A method for detecting resin loading capacity by multi-modal fusion near-infrared spectroscopy

Abstract

The application discloses a kind of multi-modal fusion near infrared spectroscopy detection resin load method, the method comprises: obtaining the near infrared spectral data of resin and at least one auxiliary modal data;Respectively extract spectral features and auxiliary modal features;Multi-level fusion is carried out to both, load is predicted based on fusion features;The multi-level fusion includes feature level fusion and / or decision level fusion, and the auxiliary modal data is used to characterize resin physical state, environmental condition or batch source, to compensate or eliminate the interference of non-chemical factors on near infrared signal.The present application takes near infrared spectroscopy as main modal, and introduces image, moisture content, temperature and humidity, sample information, batch information or other spectral data as auxiliary modal, can use multi-source information to jointly explain resin load change and near infrared scattering change source, detection speed is fast, accuracy is high, robustness is strong, suitable for resin into factory inspection and production process quality control.

Description

Technical Field

[0001] This invention relates to the field of multimodal near-infrared spectroscopy detection technology, specifically a method for detecting resin loading using multimodal fusion near-infrared spectroscopy. Background Technology

[0002] Solid-phase peptide synthesis typically uses functionalized carrier resins as starting supports. The target peptide is obtained through the reaction of active sites on the resin with protected amino acids or linkers, followed by repeated deprotection, coupling, washing, and cleavage steps. The loading amount of the carrier resin is a crucial parameter affecting the calculation of feedstock, coupling efficiency, deletion sequence ratio, crude peptide purity, and final yield. Excessive loading may lead to increased steric hindrance, incomplete coupling, and increased side reactions, while insufficient loading may reduce the yield per unit mass of resin and increase production costs.

[0003] Currently, the determination of resin loading typically relies on chemical detection methods such as Fmoc deprotection UV absorption, HPLC, and gravimetric methods. While these methods can serve as a source of standard values, they generally suffer from drawbacks such as cumbersome sample pretreatment, long detection cycles, the need for organic solvents and deprotection reagents, high dependence on operators, and difficulty in achieving rapid batch detection.

[0004] Near-infrared spectroscopy can reflect the overtone and combination frequency absorption information of chemical bonds such as CH, NH, OH, and C=O in a sample. It has advantages such as fast detection speed, simple sample preparation, non-destructive or low-destructive detection, and suitability for online or near-line analysis. However, peptide solid-phase synthesis carrier resins usually have characteristics such as significant particle scattering, sensitivity to water content, large differences in particle size and sample loading state, and relatively low content of functionalized groups. When relying solely on near-infrared spectroscopy, they are easily affected by factors such as moisture, particle size, color, sample thickness, surface smoothness, temperature and humidity, and instrument drift.

[0005] In actual production quality control, information such as resin batch, supplier, storage time, functionalization reaction conditions, sample quality, sample loading form, and ambient temperature and humidity can often explain scattering differences or non-chemical changes in near-infrared spectra; visible light or microscopic images can reflect particle color, particle size, texture, and surface morphology; moisture content and temperature and humidity data can reflect the sources of OH absorption and scattering changes. If near-infrared spectroscopy can be used as the primary mode and aligned, fused, and jointly modeled with the above auxiliary modes, the accuracy and robustness of rapid resin loading detection can be improved.

[0006] Therefore, it is necessary to provide a rapid resin loading detection method that uses near-infrared spectroscopy as the core, introduces a multimodal fusion mechanism, and simultaneously covers both the modeling and actual detection stages, in order to improve the accuracy, reproducibility, anomaly identification capability, and applicability to production quality control of peptide solid-phase synthesis carrier resin loading detection. Summary of the Invention

[0007] To address the above shortcomings, the present invention provides the following technical solution: A method for detecting resin loading using multimodal fusion near-infrared spectroscopy includes: Obtain near-infrared spectral data of the peptide solid-phase synthesis carrier resin and at least one auxiliary modal data; Spectral features are extracted based on the near-infrared spectral data, and auxiliary modal features are extracted based on the auxiliary modal data; The spectral features and the auxiliary modal features are fused in multiple levels, and the loading of the carrier resin is predicted based on the fused features. The multi-level fusion includes feature-level fusion and / or decision-level fusion. The auxiliary modal data is used to characterize the physical state, environmental conditions, or batch origin information of the carrier resin, so as to compensate for or eliminate the interference of signal changes caused by non-chemical factors in the near-infrared spectral data on the prediction results.

[0008] The near-infrared spectroscopy and auxiliary mode data can be acquired simultaneously under the same sample loading condition, or they can be acquired in association by sample number to achieve data alignment between near-infrared spectroscopy and at least one auxiliary mode.

[0009] Since near-infrared spectroscopy contains both chemical information (such as load-related absorption) and physical interference information (such as scattering changes caused by moisture, particle size, and sample loading status), auxiliary modal data can independently characterize the sources of the above-mentioned physical interference. Therefore, in the multi-level fusion process, auxiliary modal features can compensate for or eliminate the changes in non-chemical factor signals in near-infrared spectral features, thereby improving the accuracy of load prediction.

[0010] Preferably, the auxiliary modal data includes at least one of the following: image data, Raman spectral data, mid-infrared spectral data, moisture content data, particle size distribution data, sample loading status data, ambient temperature and humidity data, resin batch information or supplier information.

[0011] Specifically, it includes at least one of the following: visible light image data, microscopic image data, Raman spectral data, mid-infrared spectral data, ultraviolet-visible absorption data, sample quality data, sample thickness data, compaction degree data, moisture content data, particle size distribution data, ambient temperature data, ambient humidity data, resin supplier information, resin batch number information, functionalization reaction process parameters, or storage time data.

[0012] Preferably, the multi-level fusion of the spectral features and the auxiliary modal features includes: The spectral features and the auxiliary modal features are concatenated into a fused feature vector, and a regression model is established based on the fused feature vector; Alternatively, a first prediction model can be established based on the spectral features and a second prediction model can be established based on the auxiliary modal features, and then the outputs of the first prediction model and the second prediction model can be weighted and fused.

[0013] Preferably, the weighted fusion includes, but is not limited to, weighted averaging, stacked regression, Bayesian fusion, or confidence-weighted fusion. The splicing includes, but is not limited to, weighted splicing, splicing after principal component dimensionality reduction, splicing after partial least squares latent variable fusion, splicing after canonical correlation analysis fusion, splicing after attention-weighted fusion, or splicing after autoencoder fusion.

[0014] Preferably, the method further includes a modeling stage and an actual detection stage.

[0015] In the modeling phase, a multi-modal fusion quantitative model is established using standard samples with known loading standard values, and the fusion strategy and model parameters are fixed. Specifically, the first part of the modeling includes: selecting multiple peptide solid-phase synthesis carrier resin standard samples; adjusting the sample state of the standard samples; obtaining the loading standard values ​​of the standard samples; acquiring near-infrared spectral data of the standard samples; acquiring at least one auxiliary modal data corresponding to the standard samples; preprocessing and quality evaluation of the near-infrared spectral data and the auxiliary modal data respectively; aligning each modal data according to at least one of sample number, acquisition time, sampling location, or sample batch; extracting spectral features from the near-infrared spectral data and extracting auxiliary modal features from the auxiliary modal data; and establishing, screening, and validating a multi-modal fusion loading quantitative model based on the spectral features, the auxiliary modal features, and the loading standard values.

[0016] In the actual detection phase, the fusion strategy and model parameters fixed in the modeling phase are used to predict the actual sample. Specifically, the second part of the actual sample detection includes: adjusting the sample state of the actual resin sample to be tested in accordance with the modeling in the first part; acquiring the near-infrared spectral data and auxiliary modal data of the actual resin sample to be tested; processing the data of the actual resin sample to be tested using the preprocessing parameters, quality evaluation rules, modal alignment rules, feature extraction rules, and fusion strategy determined in the modeling in the first part; and inputting the processed fusion features or the prediction results of each modality into the multimodal fusion loading quantitative model to obtain the predicted loading value of the actual resin sample to be tested.

[0017] In the second part of the actual sample testing, the actual resin sample to be tested adopts the same or equivalently corrected sample state adjustment conditions, near-infrared acquisition conditions, auxiliary modal acquisition conditions, preprocessing parameters, modal alignment rules and fusion strategies as the standard sample in the first part of the modeling.

[0018] Preferably, the spectral features are selected from wavelength variables corresponding to overtone absorption, combination absorption, or combined absorption associated with the resin skeleton, aromatic ring, methylene, amide bond, hydroxyl group, amino group, carboxyl group, chloromethyl group, protecting group, linker arm, or coupled amino acid residue; the auxiliary modal features include at least one of resin color features, brightness features, particle morphology features, particle size statistics features, texture features, moisture content features, temperature and humidity features, sample geometry features, quality features, or batch process features.

[0019] Preferably, after predicting the loading of the carrier resin, the method further includes: outputting prediction confidence, which is determined based on the prediction dispersion of model integration or the consistency between the auxiliary modal data and the data from the modeling stage.

[0020] Specifically, when using an ensemble learning model for prediction, the stability of the prediction is measured by calculating the standard deviation or coefficient of variation of the prediction results of each sub-model. The smaller the prediction dispersion, the higher the prediction reliability. Alternatively, the reliability of the current prediction result can be evaluated by calculating the Mahalanobis distance or similarity index between the auxiliary modal features of the current test sample and the corresponding features in the modeling sample library.

[0021] Preferably, the actual detection stage further includes: comparing the near-infrared spectral data and auxiliary modal data of the actual sample with the applicable range of the model determined in the modeling stage; when the applicable range is exceeded, it is determined that chemical retesting or supplementation into the modeling sample library is required.

[0022] Specifically, when the near-infrared spectral data, auxiliary modal data, or fusion characteristics of the actual resin sample to be tested exceed the applicable range of the multimodal fusion loading quantitative model, or when the model prediction residual, leverage value, Mahalanobis distance, spectral similarity, image similarity, modal consistency index, or prediction uncertainty exceeds a preset threshold, it is determined that the actual resin sample to be tested needs to be chemically verified, re-acquired, or added to the modeling sample library.

[0023] Preferably, the near-infrared spectral data is acquired in the wavelength range of 780 nm to 2500 nm, or in the wavelength range of 10000 nm. Up to 4000 The acquisition method is at least one of diffuse reflection, transmitted reflection, transmission, diffuse reflection from an integrating sphere, diffuse reflection from a fiber optic probe, online detection, or near-line sampling detection.

[0024] Preferably, the loading amount is the content of amino, hydroxyl, chloromethyl, carboxyl, linker arms, protected amino acid residues, or active sites that can be used for polypeptide solid-phase synthesis reactions in the carrier resin, in units of mmol / g.

[0025] Preferably, the polypeptide solid-phase synthesis carrier resin includes at least one of Wang resin, RinkAmide resin, RinkAmideMBHA resin, 2-chlorotriphenylmethyl chloride resin, Merrifield resin, HMPB resin, PAL resin, SieberAmide resin, Tentagel type resin, PEG-PS type resin, polystyrene-divinylbenzene crosslinked resin, or its functionalized derivatives.

[0026] Preferably, the method is used for incoming inspection of peptide solid-phase synthesis carrier resin, supplier batch consistency evaluation, pre-synthesis feeding confirmation, resin storage stability evaluation, resin functionalization reaction endpoint determination, online or near-line quality control of peptide production process, screening of abnormal resin batches, or raw material quality traceability during peptide production scale-up process.

[0027] Preferably, the standard samples in the first part of the modeling cover the target loading detection range and include at least one source of difference from different batches, different suppliers, different degrees of functionalization, different resin matrices, different moisture contents, different particle size ranges, or different sample loading states.

[0028] Preferably, the sample condition adjustment includes at least one of drying, constant humidity equilibration, sieving, mixing, temperature control, leveling, controlling sample thickness, controlling sample mass, controlling compaction degree, or controlling sampling position, so that at least one of the following: sample moisture content, particle size range, sample thickness, sample mass, surface flatness, measurement temperature, and sampling area is within a preset range. The preset range includes: sample moisture content of 0.1% to 10.0%, particle size of 20 mesh to 400 mesh, sample thickness of 1 mm to 30 mm, and at least two near-infrared spectra of the same sample with simultaneous recording of auxiliary modal data. The simultaneous recording is performed under the same sample condition or through associated recording via sample number.

[0029] Preferably, the standard value of the loading is obtained by Fmoc deprotection ultraviolet absorption method, HPLC method, gravimetric method or a combination thereof, and is used as a supervised learning label or calibration reference value for the multimodal fusion loading quantitative model.

[0030] Preferably, the preprocessing of the near-infrared spectral data includes Savitzky-Golay smoothing, moving average smoothing, standard normal variable transformation, multiplicative scattering correction, multivariate scattering correction, normalization, detrending, baseline correction, first derivative, second derivative, wavelength selection, outlier spectral removal, or a combination thereof; the preprocessing of the auxiliary modal data includes image white balance, image color correction, image segmentation, texture enhancement, missing value imputation, outlier processing, dimensional standardization, categorical variable encoding, timestamp correction, or a combination thereof.

[0031] Preferably, the modal alignment includes establishing a one-to-one correspondence of near-infrared spectral data, image data, moisture content data, temperature and humidity data, sample loading data and batch information for the same sample, and removing, re-acquiring, interpolating, or processing missing modal data that have an acquisition time interval exceeding a preset threshold, inconsistent sample numbers, inconsistent sampling locations, or unqualified quality evaluation.

[0032] Preferably, the multimodal fusion load quantification model includes partial least squares regression model, principal component regression model, support vector regression model, ridge regression model, random forest regression model, gradient boosting tree model, extreme gradient boosting model, artificial neural network model, convolutional neural network model, Transformer model, ensemble learning model, or a combination thereof.

[0033] Preferably, establishing and validating the multimodal fusion load quantitative model includes: dividing the standard samples into a calibration set, a validation set, and an external test set; training the model based on the calibration set; selecting preprocessing parameters, feature variables, fusion strategies, and model parameters based on the validation set; evaluating the correlation coefficient, root mean square error, mean absolute error, relative prediction bias, pass / fail accuracy, prediction uncertainty, or model applicability range based on the external test set; and when the evaluation results do not meet the preset threshold, adjusting at least one of the sample state adjustment conditions, near-infrared acquisition conditions, auxiliary modal acquisition conditions, preprocessing methods, feature variables, fusion strategies, or model parameters, and re-establishing or updating the model.

[0034] After obtaining the predicted loading value of the actual resin sample to be tested, the method further includes comparing the predicted loading value with a preset quality standard, outputting a quality control result that is qualified, unqualified, pending review, or recommended to be retested by chemical method, and outputting the prediction confidence, abnormal mode prompts, or main influencing characteristics corresponding to the quality control result.

[0035] The method further includes a model update step: adding the actual resin sample to be tested after chemical verification, its near-infrared spectral data, auxiliary modal data, and loading standard value to the sample library, and retraining, transferring, or updating the multimodal fusion loading quantitative model according to a preset period or model performance triggering conditions.

[0036] The beneficial effects of this invention are as follows: 1. This invention uses near-infrared spectroscopy as the main mode and introduces images, moisture content, temperature and humidity, sample loading information, batch information, or other spectral data as auxiliary modes, which can utilize multi-source information to jointly explain the sources of resin loading changes and near-infrared scattering changes; 2. This invention, through modal quality evaluation and modal alignment rules, can reduce the impact of sample numbering errors, inconsistent acquisition times, inconsistent sampling positions, image defocusing, abnormal sample loading, or sensor malfunctions on model training and actual detection; 3. This invention, through feature-level fusion, decision-level fusion, or hybrid fusion, can improve the robustness of the model to changes in moisture, particle size, color, sample loading status, and environment while maintaining the advantages of rapid and low-destructive near-infrared detection; 4. This invention outputs prediction confidence, abnormal mode prompts, and verification suggestions along with the predicted loading value, which can reduce the risk of quality misjudgment caused by model extrapolation; 5. This invention can obtain the predicted resin loading value within seconds to minutes, reducing the consumption of organic solvents and deprotection reagents required by traditional chemical methods, and is suitable for incoming inspection, batch-to-batch consistency evaluation, pre-synthesis material feeding confirmation, and near-line quality control in the production process. Attached Figure Description

[0037] Figure 1 This is a schematic diagram of the overall process of the rapid resin loading detection method using multimodal fusion of the present invention. Figure 2 This is a schematic diagram of the original spectrum of the sample in this invention; Figure 3 This is a schematic diagram of the anomaly removal method of the present invention. Detailed Implementation

[0038] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0039] Example 1

[0040] Implementation Method 1: Multimodal Modeling of Fmoc-Ala-Rink Resin S101, 96 batches of Fmoc-Ala-Rink resin standard samples with different loading levels were taken. Each sample was vacuum dried at 35 degrees Celsius for 4 hours, and then cooled to room temperature in a desiccator. Approximately 1.00g of resin sample was weighed, placed in a quartz sample cup, gently compacted, and the surface was smoothed.

[0041] S102, collect near-infrared spectral data of each standard sample under the same sample loading condition (e.g., Figure 2 Near-infrared spectroscopy was performed in the range of 800 nm to 1900 nm, with each standard sample being acquired at least three times and the average spectrum calculated. Auxiliary modal data included sample moisture content and particle size.

[0042] S103, the standard loading values ​​of each standard sample were determined using the Fmoc deprotection UV absorption method. Specifically, approximately 10.0 mg of resin sample was weighed, 10 mL of 20% piperidine / DMF solution was added, and the mixture was shaken at room temperature for 20 min. The deprotection solution was then diluted to a suitable volume, and the absorbance was measured at 301 nm. The resin loading was calculated using the absorption or calibration curve of the dibenzo-piperidine adduct.

[0043] S104, Near-infrared spectral preprocessing. The raw near-infrared spectra are sequentially subjected to outlier removal, standard normal transformation, Savitzky-Golay first derivative, and smoothing; the Savitzky-Golay polynomial order is 2, and the window width is 11 data points. Subsequently, the spectral variables are mean-centered, and Hotelling T2 and Q residuals are used for outlier spectrum identification. Spectral records exceeding the 95% confidence limit are either reacquired or removed (e.g., ...). Figure 3 Outlier processing and standardization are performed on numerical auxiliary modal data such as moisture content and particle size.

[0044] S105, Feature Selection and Multimodal Feature Fusion. First, competitive adaptive reweighted sampling (CARS) is used to screen candidate wavelengths from the preprocessed near-infrared spectrum. Then, a secondary screening condition is used, with a partial least squares variable importance projection (VIP) greater than 1.0. The selected spectral variables are then processed using PLS to extract latent variables. The number of latent variables is determined through 10-fold cross-validation, with a candidate range of 2 to 15. The PLS latent variables, particle size statistics, and moisture content information are concatenated to form a fused feature vector. The fused feature vector is then Z-score standardized, and principal component analysis can be used to retain principal components that explain at least 95% of the cumulative variance.

[0045] S106, Integrating Model Training and Validation. In this embodiment, a gradient boosting tree regression model is preferably used to establish a quantitative load model, with parameters of 200 trees, a maximum depth of 3, a learning rate of 0.05, a subsampling ratio of 0.80, and a column sampling ratio of 0.80; simultaneously, a near-infrared single-modal PLS model is established as a control. The 96 batches of standard samples are stratified according to load level, with 70% as the calibration set, 15% as the validation set, and 15% as the external test set. The validation set is used to select the number of features and model parameters, and the external test set is used to evaluate R², RMSEP, MAE, RPD, and pass / fail accuracy. In this embodiment, the external test set R² of the multimodal fusion model is 0.989, RMSEP is 0.026 mmol / g, MAE is 0.018 mmol / g, RPD is 8.7, and the pass / fail accuracy is 98.7%; the external test set R² of the near-infrared single-modal PLS model is 0.965, RMSEP is 0.047 mmol / g, MAE is 0.034 mmol / g, RPD is 4.8, and the pass / fail accuracy is 91.3%. Therefore, the model, preprocessing parameters, feature variables, and fusion strategy are determined to be the fixed methods used in the actual detection stage.

[0046] Example 2

[0047] Implementation Method 2: Rapid Detection of Multimodal Fusion of Actual Samples S201, Sample Preparation. Take 5 batches of actual Fmoc-Ala-Rink resin samples to be tested, numbered T1 to T5. Each batch of samples is dried, sieved, loaded, and leveled according to S101 of Implementation Method 1. If the moisture content of the sample to be tested is higher than the maximum moisture content of the modeled sample library, it is dried and equilibrated before testing.

[0048] S202, Multimodal Acquisition of Actual Samples. Near-infrared spectra, moisture content, and particle size of each batch of samples are acquired according to S102 to S104 of Implementation Method 1. The acquisition software generates a detection record using the sample number as the primary key; if the time interval between any modal acquisition and near-infrared spectrum acquisition exceeds 10 minutes, the software prompts for re-acquiring the data for that modality.

[0049] S203, Preprocessing of actual sample data. The near-infrared spectra of the samples to be tested are processed using the SNV, Savitzky-Golay first derivative, and smoothing parameters fixed during the modeling stage; the moisture content and particle size are standardized using Z-scores based on the calibrated set mean and standard deviation.

[0050] S204, Model Applicability Judgment. The near-infrared spectrum of the preprocessed sample is projected onto the PLS space of the modeling stage, and Hotelling T2, Q residual, and Mahalanobis distance are calculated.

[0051] S205, Load Prediction. For samples that pass the applicable range assessment, the near-infrared PLS latent variables, moisture content, and particle size are concatenated into a fused feature vector according to the fusion order of Implementation Method 1. This vector is then input into the saved gradient boosting tree regression model to obtain the predicted load value. The model also outputs the prediction uncertainty based on the tree model ensemble prediction dispersion, which is converted into prediction confidence. When the prediction uncertainty is less than 1.5 times the RMSEP of the external test set, the confidence is high; otherwise, it is medium or low.

[0052] S206, Quality Control Judgment. The predicted load value is compared with the preset quality standard. In this embodiment, the target quality standard is 0.65 mmol / g to 0.75 mmol / g; if the predicted value is within this range and the confidence level is high or medium, it is deemed qualified; if the predicted value exceeds this range and the applicability judgment is passed, it is deemed unqualified; if the confidence level is low, the modal quality is unqualified, or the applicability judgment is not passed, it is deemed to need to be reviewed, and it is recommended to retest using the Fmoc deprotection UV absorption method.

[0053]

[0054] The method provided by this invention can be used in peptide research and development, drug production, resin supplier quality control, raw material incoming inspection, and peptide production process quality control. While retaining the advantages of near-infrared detection—rapid, low-damage, low-reagent, and suitable for batch testing—this method improves the model's adaptability to sample moisture, particle size, color, sample loading status, environmental temperature and humidity, and batch variations through multimodal fusion, thus exhibiting good industrial applicability.

[0055] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0056] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A method for detecting resin loading using multimodal fusion near-infrared spectroscopy, characterized in that, include: Obtain near-infrared spectral data of the peptide solid-phase synthesis carrier resin and at least one auxiliary modal data; Spectral features are extracted based on the near-infrared spectral data, and auxiliary modal features are extracted based on the auxiliary modal data; The spectral features and the auxiliary modal features are fused in multiple levels, and the loading of the carrier resin is predicted based on the fused features. The multi-level fusion includes feature-level fusion and / or decision-level fusion. The auxiliary modal data is used to characterize the physical state, environmental conditions, or batch origin information of the carrier resin, so as to compensate for or eliminate the interference of signal changes caused by non-chemical factors in the near-infrared spectral data on the prediction results.

2. The method according to claim 1, characterized in that, The auxiliary modal data includes at least one of the following: Image data, Raman spectral data, mid-infrared spectral data, moisture content data, particle size distribution data, sample loading status data, ambient temperature and humidity data, resin batch information or supplier information.

3. The method according to claim 1, characterized in that, Multi-level fusion of the spectral features and the auxiliary modal features includes: The spectral features and the auxiliary modal features are concatenated into a fused feature vector, and a regression model is established based on the fused feature vector; Alternatively, a first prediction model can be established based on the spectral features and a second prediction model can be established based on the auxiliary modal features, and then the outputs of the first prediction model and the second prediction model can be weighted and fused.

4. The method according to claim 1, characterized in that, It also includes a modeling stage and an actual detection stage; in the modeling stage, the multi-level fusion quantitative model is established using standard samples with multiple known loading standard values, and the fusion strategy and model parameters are fixed; in the actual detection stage, the fusion strategy and model parameters fixed in the modeling stage are used to predict the actual samples.

5. The method according to claim 1, characterized in that, The spectral features are selected from wavelength variables corresponding to overtone or combination absorptions associated with the resin skeleton, aromatic ring, methylene, amide bond, hydroxyl, amino, carboxyl group or linker arm; the auxiliary modal features include at least one of resin color features, particle morphology features, particle size statistics features, moisture content features, temperature and humidity features or sample loading geometry features.

6. The method according to claim 1, characterized in that, After predicting the loading of the carrier resin, the method further includes: outputting prediction confidence, which is determined based on the prediction dispersion of the model integration or the consistency between the auxiliary modal data and the data from the modeling stage.

7. The method according to claim 4, characterized in that, The actual testing phase also includes: The near-infrared spectral data and auxiliary modal data of the actual sample are compared with the applicable range of the model determined in the modeling stage; when the applicable range is exceeded, it is determined that chemical retesting or supplementation into the modeling sample library is required.

8. The method according to claim 1, characterized in that, The near-infrared spectral data was acquired in the wavelength range of 780 nm to 2500 nm, or 10000 nm. Up to 4000 The loading amount refers to the content of amino, hydroxyl, chloromethyl, carboxyl, or active linker groups in the carrier resin, expressed in mmol / g.

9. The method according to claim 1, characterized in that, The solid-phase synthetic support resin for the polypeptide is selected from at least one of Wang resin, RinkAmide resin, 2-chlorotriphenylmethyl chloride resin, Merrifield resin, HMPB resin, Tentagel type resin, or polystyrene-divinylbenzene crosslinking resin.

10. The method according to claim 1, characterized in that, The method is used for incoming inspection of peptide solid-phase synthesis carrier resins, supplier batch consistency evaluation, pre-synthesis material feeding confirmation, or production process quality control.

Images