KR102475057B1 - Screening method for drugs that cause cardiac arrhythmia using artificial neural networks - Google Patents

Abstract

A method for screening cardiac arrhythmia-inducing drugs using an artificial neural network includes the steps of normalizing an action potential characteristic factor index, a calcium characteristic factor index, and a net charge characteristic factor index related to arrhythmias and inputting them to an input layer node of the artificial neural network; Transferring from the input layer to the hidden layer of the neural network using an activation function ReLU function (Rectified Linear Unit), and outputting the risk of cardiac arrhythmia predicted in the hidden layer to a node in the output layer using an activation function Softmax function It is characterized by doing.

Description

Screening method for drugs that cause cardiac arrhythmia using artificial neural networks}

The present invention relates to a method for screening cardiac arrhythmia, and more particularly, to a method for screening a drug that induces cardiac arrhythmia using an artificial neural network.

In 1999, the digestive stimulant cisapride was withdrawn from the European pharmaceutical market due to Torsades de Ponte (TdP). In 2005, a guideline evaluating the possibility of inducing cardiac arrhythmias for all drugs was presented through The International Council for Harmonization of Technical Requirements for Pharmaceuticals for Human Use (ICH). Since then, the guideline has suggested that cardiotoxicity evaluation of drugs be performed according to S7B, which is a non-clinical evaluation, and E14, which is a clinical evaluation. This guideline requires extensive clinical trials, and the cardiotoxicity classification method according to these clinical trials has high sensitivity but low specificity. This means that even drugs that do not induce TdP are subject to strict regulation, negatively affecting drug development.

A Comprehensive in vitro Proarrhythmia Assay (CiPA) project was established with the participation of 13 advanced medical institutions at the Think-tank conference hosted by the FDA headquarters in 2013 with the goal of revising these existing drug development guidelines. Among the guidelines to be revised through the CiPA project, the main content to be changed in the non-clinical evaluation guideline S7B is to comprehensively evaluate the drug response of multiple ion channels using the in silico method instead of the hERG channel single analysis evaluation method through in vitro experiments. to evaluate

In addition, in 2017, the Parikh group assigned the effect of drugs to the Ord model to comprehensively evaluate the drug response of multiple ion channels, and classified the TdP-inducing drug groups using the logical regression technique. As inputs for the logical regression model, 13 electrophysiological characteristics (Upstroke velocity, Peak voltage, APD50, APD @ -60 mV, APD90, Resting voltage, AP triangulation, diastolic [Ca ²⁺ ]I, amplitude of CaT, peak [Ca ²⁺ ]I, CaTD50, CaTD90, CaT triangulation) were used individually. . Under the drug concentration condition when the IKr channel was blocked by 60%, the accuracy of predicting the presence or absence of EAD showed a verification score of at least 76 points and a maximum of 100 points.

In 2017, the Dutta group used an in silico model that modified the ORD model to reflect the effects of drugs, and analyzed the current graph over time for qNet indicators (6 ion channels (INaL, ICaL, IKr, Ito, IK1, IKs)). The sum of the areas below) was derived to establish a standard for classifying the risk of TdP occurrence of drugs into high-risk, medium-risk, and low-risk groups.

Through a linear regression model with qNet indicators as feature points, the Li group in 2018 performed drug risk classification using qNet as an input to a logical regression model. The qNet was calculated by adding a herg dynamic model to the ORD cardiomyocyte model5 modified by Dutta's group. The drug risk group classification was classified into the drug risk group with two critical values obtained through two binary classifications. Although the accuracy of classification was improved by adding the hERG-dynamic model, there were disadvantages such as large amount of data processing and mathematical complexity, such as in vitro experiments for parameter evaluation and subsequent quantification of uncertainty.

Therefore, in 2020, the Liopis-Lorente group used nine decision trees for drug risk groups as a new classification method. The indicators used as inputs are Tx (Romero et al., 2018), the ratio of drug concentration to EFTPC concentration when APD90 is increased by 10%, calculated through Dutta's modified ORD model (Romero et al., 2018), and qNet calculated when EFTPC concentration is 10 times higher. TqNet, which is the ratio between the value and the qNet value at steady state, and Ttriang, which is the ratio between the difference between APD90 and APD30 calculated at 10-fold concentration of EFTPC and the difference between APD90 and APD30 calculated at steady state. The classification accuracy of drug risk groups was 0.899 when Tx was used as an input, 0.908 when Ttriang was used as an input, and 0.917 when Tqnet was used.

The flow of preceding studies conducted so far is to derive the action potential morphology characteristics (APD) or charge characteristics (Qnet), which can be said to have a very high correlation with TdP, from the ORD in silico model, and use these to derive logical regression and decision trees. We used a classical binary classification method that classifies in a linear pattern such as However, there is no guarantee that the toxic effect of a drug should have a linear distribution according to the risk group in the action potential characteristics and charge characteristics. In addition to the action potential morphology and charge characteristics, the Ca concentration morphology proposed by Dutta also has a high correlation with TdP generation.

J. Parikh, V. Gurev, and J. J. Rice, "Novel two-step classifier for Torsades de Pointes risk stratification from direct features," Front. Pharmacol., vol. 8, no. NOV, pp. 1-18, 2017, doi: 10.3389/fphar.2017.00816.

The present invention has been proposed to solve the above technical problem, and proposes a model showing improved classification performance even when using a static model using the previously used IC50 and hill coefficient. The proposed model uses the traditional static model using IC50 and hill coefficients. Using static model data as input and nine indices highly correlated with TdP calculated through action potential simulation as input to the artificial neural network model, drug-risk groups are classified into high-risk, medium-risk, and low-risk groups, and existing studies A model with improved classification performance is presented.

In addition, the present invention proposes a screening method for cardiac arrhythmia-inducing drugs using an artificial neural network that takes 9 multi-characteristic values considering all of the action potential morphology, calcium concentration morphology, and charge characteristics as inputs in order to further increase the accuracy of drug toxicity evaluation. do.

According to an embodiment of the present invention to solve the above problem, normalizing the action potential characteristic factor index, calcium characteristic factor index, and net charge characteristic factor index related to arrhythmia and inputting them to the input layer node of the artificial neural network and transmitting from the input layer to the hidden layer of the neural network using an activation function ReLU function (Rectified Linear Unit), and outputting the risk of cardiac arrhythmia predicted in the hidden layer to an output layer node using an activation function Softmax function. A screening method for a cardiac arrhythmia-inducing drug using an artificial neural network comprising the steps is provided.

In addition, the artificial neural network in the present invention is characterized by including the input layer having 9 nodes, the hidden layer having 5 nodes, and the output layer having 3 nodes.

In addition, in the present invention, the action potential characteristic factor index is dVm/dtmax (the maximum slope when the membrane potential is depolarized in the shape of the action potential), APD50 (the period from the depolarization of the membrane potential to the point of 50% repolarization from the maximum amplitude), APD90 ( It is characterized by including the period from the depolarization of the membrane potential to the point of 90% repolarization at the maximum amplitude) and APresting (resting membrane potential).

In addition, the calcium characteristic factor index in the present invention is CaD50 (calcium transient duration 50) - the period between the 50% points based on the maximum amplitude during the calcium transient period -, CaD90 (calcium transient duration 50) - during the calcium transient period It is characterized by including the period between the 90% points based on the maximum amplitude - and , Caresting (diastolic concentration of calcium).

In addition, in the present invention, the indicator of whether the net charge is characteristic is qinward (rate of charges moving through INaL and ICaL ion channels depending on the steady state and drug state), qnet (the ratio of ion charges moving in and out of the cell through ion channels) It is characterized in that it includes the total).

In addition, in the present invention, the input data input to the input layer node is characterized in that it is preprocessed using MinMaxScaler to avoid overfitting or underfitting caused by unit differences between indices.

The artificial neural network model proposed in the present invention improved the drug risk classification accuracy of the existing model. Using one artificial neural network model, it is possible to know the individual classification accuracy of high-risk, medium-risk, and low-risk. This model is a model that can reflect changes in action potential shapes, changes in calcium concentration, and changes in moving ionic charges caused by drugs. The existing cardiotoxicity evaluation classification criteria of drugs were based on repolarization, but the changing guidelines tried to classify based on the possibility of inducing TdP. However, it is difficult to select an accurate end point to classify the possibility of arrhythmia, and it is reasonable to accept that arrhythmia of a drug is classified using a composite index highly related to arrhythmia.

The limitation of this study is that it is dependent on in vitro experimental data because it implements drug changes in electrophysiology by inputting in vitro experimental data. AI classification is classified by learning parameters, and the exact physiological meaning of the value during the calculation process cannot be known. However, more accurate drug evaluation can be expected by using the 9 input indicators of the proposed model as a combination of in silico and in vitro results. It can be seen that the two peaks in the histogram of the medium-risk group resulted in stratified results. It can be interpreted as the result that the medium-risk group has many data that are close to the two categories of high-risk group or low-risk group.

In addition, in the present invention, action potential morphology characteristics (APD90, APD50, dV/dtmax, APresting), calcium concentration morphology (CaD90, CaD50, Caresting), and charge characteristics (qNet, aInward) were all considered for the toxicity evaluation of drugs in the present invention. An artificial neural network-based deep learning algorithm with multiple feature values as input was developed. In silico simulation was performed using the modified ORD model to derive 9 characteristic values. When the performance of this classification algorithm was evaluated with the same criteria as the Li group using the same in silico model, the classification performance presented by the Li group was 10.2% higher in the high-risk group and 6.7% higher in the low-risk group AUC. In addition, by using an artificial neural network instead of the logical regression classification method used by the Li group, it was possible to explicitly classify not only high-risk and low-risk groups, but also medium-risk groups.

As a limitation of the present invention, nine parameters must be provided as input values for the toxicity evaluation classifier. This is very large compared to the one input required in previous studies. In order to acquire the 9 reliable parameters, sufficient reliability of the physiological/pharmacological in silico model of cardiomyocytes should be supported. As a second limitation, the artificial neural network proposed in the present invention does not present an explicit threshold for classification, compared to classification based on threshold values of physiologically/pharmacologically meaningful parameters such as qNet, qInward, and TqNet suggested in previous studies. that it can't This means that when a new drug development researcher uses this artificial neural network classifier for cardiotoxicity evaluation, it is difficult to causally evaluate the validity of the classified results. The above two limitations are inevitably met if the artificial neural network-based deep learning method is used. Nevertheless, it is very meaningful in the field of new drug development research to develop an algorithm with higher toxicity evaluation classification performance than previous studies.

1 is a diagram showing a method for screening cardiac arrhythmia-inducing drugs using an artificial neural network according to an embodiment of the present invention.
2 is a histogram showing the frequency of AUCs (Artificial Neuron Networks) obtained through 10,000 tests

Hereinafter, embodiments of the present invention will be described with reference to the accompanying drawings in order to describe in detail enough for those skilled in the art to easily implement the technical idea of the present invention.

1 is a diagram showing a method for screening cardiac arrhythmia-inducing drugs using an artificial neural network according to an embodiment of the present invention. Referring to FIG. 1, a method for screening cardiac arrhythmia-inducing drugs using an artificial neural network will be described as follows.

[Software and Data]

The simulation code used in the experiment to verify the present invention was used based on the data uploaded to the github website (https://github.com/FDA/CiPA/) by the CiPA project group. Patch-clamp experimental data used for Hill fit are from Crumble et al. (2016) group used the experimental results for six ion channels using a manual patch-clamp technique in a physiological temperature environment. Uncertainty quantification and simulation performance were conducted in accordance with the CiPA initiative.

[Fitting (bootstrap method)]

In order to obtain the characteristics of each ion channel generated by the drug, the patch clamp experiment data is used as an input and fit. The effect of the drug on the channel is calculated according to the Hill equation, and the change in maximum conductance that appears to the degree of blockage of the channel by the drug is calculated. The Hill fitting process was described by Chang et al. (2017) extracts 2,000 Hill coefficients and IC50 bootstrapped within a 95% confidence interval through the MCMC (Markov-chain Monte Carlo) model proposed by the group. The 2,000 IC50s and Hill coefficients calculated through this simulation are used as inputs for a single cardiomyocyte simulation. After confirming that the data published by the CiPA group and the result value calculated through the R code were the same, the R code was executed and the result data was used.

[In silico model description]

For the electrophysiological single-cell simulation model, a modified maximum conductance of the ion channel of the myocardial cell model proposed by Ohara Rudy et al. (2011) was used. The model quantifies the effect of each ion channel on the shape of the action potential, and in order to better represent the change in the duration of the action potential caused by the response of the ion channel changed by the drug, the model is based on the underestimated or overestimated maximum conductance of the channel. Corrections have been made to The modified model applied conversion factors of 2.661, 1.007, 1.013, 1.870, and 1.698 for the conductivities of the ion channels INaL, ICaL, IKr, IKs, and IK1, respectively.

[EP simulation protocol]

<Table 1>

Table 1 is a table showing 28 drugs selected by the CiPA research group into high-risk, medium-risk, and low-risk groups according to their arrhythmia-inducing potential, and 12 drugs used for machine learning training and 16 drugs used for testing. .

The single-cell electrophysiology simulation generated 1000 action potentials, and the stimulation cycle was set to 2 s. The intensity of stimulation was set to -80 μA/μF, and the duration of stimulation was set to 0.5 ms. The concentration of the drug was set to 1-4 times (increase by 1) based on the highest protein-free Cmax, which is a characteristic of each drug. The drugs used were 28 drugs with labels classified into high-risk, medium-risk, and low-risk groups according to the risk of cardiac arrhythmia in the CiPA project group. Among the 28 drugs, 12 drugs were used for training of the artificial neural network model, and 16 drugs were used for model testing. As inputs for the electrophysiology simulation, 2,000 IC50s and Hill coefficients extracted by Hill fit were used. According to the set drug concentration Cmax 1, 2, 3, 4 times, 2000 9 indicators are calculated individually. The data used for the artificial neural network is the average of the Cmax1-4 concentrations individually calculated according to the concentration, and has 2,000 result values for each indicator.

[Biomarker calculation used as an artificial neural network input]

<Equation 1>

<Equation 2>

Nine indices related to arrhythmias are derived through single-cell electrophysiological simulation. Indicators are divided into action potential characteristics, calcium characteristics, and ionic charge characteristics. Among action potential characteristics, APD90 (Action potential duration) is the period between the depolarization point and the repolarization point below 90% of the maximum amplitude in the shape of the action potential. APD50 is the period between the point of depolarization and repolarization below 50% of the maximum amplitude in the action potential profile. dvm/dtmax is the maximum slope when the membrane potential depolarizes in the action potential shape, and APresting is the resting membrane potential. Calcium characteristics include CaD90 (Calcium transient Duration), which is the period between the maximum amplitude and the point below 90% during the transient period of intracellular calcium. CaD50 The period between the maximum amplitude and the point below 50% during the transient period of intracellular calcium. Caresting is the concentration of intracellular calcium during relaxation, and qNet of the ionic charge characteristic is the total amount of ion charges moving through six ion channels (INaL, ICaL, IKr, IKs, IK1, Ito) during the action potential period, and time - Calculated as the sum of the integrals of the current graph according to Equation 1-. qNet was developed by Li et al. (2018) was used as an input index for the classification of arrhythmia-inducing risk of drugs using a logistic regression model performed by the group. qInward is the average of the drug response and steady-state ratios of charges directed into the cell through the ICaL and INaL ion channels during the action potential period and is expressed as (Equation 2). Here, _drug_AUC is the area under the graph of current change over time of each ion channel when the drug properties are reflected, and _control_AUC is the area under the graph of current change over time of each ion channel under steady state. The criterion for selecting one action potential for index calculation is to select one action potential with the largest dVm/dtrepol value at the time of repolarization, excluding failures in depolarization or repolarization for the last 250 action potentials out of 1,000 action potentials. Dog indicators are calculated.

[Description of machine learning model structure and learning/test method]

The artificial neural network model according to an embodiment of the present invention has 9 nodes that take 9 indicators (dv _m /dt _max , AP _resting , APD ₉₀ , APD ₅₀ , Ca _resting , CaD ₉₀ , CaD ₅₀ , qNet, qInward) as inputs. An artificial neural network model consisting of an input layer with , a hidden layer with 5 nodes, and an output layer with 3 nodes was constructed. MinMaxScaler was used for the input data to avoid overfitting or underfitting caused by unit differences between indicators. To prevent overfitting during the training process, 10-order cross-validation was performed. 'ReLu' was used as the activation function of the hidden layer, and 'Softmax' was used as the activation function of the output layer. Through the node of the output layer, multiple classifications were made into high-risk, medium-risk, and low-risk groups. The number of data used has 2,000 data calculated by the Hill coefficient and IC50 parameter used as input for action potential simulation, and the number of data for one drug is 18,000 with 9 indicators. Twelve drugs were used for training, and 216,600 data were used as training data for the machine learning model. The data used for verification also had 2,000 data per indicator calculated from the Hill coefficient and IC50, and 9 indicators were used. There were 16 tested drugs, and the total number of data was 288,000. The data set used for the test of the model includes 16 drugs, and 9 indicators calculated with the same Hill coefficient and IC50 for each drug are randomly selected from among 2,000 data. The test of the model is performed 10,000 times using 10,000 data sets each containing data of 16 test drugs. As a test result, the performance of the model was evaluated using AUC (Area under the curve) corresponding to the area of the receiver operating characteristic (ROC curve). For comparison of results, the 95% range and median value of the AUC confidence interval among the frequency distributions of 10,000 AUC results were set as representative values.

<Table 2>

Table 2 is a comparison table of the prediction accuracy of the arrhythmia risk group using the proposed Artificial Neuronal Network (ANN) model and the logical regression model proposed by the Li group. It is a histogram showing the frequency of ((a) high-risk group, (b) medium-risk group, (c) low-risk group)

The representative values of the AUCs obtained through 10,000 tests using the learned ANN classifier developed in the present invention were 0.958 in the high-risk group, 0.73 in the medium-risk group, and 0.927 in the low-risk group (the median value in the histogram in FIG. It is a representative value, and the 95% range of the confidence interval of the dataset is the validation range). As the classification accuracy of the logical regression model presented by the Li group, the representative AUC of the high-risk group was 0.856 and the representative AUC of the low-risk group was 0.86 (due to the methodology, the AUC of the medium-risk group could not be presented). Therefore, the accuracy of the ANN classifier developed in the present invention was 10.2% higher in the high-risk group and 6.7% higher in AUC in the low-risk group than the accuracy presented in the preceding study (Li et al. 2018). The minimum value of the confidence interval was higher for the ANN classifier by 9% in the high-risk group and 7% in the low-risk group, and the maximum value of the confidence interval was also higher in the ANN classifier by 10% in the high-risk group and 10.5% in the low-risk group.

As described above, the screening method for cardiac arrhythmia-inducing drugs using the proposed artificial neural network is

The step of normalizing the action potential characteristic factor index, calcium characteristic factor index, and net charge characteristic factor index related to arrhythmias and inputting them to the input layer node of the artificial neural network, and using the activation function ReLU function (Rectified Linear Unit) The process includes transferring from the input layer to the hidden layer of the neural network, and outputting the risk of cardiac arrhythmia predicted in the hidden layer to a node in the output layer using the Softmax activation function.

At this time, the artificial neural network includes an input layer having 9 nodes, a hidden layer having 5 nodes, and an output layer having 3 nodes.

In addition, the action potential characteristic factor index is dVm/dtmax (the maximum slope when the membrane potential is depolarized in the shape of the action potential), APD50 (the period from the depolarization of the membrane potential to the point of 50% repolarization from the maximum amplitude), APD90 (the depolarization of the membrane potential) period from maximum amplitude to the time of 90% repolarization) and APresting (resting membrane potential).

In addition, CaD50 (calcium transient duration 50) - the period between the 50% points based on the maximum amplitude during the transient period of calcium - and CaD90 (calcium transient duration 50) - the maximum amplitude during the transient period of calcium The period between the 90% points as a criterion - and , including Caresting (diastolic concentration of calcium).

In addition, as indicators of net charge characteristics, qinward (rate of charges moving through INaL and ICaL ion channels depending on the normal state and drug state) and qnet (the total sum of ion charges moving in and out of the cell through ion channels) It is characterized by including.

In addition, the input data input to the input layer node is characterized in that it is preprocessed using MinMaxScaler to avoid overfitting or underfitting caused by unit differences between indices.

As such, those skilled in the art to which the present invention pertains will be able to understand that the present invention may be embodied in other specific forms without changing its technical spirit or essential features. Therefore, the embodiments described above should be understood as illustrative in all respects and not limiting. The scope of the present invention is indicated by the following claims rather than the detailed description above, and all changes or modifications derived from the meaning and scope of the claims and their equivalent concepts should be interpreted as being included in the scope of the present invention. do.

Claims (6)

delete normalizing an action potential characteristic factor index, a calcium characteristic factor index, and a net charge characteristic factor index related to arrhythmias and inputting them to an input layer node of an artificial neural network;
transmitting from the input layer to the hidden layer of the neural network using an activation function ReLU function (Rectified Linear Unit); and
Outputting a cardiac arrhythmia risk predicted in the hidden layer to an output layer node by using an activation function Softmax function;
The artificial neural network includes the input layer having 9 nodes, the hidden layer having 5 nodes, and the output layer having 3 nodes.
normalizing an action potential characteristic factor index, a calcium characteristic factor index, and a net charge characteristic factor index related to arrhythmias and inputting them to an input layer node of an artificial neural network;
transmitting from the input layer to the hidden layer of the neural network using an activation function ReLU function (Rectified Linear Unit); and
Outputting a cardiac arrhythmia risk predicted in the hidden layer to an output layer node by using an activation function Softmax function;
The action potential characteristic factor index is dVm/dtmax (the maximum slope when the membrane potential is depolarized in the shape of the action potential), APD50 (the period from the depolarization of the membrane potential to the point of 50% repolarization from the maximum amplitude), APD90 (from the depolarization of the membrane potential) A method for screening cardiac arrhythmia-inducing drugs using an artificial neural network, characterized in that it includes period from maximum amplitude to 90% repolarization) and APresting (resting membrane potential).
normalizing an action potential characteristic factor index, a calcium characteristic factor index, and a net charge characteristic factor index related to arrhythmias and inputting them to an input layer node of an artificial neural network;
transmitting from the input layer to the hidden layer of the neural network using an activation function ReLU function (Rectified Linear Unit); and
Outputting a cardiac arrhythmia risk predicted in the hidden layer to an output layer node by using an activation function Softmax function;
The calcium characteristic factor index, CaD50 (calcium transient duration 50) - period between 50% points based on the maximum amplitude during the transient period of calcium -, CaD90 (calcium transient duration 50) - based on the maximum amplitude during the transient period of calcium A method for screening cardiac arrhythmia-inducing drugs using an artificial neural network, characterized in that it includes a period between the 90% points - and , Caresting (diastolic concentration of calcium).
normalizing an action potential characteristic factor index, a calcium characteristic factor index, and a net charge characteristic factor index related to arrhythmias and inputting them to an input layer node of an artificial neural network;
transmitting from the input layer to the hidden layer of the neural network using an activation function ReLU function (Rectified Linear Unit); and
Outputting a cardiac arrhythmia risk predicted in the hidden layer to an output layer node by using an activation function Softmax function;
The net charge characteristic index includes qinward (rate of charges moving through INaL and ICaL ion channels depending on the steady state and drug state), qnet (total sum of ion charges moving in and out of the cell through ion channels) A screening method for cardiac arrhythmia-inducing drugs using an artificial neural network, characterized in that.
normalizing an action potential characteristic factor index, a calcium characteristic factor index, and a net charge characteristic factor index related to arrhythmias and inputting them to an input layer node of an artificial neural network;
transmitting from the input layer to the hidden layer of the neural network using an activation function ReLU function (Rectified Linear Unit); and
Outputting a cardiac arrhythmia risk predicted in the hidden layer to an output layer node by using an activation function Softmax function;
The input data input to the input layer node is preprocessed using MinMaxScaler to avoid overfitting or underfitting caused by unit differences between indicators.

Images