and specificity. Its horizontal axis is the false positive rate FPR (False positive rate), which represents the ratio of positive samples that are incorrectly predicted as negative samples to all negative samples, and the vertical axis is the true positive rate TPR (True positive rate), which represents the ratio of positive samples that are correctly predicted as positive samples to all positive samples. The area under the curve is called AUC (Area under curve). The larger the area, the higher the prediction accuracy.

[0017] SE: Sensitivity, also known as the true positive rate, represents the proportion of successful samples (the true label is positive) among all positive samples predicted by the model, and is used to measure the ability of the model to predict positive samples.

[0018] SP: Specificity, which represents the proportion of successfully predicted samples (the true label is negative) among all negative samples predicted by the model, and is used to measure the model's ability to predict negative samples.

[0019] Chemical toolkit: A software tool, usually a set of computer programs for processing chemical information and molecular data. These tools can be used for tasks such as molecular modeling, computational chemical property prediction, and molecular structure analysis. Some commonly used chemical toolkits include RDKit, Open Babel, and ChemAxon. These toolkits provide a wealth of functions and algorithms that can be used to generate molecular fingerprints and descriptors for use in areas such as molecular similarity calculations, compound screening, and quantitative structure-activity relationship (QSAR) modeling.

[0020] The idea of ​​implementing the present invention is to first extract the molecular structure graph feature set MGF, the molecular fingerprint feature set MFF and the molecular descriptor feature set MDF, then use feature screening to remove the features with smaller variances in the three feature sets, then perform normalization operations, and then use a ternary mixed-level fusion convolutional neural network to build an optimal prediction network, and finally use the optimal prediction network to predict the BBB permeability of the compound. The specific steps implemented by the present invention are as follows Figure 1 As shown, the model framework is as Figure 3 shown.

[0021] S1: For the acquired compound, extract the SMILES expression of the compound, and generate the molecular structure diagram of the compound using a chemical toolkit;

[0022] S2: Perform data preprocessing operations: extract the molecular graph feature set MGF (Molecular Graph Features) of the molecular structure graph, extract the molecular fingerprint feature set MFF (Molecular Fingerprint Features) and molecular descriptor feature set MDF (Molecular Descriptor Features) of the compound according to the SMILES expression; perform feature screening on MGF to obtain the principal component structure feature set P-MGF (Principal- Molecular Graph Features), perform feature screening on MFF to obtain the principal component fingerprint feature set P-MFF (Principal-Molecular Fingerprint Features), perform feature screening on MDF to obtain the principal component descriptor feature set P-MDF (Principal- Molecular Descriptor Features); normalize P-MGF, P-MFF, and P-MDF in turn to obtain the normalized structure feature set , Normalized fingerprint feature set , normalized descriptor feature set ;

[0023] S3: Construct a ternary mixed-level fusion convolutional neural network and design a ternary module processing , , The first feature is obtained , Second feature , the third characteristic ,Will , , Get fusion features through concatenation layer , use the fully connected layer and the output layer to process , predict the BBB label of the compound;

[0024] S4: Train the ternary mixed-level fusion convolutional neural network. , , Input into the ternary mixed-level fusion convolutional neural network, and adjust parameters to obtain the best prediction network;

[0025] S5: Use the optimal prediction network to predict the BBB permeability of the compound.

[0026] Furthermore, the data preprocessing operation described in step S2 is as follows: Figure 4 As shown, the specific steps are as follows:

[0027] S21: constructing an enhanced neighbor attention network AN-GAT to extract features from the molecular structure graph to obtain MGF; using MACCS keys (Molecular ACCess System Keys) to encode the structural information of the molecule to generate the MFF of the compound; using a chemical toolkit to generate the MDF of the compound, such as shape, size, charge, etc., wherein the chemical toolkit includes but is not limited to RDKit;

[0028] Furthermore, the specific steps of constructing AN-GAT are as follows:

[0029] S211: In the molecular structure diagram, for any adjacent atoms i and j, calculate the attention score between them ; By the eigenvectors of i and j and Performing a cube root transformation reduces the gap between larger and smaller values, thereby concentrating the MGF features in a smaller range, which helps improve the stability and generalization ability of the model; adding a constant term of 1 for translation ensures that even when the feature is zero, the formula will not be zero, thus avoiding zero errors; The calculation formula is as follows:

[0030] e ij =ReLU(W⋅[ ( h i 3 +1) 2 ⊕ ( h j 3 +1) 2 ])

[0031] in, is the attention weight matrix, represents the feature concatenation operation, is the activation function, and is the eigenvector of i and j, where the eigenvector is obtained by encoding the properties of atoms and chemical bonds into numerical form. Specifically, atomic properties such as atom type and atomic number are encoded using one-hot encoding, and chemical bond properties such as bond type, bond length, and bond polarity are encoded similarly. By combining these encoded property values, a corresponding eigenvector is generated for each atom and chemical bond;

[0032] S212: Regarding the Normalize and get the attention coefficient , where the normalization operation is performed by dividing each The exponential value of all neighboring atoms is divided by This is done by summing the index values. Perform a square operation, which helps to highlight important atomic features and makes the model pay more attention to the features of neighboring atoms that are similar to the current atomic features; adjust the size of the parameter z to affect the atomic feature update, and use the following formula to update the i feature vector :

[0033]

[0034] in, represents the updated feature vector of atom i, represents the eigenvector of atom j, z represents the parameter that adjusts the amplitude of the summation result, represents the set of atoms adjacent to atom i, express The exponential function of

[0035] S213: Repeat S211 and S212 until a preset number of iterations is reached.

[0036] S22: construct an adaptive feature screening method, perform feature screening operations on MGF, MFF and MDF in turn, remove features with small variance, and obtain the screened feature sets of P-MGF, P-MFF and P-MDF in turn; wherein, feature screening is performed on MGF to obtain P-MGF, feature screening is performed on MFF to obtain P-MFF, and feature screening is performed on MDF to obtain P-MDF;

[0037] Furthermore, the specific steps of the adaptive feature screening method are as follows:

[0038] S221: Calculate initial threshold , the calculation formula is as follows:

[0039]

[0040] in, is a preset adjustment factor used to determine The size of is the average of the variances of all features in MGF, MFF, or MDF;

[0041] S222: Use Determine the importance of each feature h in MGF, MFF, and MDF. When the variance of h is lower than , then h is considered unimportant or does not contain enough information, h is discarded, and the number of h discarded is counted ;

[0042] S223: According to Dynamically adjust the threshold if If it is greater than the warning value M, it means that the threshold is too high and needs to be gradually lowered; otherwise, it means that the threshold is too low and needs to be gradually increased. The specific formula is as follows:

[0043]

[0044] in, is the current threshold, is Based on the adjusted threshold, is an adjustment factor less than 1, It is an adjustment factor greater than 1, used to control the adjustment step of the threshold; the warning value M can be set according to actual conditions.

[0045] S224: Repeat S221 to S223 until the threshold is within a relatively appropriate interval, which can be set according to actual conditions.

[0046] S23: Construct a scale adjustment normalization method, calculate the normalized values ​​of P-MGF, P-MFF, and P-MDF respectively, and obtain , , , where P-MGF is normalized to obtain , normalize P-MFF to get , normalize the P-MDF to get ; The specific normalization formula is shown as follows:

[0047]

[0048] in, is a feature in P-MGF, P-MFF or P-MDF, is the normalized feature of P-MGF, P-MFF or P-MDF, and the value range of i is {1, 2, 3}. is the average value of the corresponding feature in P-MGF, P-MFF or P-MDF, is the standard deviation of the corresponding feature, is the minimum value of the corresponding feature in P-MGF, P-MFF or P-MDF, and is an adjustment factor related to BBB permeability. is a small offset used to avoid zero denominator.

[0049] Furthermore, the ternary mixed level fusion convolutional neural network in step S3 includes a convolutional layer, a pooling layer, a splicing layer, a fully connected layer and an output layer. The process of constructing the ternary mixed level fusion convolutional neural network is as follows: Figure 5 As shown, the specific steps are as follows:

[0050] S31: construct a ternary module, the ternary module comprising Processing module, Processing module, Processing module;

[0051] The The processing module refers to the use of convolutional layer, pooling layer, and fully connected layer in sequence to process , where the convolutional layer is first used to Convolution is performed, and the pooling layer is used to perform feature pooling, and then the first feature is obtained by calculating through the fully connected layer. , the calculation formula is as follows:

[0052]

[0053] in, and Represent the weight and bias of the fully connected layer respectively, It means flattening the feature map into a one-dimensional vector. is the model smoothing adjustment factor, which is used to reduce the noise impact between features. Express Perform convolution operation, Indicates that the pooling operation is performed on the feature map after convolution;

[0054] The The processing module uses convolutional layer, pooling layer and fully connected layer in sequence to process , where the convolutional layer is first used to Convolution is performed, and the pooling layer is used to perform feature pooling, and then the fully connected layer is used to calculate to obtain the second feature , the calculation formula is as follows:

[0055]

[0056] in, Express Perform convolution operation, Indicates that the pooling operation is performed on the feature map after convolution;

[0057] The The processing module refers to the use of the fully connected layer to process , and get the third feature , the calculation formula is as follows:

[0058]

[0059] S32: Through the splicing layer , , Splice to get fusion features :

[0060]

[0061] in, Represents feature concatenation operation;

[0062] S33: Use the fully connected layer and the output layer to process , in the output layer, The activation function is used for classification, and the calculation formula is as follows:

[0063]

[0064] in, and are the weights and biases of the fully connected layer, and are the weights and biases of the output layer, Representation Matrix The transpose of represents the rectified linear unit activation function, is the predicted BBB label.

[0065] Furthermore, the specific steps of training the ternary mixed-level fusion convolutional neural network described in step S4 are as follows:

[0066] S41: , , Input into the ternary mixed-level fusion convolutional neural network to perform BBB label prediction to obtain BBB prediction label data;

[0067] S42: Calculate the loss of the BBB predicted label data and the true label. The specific calculation formula is as follows:

[0068]

[0069] in, is the sample size, is the true label of the p-th sample, is the BBB prediction label of the p-th sample, is the loss function value, which indicates the difference between the predicted label and the true label. The smaller the value, the more accurate the model prediction.

[0070] S43: Adjust the weight parameters of the ternary mixed-level fusion convolutional neural network , the specific adjustment formula is as follows:

[0071]

[0072] in, is the learning rate, is the regularization coefficient, express about The partial derivative of The neural network is The weight parameters of the layer, is the adjusted The weight parameters of the layer;

[0073] S44: Repeat S41 to S43 until When the number of iterations converges or reaches a preset number, the ternary hybrid level fusion convolutional neural network with adjusted parameters is used as the optimal prediction network, and the number of iterations can be set according to actual conditions.

[0074] Furthermore, the specific steps of predicting the BBB permeability of the compound described in step S5 are as follows:

[0075] The compound to be tested is subjected to S1 and S2 to obtain a normalized structural feature set of the compound to be tested. , Normalized fingerprint feature set , normalized descriptor feature set ,Will , , The optimal prediction network is input to predict the BBB label of the compound to be tested.

[0076] The present invention also provides a compound blood-brain barrier permeability prediction device based on a ternary mixed-level fusion convolutional neural network, comprising:

[0077] The compound data acquisition module is configured to acquire compound BBB data;

[0078] A data preprocessing module, configured to preprocess compound BBB data;

[0079] Constructing a ternary mixed level fusion convolutional neural network module, configured to construct a ternary mixed level fusion convolutional neural network according to feature data;

[0080] A training ternary mixed level fusion convolutional neural network module is configured to input feature data into the ternary mixed level fusion convolutional neural network and complete parameter adjustment to obtain an optimal prediction network;

[0081] The compound BBB permeability prediction module is configured to use the optimal prediction network to perform compound BBB permeability prediction.

[0082] The compound data acquisition module includes: extracting the SMILES expression of the acquired compound and generating the molecular structure diagram of the compound using a chemical toolkit;

[0083] The data preprocessing module includes: generating data feature sets, completing feature screening and normalization preprocessing operations;

[0084] Constructing a ternary mixed-level fusion convolutional neural network module, including: designing a ternary module to process feature data, passing the output result through the concatenation layer to obtain fusion features, sequentially using the fully connected layer and the output layer to process the fusion features, and predicting the BBB label;

[0085] Training a ternary mixed level fusion convolutional neural network module, including: inputting the preprocessed feature data into the ternary mixed level fusion convolutional neural network, adjusting parameters, and finally obtaining an optimal prediction network;

[0086] The compound BBB permeability prediction module comprises: inputting the compound data to be predicted into the optimal prediction network, and predicting the BBB label corresponding to the compound to be predicted.

[0087] Compared with the prior art, the present invention has the following advantages:

[0088] First, this patent proposes to use the AN-GAT model to extract features from the molecular structure diagram of each molecule, effectively capturing the internal structural information of the molecule and improving the expressiveness and accuracy of the prediction model.

[0089] Second, this patent constructs a ternary module, which includes Processing module, Processing module, The processing module makes full use of the information of different features and improves the prediction ability and generalization performance of the model.

[0090] Third, the ternary mixed-level fusion convolutional neural network proposed in this patent utilizes the molecular structure graph features, molecular fingerprint features and molecular descriptor features, and effectively integrates the three features together through the ternary mixed-level fusion convolutional neural network, so that the prediction model can understand the characteristics of the compound from multiple angles, thereby improving the accuracy and stability of the prediction.

[0091] Fourth, the adaptive feature screening method proposed in this patent removes features with small variance, which can automatically adapt to the changes in different features of BBB, reduce the complexity of the model, reduce the possibility of overfitting, and improve the generalization ability of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0092] Figure 1 It is a flow chart for implementing the present invention.

[0093] Figure 2 It is a schematic diagram of the structure of the device of the present invention.

[0094] Figure 3 This is a model framework diagram of the present invention.

[0095] Figure 4 Figure 2 shows the data preprocessing process.

[0096] Figure 5 To build a ternary module flow chart.

[0097] Figure 6 A diagram of an example molecule structure.

[0098] Figure 7 Figure 2 is a diagram of an example ternary module processing process. DETAILED DESCRIPTION

[0099] Various exemplary embodiments of the present invention will now be described in detail. This detailed description should not be considered as limiting the present invention, but should be understood as a more detailed description of certain aspects, features, and embodiments of the present invention.

[0100] It should be understood that the terms described in the present invention are only for describing special embodiments and are not intended to limit the present invention. In addition, for the numerical range in the present invention, it should be understood that each intermediate value between the upper and lower limits of the scope is also specifically disclosed. Each smaller range between the intermediate value in any stated value or stated range and any other stated value or intermediate value in the described range is also included in the present invention. The upper and lower limits of these smaller ranges can be independently included or excluded in the scope.

[0101] Unless otherwise indicated, all technical and scientific terms used herein have the same meanings as those of ordinary skill in the art generally understood by the present invention. Although the present invention describes only preferred methods, any method similar or equivalent to that described herein may also be used in the implementation or testing of the present invention. All documents mentioned in this specification are incorporated by reference to disclose and describe the methods related to the documents. In the event of a conflict with any incorporated document, the content of this specification shall prevail.

[0102] It will be apparent to those skilled in the art that various modifications and variations may be made to the specific embodiments of the present invention description without departing from the scope or spirit of the present invention. Other embodiments derived from the present invention description will be apparent to those skilled in the art. The present application description and examples are exemplary only.

[0103] The words “include,” “including,” “have,” “contain,” etc. used in this article are open-ended terms, meaning including but not limited to.

[0104] It should be noted that, in the absence of conflict, the embodiments and features in the embodiments of the present application can be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0105] Example 1

[0106] The implementation flow chart of a compound blood-brain barrier permeability prediction method based on a ternary mixed-level fusion convolutional neural network is as follows: Figure 1 As shown, the following steps are included:

[0107] S1, for the acquired compound, extract the SMILES expression of the compound, and generate a molecular structure diagram of the compound using a chemical toolkit;

[0108] S2, performing data preprocessing operations: extracting the molecular structure feature set MGF of the molecular structure graph; extracting the molecular fingerprint feature set MFF and the molecular descriptor feature set MDF according to the SMILES expression of the compound; performing feature screening on MGF to obtain the principal component structure feature set P-MGF, performing feature screening on MFF to obtain the principal component fingerprint feature set P-MFF, performing feature screening on MDF to obtain the principal component descriptor feature set P-MDF; normalizing P-MGF, P-MFF, and P-MDF in turn to obtain the normalized structure feature set , Normalized fingerprint feature set , normalized descriptor feature set ;

[0109] S3, build a ternary mixed-level fusion convolutional neural network and design a ternary module processing , , The first feature is obtained , Second feature , the third characteristic ,Will , , Get fusion features through concatenation layer , use the fully connected layer and the output layer to process , predict the BBB label of the compound;

[0110] S4, training the ternary mixed-level fusion convolutional neural network, , , Input into the ternary mixed-level fusion convolutional neural network, and adjust parameters to obtain the best prediction network;

[0111] S5, using the optimal prediction network to predict the BBB permeability of the compound.

[0112] Preferably, the data preprocessing process in step S2 is as follows: Figure 4 As shown, specifically:

[0113] S21, constructing an enhanced neighbor attention network AN-GAT to extract features from the molecular structure graph to obtain MGF; using MACCS keys to encode the structural information of the molecule to generate the MFF of the compound; using a chemical toolkit to generate the MDF of the compound, such as shape, size, charge, etc., wherein the chemical toolkit includes but is not limited to RDKit;

[0114] S22, constructing an adaptive feature screening method, performing feature screening operations on MGF, MFF and MDF in turn, removing features with small variance, and obtaining the screened feature sets of P-MGF, P-MFF and P-MDF in turn; wherein, P-MGF is obtained by performing feature screening on MGF, P-MFF is obtained by performing feature screening on MFF, and P-MDF is obtained by performing feature screening on MDF;

[0115] S23, construct a scale adjustment normalization method, calculate the normalized values ​​of P-MGF, P-MFF, and P-MDF respectively, and obtain , , , where P-MGF is normalized to obtain , normalizing P-MFF to obtain , normalize the P-MDF to get ; The specific normalization formula is shown as follows:

[0116]

[0117] in, is a feature in P-MGF, P-MFF or P-MDF, is the normalized feature of P-MGF, P-MFF or P-MDF, and the value range of i is {1, 2, 3}. is the average value of the corresponding feature in P-MGF, P-MFF or P-MDF, is the standard deviation of the corresponding feature, is the minimum value of the corresponding feature in P-MGF, P-MFF or P-MDF, and is an adjustment factor related to BBB permeability. is a small offset used to avoid zero denominator.

[0118] Preferably, step S21 of constructing AN-GAT is specifically as follows:

[0119] S211: In the molecular structure diagram, for any adjacent atoms i and j, calculate the attention score between them ; By the eigenvectors of i and j and Performing a cube root transformation reduces the gap between larger and smaller values, thereby concentrating the MGF features in a smaller range, which helps improve the stability and generalization ability of the model; adding a constant term of 1 for translation ensures that even when the feature is zero, the formula will not be zero, thus avoiding zero errors; The calculation formula is as follows:

[0120] e ij =ReLU(W⋅[ ( h i 3 +1) 2 ⊕ ( h j 3 +1) 2 ])

[0121] in, is the attention weight matrix, represents the feature concatenation operation, is the activation function, and is the eigenvector of i and j, where the eigenvector is obtained by encoding the properties of atoms and chemical bonds into numerical form. Specifically, atomic properties such as atom type and atomic number are encoded using one-hot encoding, and chemical bond properties such as bond type, bond length, and bond polarity are encoded similarly. By combining these encoded property values, a corresponding eigenvector is generated for each atom and chemical bond;

[0122] S212: Regarding the Normalize and get the attention coefficient , where the normalization operation is performed by dividing each The exponential value of all neighboring atoms is divided by This is done by summing the index values. Perform a square operation, which helps to highlight important atomic features and makes the model pay more attention to the features of neighboring atoms that are similar to the current atomic features; adjust the size of the parameter z to affect the atomic feature update, and use the following formula to update the i feature vector :

[0123]

[0124] in, represents the updated feature vector of atom i, represents the eigenvector of atom j, z represents the parameter that adjusts the amplitude of the summation result, represents the set of atoms adjacent to atom i, express The exponential function of

[0125] S213: Repeat S211 and S212 until a preset number of iterations is reached.

[0126] by Figure 6 Take the molecular structure diagram shown in the figure as an example, which contains three atoms O, C and H. The eigenvector of atom O is h o =[1,2,3] , the eigenvector of atom C is h c =[2,3,1] , the eigenvector of atom H is h h =[3,1,2] ;

[0127] For the attention score between atoms O and C The specific calculation process is as follows:

[0128]

[0129] Assume that the calculation result is e oc = [ 4 , 5.75 , 7.39 , 5.75 , 7.39 , 4 ] , e oh = [ 4 , 5.75 , 7.39 , 7.39 , 4 , 5.75 ] and e h c = [ 5.75 , 7.39 , 4 , 7.39 , 4 , 5.75 ] , update the feature vector of each atom according to the calculated attention score. Taking atom O as an example, its updated feature vector , assuming z=2, the preset number of iterations is 3, the calculation is as follows:

[0130] h o ' = 1 2 exp e oc exp e oc + exp e oh 2 ⋅ 2,3,1 + exp e oh exp e oc + exp e oh 2 ⋅ [3,1,2]

[0131] Calculated h o ' = [ 1.1145 , 0.3965 , 0.7394 ] , repeat the above steps until the preset number of iterations 3 is reached. In the specific implementation process, the number of iterations can be set according to actual needs and is not limited to 3. Similarly, the updated feature vectors of atoms C and H can be calculated, and the above steps are repeated multiple times until the preset number of iterations is reached.

[0132] Preferably, the feature screening in step S22 is specifically as follows:

[0133] S221: Calculate initial threshold , the calculation formula is as follows:

[0134]

[0135] in, is a preset adjustment factor used to determine The size of is the average of the variances of all features in MGF, MFF, or MDF;

[0136] S222: Use Determine the importance of each feature h in MGF, MFF, and MDF. When the variance of h is lower than , then h is considered unimportant or does not contain enough information, h is discarded, and the number of h discarded is counted ;

[0137] S223: According to Dynamically adjust the threshold if If it is greater than the warning value M, it means that the threshold is too high and needs to be gradually lowered; otherwise, it means that the threshold is too low and needs to be gradually increased. The specific formula is as follows:

[0138]

[0139] in, is the current threshold, is Based on the adjusted threshold, is an adjustment factor less than 1, It is an adjustment factor greater than 1, used to control the adjustment step of the threshold; the warning value M can be set according to actual conditions.

[0140] S224: Repeat S221 to S223 until the threshold is within a relatively appropriate interval, which can be set according to actual conditions.

[0141] Taking MGF as an example, we calculate the average value of the feature variance in the entire data set MGF, and then multiply it by a preset adjustment factor k. Assuming the average value is 10 and k = 0.5, we can get the initial threshold: ; There are 100 features, of which 30 have variances less than the initial threshold of 5. Delete these features with smaller variances to obtain a preliminary feature set and calculate the number of deleted features ,if If it is greater than the preset warning value M=20, the threshold reduction operation is performed, α is 0.8, and the new threshold is: , repeat the above steps until the preset threshold range [2, 3] is reached. In the specific implementation process, the threshold range can be set according to actual needs and is not limited to selecting [2, 3].

[0142] Preferably, the process of constructing a ternary mixed-level fusion convolutional neural network in step S3 is as follows: Figure 5 As shown, specifically:

[0143] S31: construct a ternary module, the ternary module comprising Processing module, Processing module, Processing module;

[0144] The The processing module refers to the use of convolutional layer, pooling layer, and fully connected layer in sequence to process , where the convolutional layer is first used to Convolution is performed, and the pooling layer is used to perform feature pooling, and then the first feature is obtained by calculating through the fully connected layer. , the calculation formula is as follows:

[0145]

[0146] in, and Represent the weight and bias of the fully connected layer respectively, It means flattening the feature map into a one-dimensional vector. is the model smoothing adjustment factor, which is used to reduce the noise impact between features. Express Perform convolution operation, Indicates that the pooling operation is performed on the feature map after convolution;

[0147] by Figure 7 deal with For example, suppose The feature data is a 5x5 matrix, and a 3x3 convolution kernel is used to apply it to the input features. The result of the convolution operation is a single value, called a pixel of the feature map. The entire input feature map is subjected to a sliding window convolution operation to generate a feature map A of the corresponding size. A 2x2 pooling window is used to perform a sliding window pooling operation on the feature map, and the maximum value of each region is selected as the pixel value of the feature map after pooling to obtain feature map B. The pooled feature map is flattened into a one-dimensional vector and input into the fully connected layer to obtain the first feature map. .

[0148] The The processing module uses convolutional layer, pooling layer and fully connected layer in sequence to process , where the convolutional layer is first used to Convolution is performed, and the pooling layer is used to perform feature pooling, and then the fully connected layer is used to calculate to obtain the second feature , the calculation formula is as follows:

[0149]

[0150] in, Express Perform convolution operation, Indicates that the pooling operation is performed on the feature map after convolution;

[0151] The The processing module refers to the use of the fully connected layer to process , and get the third feature , the calculation formula is as follows:

[0152]

[0153] S32: Through the splicing layer , , Splice to get fusion features , expressed as:

[0154]

[0155] in, Represents a feature connection operation;

[0156] S33: Use the fully connected layer and the output layer to process , in the output layer, The activation function is used for classification to obtain the predicted BBB label, and its calculation formula is as follows:

[0157]

[0158] in, and are the weights and biases of the fully connected layer, and are the weights and biases of the output layer, Representation Matrix The transpose of represents the rectified linear unit activation function, is the predicted BBB label.

[0159] Preferably, the specific process of training the ternary mixed-level fusion convolutional neural network described in step S4 is:

[0160] S41: , , Input into the ternary mixed-level fusion convolutional neural network to perform BBB label prediction to obtain BBB prediction label data;

[0161] S42: Calculate the loss of the BBB predicted label data and the true label. The specific calculation formula is as follows:

[0162]

[0163] in, is the sample size, is the true label of the p-th sample, is the BBB prediction label of the p-th sample, is the loss function value, which indicates the difference between the predicted label and the true label. The smaller the value, the more accurate the model prediction.

[0164] S43: Adjust the weight parameters of the ternary mixed-level fusion convolutional neural network , the specific adjustment formula is as follows:

[0165]

[0166] in, is the learning rate, is the regularization coefficient, express about The partial derivative of The neural network The weight parameters of the layer, is the adjusted The weight parameters of the layer;

[0167] S44: Repeat S41 to S43 until When the number of iterations converges or reaches a preset number, the ternary hybrid level fusion convolutional neural network with adjusted parameters is used as the optimal prediction network, and the number of iterations can be set according to actual conditions.

[0168] The following specific example explains the adjustment process. Enter the following data: Sample size , the true label y=[1,0,1] , predicted label y'=[0.9,0.2,0.8] , calculate the loss of each sample, sample 1 is ; Sample 2 is ; Sample 3 is ; Calculate the total loss as , assuming the current weight ,gradient , learning rate , regularization coefficient The adjustment formula is: , check whether it converges or continues to iterate, assuming that the threshold is set to 10 -4 The preset number of iterations is 5. If the total loss is less than 10 -4 Or when it reaches 5 times, the algorithm can be judged to have converged. In the specific implementation process, the preset threshold and number of iterations can be set according to actual needs and are not limited to selecting 10. -4 and 5.

[0169] Preferably, the specific steps of predicting the BBB permeability of the compound described in step S5 are as follows:

[0170] The compound to be tested is subjected to S1 and S2 to obtain a normalized structural feature set of the compound to be tested. , Normalized fingerprint feature set , normalized descriptor feature set ,Will , , The optimal prediction network is input to predict the BBB label of the compound to be tested.

[0171] Example 2 Please refer to Figure 2 The present invention also provides an embodiment of a compound blood-brain barrier permeability prediction device based on a ternary mixed-level fusion convolutional neural network, comprising:

[0172] Data preprocessing module, used to preprocess compound BBB data;

[0173] Constructing a ternary mixed level fusion convolutional neural network module, configured to construct a ternary mixed level fusion convolutional neural network according to feature data;

[0174] A training ternary mixed level fusion convolutional neural network module is configured to input feature data into the ternary mixed level fusion convolutional neural network and complete parameter adjustment to obtain an optimal prediction network;

[0175] A compound BBB permeability prediction module is configured to use the optimal prediction network to perform compound BBB permeability prediction;

[0176] In a preferred embodiment, the data preprocessing module includes: generating a feature set of data, completing feature screening and normalization preprocessing operations;

[0177] The preferred implementation method is to construct a ternary mixed-level fusion convolutional neural network module, including: designing a ternary module to process feature data, passing the output result through a concatenation layer to obtain fusion features, sequentially using a fully connected layer and an output layer to process the fusion features, and predicting a BBB label;

[0178] A preferred implementation method, training a ternary mixed level fusion convolutional neural network module, comprises: inputting the preprocessed feature data into the ternary mixed level fusion convolutional neural network, adjusting parameters, and finally obtaining an optimal prediction network;

[0179] In a preferred embodiment, the compound BBB permeability prediction module comprises: inputting the compound data to be predicted into the optimal prediction network, and predicting the BBB label corresponding to the compound to be predicted.

[0180] The compound blood-brain barrier permeability prediction device based on the ternary mixed-level fusion convolutional neural network provided in this embodiment can reduce the training complexity of the model and improve its prediction accuracy for a few categories by preprocessing the data set; at the same time, considering that the traditional machine learning prediction is too single, the ternary mixed-level fusion convolutional neural network is used to increase the diversity of the model, make full use of the information of different features, and improve the prediction ability and generalization performance of the model; this model effectively integrates the molecular structure map features, molecular fingerprint features and molecular descriptor features, so that the prediction model can understand the characteristics of the compound from multiple angles, and improves the accuracy and stability of the prediction; this model can adapt to large data sets, and the prediction performance gap for different categories of compounds is small, which helps to assist compound researchers in designing suitable compounds, improve the fault tolerance of research and development, and reduce cost consumption.

[0181] Those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0182] In the embodiments provided in the present application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are only schematic. For example, the division of the units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be an indirect coupling or communication connection through some interfaces, devices or units, which can be electrical, mechanical or other forms.

[0183] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0184] In addition, each functional unit in each embodiment of the present invention may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0185] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk and other media that can store program codes.

[0186] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting compound blood-brain barrier permeability based on a ternary mixed-level fusion convolutional neural network, characterized in that The method uses molecular structure graph features, molecular fingerprint features and molecular descriptor features, and uses a ternary mixed-level fusion convolutional neural network to predict the permeability of compounds through the blood-brain barrier. The method includes the following steps: S1. For the acquired compound, extract the SMILES expression of the compound, and generate the molecular structure diagram of the compound using a chemical toolkit; S2, perform data preprocessing operations: extract the molecular structure feature set MGF of the molecular structure graph; extract the molecular fingerprint feature set MFF and the molecular descriptor feature set MDF according to the SMILES expression of the compound; perform feature screening on MGF to obtain the principal component structure feature set P-MGF, perform feature screening on MFF to obtain the principal component fingerprint feature set P-MFF, perform feature screening on MDF to obtain the principal component descriptor feature set P-MDF; normalize P-MGF, P-MFF, and P-MDF in turn to obtain the normalized structure feature set , Normalized fingerprint feature set , normalized descriptor feature set ; The specific steps of the data preprocessing operation include: S21, constructing an enhanced neighbor attention network AN-GAT to extract features from the molecular structure graph to obtain MGF; using a chemical toolkit to generate MFF and MDF of the compound; The specific contents of constructing AN-GAT include: S211. In the molecular structure diagram, for any adjacent atoms i and j, calculate the attention score between them. , The calculation formula is as follows: e ij =ReLU(W⋅[ ( h i 3 +1) 2 ⊕ ( h j 3 +1) 2 ]) in, is the attention weight matrix, represents the feature concatenation operation, is the activation function, and are the eigenvectors of i and j, where the eigenvectors are obtained by encoding the properties of atoms and chemical bonds into numerical form; S212. Update the eigenvector of i using the following formula: in, represents the updated feature vector of atom i, represents the eigenvector of atom j, z represents the parameter that adjusts the amplitude of the summation result, represents the set of atoms adjacent to atom i, express The exponential function of S213, repeatedly executing S211 and S212 until a preset number of iterations is reached; S22, performing feature screening operations on MGF, MFF and MDF in turn, removing features with smaller variances, and obtaining the screened feature sets of P-MGF, P-MFF and P-MDF in turn; S23, calculate the normalized values ​​of P-MGF, P-MFF, and P-MDF respectively, and obtain , , , the specific normalization formula is shown as follows: in, is a feature in P-MGF, P-MFF or P-MDF, is the normalized feature of P-MGF, P-MFF or P-MDF, and the value range of i is {1, 2, 3}. is the average value of the corresponding feature in P-MGF, P-MFF or P-MDF, is the standard deviation of the corresponding feature, is the minimum value of the corresponding feature in P-MGF, P-MFF or P-MDF, and is an adjustment factor related to BBB permeability. is a small offset used to avoid zero denominator; S3. Build a ternary mixed-level fusion convolutional neural network and design a ternary module for processing , , The first feature is obtained , Second feature , the third characteristic ,Will , , Get fusion features through concatenation layer , use the fully connected layer and the output layer to process , predict the BBB label of the compound; The ternary mixed level fusion convolutional neural network includes a convolutional layer, a pooling layer, a splicing layer, a fully connected layer and an output layer. The specific steps of constructing the ternary mixed level fusion convolutional neural network include: S31, constructing a ternary module, the ternary module comprising Processing module, Processing module, Processing module; The The processing module refers to the use of convolutional layer, pooling layer, and fully connected layer in sequence to process , where the convolutional layer is first used to Convolution is performed, and the pooling layer is used to perform feature pooling, and then the first feature is obtained by calculating through the fully connected layer. , the calculation formula is as follows: in, and denote the weight and bias of the fully connected layer respectively, It means flattening the feature map into a one-dimensional vector. is the model smoothing adjustment factor, which is used to reduce the noise impact between features. Express Perform convolution operation, Indicates that the pooling operation is performed on the feature map after convolution; The The processing module uses convolutional layer, pooling layer and fully connected layer in sequence to process , where the convolutional layer is first used to Convolution is performed, and the pooling layer is used to perform feature pooling, and then the fully connected layer is used to calculate to obtain the second feature , the calculation formula is as follows: in, Express Perform convolution operation, Indicates that the pooling operation is performed on the feature map after convolution; The The processing module refers to the use of the fully connected layer to process , and get the third feature , the calculation formula is as follows: S32, through the splicing layer , , Splice to get fusion features : in, Represents feature concatenation operation; S33, use the fully connected layer and output layer to process , in the output layer, The activation function is used for classification, and the calculation formula is as follows: in, and are the weights and biases of the fully connected layer, and are the weights and biases of the output layer, Representation Matrix The transpose of represents the rectified linear unit activation function, is the predicted BBB label; S4, training the ternary mixed-level fusion convolutional neural network, , , Input into the ternary mixed-level fusion convolutional neural network, and adjust parameters to obtain the best prediction network; S5. Subject the compound to be tested to S1 and S2 to obtain a normalized structural feature set of the compound to be tested. , Normalized fingerprint feature set , normalized descriptor feature set ,Will , , The optimal prediction network is input to predict the BBB label of the compound to be tested.

2. The method for predicting compound blood-brain barrier permeability based on ternary mixed-level fusion convolutional neural network according to claim 1, characterized in that: In the above-mentioned S22, the specific steps of the feature screening method include: S221. Calculate initial threshold , the calculation formula is as follows: in, is a preset adjustment factor used to determine The size of is the average of the variances of all features in MGF, MFF, or MDF; S222, Use Determine the importance of each feature h in MGF, MFF and MDF. When the variance of h is lower than , then h is considered unimportant or does not contain enough information, h is discarded, and the number of h discarded is counted ; S223, according to Dynamically adjust the threshold if If it is greater than the warning value M, it means that the threshold is too high and needs to be gradually lowered; otherwise, it means that the threshold is too low and needs to be gradually increased. The specific formula is as follows: in, is the current threshold, is Based on the adjusted threshold, is an adjustment factor less than 1, is an adjustment factor greater than 1, used to control the adjustment step size of the threshold; S224. Repeat S221 to S223 until the threshold is within a suitable range.

3. The method for predicting compound blood-brain barrier permeability based on ternary mixed-level fusion convolutional neural network according to claim 1, characterized in that: In the above S4, the specific steps of training the ternary mixed-level fusion convolutional neural network include: S41, will , , Input into the ternary mixed-level fusion convolutional neural network to perform BBB label prediction to obtain BBB prediction label data; S42, performing loss calculation on the BBB predicted label data and the true label, the calculation formula is as follows: in, is the sample size, is the true label of the p-th sample, is the BBB prediction label of the p-th sample, is the loss function value, which indicates the difference between the predicted label and the true label. The smaller the value, the more accurate the model prediction. S43, adjusting the weight parameters of the ternary mixed-level fusion convolutional neural network , the specific adjustment formula is as follows: in, is the learning rate, is the regularization coefficient, express about The partial derivative of The neural network is The weight parameters of the layer, is the adjusted The weight parameters of the layer; S44, repeat S41 to S43 until When convergence or reaching a preset number of iterations, the ternary mixed-level fusion convolutional neural network with adjusted parameters is used as the optimal prediction network.

4. A compound blood-brain barrier permeability prediction device based on a ternary mixed-level fusion convolutional neural network, used to run the compound blood-brain barrier permeability prediction method based on a ternary mixed-level fusion convolutional neural network according to any one of claims 1 to 3, characterized in that: It includes a compound data acquisition module, a data preprocessing module, a ternary mixed level fusion convolutional neural network construction module, a ternary mixed level fusion convolutional neural network training module and a compound BBB permeability prediction module; The compound data acquisition module is used to: extract the SMILES expression of the acquired compound and generate the molecular structure diagram of the compound using a chemical toolkit; The data preprocessing module is used to: extract the molecular structure feature set MGF of the molecular structure graph, extract the molecular fingerprint feature set MFF and the molecular descriptor feature set MDF according to the SMILES expression of the compound; perform feature screening on MGF to obtain the principal component structure feature set P-MGF, perform feature screening on MFF to obtain the principal component fingerprint feature set P-MFF, perform feature screening on MDF to obtain the principal component descriptor feature set P-MDF; normalize P-MGF, P-MFF, and P-MDF in turn to obtain the normalized structure feature set , Normalized fingerprint feature set , normalized descriptor feature set The ternary mixed level fusion convolutional neural network module is used to construct a ternary mixed level fusion convolutional neural network and design a ternary module processing , , , and get the first feature , Second feature , the third characteristic ,Will , , Get fusion features through concatenation layer , use the fully connected layer and the output layer to process , predicting the BBB label of the compound; the training ternary mixed level fusion convolutional neural network module is used to: train the ternary mixed level fusion convolutional neural network, , , The input is into the ternary mixed-level fusion convolutional neural network, and the parameters are adjusted to obtain the best prediction network; the compound BBB permeability prediction module is used to: use the best prediction network to predict the sample of the compound to be tested.

Images