A method for detecting monoclonal antibody drugs

Abstract

The present invention discloses a method for detecting monoclonal antibody drugs, comprising: step S1: collecting Raman spectrum data of the drug to be tested; step S2: using a preprocessing method to generate a spectrum feature array for the Raman spectrum data; step S3: using a pre-built neural network model to classify the spectrum feature array to generate a classification result. The beneficial effect of the present invention is that the processing of Raman spectrum data based on the spectrum feature array can effectively improve the recognition ability of low-concentration drugs, and realize feature extraction through a relatively simple method, which facilitates the neural network model to achieve a better recognition effect, avoiding the defect of the prior art that it is difficult to detect monoclonal antibody drugs.

Description

Technical Field

[0001] The present invention relates to the technical field of drug detection, and in particular to a method for detecting a monoclonal antibody drug. Background Art

[0002] Monoclonal antibody drugs are highly uniform antibody proteins produced by single cell clones and targeting specific antigen epitopes. Their R&D and production costs are much higher than chemical drugs. The market price of each 100 mg monoclonal antibody drug is as high as tens of thousands of yuan, and the price of each monoclonal antibody drug of each variety ranges from 5,000 to 30,000 yuan. The high price can bring high profits to counterfeiting, and monoclonal antibody drugs are easily targeted by criminals for counterfeiting and trafficking. Therefore, it is of great significance to conduct rapid and accurate detection of monoclonal antibody drugs.

[0003] Currently, commonly used methods for detecting monoclonal antibody drugs include: UV method for determining protein content, capillary isoelectric focusing method for determining protein isoelectric point, and liquid chromatography-mass spectrometry method for determining protein primary structure. These methods can detect different properties of monoclonal antibody drug proteins from different levels and angles, and can be used in combination to identify the authenticity of monoclonal antibody drugs. However, the above methods are cumbersome to operate and have long experimental cycles: UV method for determining protein content takes about 30 minutes, capillary isoelectric focusing method for determining protein isoelectric point takes about 60 minutes, and liquid chromatography-mass spectrometry method for determining protein primary structure takes about 2 to 3 days. In addition, the equipment involved is large in size and has certain requirements for the detection environment. It needs to be completed in the laboratory, making it difficult to achieve the goal of on-site sampling and rapid release.

[0004] Raman spectroscopy has the technical advantages of simple operation, fast detection speed, high reproducibility, and non-destructiveness. Different compounds can be accurately identified based on the "fingerprint" characteristics of Raman spectroscopy. At present, Raman spectroscopy technology has been widely used in chemical drug production, chemical drug sampling, etc. Compared with chemical drugs, monoclonal antibody drugs are essentially proteins with complex structures and low content in drugs. The following difficulties exist in the identification of monoclonal antibody drugs using Raman spectroscopy technology:

[0005] 1. The similarity between different types of monoclonal antibody drugs is high: Monoclonal antibody drugs are all immunoglobulins, and different types of monoclonal antibody drugs only have certain structural differences in the Fab segment;

[0006] 2. The protein content of monoclonal antibody drugs is at the milligram level: The protein content in the monoclonal antibody drug prescription is about tens of milligrams per milliliter, for example, the content of commercially available adalimumab monoclonal antibody is 40mg / 0.8mL, bevacizumab monoclonal antibody is 100mg / 4mL, and ranibizumab monoclonal antibody is 10mg / mL, which is lower than the molar concentration of the excipients;

[0007] 3. Similarity in monoclonal antibody drug prescriptions: The types of excipients used in monoclonal antibody drug prescriptions are relatively similar, such as polysorbate 80, D-trehalose, methionine, etc.;

[0008] Therefore, the Raman spectra of monoclonal antibody drugs have a high degree of similarity, and it is difficult to achieve high-accuracy identification of monoclonal antibody drugs using existing Raman spectroscopic detection methods and systems for chemical drugs. Summary of the invention

[0009] In view of the above problems existing in the prior art, a method for detecting a monoclonal antibody drug is now provided.

[0010] The specific technical solutions are as follows:

[0011] A method for detecting a monoclonal antibody drug, comprising:

[0012] Step S1: collecting Raman spectrum data of the drug to be tested;

[0013] Step S2: applying a preprocessing method to the Raman spectrum data to generate a spectrum feature array;

[0014] Step S3: using a pre-built neural network model to classify the graph feature array to generate a classification result.

[0015] Preferably, the method for constructing the neural network model in step S3 includes:

[0016] Step A1: collecting the Raman spectral data of a plurality of non-monoclonal antibody drugs to generate a first data set, and collecting the Raman spectral data of a plurality of monoclonal antibody drugs to generate a second data set;

[0017] Step A2: using the preprocessing method to generate a plurality of first arrays and a plurality of second arrays for the first data set and the second data set;

[0018] The number of the first arrays is the same as the number of the second arrays;

[0019] Step A3: generating a training set and a validation set according to a plurality of the first arrays and a plurality of the second arrays;

[0020] Step A4: Generate a model to be verified based on the training set, and then perform cross-validation based on the model to be verified and the verification set to generate the neural network model.

[0021] Preferably, the preprocessing method in step S2 includes:

[0022] Step S21: intercepting the characteristic area of ​​the spectrum of the Raman spectrum data to generate an intercepted spectrum;

[0023] Step S22: extracting the spectrum range of the intercepted spectrum to generate a filtered image;

[0024] Step S23: extracting spectrum peak extreme points from the filtered image to generate the spectrum feature array.

[0025] Preferably, the spectrum range extracted in step S22 is 153 cm -1 ~1792cm -1 .

[0026] Preferably, the spectrum range extraction in step S22 includes:

[0027] Step S221: performing Fourier transform on the intercepted spectrum to generate a frequency domain image;

[0028] Step S222: extracting a frequency band of 0-200 Hz from the frequency domain image, and then performing an inverse Fourier transform to generate the filtered image.

[0029] Preferably, the neural network model comprises:

[0030] An embedding layer, the embedding layer receiving the graph feature array and performing dimensionality reduction on the graph feature array to generate a reduced dimensionality array;

[0031] A long short-term memory network, the long short-term memory network receiving the reduced-dimensional array and generating a recursive result;

[0032] A fully connected layer receives the recursive result and outputs the classification result.

[0033] Preferably, in step S1, a Raman spectrometer is used to collect the Raman spectrum data;

[0034] The Raman spectrometer has a laser wavelength of 532 nm, a laser energy of 10 mW, a grating of 900 lines / mm, and a spectral resolution of less than 8.3 cm -1 , spectral scanning range 58cm -1 ~3350cm -1 .

[0035] Preferably, in the step A1, three production batches of the non-monoclonal antibody drug or the monoclonal antibody drug are sampled respectively;

[0036] 60 sets of Raman spectral data were collected for each production batch.

[0037] Preferably, the spectrum feature array is 50 bits.

[0038] The above technical solution has the following advantages or beneficial effects: by processing Raman spectral data based on the spectrum feature array, the recognition ability of low-concentration drugs can be effectively improved, and feature extraction can be achieved through a relatively simple method, which facilitates the neural network model to achieve better recognition effects, avoiding the defect of the existing technology that it is difficult to detect monoclonal antibody drugs. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] The embodiments of the present invention will be described more fully with reference to the attached drawings, which are provided for illustration and description only and are not intended to limit the scope of the present invention.

[0040] Figure 1 Schematic diagram of a detection method according to an embodiment of the present invention;

[0041] Figure 2 A schematic diagram of a method for constructing a neural network model according to an embodiment of the present invention;

[0042] Figure 3 The Raman spectrum data of some drugs in the embodiments of the present invention;

[0043] Figure 4 Schematic diagram of a preprocessing method according to an embodiment of the present invention;

[0044] Figure 5A This is a diagram of the pretreatment results of some drugs in the embodiments of the present invention;

[0045] Figure 5B This is a diagram of the pretreatment results of some drugs in the embodiments of the present invention;

[0046] Figure 5C This is a diagram of the pretreatment results of some drugs in the embodiments of the present invention;

[0047] Figure 6 This is a schematic diagram of sub-steps of step S22 of an embodiment of the present invention;

[0048] Figure 7 A schematic diagram of a neural network model according to an embodiment of the present invention;

[0049] Figure 8 A schematic diagram of the number of iterations in an embodiment of the present invention;

[0050] Fig. 9 Schematic diagram of comparison results of an embodiment of the present invention. DETAILED DESCRIPTION

[0051] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0052] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features in the embodiments may be combined with each other.

[0053] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments, but they are not intended to limit the present invention.

[0054] The present invention comprises:

[0055] A method for detecting a monoclonal antibody drug, such as Figure 1 As shown, including:

[0056] Step S1: collecting Raman spectrum data of the drug to be tested;

[0057] Step S2: applying a preprocessing method to the Raman spectrum data to generate a spectrum feature array;

[0058] Step S3: Use a pre-built neural network model to classify the graph feature array to generate a classification result.

[0059] In a preferred embodiment, Figure 2 As shown, the method for constructing the neural network model in step S3 includes:

[0060] Step A1: collecting Raman spectral data of a plurality of non-monoclonal antibody drugs to generate a first data set, and collecting Raman spectral data of a plurality of monoclonal antibody drugs to generate a second data set;

[0061] Step A2: using a preprocessing method on the first data set and the second data set to generate a plurality of first arrays and a plurality of second arrays;

[0062] The number of the first array is the same as the number of the second array;

[0063] Step A3: generating a training set and a validation set according to the plurality of first arrays and the plurality of second arrays;

[0064] Step A4: Generate a model to be verified based on the training set, and then perform cross-validation based on the model to be verified and the verification set to generate a neural network model.

[0065] Specifically, the non-monoclonal antibody drugs used as controls in this embodiment include: water, normal saline, glucose injection, invert sugar electrolyte injection, potassium chloride / sodium chloride injection, intravenous human immunoglobulin injection, tetanus human immunoglobulin, atosiban acetate, potassium magnesium aspartate, compound aminbarbital injection, dopamine hydrochloride injection, lidocaine injection, tramadol hydrochloride injection, nikethamide injection, aminobenzoic acid injection, epinephrine hydrochloride injection, sterile water for injection, bupivacaine hydrochloride injection, potassium chloride injection non-monoclonal antibody injection, all of which are domestically available products.

[0066] In this embodiment, the number of monoclonal antibody drugs used to construct the neural network model is 15, including: adalimumab monoclonal antibody injection (number 0), bevacizumab monoclonal antibody injection (number 1), ranibizumab monoclonal antibody injection (number 2), tocilizumab monoclonal antibody injection (number 3), elouzumab monoclonal antibody injection (number 4), secukiyu monoclonal antibody injection (number 5), rituximab monoclonal antibody injection (number 6), trastuzumab monoclonal antibody injection (number 7), pertuzumab monoclonal antibody injection (number 8), denosumab monoclonal antibody injection (number 9), ixetuzumab monoclonal antibody injection (number 10), ustekinumab monoclonal antibody injection (number 11), gusekiyu monoclonal antibody injection (number 12), emicizumab monoclonal antibody injection (number 13), etanercept monoclonal antibody injection (number 14). The above 15 drugs are all domestically marketed products. Since the domestic market usage rate of the above 15 drugs accounts for 70% of all monoclonal antibody drugs on the market, they have strong detection representativeness. Therefore, the neural network model constructed based on the above 15 drugs can effectively cover them.

[0067] In this embodiment, the Raman spectrometer used is the American Thermo scientific DXR micro Raman spectrometer, which is equipped with a 532nm laser, a laser energy of 10mW, an aperture slit of 50μm, a grating of 900lines / mm, and a scanning range of 56cm. -1 ~3350cm -1 The single acquisition exposure time is 10s, and the number of exposures for a single sample is 20 times.

[0068] As an optional implementation, the method for acquiring Raman spectrum data includes:

[0069] Take an appropriate amount of the drug and place it in a 200μL quartz dish, so that the liquid surface is flush with the mouth of the quartz dish, and then collect the Raman spectrum data of the drug. For a single drug, collect its 3 different production batches, and generate 60 sets of Raman spectrum data for the drug in each production batch. Repeat the above steps to generate the first data set and the second data set. Among them, when the above 15 monoclonal antibody drugs are selected, the Raman spectrum graph generated is as follows Figure 3 shown. Figure 3 The displayed results show that in this embodiment, the above acquisition method has good response values ​​for the above 15 monoclonal antibody drugs, without the need to apply surface enhancement technology or use any other pre-treatment methods, and the direct sampling operation step takes about 30 seconds, which makes the present invention have a higher detection speed.

[0070] In a preferred embodiment, Figure 4 As shown, the preprocessing method in step S2 includes:

[0071] Step S21: intercepting the characteristic area of ​​the Raman spectrum data to generate an intercepted spectrum;

[0072] Step S22: extracting the spectrum range of the intercepted spectrum to generate a filtered image;

[0073] Step S23: extracting spectrum peak extreme points from the filtered image to generate a spectrum feature array.

[0074] Specifically, Figure 5A , Figure 5B and Figure 5C As shown, in the embodiment of selecting the above 15 kinds of drugs, the 2700 collected Raman spectra were intercepted to generate a spectrum range of 153cm -1 ~1792cm -1 The intercepted image is then converted to the frequency domain using Fourier transform. After intercepting the 0-200Hz frequency band in the frequency domain, the inverse Fourier transform is used to convert it back to the time domain to complete the filtering and smoothing operation. Then, by extracting the peak extreme points of the filtered spectrum, the position serial number of each extreme point is recorded and the serial numbers are arranged in reverse order according to the corresponding intensity of the extreme point to generate a spectrum feature array. When the number of extreme point serial numbers exceeds 50, the first 50 serial numbers are taken. When the number of serial numbers is less than 50, "0" is used to fill it to 50 bits to meet the input characteristics of the long short-term memory network.

[0075] In a preferred embodiment, the spectrum range extracted in step S22 is 153 cm -1 ~1792cm -1 .

[0076] In a preferred embodiment, Figure 6 As shown, the spectrum range extraction in step S22 includes:

[0077] Step S221: performing Fourier transform on the intercepted spectrum to generate a frequency domain image;

[0078] Step S222: extracting a 1-200 Hz frequency band from the frequency domain image, and then performing an inverse Fourier transform to generate a filtered image.

[0079] In a preferred embodiment, Figure 7 As shown, the neural network model includes:

[0080] Embedding layer 1, the embedding layer receives the graph feature array and reduces the dimension of the graph feature array to generate a reduced dimension array;

[0081] Long short-term memory network 2, the long short-term memory network receives the reduced dimension array and generates a recursive result;

[0082] Fully connected layer 3, the fully connected layer receives the recursive results and outputs the classification results.

[0083] Specifically, during the training process of the neural network model, the first data set and the second data set are exported in a 1:1 ratio of the total number, and randomly divided into a training set and a validation set in a 4:1 ratio. The training set is used to train the above neural network model, and the validation set is used for cross-validation to obtain a better training effect. Among them, the number of long short-term memory networks 2 is 70, the fully connected layer 3 is output as a single hidden layer, and the number of nodes in the fully connected layer 3 is 37. Based on the above settings, a better neural network model training effect can be achieved.

[0084] Furthermore, if Figure 8 As shown in the figure, when the number of epochs is equal to 5, the model loss is 0.0001, the prediction accuracy is 0.999, and the recall rate is 1. The neural network model at this time is the optimal model, that is, when the number of iterations of the long short-term memory network is 5, the output neural network model is the optimal model.

[0085] In a preferred embodiment, in step S1, a Raman spectrometer is used to collect Raman spectrum data;

[0086] The Raman spectrometer has a laser wavelength of 532nm, a laser energy of 10mW, a grating of 900lines / mm, and a spectral resolution of less than 8.3cm -1 , spectral scanning range 58cm -1 ~3350cm -1 .

[0087] In a preferred embodiment, in step A1, three production batches of non-monoclonal antibody drugs or monoclonal antibody drugs are sampled respectively;

[0088] 60 Raman spectra data were collected for each production batch.

[0089] Specifically, by sampling three production batches of drugs, the impact of differences in a single batch of drugs on the results can be effectively avoided, and by collecting 60 Raman spectral data to obtain more samples, the accuracy of judgment can be further improved.

[0090] In a preferred embodiment, the spectrum feature array is 50 bits.

[0091] Specifically, by setting the number of bits of the spectrum feature array to 50 bits, the input layer of the neural network model can be better matched. To achieve this effect, in the specific implementation process, when the number of spectrum peak extreme value points extracted from the filtered image is greater than 50, the first 50 bits are taken to generate the spectrum feature array; when the number of spectrum peak extreme value points extracted from the filtered image is less than 50, the remaining bits are set to 0 to generate the spectrum feature array.

[0092] As an optional implementation, in step S3, a neural network model is used to calculate the probability value of the classification of each drug, and the one with the largest probability and a probability value greater than 0.5 is selected as the classification result output.

[0093] As an optional implementation, the output layer of the neural network model is 1 and 0, which are respectively used to represent monoclonal antibody drugs of this variety and monoclonal antibody drugs of other varieties.

[0094] Furthermore, in order to verify the higher accuracy of the present invention compared with the prior art, 20 Raman spectra of 15 kinds of monoclonal antibody samples were collected respectively, and after being exported as csv files, the spectrum matching algorithms built in the present invention and Thermo Fisher OMNIC8 software were used for identification, and the accuracy results were recorded.

[0095] The results are as follows Fig. 9As shown in the figure, the recognition scores of the method and OMNIC 8.3 software for 15 monoclonal antibody drugs are respectively 100 and 80 for adalimumab monoclonal antibody injection (No. 0), 100 and 85 for bevacizumab monoclonal antibody injection (No. 1), 100 and 95 for ranibizumab monoclonal antibody injection (No. 2), 100 and 15 for tocilizumab monoclonal antibody injection (No. 3), 100 and 80 for elouvir monoclonal antibody injection (No. 4), 100 and 80 for secukinumab monoclonal antibody injection (No. 5), 100 and 90 for rituximab monoclonal antibody injection (No. 6), 100 and 85 for trastuzumab monoclonal antibody injection (No. 7), and 100 and 95 for trastuzumab monoclonal antibody injection (No. 8). Monoclonal antibody injection (No. 7) 100 and 95, Pertuzumab monoclonal antibody injection (No. 8) 100 and 0, Dishu monoclonal antibody injection (No. 9) 100 and 90, Equizumab monoclonal antibody injection (No. 10) 100 and 55, Ustena monoclonal antibody injection (No. 11) 100 and 90, Gusekiu monoclonal antibody injection (No. 12) 100 and 50, Amicizumab monoclonal antibody injection (No. 13) 100 and 95, Etanercept monoclonal antibody injection (No. 14) 100 and 80. The recognition scores of the 15 monoclonal antibodies by this method were all 100. The matching scores of OMNIC 8.3 software for different varieties were quite different. The recognition scores of the 15 polyclonal antibody drugs by this method were all better than those of OMNIC 8.3 software, indicating that the method has higher accuracy than the existing technology.

[0096] The beneficial effects of the present invention are:

[0097] 1. Using commercially available monoclonal antibody drugs for spectrum collection can obtain a large amount of authentic Raman spectral data. Collecting water, saline, glucose injection and other non-monoclonal antibody drug spectra for model training can obtain better model generalization ability;

[0098] 2. By extracting the unique spectral feature area of ​​monoclonal antibody drugs in the range of 153cm-1 to 1792cm-1 for identification, the feature recognition ability of low-concentration monoclonal antibody drugs can be greatly improved compared with the existing complete spectrum similarity comparison scheme;

[0099] 3. Use Fourier transform to convert the extracted characteristic area data in the range of 153cm-1 to 1792cm-1 into frequency domain signals, extract the 0-200Hz frequency band, and perform inverse Fourier transform to obtain spectrum filtering smoothing and dimensionality reduction effects at the same time, reduce the impact of spectrum noise and redundant data points on subsequent calculations, and improve recognition accuracy and computing efficiency;

[0100] 4. By extracting the position serial number of the peak extreme point of the characteristic spectrum, the effect of spectrum data normalization and feature extraction can be achieved in a simple calculation method, thereby improving the calculation efficiency; the serial numbers are then arranged in descending order according to the spectrum peak intensity of the serial number position, so that the consideration of the spectrum peak signal intensity characteristics can be incorporated into the feature array, making full use of the spectrum characteristics; for arrays with a total number of serial numbers less than 50, the end is padded with 0 to 50 bits to form a spectrum feature array, and for arrays with a total number of serial numbers exceeding 50, the first 50 bits are taken to form a spectrum feature array, so that the feature array can be compatible with the input layer of the long short-term memory artificial neural network model;

[0101] 5. The long short-term memory artificial neural network model is used to identify the feature array after Raman spectrum extraction, which can realize intelligent judgment of the spectrum. When the detection conditions change within a certain range and cause differences in the spectrum, the long short-term memory artificial neural network model has good model robustness and can still make accurate judgments.

[0102] The above are only preferred embodiments of the present invention, and are not intended to limit the implementation methods and protection scope of the present invention. Those skilled in the art should be aware that all solutions obtained by equivalent substitutions and obvious changes made using the description and illustrations of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for detecting a monoclonal antibody drug, characterized in that: include: Step S1: collecting Raman spectrum data of the drug to be tested; Step S2: applying a preprocessing method to the Raman spectrum data to generate a spectrum feature array; Step S3: using a pre-built neural network model to classify the graph feature array to generate a classification result; The drug to be tested is a monoclonal antibody drug; The pre-processing method in step S2 includes: Step S21: intercepting the characteristic area of ​​the spectrum of the Raman spectrum data to generate an intercepted spectrum; Step S22: extracting the spectrum range of the intercepted spectrum to generate a filtered image; Step S23: extracting spectrum peak extreme points from the filtered image to generate the spectrum feature array; The step S23 comprises: extracting the peak extreme points of the filtered spectrum from the filtered image, recording the position serial number of each extreme point and arranging the serial numbers in reverse order according to the corresponding intensity of the extreme point to generate the spectrum feature array; The spectrum range extraction in step S22 includes: Step S221: performing Fourier transform on the intercepted spectrum to generate a frequency domain image; Step S222: extracting a frequency band of 0-200 Hz from the frequency domain image, and then performing an inverse Fourier transform to generate the filtered image.

2. The detection method according to claim 1, characterized in that: The method for constructing the neural network model in step S3 includes: Step A1: collecting the Raman spectral data of a plurality of non-monoclonal antibody drugs to generate a first data set, and collecting the Raman spectral data of a plurality of monoclonal antibody drugs to generate a second data set; Step A2: using the preprocessing method to generate a plurality of first arrays and a plurality of second arrays for the first data set and the second data set; The number of the first arrays is the same as the number of the second arrays; Step A3: generating a training set and a validation set according to a plurality of the first arrays and a plurality of the second arrays; Step A4: Generate a model to be verified based on the training set, and then perform cross-validation based on the model to be verified and the verification set to generate the neural network model.

3. The detection method according to claim 1, characterized in that: The spectrum range extracted in step S22 is 153 cm -1 ~1792cm -1 .

4. The detection method according to claim 1, characterized in that: The neural network model includes: An embedding layer, the embedding layer receiving the graph feature array and performing dimensionality reduction on the graph feature array to generate a reduced dimensionality array; A long short-term memory network, the long short-term memory network receiving the reduced-dimensional array and generating a recursive result; A fully connected layer receives the recursive result and outputs the classification result.

5. The detection method according to claim 1, characterized in that: In the step S1, a Raman spectrometer is used to collect the Raman spectrum data; The Raman spectrometer has a laser wavelength of 532 nm, a laser energy of 10 mW, a grating of 900 lines / mm, and a spectral resolution of less than 8.3 cm -1 , spectral scanning range 58cm -1 ~3350cm -1 .

6. The detection method according to claim 2, characterized in that: In the step A1, three production batches of the non-monoclonal antibody drug or the monoclonal antibody drug are sampled respectively; 60 sets of Raman spectral data were collected for each production batch.

7. The detection method according to claim 1, characterized in that: The spectrum feature array has 50 bits.

Images