A method for predicting the biological activity of ERα antagonists based on the RegNet_1d model and the integrated gradient method

Abstract

The present invention discloses a method for predicting the biological activity of an ERα antagonist based on a RegNet_1d model and an integral gradient method, which first builds a RegNet_1d deep learning model, then uses m molecular descriptor information of the ERα antagonist to predict its biological activity on ERα, and optimizes the data structure based on the integral gradient method, attributes the prediction results to the input features, obtains the correlation distribution of the effects of the m variables on the biological activity, and thereby screens suitable molecular descriptor variables for reconstructing the data structure to train the prediction model. The present invention uses a deep learning network to automatically learn the representation model, improves the prediction efficiency, and at the same time, the prediction accuracy of the optimized model is slightly reduced, but the amount of data required for measurement is greatly reduced, saving the time and cost of drug development.

Description

Technical Field

[0001] The present invention relates to the field of ERα antagonist activity prediction and virtual screening thereof, in particular to a method for predicting the biological activity of an ERα antagonist and an optimization method thereof based on a RegNet_1d model and an integrated gradient method. Background Art

[0002] At present, in drug development, in order to save time and cost, the method of establishing a compound activity prediction model is usually used to screen potential active compounds. The specific approach is: for a certain target, a series of compounds acting on the target and their biological activity data are collected, and then a series of molecular structure descriptors are used as independent variables, and the biological activity value of the compound is used as the dependent variable to construct a quantitative structure-activity relationship (Quantitative Structure-Activity Relationship, QSAR) model of the compound, and then the model is used to predict new compound molecules with better biological activity, or to guide the structural optimization of existing active compounds. The existing technology is more likely to use traditional modeling methods for prediction, and the accuracy rate is not high; and most of the modeling methods are based on experimental paradigms and rely on manual design, while all experimental tests are costly and the cycle is too long, which is not conducive to drug development iteration. Summary of the invention

[0003] In order to overcome the shortcomings of the prior art, the present invention proposes a method for predicting the biological activity of ERα antagonists based on the RegNet_1d model and the integrated gradient method, in order to reduce the amount of measured data while improving the prediction efficiency, thereby saving time and cost in drug development.

[0004] In order to achieve the above-mentioned object of the invention, the present invention adopts the following technical scheme:

[0005] The characteristics of the method for predicting the biological activity of ERα antagonists based on the RegNet_1d model and the integrated gradient method of the present invention include the following steps:

[0006] S1: Collect the biological activity data of a series of antagonist compounds acting on the target ERα and m molecular structure descriptors;

[0007] The m molecular descriptors of the target ERα antagonist compound are used as m independent variables. After data standardization operations are performed on the m independent variables, the characteristic data obtained is recorded as: X = [x 1 ,x 2 ,…,x i ,…,x m ], x i represents the value of the i-th molecular descriptor; the negative logarithm pIC of the biological activity value data of the antagonist compound against the target ERα 50is the dependent variable, denoted as Y; the feature data X and the dependent variable Y are combined into a data set and divided into a training set D train and validation set D val ;

[0008] S2: Build the RegNet_1d prediction model consisting of input module, core module and output module:

[0009] S2.1: The input module includes a convolution layer Conv1d, a pooling layer Pool1d and a batch normalization layer BatchNorm1d in sequence; the number of channels of the input data is set to m;

[0010] S2.2: The core module is based on the RegNet network architecture and consists of a first module block 1 , b second module block 2 , c third module blocks 3 The first module block 1 Contains one-dimensional convolution layer Conv1d, one-dimensional batch normalization layer BatchNorm1d; the second module block 2 Contains one-dimensional convolution layer Conv1d, one-dimensional batch normalization layer BatchNorm1d, activation function layer ReLU; the third module block 3 Including Conv1d, BatchNorm1d, ReLU, and bottleneck layer Bottleneck;

[0011] S2.3: The output module includes an adaptive pooling layer AdaptiveAvgPool1d and a fully connected layer;

[0012] S3: Training and selecting models:

[0013] S3.1: Initialize the learning rate to lr, the current number of iterations to epoch, and the optimal determination coefficient to R2 max , the learning rate adjustment iteration value t = 0, set the adjustment cycle threshold to t max ;

[0014] S3.2: In the epoch iteration, the training set D train The size of each batch is bs, which is input into the RegNet_1d neural network model for training. The gradients of the m channels are solved after calculating the MSE loss L, and the weight parameters in the gradient are optimized using the Adam optimizer based on the learning rate lr, so as to obtain the model trained for the epoch.

[0015] S3.3: After the epoch iteration training, on the validation set D valThe model trained in the epoch is verified with a batch size of bs, and the determination coefficient of the model trained in the current epoch is calculated. As an evaluation indicator, if Then Assign to R 2max , and save the parameters of the model trained in the current epoch. If Then after assigning t+1 to t, we can judge that t=t max Is it true? If so, adjust the learning rate lr to 0.5lr; otherwise, keep the learning rate lr;

[0016] S3.4: After assigning epoch+1 to epoch, return to step S3.2 until the determination coefficient no longer increases, stop training and use the last trained model as the optimal RegNet_1d neural network model F;

[0017] S4: Based on the optimal RegNet_1d neural network model F, the Integrated Gradients algorithm is used to calculate the correlation distribution vector A=[A 1 ,A 2 ,…,A i ,…,A m ], where A i is the correlation value between the ith molecular descriptor and the dependent variable Y;

[0018] S5: Eliminate the independent variables with correlation of 0 in the correlation distribution vector A, obtain n molecular descriptors, and thus construct the optimized feature data X′=[x′ 1 ,x′ 2 ,…,x′ j ,…,x′ n ],x′ j represents the value of the optimized j-th molecular descriptor. The optimized feature data X′ and dependent variable Y are combined into a new data set and re-divided into a training set D′ train and validation set D′ val ;

[0019] S6: Set the number of input channels of the RegNet_1d prediction model to n;

[0020] S7: In the new training set D′ train and validation set D′ val According to the process of step S3, the optimal prediction model F′ after data structure optimization is obtained;

[0021] S8: Input the biological activity data of the target ERα antagonist compound to be tested into the optimal prediction model F′ and output the prediction result.

[0022] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0023] 1. The present invention uses the biological activity of different compounds on ERα as the dependent variable and the m molecular descriptor information of the corresponding compounds as the independent variable to create a data set, and builds a RegNet_1d deep learning model to predict the biological activity of ERα antagonists; and improves and designs the RegNet_1d model for the biological activity prediction task of ERα antagonists, which can achieve a prediction accuracy of 71%; at the same time, it has the characteristics of automatic learning representation of deep learning networks, and the prediction efficiency is higher;

[0024] 2. While designing the prediction model, the present invention also optimizes its data structure and uses the idea of ​​integrated gradients. By characterizing the interpretability of the deep learning model, the correlation between the independent variable characteristics and the dependent variable in the model is calculated, thereby selecting important molecular structure descriptors and optimizing the modeling prediction process, thereby greatly saving the time and cost of drug research and development. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 It is a flow chart of the biological activity prediction and optimization method of ERα antagonists based on the RegNet_1d model and the integrated gradient method of the present invention.

[0026] Figure 2 It is a schematic diagram of the original data structure.

[0027] Figure 3 This is a schematic diagram of the distribution of the top 20 molecular descriptors with the most significant impact on biological activity and their correlation with biological activity. DETAILED DESCRIPTION

[0028] In this embodiment, a method for predicting the biological activity of an ERα antagonist based on a RegNet_1d model and an integrated gradient method is to use the RegNet_1d model to complete the biological activity prediction task of the ERα antagonist, and use the integrated gradient method to calculate the correlation distribution between the independent variable molecular descriptor and the dependent variable biological activity, and after sorting, filter out independent variables with no correlation or correlation close to 0, thereby simplifying the data structure, and retrain the prediction model to complete the optimization modeling of the biological activity prediction of the ERα antagonist, which has the advantages of high accuracy, saving time and experimental costs. Specifically, Figure 1 As shown, the steps are as follows:

[0029] S1: Collect the biological activity data of a series of antagonist compounds acting on the target ERα and m molecular structure descriptors;

[0030] The m molecular descriptors of the target Erα antagonist compound are used as m independent variables. After data standardization operations are performed on the m independent variables, the characteristic data obtained is recorded as: X = [x 1 ,x 2 ,…,x i ,…,x m ], x i represents the value of the i-th molecular descriptor; the negative logarithm pIC of the biological activity value data of the antagonist compound against the target ERα 50 is the dependent variable, denoted as Y; the feature data X and the dependent variable Y are combined into a data set and divided into a training set D train and validation set D val ; In this embodiment, m=729, that is, 729 molecular descriptors of ERα antagonist compounds are collected, including physicochemical properties (such as molecular weight, LogP, etc.), topological structure characteristics (such as the number of hydrogen bond donors, the number of hydrogen bond acceptors, etc.), etc.; If the variance of a certain feature is much larger than the variance of other features, then it will occupy a dominant position in the algorithm learning, resulting in the learner being unable to learn other features as expected, which will cause the final model to converge slowly or even not converge. Therefore, the original data is first subjected to z-score standardization, and then the data structure of each compound in sequence is {X: [1974, 729, 1], Y: [1974, 1]}; the training set and the validation set are generated by dividing the data in a ratio of 9:1;

[0031] S2: Build the RegNet_1d prediction model consisting of input module, core module and output module:

[0032] S2.1: The input module includes a convolution layer Conv1d, a pooling layer Pool1d and a batch normalization layer BatchNorm1d in sequence; the number of channels of the input data is set to m; in this embodiment, m=729;

[0033] S2.2: The core module is based on the RegNet network architecture and consists of a first module block 1 , b second module block 2 , c third module blocks 3 The first module block 1 Contains one-dimensional convolution layer Conv1d, one-dimensional batch normalization layer BatchNorm1d; the second module block 2 Contains one-dimensional convolution layer Conv1d, one-dimensional batch normalization layer BatchNorm1d, activation function layer ReLU; the third module block 3Including Conv1d, BatchNorm1d, ReLU, and bottleneck layer Bottleneck; in this embodiment, a=4, b=33, c=16;

[0034] S2.3: The output module includes an adaptive pooling layer AdaptiveAvgPool1d and a fully connected layer; in this example, the input data is processed by the core architecture and sent to the adaptive pooling layer and the fully connected layer, and the result Y′ is obtained as the prediction output;

[0035] S3: Training and selecting models:

[0036] S3.1: Initialize the learning rate to lr, the current number of iterations to epoch, and the optimal determination coefficient to R2 max , the learning rate adjustment iteration value t = 0, set the adjustment cycle threshold to t max ; In this embodiment, the initial learning rate lr = 0.01, epoch = 1, R2 max =0,t max =2, that is, when the cumulative value of R2 does not increase twice, the learning rate is reduced for adjustment, and t is reset to 0;

[0037] S3.2: In the epoch iteration, the training set D train According to the size of each batch bs, it is input into the RegNet_1d neural network model for training, and the gradient of m channels is solved after calculating the MSE loss L, and then the weight parameters in the gradient are optimized by the Adam optimizer based on the learning rate lr, so as to obtain the model trained for the epoch; in this embodiment, bs=32 is set, and the schematic diagram of the original data structure of the model input is as follows Figure 2 As shown, the data is normalized and then input into the model for training;

[0038] S3.3: After the epoch iteration training, on the validation set D val The model trained in the epoch is verified with a batch size of bs, and the determination coefficient of the model trained in the current epoch is calculated. As an evaluation indicator, if Then Assign to R 2max , and save the parameters of the model trained in the current epoch. If Then after assigning t+1 to t, we can judge that t=t max Is it true? If true, adjust the learning rate lr to 0.5lr; otherwise, keep the learning rate lr; In this embodiment, the determination coefficient R 2It reflects the percentage of Y fluctuation that can be described by X fluctuation. After linear regression of the model, it is used to evaluate the goodness of fit of the regression model coefficients. 2 The calculation formula is:

[0039]

[0040] S3.4: After assigning epoch+1 to epoch, return to step S3.2 until the determination coefficient no longer increases, stop training and use the last trained model as the optimal RegNet_1d neural network model F;

[0041] To further illustrate the advantages of this method, other deep learning models were built for training comparison. The results are shown in Table 1, where Loss represents the loss value, and the smaller the value, the better the prediction effect; R 2 Equivalent to accuracy, the value range is 0 to 1, and the larger the value, the better the prediction effect.

[0042] Table 1 Comparison data of the present invention and other deep learning model methods

[0043]

[0044] S4: Based on the optimal RegNet_1d neural network model F, the Integrated Gradients algorithm is used to calculate the correlation distribution vector A=[A 1 ,A 2 ,…,A i ,…,A m ], where A i is the correlation value between the ith molecular descriptor and the dependent variable Y; in this embodiment, the correlation distribution values ​​of 729 molecular descriptors and biological activities are output, among which the top 20 molecular descriptors with the most significant impact on biological activity and their correlation values ​​with biological activity are as follows: Figure 3 As shown, negative values ​​represent negative correlation, positive values ​​represent positive correlation, 0 represents no correlation, and the larger the absolute value, the stronger the correlation;

[0045] S5: Eliminate the independent variables with correlation of 0 in the correlation distribution vector A, obtain n molecular descriptors, and thus construct the optimized feature data X′=[x′ 1 ,x′ 2 ,…,x′ j ,…,x′ n ],x′ j represents the value of the optimized j-th molecular descriptor. The optimized feature data X′ and dependent variable Y are combined into a new data set and re-divided into a training set D′ train and validation set D′ val; In this embodiment, in order to compare the effect after optimization, n=20 is taken, that is, only the top 20 molecular descriptor data with the most significant impact on biological activity are retained as X′ to construct a data set;

[0046] S6: Set the number of input channels of the RegNet_1d prediction model to n; in this embodiment, n=20;

[0047] S7: In the new training set D′ train and validation set D′ val According to the process of step S3, the optimal prediction model F′ after data structure optimization is obtained;

[0048] S8: Inputting the biological activity data of the target ERα antagonist compound to be tested into the optimal prediction model F′ and outputting the prediction results; the prediction effect comparison of the optimized model in this embodiment is shown in Table 2;

[0049] Table 2 Comparison data of RegNet_1d model method before and after data structure optimization

[0050]

[0051] As can be seen from Table 2, although the prediction effect of the optimized prediction model on the validation set is slightly reduced, only 20 molecular descriptors are used here for demonstration, which greatly reduces the cost required for collecting 729 features. It can be seen that the present invention can determine the feature data used according to needs and actual conditions, thereby achieving the requirements of high accuracy and low cost at the same time.

Claims

1. A method for predicting the biological activity of ERα antagonists based on the RegNet_1d model and the integrated gradient method, characterized in that: The following steps are involved: S1: Collect the biological activity data of a series of antagonist compounds acting on the target ERα and m molecular structure descriptors; The m molecular descriptors of the target ERα antagonist compound are used as m independent variables. After data standardization operations are performed on the m independent variables, the characteristic data obtained are recorded as: X = [x1, x2, ..., x i ,…,x m ], x i represents the value of the i-th molecular descriptor; the negative logarithm pIC of the biological activity value data of the antagonist compound against the target ERα 50 is the dependent variable, denoted as Y; the feature data X and the dependent variable Y are combined into a data set and divided into a training set D train and validation set D val ; S2: Build the RegNet_1d prediction model consisting of input module, core module and output module: S2.1: The input module includes a convolution layer Conv1d, a pooling layer Pool1d and a batch normalization layer BatchNorm1d in sequence; the number of channels of the input data is set to m; S2.2: The core module is based on the RegNet network architecture and is composed of a first module block1, b second modules block2, and c third modules block3. The first module block1 includes a one-dimensional convolution layer Conv1d and a one-dimensional batch normalization layer BatchNorm1d; the second module block2 includes a one-dimensional convolution layer Conv1d, a one-dimensional batch normalization layer BatchNorm1d, and an activation function layer ReLU; the third module block3 includes Conv1d, BatchNorm1d, ReLU, and a bottleneck layer Bottleneck; S2.3: The output module includes an adaptive pooling layer AdaptiveAvgPool1d and a fully connected layer; S3: Training and selecting models: S3.1: Initialize the learning rate to lr, the current number of iterations to epoch, and the optimal determination coefficient to R2 max , the learning rate adjustment iteration value t = 0, set the adjustment cycle threshold to t max ; S3.2: In the epoch iteration, the training set D train The size of each batch is bs, which is input into the RegNet_1d neural network model for training. The gradients of the m channels are solved after calculating the MSE loss L, and the weight parameters in the gradient are optimized using the Adam optimizer based on the learning rate lr, so as to obtain the model trained for the epoch. S3.3: After the epoch iteration training, on the validation set D val The model trained in the epoch is verified with a batch size of bs, and the determination coefficient of the model trained in the current epoch is calculated. As an evaluation indicator, if Then Assign to R 2max , and save the parameters of the model trained in the current epoch. If Then after assigning t+1 to t, we can judge that t=t max Is it true? If so, adjust the learning rate lr to 0.5lr; otherwise, keep the learning rate lr; S3.4: After assigning epoch+1 to epoch, return to step S3.2 until the determination coefficient no longer increases, stop training and use the last trained model as the optimal RegNet_1d neural network model F; S4: Based on the optimal RegNet_1d neural network model F, the Integrated Gradients algorithm is used to calculate the correlation distribution vector A = [A1, A2, ..., A i ,…,A m ], where A i is the correlation value between the ith molecular descriptor and the dependent variable Y; S5: Eliminate the independent variables with correlation of 0 in the correlation distribution vector A, obtain n molecular descriptors, and thus construct the optimized feature data X′=[x′1,x′2,…,x′ j ,…,x′ n ],x′ j represents the value of the optimized j-th molecular descriptor. The optimized feature data X′ and dependent variable Y are combined into a new data set and re-divided into a training set D′ train and validation set D′ val ; S6: Set the number of input channels of the RegNet_1d prediction model to n; S7: In the new training set D′ train and validation set D′ val According to the process of step S3, the optimal prediction model F′ after data structure optimization is obtained; S8: Input the biological activity data of the target ERα antagonist compound to be tested into the optimal prediction model F′ and output the prediction result.

Images