A Machine Learning-Based Method for Quantitative Analysis of β-Sheet Structure in Silk Fibroin Materials
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- FAVORSUN MEDICAL TECH (SUZHOU) CO LTD
- Filing Date
- 2024-09-14
- Publication Date
- 2026-05-26
Smart Images
Figure CN119741964B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of materials characterization and biomedical materials technology, and in particular to a quantitative analysis method for the microstructure of silk fibroin materials based on machine learning. Background Technology
[0002] The molecular assembly structure of proteins, both in vivo and in vitro, significantly influences their performance and function. In the biomedical field, structural proteins such as collagen and silk fibroin have been developed into biomedical materials, exhibiting excellent biocompatibility. However, the complex molecular interactions inherent in protein topology, including hydrogen bonds, van der Waals forces, and hydrophobic interactions, coupled with the sensitivity of protein structures to factors such as heat and force, mean that the precise and quantitative analysis of their microstructure remains a bottleneck restricting their application.
[0003] Silk fibroin is a protein extracted from natural silkworm silk, exhibiting typical β-sheet, helical, and random coil conformations. Materials made from silk fibroin, such as films, solutions, gels, and scaffolds, have become an important category in biomedical materials. The amino acid sequence of silk fibroin is mostly composed of GAGAGS repeating modules, with up to 432 repetitions. When two GAGAGS segments are aligned in parallel and locked by hydrogen bonds, they form an antiparallel β-sheet structure. This structure can be considered the smallest repeating unit of a β-sheet structure (see appendix). Figure 1 In silk fibroin materials, such as in solutions / gels, the size of their β-sheet structures depends on the preparation conditions and evolves over time. Different silk fibroin materials contain β-sheet structures of different sizes, which can be considered as sheets or cuboid structures formed by the stacking of minimal repeating units. (See attached...) Figure 2 In this diagram, the long side of the cuboid represents the 'a' direction, where hydrogen bonds form between molecular chains; the wide side represents the 'b' direction, where the layers stack; and the height represents the 'c' direction, where the chains extend. Currently, information on inter-chain and inter-layer spacing has been obtained using wide-angle X-ray diffraction (WAXD), but quantitative analysis methods are still lacking for β-sheet structures of sheets and cuboids with dimensions ranging from 10 nm to 500 nm.
[0004] Small-angle X-ray diffraction (SAXS) involves exposing a sample to high-energy, hard X-rays with very small wavelengths (0.07–0.2 nm) and photon energies above ~10 keV. This high energy allows for sufficiently strong scattered light signals even when the sample is only 3 to 4 meters away from the X-ray source, at a very small scattering angle. Therefore, SAXS has become an important tool for resolving protein structures. This method is applicable to solution, gas, and solid samples and is a non-invasive, non-destructive characterization technique. Each sample produces a unique scattering signal determined by its internal structure, allowing for precise and efficient inference of the sample's microstructure from the scattering results. Small-angle scattering includes SAXS and small-angle neutron scattering (SANS), and is commonly used to study the shape, structure, and interactions of proteins and polymers in solution.
[0005] Machine learning is a data-driven approach with significant advantages in solving multivariate nonlinear problems. Supervised learning in machine learning encompasses two main algorithms: regression and classification, used for predicting continuous and discrete results respectively, effectively assisting research in natural sciences. These mainstream algorithms include Decision Tree, Random Forest, Support Vector Machine (SVM), and Naive Bayes. Machine learning, as a key technology in the field of artificial intelligence, has had a significant impact on areas such as computer graphics, natural language processing, and content generation. In materials science, the application of machine learning models based on big data is also becoming increasingly widespread. By constructing machine learning models to predict the size of characteristic β-sheet structures in silk fibroin, experimental costs can be effectively reduced, and the materials development and optimization process can be accelerated. Examples include Bayesian learning, ensemble learning, and multilayer perceptrons (MLP). Summary of the Invention
[0006] This invention aims to protect a method for quantitative analysis of the β-sheet structure in silk fibroin materials. This method uses SAXS as the experimental technique, combined with machine learning methods and models, to obtain the shape and size information of the β-sheet structure in silk fibroin solutions, gels, and other materials. This quantitative method has important reference value for constructing structure-activity relationships of silk fibroin materials, optimizing the preparation process of silk fibroin materials, and expanding their application scenarios.
[0007] To achieve the above-mentioned objectives, the present invention provides the following technical solution:
[0008] The specific steps are as follows (see appendix) Figure 3 ):
[0009] (1) Data Collection and Processing. This invention uses SAXS experiments to measure the scattering curve data (i.e., the data on the variation of scattering intensity I with scattering vector q) of a large number of silk fibroin solutions, and performs preliminary processing on the scattering data. Then, the β-sheet structure size of the measured samples is characterized using an independent method. Assuming the β-sheet structure is a spherical particle, the scattering signal satisfies Guinier's law in the small scattering angle region. Therefore, this invention uses Guinier's law to calculate the β-sheet structure size of the samples, and uses this calculated value as the baseline size of the β-sheet structure. Based on this method, the dataset needed to train the machine learning model is obtained.
[0010] (2) Feature Extraction and Size Calibration. This invention extracts characteristic indices from the SAXS scattering curve using various numerical analysis methods. These features are all related to the physical and chemical processes of protein β-sheet in silk fibroin solution, and also reflect the morphological characteristics of the β-sheet structure to a certain extent. Related features include, but are not limited to:
[0011] a) Scattering curve characteristics: such as the slope of the scattering curve, the maximum / minimum value of the scattering intensity, etc.
[0012] b) Radial distribution function: The characteristic peak value is obtained by Fourier transform.
[0013] c) Structure factor S(q): The structure factor reflects the short-to-medium range order of atoms or molecules in the system and can provide information about the correlation between atoms.
[0014] All features involved in the modeling process are standardized using various methods to ensure feature quality and consistency.
[0015] (3) AI Model Training. Select a suitable traditional machine learning model for training. Possible models include:
[0016] a) Random Forest: Random forest is an ensemble learning algorithm composed of multiple decision trees. It can construct different decision trees through multiple random samplings and take the average of them for prediction. Random forest has the advantages of being resistant to overfitting, handling high-dimensional data, and handling nonlinear relationships, and is suitable for the problem of predicting the size of β-fold structures.
[0017] b) Support Vector Machine (SVM): Support Vector Machine is a machine learning model used for classification and regression. SVM achieves efficient prediction by finding the optimal hyperplane, and is particularly effective in capturing the complex relationship between the size and structure of β-fold structures when dealing with nonlinear problems using kernel functions (such as the RBF kernel).
[0018] During model training, firstly, all collected data is split into training and test sets according to a certain ratio (e.g., 8:2). Then, the model is fitted using the training set data, and preliminary evaluation and hyperparameter tuning are performed using the test set. During model training, cross-validation or Bayesian optimization methods are used to search for the optimal combination of model hyperparameters. Model hyperparameters include: the number and depth of decision trees in the random forest, and the kernel function selection and penalty parameter C in the SVM.
[0019] The trained machine learning model is evaluated using mean squared error (MSE) and goodness-of-fit (R²).
[0020]
[0021]
[0022] Where n represents the total number of protein sample datasets, and y i These are the model prediction and theoretical calculation values for the i-th data point, respectively. All predicted values The average value.
[0023] (4) AI Model Evaluation and Optimization. The machine learning model obtained through training can be used to predict the β-sheet structure size in silk fibroin solutions. To evaluate the model's performance, a SAXS experiment needs to be performed on new solution samples. The scattering curve data of the new samples are processed in the same way as the training dataset, and the extracted features and the trained model are used to predict the β-sheet structure size of the new samples. The model prediction results are compared with the results calibrated by Guinier's law. If the results reach a certain level of accuracy, it indicates that the AI model has reached the level of industrial application. The next step is to deploy the trained model in the production environment to effectively assist in the structural design and synthesis experiments of silk fibroin, reducing the cost and time of experimental measurements. If the accuracy of the results does not reach a usable level, it is necessary to continuously improve the model by modifying the model training parameters, increasing the amount of data, and adjusting feature extraction.
[0024] This invention proposes a method for quantitative analysis of the shape and size information of β-sheet structures in silk fibroin materials based on machine learning. The method provides the shape and size information of β-sheet structures based on SAXS experimental data and machine learning models.
[0025] The silk fibroin materials mentioned include, but are not limited to, macroscopic forms such as silk fibroin solutions, gels, membranes, powders, sponges, blocks, and fibers. A suitable SAXS sample stage / injection system is selected based on the material morphology.
[0026] In the SAXS experiment, the scattering angle ranged from 0.1 to 5°, the scattering vector was q = 4πsinθ / λ, and the β-sheet structure in the silk fibroin material was used as the scatterer.
[0027] The machine learning model described above takes the SAXS experimental results, i.e., the Iq data curve, as input and predicts the shape and size of the β-sheet structure in silk fibroin materials by extracting the mathematical features of the curve.
[0028] The machine learning model described above is based on a training set containing more than 100 sets of data, and selects the machine learning model with the highest prediction accuracy from the AI models.
[0029] The dataset of >100 sets includes Iq curve data obtained from SAXS experiments on >100 samples, the gyration radius of the scatterer calculated based on Guinier's law, and the power law of q and I in a specific interval. This is specifically embodied in the embodiments of the present invention.
[0030] The machine learning model uses feature data extracted from the scattering curves characterized by the SAXS experiment for model training, and the radius of gyration fitted according to Guinier's law as the target for model prediction.
[0031] The machine learning model described uses k-fold cross-validation to improve its generalization ability and Bayesian optimization to search for the optimal model hyperparameters.
[0032] The machine learning model described uses mean squared error and R0. 2 Values are used to evaluate the performance of machine learning models.
[0033] The invention has a wide range of applications, including obtaining the mechanical structure-property relationship of materials or the relationship between the biological properties and structure of materials or assisting in process optimization.
[0034] For example, after quantitative analysis of the β-sheet structure of silk fibroin materials prepared under specific process conditions, their mechanical properties, such as modulus, can be further determined to obtain the mechanical-structure-property relationship of the material. After quantitative analysis of the β-sheet structure of silk fibroin materials prepared under specific process conditions, their biological properties, such as in vivo degradation / absorption rate based on a rat subcutaneous animal model, can be further determined to obtain the relationship between the biological properties and structure of the material. In the process of new material research and development, the relationship between process conditions, such as solution concentration and temperature, and β-sheet structure parameters can be established to assist in the optimization of the process.
[0035] Beneficial effects of this invention:
[0036] (1) This invention proposes a measurement method at the submicron level that can quantify the size of β-sheet structures in silk fibroin solutions / gels. This invention uses SAXS as the detection method, and the stacking of β-sheets and their corresponding sizes within the material can be obtained upon receiving the SAXS signal.
[0037] (2) After accumulating a large amount of SAXS detection data of various silk fibroin solutions / gels, the present invention obtains the β-sheet size of the corresponding silk fibroin solution / gel according to the corresponding law. A deep learning model can be established based on the accumulated data, and the β-sheet size of the silk fibroin solution / gel to be tested can be calculated based on the model.
[0038] (3) After the β-sheet structure size of the silk fibroin solution / gel system is obtained by simulation calculation based on the machine learning model, the physicochemical properties of the relevant materials can be inferred based on this information, which provides a powerful aid for the efficient design of silk fibroin materials.
[0039] (4) Compared with traditional experimental methods, the machine learning algorithm of this invention can predict the β-sheet structure size of a large number of protein sequences in a few minutes, which greatly shortens the research cycle and reduces the overall cost of research investment. Attached Figure Description
[0040] Figure 1 The repeating sequence GAGAGS in silk fibroin is the smallest repeating unit of the β-sheet structure.
[0041] Figure 2 This is a schematic diagram of the sheet-like, cuboid β-sheet structure formed by the stacking of minimal repeating units in silk fibroin.
[0042] Figure 3 Flowchart for applying a machine learning-assisted model for predicting the size of β-folded structures;
[0043] Figure 4 Iq signal graph of a simple mixed solvent system in SAXS test;
[0044] Figure 5 Iq signal graph of a simple mixed solvent system in SAXS test;
[0045] Figure 6 Iq signal diagrams for two types of brackets under SAXS testing;
[0046] Figure 7 Image of silk fibroin sample;
[0047] Figure 8 Iq signal diagram of SAXS test for silk fibroin aqueous solution. Detailed implementation method:
[0048] The present invention will be further described below with reference to specific embodiments, but the scope of protection of the present invention is not limited thereto:
[0049] Example 1:
[0050] After dissolving silk fibroin in hexafluoroisopropanol (HFIP) to form a completely transparent homogeneous solution, water was added to induce a conformational change in the silk fibroin chains. The mass ratio of the three phases in the solution was SF:HFIP:H2O = 0.4%:61.3%:38.3%. After 12 hours of physical gelation, tiny β-sheet structures were generated in the system. Using the Xenocs Xeuss 3.0 platform, the sample was placed in the liquid window, and two detectors were set at distances of 1798 mm and 300 mm to cover a wide range of q values. SAXS tests were performed on the heterogeneous solution system containing β-sheet structures and the simple mixed solvent system to obtain the Iq signal (see attached). Figure 4 Assuming the β-fold structure consists of spherical particles, in the low q-value (small scattering angle) region, the SAXS curve in the Guinier region (qR) g <1.3) Follows Guinier's Law:
[0051]
[0052] By analyzing logI to q 2 The radius of gyration R of the particle is obtained by using linear fitting to calculate the slope. g ~0.5nm, which translates to a particle diameter D of ~1.3nm, consistent with the β-sheet structure size of the smallest repeating structure predicted by the amino acid sequence.
[0053] This embodiment demonstrates through testing that, theoretically, the dimensions of the β-sheet structure can be obtained from the structural information of the chemical bonds, while simultaneously following Guinier's law.
[0054] Example 2:
[0055] To explore the effect of concentration, different concentrations of silk fibroin HFIP solutions were prepared (as shown in Table 1), resulting in samples 2-1, 2-2, and 2-3. Using the Xenocs Xeuss 3.0 platform, samples were placed in the liquid window, and two detectors were set at distances of 1798 mm and 300 mm to cover a wide range of q values. SAXS tests were performed on heterogeneous solution systems of different concentrations and simple mixed solvent systems to obtain the Iq signals (see attached table). Figure 5 Assuming the β-sheet structure is a spherical particle, the size of the β-sheet structure in each sample is calculated using Guinier's law in the small scattering angle region of the SAXS curve. A suitable low q-value space qR is selected. g<1.3, and ensure that the data does not contain areas with strong noise or background scattering; then, process the Iq curve to calculate lnI(q) and q 2 The relationship between lnI(q) and q; third, using the least squares method to compare lnI(q) and q 2 Perform linear fitting, calculate the slope m and intercept c of the fitted line; finally, according to the formula... Calculate the radius of gyration R from the slope obtained from the fitting. g .
[0056] Table 1. β-sheet size corresponding to different concentrations of silk fibroin solutions
[0057] Serial Number name concentration SAXS theoretical calculation of β-fold size 2-1 Silk fibroin solution 0.4％ 2.1nm 2-2 Silk fibroin solution 1.9％ 3.4nm 2-3 Silk fibroin solution 3.6％ 4.8nm
[0058] This embodiment demonstrates that by characterizing silk fibroin solutions of different concentrations using SAXS and calculating β-sheet size data using the Guinier formula, the β-sheet size corresponding to different concentrations of silk fibroin solutions is different.
[0059] Example 3:
[0060] Silk fibroin scaffolds with physicochemical double cross-linking structures were prepared by cryogelation using small molecule cross-linking agents and compared with pure silk fibroin scaffold samples. SAXS measurements were performed on both types of scaffolds at the 1W2A beamline of the Beijing Synchrotron Radiation Facility to obtain the Iq signal (see attached image). Figure 6 The dimensions of the β-fold structure were calculated using SAXS theory, i.e., Guinier's law, as shown in Table 2.
[0061] Table 2. β-sheet dimensions of silk fibroin scaffolds prepared by different cross-linking methods
[0062]
[0063]
[0064] This embodiment demonstrates that, through SAXS characterization and calculation using the Guinier formula, the β-sheet size of silk fibroin scaffolds prepared by different crosslinking methods is different.
[0065] Example 4: In this example, a machine learning method and model are established. Different silk fibroin materials are characterized by SAXS. The scattering curves of SAXS are input into the trained machine learning model to obtain the size information of the β-sheet structure.
[0066] (1) Prepare 120 silk fibroin samples, including solutions, gels, membranes, powders, sponges, blocks, and fibers (such as...). Figure 7 ).
[0067] (2) Scattering curve data (including scattering intensity I and scattering vector q) of all silk fibroin samples were measured by SAXS experiment. These Iq curves were used to characterize the internal structure of the samples and provide microscopic information about particle size, shape, distribution and other aspects.
[0068] (3) The obtained Iq curve data needs to be preprocessed before it can be input into the machine learning model. Specific preprocessing steps include:
[0069] - Normalization processing: Normalize all Iq curve data to ensure that the I values of different samples are within the same order of magnitude, preventing model training instability caused by excessively large numerical ranges.
[0070] -Since the number and interval of Iq curve data points may differ for different samples, it is necessary to standardize the q values by using interpolation methods to make the q values of each sample uniform within the same range and step size.
[0071] (4) According to Guinier's law, select Iq curve data in the low q region and obtain the radius of gyration R by fitting the Guinier formula. g This parameter is an important physical quantity that characterizes the geometry of a β-folded scatterer.
[0072] (5) Based on the preprocessed Iq curve and radius of gyration R g Construct features for machine learning modeling, where the features for each sample include:
[0073] -Iq curve characteristics: I values corresponding to a set of standardized q values.
[0074] - Scattering curve characteristics: maximum and minimum scattering intensity.
[0075] -Radial distribution function: The positions of all characteristic peaks are obtained by performing a Fourier transform on the Iq curve.
[0076] -Structure factor: The structure factor S(q) of the sample is separated from the experimental data. This feature can provide information on the correlation between atoms.
[0077] (6) Using the obtained silk fibroin sample database, the dataset was split into 80% training set and 20% test set. Machine learning models were built using the following five methods: Random Forest, Support Vector Machine, Naive Bayes, Multilayer Perceptron, and XGBoost. During model training, cross-validation was used to update the training and test sets to improve the model's generalization ability. The model hyperparameters were adjusted to ensure optimal performance.
[0078] (7) Calculate the performance metrics for each model on the test set, including:
[0079] - Mean Squared Error (MSE) is used to measure the difference between the model's predicted values and the values calculated by the Guinier formula.
[0080] -R 2 R value: Reflects the goodness of fit of the model to the data. 2 The closer the value is to 1, the better the model fits.
[0081] (8) Among the five machine learning models, the model with the better performance metrics is selected as the model used in the prediction stage. In this embodiment, the random forest model exhibits the highest prediction accuracy, with the lowest MSE and R... 2 Since the value is close to 1, the random forest model was chosen as the final model.
[0082] Table 3 shows the β-sheet size range of silk fibroin materials in solution, gel, membrane, powder, sponge, block, and fiber forms, statistically analyzed using a machine learning model.
[0083] Serial Number name β-fold size range 1 solution 1-8nm 2 gel 3-10nm 3 membrane 30-130nm 4 powder 1-2nm 5 foam support 10-14nm 6 block material 50-130nm 7 fiber 100-300nm
[0084] This embodiment demonstrates that machine learning models can effectively measure the β-sheet size of silk fibroin materials in solution, gel, membrane, powder, sponge, block, and fiber forms.
[0085] Example 5:
[0086] This embodiment verifies the machine learning method and model. A 2.4% silk fibroin aqueous solution was prepared, the Iq curve of the SAXS was obtained, and theoretical calculations were performed based on Guinier's law to obtain R. g =4.8nm (with) Figure 8 By inputting the Iq curve into the trained machine learning model, the size of the β-fold structure was found to be 4.9 nm, with a prediction accuracy of 98%.
Claims
1. A method for quantitatively analyzing the size information of β-sheet structures in silk fibroin materials based on machine learning, characterized in that: The method provides size information of β-folded structures based on SAXS experimental data and machine learning models, where SAXS represents small-angle X-ray scattering. The machine learning model described above takes the SAXS experimental results, i.e., the Iq data curve, as input and predicts the size of the β-sheet structure in silk fibroin material by extracting the mathematical features of the curve. Feature data for model training was extracted from the scattering curves characterized by the SAXS experiment, and the radius of gyration fitted according to the Guinier law was used as the target for model prediction. Specifically, this includes measuring the scattering curve data of a large number of silk fibroin solutions using SAXS experiments, i.e., the data on the change of scattering intensity I as a function of scattering vector q, and performing preliminary processing on the scattering data; then characterizing the β-sheet structure size of the measured samples based on an independent method. Feature extraction and size calibration: Feature indices of the SAXS scattering curve were extracted using various numerical analysis methods. Relevant features include: a) Scattering curve characteristics: including the slope of the scattering curve and the maximum / minimum value of the scattering intensity; b) Radial distribution function: The characteristic peak value is obtained by Fourier transform; c) Structure factor S(q): The structure factor reflects the short-to-medium range order of atoms or molecules in the system and can provide information about the correlation between atoms; All features involved in the modeling process are standardized to ensure feature quality and consistency.
2. The method according to claim 1, wherein the silk fibroin material includes macroscopic forms such as silk fibroin solution, gel, membrane, powder, sponge, block, and fiber, and a suitable SAXS sample stage / injection system is selected according to the material form.
3. The method according to claim 1, wherein the scattering angle θ in the SAXS experiment ranges from... scattering vector is λ is the X-ray wavelength, and the β-sheet structure in the silk fibroin material acts as a scatterer.
4. The method according to claim 3, wherein the machine learning model is selected from 5 machine learning models based on a training set containing more than 100 sets of data, and the machine learning model with the highest prediction accuracy is selected. The dataset of more than 100 sets includes Iq curve data obtained from SAXS experiments on more than 100 samples, the gyration radius of the scatterer calculated based on Guinier's law, and the power law of q and I in a specific interval.
5. The method according to claim 1, wherein the machine learning model uses k-fold cross-validation to improve the model's generalization ability and uses Bayesian optimization to search for the optimal model hyperparameters.
6. The method of claim 1, wherein the machine learning model uses mean squared error and R0. 2 Values are used to evaluate the performance of machine learning models.
7. The method according to claim 1, comprising the following steps: (1) AI model training, including selecting a suitable traditional machine learning model for training. Optional models include: a) Random Forest: Random forest is an ensemble learning algorithm composed of multiple decision trees. It can construct different decision trees through multiple random samplings and take the average of them for prediction. b) Support Vector Machine: Support Vector Machine is a machine learning model used for classification and regression. Support Vector Machine achieves efficient prediction by finding the optimal hyperplane. When using kernel functions to handle nonlinear problems, Support Vector Machine can capture the complex relationship between the size and structure of β-fold structures. During model training, firstly, all collected data is divided into training and test sets according to a certain ratio; then, the model is fitted using the training set data, and preliminary evaluation and parameter tuning are performed using the test set. During model training, cross-validation or Bayesian optimization methods are used to search for the optimal combination of model hyperparameters. Model hyperparameters include: the number and depth of decision trees in the random forest, and the kernel function selection and penalty parameter C in the SVM. ； ； The model is evaluated using the mean squared error (MSE) and the goodness of fit (R²). 2 Evaluate the trained machine learning model: in, n This represents the total number of protein sample datasets. and They are the first i The model predictions and theoretical calculations for each data point. All predicted values The average value; (2) AI model evaluation and optimization; The machine learning model obtained through training can be used to predict the β-sheet structure size in silk fibroin solution; To evaluate the model effect, it is necessary to perform SAXS experiments on new solution samples, process the scattering curve data of new samples in the same way as the training dataset, and use the extracted features and the trained model to predict the β-sheet structure size of new samples; Compare the model prediction results with the results calibrated by Guinier's law. If the results reach a certain level of accuracy, it indicates that the AI model has achieved the level of industrial application. The next step is to deploy the trained model in the production environment to effectively assist in the structural design and synthesis experiments of silk fibroin and reduce the cost and time of experimental measurement; If the accuracy of the results does not reach the usable level, it is necessary to continuously improve the model by modifying the model training parameters, increasing the amount of data, and adjusting the feature extraction method.
8. The method according to any one of claims 1 to 7, wherein the use includes: To obtain the mechanical structure-property relationship of materials, the relationship between the biological properties and structure of materials, or to assist in the optimization of processes.