Regression model evaluation device, regression model evaluation method, and regression model evaluation program

Abstract

The accuracy of the mathematical parameters output from the regression model M is improved. [Solution] The ion channel evaluation device 7 includes an evaluation unit 80 that evaluates the estimated value of mathematical parameters 70 output in correspondence with first data 41 from a regression model M to which first data 41 indicating the ion current flowing through the ion channel 11 is input, using a loss function. The evaluation unit 80 considers the loss of the estimated value of mathematical parameters 70 to be the minimum value when the input value of mathematical parameters 70 corresponding to the first data 41 relating to the ion current I satisfies predetermined conditions.

Description

Technical Field

[0001] The present disclosure relates to an evaluation device for a regression model, an evaluation method for a regression model, and an evaluation program for a regression model.

Background Art

[0002] There are transmembrane proteins called ion channels in the cell membrane of a living body (for example, Patent Document 1). Ion channels serve as passageways for ions in the cell membrane. Ions carry electric charges. In the cell membrane, ions are passively permeated through ion channels. Ion channels open and close.

[0003] When an ion channel opens, the ion current flowing through the ion channel increases. When an ion channel closes, the ion current flowing through the ion channel decreases.

Prior Art Documents

Patent Documents

[0004]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0005] By the way, it is conceivable to analyze the state of an ion channel based on first data (such as a current waveform or a histogram) obtained from the ion current flowing through the ion channel.

[0006] In order to analyze the state of an ion channel based on data obtained from the ion current flowing through the ion channel, it is necessary to obtain a mathematical parameter corresponding to the data, but it has been difficult to obtain such a mathematical parameter.

[0007] Therefore, it is conceivable to generate a regression model in which, when the first data obtained from ion current is input, mathematical parameters corresponding to the first data are output.

[0008] However, when generating such regression models, training the model using predetermined first data as training data sometimes resulted in a decrease in the accuracy of the output mathematical parameters.

[0009] This disclosure has been made in view of the above, and its purpose is to provide a regression model evaluation device, a regression model evaluation method, and a regression model evaluation program that can improve the accuracy of mathematical parameters output from the regression model. [Means for solving the problem]

[0010] The regression model evaluation device according to this disclosure includes an evaluation unit that evaluates a second mathematical parameter, which is an estimated value of a mathematical parameter output in correspondence with the first data, from a regression model to which first data representing a first ion current flowing through a first ion channel is input, using a loss function. The evaluation unit considers the loss of the second mathematical parameter to be at its minimum value when the first mathematical parameter, which is the input value of the mathematical parameter corresponding to the first data, satisfies predetermined conditions. [Effects of the Invention]

[0011] According to this disclosure, the accuracy of mathematical parameters output from regression models can be improved. [Brief explanation of the drawing]

[0012] [Figure 1] Figure 1 shows the ion channels formed in the artificial cell membrane. [Figure 2] Figure 2 shows the current waveform of the ion current. [Figure 3] Figure 3 shows the histogram of ion current. [Figure 4] Figure 4 shows the regression model M according to the embodiment. [Figure 5] FIG. 5 shows the current waveform of the ion current I when the opening probabilities of the ion channels 11 and 12 are 0%. [Figure 6] FIG. 6 shows the current waveform of the ion current I when the opening probabilities of the ion channels 11 and 12 are 100%. [Figure 7] FIG. 7 shows the current waveform of the ion current I when the ion current widths of the ion channels 11 and 12 are the same.

BEST MODE FOR CARRYING OUT THE INVENTION

[0013] Hereinafter, embodiments of the present disclosure will be described in detail based on the drawings. The following description of the preferred embodiments is merely exemplary in nature and is in no way intended to limit the present disclosure, its applications, or its uses.

[0014] Note that the ion current data generation method and the ion current data generation program of the present disclosure are implemented as a function of an ion current data generation device.

[0015] <Embodiment> (Ion Channel Analyzer) The ion channel analyzer 1 will be described. The ion channel analyzer 1 is applied to the ion channel 10. The ion channel analyzer 1 analyzes the state of the ion channel 10. In this example, the ion channel 10 is formed in the artificial cell membrane 5. Note that in FIG. 1, a case where there are two ion channels 11 and 12 formed in the artificial cell membrane 5 is shown.

[0016] Figure 1 shows an ion channel 10 formed in an artificial cell membrane 5. The artificial cell membrane 5 mimics the cell membrane of a living body and is artificially formed. When water droplets 4 are dropped into an organic solvent (oil) in which lipid molecules 2 are dispersed, a monolayer of lipid molecules 2 is spontaneously formed around the water droplets 4. When two water droplets 4 are brought into contact, the monolayers of lipid molecules 2 overlap at the contact portion between the two water droplets 4. An artificial cell membrane 5 composed of a bilayer of lipid molecules 2 is formed between the two water droplets 4. One of the two water droplets 4 mimics the inside of a living cell, and the other of the two water droplets 4 mimics the outside of a living cell.

[0017] The protein 3 that serves as the origin of the ion channel 10 is dispersed in the water droplet 4. When the protein 3 adheres to the artificial cell membrane 5, the ion channel 10 is formed.

[0018] The ion channel 10 is a passage that penetrates the artificial cell membrane 5. The ion channel 10 connects the two water droplets 4 to each other. Ions passively pass through the ion channel 10 between the two water droplets 4. The ion channel 10 is a passage for ions in the artificial cell membrane 5. Ions carry a charge. Examples of ions include potassium ions and sodium ions.

[0019] The ion channel 10 opens and closes. When the ion channel 10 opens, the ion current I flowing through the ion channel 10 between the two water droplets 4 increases. When the ion channel 10 closes, the ion current I flowing through the ion channel 10 between the two water droplets 4 decreases.

[0020] The ion channel analyzer 1 includes a detection unit 20 (measuring instrument) and an analysis unit 30. The detection unit 20 is a known ammeter. The detection unit 20 detects the ion current I flowing through the ion channel 10. The detection unit 20 is connected in series to both ends of the ion channel 10 (one water droplet 4 side and the other water droplet 4 side). In addition, the detection unit 20 has a function of applying a command voltage to the ion channel 10.

[0021] The analysis unit 30 is connected to the detection unit 20. The analysis unit 30 is built into the main body of the ion channel analyzer 1. The analysis unit 30 is, for example, a computer. The analysis unit 30 includes, for example, a processor mounted on a substrate and a memory device that stores software for operating the processor. The analysis unit 30 performs the calculation processing described later.

[0022] The analysis unit 30 obtains first data 40 related to the ion current I detected by the detection unit 20. Based on the first data 40 related to the ion current I, the analysis unit 30 analyzes the state of the ion channel 10. In particular, in this example, the analysis unit 30 analyzes the open / closed state of the ion channel 10.

[0023] (Current waveform) The first data set 40 includes a current waveform 50. The current waveform 50 shows the time variation of the ion current I. Figure 2 shows the current waveform 50 of the ion current I. In Figure 2, the horizontal axis represents time t, and the vertical axis represents the value of the ion current I. The unit of time is, for example, [s]. The unit of the ion current I is, for example, [A]. The current waveform 50 is a collection of sets of data for time t and data for the ion current I.

[0024] In the current waveform 50, the ion current I includes the leakage current value IL as the base current value and the peak current value IP. The leakage current value IL corresponds to the value of the ion current I when the ion channel 10 is closed. Even when the ion channel 10 is closed, the ion current I does not become completely zero, but flows a small amount through the ion channel 10 by leaking through the artificial cell membrane 5. For this reason, the leakage current value IL is slightly greater than zero.

[0025] The magnitude of the leakage current IL in the ion current I is determined in proportion to the magnitude of the command voltage applied to the ion channel 10. If the applied command voltage is at a negative potential, the leakage current IL in the ion current I will also be a negative value.

[0026] The peak current value IP is greater than the leakage current value IL in absolute terms. The peak current value IP corresponds to the ion current I when ion channel 10 is open. When ion channel 10 is open, a large amount of ion current I flows through ion channel 10. Therefore, the peak current value IP is greater than the leakage current value IL in absolute terms.

[0027] As shown in Figure 2, the ion current I fluctuates between a leakage current value IL and a peak current value IP. The leakage current value IL is not perfectly constant, but fluctuates randomly within the range of noise σ. The peak current value IP is not perfectly constant, but fluctuates randomly within the range of noise σ.

[0028] When the ion channel 10 changes from closed to open, that is, it takes time for the ion current I to go from the leakage current value IL to the peak current value IP by a first time constant τo. When the ion channel 10 changes from open to closed, that is, it takes time for the ion current I to go from the peak current value IP to the leakage current value IL by a second time constant τc.

[0029] The ionic current I persists at the leakage current value IL for a closed duration Tc. The closed durations Tc at the leakage current values ​​IL for the first to fourth valleys are denoted by Tc1, Tc2, Tc3, and Tc4. The total closed duration Tc for the period in question is the sum of Tc1 to Tc4. That is, Tc = Tc1 + Tc2 + Tc3 + Tc4.

[0030] The ion current I is maintained at a peak current value IP for an open duration To. The open durations To at the peak current values ​​IP of the first to fourth peaks are denoted by To1, To2, To3, and To4. The total open duration To for the period is the sum of To1 to To4. That is, To = To1 + To2 + To3 + To4.

[0031] The open probability Po represents the open state of ion channel 10. The open probability Po is the ratio of the open duration To of ion channel 10 to the given period. The open probability Po can be roughly obtained as To / (Tc+To) (Po=To / (Tc+To)).

[0032] The closure probability Pc represents the open / closed state of ion channel 10. The closure probability Pc is the ratio of the closed duration Tc of ion channel 10 to the given period. The closure probability Pc can be roughly obtained as Tc / (Tc+To) (Pc=Tc / (Tc+To)).

[0033] The sum of the open probability Po and the closed probability Pc is 1 (Po + Pc = 1). If one of the open probability Po or closed probability Pc is determined, the other is also determined. Finding the open probability Po and finding the closed probability Pc are essentially the same thing. The open probability Po will be explained below.

[0034] The period in question includes not only the closed duration Tc and the open duration To, but also the first time constant τo and the second time constant τc. In order to accurately determine the open probability Po, the first time constant τo and the second time constant τc must be allocated to either the closed duration Tc or the open duration To.

[0035] For example, a suitable threshold IA can be set for the ion current I. The time when the ion current I is less than the threshold IA can be allocated to the closed duration Tc, and the time when the ion current I is greater than the threshold IA can be allocated to the open duration To.

[0036] For a human, a threshold IA can be set visually each time, midway between the leakage current value IL and the peak current value IP, for a given current waveform 50. This allows a human to accurately determine the open probability Po, taking into account the first time constant τo and the second time constant τc.

[0037] However, it is difficult for a computer to automatically determine the accurate open probability Po based on the current waveform 50. A computer cannot even recognize the leakage current value IL and the peak current value IP, even if the current waveform 50 is provided. Therefore, the computer cannot appropriately set a threshold IA between the leakage current value IL and the peak current value IP for each given current waveform 50. Furthermore, the influence of noise σ also makes it difficult for the computer to recognize the leakage current value IL and the peak current value IP.

[0038] While it is conceivable to pre-fix the threshold value IA to an appropriate predetermined value, the characteristics of the current waveform 50 vary each time depending on the type of ion channel 10 and the magnitude of the command voltage applied to the ion channel 10. For example, both the leakage current value IL and the peak current value IP may increase or decrease (the entire current waveform 50 may shift upward or downward). Therefore, if the threshold value IA is pre-fixed to an appropriate predetermined value, both the leakage current value IL and the peak current value IP in the current waveform 50 may increase or decrease beyond the threshold value IA, making it impossible to determine an accurate open probability Po.

[0039] Thus, it is difficult for a computer to automatically determine the exact open probability Po based on the current waveform 50.

[0040] (histogram) The first data set 40 includes a histogram 60. The histogram 60 corresponds to the current waveform 50. Specifically, the histogram 60 is converted from the current waveform 50. The histogram 60 shows the frequency F for each value of ion current I. Figure 3 shows the histogram 60 for ion current I. In Figure 2, the horizontal axis represents the ion current I, and the vertical axis represents the frequency F. The unit of frequency F is dimensionless [-]. Frequency F is also a probability.

[0041] Histogram 60 is inherently a bar graph, but by increasing the number of data points, it becomes a curve as shown in Figure 3. In this example, histogram 60 is constructed using a Gaussian distribution. A Gaussian distribution is also called a normal distribution.

[0042] In the histogram 60, a closed-side peak 61 corresponding to the closed-side duration Tc (leakage current IL) is formed on the smaller side (left side) of the ion current I, and an open-side peak 62 corresponding to the open-side duration To (peak current IP) is formed on the larger side (right side) of the ion current I. In the histogram 60, a tail 63 corresponding to the first time constant τo, the second time constant τc, and noise σ is formed between the closed-side peak 61 and the open-side peak 62.

[0043] The closed side peak 61 has a closed area A1. The open side peak 62 has an open area A2. The base 63 has a small base area A3.

[0044] The open probability Po of ion channel 10 is roughly given by A2 / (A1+A2) (Po=A2 / (A1+A2)). The closed probability Pc of ion channel 10 is roughly given by A1 / (A1+A2) (Pc=A1 / (A1+A2)).

[0045] As mentioned above, finding the open probability Po and finding the closed probability Pc are essentially the same thing, so we will explain the open probability Po below.

[0046] To accurately determine the open probability Po, the tail area A3 must be allocated to either the closed area A1 or the open area A2. For example, a suitable threshold IA can be set for the ion current I. When the ion current I is less than the threshold IA, the area can be allocated to the closed area A1, and when the ion current I is greater than the threshold IA, the area can be allocated to the open area A2.

[0047] A human can visually identify the closed-side peaks 61 and the open-side peaks 62 in a given histogram 60 and set a threshold IA in the middle (tail 63) between them each time. This allows a human to accurately determine the open probability Po, taking into account the tail area A3 of the tail 63.

[0048] However, it is difficult for a computer to automatically determine the accurate open probability Po based on the histogram 60. Even with the histogram 60 provided, a computer cannot distinguish between the closed-side peak 61 and the open-side peak 62. Therefore, the computer cannot appropriately set a threshold IA in the intermediate area (tail 63) between the closed-side peak 61 and the open-side peak 62 each time.

[0049] It is conceivable to pre-fix the threshold IA to an appropriate predetermined value, but the appearance of the histogram 60 will vary each time depending on the type of ion channel 10 and the magnitude of the command voltage applied to the ion channel 10. For example, if both the leakage current value IL and the peak current value IP increase or decrease, the entire histogram 60 will shift to the right or to the left. Therefore, if the threshold IA is pre-fixed to an appropriate predetermined value, both the closed-side peak 61 and the open-side peak 62 in the histogram 60 may become larger or smaller than the threshold IA (they may be located to the right or to the left), making it impossible to determine the accurate open probability Po.

[0050] Thus, it is difficult for a computer to automatically determine the exact open probability Po based on the histogram 60.

[0051] (Regression model) Figure 4 shows a regression model M according to the first embodiment. The analysis unit 30 of the ion channel analyzer 1 uses the regression model M. The regression model M is located on an external server 6. Specifically, the regression model M is stored on the external server 6. The analysis unit 30 accesses the regression model M on the external server 6 to input data into the regression model M and to receive data output by the regression model M.

[0052] A regression model M is a supervised machine learning model. Specific examples of regression models M include ridge regression, GDBT (Gradient Boosting Decision Tree), multilayer perceptron (MLP), CNN (Convolutional Neural Network) models, and Transformer models.

[0053] The regression model M is trained as follows: The regression model M uses known first data 40 and known mathematical parameters 70 as training data. The known mathematical parameters 70 correspond to the known first data 40.

[0054] During the training of the regression model M, the known first data 40 is provided as training data to the input layer of the regression model M, and the known mathematical parameters 70 are provided as training data to the output layer of the regression model M.

[0055] For example, a regression model M uses a known current waveform 50 and known mathematical parameters 70 as training data. In this case, the known mathematical parameters 70 correspond to the known current waveform 50. The known current waveform 50 is provided as training data to the input layer of the regression model M, and the known mathematical parameters 70 are provided as training data to the output layer of the regression model M.

[0056] Alternatively, the regression model M uses a known histogram 60 and known mathematical parameters 70 as training data. In this case, the known mathematical parameters 70 correspond to the known histogram 60. Furthermore, the known histogram 60 is provided as training data to the input layer of the regression model M, and the known mathematical parameters 70 are provided as training data to the output layer of the regression model M.

[0057] The mathematical parameter 70 is an input variable corresponding to the first data 40 (current waveform 50 and histogram 60). The known first data 40 (current waveform 50 and histogram 60) is generated by numerical simulation based on the known mathematical parameter 70. Numerous mathematical parameters 70 are set by changing the conditions of the mathematical parameter 70. For each mathematical parameter 70 with different conditions, the corresponding first data 40 (current waveform 50 and histogram 60) is generated. This provides numerous sets of first data 40 (current waveform 50 and histogram 60) and mathematical parameter 70 depending on the conditions. Note that the known first data 40 may be the measured value of the ion current I. In this case, the known mathematical parameter 70 is determined based on the measured value of the ion current I (first data 40). Furthermore, the known first data 40 and mathematical parameter 70 may be generated by numerical simulation, based on measured values, or include both.

[0058] The mathematical parameter 70 contains information about the open / closed state of the ion channel 10. The open / closed state of the ion channel 10 refers to the state of whether the ion channel 10 is open or closed, and the degree to which it is open or closed. Specifically, the mathematical parameter 70 contains information about the open probability Po as the open / closed state of the ion channel 10.

[0059] More specifically, the mathematical parameter 70 includes, as information about the open probability Po of the ion channel 10, the leakage current value IL, the peak current value IP, the first time constant τo from the leakage current value IL to the peak current value IP, the second time constant τc from the peak current value IP to the leakage current value IL, the closed duration Tc at the leakage current value IL, the open duration To at the peak current value IP, the standard deviation of the noise σ, and the cutoff frequency.

[0060] Furthermore, the mathematical parameter 70 may include the open probability Po itself as information about the open probability Po of the ion channel 10.

[0061] Furthermore, regardless of whether known current waveforms 50 or known histograms 60 are used as training data for the input layer, it is preferable to provide these known mathematical parameters 70 (items not in parentheses in Figure 4) to the output layer.

[0062] The mathematical parameter 70 may include, as information regarding the opening probability Po of the ion channel 10, the number of opening / closing events D (the number of times the ion channel 10 opens and closes during the target period), the opening / closing timing E (the time when the ion channel 10 opens and closes), the closing duration Tc for each opening / closing event (e.g., Tc1 to Tc4), and the opening duration To for each opening / closing event (e.g., To1 to To4).

[0063] When using known current waveforms 50 as training data for the input layer, it is preferable to provide these known mathematical parameters 70 (items in parentheses in Figure 4) to the output layer. However, when using known histograms 60 as training data for the input layer, it is not essential to provide these known mathematical parameters 70 (items in parentheses in Figure 4) to the output layer.

[0064] The analysis unit 30 executes the regression model M as follows: The analysis unit 30 inputs the first data 40 related to the ion current I detected by the detection unit 20 into the regression model M, and outputs mathematical parameters 70 corresponding to the input first data 40. The analysis unit 30 accesses the regression model M on the external server 6 and inputs the first data 40 related to the detected ion current I into the regression model M. The analysis unit 30 obtains the mathematical parameters 70 (corresponding to the first data 40) output by the regression model M on the external server 6. The analysis unit 30 outputs the obtained mathematical parameters 70.

[0065] As the mathematical parameter 70 corresponding to the first data 40, information regarding the open probability Po (open / closed state) of the ion channel 10 is output.

[0066] For example, the analysis unit 30 outputs mathematical parameters 70 corresponding to the input current waveform 50 by inputting the current waveform 50 related to the ion current I detected by the detection unit 20 into the regression model M.

[0067] The mathematical parameters 70 corresponding to the current waveform 50 are output as follows: leakage current value IL, peak current value IP, first time constant τo, second time constant τc, closed duration Tc, open duration To, open probability Po itself, standard deviation of noise σ, and cutoff frequency. Furthermore, the number of switching events D, switching timing E, closed duration Tc for each switching event (e.g., Tc1 to Tc4), and open duration To for each switching event (e.g., To1 to To4) are output as mathematical parameters 70 corresponding to the current waveform 50.

[0068] Alternatively, the analysis unit 30 outputs mathematical parameters 70 corresponding to the input histogram 60 by inputting the histogram 60 related to the ion current I detected by the detection unit 20 into the regression model M.

[0069] The mathematical parameters 70 corresponding to the histogram 60 are output as follows: leakage current value IL, peak current value IP, first time constant τo, second time constant τc, closed duration Tc, open duration To, open probability Po itself, standard deviation of noise σ, and cutoff frequency.

[0070] The analysis unit 30 analyzes the state of the ion channel 10 based on the output mathematical parameters 70. Specifically, the analysis unit 30 analyzes the open / closed state of the ion channel 10 based on the output mathematical parameters 70.

[0071] More specifically, the analysis unit 30 determines the open probability Po of the ion channel 10 based on the output mathematical parameters 70. Since specific numerical values ​​are given as mathematical parameters 70, the analysis unit 30 can easily determine the open probability Po.

[0072] For example, the analysis unit 30 calculates the open probability Po based on mathematical parameters 70 corresponding to the histogram 60 (see Figure 3). In the mathematical parameters 70 corresponding to the histogram 60, the leakage current value IL and the peak current value IP are given as specific numerical values. The analysis unit 30 calculates the intermediate value between the leakage current value IL and the peak current value IP and sets this intermediate value as the threshold value IA.

[0073] The analysis unit 30 allocates the area of ​​the portion on the left side (where the ion current I is smaller than the threshold IA) of the histogram 60 to the closed area A1 of the closed peak 61, and the area of ​​the portion on the right side (where the ion current I is larger than the threshold IA) to the open area A2 of the open peak 62. At this time, the tail area A3 of the tail portion 63 is allocated between the closed area A1 and the open area A2. As a result, the analysis unit 30 can easily determine the open probability Po using the formula Po = A2 / (A1+A2).

[0074] Furthermore, the mathematical parameter 70 corresponding to the histogram 60 is given by the regression model M, which provides the open probability Po itself. The analysis unit 30 may directly apply the open probability Po, which is given by the regression model M as the mathematical parameter 70.

[0075] For example, the validity (reliability) of the mathematical parameter 70 obtained from the regression model M can be confirmed by comparing the open probability Po obtained using the threshold IA described above with the open probability Po given as a mathematical parameter 70 by the regression model M.

[0076] Alternatively, the analysis unit 30 may determine the open probability Po based on mathematical parameters 70 corresponding to the current waveform 50 (see Figure 2). Note that when using mathematical parameters 70 corresponding to the current waveform 50, the number of mathematical parameters 70 increases compared to when using mathematical parameters 70 corresponding to the histogram 60 (see Figure 4), making the calculation more complex.

[0077] In the case based on mathematical parameters 70 corresponding to the current waveform 50, as in the case based on mathematical parameters 70 corresponding to the histogram 60, the open probability Po given by the regression model M as mathematical parameter 70 may be applied directly.

[0078] (Evaluation tool for regression models) As shown in Figure 1, the evaluation device 7 (evaluation device for regression model M) is connected to the server device 6. The evaluation device 7 is, for example, a computer. The evaluation device 7 includes, for example, a processor mounted on a circuit board and a memory device that stores software for operating the processor.

[0079] The evaluation device 7 includes an evaluation unit 80. When the regression model M learns, the evaluation unit 80 of the evaluation device 7 evaluates the mathematical parameters 72 (second mathematical parameters, which are estimated values ​​of mathematical parameters) output from the regression model M in accordance with the first data 41 related to the ion current I, using a loss function. Specifically, the evaluation unit 80 considers that the loss of mathematical parameter 72 is at its minimum value (specifically, 0) if the mathematical parameter 71 (first mathematical parameter), which is the input value of the mathematical parameter corresponding to the first data 41 related to the ion current I, satisfies predetermined conditions. The first data 41 becomes the training data for the regression model M as the known first data 40 (it is input to the regression model M). In addition, some or all of the set mathematical parameters 71 become the training data for the regression model M as the known mathematical parameters 70 (it is input to the regression model M). Furthermore, mathematical parameter 72 corresponds to the mathematical parameter 70 in Figure 4.

[0080] As described above, the first data 41 (known first data 40) and mathematical parameters 71 (known mathematical parameters 70) may be generated by numerical simulation, based on measured values, or may include both.

[0081] Here, mathematical parameters 71 and 72 include parameters common to mathematical parameter 70. Specifically, mathematical parameters 71 and 72, like mathematical parameter 70, include the average values ​​of the ion channel's open probability Po, leakage current value IL, peak current value IP, closed duration Tc, and open duration To. Mathematical parameters 71 and 72 also include the command voltage for the ion channel, the ion current width (or conductance) in the ion channel, the sampling frequency of the first data 41, the block size (sec) of the first data 41, the standard deviation of the noise σ, and the cutoff frequency in the first data 41.

[0082] In the following explanation, it will be assumed that the first data 41 includes the current waveforms of two ion channels (ion channels 11 and 12).

[0083] The evaluation unit 80 uses a loss function that includes the following conditions (1) to (3). Specifically, condition (1): The evaluation unit 80 considers the loss of ion current width, switching duration, and cutoff frequency in the mathematical parameter 72 output from the regression model M to be at its minimum value when the opening probability of at least one of ion channels 11 and 12 is at least one of 0% and 100% in the mathematical parameter 71. Condition (2): The evaluation unit 80 considers the loss of leakage current in the mathematical parameter 72 output from the regression model M to be at its minimum value when the opening probability of at least one of ion channels 11 and 12 is 100% in the mathematical parameter 71. Condition (3): The evaluation unit 80 evaluates the loss without distinguishing between the opening probability Po of ion channel 11 and the opening probability Po of ion channel 12.

[0084] Figure 5 shows the current waveform of the ion current I when the open probability of ion channels 11 and 12 is 0%. As shown in Figure 5, when the open probability of ion channels 11 and 12 is 0%, both ion channels 11 and 12 remain closed, so the current waveform of the ion current I does not show any features other than the open probability, leakage current, and noise. In other words, the current waveform of the ion current I does not show any features of ion current width, open / closed duration, or cutoff frequency. For this reason, using condition (1) in the loss function improves the learning accuracy of the regression model M.

[0085] Figure 6 shows the current waveform of the ion current I when the open probability of ion channels 11 and 12 is 100%. As shown in Figure 6, when the open probability of ion channels 11 and 12 is 100%, both ion channels 11 and 12 remain in the open state, so no features other than the open probability and noise appear in the current waveform of the ion current I. In other words, no leakage current features appear in the current waveform of the ion current I. For this reason, the learning accuracy of the regression model M is improved by using condition (2) in the loss function.

[0086] Figure 7 shows the current waveform of the ion current when the open probability of ion channel 11 is 10%, the open probability of ion channel 12 is 80%, and the ion current width (or conductance) of ion channels 11 and 12 are the same. As shown in Figure 7, when the ion current widths of ion channels 11 and 12 are the same, it is not possible to determine which of ion channels 11 or 12 is in the open state. Therefore, by using condition (3) in the loss function, the learning accuracy of the regression model M is improved.

[0087] For example, the loss functions described in specific examples 1-3 below can be used. Note that the following explanation uses the Absolute Mean Error (MAE) as the loss function, but other loss functions can also be used.

[0088] (Specific example of a loss function 1) First, we will explain a specific example of the loss function used by the evaluation unit 80.

[0089] The loss function MAE is expressed as follows:

[0090]

number

[0091] However, n is the number of parameters, y i This is the correct answer.

[0092]

number

[0093] is a predicted value. In this example, y i ,

[0094]

number

[0095] Each of these can be represented by eight parameters (n=8), as shown below.

[0096]

number

[0097] However, istep1 is the ion current width of ion channel 11, istep2 is the ion current width of ion channel 12, ileak is the leakage current, noise is the standard deviation of the noise, fc is the cutoff frequency, dtime is the average value of the open / closed durations of ion channels 11 and 12, pol is the open probability of the ion channel with the lower open probability (low open probability), and poh is the open probability of the ion channel with the higher open probability (high open probability). Since pol is the low open probability and poh is the high open probability, the MAE loss function in this example does not distinguish between the open probabilities of ion channels 11 and 12 in the loss calculation.

[0098] Based on the above, the loss function MAE can be expressed as follows.

[0099]

Number

[0100] And then, the following conditional expressions (a) to (d) are added to the loss function MAE. Conditional expression (a) is when (pol = 0 ∩ poh = 0) ∪ (pol = 1 ∩ poh = 1),

[0101]

Number

[0102] it is. Conditional expression (b) is when (pol = 0) ∩ (0 < poh < 1),

[0103]

Number

[0104] it is. Conditional expression (c) is when (poh = 1) ∩ (0 < pol < 1),

[0105]

Number

[0106] it is. Conditional expression (d) is when poh = 1,

[0107]

Number

[0108] Condition (a) is the condition when the open probability of both ion channels 11 and 12 is either 0% or 100%. Condition (b) is the condition when the open probability of only one of ion channels 11 or 12 is 0%. Condition (c) is the condition when the open probability of only one of ion channels 11 or 12 is 100%. Condition (d) is the condition when the open probability of at least one of ion channels 11 or 12 is 100%. Note that in the above conditions (a) to (c), the low open probability pol and the high open probability poh are shown as percentages.

[0109] By using the above loss function MAE, conditions (1) to (3) are met in the loss function used by the evaluation unit 80.

[0110] (Example 2 of loss functions) Next, we will explain a specific example of the loss function used by the evaluation unit 80.

[0111] The loss function MAE is expressed as follows:

[0112]

number

[0113] However, n is the number of parameters, y i This is the correct answer.

[0114]

number

[0115] is a predicted value. In this example, y i ,

[0116]

number

[0117] Each of these can be represented by seven parameters (n=7), as shown below.

[0118]

Number

[0119] However, pom is the average value of the opening probabilities of ion channels 11 and 12. Since pom is the average value of the opening probabilities of ion channels 11 and 12, in the loss function MAE, the distinction between the opening probabilities of ion channels 11 and 12 is not made in the loss calculation.

[0120] From the above, the loss function MAE is expressed as follows.

[0121]

Number

[0122] And the following conditional expressions (a) to (d) are added to the loss function MAE. Conditional expression (a) is for the case of (pom = 0) ∪ (pom = 1),

[0123]

Number

[0124] is. Conditional expression (b) is for the case of (po1(po2) = 0) ∩ (0 < po2(po1) < 1),

[0125]

Number

[0126] is. Conditional expression (c) is for the case of (po1(po2) = 1) ∩ (0 < po2(po1) < 1),

[0127]

Number

[0128] is. Conditional expression (d) is for the case of (po1 = 1) ∪ (po2 = 1),

[0129]

number

[0130] This is the case where po1 is the probability of ion channel 11 being open, and po2 is the probability of ion channel 12 being open. Condition (a) is the condition when the probability of opening both ion channels 11 and 12 is 0% or 100%. Condition (b) is the condition when the probability of opening only one of ion channels 11 or 12 is 0%. Condition (c) is the condition when the probability of opening only one of ion channels 11 or 12 is 100%. Condition (d) is the condition when the probability of opening at least one of ion channels 11 or 12 is 100%. Note that in the above conditions (a) to (c), the average value of the probability of opening ion channels 11 and 12, pom, the probability of opening ion channel 11 po1, and the probability of opening ion channel 12 po2 are shown as percentages.

[0131] By using the above loss function MAE, conditions (1) to (3) are met in the loss function used by the evaluation unit 80.

[0132] Furthermore, the specific example 2 of the loss function described above is more effective when the open probability po1 of ion channel 11 and the open probability po2 of ion channel 12 are close in value. For example, if the open probability po1 of ion channel 11 and the open probability po2 of ion channel 12 are significantly different, evaluating using the average value pom of the open probabilities of ion channels 11 and 12 may result in multiple solutions being generated, potentially preventing the regression model M from converging. For this reason, the specific example 2 of the loss function is more effective when the open probability po1 of ion channel 11 and the open probability po2 of ion channel 12 are close in value.

[0133] (Example 3 of loss functions) Next, we will explain a specific example 3 of the loss function used by the evaluation unit 80.

[0134] The loss function MAE is expressed as follows:

[0135]

number

[0136] However, n is the number of parameters, y i This is the correct answer.

[0137]

number

[0138] is a predicted value. In this example, y i ,

[0139]

number

[0140] This can be expressed as follows:

[0141]

number

[0142] The loss function MAE is expressed as follows:

[0143]

number

[0144] At this time,

[0145]

number

[0146] It is as follows. S(δ1) is the sum of the difference between the correct value and the predicted value of the opening probability of ion channel 11 and the difference between the correct value and the predicted value of the opening probability of ion channel 12. S(δ2) is the sum of the difference between the correct value of the opening probability of ion channel 11 and the predicted value of ion channel 12 and the difference between the correct value of the opening probability of ion channel 12 and the predicted value of ion channel 12. That is, in the loss function MAE, since the absolute value of the difference between the correct value and the predicted value of ion channels 11 and 12 is minimized, the distinction between the opening probabilities of ion channels 11 and 12 is not made in the loss calculation.

[0147] And the following conditional expressions (a) to (d) are added to the loss function MAE. Conditional expression (a) is for the case of (po1 = 0 ∩ po2 = 0) ∪ (po1 = 1 ∩ po2 = 1),

[0148]

Number

[0149] It is as follows. Conditional expression (b) is for the case of (po1(po2) = 0) ∩ (0 < po2(po1) < 1),

[0150]

Number

[0151] It is as follows. Conditional expression (c) is for the case of (po1(po2) = 1) ∩ (0 < po2(po1) < 1),

[0152]

Number

[0153] It is as follows. Conditional expression (d) is for the case of (po1 = 1) ∪ (po2 = 1),

[0154]

Number

[0155] Condition (a) is the condition when the open probabilities of both ion channels 11 and 12 are either 0% or 100%. Condition (b) is the condition when the open probability of only one of ion channels 11 or 12 is 0%. Condition (c) is the condition when the open probability of only one of ion channels 11 or 12 is 100%. Condition (d) is the condition when the open probability of at least one of ion channels 11 or 12 is 100%. Note that in the above conditions (a) to (c), the open probability po1 of ion channel 11 and the open probability po2 of ion channel 12 are shown as percentages.

[0156] By using the above loss function MAE, conditions (1) to (3) are met in the loss function used by the evaluation unit 80.

[0157] In this embodiment, the first data 41 was described as including the current waveforms of ion channels 11 and 12, but the first data 41 may also include a histogram of the current waveforms of ion channels 11 and 12.

[0158] (Effects of the embodiment) The ion channel evaluation device 7 according to this embodiment includes an evaluation unit 80 that evaluates mathematical parameters 72 (second mathematical parameters), which are estimated values ​​of mathematical parameters output in correspondence with first data 41, from a regression model M to which first data 41 indicating the ion current flowing through the ion channel 11 (first ion channel), using a loss function. The evaluation unit 80 considers that the loss of mathematical parameter 72 is at its minimum value if mathematical parameter 71 (first mathematical parameter), which is the measured value of mathematical parameter 70 corresponding to the first data 41 related to the ion current I, satisfies predetermined conditions.

[0159] Under predetermined conditions, the current waveform and histogram of the ion current I may not show features such as leakage current. When such first data 41 (ion current I) is used as a training model for the regression model M, the accuracy of the output mathematical parameters 70 decreases. Therefore, when training the regression model M using training data, if the measured value of the corresponding mathematical parameter 70 (mathematical parameter 71) in the first data 41 used as training data satisfies predetermined conditions, the accuracy of the mathematical parameters output from the regression model M is improved by minimizing the loss of the estimated value of the mathematical parameter 70 (mathematical parameter 72) output by the regression model M.

[0160] Furthermore, the evaluation unit 80 considers that the losses in the ion current width, switching duration, and cutoff frequency in the mathematical parameters 72 output from the regression model M are at their minimum values ​​if the probability of opening at least one of the ion channels 11 and 12 is at least one of 0% or 100% in the mathematical parameters 71.

[0161] When the open probability of at least one of the ion channels 11 and 12 is 0%, the ion channel remains closed, and therefore, no features other than the open probability, leakage current, and noise appear in the current waveform of the ion current I. In other words, the features of the ion current width, switching duration, and cutoff frequency do not appear in the current waveform of the ion current I. For this reason, when the open probability of at least one of the ion channels 11 and 12 is at least one of 0% or 100% in mathematical parameter 71, the learning accuracy of the regression model M is improved by minimizing the loss of the ion current width, switching duration, and cutoff frequency in mathematical parameter 72 output from the regression model M.

[0162] The evaluation unit 80 considers that the leakage current loss in the mathematical parameter 72 output from the regression model M is at its minimum value when the probability of opening at least one of the ion channels 11 and 12 is 100% in the mathematical parameter 71.

[0163] When the open probability of at least one of the ion channels 11 and 12 is 100%, that ion channel remains open, and therefore, no features other than the open probability and noise appear in the current waveform of the ion current I. In other words, no leakage current features appear in the current waveform of the ion current I. For this reason, when the open probability of at least one of the ion channels 11 and 12 is 100% in mathematical parameter 71, the learning accuracy of the regression model M is improved by minimizing the leakage current loss in mathematical parameter 72 output from the regression model M.

[0164] The evaluation unit 80 evaluates the loss without distinguishing between the open probability of ion channel 11 and the open probability of ion channel 12. For example, if the conductances of ion channels 11 and 12 are the same, it is not possible to determine which of ion channels 11 or 12 is in the open state. Therefore, by evaluating the loss without distinguishing between the open probability of ion channel 11 and the open probability of ion channel 12, the learning accuracy of the regression model M is improved.

[0165] <Other Embodiments> Although this disclosure has been described above with reference to preferred embodiments, this description is not limiting, and various modifications, substitutions, or combinations are, of course, possible.

[0166] The analysis unit 30 may be provided outside the ion channel analyzer 1, rather than being built into the ion channel analyzer 1 main body.

[0167] The evaluation device 7 and the server device 6 may be configured by a single computer.

[0168] In the above embodiment, the ion channel analyzer 1 was applied to an artificial cell membrane 5, but it is not limited to this and may be applied to, for example, a biological cell membrane.

[0169] In the above embodiment, the case where there are two ion channels 10 is illustrated, but the embodiment is not limited to this. There may be one ion channel 10 or three or more. In this case, the current waveform 50 may be a superposition of the currents I flowing through multiple ion channels 10. [Industrial applicability]

[0170] This disclosure is extremely useful and has high industrial applicability because it can be applied to the regression model evaluation device 7 (ion channel analyzer 1). [Explanation of symbols]

[0171] Ionic current IL Leakage Current Value (Base Current Value) IP Peak Current Value IA threshold t time τo First time constant (time constant) τc Second time constant (time constant) To Open Duration (Duration) Tc Closure duration (duration) σ Noise Po Open Probability Pc closed probability F frequency A1 closed area A2 open area A3 Hem Area D Number of opening / closing events E Opening and closing timing 1. Ion channel analyzer 2 Lipid molecules 3 Organic solvents 4 water drops 5 Artificial cell membrane 6. External Servers 7. Evaluation device 10 Ion Channels 11 Ion Channels (First Ion Channel) 12 Ion channels (second ion channels) 20 Detection unit 30 Analysis Department 40 Data 1 50 Current waveform 60 histograms 61 Closed side mountain part 62 Open-sided mountain section 63 Hem 70 Mathematical Parameters 71. Mathematical Parameters (First Mathematical Parameter) 72. Mathematical Parameters (Second Mathematical Parameter) 80 Evaluation Department

Claims

1. The system includes an evaluation unit that evaluates, using a loss function, a second mathematical parameter, which is an estimated value of a mathematical parameter output in correspondence with the first data, from a regression model to which first data representing the first ion current flowing through the first ion channel is input. The evaluation unit is a regression model evaluation device that considers the loss of the second mathematical parameter to be at its minimum value when the first mathematical parameter, which is an input value of a mathematical parameter corresponding to the first data, satisfies predetermined conditions.

2. The regression model evaluation device according to claim 1, wherein the evaluation unit considers the loss of the second mathematical parameter to be at least the minimum value when the opening probability of the ion channel in the first mathematical parameter is 0% and 100%.

3. The regression model evaluation device according to claim 1, wherein the evaluation unit considers the loss in the second mathematical parameter to be at its minimum value when the probability of opening the ion channel is 100% in the first mathematical parameter.

4. The first data includes the second ion current flowing through the second ion channel. The evaluation unit is a regression model evaluation device that evaluates the loss without distinguishing between the opening probability of the first ion channel and the opening probability of the second ion channel.

5. The process includes a first step of evaluating, using a loss function, a second mathematical parameter, which is an estimated value of a mathematical parameter output in correspondence with the first data, from a regression model to which first data representing the first ion current flowing through the first ion channel is input. The first step is a method for evaluating a regression model in which, if the first mathematical parameter, which is an actual measured value of the mathematical parameter corresponding to the first data, satisfies a predetermined condition, the loss of the second mathematical parameter is considered to be at its minimum value.

6. On the computer, The first step involves performing a regression model to which first data representing the first ion current flowing through the first ion channel is input. The second mathematical parameter, which is an estimate of the mathematical parameter output in correspondence with the first data, is evaluated using a loss function. The first step is a regression model evaluation program that, if the first mathematical parameter, which is an actual measured value of the mathematical parameter corresponding to the first data, satisfies predetermined conditions, then considers the loss of the second mathematical parameter to be at its minimum value.

Images