Model-driven intelligent analysis system for drug crystal forms

Abstract

The present application relates to the technical field of drug crystal form intelligent analysis, in particular to a drug crystal form intelligent analysis system based on model driving, which comprises a Raman scattering diagram construction module, a crystal form characteristic peak intelligent analysis module and a crystal form intelligent recognition closed loop verification module, wherein: the Raman scattering diagram construction module is used for constructing a Raman spectrum diagram; the crystal form characteristic peak intelligent analysis module is used for judging whether a characteristic peak is a crystal form characteristic peak, defining a crystal form characteristic diagram, setting a peak position and a peak intensity weight coefficient, sequentially calculating a comprehensive matching degree, and determining a drug crystal form of the characteristic peak in the crystal form characteristic diagram; and the crystal form intelligent recognition closed loop verification module is used for establishing a crystal form analysis model, judging a crystal form type when a drug crystal form cannot be recognized, adjusting a method for judging a crystal form characteristic peak in the crystal form characteristic peak intelligent analysis module, adjusting the peak position and the peak intensity weight coefficient according to the crystal form type, and again calculating a drug comprehensive matching degree; if the drug crystal form still cannot be recognized, an unrecognizable signal is output.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent analysis of drug crystal forms, and in particular to a model-driven intelligent analysis system for drug crystal forms. Background Art

[0002] Drug crystal forms refer to the different crystal forms formed when drug molecules form a solid state due to different molecular arrangements and stacking structures. The same drug may have multiple crystal forms, and the differences between the multiple crystal forms at the molecular level will be reflected through the vibration modes of the chemical bonds within the molecules (such as stretching and bending vibrations). In actual scenarios, drug crystal forms can be divided into normal crystal forms and mixed crystal forms. Normal crystal forms have regular molecular arrangement, and their peak positions are determined by the inherent vibrations of chemical bonds, which is more critical for identifying the nature of the crystal form. Mixed crystal forms are caused by the superposition of multiple crystal forms, and the peak intensity is easily affected by the crystal proportions, mixing state, etc., and the peak position may deviate due to fusion.

[0003] In order to accurately identify the crystal form of a drug, the existing technology usually uses the Raman scattering phenomenon generated by emitting a laser of a specific wavelength to collide with the drug molecules, and uses the laser photons to separate the Stokes scattered light, calculate the frequency difference between the Stokes scattered light and the incident light to obtain the Raman shift wave number, and convert the Stokes scattered light into Raman shift light intensity through the photoelectric effect. Then, a Raman spectrum is constructed with the Raman shift wave number as the horizontal axis and the Raman shift light intensity as the vertical axis. Then, the characteristic peaks in the Raman spectrum are extracted by setting a threshold, and then the extracted characteristic peaks are compared one by one with the characteristic peaks in the standard spectrum to distinguish the drug crystal form corresponding to the characteristic peaks in the Raman spectrum.

[0004] However, impurities are easily introduced in drug preparation (such as residual raw materials and intermediates from incomplete synthesis reactions) and storage (such as degradation reactions triggered by the storage environment) due to insufficient raw material purity, generation of reaction by-products, and environmental factors-induced degradation. These impurity molecules have unique vibration modes and will produce additional worthless characteristic peaks in the Raman spectrum, interfering with the identification of characteristic peaks of the crystal form. At the same time, drug crystal forms are highly complex. Mixed crystal forms may cause abnormal peak heights and widened half-peak widths due to crystal form superposition, and defective crystal forms may reduce peak spacing due to disordered molecular arrangement. The characteristic peaks of different crystal forms (such as mixed crystal forms and defective crystal forms) may show special distributions in peak height, half-peak width, and peak spacing, resulting in misjudgment or omission when setting thresholds to extract characteristic peaks in the Raman spectrum, which will further lead to the inability to distinguish the corresponding drug crystal forms during the comparison process. In view of this, we propose a model-driven intelligent analysis system for drug crystal forms. Summary of the Invention

[0005] The present application aims to solve the problem that impurities are easily introduced in the preparation and storage of drugs, and the molecular vibration generates additional characteristic peaks to interfere; the characteristic peaks of mixed and defective crystal forms are specially distributed in peak height, half-peak width and peak spacing, and the general threshold extraction is easy to misjudge or miss detection, resulting in the problem that the corresponding drug crystal form cannot be distinguished when compared.

[0006] To achieve the above-mentioned purpose, the present application provides a model-driven intelligent analysis system for drug crystal form, which comprises a Raman scattering diagram construction module, an intelligent analysis module for crystal form characteristic peaks and a closed-loop verification module for intelligent identification of crystal form, wherein:

[0007] The Raman scattering diagram construction module is used to construct a Raman spectrum diagram; the intelligent analysis module for crystal form characteristic peaks is used to determine whether the characteristic peaks are crystal form characteristic peaks, and if not, the other characteristic peaks are removed to define a crystal form characteristic diagram; the peak position and peak intensity weight coefficient are set, and the comprehensive matching degree of the characteristic peaks in the crystal form characteristic diagram and the characteristic peaks in the professional crystal database is calculated in turn, if the comprehensive matching degree is greater than or equal to the crystal judgment threshold, it is determined that the drug crystal form of the characteristic peaks in the crystal form characteristic diagram is the corresponding drug crystal form in the professional crystal database, otherwise it is determined that the drug crystal form cannot be identified;

[0008] The closed-loop verification module for intelligent identification of crystal form is used to establish a crystal analysis model, receive a signal that the drug crystal form cannot be identified, determine the crystal type when the drug crystal form cannot be identified, adjust the method for determining the crystal form characteristic peaks in the intelligent analysis module for crystal form characteristic peaks, adjust the peak position and peak intensity weight coefficient according to the crystal type, and calculate the comprehensive matching degree of the drug again, if the drug crystal form still cannot be identified, output the signal that the drug crystal form cannot be identified.

[0009] As a further improvement of the present technical solution, the Raman scattering diagram construction module emits specific wavelength laser to the drug sample to be detected, and the laser photons collide with the drug molecules in a dynamic process; in the collision process, the laser photons produce elastic scattering and Raman scattering; when the photon energy matches the vibration energy level difference of the drug molecules, the Raman scattering appears to be offset, forming mixed frequency Raman scattering light, which includes anti-Stokes scattering light and Stokes scattering light.

[0010] As a further improvement of the present technical solution, the Raman scattering diagram construction module separates multiple Stokes scattering lights from the mixed frequency Raman scattering light by using a grating: after the laser photons collide with the drug molecules, the mixed frequency Raman scattering light passes through the grating, and multiple Stokes scattering lights are separated from the Raman scattering light due to the difference in grating diffraction angle;

[0011] The difference between the frequency of each Stokes scattered light and the frequency of the incident light is calculated as the Raman shift wavenumber, and each Stokes scattered light is converted into an electrical signal through the photoelectric effect, that is, the Raman shift intensity of each Stokes scattered light; with the Raman shift wavenumber as the horizontal axis and the Raman shift intensity as the vertical axis, multiple Stokes scattered lights are constructed into a Raman spectrum.

[0012] The beneficial effect of the above-mentioned further scheme is to provide high-quality spectral data for subsequent crystal form analysis, ensure the accuracy of characteristic peak extraction and crystal form identification, effectively deal with the interference caused by impurities and complex crystal forms in drug preparation and storage, and improve the reliability and accuracy of the intelligent analysis system of drug crystal forms.

[0013] On the basis of the above technical solution, the present invention can also be improved as follows:

[0014] The crystal form characteristic peak intelligent analysis module includes a crystal form characteristic peak threshold screening unit and a crystal form matching degree multi-parameter fusion calculation unit; the crystal form characteristic peak threshold screening unit receives a Raman spectrum, and uses the light value of the Raman shift light intensity corresponding to the characteristic peak vertex in the Raman spectrum minus the baseline light intensity at the position of the characteristic peak as the peak height of the characteristic peak; when the light intensity on the right side of the characteristic peak is half of the peak height, the difference between the high wavenumber and the low wavenumber of the corresponding Raman shift wavenumber is the half-peak width; the absolute value of the difference between the Raman shift wavenumbers of the two characteristic peak vertices is the peak spacing.

[0015] As a further improvement of the present technical solution, the crystal form characteristic peak threshold screening unit sets a peak height threshold, a half-peak width threshold, and a peak spacing threshold. If a characteristic peak in the Raman spectrum has a corresponding peak height greater than the peak height threshold, the half-peak width is within the half-peak width threshold, and the peak spacing is ≥ the peak spacing threshold, then the characteristic peak is determined to be a crystal form characteristic peak; otherwise, the characteristic peak is determined to be other characteristic peaks.

[0016] As a further improvement of the present technical solution, the multi-parameter fusion calculation unit for crystal form matching first calculates the peak position matching and the peak intensity matching, calculates the difference between the detected peak position and the standard peak position, and then divides it by the allowed difference, and calculates the average value, which is the peak position matching; the peak intensity matching is to calculate the difference between the detected peak intensity and the standard peak intensity for multiple characteristic peaks, and then divide it by the allowed difference, and calculate the average value, which is the peak intensity matching;

[0017] The comprehensive matching degree is calculated by multiplying the peak position matching degree by the corresponding weight coefficient and adding the peak intensity matching degree by the corresponding weight coefficient.

[0018] The beneficial effect of the above further scheme is that the Raman spectrum is received by the crystal characteristic peak threshold screening unit, and the peak height, half-peak width and peak spacing of the characteristic peak are calculated, and the threshold is set to screen out the crystal characteristic peak, which can accurately eliminate the interference of impurity peaks and invalid peaks, and ensure the purity of the characteristic peaks used in subsequent analysis; the crystal form matching multi-parameter fusion calculation unit first calculates the peak position matching and peak intensity matching, and then calculates the comprehensive matching based on the weight coefficient. Because the peak position is determined by the inherent vibration of the molecular chemical bond, it is more critical to the identification of the crystal form essence. Giving the peak position a higher weight can highlight the core characteristics and reduce the error of the peak intensity caused by interference from the crystal form ratio, mixing state, etc., so that the comprehensive matching is more in line with the real difference of the crystal form, greatly improving the accuracy of crystal form identification, and effectively solving the identification problems caused by impurity interference and the difficulty in distinguishing complex crystal form characteristics, providing reliable support for the precise analysis of drug crystal forms.

[0019] On the basis of the above technical solution, the present invention can also be improved as follows:

[0020] The crystal form intelligent identification closed-loop verification module includes a crystal form machine learning classification model unit and a complex crystal form adaptive identification optimization unit; the crystal form machine learning classification model unit receives the peak intensity of each characteristic peak in the professional crystal database in the crystal form matching multi-parameter fusion calculation unit. and peak positions, and the peak positions are arranged in a fixed order Peak intensity, forming a local eigenvector;

[0021] Learn the relationship between the peak intensities and peak positions corresponding to the characteristic peaks of normal crystal forms and mixed crystal forms: perform linear transformation on the local eigenvectors through the weight matrix. When learning the weight matrix, it automatically assigns high weights to key features based on the impact of different peak position and peak intensity combinations on crystal form classification; add bias adjustment, use activation function to perform nonlinear processing on the linear transformation results, and finally output global correlation features.

[0022] As a further improvement of the present technical solution, the complex crystal form adaptive identification optimization unit receives a crystal form characteristic peak B drug crystal form unrecognizable signal, the crystal form characteristic peak B that cannot be identified by the crystal form matching multi-parameter fusion calculation unit, and the Raman spectrum constructed by the Raman scattering map construction module; sets an adjustment coefficient, calls out the corresponding position of the crystal form characteristic peak B in the Raman spectrum, reduces the peak height threshold in the crystal form characteristic peak threshold screening unit by adjusting the coefficient, increases the half-peak width threshold in the crystal form characteristic peak threshold screening unit, reduces the peak spacing threshold in the crystal form characteristic peak threshold screening unit, and re-selects the crystal form characteristic peak B through the adjusted threshold again, and inputs the re-selected crystal form characteristic peak B into the crystal form analysis model, the crystal form analysis model constructs a corresponding local feature vector, uses a weight matrix to perform a linear transformation on the vector and adds a bias term, and then processes it through the Softmax function to calculate the classification probability;

[0023] If the classification probability is greater than 0.5, the characteristic peak is determined to be a mixed crystal form; if the classification probability is less than or equal to 0.5, the characteristic peak is determined to be a normal crystal form.

[0024] When the complex crystal form adaptive identification and optimization unit analyzes that the crystal form category in the crystal form characteristic diagram is a normal crystal form, a control signal is output to the crystal form matching multi-parameter fusion calculation unit to increase the peak position weight coefficient and reduce the peak intensity weight coefficient in the crystal form matching multi-parameter fusion calculation unit; if the identified crystal form category is a mixed crystal form, the peak position weight coefficient in the crystal form matching multi-parameter fusion calculation unit is reduced and the peak intensity weight coefficient is increased.

[0025] The beneficial effect of the above-mentioned further scheme is that when missed detections and misjudgments occur under the initial threshold due to interference from impurities, special distribution of complex crystal characteristics, etc., the threshold can be dynamically adjusted to recapture the crystal characteristic peaks that are easily missed, thereby broadening the adaptability of this system to complex scenarios (such as mixed crystal forms, defective crystal forms, etc.); by using the model to construct feature vectors and combining them with the Softmax function to accurately calculate the classification probability, it can more scientifically distinguish between mixed crystal forms and normal crystal forms, effectively solving the limitations of traditional single threshold recognition, significantly improving the accuracy and robustness of crystal form recognition under complex conditions, and ensuring the reliability of drug crystal form analysis in a variety of actual scenarios.

[0026] On the basis of the above technical solution, the present invention can also be improved as follows:

[0027] The crystal form intelligent identification closed-loop verification module also includes a crystal form identification result output unit, which is used to sense the comprehensive matching degree and crystal judgment threshold comparison signal. Before comparison, the characteristic peak in the Raman spectrum diagram in the Raman scattering diagram construction module is used to build a crystal form analysis model to the crystal form machine learning classification model unit, and output the crystal form type of the crystal form characteristic diagram. After the crystal form matching degree multi-parameter fusion calculation unit calculates the comprehensive matching degree, it preferentially calls out the professional crystal database and compares the corresponding crystal form types.

[0028] The beneficial effect of the above further scheme is that by giving priority to calling out the corresponding crystal types from the professional crystal database for comparison, the corresponding crystal types can be compared, thereby avoiding indiscriminate full search of the professional crystal database when the multi-parameter fusion calculation unit of the crystal type matching calculates the comprehensive matching degree, reducing the amount of data call and calculation, speeding up the matching process, and making crystal type identification more efficient.

[0029] In addition to the above-described objects, features and advantages, the present invention has other objects, features and advantages. The present invention will be further described in detail below with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 This is the overall module principle diagram of the present invention;

[0031] Figure 2 This is a flow chart of the working principle of the intelligent analysis module of crystal characteristic peaks of the present invention;

[0032] Figure 3 This is a flow chart showing the working principle of the crystal form intelligent identification closed-loop verification module of the present invention;

[0033] Figure 4 It is a flow chart of the overall working principle of the present invention;

[0034] Figure 5 Schematic diagram of normal drug crystal form and mixed drug crystal form in the present invention.

[0035] The meaning of each number in the figure is:

[0036] 100. Raman scattering image construction module; 200. Crystal form characteristic peak intelligent analysis module; 210. Crystal form characteristic peak threshold screening unit; 220. Crystal form matching degree multi-parameter fusion calculation unit; 300. Crystal form intelligent identification closed-loop verification module; 310. Crystal form machine learning classification model unit; 320. Complex crystal form adaptive identification and optimization unit; 330. Crystal form identification result output unit. DETAILED DESCRIPTION

[0037] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. 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 making any creative efforts shall fall within the scope of protection of the present invention.

[0038] refer to Figure 1-Figure 5 As shown, the model-driven drug crystal form intelligent analysis system includes a Raman scattering image construction module 100, a crystal form characteristic peak intelligent analysis module 200 and a crystal form intelligent identification closed-loop verification module 300, wherein:

[0039] The Raman scattering pattern construction module 100 precisely emits a laser of a specific wavelength to the drug sample to be tested, and the laser photons undergo a dynamic collision process with the drug molecules. During the collision process, the laser photons produce elastic scattering (photon energy remains unchanged) and Raman scattering (inelastic scattering, photon energy changes). Because drug molecules have multiple vibration modes (such as chemical bond stretching and bending vibrations), corresponding to different vibration energy levels, when the photon energy matches the molecular vibration energy level difference, it will cause the Raman scattering to shift, forming mixed-frequency Raman scattered light, which includes anti-Stokes scattered light and Stokes scattered light.

[0040] At the same time, since drug molecules (especially polyatomic molecules) do not have only one vibration mode, but have multiple independent vibration modes, the Raman scattering map construction module 100 uses a grating to separate multiple Stokes scattered lights from the mixed-frequency Raman scattered light: after the laser photon collides with the drug molecule, the mixed-frequency Raman scattered light passes through the grating, and due to the difference in diffraction angle, multiple Stokes scattered lights are separated from the mixed-frequency Raman scattered light, following the grating equation: ,in is the grating constant, is the diffraction angle, is the interference level, is the wavelength of incident light;

[0041] The Raman shift of Stokes scattered light is the wave number difference between the incident light and the scattered light, reflecting the characteristics of molecular vibration energy level transitions. The calculation formula is:

[0042]

[0043] is the incident laser wavelength, is the wavelength of Stokes scattered light. Stokes scattered light is converted into an electrical signal by a photodetector (such as a CCD). Its intensity (measured light intensity, the original signal output by the detector, influencing factors include the scattering cross-sectional area of ​​the molecular vibration mode, the incident laser power, and the instrument detection sensitivity) reflects the change in the polarizability of the molecular vibration mode. The characteristic peak of each Stokes scattered light is plotted with the Raman shift wave number as the horizontal axis and the Raman shift light intensity (measured light intensity) as the vertical axis;

[0044] At room temperature, most drug molecules are in low vibrational energy levels. When laser photons collide with them, the molecules absorb energy and jump to high energy levels, causing Stokes scattered light. The signal intensity of Stokes scattered light is much higher than the anti-Stokes scattered light generated when the molecules fall back from high energy levels (the latter requires the molecules to be in high energy levels first, which has an extremely low probability and weak signal in the natural state, making it difficult to collect effectively). In addition, the frequency change of Stokes scattered light (Raman shift) is strictly determined by the vibration mode of the drug molecules themselves. Different chemical bonds and functional group vibrations will correspond to unique displacement values, like "molecular fingerprints", which stably reflect structural information. Therefore, Stokes scattered light is used to construct the characteristic peak in the Raman spectrum.

[0045] In order to avoid the presence of impurities in the drug sample, the molecular vibration of the impurities in the drug sample will produce additional frequency differences, resulting in the generation of worthless characteristic peaks in the Raman spectrum constructed by the Raman scattering map construction module 100. Therefore, the crystal form characteristic peak intelligent analysis module 200 includes a crystal form characteristic peak threshold screening unit 210 and a crystal form matching degree multi-parameter fusion calculation unit 220;

[0046] The crystal form characteristic peak threshold screening unit 210 receives the Raman spectrum and calculates the peak height of each characteristic peak in the Raman spectrum. , half-peak width and peak spacing : Peak height It is the difference between the peak apex and the baseline intensity on the vertical axis of the Raman scattering spectrum (the baseline intensity is the fitted straight line on the left and right sides of a randomly selected characteristic peak in the Raman scattering spectrum): ,in is the light intensity value of the Raman shift light intensity corresponding to the characteristic peak apex, is the baseline intensity at the characteristic peak position; half-peak width is the peak width corresponding to half, specifically the width on the horizontal axis of the Raman scattering spectrum: ,in The light intensity on the right side of the characteristic peak = high wavenumbers corresponding to Raman shift wavenumbers; The light intensity on the right side of the characteristic peak = Low wavenumber corresponding to the Raman shift wavenumber; peak spacing The wave number difference of the Raman shift wave number between the two characteristic peak vertices on the horizontal axis ,in 、 is the wave number of the Raman shift wave number at the top of the two characteristic peaks.

[0047] The present invention further analyzes that impurities may be introduced into the drug during drug preparation and storage (such as residual raw materials and intermediates due to incomplete synthesis reactions, degradation reactions induced by the storage environment, etc.) due to insufficient raw material purity, generation of reaction by-products, and degradation induced by environmental factors. Since impurity molecules have unique vibration modes, during Raman spectroscopy detection, they will produce Raman shifts of specific frequencies due to vibrations such as stretching and bending of molecular bonds, thereby forming additional characteristic peaks (i.e., characteristic peaks corresponding to impurities) in the Raman spectrum. In order to avoid interference of impurity characteristic peaks with the identification of drug crystal characteristic peaks, the peak height threshold is set by the crystal characteristic peak threshold screening unit 210. , half-peak width threshold ( , ), peak spacing threshold , if there is a characteristic peak in the Raman spectrum, the corresponding peak height >Peak height threshold , half-peak width threshold lower limit ≤Full width at half peak ≤ Upper limit of half-peak width threshold , and the peak spacing ≥Peak spacing threshold , then the characteristic peak is determined to be a crystal form characteristic peak, otherwise the characteristic peak is determined to be other characteristic peaks, until all the characteristic peaks in the Raman spectrum are compared, and then other characteristic peaks are removed from the Raman spectrum, and the Raman spectrum after removing other characteristic peaks is defined as a crystal form characteristic graph, laying a reliable foundation for the crystal form matching degree multi-parameter fusion calculation unit 220 to identify the drug crystal form.

[0048] The multi-parameter fusion calculation unit 220 senses the professional crystal database and sequentially calculates the comprehensive matching degree between the characteristic peaks in the crystal characteristic graph and the characteristic peaks in multiple spectra (normal spectra and mixed spectra) in the professional crystal database. ,in:

[0049] Set the crystal judgment threshold If the crystal characteristic peak A (characteristic peak A of the spectrum in the professional crystal database) corresponds to the comprehensive matching degree of crystal characteristic peak B (characteristic peak B in the crystal characteristic graph) ≥Crystal judgment threshold , then the drug crystal form with characteristic crystal form peak B is determined to be the drug crystal form corresponding to characteristic crystal form peak A in the professional crystal database; if the comprehensive matching degree <Crystal judgment threshold , it means that the degree of matching between the characteristic peak B of the crystal form and the characteristic peaks of all spectra in the professional crystal database in spectral characteristics (peak position, peak intensity) is low, and does not meet the similarity requirements for being determined as a known crystal form. Therefore, it is determined that the drug crystal form with the characteristic peak B of the crystal form cannot be identified.

[0050] The comprehensive matching degree in the multi-parameter fusion calculation unit 220 of the crystal form matching degree :

[0051] Set peak position matching [Measure the matching degree between the sample to be tested and the standard crystal form in Raman shift wave number (peak position), the weight coefficient is 】、Peak intensity matching [Measure the degree of matching between the sample to be tested and the standard crystal form in terms of peak intensity. The weight coefficient is (1- ) ], and the peak weight coefficient >Peak intensity weight coefficient Secondly, the weight coefficient is used to adjust the peak position (peak position is the Raman shift wave number in the crystal characteristic diagram) matching degree and peak intensity (peak intensity is the Raman shift light intensity in the crystal characteristic diagram) matching degree to the comprehensive matching degree. The impact of

[0052]

[0053] It is a comprehensive matching method that can effectively quantify the crystal form matching degree, taking into account the contribution of peak position and peak intensity, and is suitable for scenarios such as drug polymorph identification;

[0054]

[0055] is the peak position matching degree, is the Raman shift wave number of the i-th characteristic peak of the sample to be tested, is the wave number of the i-th characteristic peak of the standard crystal form, is the allowable wave number deviation range (tolerance value preset by the user), The number of characteristic peaks involved in the matching. If the wave number to be measured is exactly the same as the standard wave number (the difference is 0), this item is 1 (full score). If the difference exceeds When , the result is negative (complete mismatch), and when the result is in the interval [0,1], it means that the matching degree decreases linearly with the deviation;

[0056]

[0057] is the peak intensity matching, is the intensity value of the i-th characteristic peak of the sample to be tested, is the intensity value of the i-th characteristic peak of the standard crystal form, is the number of characteristic peaks involved in the matching. If the measured intensity is completely consistent with the standard, the result is 1. The greater the deviation, the lower the matching degree (if the measured intensity is 0, the result is 0);

[0058] Peak position matching in the crystal form matching multi-parameter fusion calculation unit 220 It reflects the degree of fit of the characteristic peak wave number of the drug crystal form, and can define the crystal form from the essence of molecular chemical bond vibration, peak intensity matching Reflecting the difference in light intensity can supplement the information brought by test fluctuations, crystal stacking, etc., and the peak position in drug crystal form analysis is more essential for crystal form identification (drug molecules are composed of various chemical bonds (such as CC, C=O, CH, etc.), different chemical bonds have unique vibration modes, and the peak position is determined by the vibration frequency of the chemical bonds within the molecule). Therefore, the weight coefficient , so that the peak position matching dominates the comprehensive results, through the peak position matching Peak intensity matching Weighted fusion calculates the comprehensive matching degree It not only grasps the core basis of crystal form identification, but also takes into account the complex situations of actual testing, so as to achieve more accurate, stable and drug R&D-oriented crystal form matching judgment, effectively avoiding the limitations of a single parameter (peak position is easily misjudged due to common overlap of functional groups, and peak intensity is easily affected by test conditions), and allowing crystal form identification to move from experience to data and standardization. Through the crystal form matching multi-parameter fusion calculation unit 220, the reliability and efficiency of drug crystal form analysis in R&D, quality inspection and other links are improved.

[0059] In order to more accurately identify drug crystal forms, the crystal form intelligent identification closed-loop verification module 300 includes a crystal form machine learning classification model unit 310 and a complex crystal form adaptive identification optimization unit 320;

[0060] The crystal form machine learning classification model unit 310 receives the professional crystal database obtained by the crystal form matching multi-parameter fusion calculation unit 220,

[0061] Among them, the peak intensity of the characteristic peak of each spectrum in the professional crystal database is received Peak position , the peaks are arranged in a fixed order and the corresponding peak intensity , forming a local feature vector ,in is the peak wave number, Peak strength;

[0062] Learn the relationship between the peak intensity and peak position of the characteristic peaks of normal crystal form and mixed crystal form: through the weight matrix For local eigenvectors Do linear transformation, weight matrix During learning, based on the influence of different peak positions and peak intensity combinations on crystal classification, high weights are automatically assigned to key features to highlight effective correlations; plus the bias term Adjustment, and then use activation function to perform nonlinear processing on the linear transformation results to establish a crystal form analysis model. The crystal form analysis model classifies the characteristic peaks into normal crystal form and mixed crystal form by learning the peak position and peak intensity of each characteristic peak in the normal spectrum and mixed spectrum that have been marked in the professional crystal database, so that the crystal form analysis model can learn more complex associations and finally output global correlation features. , in order to distinguish different crystal forms and assist in drug crystal form analysis, the specific expression is: ,in Is the activation function, used to perform nonlinear processing on the linear transformation results , enhancing the ability of the crystal analysis model to learn complex associations.

[0063] The present invention further considers that the peak height threshold value set in the crystal characteristic peak threshold screening unit 210 , half-peak width threshold value (FWHM) , , peak interval threshold value are formulated based on the general characteristic peak rules adapted to a large number of conventional cases. However, drug crystal forms have high complexity, and the characteristic peaks of different crystal forms (such as mixed crystal forms, new crystal forms, and crystal forms containing defects) can present special distributions in peak height, half-peak width, and peak interval (such as the abnormal peak height of mixed crystal forms due to crystal form superposition, the broadening of half-peak width, and the reduction of peak interval due to the disordered arrangement of molecules in crystal forms containing defects), which can cause the following problems: when the characteristic peaks are identified by the crystal form characteristic peak threshold value screening unit 210, some characteristic peaks belonging to the drug crystal form (such as the weak peaks of the secondary crystal form in the mixed crystal form and the special wide peaks of the new crystal form) are mistakenly determined as non-crystal form characteristic peaks and are removed, and finally the crystal form matching degree multi-parameter fusion calculation unit 220 cannot identify the crystal form type due to the lack of key characteristic peak data, causing crystal form identification errors or failures;

[0064] Therefore, the complex crystal form adaptive identification optimization unit 320 receives the drug crystal form unidentifiable signal of the crystal form characteristic peak B, the crystal form characteristic peak B that cannot be identified by the crystal form matching degree multi-parameter fusion calculation unit 220, and the Raman spectrum constructed by the Raman scattering diagram construction module 100; sets an adjustment coefficient, calls out the position corresponding to the crystal form characteristic peak B in the Raman spectrum, subtracts the adjustment coefficient by 1, and then multiplies the peak height threshold value , thereby reducing the peak height threshold value in the crystal form characteristic peak threshold value screening unit 210; adds the adjustment coefficient by 1, and then multiplies the half-peak width threshold value (FWHM) , , thereby increasing the half-peak width threshold value (FWHM) , in the crystal form characteristic peak threshold value screening unit 210; subtracts the adjustment coefficient by 1, and then multiplies the peak interval threshold value , thereby reducing the peak interval threshold value in the crystal form characteristic peak threshold value screening unit 210; the crystal form characteristic peak B is reselected through the adjusted threshold value, and the reselected crystal form characteristic peak B is input into the crystal form analysis model. The crystal form analysis model constructs the corresponding local feature vector and calculates the classification probability , wherein is the output layer weight of the crystal form analysis model, is the output layer weight of the crystal form analysis model;

[0065] If the classification probability > 0.5, the characteristic peak is determined to be a mixed crystal form. If the classification probability ≤ 0.5, the characteristic peak is determined to be a normal crystal form.

[0066] By adjusting the threshold in the crystal form characteristic peak threshold screening unit 210, the problem of missed detection and misjudgment of characteristic peaks caused by impurity interference and other reasons is solved, and the accuracy and robustness of crystal form identification in complex scenarios are guaranteed.

[0067] If the comprehensive matching degree in the crystal form matching degree multi-parameter fusion calculation unit 220 <Crystal judgment threshold When the drug crystal form in the crystal form characteristic diagram cannot be determined, the complex crystal form adaptive recognition optimization unit 320 receives the characteristic peak category determined by the crystal form machine learning classification model unit 310:

[0068] If the crystal type in the crystal characteristic diagram is normal, a control signal is output to the crystal matching multi-parameter fusion calculation unit 220 to adjust the peak position weight coefficient in the crystal matching multi-parameter fusion calculation unit 220. for , the peak intensity weight coefficient is ,and , ,in is the weight adjustment coefficient, that is, increasing the peak position weight coefficient and reducing the peak intensity weight coefficient in the crystal form matching multi-parameter fusion calculation unit 220, recalculating the comprehensive matching degree, and outputting a comparison signal between the comprehensive matching degree and the crystal judgment threshold;

[0069] If the crystal type is identified as a mixed crystal type, a control signal is output to the crystal type matching multi-parameter fusion calculation unit 220 to adjust the peak position weight coefficient in the crystal type matching multi-parameter fusion calculation unit 220. for , peak intensity weight coefficient for That is, the peak position weight coefficient in the crystal form matching multi-parameter fusion calculation unit 220 is reduced, the peak intensity weight coefficient is increased, the comprehensive matching degree is recalculated, and a comparison signal of the comprehensive matching degree and the crystal judgment threshold is output. If the crystal form matching multi-parameter fusion calculation unit 220 still cannot identify the drug crystal form after adjusting the weight coefficient according to the crystal form category, an unrecognizable signal is output;

[0070] Because the crystal forms in the mixed crystal form are fused with each other, the peak intensity is affected by the crystal form ratio, mixing state, etc., and has a weak reaction to the crystal form nature (intramolecular chemical bond vibration). The peak intensity weight needs to be increased to Assist in identifying mixed features and correcting the deviation of peak positions caused by fusion;

[0071] Normal crystal molecules are arranged regularly, and the peak position is determined by the inherent vibration of the chemical bond, which is more critical for identifying the nature of the crystal form. Therefore, the peak position weight is increased to The function is to make the weights adapt to the characteristics of different crystal forms, accurately identify mixed crystal forms (using peak intensity to supplement mixed information) and normal crystal forms (relying on peak position to grasp the essence), and the purpose is to improve the recognition accuracy and adaptability of the crystal form analysis system to complex crystal form scenarios (including mixed and normal crystal forms), avoid misjudgment of different crystal form categories by a single weight, make the comprehensive matching degree calculation more in line with the real characteristics of the crystal form, and ensure the reliability of drug crystal form analysis and judgment.

[0072] When the comprehensive matching degree is calculated by the weight coefficient adjusted by the complex crystal form adaptive recognition optimization unit 320, when the comprehensive matching degree of the drug is calculated again, the weight coefficient that has not been adjusted in the original crystal form matching degree multi-parameter fusion calculation unit 220 is used for calculation; since the initial weight (peak position weight coefficient) of the crystal form matching degree multi-parameter fusion calculation unit 220 is used in the normal calculation of the comprehensive matching degree , peak intensity weight coefficient , because the weight is set based on the general crystal form recognition logic - the peak position is determined by the nature of the chemical bond vibration within the crystal form molecule, and is more specific for distinguishing crystal form categories (the difference in peak position of different crystal forms directly reflects the difference in molecular structure), giving priority to ensuring a high weight ratio of peak position matching, which is in line with the analysis logic of "taking molecular structure characteristics as the core" in conventional crystal form recognition, and can stably adapt to the basic recognition scenarios of most normal crystal forms and mixed crystal forms; the weight adjusted by the complex crystal form adaptive recognition optimization unit 320 is for special scenarios where recognition fails (such as the peak intensity of the mixed crystal form is greatly affected by the proportion / fusion, and the normal The "compensatory correction" of crystal form (missing peak positions due to misjudgment of threshold values) belongs to the optimization link of "initial recognition failure → analysis of causes → targeted adjustment" in the recognition process. If the adjusted weights are used in normal calculations, the stability of the general recognition logic will be destroyed. It may also cause deviations in conventional crystal form recognition due to over-adaptation to special scenarios. Therefore, the initial weights of the crystal form matching multi-parameter fusion calculation unit 220 must be used to ensure basic recognition. Only when recognition fails will the complex crystal form adaptive recognition optimization unit 320 intervene to make adjustments, forming a "general → special" hierarchical recognition mechanism to balance recognition accuracy and stability.

[0073] The crystal form intelligent recognition closed-loop verification module 300 also includes a crystal form recognition result output unit 330, which is used to preset high-quality analysis items. When the high-quality analysis item is triggered (the user clicks the high-quality analysis item), when the perceived comprehensive matching degree is and crystal judgment threshold Comparing the signals, the characteristic peaks in the Raman spectrum in the Raman scattering diagram construction module 100 are preferentially converted to the crystal form analysis model established by the crystal form machine learning classification model unit 310, and the crystal form type of the crystal form characteristic diagram is output. The crystal form matching degree multi-parameter fusion calculation unit 220 calculates the comprehensive matching degree. Afterwards, the professional crystal database is preferentially called out, the complex crystal adaptive recognition optimization unit 320 is compared according to the crystal type, a comprehensive matching degree and a crystal judgment threshold comparison signal are output, corresponding crystal type data are compared by preferentially calling out from the professional crystal database, indiscriminate full search of the professional crystal database is avoided when the comprehensive matching degree is calculated by the complex crystal matching degree multi-parameter fusion calculation unit 220, the data calling amount and the calculation amount are reduced, the matching process is accelerated, the crystal recognition is more efficient;

[0074] Wherein, the complex crystal adaptive recognition optimization unit 320 outputs the comprehensive matching degree and the crystal judgment threshold comparison signal, including if the comprehensive matching degree is greater than or equal to the crystal judgment threshold, then the drug crystal type of the characteristic peak in the crystal characteristic map is determined as the corresponding drug crystal type in the professional crystal database;

[0075] If the comprehensive matching degree is less than the crystal judgment threshold, it is judged that the drug crystal type cannot be recognized, and intelligent analysis of the drug crystal type is realized.

[0076] The above shows and describes the basic principles, main features and advantages of the present application. It should be understood by those skilled in the art that the present application is not limited by the above examples, and the above examples and descriptions in the specification are only preferred examples of the present application and are not intended to limit the present application. Without departing from the spirit and scope of the present application, various changes and improvements can be made to the present application, and these changes and improvements all fall within the scope of the present application. The scope of protection of the present application is defined by the appended claims and their equivalents.

Claims

1. Model-driven intelligent analysis system for drug crystal forms, characterized by: The system comprises a Raman scattering diagram construction module (100), a crystal form characteristic peak intelligent analysis module (200), and a crystal form intelligent identification closed-loop verification module (300), wherein: The Raman scattering image construction module (100) is used to construct a Raman spectrum image, and the Raman spectrum image is composed of multiple characteristic peaks; the crystal form characteristic peak intelligent analysis module (200) determines whether the characteristic peak in the Raman spectrum image is a crystal form characteristic peak, and if it is not a crystal form characteristic peak, then the other characteristic peaks are removed and the Raman spectrum image is defined as a crystal form characteristic image; the peak position and peak intensity weight coefficients are set, and the comprehensive matching degree between the characteristic peak in the crystal form characteristic image and the characteristic peak in the professional crystal database is calculated in sequence. If the comprehensive matching degree is greater than or equal to the crystal judgment threshold, then the drug crystal form of the characteristic peak in the crystal form characteristic image is determined to be the corresponding drug crystal form in the professional crystal database, otherwise it is determined that the drug crystal form cannot be identified; The crystal form intelligent identification closed-loop verification module (300) is used to establish a crystal form analysis model, receive a drug crystal form unrecognizable signal, determine the type of crystal form when the drug crystal form cannot be identified, adjust the method for determining the crystal form characteristic peak in the crystal form characteristic peak intelligent analysis module (200), adjust the peak position and peak intensity weight coefficient according to the crystal form type, calculate the comprehensive drug matching degree again, and output an unrecognizable signal if the drug crystal form still cannot be identified.

2. The model-driven intelligent analysis system for drug crystal forms according to claim 1, characterized in that: The Raman scattering pattern building module (100) precisely emits a laser of a specific wavelength to the drug sample to be detected, and the laser photons undergo a dynamic collision process with the drug molecules. During the collision process, the laser photons produce elastic scattering and Raman scattering. When the photon energy matches the vibration energy level difference of the drug molecules, the Raman scattering is offset, forming mixed-frequency Raman scattering light, which includes anti-Stokes scattering light and Stokes scattering light.

3. The model-driven intelligent analysis system for drug crystal forms according to claim 2, characterized in that: The Raman scattering diagram construction module (100) uses a grating to separate multiple Stokes scattered lights from the mixed-frequency Raman scattered light: after the laser photon collides with the drug molecule, when the mixed-frequency Raman scattered light passes through the grating, the multiple Stokes scattered lights are separated from the Raman scattered light due to the difference in the grating diffraction angle; The difference between the frequency of each Stokes scattered light and the frequency of the incident light is calculated as the Raman shift wavenumber, and each Stokes scattered light is converted into an electrical signal through the photoelectric effect, that is, the Raman shift intensity of each Stokes scattered light; with the Raman shift wavenumber as the horizontal axis and the Raman shift intensity as the vertical axis, multiple Stokes scattered lights are constructed into a Raman spectrum.

4. The model-driven intelligent analysis system for pharmaceutical crystal forms according to claim 3, characterized in that: The crystal form characteristic peak intelligent analysis module (200) includes a crystal form characteristic peak threshold screening unit (210) and a crystal form matching degree multi-parameter fusion calculation unit (220); the crystal form characteristic peak threshold screening unit (210) receives a Raman spectrum, and uses the light value of the Raman shift light intensity corresponding to the characteristic peak vertex in the Raman spectrum minus the baseline light intensity at the position of the characteristic peak as the peak height of the characteristic peak; when the light intensity on the right side of the characteristic peak is half of the peak height, the difference between the high wave number of the corresponding Raman shift wave number and the low wave number is the half-peak width; The absolute value of the difference in the Raman shift wavenumbers of the two characteristic peak vertices is the peak distance.

5. The model-driven intelligent analysis system for pharmaceutical crystal forms according to claim 4, characterized in that: The crystal form characteristic peak threshold screening unit (210) sets a peak height threshold, a half-peak width threshold and a peak spacing threshold. If a characteristic peak in the Raman spectrum has a corresponding peak height greater than the peak height threshold and a half-peak width within the half-peak width threshold, and the peak spacing is greater than or equal to the peak spacing threshold, the characteristic peak is determined to be a crystal form characteristic peak; otherwise, the characteristic peak is determined to be another characteristic peak.

6. The model-driven intelligent analysis system for pharmaceutical crystal forms according to claim 4, characterized in that: The crystal form matching degree multi-parameter fusion calculation unit (220) first calculates the peak position matching degree and the peak intensity matching degree, calculates the difference between the detected peak position and the standard peak position, and then divides it by the allowed difference, and calculates the average value, which is the peak position matching degree; Peak intensity matching is to calculate the difference between the detected peak intensity and the standard peak intensity for multiple characteristic peaks, then divide it by the allowed difference and calculate the average value, which is the peak intensity matching; The comprehensive matching degree is calculated by multiplying the peak position matching degree by the corresponding weight coefficient, and adding the peak intensity matching degree by the corresponding weight coefficient. If the similarity requirement is not met, the drug crystal form with the output crystal characteristic peak B cannot be identified.

7. The model-driven intelligent analysis system for pharmaceutical crystal forms according to claim 6, characterized in that: The crystal form intelligent identification closed-loop verification module (300) includes a crystal form machine learning classification model unit (310); The crystal form machine learning classification model unit (310) is used to receive the professional crystal database obtained by the crystal form matching multi-parameter fusion calculation unit (220) to learn the relationship between the corresponding peak intensities and peak positions of the normal crystal form characteristic peaks and the mixed crystal form characteristic peaks: a linear transformation is performed on the local feature vectors through the weight matrix, and when learning, the weight matrix automatically assigns high weights to key features based on the influence of different peak positions and peak intensities on the crystal form classification; in addition, the bias item is adjusted, and the activation function is used to perform nonlinear processing on the linear transformation results to establish a crystal form analysis model. The crystal form analysis model classifies the normal crystal form and the mixed crystal form by learning the peak position and peak intensity of each characteristic peak in the normal spectrum and the mixed spectrum that have been marked in the professional crystal database.

8. The model-driven intelligent analysis system for pharmaceutical crystal forms according to claim 7, characterized in that: The crystal form intelligent identification closed-loop verification module (300) further includes a complex crystal form adaptive identification optimization unit (320) for receiving the unidentifiable signal output by the crystal form matching degree multi-parameter fusion calculation unit (220) and the Raman spectrum constructed by the Raman scattering map construction module (100); and setting an adjustment coefficient, calling out the position corresponding to the crystal form characteristic peak B in the Raman spectrum, lowering the peak height threshold in the crystal form characteristic peak threshold screening unit (210), increasing the half-peak width threshold in the crystal form characteristic peak threshold screening unit (210), and lowering the peak spacing threshold in the crystal form characteristic peak threshold screening unit (210) by adjusting the coefficient, reselecting the crystal form characteristic peak B again by adjusting the threshold, and inputting the reselected crystal form characteristic peak B into the crystal form analysis model, the crystal form analysis model constructing the corresponding local feature vector, performing a linear transformation on the vector using a weight matrix and adding a bias term, and then processing the vector through a Softmax function to calculate the classification probability; If the classification probability is greater than 0.5, the characteristic peak is determined to be a mixed crystal form; if the classification probability is less than or equal to 0.5, the characteristic peak is determined to be a normal crystal form.

9. The model-driven intelligent analysis system for pharmaceutical crystal forms according to claim 8, characterized in that: When the complex crystal form adaptive identification optimization unit (320) determines that the complex crystal form is a normal crystal form, a control signal is output to the crystal form matching degree multi-parameter fusion calculation unit (220), the peak position weight coefficient in the crystal form matching degree multi-parameter fusion calculation unit (220) is increased and the peak intensity weight coefficient is reduced, the comprehensive matching degree is recalculated, and a comparison signal of the comprehensive matching degree and the crystal judgment threshold is output; When the complex crystal form adaptive identification optimization unit (320) determines that the complex crystal form is a mixed crystal form, the peak position weight coefficient in the crystal form matching multi-parameter fusion calculation unit (220) is reduced, the peak intensity weight coefficient is increased, the comprehensive matching degree is recalculated, and a comparison signal of the comprehensive matching degree and the crystal judgment threshold is output.

10. The model-driven intelligent analysis system for pharmaceutical crystal forms according to claim 7, characterized in that: The crystal form intelligent identification closed-loop verification module (300) also includes a crystal form identification result output unit (330), which is used to preset high-quality analysis items. When the high-quality analysis items are triggered, the characteristic peaks in the Raman spectrum in the Raman scattering image construction module (100) are preferentially input into the crystal form analysis model established by the crystal form machine learning classification model unit (310), and the crystal form type of the crystal form characteristic image is output, so that the complex crystal form adaptive identification optimization unit (320) can be matched according to the crystal form type.

Images