Living volt-ampere analysis method based on generative deep learning model
By applying intelligent algorithms of generative deep learning models in live voltammetry analysis methods, the problem of signal overlapping interference in complex brain neurochemical environments is solved, and high accuracy and high throughput quantitative detection of a variety of neurochemical molecules are achieved.
Patent Information
- Application Number
- CN202510078121.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-01-17
- Publication Date
- 2025-05-16
AI Technical Summary
In situ electrochemical voltammetry analysis method in living organisms faces the problems of signal overlapping interference and low detection accuracy in complex brain neurochemical environments.
Intelligent algorithm based on the generative deep learning model is adopted to collect data through rapid scanning cyclic voltammetry, preprocess data, filter interference signals using the generative deep learning model, and quantitative analysis is performed using the deep learning regression model to achieve simultaneous quantitative detection of multiple neurochemical molecules.
It effectively removes interference from electrochemical signals in the living environment, improves the accuracy and flux of measurement results, and is suitable for quantitative analysis of complex brain neurochemicals.
Smart Images
Figure CN120015177A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of in vivo analysis, and in particular to an in vivo voltammetric analysis method based on a generative deep learning model. Background Art
[0002] At present, the methods for in situ detection of brain neurochemical molecules in the field of in vivo analysis are mainly through electrochemical and fluorescence analysis methods. Among them, electrochemical methods have attracted much attention because they do not require gene transfection and have high sensitivity. However, the complex chemical environment in the brain nervous system poses a huge challenge to the design of analytical methods. In order to accurately and comprehensively describe the process of brain neurochemical changes, it is still necessary to develop analytical detection methods with higher selectivity and higher throughput.
[0003] A major challenge facing in vivo in situ electrochemical analysis methods is the mutual interference of electrochemical signals of different neurochemical substances in the living environment. Taking fast scanning cyclic voltammetry as an example, although this method has been widely used in the detection of dopamine release process in vivo, in complex physiological and pathological processes, when there are changes in the concentration of other neurochemical molecules such as ascorbic acid and ions, obvious interference signals will be generated, resulting in distortion of the voltammetric peak shape and deviation in the measurement results.
[0004] For in situ electrochemical voltammetric data of complex brain neural processes, traditional methods such as electrode material design and measurement condition optimization can no longer meet the analysis needs. Considering that each interference factor has a unique voltammetric characteristic, it is possible to use an intelligent algorithm based on a generative deep learning model to filter out interference signals in the original voltammetric peak shape and simultaneously quantitatively detect multiple different neurochemical molecules, thereby improving the accuracy of the measurement results of the in situ analysis method in vivo and realizing multi-substance high-throughput sensing. Therefore, it is very important to develop an in vivo voltammetric analysis method based on a generative deep learning model. Summary of the invention
[0005] In view of the above-mentioned deficiencies in the prior art, the present invention provides an in vivo voltammetric analysis method based on a generative deep learning model, which solves the problems of signal overlapping interference, low detection accuracy and single detection object in the detection process of the prior art.
[0006] In order to achieve the above-mentioned object of the invention, the technical solution adopted by the present invention is:
[0007] An in vivo voltammetric analysis method based on a generative deep learning model comprises the following steps:
[0008] S1, collect raw measurement data using fast scan cyclic voltammetry;
[0009] S2, preprocessing the original measurement data to obtain the in-situ measurement data to be analyzed;
[0010] S3. Using the trained generative deep learning model to filter out interference signals from the in vivo in-situ measurement data to be analyzed, to obtain the in vivo in-situ measurement data after the interference signals are filtered out;
[0011] S4. Use the trained deep learning regression model to quantitatively analyze the in vivo in situ measurement data after filtering out interference signals to obtain the concentration values of different neurochemical molecules to be tested.
[0012] The beneficial effects of the present invention are:
[0013] 1. The generative deep learning model constructed is used to remove interfering signals in the data to be tested, which helps to solve the problems of ion signal interference, background current drift, potential shift, and overlapping interference of electrochemical voltammetric signals of different neurochemical substances during long-term recording. It has strong adaptability and promotes the widespread application of electrochemical methods in complex brain neurochemistry.
[0014] 2. Semi-supervised training and unsupervised training are used respectively in the model training process, which not only improves the performance and adaptability of the model, but also reduces the cost of data calibration to a certain extent. It has strong applicability and is suitable for the quantitative analysis of a variety of brain neurochemical substances. BRIEF DESCRIPTION OF THE DRAWINGS
[0015] Figure 1 A flow chart of the method proposed by the present invention;
[0016] Figure 2 A schematic diagram of the network structure of a generative deep learning model or a deep learning regression model;
[0017] Figure 3 A typical voltammetric response diagram of the dopamine solution provided in Example 1;
[0018] Figure 4 A typical voltammetric response diagram of the vitamin C solution provided in Example 1;
[0019] Figure 5 The cyclic voltammetric response diagram of the sodium chloride ion concentration change provided in Example 1;
[0020] Figure 6 A typical cyclic voltammogram of the mixed solution provided in Example 1;
[0021] Figure 7 A typical voltammogram of the interference signal prediction result provided in Example 1;
[0022] Figure 8 A typical voltammogram of the data after the interference signal provided in Example 1 is subtracted;
[0023] Fig. 9This is a diagram showing the quantitative analysis results of various neurochemical substances provided in Example 1;
[0024] Fig.10 The peak current-time curves of dopamine voltammetric response of different solutions provided for comparative experiments;
[0025] Fig.11 Current-time curves at the same potential in mixed solutions provided for comparative experiments;
[0026] Fig.12 The results of dopamine quantitative analysis using the traditional method provided for comparative experiments;
[0027] Fig.13 The in vivo in situ data and the typical voltammogram of the ion interference signal predicted by the model provided in Example 2;
[0028] Fig.14 A typical voltammogram after filtering out the interference signal provided in Example 2;
[0029] Fig.15 This is the multi-substance analysis result of the in vivo in situ measurement data provided in Example 2. DETAILED DESCRIPTION
[0030] The specific implementation modes of the present invention are described below so that those skilled in the art can understand the present invention. However, it should be clear that the present invention is not limited to the scope of the specific implementation modes. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the attached claims, these changes are obvious, and all inventions and creations utilizing the concept of the present invention are protected.
[0031] like Figure 1 As shown, the in vivo voltammetric analysis method based on a generative deep learning model proposed in the present invention comprises the following steps:
[0032] S1. Use fast scan cyclic voltammetry to collect raw measurement data, where the neurochemical molecules to be measured are the electrically active neurotransmitters (such as dopamine), electrically active neuromodulators (such as vitamin C), ions, etc. involved in the brain nerve process under study;
[0033] S2, preprocessing the original measurement data to obtain the in-situ measurement data to be analyzed;
[0034] S3. Using the trained generative deep learning model to filter out interference signals from the in vivo in-situ measurement data to be analyzed, to obtain the in vivo in-situ measurement data after the interference signals are filtered out;
[0035] S4. Use the trained deep learning regression model to quantitatively analyze the in vivo in situ measurement data after filtering out interference signals to obtain the concentration values of different neurochemical molecules to be tested.
[0036] The interference signal filtering described in step S3 is specifically as follows: inputting the in-situ measurement data of the living body to be analyzed into the encoder of the generative deep learning model to obtain the output result of the encoder of the generative deep learning model; using the current difference at the potential reversal point in the in-situ measurement data of the living body to be analyzed as the characteristic parameter of the interference signal, and inputting it together with the output result of the encoder of the generative deep learning model into the decoder of the generative deep learning model to obtain the interference signal volt-ampere prediction result output by the decoder of the generative deep learning model; converting the volt-ampere prediction result output by the decoder of the generative deep learning model according to the normalization coefficient to obtain the interference signal volt-ampere peak prediction result; subtracting the corresponding interference signal volt-ampere peak prediction result from the in-situ measurement data of the living body to be analyzed to obtain the in-situ measurement data of the living body after filtering out the interference signal.
[0037] Among them. Based on electrochemical expertise, the current difference at the potential reversal point in the in-situ measurement data to be analyzed is used as the characteristic parameter of the interference signal. According to theoretical understanding, the current difference at the potential reversal point is only related to the change in ion concentration and has nothing to do with other substances involved in the redox process. The current at other potentials may have multiple signals superimposed. Therefore, this parameter is suitable for use as a characteristic signal indicating the degree of ion interference for model establishment.
[0038] The quantitative analysis described in step S4 is specifically as follows: the in vivo in situ measurement data after filtering out the interference signal is normalized and then input into the deep learning regression model to obtain the output result of the deep learning regression model; the output result of the deep learning regression model is converted according to the normalization coefficient to obtain the concentration values of different neurochemical molecules.
[0039] like Figure 2 As shown, both the generative deep learning model and the deep learning regression model are deep learning models with a variational autoencoder structure, including an encoder and a decoder; wherein the encoder includes a first convolutional layer, a first ReLu layer, a first pooling layer, a second convolutional layer, a second ReLu layer, a second pooling layer, a third convolutional layer, a third ReLu layer, a third pooling layer, a first fully connected layer, a second fully connected layer, and a third fully connected layer connected in sequence, the input of the first convolutional layer is the input of the encoder, and the output of the third fully connected layer is the output of the encoder; the decoder includes a first deconvolutional layer, a fourth ReLu layer, a second deconvolutional layer, a fifth ReLu layer, a third deconvolutional layer, a sixth ReLu layer, a fourth deconvolutional layer, and a seventh ReLu layer connected in sequence, the input of the first deconvolutional layer is the input of the decoder, and the output of the seventh ReLu layer is the output of the decoder; the output end of the encoder is connected to the input end of the decoder.
[0040] Among them, the training of the generative deep learning model is unsupervised training, which includes the following sub-steps:
[0041] A1. Use fast scanning cyclic voltammetry to collect raw measurement data, and pre-process the collected data to obtain a training set;
[0042] A2. Input the training set into the encoder of the generative deep learning model to obtain the output result of the encoder of the generative deep learning model;
[0043] A3. The current difference at the potential reversal point in the training set data is used as the interference signal feature value and input into the decoder of the generative deep learning model together with the output result of the encoder of the generative deep learning model to obtain the interference signal prediction result;
[0044] A4. Calculate the loss value and use the Adam algorithm to update the parameters of the generative deep learning model. When the loss value converges, the training ends, otherwise return to step A2.
[0045] The loss value The expression is:
[0046]
[0047] in, are intermediate parameters; α1 and β1 are hyperparameters, x1 is the input data of the encoder of the generative deep learning model, g(x1) is the output result of the decoder of the generative deep learning model, y1 is the expected output related to x; g′(x1) and y1′ are the differential data of g(x1) and y1 respectively, and MSE(·) is the mean square error function; σ i1 is the variance of the i-th variable output by the encoder of the generative deep learning model, μ i1 is the mean of the i-th variable output by the encoder of the generative deep learning model, K1 is the number of variables output by the encoder of the generative deep learning model, and ln represents the natural logarithm.
[0048] The training of deep learning regression model is semi-supervised training, which includes the following sub-steps:
[0049] B1. Collect in vitro calibration data using fast scan cyclic voltammetry; the in vitro calibration data is electrochemical voltammetry data recorded by placing the microelectrode in a flow electrochemical cell and using fast scan cyclic voltammetry in a solution environment containing different concentrations of the neurochemical molecule to be tested;
[0050] B2. Preprocess the in vitro calibration data and combine it with the in vivo in situ measurement data after filtering out interference signals to obtain a training set;
[0051] B3. Input the training set into the deep learning regression model to obtain the concentration values of different neurochemical molecules;
[0052] B4. Calculate the loss value and use the Adam algorithm to update the parameters of the deep learning regression model. When the loss value converges, the training ends, otherwise return to step B3.
[0053] The loss value The expression is:
[0054]
[0055] Among them, α2, β2 are hyperparameters, x2 is the encoder input data of the deep learning regression model, g(x2) is the decoder output result of the deep learning regression model; g′(x2) and x2′ are the differential data of g(x2) and x2 respectively; γ is a hyperparameter, f(x2) is the encoder output result of the deep learning regression model; y2 is the concentration mark value of the in vitro calibration data; σ i2 is the variance of the i-th variable output by the encoder of the deep learning regression model, μ i2 is the mean of the i-th variable output by the encoder of the deep learning regression model, K2 is the number of variables output by the encoder of the deep learning regression model, and ln represents the natural logarithm; and All are intermediate parameters.
[0056] In this embodiment, the additional reconstruction loss (L Re ) and KL loss (L KL ) These two parts mainly describe the data distribution, that is, to reduce the prediction error of the model for the training set while making the network model learn the characteristics of the training set data. This method can improve the generalization ability of the model for location data outside the training set and prevent overfitting.
[0057] The data preprocessing includes preprocessing of the original measurement data and preprocessing of the in vitro calibration data. The preprocessing of the original measurement data is specifically as follows: using the voltammetric data within 5 to 30 seconds before stimulating the organism to release neurochemical molecules as the background signal to perform subtraction with the original measurement data to obtain the original measurement data after the background is subtracted; filtering out the data with ion concentration changes in the original measurement data after the background is subtracted and normalizing the filtered data to obtain the in-situ measurement data to be analyzed;
[0058] The in vitro calibration data were preprocessed as follows: the average value of the electrochemical voltammetric data recorded when the solution composition in each group of in vitro calibration data was close to the cerebrospinal fluid was used as the background signal, and the voltammetric data measured under other concentration conditions in the same group of data were subtracted according to the potential to obtain the in vitro calibration data after the background was subtracted; the data with ion concentration changes in the in vitro calibration data after the background was subtracted were screened out and normalized so that their value range was between -1 and 1, and the normalization coefficient was recorded to obtain the preprocessed in vitro calibration data.
[0059] Embodiment 1:
[0060] The training data was collected according to the following process: the carbon fiber microelectrode was placed in a flow electrochemical cell, artificial cerebrospinal fluid was pumped into the electrode, and rapid scanning cyclic voltammetry data recording was started. Artificial cerebrospinal fluid containing 1, 2, 3, 4, and 5 μmol / L dopamine was pumped into the electrode every 1 min, and rapid scanning cyclic voltammetry data was recorded in real time. The typical voltammetric response was measured as follows: Figure 3 As shown. Artificial cerebrospinal fluid containing 50, 100, 150, and 200 μmol / L vitamin C was pumped to the electrode at 1-min intervals, and the fast scan cyclic voltammetry data was recorded in real time. The typical voltammetric response was measured as follows: Figure 4 As shown. Artificial cerebrospinal fluid containing 1, 5, 15, 50, 100, and 200 mmol / L sodium chloride was pumped to the electrode every 1 minute, and the fast scan cyclic voltammetry data was recorded in real time. The typical voltammetric response was measured as shown in Figure 5 shown.
[0061] The fast scan voltammetric data recorded in the above process are preprocessed as a training set, and the training set is input into the generative deep learning model for training.
[0062] The test set was constructed in the following way: a mixed solution containing 100 mmol / L vitamin C, 2 μmol / L dopamine, and 50 mmol / L sodium chloride was pumped to the electrodes in a flow electrochemical cell to simulate the release of multiple neurochemical substances, and fast scan cyclic voltammetry data was recorded in real time. The typical voltammetric response was measured as follows Figure 6 The test set data is normalized and input into the generative deep learning model to obtain the ion interference signal data as shown in Figure 7 As shown in the figure, the final result of the generative deep learning model output after filtering out the interference signal is as follows Figure 8 The result after filtering out the interference signal is input into the deep learning regression model to obtain the quantitative analysis result of the test set, as shown in Fig. 9 shown.
[0063] In order to verify that the method proposed in the present invention has higher measurement accuracy, a comparative experiment is carried out:
[0064] The rapid scan cyclic voltammetry data were collected in the flow electrochemical cell in the same manner as in Example 1 for collecting the training set data. The voltammetry data recorded by the electrode under the action of different concentrations of dopamine solutions were analyzed by conventional methods, background subtraction was performed, and the peak current-time curve was extracted, such as Fig.10 The current-time curve of the electrode at the same potential in a mixed solution containing 100mmol / L vitamin C, 2μmol / L dopamine, and 50mmol / L sodium chloride is shown in Fig.11 Substituting this data into the conventional linear calibration model, the concentration conversion result can be obtained as shown in Fig.12 shown.
[0065] From the comparison of experimental data, it can be seen that the dopamine response measured using the conventional method has a large deviation from the actual concentration of the added substance (2 μmol / L).
[0066] Embodiment 2:
[0067] Microelectrodes were implanted in the striatum of SD rats, and multiple sets of in vivo in situ voltammetric data were recorded before and after local high potassium stimulation. The voltammetric response curves during the stimulation release process were obtained by subtracting the pre-stimulation data as the background. Typical voltammetric response data are shown in Figure 2. Fig.13 As shown in the middle curve a.
[0068] The in vivo voltammetric data were input into the generative deep learning model to obtain the ion interference data prediction results, where the telecommunication data such as Fig.13 The result after the interference signal is filtered out can be obtained by subtracting the in vivo in situ data from the model prediction results. Typical data are shown in Fig.14 shown.
[0069] The result after filtering out the interference signal is input into the deep learning regression model to obtain the quantitative analysis result, such as Fig.15 This shows that the deep learning regression model can effectively distinguish and quantify the voltammetric signals of different neurochemical substances.
[0070] In summary, the present invention can not only accurately detect the quantitative neurochemical molecules in the living brain on the basis of filtering out interference signals, but also realize multi-substance analysis.
Claims
1. A method for in vivo voltammetric analysis based on a generative deep learning model, characterized in that: The following steps are involved: S1, collect raw measurement data using fast scan cyclic voltammetry; S2, preprocessing the original measurement data to obtain the in-situ measurement data to be analyzed; S3. Using the trained generative deep learning model to filter out interference signals from the in vivo in-situ measurement data to be analyzed, to obtain the in vivo in-situ measurement data after the interference signals are filtered out; S4. Use the trained deep learning regression model to quantitatively analyze the in vivo in situ measurement data after filtering out interference signals to obtain the concentration values of different neurochemical molecules to be tested.
2. The in vivo voltammetric analysis method based on a generative deep learning model according to claim 1, characterized in that: Both the generative deep learning model and the deep learning regression model are deep learning models with a variational autoencoder structure, including an encoder and a decoder; wherein the encoder includes a first convolutional layer, a first ReLu layer, a first pooling layer, a second convolutional layer, a second ReLu layer, a second pooling layer, a third convolutional layer, a third ReLu layer, a third pooling layer, a first fully connected layer, a second fully connected layer, and a third fully connected layer connected in sequence, the input of the first convolutional layer is the input of the encoder, and the output of the third fully connected layer is the output of the encoder; the decoder includes a first deconvolutional layer, a fourth ReLu layer, a second deconvolutional layer, a fifth ReLu layer, a third deconvolutional layer, a sixth ReLu layer, a fourth deconvolutional layer, and a seventh ReLu layer connected in sequence, the input of the first deconvolutional layer is the input of the decoder, and the output of the seventh ReLu layer is the output of the decoder; the output end of the encoder is connected to the input end of the decoder.
3. The in vivo voltammetric analysis method based on a generative deep learning model according to claim 2, characterized in that: The training of the generative deep learning model is unsupervised training, which includes the following sub-steps: A1. Using fast scanning cyclic voltammetry to collect training set data, and preprocessing the collected data to obtain a training set; A2. Input the training set into the encoder of the generative deep learning model to obtain the output result of the encoder of the generative deep learning model; A3. The current difference at the potential reversal point in the training set data is used as the interference signal feature value and input into the decoder of the generative deep learning model together with the output result of the encoder of the generative deep learning model to obtain the interference signal prediction result; A4. Calculate the loss value and use the Adam algorithm to update the parameters of the generative deep learning model. When the loss value converges, the training ends, otherwise return to step A2. The loss value The expression is: in, are intermediate parameters; α1 and β1 are hyperparameters, x1 is the input data of the encoder of the generative deep learning model, g(x1) is the output result of the decoder of the generative deep learning model, y1 is the expected output related to x; g′(x1) and y1′ are the differential data of g(x1) and y1 respectively, and MSE(·) is the mean square error function; σ i1 is the variance of the i-th variable output by the encoder of the generative deep learning model, μ i1 is the mean of the i-th variable output by the encoder of the generative deep learning model, K1 is the number of variables output by the encoder of the generative deep learning model, and ln represents the natural logarithm.
4. The in vivo voltammetric analysis method based on a generative deep learning model according to claim 2, characterized in that: The training of deep learning regression model is semi-supervised training, which includes the following sub-steps: B1. Collect in vitro calibration data using fast scan cyclic voltammetry; The in vitro calibration data are electrochemical voltammetry data recorded by placing the microelectrode in a flow electrochemical cell and using rapid scanning cyclic voltammetry in a solution environment containing different concentrations of the neurochemical molecules to be tested. B2. Preprocess the in vitro calibration data and combine it with the in vivo in situ measurement data after filtering out interference signals to obtain a training set; B3. Input the training set into the deep learning regression model to obtain the concentration values of different neurochemical molecules; B4. Calculate the loss value and use the Adam algorithm to update the parameters of the deep learning regression model. When the loss value converges, the training ends, otherwise return to step B3. The loss value The expression is: Among them, α2, β2 are hyperparameters, x2 is the encoder input data of the deep learning regression model, g(x2) is the decoder output result of the deep learning regression model; g′(x2) and x2′ are the differential data of g(x2) and x2 respectively; γ is a hyperparameter, f(x2) is the encoder output result of the deep learning regression model; y2 is the concentration mark value of the in vitro calibration data; σ i2 is the variance of the i-th variable output by the encoder of the deep learning regression model, μ i2 is the mean of the i-th variable output by the encoder of the deep learning regression model, K2 is the number of variables output by the encoder of the deep learning regression model, and ln represents the natural logarithm; and All are intermediate parameters.
5. The in vivo voltammetric analysis method based on a generative deep learning model according to claim 1, characterized in that: The original measurement data are electrochemical voltammetric data recorded in vivo using a microelectrode and rapid scanning cyclic voltammetry.
6. The in vivo voltammetric analysis method based on a generative deep learning model according to claim 1, 3 or 4, characterized in that: The pre-processing is specifically as follows: Preprocess the raw measurement data: The voltammetric data before stimulating the organism to release neurochemical molecules is used as a background signal to perform subtraction with the original measurement data to obtain the original measurement data after the background is deducted; the data with ion concentration changes in the original measurement data after the background is deducted are screened out and the screened data are normalized to obtain the in-situ measurement data to be analyzed; Preprocessing of in vitro calibration data: The average value of the electrochemical voltammetric data recorded when the solution composition in each group of in vitro calibration data was close to cerebrospinal fluid was taken as the background signal, and the voltammetric data measured under other concentration conditions in the same group of data were subtracted according to the potential to obtain the in vitro calibration data after deducting the background; the data with changes in ion concentration in the in vitro calibration data after deducting the background were screened out and normalized to obtain the preprocessed in vitro calibration data.
7. The in vivo voltammetric analysis method based on a generative deep learning model according to claim 2, characterized in that: The interference signal filtering in step S3 is specifically as follows: Input the in-situ measurement data of the living body to be analyzed into the encoder of the generative deep learning model to obtain the output result of the encoder of the generative deep learning model; use the current difference at the potential reversal point in the in-situ measurement data of the living body to be analyzed as the characteristic parameter of the interference signal, input it together with the output result of the encoder of the generative deep learning model into the decoder of the generative deep learning model to obtain the volt-ampere prediction result of the interference signal output by the decoder of the generative deep learning model; convert the volt-ampere prediction result output by the decoder of the generative deep learning model according to the normalization coefficient to obtain the volt-ampere peak prediction result of the interference signal; The in-vivo in-situ measurement data to be analyzed is subtracted from the corresponding interference signal volt-ampere peak prediction result to obtain the in-vivo in-situ measurement data after filtering out the interference signal.
8. The in vivo voltammetric analysis method based on a generative deep learning model according to claim 1, characterized in that: The quantitative analysis in step S4 is specifically as follows: The in vivo in situ measurement data after filtering out interference signals is normalized and then input into the deep learning regression model to obtain the output result of the deep learning regression model; the output result of the deep learning regression model is converted according to the normalization coefficient to obtain the concentration values of different neurochemical molecules.