CN109722466B - Method for rapidly detecting strain producing carbapenemase by using AI model - Google Patents

Abstract

The invention discloses a method for rapidly detecting carbapenemase-producing strains by using an AI model. By using a bromothymol blue solution, the OD value of a mixed solution is measured, and finally the AI model is used to analyze the changing trend of the OD value of the mixed solution. Judging whether the strain can produce carbapenemase; the invention reduces errors by detecting the trend of the OD value of the strain changing with time; by detecting the OD value of the strain solution, it can quickly judge whether the strain produces carbapenemase in a very short time Enzyme, which greatly saves time; for weak enzyme-producing strains, the AI model is used to analyze the change of the OD value of the strain solution and quickly judge; import the OD value into the LSTM model, and you can immediately get whether the strain can produce enzymes and The enzyme-producing ability is strong and weak, and the number of positive and negative enzyme-producing strains in a batch can be intuitively seen, and the enzyme-producing results can be directly derived, without the need to manually record the experimental results, which is efficient and convenient.

Description

A method for rapid detection of carbapenemase-producing strains using an AI model

technical field

The invention relates to the research fields of artificial intelligence and biotechnology, in particular to a method for rapidly detecting carbapenemase-producing strains by using an AI model.

Background technique

Carbapenems are currently one of the last lines of defense in the clinical treatment of Gram-negative bacterial infections, especially for Carbapenemase-producing Enterobacteriaceae (CRE). The drug resistance rate is increasing year by year. Therefore, developing a method for accurate and rapid detection of carbapenemase is crucial for clinical treatment.

In 2012, Nordmann et al. reported that the Carba-NP method detects whether the strain produces carbapenemase by using the color change of the phenol red indicator under different conditions. The carbapenemase-producing strain has hydrolysis of its β-lactam ring. When the pH value decreases, carboxyl groups will be generated, and the color of the phenol red indicator will change from red to yellow with the decrease of pH, so as to determine whether the strain produces carbapenemase. It is not large enough, and for some strains with weak enzyme production, it is not very good to judge whether it produces carbapenemase. In 2013, J.Pires et al. reported the Blue-Carba method, which uses the wide discoloration range of bromothymol blue (pH=6.0-7.6) to change from blue to yellow, making it easier to judge whether the strain produces carbon blue. The disadvantage is that: the strains with weak enzyme production cannot be well judged, and the strain growth time of 2h is also required; in 2017, Jeremy Surre et al. detected the OD value of the corresponding solution at 558nm by using a UV spectrophotometer. The change of OD value is used to judge whether the strain produces enzyme. The disadvantage is that it still takes 2 hours to completely detect whether the strain produces carbapenemase by using UV spectrophotometer; Lisong Shen et al. detected the carbon production of blood culture by MALDI-TOF. Enterobacteriaceae strains of penemase, the method is a hydrolysis test method of imipenem with Mysterella as the substrate, and the logQ value is calculated according to the peak intensity of imipenem and its hydrolyzed products to distinguish whether the strain is not Carbapenemase production, this method can be used for rapid identification and drug susceptibility test report of blood samples, but 41.7% of Enterobacteriaceae strains of NDM-1 subtype were judged negative, so the sensitivity of the results is not very high .

SUMMARY OF THE INVENTION

The main purpose of the present invention is to overcome the shortcomings and deficiencies of the prior art, and to provide a method for rapidly detecting carbapenemase-producing strains by using an AI model, which can automatically analyze whether the strains produce carbapenemase and whether the strains produce carbapenemase-positive strains. Quantity, and then quickly determine whether the strain produces carbapenemase and the strength of the strain's enzyme-producing ability, and can also quickly determine whether the strain produces carbapenemase for some weak enzyme-producing strains, and export the experimental results directly to the computer. , reducing the time for manual recording of experimental results. It solves the problems of the insensitivity of human visual observation of color changes and the long time for enzyme production.

The object of the present invention is achieved through the following technical solutions:

A method for rapidly detecting carbapenemase-producing strains using an AI model, comprising the following steps:

Use the Blue-Carba method to prepare the solution solution; namely, add imipenem to the bromothymol blue solution, then add ZnSO4 solution, and finally adjust the pH of the solution to obtain the solution solution;

Select wild strains and prepare bacterial suspension with PBS solution: isolate and purify the samples on MacConkey agar medium, then streaked on LB agar medium, cultivate for a certain period of time, and identify the strain species by MALDI-TOF; Scrape a circle of bacterial moss into the PBS solution, vortex to mix, and adjust the OD value of the solution to obtain a bacterial suspension;

Mix the solution solution and the bacterial suspension; use a discharge gun to draw the solution solution and add it to the 96-well plate, and then draw the bacterial suspension and mix it with the solution solution;

Use an automatic microplate reader to detect the OD value of the mixed solution; put the 96-well plate of the mixed solution into the automatic microplate reader, set the temperature, detect the OD value of the mixed solution, and export the original data;

Select the AI model, and perform model training and verification to obtain the optimal AI model, import the raw data of the OD value of the mixed solution into the optimal AI model, and cyclically calculate through the LSTM cyclic neurons, output the feature vector, and then go through a fully connected network. , and analyze the results of carbapenemase production by the strain;

Further, the described model training and verification, the specific process is:

Use 80% of the data as the learning sample to train the model, and use 20% of the data as the validation data;

The model training adopts cross entropy as the loss function, adopts the Adam gradient descent method, and uses the training data to perform multiple iterations to optimize the model parameters to minimize the loss function;

S1, the calculation process of LSTM:

LSTM takes the OD value of each time point as input, and the number of time points is denoted as n. When LSTM calculates, it traverses each time point and uses it as input to perform a loop calculation; set the number of neurons of LSTM to k; set t is the current point in time;

LSTM has two transmission states, one C _t and one h _t ; the current input x _t of LSTM and the h _t-1 passed from the previous state are concatenated to obtain four states:

f _t =σ(W _f *[h _t-1 ,x _t ]+b _f ),

i _t =σ(W _i *[h _t-1 ,x _t ]+ _bi ),

o _t =σ(W _o *[h _t-1 ,x _t ]+b _o ),

为输入数据，o_t为控制当前输出；W_f为忘记门控的权重矩阵，b_f为忘记门控的偏置；W_i为记忆门控的权重矩阵，b_i为记忆门控的偏置；W_c为输入数据的权重矩阵，b_c为输入数据的偏置；W_o为控制当前输出的权重矩阵，b_o为控制当前输出的偏置；σ为激活函数；Among them, f _t is the forget gate, it _is the memory gate,

is the input data, o _t is the control current output; W _f is the weight matrix of forget gate, b _f is the bias of forget gate; Wi is the weight matrix of memory gate _{, b i is} the bias of memory gate ; W _c is the weight matrix of the input data, b _c is the bias of the input data; W _o is the weight matrix that controls the current output, _bo is the bias that controls the current output; σ is the activation function;

Different states have different weight matrices W and bias b, and the optimization parameters need to be trained to obtain the training parameters; the weight matrix is a (k+1)*k matrix, which is multiplied by [h _t-1 ,x _t ] to obtain a length of A vector of k, that is, k neurons; where h _t-1 is a vector of length k; x _t is the OD value at time t; the initial values of all weight matrices and biases are based on a mean of 0 and a variance of 1. The normal distribution is randomly generated;

则C_t由忘记门控和记忆门控为依据进行更新：f _t is the forget gate, which is the previous state C _t-1 for the object; i _t is the memory gate, which is the current input for the object

Then C _t is updated according to forget gate and memory gate:

o _t controls the current output, while h _t is updated according to o _t :

h _t =o _t *tanh(C _t ),

After the LSTM calculation is completed, through a fully connected network, the final output node is z, indicating whether the enzyme is produced or not:

z=W _z *h _n +b _z ;

Among them, W _z is the weight matrix of the fully connected network, b _z is the bias of the fully connected network, and h _n is the final output of the above LSTM;

S2, the calculation process of the loss function:

Use cross entropy as the loss function, record cross entropy as LOSS, and the calculation process is as follows:

为预测值，经softmax计算得到，为c*2的矩阵；y为真实值，其同样为c*2的矩阵，对于每个样本，若产酶，则有向量[1,0]，若不产酶，则有向量[0,1]；y^T为矩阵y的转置；

为单位矩阵；Among them, c is the sample size; z is the matrix formed by all samples after the calculation in the previous step, each sample is calculated to obtain a vector, the vector length is 2, which means enzyme production or non-enzyme production, and c samples get the size is a matrix of c*2; exp is an exponential function with the base of natural constant e; the softmax function is used to convert the range of the input value to between 0 and 1 to simulate probability;

is the predicted value, which is calculated by softmax and is a matrix of c*2; y is the real value, which is also a matrix of c*2. For each sample, if enzyme is produced, there will be a vector [1,0], if not If the enzyme is produced, there is a vector [0,1]; y ^T is the transpose of the matrix y;

is the unit matrix;

S3. The calculation process of the training algorithm:

The parameter optimization algorithm Adam is used as the training algorithm, and the parameters are set as follows: learning rate α, α=0.001; exponential decay rate β ₁ , β ₁ =0.9 for first-order moment estimation; exponential decay rate β ₂ , β ₂ for first-order moment estimation =0.999; parameters ε, ε=10e-8; training times threshold THRESHOLD, THRESHOLD=2000;

The parameter optimization algorithm Adam is aimed at the loss function. The loss function takes all the training parameters as independent variables, that is, all weight matrices and biases; the gradient of the loss function is calculated ▽ _θ F _t (θ _t-1 );

Assuming that the current training times is t, the calculation process of the parameter optimization algorithm Adam is as follows:

g _t =▽ _θ F _t (θ _t-1 ),

m _t =β ₁ *m _t-1 +(1-β ₁ )*g _t ,

为二阶矩估计；α为学习率；β₁为一阶矩估计的指数衰减率；β₂为二阶矩估计的指数衰减率；ε为非常小的数，其为了防止在实现中除以零；m_t，v_t为指数加权移动平均；

为偏差修正；θ_t-1为上一轮训练完成的参数集合；θ_t为此轮训练完成的参数集合；where g _t is the first-order moment estimate,

is the second-order moment estimation; α is the learning rate; β1 is the exponential decay rate of the _first -order moment estimation; β2 is the exponential decay rate of the _second -order moment estimation; zero; m _t , v _t are exponentially weighted moving averages;

is the bias correction; θ _t-1 is the parameter set completed in the previous round of training; θ _t is the parameter set completed in this round of training;

When the number of training reaches THRESHOLD, the training is completed;

The verification data adopts the method of cross-validation, repeats the training and verification in multiple times, takes the average value of the verification indicators as the quantitative display of the model performance, and obtains the standard deviation of the accuracy rate of the model;

Further, the AI model includes: decision tree model, AdaBoost model, SVM model, LSTM model

Further, the optimal AI model is an LSTM model;

Further, the bromothymol blue solution is 0.4%, PH=6.0;

Further, in the solution solution, the imipenem content is 3mg/mL; in the solution solution, the ^Zn concentration is 0.1mmol/mL; The pH of the final adjustment solution is 7.0;

Further, the wild strain is one of Escherichia coli, Pseudomonas putida, Klebsiella pneumoniae, Citrobacter freundii, Enterobacter cloacae, and Enterobacter myxa;

Further, the culturing for a certain period of time, specifically, culturing at 37°C for 16h-24h;

Further, the specific configuration process of the bacterial suspension is as follows: scrape a bacterial lawn with a 10uL inoculation ring into a 500uL PBS solution, the pH of the PBS solution is 7.4, vortex and mix, and adjust the OD value of the bacterial suspension at a wavelength of 600nm. Between 1.3OD-1.7OD;

Further, the OD value of the detection mixed solution is specifically: put the 96-well plate into the automatic microplate reader, set the temperature of the microplate reader to 37°, and set the temperature of the microplate reader at 615nm and 720nm every 5 minutes. The wavelength is out to detect the OD value of the mixed solution, the detection time is 1h, and the original data of the OD value of the mixed solution is derived.

Compared with the prior art, the present invention has the following advantages and beneficial effects:

1, the present invention greatly reduces the error caused by the change of the color of the solution observed by the naked eye by detecting the trend of the OD value of the strain changing with time, and has higher objectivity;

2. In the present invention, it is only necessary to import the OD value into the LSTM model, and it is possible to immediately obtain whether the strain can produce enzymes and the strength of the enzyme production ability, and can intuitively see the number of positive strains producing enzymes in a batch and the enzyme production capacity. The number of negatives is efficient and convenient;

3. The present invention only detects the change of the OD value at two time points to obtain the result of whether the strain can produce carbapenemase, which can save time;

4. The present invention intuitively sees the number of positive and negative enzyme-producing strains in a batch, and can directly derive the enzyme-producing results of a batch of strains without manually recording the experimental results, which is efficient and convenient;

5. The present invention can quickly obtain the result of whether the weak enzyme-producing strain produces carbapenemase by detecting the change of the OD value at two time points.

Description of drawings

Fig. 1 is a kind of method flow chart of the present invention that utilizes AI model to rapidly detect the method for producing carbapenemase strain;

Fig. 2 is the test result diagram of decision tree model in the described embodiment of the present invention;

Fig. 3 is the test result diagram of AdaBoost model in the described embodiment of the present invention;

Fig. 4 is the test result diagram of SVM model in the described embodiment of the present invention;

Fig. 5 is the test result diagram of LSTM model in the described embodiment of the present invention;

Fig. 6 is the accuracy and standard deviation contrast figure of four mixes in the described embodiment of the present invention;

Fig. 7 is the LSTM model structure diagram in the described embodiment of the present invention;

Fig. 8 is the LSTM training curve diagram in the described embodiment of the present invention;

9 is a schematic diagram of 10-fold cross-validation in the embodiment of the present invention;

Fig. 10 is a training process diagram of 8 groups of time points in the embodiment of the present invention;

FIG. 11 is a comparison chart of the accuracy rate of training results of 8 groups of time points in the embodiment of the present invention;

Fig. 12 is the variation trend diagram of OD value of LSTM model analysis strain over time in the embodiment of the present invention;

FIG. 13 is a schematic diagram of the enzyme production results obtained by analyzing the OD value of the strain by the LSTM model in the embodiment of the present invention.

Detailed ways

The present invention will be described in further detail below with reference to the embodiments and the accompanying drawings, but the embodiments of the present invention are not limited thereto.

Example:

A method for rapidly detecting carbapenemase-producing strains using an AI model, as shown in Figure 1, includes the following steps:

Step 1: Prepare solutionA solution:

Using the Blue-Carba method: add imipenem to the prepared 0.4% bromothymol blue solution (pH=6.0), so that the final drug content of imipenem in the solution is 3 mg/mL, and then add ZnSO4 solution , the final concentration of Zn ²⁺ in the solution was 0.1 mmol/mL, and the pH of the solution was finally adjusted to 7.0.

Step 2: Prepare bacterial suspension with PBS solution:

Strains: The tested strains are from wild strains isolated in this experiment, including Escherichia coli, Pseudomonas putida, Klebsiella pneumoniae, Citrobacter freundii, Enterobacter cloacae, and Enterobacter myxa. The samples were first isolated and purified on MacConkey agar medium, then streaked on LB agar medium, cultured at 37°C for 16-24h, and the strains were identified by MALDI-TOF. The negative control strain ATCC25922 was kept in the laboratory.

Preparation of bacterial suspension: scrape a loop of bacterial lawn with a 10uL inoculation ring into 500uL PBS (pH=7.4) solution, vortex and mix, and adjust the OD value of the bacterial suspension at 600nm wavelength between 1.3OD-1.7OD.

Step 3: Mix the bacterial suspension with the solution

Pipette 100ul of solutionA solution into the 96-well plate with a drain gun, and then pipette 100ul of bacterial suspension and mix with solutionA solution.

Step 4: Put the 96-well plate into the automatic microplate reader to detect the OD value of the mixed solution:

Put the 96-well plate into which the mixed solution was added in the previous step into the automatic microplate reader, set the temperature of the microplate reader to 37°, and detect the OD value of the mixed solution at the wavelengths of 615nm and 720nm every 5min, detect for 1h, and derive the mixed solution. Raw data of solution OD value.

Step 5: Select the AI model, and perform model training and verification to obtain the optimal AI model, import the raw data of the OD value of the mixed solution into the optimal AI model, select it for cyclic calculation, output the feature vector, and then go through a full Connect the network and analyze the results of carbapenemase production by strains

Design of LSTM model:

In order to more comprehensively verify the effectiveness of the analysis AI model in enzyme production discrimination, four common intelligent models were selected in this experiment: decision tree model, AdaBoost model, SVM model, and LSTM model. The reason for choosing a decision tree is that it is highly explanatory and can train an easy-to-understand model. The SVM is a commonly used algorithm in the small sample classification algorithm, which is suitable for the sample situation of this experiment. AdaBoost is one of the representatives of ensemble learning, which can be understood as a comprehensive decision-making model that integrates multiple decision trees. LSTM is a representative of deep learning time series models, and it has outstanding effects in mining sample time series features. More specifically, the decision tree model uses the CART classification regression tree algorithm. The AdaBoost model used 50 decision trees for comprehensive decision making. The SVM model uses a Gaussian kernel function to deal with nonlinear data. The LSTM model used a single-layer LSTM with 128 neurons. The above model settings are the best choices obtained by repeatedly validating and adjusting on the training set.

The experiment collected the OD value data of the enzyme production experiment of 517 strains at 13 time points within 1 hour, and used 80% of the data as the learning sample of the AI model to train the model to achieve a high accuracy. Build the final model. The remaining 20% of the data is used as the final data to verify the accuracy of the model. In order to further accurately evaluate the performance of the model, the experiment uses the method of cross-validation, repeats the training and verification for many times, and takes the average value of the verification indicators as the quantitative display of the model performance. The standard deviation of the rate to assess the stability of the model.

Among them, Figure 2 shows the test results of the decision tree model. The accuracy and standard deviation are 0.9766 and 0.0218 respectively. During the verification process, the number of samples with wrong decision-making is 12. It can be seen from the samples that identify the wrong decision tree. The time series relationship leads to a high error rate; Figure 3 shows the test results of the AdaBoost model. The accuracy and standard deviation are 0.9883 and 0.0157 respectively. During the verification process, the number of wrong samples is 6, which is twice as much as the decision tree; Figure 4 shows the test results of the SVM model. The accuracy and standard deviation are 0.9902 and 0.01312, respectively. SVM obtains the same level of effect as AdaBoost, and only detects one more correct sample than AdaBoost, and the number of wrong samples is 5; Figure 5 shows In the test results of the LSTM model, the accuracy and standard deviation are 0.9941 and 0.01251, respectively. It can be seen that the accuracy has increased by an order of magnitude. During the verification process, the number of wrong samples is only 3; Figure 6 shows the accuracy and standard deviation of the 4 models. Compared.

Since the data is time series data, LSTM is used as the experimental model. LSTM is a recurrent neural network that retains useful long-term information and removes useless short-term information through a threshold mechanism to realize the mining of time series information.

The OD value at each moment is input into the cyclic neural module of LSTM in turn. After cyclic calculation, LSTM outputs the feature vector, and then passes through a fully connected network to finally obtain the positive result of enzyme production.

The detailed structure diagram of the LSTM model is shown in Figure 7.

Model training and validation

The experiment collected the original OD value data of the enzyme production experiment of 517 strains at 13 time points within 1 hour, and used 80% of the data as the learning sample of the AI model to train the model to achieve high accuracy. to build the final model. The remaining 20% of the data is used as the final data to verify the accuracy of the model.

In the training plan, the Adam gradient descent function is used as the model parameter optimizer, the crossentropy is used as the loss function, the gradient truncation threshold is 1.0, the learning rate is 0.001, and the number of iterations is 2000.

Let x be the input OD value, that is, the training data, W is the parameter of the LSTM model, y is the output result, y' is the label, that is, the correct result, the training process is as follows:

The specific process of model training and verification is as follows:

Use 80% of the data as the learning sample to train the model, and use 20% of the data as the validation data;

The model training adopts cross entropy as the loss function, adopts the Adam gradient descent method, and uses the training data to perform multiple iterations to optimize the model parameters to minimize the loss function;

S1, the calculation process of LSTM:

LSTM takes the OD value of each time point as input, and the number of time points is denoted as n. When LSTM calculates, it traverses each time point and uses it as input to perform a loop calculation; set the number of neurons of LSTM to 8; set t is the current point in time;

LSTM has two transmission states, one C _t and one h _t ; the current input x _t of LSTM and the h _t-1 passed from the previous state are concatenated to obtain four states:

f _t =σ(W _f *[h _t-1 ,x _t ]+b _f ),

i _t =σ(W _i *[h _t-1 ,x _t ]+ _bi ),

o _t =σ(W _o *[h _t-1 ,x _t ]+b _o ),

为输入数据，o_t为控制当前输出；W_f为忘记门控的权重矩阵，b_f为忘记门控的偏置；W_i为记忆门控的权重矩阵，b_i为记忆门控的偏置；W_c为输入数据的权重矩阵，b_c为输入数据的偏置；W_o为控制当前输出的权重矩阵，b_o为控制当前输出的偏置；σ为激活函数；Among them, f _t is the forget gate, it _is the memory gate,

is the input data, o _t is the control current output; W _f is the weight matrix of forget gate, b _f is the bias of forget gate; Wi is the weight matrix of memory gate _{, b i is} the bias of memory gate ; W _c is the weight matrix of the input data, b _c is the bias of the input data; W _o is the weight matrix that controls the current output, _bo is the bias that controls the current output; σ is the activation function;

Different states have different weight matrices W and bias b, and the optimization parameters need to be trained to obtain the training parameters; the weight matrix is a (k+1)*k matrix, which is multiplied by [h _t-1 ,x _t ] to obtain a length of A vector of k, that is, k neurons; where h _t-1 is a vector of length k; x _t is the OD value at time t; the initial values of all weight matrices and biases are based on a mean of 0 and a variance of 1. The normal distribution is randomly generated;

则C_t由忘记门控和记忆门控为依据进行更新：f _t is the forget gate, which is the previous state C _t-1 for the object; i _t is the memory gate, which is the current input for the object

Then C _t is updated according to forget gate and memory gate:

o _t controls the current output, while h _t is updated according to o _t :

h _t =o _t *tanh(C _t ),

After the LSTM calculation is completed, through a fully connected network, the final output node is z, indicating whether the enzyme is produced or not:

z=W _z *h _n +b _z ;

Among them, W _z is the weight matrix of the fully connected network, b _z is the bias of the fully connected network, and h _n is the final output of the above LSTM;

S2, the calculation process of the loss function:

Use cross entropy as the loss function, record cross entropy as LOSS, and the calculation process is as follows:

为预测值，经softmax计算得到，为c*2的矩阵；y为真实值，其同样为c*2的矩阵，对于每个样本，若产酶，则有向量[1,0]，若不产酶，则有向量[0,1]；y^T为矩阵y的转置；

为单位矩阵；Among them, c is the sample size; z is the matrix formed by all samples after the calculation in the previous step, each sample is calculated to obtain a vector, the vector length is 2, which means enzyme production or non-enzyme production, and c samples get the size is a matrix of c*2; exp is an exponential function with the base of natural constant e; the softmax function is used to convert the range of the input value to between 0 and 1 to simulate probability;

is the predicted value, which is calculated by softmax and is a matrix of c*2; y is the real value, which is also a matrix of c*2. For each sample, if enzyme is produced, there will be a vector [1,0], if not If the enzyme is produced, there is a vector [0,1]; y ^T is the transpose of the matrix y;

is the unit matrix;

S3, the calculation process of the training algorithm

The parameter optimization algorithm Adam is used as the training algorithm, and the parameters are set as follows: learning rate α, α=0.001; exponential decay rate β ₁ , β ₁ =0.9 for first-order moment estimation; exponential decay rate β ₂ , β ₂ for first-order moment estimation =0.999; parameters ε, ε=10e-8; training times threshold THRESHOLD, THRESHOLD=2000;

The parameter optimization algorithm Adam is aimed at the loss function. The loss function takes all the training parameters as independent variables, that is, all weight matrices and biases; the gradient of the loss function is calculated ▽ _θ F _t (θ _t-1 );

Assuming that the current training times is t, the calculation process of the parameter optimization algorithm Adam is as follows:

g _t =▽ _θ F _t (θ _t-1 ),

m _t =β ₁ *m _t-1 +(1-β ₁ )*g _t ,

为二阶矩估计；α为学习率；β₁为一阶矩估计的指数衰减率；β₂为二阶矩估计的指数衰减率；ε为非常小的数，其为了防止在实现中除以零；m_t，v_t为指数加权移动平均；

为偏差修正；θ_t-1为上一轮训练完成的参数集合；θ_t为此轮训练完成的参数集合；where g _t is the first-order moment estimate,

is the second-order moment estimation; α is the learning rate; β1 is the exponential decay rate of the _first -order moment estimation; β2 is the exponential decay rate of the _second -order moment estimation; zero; m _t , v _t are exponentially weighted moving averages;

is the bias correction; θ _t-1 is the parameter set completed in the previous round of training; θ _t is the parameter set completed in this round of training;

When the number of training reaches THRESHOLD, the training is completed; the training curve is shown in Figure 8.

Model validation uses the method of cross-validation to repeat the training and validation in multiple times, and takes the average value of the validation indicators as the quantitative display of the model performance. stability. Figure 9 is a schematic diagram of 10-fold cross-validation, that is, each validation uses one tenth of the different data in the training set as validation, and the rest as training.

The final result of cross-validation is:

Accuracy: 0.994174757282

Standard deviation: 0.0125143666203

Accuracy: 0.997347480106

Regression: 0.994708994709

Model Validation Results Analysis

The final test results of the LSTM model are shown in Figure 6, which are presented in terms of accuracy, standard deviation, and samples with detected errors.

LSTM supplementary experiment

In the above experiment, when all the OD value data of 13 time points were used, the correct rate of the LSTM model was very high, so the experiment further tried to use 0min and 5min, 10min, 15min, 20min, 25min, 30min, 35min, 40min, respectively. The OD values of the 8 batches of two time points at 45min are used as the input of the LSTM neural network in the form of two-tuples. Finally, 8 corresponding LSTM models are obtained, and the training process is shown in Figure 10.

Using the verification data to analyze the recognition effect of LSTM, the results are shown in Figure 11. It can be seen from Figure 10 that the accuracy rates corresponding to different inputs are all above 92%. Among them, the accuracy rate of the LSTM model with the OD value of 0min and 5min as input is 92%, which is the lowest among all models, and the accuracy rates of the other models are all The accuracy rate of the LSTM model with OD values of 0min and 35min as input is 98.6%, which is the highest among all models.

After the time is greater than 10min, the accuracy of any two time points is above 95%, the accuracy of 0min-20min is 97.5%, and the accuracy of the two time points of 0min-35min is up to 99%. Therefore, the computer model is used. It can be quickly judged whether the strain produces carbapenemase by analyzing the OD value at any time point after 0min and 10min. This method greatly reduces the manpower, material resources and time for judging carbapenem-resistant strains in clinical treatment. cost. The horizontal axis in the figure represents the number of training iterations, and the vertical axis represents the error of the model, which is represented by the value of the cross entropy loss function.

At the beginning of the experiment, the OD value of the strain at 13 time points within 1 hour was analyzed, and the final accuracy compared with Blue-Carba was 99.41%, indicating that the AI model was used to analyze the OD value of the strain at a specific wavelength. The method of judging whether the strain is a carbapenemase-positive strain is feasible, and has extremely high accuracy and high efficiency.

LSTM result analysis

After importing the data into the LSTM model, through the in-depth analysis of the LSTM model, the trend of changes in OD among strains can be immediately and intuitively seen. The OD value of the strain solution gradually decreases with time, and the OD value of the positive control (NDM-5) is basically From the beginning, it was between 0.2OD-.05OD, and the curve was a stable straight line. The OD value of the negative control strain was around 1.5OD, and the OD value of the other enzyme-producing strains gradually decreased from 1.5OD to about 0.2OD. Basically in 20min, it is possible to roughly judge whether the strain produces carbapenemase. And the OD value change curve of strains with different enzyme production intensity is obviously different, the OD value change curve of VIM-2 drug resistance gene is significantly slower than that of NDM-1 and NDM-5, and the strains carrying NDM-1 and VIM-4 at the same time The change of the OD value of the strain was significantly slower than that of other strains carrying only a single drug resistance gene, and it has been showing a slow upward trend since then, which is worthy of careful study. Therefore, the enzyme-producing ability of the strain can also be judged by the change of the OD value of the strain. As shown in Figure 12, and the change trend of different OD values also intuitively indicates the strength of the strain's ability to produce carbapenemase.

As shown in Figure 13, the LSTM model can detect the combination of multiple binary time points, so in the experiment, it can be determined whether the strain produces carbapenemase by detecting the change of the OD value of the strain at any two time points, and The number of enzyme-producing strains in a batch of bacteria can be directly obtained.

The above-mentioned embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited by the above-mentioned embodiments, and any other changes, modifications, substitutions, combinations, The simplification should be equivalent replacement manners, which are all included in the protection scope of the present invention.

Claims (6)

1. a method of using AI model to rapidly detect and produce carbapenemase strain, is characterized in that, comprises the following steps: Use the Blue-Carba method to prepare solutionA solution; namely, add imipenem to the bromothymol blue solution, then add ZnSO4 solution, adjust the pH of the solution to 7.0, and obtain solutionA solution; Select wild strains and prepare bacterial suspension with PBS solution: isolate and purify the samples on MacConkey agar medium, then streak and cultivate on LB agar medium, identify strain species by MALDI-TOF; use 10ul inoculation loop Scrape the semicircular bacteria moss into the PBS solution, vortex and mix, and adjust the OD value of the solution to obtain a bacterial suspension; the wild strains include Escherichia coli, Pseudomonas putida, Klebsiella pneumoniae, and citric acid Freund bacteria, Enterobacter cloacae, Enterobacter myxa; Mix the solutionA solution and the bacterial suspension: add the solutionA solution to the 96-well plate, and then draw the bacterial suspension and mix it with the solutionA solution; Use an automatic microplate reader to detect the OD value of the mixed solution: put the 96-well plate of the mixed solution into the automatic microplate reader, set the detection temperature and detection time, and detect the OD value of the mixed solution at the wavelengths of 615nm and 720nm, respectively. and export raw data; Select an AI model, and perform model training and validation, The specific process of model training and verification is as follows: Use 80% of the data as the learning sample to train the model, and use 20% of the data as the validation data; The model training adopts cross entropy as the loss function, adopts the Adam gradient descent method, and uses the training data to perform multiple iterations to optimize the model parameters to minimize the loss function; S1, the calculation process of LSTM: LSTM takes the OD value of each time point as input, and the number of time points is denoted as n. When LSTM calculates, it traverses each time point and uses it as input to perform a loop calculation; set the number of neurons of LSTM to k; set t is the current point in time; LSTM has two transmission states, one C _t and one h _t ; the current input x _t of LSTM and the h _t-1 passed from the previous state are concatenated to obtain four states: f _t =σ(W _f *[h _t-1 ,x _t ]+b _f ), i _t =σ(W _i *[h _t-1 ,x _t ]+ _bi ),

o _t =σ(W _o *[h _t-1 ,x _t ]+b _o ), Among them, f _t is the forget gate, it _is the memory gate,

is the input data, o _t is the control current output; W _f is the weight matrix of forget gate, b _f is the bias of forget gate; Wi is the weight matrix of memory gate _{, b i is} the bias of memory gate ; W _c is the weight matrix of the input data, b _c is the bias of the input data; W _o is the weight matrix that controls the current output, _bo is the bias that controls the current output; σ is the activation function; Different states have different weight matrices W and bias b, and the optimization parameters need to be trained to obtain the training parameters; the weight matrix is a (k+1)*k matrix, which is multiplied by [h _t-1 ,x _t ] to obtain a length of A vector of k, that is, k neurons; where h _t-1 is a vector of length k; x _t is the OD value at time t; the initial values of all weight matrices and biases are based on a mean of 0 and a variance of 1. The normal distribution is randomly generated; f _t is the forget gate, which is the previous state C _t-1 for the object; i _t is the memory gate, which is the current input for the object

Then C _t is updated according to forget gate and memory gate:

o _t controls the current output, while h _t is updated according to o _t : h _t =o _t *tanh(C _t ), After the LSTM calculation is completed, through a fully connected network, the final output node is z, indicating whether the enzyme is produced or not: z=W _z *h _n +b _z ; Among them, W _z is the weight matrix of the fully connected network, b _z is the bias of the fully connected network, and h _n is the final output of the above LSTM; S2, the calculation process of the loss function: Use cross entropy as the loss function, record cross entropy as LOSS, and the calculation process is as follows:

Among them, c is the sample size; z is the matrix formed by all samples after the calculation in the previous step, each sample is calculated to obtain a vector, the vector length is 2, which means enzyme production or non-enzyme production, and c samples get the size is a matrix of c*2; exp is an exponential function with the base of natural constant e; the softmax function is used to convert the range of the input value to between 0 and 1 to simulate probability;

is the predicted value, which is calculated by softmax and is a matrix of c*2; y is the real value, which is also a matrix of c*2. For each sample, if enzyme is produced, there will be a vector [1,0], if not If the enzyme is produced, there is a vector [0,1]; y ^T is the transpose of the matrix y;

is the unit matrix; S3. The calculation process of the training algorithm: The parameter optimization algorithm Adam is used as the training algorithm, and the parameters are set as follows: learning rate α, α=0.001; exponential decay rate β ₁ , β ₁ =0.9 for first-order moment estimation; exponential decay rate β ₂ , β ₂ for first-order moment estimation =0.999; parameters ε, ε=10e-8; training times threshold THRESHOLD, THRESHOLD=2000; The parameter optimization algorithm Adam is aimed at the loss function. The loss function takes all the training parameters as independent variables, that is, all weight matrices and biases; the gradient of the loss function is calculated.

Assuming that the current training times is t, the calculation process of the parameter optimization algorithm Adam is as follows:

m _t =β ₁ *m _t-1 +(1-β ₁ )*g _t ,

where g _t is the first-order moment estimate,

is the second-order moment estimation; α is the learning rate; β1 is the exponential decay rate of the _first -order moment estimation; β2 is the exponential decay rate of the _second -order moment estimation; zero; m _t , v _t are exponentially weighted moving averages;

is the bias correction; θ _t-1 is the parameter set completed in the previous round of training; θ _t is the parameter set completed in this round of training; When the number of training reaches THRESHOLD, the training is completed; The verification data adopts the method of cross-validation, repeats the training and verification in multiple times, takes the average value of the verification indicators as the quantitative display of the model performance, and obtains the standard deviation of the accuracy rate of the model; The optimal AI model was obtained as the LSTM model, and the original data of the OD value of the mixed solution was imported into the above-mentioned optimal AI model, and the results of carbapenemase production by the strain were obtained by analysis. 2 . A method for rapidly detecting a carbapenemase-producing strain by using an AI model according to claim 1 , wherein the bromothymol blue solution is 0.4%, PH=6.0. 3 . 3. a kind of method that utilizes AI model to rapidly detect and produce carbapenemase strain according to claim 1, is characterized in that, in described solutionA solution, imipenem content is 3mg/mL; Described solutionA solution , the Zn ²⁺ concentration was 0.1 mmol/mL. 4 . The method for rapidly detecting carbapenemase-producing strains by using an AI model according to claim 1 , wherein the culturing is specifically culturing at 37° C. for 16h-24h. 5 . 5. a kind of method that utilizes AI model to rapidly detect and produce carbapenemase strain according to claim 1, is characterized in that, described bacterial suspension, concrete configuration process is: scrape a circle of bacteria with the inoculation ring of 10uL Moss to 500uL PBS solution, the pH of PBS solution is 7.4, vortex to mix, adjust the OD value of bacterial suspension at 600nm wavelength between 1.3OD-1.7OD. 6. a kind of method that utilizes AI model to rapidly detect and produce carbapenemase strain according to claim 1, is characterized in that, the OD value of described detection mixed solution is specially: will add the 96-well plate of mixed solution Put it into the automatic microplate reader, set the temperature of the microplate reader to 37°, and detect the OD value of the mixed solution at the wavelengths of 615nm and 720nm every 5 minutes. The detection time is 1h, and derive the original data of the OD value of the mixed solution.

Images