A method using non-linear dynamic analysis to assist neural networks for evaluating drug-induced cardiotoxicity

Abstract

The present invention belongs to the technical field of drug toxicity analysis, and discloses a method for using non-linear dynamic analysis assisted neural network to evaluate the cardiotoxicity caused by drugs. This method uses a neural network assisted by non-linear dynamic analysis to extract non-linear dynamic features from the mechanical pulsation signals of cardiomyocytes, and then analyzes through a neural network model to evaluate the cardiotoxicity caused by drugs, achieving a high accuracy rate for drug classification and concentration prediction. Moreover, this neural network model can avoid segmenting the pulsation signal into multiple cycles for analysis, directly process the pulsation time series with a fixed duration and extract dynamic features. While the method of the present invention has extremely high accuracy in identifying the cardiotoxicity caused by drugs, it also has excellent generalization performance, and has achieved a high accuracy rate in identifying the cardiotoxicity of drugs not in the neural network training library. The present invention can screen the cardiotoxicity of newly developed drugs, which is helpful for drug development.

Description

Technical Field

[0001] The present invention belongs to the technical field of drug toxicity analysis, and relates to a method for evaluating drug-induced cardiotoxicity by using non-linear dynamic analysis to assist neural networks. Background Art

[0002] The incidence and severity of drug-induced cardiovascular toxicity are the highest among drug adverse reactions. The application of some drugs in the treatment of cardiovascular diseases will induce serious arrhythmias due to their significant cardiotoxic side effects. For example, astemizole, designed in 1977 to relieve allergic symptoms, was withdrawn from the market in 1999 because it had the potential to block the human hERG potassium channel and cause heart failure. Therefore, in drug development, in order to avoid huge losses of life and money caused by drug adverse reactions, cardiotoxicity assessment is necessary.

[0003] Therefore, various techniques have been developed for the preclinical assessment of induced cardiotoxicity, including in vivo and in vitro drug assessment models. Due to the high cost and low throughput of in vivo assessment models, which limit their application in drug assessment, in vitro cell-based techniques, including interdigitated electrodes and microelectrode arrays, have been developed to achieve high-throughput drug screening and real-time monitoring of cell activities after drug treatment. To process the data collected from in vitro high-throughput experiments and evaluate drug-induced cardiotoxicity, researchers have adopted many artificial intelligence algorithms, including linear regression, logistic regression, neural networks, etc. Among them, neural networks have shown relatively superior accuracy compared to other methods.

[0004] In order to extract features from the mechanical beating signals of cardiomyocytes recorded by interdigitated electrodes for further artificial neural network analysis for drug cardiotoxicity assessment, the traditional method is to segment the signals into beat cycles and find feature points in the beat cycles. However, for different signal shapes, the segmentation of signals and the positioning of feature points require the design of specific algorithms, which may reduce the efficiency of drug cardiotoxicity assessment. And because biological signals are non-linear and dynamic, extracting time-domain feature localization through signal segmentation and feature points cannot extract all the information contained in biological signals. Summary of the Invention

[0005] The purpose of the present invention is to provide a method for evaluating drug-induced cardiotoxicity by using non-linear dynamic analysis to assist neural networks to solve the above technical problems.

[0006] To solve the above technical problems, the specific technical solution of a method for evaluating drug-induced cardiotoxicity by using non-linear dynamic analysis to assist neural networks of the present invention is as follows:

[0007] A method for evaluating drug-induced cardiotoxicity using non-linear dynamic analysis assisted neural network, comprising the following steps:

[0008] Step 1: Record the beating signals of cardiomyocytes using an electrocardiogram system with interdigitated electrodes;

[0009] Step 2: Denoise the original beating signals of cardiomyocytes through a filter and calculate 15 non-linear dynamic parameters for each time series;

[0010] Step 3: Input the non-linear parameters into a neural network for training, classify the drugs and predict the drug concentration level to indicate the toxicity intensity of cardiotoxic drugs.

[0011] Furthermore, the said Step 1 includes the following specific steps:

[0012] Measure the beating signals of cardiomyocytes: Measure the beating signals of cardiomyocytes by combining a 96-well plate with interdigitated electrodes at the bottom of each well and a high-resolution electrocardiogram system. As the cardiomyocytes contract and relax rhythmically, the coupling between the cardiomyocytes and the interdigitated electrodes also changes regularly, so that the measuring device can measure the change in impedance signal to reflect the beating state of the cardiomyocytes.

[0013] Furthermore, the said Step 2 includes the following specific steps:

[0014] Calculate non-linear dynamic parameters: The beating signals of cardiomyocytes are represented by a time series {x(i), i =

[0015] 0, 1, 2, …, N}, where N is the length of the time series. Calculate the non-linear dynamic parameters of each beating signal time series without dividing the time series into beating cycles. The 15 parameters are calculated based on the concepts of chaos, fractal and complexity.

[0016] Furthermore, the said Step 3 includes the following specific steps:

[0017] Step 3: Neural network training and analysis of drug types and concentrations: Establish a neural network model and train this network model with non-linear dynamic parameters. The neural network model consists of five layers, including an input layer, three hidden layers and an output layer. The input layer has 15 nodes corresponding to 15 non-linear dynamic parameters. The first, second and third hidden layers of the neural network contain 105, 120 and 105 nodes respectively. The input of this model is a feature vector, and each vector has 23 elements:

[0018] [F 1, F 2, F3, …, F 15 , LB1, LB2, LB3, …, LB7, DC]

[0019] Where F N (N = 1, 2, 3, …, 15) represents the non-linear dynamic characteristics, which are marked by the binary element LB N (N = 1, 2, 3,..., 7), representing the drug type, and a scalar element DC representing the drug concentration.

[0020] Furthermore, it includes a training step:

[0021] The model was trained for 8000 epochs using the Adam optimizer with a learning rate of 0.0002. The ReLU function was used as the activation function for each node in the hidden layer to avoid gradient vanishing and reduce computational cost. Half of all the sample numbers were used for training, and the other half were used for testing the neural network model. The classification ability of this model was evaluated by the receiver operating characteristic curve and the confusion matrix.

[0022] A method for evaluating drug-induced cardiotoxicity using non-linear dynamic analysis assisted neural network according to the present invention has the following advantages: The neural network algorithm based on non-linear dynamic analysis assisted by the present invention can identify drug-induced cardiotoxicity with an accuracy exceeding 0.99. The classification accuracy of a single drug exceeds 0.94, and the level of cardiotoxicity caused by the drug is also accurately predicted. When evaluating the cardiotoxicity of newly developed drugs, the toxicity recognition accuracy for cardiotoxic drugs with different drug concentrations reaches 0.85 - 0.95, with relatively high drug classification and concentration prediction accuracy. And this neural network model can avoid segmenting the pulsation signal into multiple cycles for analysis, directly process the pulsation time series with a fixed duration, and extract dynamic features. The method of the present invention has extremely high accuracy in identifying drug-induced cardiotoxicity, and also has excellent generalization performance, achieving a relatively high accuracy in identifying the cardiotoxicity of drugs not in the neural network training library. Therefore, the present invention can screen the cardiotoxicity of newly developed drugs, which is helpful for drug development. Description of the Drawings

[0023] Figure 1 It is a schematic diagram of the method flow of the present invention;

[0024] Figure 2a It is a schematic diagram of the training and test accuracy, and the training loss value curve of the first specific embodiment of the present invention;

[0025] Figure 2b It is a schematic diagram of the false positive rate and true positive rate ROC curve of the first specific embodiment of the present invention;

[0026] Figure 2c It is a schematic diagram of the confusion matrix of the accuracy of classifying drugs into non-cardiotoxic drugs, cardiotoxic drugs, and control groups in the first specific embodiment of the present invention;

[0027] Figure 2d Schematic diagram of the confusion matrix of the classification accuracy of each drug in the first specific embodiment of the present invention;

[0028] Figure 2e Schematic diagram of the predicted logarithm of drug concentration normalized between 0 and 10 in the first specific embodiment of the present invention;

[0029] Figure 3a Schematic diagram of the training and test accuracy and training loss value curves in the second specific embodiment of the present invention;

[0030] Figure 3b Schematic diagram of the ROC curve of false positive rate and true positive rate in the second specific embodiment of the present invention;

[0031] Figure 3c Schematic diagram of the confusion matrix of the classification accuracy of classifying drugs into non-cardiotoxic drugs, cardiotoxic drugs and control groups in the second specific embodiment of the present invention;

[0032] Figure 3d Schematic diagram of the confusion matrix of the classification accuracy of each drug in the second specific embodiment of the present invention;

[0033] Figure 3e Schematic diagram of the predicted logarithm of drug concentration normalized between 0 and 10 in the second specific embodiment of the present invention;

[0034] Figure 4a Schematic diagram of the accuracy of the drug cardiotoxicity assessment based on the pulsation cycle of the present invention;

[0035] Figure 4b Schematic diagram of the accuracy of the drug cardiotoxicity assessment based on time series of the present invention. Detailed implementation manners

[0036] In order to better understand the purpose, structure and function of the present invention, the following further describes in detail a method for using non-linear dynamic analysis to assist a neural network for evaluating drug-induced cardiotoxicity according to the present invention with reference to the accompanying drawings.

[0037] As Figure 1 shown, a method for using non-linear dynamic analysis to assist a neural network for evaluating drug-induced cardiotoxicity according to the present invention includes the following steps:

[0038] Step 1: Measure the beating signal of cardiomyocytes. The beating signal of cardiomyocytes is measured by combining a 96-well plate with interdigital electrodes at the bottom of each well and a high-resolution electrocardiogram system. As the cardiomyocytes contract and relax rhythmically, the coupling between the cardiomyocytes and the interdigital electrodes also changes regularly, enabling the measuring device to detect the change in impedance signal, which is used to reflect the beating state of cardiomyocytes.

[0039] Step 2: Calculate the nonlinear dynamics parameters. The beating signal of cardiomyocytes is represented by a time series {x(i), i = 0, 1, 2, …, N}, where N is the length of the time series. In the first embodiment, the time series is segmented into multiple beating cycle signals by an algorithm and its nonlinear dynamics parameters are calculated. In the second embodiment, the nonlinear dynamics parameters of each beating signal time series are calculated without segmenting the time series into beating cycles. 15 parameters are calculated based on the concepts of chaos, fractal, and complexity.

[0040] Step 3: Train the neural network and analyze the drug type and concentration. A neural network model is established and trained with the nonlinear dynamics parameters. The neural network model consists of five layers, including an input layer, three hidden layers, and an output layer. The input layer has 15 nodes corresponding to 15 nonlinear dynamic parameters. The first, second, and third hidden layers of the neural network contain 105, 120, and 105 nodes respectively. The input of this model is a feature vector, and each vector has 23 elements:

[0041] [F 1, F 2, F3, …, F 15 , LB1, LB2, LB3, …, LB7, DC]

[0042] where F N (N = 1, 2, 3, …, 15) represents the nonlinear dynamics feature. The nonlinear dynamics feature is marked by binary elements LB N (N = 1, 2, 3,..., 7), representing the drug types (three cardiotoxic drugs, three non-cardiotoxic drugs, and one control (DMSO)), and a scalar element DC representing the drug concentration.

[0043] The model was trained for 8000 epochs using the Adam optimizer with a learning rate of 0.0002. We adopted the ReLU function as the activation function for each node in the hidden layer to avoid the vanishing gradient and reduce the computational cost. In the drug evaluation study based on the nonlinear dynamic parameters of the cardiac cycle, we used the cardiac cycles of 13750 cardiomyocytes as training and test samples. In the drug evaluation study based on the cardiac time series, 5000 cardiac time series were used for analysis. For both forms of signals, half of the samples were used for training and the other half for testing the neural network model. The classification ability of the model was evaluated by the receiver operating characteristic curve (ROC) and the confusion matrix.

[0044] In the first specific embodiment, a neural network model assisted by nonlinear dynamic analysis was used to evaluate the cardiotoxicity caused by drugs using a single cardiac cycle. The cardiac cycles of single cardiomyocytes were used to train the neural network model. After training, the overall training accuracy of all drugs reached above 0.99, the test accuracy was higher than 0.98, and the drug concentration was accurately predicted, indicating the toxicity level of toxic drugs. The results are as Figures 2a - 2e shown. As Figure 2a shown, after 8000 epochs of training, the training and test accuracies exceeded 0.98, and the training loss value was lower than 0.02. As Figure 2b shown, the ROC curve shows that when the false positive rate is lower than 0.1, the true positive rate exceeds 0.98, and the area under the ROC curve exceeds 0.998. As Figure 2c shown, the confusion matrix shows that the accuracy of classifying drugs into non-cardiotoxic drugs, cardiotoxic drugs, or control groups exceeds 0.99. As Figure 2d shown, the confusion matrix indicates that the classification accuracy of each drug is above 0.96. As Figure 2e shown, for non-cardiotoxic drugs, the difference between the average predicted concentration and the true value is between 0 and 6%. For cardiotoxic drugs, the difference between the average predicted concentration and the true value is in the range of 0 to 5%. The black dashed line in the figure represents the true value of the drug concentration. The logarithmic values of drug concentrations 3.3, 4.7, 6.1, 7.5 correspond to the actual drug concentrations 0.000432 μM, 0.0216 μM, 0.11 μM, 0.54 μM respectively, and the logarithmic values of drug concentrations 4.4, 5.8, 7.2, 8.6 correspond to the actual drug concentrations 0.016 μM, 0.08 μM, 0.4 μM, 2 μM respectively. ‘***’ indicates the significance level p < 0.001 of the t-test.

[0045] In the second specific embodiment, a neural network model assisted by non-linear dynamics analysis is used to evaluate drug-induced cardiotoxicity by using a pulsatile time series of a fixed time length. After training for 8000 epochs, drugs are classified, and the overall test accuracy is 0.97. At the same time, the drug concentration is accurately predicted. The results are as Figures 3a - 3e shown. As Figure 3a shown, after training for 8000 epochs, the training and test accuracies exceed 0.97, and the training loss is lower than 0.016. As Figure 3b shown, the ROC curve shows that when the false positive rate is lower than 0.1, the true positive rate exceeds 0.99, and the area under the ROC curve exceeds 0.996. As Figure 3c shown, the confusion matrix shows that the accuracy of classifying drugs into non-cardiotoxic drugs, cardiotoxic drugs or control groups exceeds 0.96. As Figure 3d shown, the confusion matrix shows that the classification accuracy of each drug is above 0.94. As Figure 3e shown, for non-cardiotoxic drugs, the difference between the average predicted concentration and the true value is between 1% and 17%. For cardiotoxic drugs, the difference between the average predicted concentration and the true value is in the range of 0 to 10%. The black dotted line in the figure represents the true value of the drug concentration. The logarithmic values of the drug concentrations 3.3, 4.7, 6.1, 7.5 correspond to the actual drug concentrations 0.000432 μM, 0.0216 μM, 0.11 μM, 0.54 μM respectively, and the logarithmic values of the drug concentrations 4.4, 5.8, 7.2, 8.6 correspond to the actual drug concentrations 0.016 μM, 0.08 μM, 0.4 μM, 2 μM respectively. ‘***’ indicates that the significance level p of the t-test < 0.001.

[0046] In the third specific embodiment, a neural network model assisted by non-linear dynamics analysis is used to analyze the cardiotoxicity caused by drugs not in this neural network training library, and the generalization performance of this neural network is tested. The newly introduced drugs include a non-cardiotoxic drug (carvedilol) and a cardiotoxic drug (droperidol). By analyzing the pulsatile time series of cardiomyocytes with a fixed time length, the recognition accuracy of the cardiotoxicity of carvedilol is between 0.92 and 0.99, and the recognition accuracy of droperidol is between 0.85 and 0.95. The results are as Figures 4a - 4b shown. As Figure 4a shown, in the evaluation of the cardiotoxicity of drugs based on the pulsation period, the accuracy of classifying four concentrations of carvedilol as non-cardiotoxic drugs is between 0.77 - 0.95, and the accuracy of classifying four concentrations of droperidol as cardiotoxic drugs is between 0.64 - 0.99. As Figure 4bAs shown, in the assessment of drug cardiotoxicity based on time series, the accuracy of classifying carvedilol at four concentrations as a drug without cardiotoxicity is between 0.92 and 0.99, and the accuracy of classifying droperidol at four concentrations as a drug with cardiotoxicity is between 0.85 and 0.95.

[0047] From the description of the above embodiments, it can be seen that the neural network model assisted by the nonlinear kinetic analysis has a quite high accuracy in the application of assessing drug-induced cardiotoxicity. This model can also successfully identify the additionally introduced drugs in the general ability test. It shows that this model can be applied to screen newly developed drugs and predict whether they cause cardiotoxicity, which is helpful for drug development.

[0048] It can be understood that the present invention is described through some embodiments. Those skilled in the art know that without departing from the spirit and scope of the present invention, various changes or equivalent substitutions can be made to these features and embodiments. Additionally, under the teaching of the present invention, these features and embodiments can be modified to adapt to specific situations and materials without departing from the spirit and scope of the present invention. Therefore, the present invention is not limited by the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application belong to the scope protected by the present invention.

Claims

1. A method using non-linear dynamic analysis to assist neural networks for evaluating drug-induced cardiotoxicity, characterized in that, It includes the following steps: Step 1: Use an electrocardiogram system with interdigital electrodes to record myocardial cell beating signals; Step 2: Denoise the original myocardial cell beating signals through a filter and calculate 15 non-linear dynamic parameters for each time series; the 15 non-linear dynamic parameters include: time delay, embedding dimension, correlation dimension, maximum Lyapunov exponent, Shannon entropy, approximate entropy, spectral entropy, box dimension, C0 complexity, Poincaré plot parameters, detrended fluctuation analysis parameters, and recurrence quantification analysis parameters; Calculate non-linear dynamics parameters: The myocardial cell beating signals are represented by the time series {x(i), i = 0, 1, 2, …, N}, where N is the length of the time series. Without dividing the time series into beating cycles, calculate the non-linear dynamics parameters for each beating signal time series. The non-linear dynamics parameters, and at the same time, the 15 parameters are calculated according to the concepts of chaos, fractal, and complexity; Step 3: Input the non-linear parameters into neural network training to classify drugs and predict drug concentration levels, indicating the toxicity intensity of cardiotoxic drugs; Neural network training and analysis of drug types and concentrations: Establish a neural network model and train this network model with non-linear dynamics parameters. The neural network model consists of five layers, including an input layer, three hidden layers, and an output layer. The input layer has 15 nodes corresponding to 15 non-linear dynamic parameters. The first, second, and third hidden layers of the neural network contain 105, 120, and 105 nodes respectively. The input of this model is a feature vector, and each vector has 23 elements: [F 1, F 2, F3,…,F 15 ,LB1,LB2,LB3,…,LB7,DC] where F N represents the non-linear kinetic characteristics, N = 1, 2, 3, …, 15; the non-linear kinetic characteristics are marked by the binary element LB N , N = 1, 2, 3, ..., 7, LB N represents the drug type, and a scalar element DC representing the drug concentration.

2. The method for evaluating drug-induced cardiotoxicity using non-linear dynamic analysis to assist neural networks according to claim 1, characterized in that, The said Step 1 includes the following specific steps: Measure myocardial cell beating signals: Measure myocardial cell beating signals by combining a 96-well plate with interdigital electrodes at the bottom of each well and a high-resolution electrocardiogram system. As the myocardial cells contract and relax rhythmically, the coupling between the myocardial cells and the interdigital electrodes also changes regularly, so that the measuring device can measure the change in impedance signal to reflect the beating state of the myocardial cells.

3. The method for evaluating drug-induced cardiotoxicity using a non-linear dynamic analysis-assisted neural network according to claim 1, characterized in that, It includes a training step: The model was trained for 8000 epochs using the Adam optimizer with a learning rate of 0.0002. The ReLU function was used as the activation function for each node in the hidden layer to avoid gradient disappearance and reduce computational costs. Half of all the sample numbers were used for training, and the other half was used for testing the neural network model. The classification ability of this model was evaluated through the receiver operating characteristic curve and the confusion matrix.

Images