Preparation method of ochratoxin A monoclonal antibody conjugate

Abstract

The present invention relates to the technical field of antibody-drug conjugates, and in particular to a method for preparing an ochratoxin A monoclonal antibody conjugate, comprising step S01: obtaining preparation process data and measured control parameter data in a preparation process at preset time intervals, and obtaining a real-time preparation process coefficient based on the preparation process data; step S02: obtaining a preparation time ratio Q / M, inputting the component ratio data of the preparation raw materials, the real-time preparation parameters and the preparation time ratio Q / M into a pre-constructed first machine learning model for predicting the preparation process coefficient, predicting the preparation process coefficient at time P, and recording it as the predicted preparation process coefficient. The method for preparing an ochratoxin A monoclonal antibody conjugate provided by the present invention performs real-time adaptive adjustment on the preparation parameters of the preparation process according to the preparation process coefficient in the preparation process, thereby avoiding the situation of over-coupling or under-coupling when preparing the antibody conjugate within a given time.

Description

Technical Field

[0001] The invention relates to the technical field of antibody-coupled drug preparation, in particular to a method for preparing an ochratoxin A monoclonal antibody conjugate. Background Art

[0002] 6-Atomic Arm-Red Microsphere-Ochratoxin A Monoclonal Antibody Conjugate is an immunoassay system used in biomedical and immunological research, which combines red microspheres (usually fluorescent or magnetic microspheres), monoclonal antibodies (against ochratoxin A, OTA) and atomic arm technology. The conjugate has high specificity and sensitivity and is widely used in immunoassay, immunoadsorption, immunoassay, biosensors and in vitro diagnostics.

[0003] A Chinese patent with authorization publication number CN117180449B discloses a method for preparing an antibody-drug conjugate, comprising: (i) reacting an antibody or an antigen-binding fragment thereof with a reducing agent in a buffer to reduce interchain disulfide bonds of the antibody or the antigen-binding fragment thereof, wherein the reducing agent is tris(2-carboxyethyl)phosphine hydrochloride;

[0004] (ii) reacting the linker-payload with the antibody or antigen-binding fragment thereof having a thiol group obtained in step (i);

[0005] Wherein, the reaction temperature in step (i) is 10.5°C to 14°C, the reaction time is 2 to 3 hours, and the ratio of the amount of the reducing agent to the antibody or its antigen-binding fragment is 2:1 to 3:1; the reaction temperature in step (ii) is 10.5°C to 14°C, the reaction time is 2 to 3 hours, and the ratio of the amount of the linker-payload to the antibody or its antigen-binding fragment is 4:1 to 6:1.

[0006] As disclosed in the prior art above, the existing preparation of antibody-drug conjugates generally controls the preparation process based on established control parameters (temperature, pH value, concentration of cross-linking agent and preparation time). However, in the actual preparation process, other factors (antibody activity, antibody uniformity) may affect the preparation process. If the preparation is always carried out with established control parameters, over-coupling or under-coupling may occur, thereby affecting the stability and effect of the final product. Summary of the invention

[0007] In order to solve the above problems, the present invention provides a method for preparing an ochratoxin A monoclonal antibody conjugate.

[0008] The present invention adopts the following technical scheme, a method for preparing an ochratoxin A monoclonal antibody conjugate, which is applied to the process of combining an anti-ochratoxin A monoclonal antibody with a microsphere that has been conjugated with ochratoxin A, comprising the following steps:

[0009] Step S01: obtaining component ratio data of the preparation raw materials, and obtaining preparation process data and measured control parameter data in the preparation process at preset time intervals during the process of preparing the antibody conjugate using the preparation raw materials under predetermined preparation parameters, and obtaining a real-time preparation process coefficient based on the preparation process data;

[0010] Step S02: obtaining the preparation time ratio Q / M, inputting the component ratio data of the preparation raw materials, the real-time preparation parameters and the preparation time ratio Q / M into the pre-built first machine learning model for predicting the preparation progress coefficient, predicting the preparation progress coefficient at time P, and recording it as the predicted preparation progress coefficient;

[0011] Step S03: obtaining a real-time preparation process coefficient, and judging whether the antibody conjugate prepared by the preparation raw material within the predetermined total preparation time M of the antibody conjugate meets the preparation process standard according to the real-time preparation process coefficient; if so, continuing to control the preparation reaction device according to the measured control parameter data;

[0012] Step S04: If not, the preparation process coefficient, the preparation time ratio and the measured control parameter data are input into a second machine learning model pre-configured for feedback correction of the control parameters, to obtain the corrected preparation parameters, and the preparation reaction is controlled based on the corrected preparation parameters;

[0013] Step S05: Control the preparation of antibody conjugates according to the measured control parameters or the modified preparation parameters.

[0014] As a further description of the above technical solution: the preparation process data includes: the binding efficiency, binding amount and protein concentration of the conjugate;

[0015] The method for obtaining the binding efficiency and binding amount of the conjugate comprises:

[0016] Add OTA solution of known concentration into the cuvette and measure its absorbance at 332 nm, which is recorded as A. 332（OTA） ;

[0017] By sampling, the coupled antibody-OTA solution was added to the cuvette, and its absorbance values ​​at 332nm and 280nm were measured respectively, which were recorded as A 332（偶联物） and A 280（偶联物）；

[0018] The conjugate binding efficiency and binding amount were calculated based on the formula.

[0019] As a further description of the above technical solution: the method for obtaining the protein concentration includes:

[0020] Use a protein standard solution (such as BSA or other standard protein) of known concentration to construct a calibration curve;

[0021] By taking samples, the absorbance of the samples at 280 nm was measured using a UV-Vis spectrophotometer;

[0022] Calculate protein concentration based on the calibration curve and absorbance.

[0023] As a further description of the above technical solution: the method for obtaining the real-time preparation process coefficient based on the preparation process data is to perform weighted calculation on the protein concentration, conjugate binding efficiency and binding amount to obtain the real-time preparation process coefficient.

[0024] As a further description of the above technical solution: the method for constructing the first machine learning model includes:

[0025] Initialize the first machine learning model structure. The first machine learning model structure adopts an MLP type multi-layer forward network structure with one input layer, two hidden layers, and one output layer; the input layer includes the first input layer, the second input layer, and the third input layer. The number of nodes in the first input layer of the first input layer is 1, corresponding to the composition ratio data of the prepared raw materials; the number of nodes in the second input layer is 3, corresponding to the reaction temperature, reaction pH value, and concentration of the cross-linking agent of the real-time control parameters; the number of nodes in the third input layer is 1, corresponding to the preparation time ratio; the hidden layer includes the first hidden layer and the second hidden layer. The number of nodes in the first hidden layer of the first hidden layer is 128, and the Relu function is used as the activation function; the number of nodes in the second hidden layer of the second hidden layer is 64, and the Relu function is used as the activation function; the output layer is the first output layer, and the number of nodes in the first output layer of the first output layer is 1, which is the predicted preparation process coefficient.

[0026] After initializing the first machine learning model structure, the first machine learning model is trained using the ingredient ratio data of the preparation raw materials, the real-time preparation parameters and the preparation time ratio Q / M, and the preparation process coefficients corresponding to the ingredient ratio data of the preparation raw materials, the real-time preparation parameters and the preparation time ratio Q / M. The ingredient ratio data of the preparation raw materials, the real-time preparation parameters and the preparation time ratio Q / M, and the preparation process coefficients corresponding to the ingredient ratio data of the preparation raw materials, the real-time preparation parameters and the preparation time ratio Q / M are obtained from a database. The database records the data of each preparation of the 6-atomic arm-red microsphere-ochratoxin A monoclonal antibody conjugate, which is used to optimize the model. The optimizer is Adam, the loss function is the MSE loss function, the batch size is set to 32, the number of iterations is 200 rounds, and the training is ended when the loss function converges, indicating that the training is completed.

[0027] As a further description of the above technical solution: the first machine learning model training method includes:

[0028] The ingredient ratio data of the preparation raw materials, the real-time preparation parameters and the preparation time ratio Q / M are converted into a corresponding set of feature vectors.

[0029] The ingredient ratio data, real-time preparation parameters and preparation time ratio Q / M of the preparation raw materials are taken as inputs of the machine learning model, and the machine learning model takes a group of preparation process coefficients corresponding to the ingredient ratio data, real-time preparation parameters and preparation time ratio Q / M of each group of preparation raw materials as outputs, and takes the preparation process coefficients actually corresponding to the ingredient ratio data, real-time preparation parameters and preparation time ratio Q / M of each group of preparation raw materials as prediction targets, and takes minimizing the loss function value of the machine learning model as training target; training stops when the loss function value of the machine learning model is less than or equal to the preset target loss value.

[0030] As a further description of the above technical solution: the method for judging whether the antibody conjugate prepared by the preparation raw materials within the total preparation time M of the established antibody conjugate meets the preparation process standard according to the real-time preparation process coefficient includes:

[0031] A preparation process difference threshold is preset, and the real-time preparation process coefficient is subtracted from the predicted preparation process coefficient to obtain an absolute value, which is recorded as the process loss value. When the process loss value is less than or equal to the preset preparation process difference threshold, it indicates that the preparation process standard is met; conversely, when the process loss value is greater than the preset preparation process difference threshold, it indicates that the preparation process standard is not met.

[0032] As a further description of the above technical solution: the preparation time ratio Q / M is generated based on the total preparation time M of the antibody conjugate and the start time, and the preparation time Q is obtained by obtaining the current time and the start time, and the preparation time Q is divided by the total preparation time M.

[0033] As a further description of the above technical solution: the second machine learning model is a reinforcement learning model, using a Q network structure, and the construction steps are as follows:

[0034] Step 1: Design state representation dimensions:

[0035] The design state representation dimensions include design pH state discreteness, design temperature state discreteness and design cross-linking agent concentration state discreteness;

[0036] The pH state is discrete as follows: (5.5-7] (6-7] (7-7.5], using One-hot coding; the temperature state is discrete as follows: (25°C-35°C] (30°C-35°C] (30°C-37°C], using One-hot coding; the concentration state of the cross-linking agent is discrete as follows: (0.1mM-0.8mM] (0.5mM-0.8mM] (0.5mM-1mM], the state representation dimension is One-hot coding, and is the sum of the pH state One-hot dimension, the temperature state One-hot dimension, and the cross-linking agent concentration state One-hot dimension.

[0037] Step 2: Design the action representation dimension:

[0038] The design action representation dimensions include a design pH adjustment action range, a design temperature adjustment action range, and a design cross-linking agent concentration adjustment action range;

[0039] The pH adjustment action range is: -0.5, -0.2, 0, 0.2, 0.5, using One-hot encoding; the temperature adjustment action range is: -5°C, -2°C, 0°C, 2°C, 5°C, using One-hot encoding; the cross-linker concentration adjustment action range is: -0.01mM, -0.05mM, 0mM, 0.01mM, 0.05mM, the action representation dimension is One-hot encoding, and is the sum of the pH action One-hot dimension, the temperature action One-hot dimension, and the cross-linker concentration action One-hot dimension.

[0040] It is particularly noted that the state representation dimension and the action representation dimension are not the state and action under normal conditions, but are improvements made according to the particularity of the state and action of the present invention. For example, after the pH state is one-hot encoded, each state is represented by a one-hot vector, such as: (0-3] corresponds to the vector [1, 0, 0, 0], (3-5] corresponds to the vector [0, 1, 0, 0], (5-7] corresponds to the vector [0, 0, 1, 0], and (7-14] corresponds to the vector [0, 0, 0, 1]; the temperature state is also represented by the same regular one-hot vector after one-hot encoding. The state representation dimension refers to the sum of the dimensions of each state vector after one-hot encoding of the pH state, plus the sum of the dimensions of each state vector after one-hot encoding of the temperature state. For example, there are 4 pH states, each of which is a 4-dimensional one-hot vector; there are 3 temperature states, each of which is a 3-dimensional one-hot vector, then the state representation dimension is:

[0041] pH state One-hot dimension sum = 4 × 4 = 16;

[0042] Temperature state One-hot dimension sum = 3×3=9;

[0043] The state representation dimension = pH state One-hot dimension sum + temperature state One-hot dimension sum = 16 + 9 = 25, and the same is true for the action representation dimension.

[0044] Step 3: Build the Q network:

[0045] The Q network includes 1 input layer, 2 hidden layers and 1 output layer. The input layer is the second input layer, and the number of nodes is 4+3=7. The hidden layer includes the third hidden layer and the fourth hidden layer. The activation function of the third hidden layer is the Relu function, and the number of nodes is 14. The structure of the fourth hidden layer is the same as the third hidden layer. The output layer is the fifth output layer, and the number of nodes in the fifth output layer is 10, which is linearly activated.

[0046] Step 4. Set the reward function:

[0047] The reward function formula is:

[0048] ;

[0049] In the formula, is the target Q value, =1-(preparation process coefficient-predicted preparation process coefficient) normalized value, For the next state, For the next state The best action under is the discount rate, the initial value is set to 0.9, is the maximum target Q value of the optimal action for the next state.

[0050] Step 5: Set the optimized execution plan:

[0051] The optimizer is Adam, the learning rate is 0.001, the experience replay pool capacity is 10w, the Q network is updated every 100 steps, and the ε-greedy strategy is used to select actions.

[0052] As a further description of the above technical solution: the training optimization method of the second machine learning model includes:

[0053] Call i preparation parameter adjustment records in the database as c group historical training data, the preparation parameter adjustment records include pH value, temperature, concentration of cross-linking agent and optimal adjustment strategy, divide c group historical training data into training set and test set, the ratio is 5:1; initialize the experience replay pool, start from a random initial state, adopt ε-greedy strategy to select action, execute action, observe reward and next state, store experience in the replay pool, every δ steps, δ is a preset value, determined by technicians in this field according to data fitting, randomly sample experience samples from the replay pool for learning, the learning method includes: sampling a (s, a, a', s'), is the current state, For the current action, calculate , calculate the loss function, loss function The formula is:

[0054] ;

[0055] In the formula, is the Q network parameter, initially 0; It is the actual preset reward value obtained by using the current action in the current state according to the current Q network parameters.

[0056] Beneficial effects:

[0057] In the above technical scheme, the preparation method of the ochratoxin A monoclonal antibody conjugate provided by the present invention, during the preparation process, according to the preparation process coefficient in the preparation process, the preparation parameters of the preparation process are adaptively adjusted in real time, thereby avoiding the preparation of the antibody conjugate within a given time, and the occurrence of over-coupling or under-coupling, and through the intelligent analysis and planning of the machine learning model, more accurate preparation parameter adjustment can be achieved, thereby improving the efficiency of preparing the antibody conjugate, and greatly reducing human manipulation, and improving the automation level of preparing the antibody conjugate; the model learns historical experience knowledge, can select the optimal solution for different conditions, so that the quality of the prepared antibody conjugate is more stable and controllable. BRIEF DESCRIPTION OF THE DRAWINGS

[0058] The present invention will be further explained below in conjunction with the accompanying drawings and embodiments:

[0059] Figure 1 A flow chart of a method for preparing an ochratoxin A monoclonal antibody conjugate provided in an embodiment of the present invention;

[0060] Figure 2 A flow chart of a method for obtaining the binding efficiency and binding amount of a conjugate is provided for an embodiment of the present invention;

[0061] Figure 3 A flowchart for constructing a second machine learning model provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0062] In order to make the technical means, creative features, objectives and effects of the present invention easy to understand, the present invention is further described below with reference to specific diagrams. It should be noted that the embodiments and features in the embodiments of the present application can be combined with each other without conflict.

[0063] Example 1

[0064] See also Figure 1-Figure 3 The embodiment of the present invention provides a technical solution: a method for preparing an ochratoxin A monoclonal antibody conjugate, which is applied to the process of combining the monoclonal antibody of ochratoxin A with the microspheres that have been conjugated with ochratoxin A, and is characterized in that it comprises the following steps:

[0065] It should be noted that the preparation process of the 6-atom arm-red microsphere-ochratoxin A monoclonal antibody conjugate includes:

[0066] The pretreated red microspheres and ochratoxin A are added to a reaction vessel for reaction, and the cross-linking agent (EDC / NHS) forms a covalent bond between the ochratoxin A molecules and the amino groups on the surface of the microspheres;

[0067] Using cross-linking agents such as EDC / NHS, the monoclonal antibody against ochratoxin A is combined with the microspheres that have been coupled with ochratoxin A, and the antibody and the microspheres coupled with ochratoxin A are connected by covalent bonds.

[0068] Step S01: obtaining component ratio data of the preparation raw materials, and obtaining preparation process data and measured control parameter data in the preparation process at preset time intervals during the process of preparing the antibody conjugate using the preparation raw materials under predetermined preparation parameters, and obtaining a real-time preparation process coefficient based on the preparation process data;

[0069] The predetermined preparation parameters include reaction temperature, reaction pH value, concentration of cross-linking agent, and total preparation time M of antibody conjugate;

[0070] The measured control parameter data include reaction temperature, reaction pH value and cross-linking agent concentration;

[0071] It should be noted that the composition ratio data of the preparation raw materials, that is, the type of preparation raw materials and the proportion of various raw materials, wherein the type of preparation raw materials includes red microspheres (microspheres with polystyrene or polymer substrates are selected, and the surface can be modified to have functional groups with certain hydrophilicity, such as amino groups and carboxyl groups), ochratoxin A (Ochratoxin A, OTA), target molecules for antigen coupling, monoclonal antibodies: antibodies against ochratoxin A, chemical cross-linking agents, such as EDC (1-ethyl-3-(3-dimethylaminopropyl) carbodiimide) / NHS (N-hydroxysuccinimide), used for cross-linking reactions of coupling antibodies to microspheres, buffers and solvents, such as PBS (phosphate buffer solution) and purified water.

[0072] Step S02: obtaining the preparation time ratio Q / M, inputting the component ratio data of the preparation raw materials, the real-time preparation parameters and the preparation time ratio Q / M into the pre-built first machine learning model for predicting the preparation progress coefficient, predicting the preparation progress coefficient at time P, and recording it as the predicted preparation progress coefficient;

[0073] It should be noted that the preparation time ratio Q / M is generated based on the total preparation time M and the start time of the antibody conjugate. The preparation time Q is obtained by obtaining the current time and the start time, and the preparation time Q is divided by the total preparation time M. For example, if the start time is 7:00 and the T time is 8:00, the preparation time Q is 60 minutes; the preparation time ratio is Q / M;

[0074] The method for constructing the first machine learning model includes:

[0075] Initialize the first machine learning model structure. The first machine learning model structure adopts an MLP type multi-layer forward network structure with one input layer, two hidden layers, and one output layer; the input layer includes the first input layer, the second input layer, and the third input layer. The number of nodes in the first input layer of the first input layer is 1, corresponding to the composition ratio data of the prepared raw materials; the number of nodes in the second input layer is 3, corresponding to the reaction temperature, reaction pH value, and concentration of the cross-linking agent of the real-time control parameters; the number of nodes in the third input layer is 1, corresponding to the preparation time ratio; the hidden layer includes the first hidden layer and the second hidden layer. The number of nodes in the first hidden layer of the first hidden layer is 128, and the Relu function is used as the activation function; the number of nodes in the second hidden layer of the second hidden layer is 64, and the Relu function is used as the activation function; the output layer is the first output layer, and the number of nodes in the first output layer of the first output layer is 1, which is the predicted preparation process coefficient.

[0076] After initializing the first machine learning model structure, the first machine learning model is trained using the ingredient ratio data of the preparation raw materials, the real-time preparation parameters and the preparation time ratio Q / M, and the preparation process coefficients corresponding to the ingredient ratio data of the preparation raw materials, the real-time preparation parameters and the preparation time ratio Q / M. The ingredient ratio data of the preparation raw materials, the real-time preparation parameters and the preparation time ratio Q / M, and the preparation process coefficients corresponding to the ingredient ratio data of the preparation raw materials, the real-time preparation parameters and the preparation time ratio Q / M are obtained from a database. The database records the data of each preparation of the 6-atomic arm-red microsphere-ochratoxin A monoclonal antibody conjugate, which is used to optimize the model. The optimizer is Adam, the loss function is the MSE loss function, the batch size is set to 32, the number of iterations is 200 rounds, and the training is ended when the loss function converges, indicating that the training is completed.

[0077] It should be noted that the Adam optimizer is one of the five major optimizers commonly used in machine learning, and its full name is Adaptive Moment Estimation.

[0078] The first machine learning model training method comprises:

[0079] The ingredient ratio data of the preparation raw materials, the real-time preparation parameters and the preparation time ratio Q / M are converted into a corresponding set of feature vectors.

[0080] The ingredient ratio data, real-time preparation parameters and preparation time ratio Q / M of the preparation raw materials are taken as inputs of the machine learning model, and the machine learning model takes a group of preparation process coefficients corresponding to the ingredient ratio data, real-time preparation parameters and preparation time ratio Q / M of each group of preparation raw materials as outputs, and takes the preparation process coefficients actually corresponding to the ingredient ratio data, real-time preparation parameters and preparation time ratio Q / M of each group of preparation raw materials as prediction targets, and takes minimizing the loss function value of the machine learning model as training target; training stops when the loss function value of the machine learning model is less than or equal to the preset target loss value.

[0081] The loss function value of the first machine learning model is the mean square error;

[0082] Mean square error is one of the commonly used loss functions. By transforming the loss function formula

[0083] = The model is trained by minimizing the target, so that the machine learning model can better fit the data, thereby improving the performance and accuracy of the model.

[0084] In the loss function is the loss function value of the first machine learning model, x is the feature vector group number; e is the number of feature vector groups; is a set of preparation process coefficients of the first machine learning model output corresponding to the x-th set of feature vectors, is the preparation process coefficient actually corresponding to the xth group of eigenvectors.

[0085] Step S03: obtaining a real-time preparation process coefficient, and judging whether the antibody conjugate prepared by the preparation raw material within the predetermined total preparation time M of the antibody conjugate meets the preparation process standard according to the real-time preparation process coefficient; if so, continuing to control the preparation reaction device according to the measured control parameter data;

[0086] The method for judging whether the antibody conjugate prepared by the preparation raw material within the total preparation time M of the established antibody conjugate meets the preparation process standard according to the real-time preparation process coefficient includes:

[0087] A preparation process difference threshold is preset, and the real-time preparation process coefficient is subtracted from the predicted preparation process coefficient to obtain an absolute value, which is recorded as the process loss value. When the process loss value is less than or equal to the preset preparation process difference threshold, it indicates that the preparation process standard is met; conversely, when the process loss value is greater than the preset preparation process difference threshold, it indicates that the preparation process standard is not met.

[0088] Step S04: If not, the preparation process coefficient, the preparation time ratio and the measured control parameter data are input into a second machine learning model pre-configured for feedback correction of the control parameters, to obtain the corrected preparation parameters, and the preparation reaction is controlled based on the corrected preparation parameters;

[0089] The second machine learning model is a reinforcement learning model, using a Q network structure, and the construction steps are as follows:

[0090] Step 1: Design state representation dimensions:

[0091] The design state representation dimensions include design pH state discreteness, design temperature state discreteness and design cross-linking agent concentration state discreteness;

[0092] The pH state is discrete as follows: (5.5-7] (6-7] (7-7.5], using One-hot coding; the temperature state is discrete as follows: (25°C-35°C] (30°C-35°C] (30°C-37°C], using One-hot coding; the concentration state of the cross-linking agent is discrete as follows: (0.1mM-0.8mM] (0.5mM-0.8mM] (0.5mM-1mM], the state representation dimension is One-hot coding, and is the sum of the pH state One-hot dimension, the temperature state One-hot dimension, and the cross-linking agent concentration state One-hot dimension.

[0093] Step 2: Design the action representation dimension:

[0094] The design action representation dimensions include a design pH adjustment action range, a design temperature adjustment action range, and a design cross-linking agent concentration adjustment action range;

[0095] The pH adjustment action range is: -0.5, -0.2, 0, 0.2, 0.5, using One-hot encoding; the temperature adjustment action range is: -5°C, -2°C, 0°C, 2°C, 5°C, using One-hot encoding; the cross-linker concentration adjustment action range is: -0.01mM, -0.05mM, 0mM, 0.01mM, 0.05mM, the action representation dimension is One-hot encoding, and is the sum of the pH action One-hot dimension, the temperature action One-hot dimension, and the cross-linker concentration action One-hot dimension.

[0096] It is particularly noted that the state representation dimension and the action representation dimension are not the state and action under normal conditions, but are improvements made according to the particularity of the state and action of the present invention. For example, after the pH state is one-hot encoded, each state is represented by a one-hot vector, such as: (0-3] corresponds to the vector [1, 0, 0, 0], (3-5] corresponds to the vector [0, 1, 0, 0], (5-7] corresponds to the vector [0, 0, 1, 0], and (7-14] corresponds to the vector [0, 0, 0, 1]; the temperature state is also represented by the same regular one-hot vector after one-hot encoding. The state representation dimension refers to the sum of the dimensions of each state vector after one-hot encoding of the pH state, plus the sum of the dimensions of each state vector after one-hot encoding of the temperature state. For example, there are 4 pH states, each of which is a 4-dimensional one-hot vector; there are 3 temperature states, each of which is a 3-dimensional one-hot vector, then the state representation dimension is:

[0097] pH state One-hot dimension sum = 4 × 4 = 16;

[0098] Temperature state One-hot dimension sum = 3×3=9;

[0099] The state representation dimension = pH state One-hot dimension sum + temperature state One-hot dimension sum = 16 + 9 = 25, and the same is true for the action representation dimension.

[0100] Step 3: Build the Q network:

[0101] The Q network includes 1 input layer, 2 hidden layers and 1 output layer. The input layer is the second input layer, and the number of nodes is 4+3=7. The hidden layer includes the third hidden layer and the fourth hidden layer. The activation function of the third hidden layer is the Relu function, and the number of nodes is 14. The structure of the fourth hidden layer is the same as the third hidden layer. The output layer is the fifth output layer, and the number of nodes in the fifth output layer is 10, which is linearly activated.

[0102] Step 4. Set the reward function:

[0103] The reward function formula is:

[0104] ;

[0105] In the formula, is the target Q value, =1-(preparation process coefficient-predicted preparation process coefficient) normalized value, For the next state, For the next state The best action under is the discount rate, the initial value is set to 0.9, is the maximum target Q value of the optimal action for the next state.

[0106] Step 5: Set the optimized execution plan:

[0107] The optimizer is Adam, the learning rate is 0.001, the experience replay pool capacity is 10w, the Q network is updated every 100 steps, and the ε-greedy strategy is used to select actions.

[0108] The training optimization method of the second machine learning model includes:

[0109] Call i preparation parameter adjustment records in the database as c group historical training data, the preparation parameter adjustment records include pH value, temperature, concentration of cross-linking agent and optimal adjustment strategy, divide c group historical training data into training set and test set, the ratio is 5:1; initialize the experience replay pool, start from a random initial state, adopt ε-greedy strategy to select action, execute action, observe reward and next state, store experience in the replay pool, every δ steps, δ is a preset value, determined by technicians in this field according to data fitting, randomly sample experience samples from the replay pool for learning, the learning method includes: sampling a (s, a, a', s'), is the current state, For the current action, calculate , calculate the loss function, loss function The formula is:

[0110] ;

[0111] In the formula, is the Q network parameter, initially 0; It is the actual preset reward value obtained by using the current action in the current state according to the current Q network parameters.

[0112] Special attention should be paid to: It is determined by those skilled in the art as a value guide to give a training direction so that the direction of model change turns to the development direction of pre-solving the problem of the present invention.

[0113] Will Back propagation is used to optimize the Q network parameters. The Adam optimizer is used. ε gradually decays from 1 to the minimum value. The Q network parameters obtained through training are enhanced. The effect of the Q network is tested every δ steps. Training is continued until the loss function converges. The Q network parameters are output and the control effect is evaluated on the test set using the Q network parameters. When the average reward value is higher than the preset threshold of 0.9 for N consecutive evaluation cycles, the training can be terminated.

[0114] Step S05: Control the preparation of antibody conjugates according to the measured control parameters or the modified preparation parameters.

[0115] In this embodiment, during the preparation process, according to the preparation process coefficient in the preparation process, the preparation parameters of the preparation process are adaptively adjusted in real time, thereby avoiding the situation of over-coupling or under-coupling when preparing antibody conjugates within a given time, that is, dynamic adjustment of preparation parameters is achieved to ensure that the quality of antibody conjugates prepared within a given time is more stable, and through intelligent analysis and planning of machine learning models, more accurate preparation parameter adjustment can be achieved, thereby improving the efficiency of preparing antibody conjugates, greatly reducing human manipulation, and improving the automation level of preparing antibody conjugates; the model learns historical experience knowledge and can select the optimal solution for different conditions, so that the quality of prepared antibody conjugates is more stable and controllable.

[0116] Example 2

[0117] See also Figure 1 and Figure 2 , the preparation process data include: binding efficiency, binding amount and protein concentration of the conjugate;

[0118] The method for obtaining the binding efficiency and binding amount of the conjugate comprises:

[0119] Add OTA solution of known concentration into the cuvette and measure its absorbance at 332 nm, which is recorded as A. 332（OTA） ;

[0120] By sampling, the coupled antibody-OTA solution was added to the cuvette, and its absorbance values ​​at 332nm and 280nm were measured respectively, which were recorded as A 332（偶联物） and A 280（偶联物）；

[0121] The binding efficiency and binding amount of the conjugate are obtained based on the calculation formula;

[0122] The calculation formula for the conjugate binding efficiency is:

[0123] The calculation formula for the conjugate binding amount is: ;

[0124] In the formula, is the binding efficiency of the conjugate, is the absorbance of the conjugate at 332 nm, indicating the OTA content in the antibody-OTA complex. is the absorbance of OTA solution at 332 nm, indicating the initial concentration of OTA, is the binding amount of the conjugate, is the molar absorption coefficient of anti-ochratoxin A monoclonal antibody solution at 280 nm, is the concentration of anti-ochratoxin A monoclonal antibody solution.

[0125] In the embodiment of the present application, the concentration is measured in combination with the wavelengths of 332nm and 280nm. There are specific chromophores in the molecular structure of ochratoxin A, which can show a special absorption peak at 332nm, which can be used as an ideal marker to track the dynamic changes of OTA in the entire coupling reaction system. When OTA is not bound to the antibody, the absorbance of the solution at 332nm directly reflects the concentration level of free OTA. As the coupling reaction progresses, OTA molecules continue to covalently link with antibodies, the number of free OTA decreases accordingly, and the absorbance of the solution at this wavelength also decreases accurately. In this way, the feeding consumption rate and binding conversion degree of OTA at each time node can be determined, and then it can be accurately judged whether the coupling reaction is carried out according to the expected kinetic path.

[0126] In addition, the present application cooperates with 332nm and 280nm wavelength correlation detection. Aromatic amino acid residues in proteins, such as tyrosine, tryptophan, etc., have strong light absorption characteristics near 280nm. For monoclonal antibodies involved in the coupling reaction, the 280nm wavelength can monitor their state changes. Before the coupling reaction is started, the absorbance of the pure antibody solution at 280nm is measured to accurately determine the initial concentration and purity of the antibody. Once OTA begins to bind to the antibody, not only will the spatial conformation of the antibody molecule itself change subtly, but also due to the chemical modification of OTA, it will further affect the microenvironment in which the aromatic amino acid residues are located, causing the overall absorbance at 280nm to produce characteristic fluctuations. This fluctuation not only contains information about the reduction of antibody concentration due to the binding reaction, but also contains the "fingerprint" of subtle changes in the molecular structure. Through the in-depth analysis of the absorbance changes at this wavelength, the integrity of the antibody during the coupling process, the maintenance of activity, and the dynamic changes of the binding site with OTA can be fully grasped. Key details such as dynamic changes.

[0127] In a preferred embodiment of the present application, cross-validation of real-time data can also be performed. Specifically, at every moment of the coupling reaction, absorbance data at 332nm and 280nm are collected simultaneously to form an instant "data pair". These two sets of data are closely related and confirm each other. For example, when the absorbance at 332nm shows a rapid decrease in OTA concentration, theoretically, the absorbance change pattern caused by antibody binding to OTA should be observed simultaneously at 280nm. If the change trends of the two data are inconsistent, such as a sudden drop in absorbance at 332nm but a nearly stable absorbance at 280nm, it indicates that there may be abnormalities, such as non-specific adsorption of OTA, premature loss of antibody activity, and other problems.

[0128] Further, in the preferred embodiment of the present application, the dynamic change curve of the coupling reaction can be constructed by continuously accumulating the dual-wavelength absorbance data at different time points. Using the data fitting and analysis algorithm, it is not only possible to trace back the reaction process and accurately review the state of the reactants at each key node, but more importantly, the trend of the coupling reaction in the future in a short period of time can be predicted based on the current data trend. For example, if it is found that the rate of decrease of the absorbance at 332nm is gradually slowing down, and the fluctuation of the absorbance at 280nm tends to be stable, the reaction is about to enter the plateau period, and the adjustment mechanism can be automatically triggered at this time. According to the preset optimization strategy, the reaction temperature is moderately increased, the reaction time is extended, or the concentration ratio of the reactants is fine-tuned, and the coupling reaction is actively guided to move in the direction of higher efficiency and better quality, thereby improving the success rate and stability of the preparation of the conjugate.

[0129] In addition, in some embodiments, the method for obtaining the protein concentration includes:

[0130] Use a protein standard solution (such as BSA or other standard protein) of known concentration to construct a calibration curve;

[0131] By taking samples, the absorbance of the samples at 280 nm was measured using a UV-Vis spectrophotometer;

[0132] Calculate protein concentration based on the calibration curve and absorbance.

[0133] The calculation formula is: ; In the formula, is the protein concentration, is the absorbance at 280 nm, is the molar absorptivity of the protein (usually requires experimental determination for a specific protein), is the optical path length.

[0134] It should be noted that during the binding reaction, the protein concentration will change. The binding progress can be evaluated by regularly measuring the changes in protein concentration. If the protein concentration gradually decreases during the binding reaction, it means that the protein has bound to the microspheres, that is, the lower the protein concentration, the faster the reaction progresses.

[0135] The method for obtaining the real-time preparation process coefficient based on the preparation process data is to perform weighted calculation on the protein concentration, the binding efficiency of the conjugate and the binding amount of the conjugate to obtain the real-time preparation process coefficient.

[0136] The calculation method of the real-time preparation process coefficient is:

[0137] ;

[0138] In the formula, is the preparation process coefficient, is the protein concentration, , and is the weight coefficient and is greater than 0.

[0139] It should be noted that the preparation process coefficient is used to indicate the progress of preparing antibody conjugates, wherein the larger the preparation process coefficient, the faster the surface preparation process. It should be understood that within a given preparation time, a preparation process that is too fast or too slow will affect the quality of the final antibody conjugate, because preparation that is often too fast will cause over-coupling, and vice versa will cause insufficient coupling.

[0140] It should be noted that the size of the weight coefficient is a specific value obtained by quantifying each data to facilitate subsequent comparison. The size of the weight coefficient depends on the number of comprehensive parameters and the preliminary setting of the corresponding weight coefficient for each set of comprehensive parameters by technical personnel in this field.

[0141] In this embodiment, by collecting preparation process data, the preparation process data includes: binding efficiency, binding amount and protein concentration of the conjugate, and then obtaining a real-time preparation process coefficient based on the preparation process data, the user can intuitively understand the process of preparing the antibody conjugate, which is convenient for adjusting the preparation parameters based on the process of preparing the antibody conjugate.

[0142] The basic principles, main features and advantages of the present invention are shown and described above. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The above embodiments and the description in the specification are only to illustrate the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which fall within the scope of the present invention to be protected. The scope of protection of the present invention is defined by the attached claims and their equivalents.

Claims

1. A method for preparing an ochratoxin A monoclonal antibody conjugate, which is applied to the process of combining an anti-ochratoxin A monoclonal antibody with a microsphere conjugated with ochratoxin A, characterized in that: The following steps are involved: Step S01: obtaining component ratio data of the preparation raw materials, and obtaining preparation process data and measured control parameter data in the preparation process at preset time intervals during the process of preparing the antibody conjugate using the preparation raw materials under predetermined preparation parameters, and obtaining a real-time preparation process coefficient based on the preparation process data; The preparation process data include: the binding efficiency, binding amount and protein concentration of the conjugate; The method for obtaining the real-time preparation process coefficient based on the preparation process data is: performing weighted calculation on the protein concentration, conjugate binding efficiency and binding amount to obtain the real-time preparation process coefficient; Step S02: obtaining the preparation time ratio Q / M, inputting the component ratio data of the preparation raw materials, the real-time preparation parameters and the preparation time ratio Q / M into the pre-built first machine learning model for predicting the preparation progress coefficient, predicting the preparation progress coefficient at time P, and recording it as the predicted preparation progress coefficient; The preparation time ratio Q / M is obtained by obtaining the preparation time Q by the current time and the start time, and dividing the preparation time Q by the total preparation time M; Step S03: obtaining a real-time preparation process coefficient, and judging whether the antibody conjugate prepared by the preparation raw material within the predetermined total preparation time M of the antibody conjugate meets the preparation process standard according to the real-time preparation process coefficient; if so, continuing to control the preparation reaction device according to the measured control parameter data; Step S04: If not, the preparation process coefficient, the preparation time ratio and the measured control parameter data are input into a second machine learning model pre-configured for feedback correction of the control parameters, to obtain the corrected preparation parameters, and the preparation reaction is controlled based on the corrected preparation parameters; Step S05: Control the preparation of antibody conjugates according to the measured control parameters or the modified preparation parameters.

2. The method for preparing the ochratoxin A monoclonal antibody conjugate according to claim 1, characterized in that: The method for obtaining the binding efficiency and binding amount of the conjugate comprises: Add OTA solution of known concentration into the cuvette and measure its absorbance at 332nm, which is recorded as A. 332（OTA） ; By sampling, the coupled antibody-OTA solution was added to the cuvette, and its absorbance values ​​at 332nm and 280nm were measured respectively, which were recorded as A 332（偶联物） and A 280（偶联物）； The conjugate binding efficiency and binding amount were calculated based on the formula.

3. The method for preparing the ochratoxin A monoclonal antibody conjugate according to claim 2, characterized in that: The method for obtaining the protein concentration comprises: A calibration curve was constructed using protein standard solutions of known concentration; By taking samples, the absorbance of the samples at 280 nm was measured using a UV-Vis spectrophotometer; Calculate protein concentration based on the calibration curve and absorbance.

4. The method for preparing the ochratoxin A monoclonal antibody conjugate according to claim 1, characterized in that: The method for constructing the first machine learning model includes: Initialize the first machine learning model structure. The first machine learning model structure adopts an MLP type multi-layer forward network structure, with one input layer, two hidden layers, and one output layer; the input layer includes the first input layer, the second input layer, and the third input layer. The number of nodes in the first input layer of the first input layer is 1, corresponding to the composition ratio data of the prepared raw materials; the number of nodes in the second input layer is 3, corresponding to the reaction temperature, reaction pH value, and concentration of the cross-linking agent of the real-time control parameters; the number of nodes in the third input layer is 1, corresponding to the preparation time ratio; the hidden layer includes the first hidden layer and the second hidden layer, the number of nodes in the first hidden layer of the first hidden layer is 128, and the Relu function is used as the activation function; the number of nodes in the second hidden layer of the second hidden layer is 64, and the Relu function is used as the activation function; the output layer is the first output layer, and the number of nodes in the first output layer of the first output layer is 1, which is the predicted preparation process coefficient; After initializing the first machine learning model structure, the first machine learning model is trained using the ingredient ratio data of the preparation raw materials, the real-time preparation parameters and the preparation time ratio Q / M, and the preparation process coefficients corresponding to the ingredient ratio data of the preparation raw materials, the real-time preparation parameters and the preparation time ratio Q / M. The ingredient ratio data of the preparation raw materials, the real-time preparation parameters and the preparation time ratio Q / M, and the preparation process coefficients corresponding to the ingredient ratio data of the preparation raw materials, the real-time preparation parameters and the preparation time ratio Q / M are obtained from a database. The database records the data of each preparation of the 6-atomic arm-red microsphere-ochratoxin A monoclonal antibody conjugate, which is used to optimize the model. The optimizer is Adam, the loss function is the MSE loss function, the batch size is set to 32, the number of iterations is 200 rounds, and the training is ended when the loss function converges, indicating that the training is completed.

5. The method for preparing the ochratoxin A monoclonal antibody conjugate according to claim 4, characterized in that: The first machine learning model training method comprises: Converting the ingredient ratio data of the raw materials, the real-time preparation parameters and the preparation time ratio Q / M into a corresponding set of feature vectors; The ingredient ratio data, real-time preparation parameters and preparation time ratio Q / M of the preparation raw materials are taken as inputs of the machine learning model, and the machine learning model takes a group of preparation process coefficients corresponding to the ingredient ratio data, real-time preparation parameters and preparation time ratio Q / M of each group of preparation raw materials as outputs, and takes the preparation process coefficients actually corresponding to the ingredient ratio data, real-time preparation parameters and preparation time ratio Q / M of each group of preparation raw materials as prediction targets, and takes minimizing the loss function value of the machine learning model as training target; training stops when the loss function value of the machine learning model is less than or equal to the preset target loss value.

6. The method for preparing the ochratoxin A monoclonal antibody conjugate according to claim 1, characterized in that: The method for judging whether the antibody conjugate prepared by the preparation raw materials within the predetermined total preparation time M of the antibody conjugate meets the preparation process standard according to the real-time preparation process coefficient includes: A preparation process difference threshold is preset, and the real-time preparation process coefficient is subtracted from the predicted preparation process coefficient to obtain an absolute value, which is recorded as the process loss value. When the process loss value is less than or equal to the preset preparation process difference threshold, it indicates that the preparation process standard is met; conversely, when the process loss value is greater than the preset preparation process difference threshold, it indicates that the preparation process standard is not met.

7. The method for preparing the ochratoxin A monoclonal antibody conjugate according to claim 1, characterized in that: The second machine learning model is a reinforcement learning model, using a Q network structure, and the construction steps are as follows: Step 1: Design state representation dimensions: The design state representation dimensions include design pH state discreteness, design temperature state discreteness and design cross-linking agent concentration state discreteness; The pH state is discrete as follows: (5.5-7] (6-7] (7-7.5], using One-hot coding; the temperature state is discrete as follows: (25°C-35°C] (30°C-35°C] (30°C-37°C], using One-hot coding; the concentration state of the cross-linking agent is discrete as follows: (0.1mM-0.8mM] (0.5mM-0.8mM] (0.5mM-1mM], the state representation dimension is One-hot coding, and is the sum of the pH state One-hot dimension, the temperature state One-hot dimension, and the cross-linking agent concentration state One-hot dimension; Step 2: Design the action representation dimension: The design action representation dimensions include a design pH adjustment action range, a design temperature adjustment action range, and a design cross-linking agent concentration adjustment action range; The pH adjustment action range is: -0.5, -0.2, 0, 0.2, 0.5, using One-hot coding; the temperature adjustment action range is: -5°C, -2°C, 0°C, 2°C, 5°C, using One-hot coding; the cross-linking agent concentration adjustment action range is: -0.01mM, -0.05mM, 0mM, 0.01mM, 0.05mM, the action representation dimension is One-hot coding, and is the sum of the pH action One-hot dimension, the temperature action One-hot dimension, and the cross-linking agent concentration action One-hot dimension. Step 3: Build the Q network: The Q network includes 1 input layer, 2 hidden layers and 1 output layer. The input layer is the second input layer, the number of nodes is 4+3=7, the hidden layer includes the third hidden layer and the fourth hidden layer, the activation function of the third hidden layer is the Relu function, the number of nodes is 14, the structure of the fourth hidden layer is the same as the third hidden layer, the output layer is the fifth output layer, the number of nodes in the fifth output layer is 10, and the linear activation is used; Step 4. Set the reward function: The reward function formula is: ； In the formula, is the target Q value, =1-(preparation process coefficient-predicted preparation process coefficient) normalized value, For the next state, For the next state The best action under is the discount rate, the initial value is set to 0.9, is the maximum target Q value of the optimal action for the next state; Step 5: Set the optimized execution plan: The optimizer is Adam, the learning rate is 0.001, the experience replay pool capacity is 10w, the Q network is updated every 100 steps, and the ε-greedy strategy is used to select actions.

8. The method for preparing the ochratoxin A monoclonal antibody conjugate according to claim 7, characterized in that: The training optimization method of the second machine learning model includes: Call i preparation parameter adjustment records in the database as c group historical training data, the preparation parameter adjustment records include pH value, temperature, cross-linking agent concentration and optimal adjustment strategy, divide c group historical training data into training set and test set, the ratio is 5:1; initialize the experience replay pool, start from a random initial state, use ε-greedy strategy to select actions, execute actions, observe rewards and next states, store experience in the replay pool, every δ steps, δ is a preset value, randomly sample experience samples from the replay pool for learning, the learning method includes: sampling a (s, a, a', s'), is the current state, For the current action, calculate , calculate the loss function, loss function The formula is: ； In the formula, is the Q network parameter, initially 0; It is the actual preset reward value obtained by using the current action in the current state according to the current Q network parameters.

Images