Artificial intelligence prediction method and system for self-assembly behavior of traditional chinese medicine monomer components

Abstract

The present invention provides an artificial intelligence prediction method and system for the self-assembly behavior of traditional Chinese medicine monomer components. The method includes: acquiring multidimensional feature data of a target traditional Chinese medicine monomer; inputting the multidimensional feature data into a trained prediction model; obtaining a molecular structure embedding vector, an environmental feature vector, and an image feature vector of the target traditional Chinese medicine monomer based on the multidimensional feature data; fusing the molecular structure embedding vector, the environmental feature vector, and the image feature vector to obtain a fused feature vector; performing multi-task prediction based on the fused feature vector; and outputting the self-assembly behavior prediction result. This method employs a three-modal feature fusion strategy of molecular structure features, environmental parameter features, and characterizing image features to achieve simultaneous multi-task prediction of traditional Chinese medicine monomers. It overcomes the limitations of traditional experimental trial-and-error methods, significantly shortens the R&D cycle of self-assembled traditional Chinese medicine formulations, reduces R&D costs, and realizes the efficient development of self-assembled drugs from traditional Chinese medicine monomers.

Description

Technical Field

[0001] This invention relates to the interdisciplinary fields of artificial intelligence, modernization of traditional Chinese medicine, and pharmaceutical formulation, and in particular to an artificial intelligence prediction method and system for the self-assembly behavior of single components of traditional Chinese medicine. Background Technology

[0002] Monomers of traditional Chinese medicine (TCM) are the core material basis for the efficacy of TCM, possessing advantages such as diverse structures, significant biological activity, and low toxicity. Self-assembly technology, as a novel formulation technology, enables TCM monomers to spontaneously form ordered nanoscale aggregates (such as micelles, nanoparticles, and nanofibers) under specific conditions, significantly improving the water solubility, bioavailability, and targeting of TCM monomers. This has become an important direction for the modern formulation research and development of TCM.

[0003] However, the self-assembly behavior of traditional Chinese medicine monomers is complexly influenced by multiple factors, including molecular structure (molecular weight, functional group type, spatial configuration) and environmental conditions (concentration, pH value, temperature, solvent type), resulting in poor controllability of the self-assembly process and large fluctuations in product characteristics. Traditional methods rely on experimental trial and error, searching for the optimal solution through a large number of screening conditions, which leads to long development cycles and high experimental costs for self-assembled traditional Chinese medicine formulations.

[0004] Therefore, existing technologies have shortcomings and need to be improved and developed. Summary of the Invention

[0005] This application provides an artificial intelligence prediction method and system for the self-assembly behavior of traditional Chinese medicine monomer components, in order to solve the technical problems of long development cycle and high experimental cost of traditional Chinese medicine self-assembly formulations in related technologies.

[0006] To achieve the above objectives, this application adopts the following technical solution: An artificial intelligence prediction method for the self-assembly behavior of monomeric components in traditional Chinese medicine, the method comprising: The multidimensional feature data of the target Chinese herbal medicine monomer is obtained, and the multidimensional feature data is input into the trained prediction model. Based on the multidimensional feature data, the molecular structure embedding vector, environmental feature vector and image feature vector of the target Chinese herbal medicine monomer are obtained. The molecular structure embedding vector, environmental feature vector, and image feature vector are fused to obtain a fused feature vector; Multi-task prediction is performed based on the fused feature vector, and the self-assembly behavior prediction result is output.

[0007] In one embodiment of this application, obtaining the molecular structure embedding vector, environmental feature vector, and image feature vector of the target traditional Chinese medicine monomer based on the multidimensional feature data includes: Obtain the expression of the Chinese herbal monomer in the multidimensional feature data, convert the expression of the Chinese herbal monomer into a molecular graph structure, construct graph data with atoms as nodes and chemical bonds as edges, learn the topological relationship between atoms through the graph convolutional layer of the graph neural network, and output the molecular structure embedding vector of a predetermined dimension through the global average pooling layer. Obtain environmental parameters of various types from the multidimensional feature data, and encode each type of environmental parameter based on the pre-constructed correspondence between types and encoding methods to obtain an environmental feature vector of a predetermined dimension; The transmission electron microscope (TEM) image in the multidimensional feature data is acquired, the TEM image is preprocessed, deep texture and morphological features are extracted from the preprocessed TEM image, and a predetermined dimension image feature vector is output by global average pooling.

[0008] In one embodiment of this application, the molecular structure embedding vector, environmental feature vector, and image feature vector are fused to obtain a fused feature vector, including: The first attention weight corresponding to the molecular structure embedding vector, the second attention weight corresponding to the environmental feature vector, and the third attention weight corresponding to the image feature vector are calculated using a pre-constructed modal attention weight calculation function. Based on the first attention weight, the second attention weight, and the third attention weight, the molecular structure embedding vector, the environmental feature vector, and the image feature vector are weighted and fused to obtain a fusion vector of a predetermined dimension. The fusion vector is concatenated with the molecular structure embedding vector, environmental feature vector, and image feature vector to obtain the fusion feature vector.

[0009] In one embodiment of this application, multi-task prediction is performed based on the fused feature vector to output a self-assembly behavior prediction result, including: The fused feature vector is positionally encoded, and the positionally encoded fused feature vector is used for multi-task prediction to output the self-assembly behavior prediction result. The self-assembly behavior prediction results include: self-assembly occurrence probability, product particle size, zeta potential, morphology type, and in vitro stability period.

[0010] In one embodiment of this application, the training step of the prediction model includes: Construct a multi-dimensional database, which includes several Chinese herbal medicine monomers and their corresponding multi-dimensional feature data; The multidimensional feature data in the multidimensional database is preprocessed, and the preprocessed multidimensional feature data is divided into molecular structure features, environmental parameter features and characterization image features. Feature encoding and extraction were performed on molecular structure features, environmental parameter features, and characterization image features respectively to obtain molecular structure embedding vectors, environmental feature vectors, and image feature vectors corresponding to Chinese herbal monomers in a multi-dimensional database. The molecular structure embedding vector, environmental feature vector, and image feature vector corresponding to the Chinese herbal monomers in the multi-dimensional database are fused to obtain the fused feature vector. The fused feature vector is then positionally encoded to generate the input sequence. The input sequence is divided into a training set, a validation set, and a test set. The constructed prediction model is pre-trained and fine-tuned using a transfer learning strategy. The hyperparameters are optimized using the validation set to obtain the trained prediction model. The trained prediction model is then evaluated using the test set.

[0011] In one embodiment of this application, the prediction model includes: a feature encoding layer, a modality fusion layer, a multi-task prediction layer, and a loss function optimization layer; the feature encoding layer includes: a convolutional neural network sublayer, a graph neural network sublayer, and a fully connected encoding layer; Feature encoding and extraction were performed on molecular structure features, environmental parameter features, and characterization image features, respectively, to obtain molecular structure embedding vectors, environmental feature vectors, and image feature vectors corresponding to traditional Chinese medicine monomers in a multi-dimensional database, including: The molecular structure features are encoded using the sublayers of the convolutional neural network to obtain the molecular structure embedding vector; The image feature vector is obtained by encoding the image features using the graph neural network sublayer. The fully connected coding layer is used to perform deep feature extraction on environmental parameter features to obtain an environmental feature vector.

[0012] In one embodiment of this application, the molecular structure embedding vector, environmental feature vector, and image feature vector corresponding to traditional Chinese medicine monomers in a multi-dimensional database are fused to obtain a fused feature vector. The fused feature vector is then positionally encoded to generate an input sequence, including: The attention weights corresponding to the molecular structure embedding vector, environmental feature vector, and image feature vector of traditional Chinese medicine monomers in the multi-dimensional database are calculated using a pre-constructed modal attention weight calculation function. Based on the attention weights corresponding to molecular structure embedding vector, environmental parameter features, and image feature features, the molecular structure embedding vector, environmental feature vector, and image feature vector are weighted and fused. The weighted and fused feature vector is concatenated with the corresponding molecular structure embedding vector, environmental feature vector, and image feature vector to obtain the fused feature vector. Position encoding is performed on the fused feature vectors to generate the input sequence.

[0013] This application also provides an artificial intelligence prediction system for the self-assembly behavior of monomeric components in traditional Chinese medicine, the system comprising: The input module is used to acquire multidimensional feature data of the target Chinese medicine monomer, input the multidimensional feature data into the trained prediction model, and obtain the molecular structure embedding vector, environmental feature vector and image feature vector of the target Chinese medicine monomer based on the multidimensional feature data. The fusion module is used to fuse the molecular structure embedding vector, environmental feature vector, and image feature vector to obtain a fused feature vector; The prediction module is used to perform multi-task prediction based on the fused feature vector and output the self-assembly behavior prediction result.

[0014] This application also provides a terminal, including: a memory, a processor, and an artificial intelligence prediction program for the self-assembly behavior of traditional Chinese medicine monomer components stored in the memory and executable on the processor. When the artificial intelligence prediction program for the self-assembly behavior of traditional Chinese medicine monomer components is executed by the processor, it implements the steps of the artificial intelligence prediction method for the self-assembly behavior of traditional Chinese medicine monomer components as described above.

[0015] This application also provides a computer-readable storage medium storing a computer program that can be executed to implement the steps of the artificial intelligence prediction method for the self-assembly behavior of traditional Chinese medicine monomer components as described above.

[0016] The beneficial effects of this invention are as follows: The method of this embodiment acquires multidimensional feature data of the target Chinese herbal medicine monomer, inputs the multidimensional feature data into a trained prediction model, and obtains the molecular structure embedding vector, environmental feature vector, and image feature vector of the target Chinese herbal medicine monomer based on the multidimensional feature data; the molecular structure embedding vector, environmental feature vector, and image feature vector are fused to obtain a fused feature vector; multi-task prediction is performed based on the fused feature vector, and the self-assembly behavior prediction result is output. By adopting a three-modal feature fusion strategy of molecular structure features, environmental parameter features, and characterizing image features, multi-task synchronous prediction of Chinese herbal medicine monomers is achieved, breaking through the limitations of traditional experimental trial and error methods, significantly shortening the R&D cycle of self-assembled Chinese herbal medicine preparations, reducing R&D costs, and realizing the efficient development of self-assembled drugs from Chinese herbal medicine monomers. Attached Figure Description

[0017] Figure 1 This is a flowchart of a preferred embodiment of the artificial intelligence prediction method for the self-assembly behavior of traditional Chinese medicine monomer components in this invention.

[0018] Figure 2 This is a schematic diagram of the three-modal feature fusion process of the artificial intelligence prediction method for the self-assembly behavior of traditional Chinese medicine monomer components in this invention.

[0019] Figure 3 This is a schematic diagram of the CNN-GNN-Transformer hybrid model structure of the artificial intelligence prediction method for the self-assembly behavior of traditional Chinese medicine monomer components in this invention.

[0020] Figure 4 This is a diagram of the artificial intelligence system module architecture in this invention.

[0021] Figure 5 This is a functional principle block diagram of a preferred embodiment of the artificial intelligence prediction system for the self-assembly behavior of traditional Chinese medicine monomer components in this invention.

[0022] Figure 6 This is a functional principle block diagram of a preferred embodiment of the terminal in this invention. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of this invention clearer and more explicit, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0024] Traditional R&D methods rely on experimental trial and error, searching for the optimal solution through a large number of combinations of screening conditions, which has the following significant drawbacks: First, the research and development cycle is long: the screening of self-assembly conditions for single Chinese medicine monomers takes several months to several years, which is difficult to meet the needs of rapid research and development. Second, the experimental costs are high: it involves a large amount of reagent consumption, instrument use and manpower input, and the experimental costs of rare Chinese herbal medicine monomers are even higher; Third, low prediction accuracy: lack of systematic analysis of multi-factor interactions makes it impossible to accurately predict whether self-assembly will occur and the characteristics of the products; Fourth, low data utilization: Existing self-assembled data is scattered across literature and experimental reports, lacking a unified database and thus unable to be effectively reused. In recent years, the application of artificial intelligence technology in drug development has deepened, but existing methods have significant shortcomings: Fifth, lack of specificity: Existing models are designed for chemically synthesized drugs and do not take into account the diversity of monomer structures and the complexity of physicochemical properties of traditional Chinese medicine, making them difficult to apply directly to the field of traditional Chinese medicine. Sixth, single feature dimension: It only focuses on molecular structure or a single environmental parameter, without integrating visual data such as characterization images, and cannot fully reflect multi-dimensional influencing factors; Seventh, limitations of model architecture: using a single machine learning model (such as random forest or single Transformer) makes it difficult to handle multimodal features; Eighth, the problem of data scarcity: There is a lack of experimental data on the self-assembly of Chinese herbal monomers, and existing models do not adopt effective transfer learning strategies, resulting in insufficient training and poor generalization ability.

[0025] Therefore, developing self-assembly behavior prediction methods and systems that target the characteristics of Chinese medicine monomers, integrate multimodal features, and adopt innovative hybrid model architectures are of great significance for promoting the efficient research and development of self-assembled Chinese medicine formulations and advancing the modernization of Chinese medicine.

[0026] The method provided in this application can solve the problems of long development cycle, high cost, low prediction accuracy and low data utilization of traditional Chinese medicine monomer self-assembly in the prior art. It is applicable to the research and development process of traditional Chinese medicine monomer self-assembly nano-formulation and can realize accurate prediction of self-assembly behavior and product characteristics.

[0027] Please see Figure 1 The artificial intelligence prediction method for the self-assembly behavior of traditional Chinese medicine monomer components described in this embodiment of the invention includes the following steps: Step S100: Obtain multidimensional feature data of the target Chinese medicine monomer, input the multidimensional feature data into the trained prediction model, and obtain the molecular structure embedding vector, environmental feature vector and image feature vector of the target Chinese medicine monomer based on the multidimensional feature data.

[0028] The multidimensional feature data includes: basic information on Chinese herbal monomers, physicochemical property data, self-assembly environmental conditions data, self-assembly product characteristic data, and characterization and detection data.

[0029] Specifically, the basic information of Chinese medicine monomers includes: monomer name, source of Chinese medicine, chemical structure type (flavonoids, alkaloids, terpenes, phenolic acids, etc.), CAS number, and molecular expression of SMILES.

[0030] Physicochemical properties include: molecular weight (MW, g / mol, range: 100-1500), lipid-water partition coefficient (logP, range: -5-10), number of hydrogen bond donors (HBD, range: 0-10), number of hydrogen bond acceptors (HBA, range: 1-20), topological polar surface area (TPSA, Ų, range: 10-500), number of rotatable bonds (RB, range: 0-30), dissociation constant (pKa, range: 2-12), solubility (S, mg / mL, range: 0.001-100), melting point (MP, °C, range: 50-350), and boiling point (BP, °C, range: 100-600).

[0031] The self-assembly environmental conditions data include: solution concentration (C, unit: mg / mL, range: 0.01-50), pH value (range: 2-10), temperature (T, unit: ℃, range: 4-60), solvent type (aqueous phase, ethanol phase, propylene glycol phase, mixed phase), ionic strength (I, unit: mol / L, range: 0-0.5), stirring rate (R, unit: rpm, range: 0-500), and incubation time (t, unit: h, range: 0.5-72).

[0032] Self-assembled product characteristic data include: self-assembly occurrence indicator (yes / no), product type (micelles, nanoparticles, nanofibers, vesicles, etc.), particle size (D, unit: nm, range: 10-500), particle size distribution (PDI, range: 0.1-0.8), zeta potential (ZP, unit: mV, range: -60-60), critical micelle concentration (CMC, unit: mg / mL, range: 0.001-10), and in vitro stability period (Tstab, unit: d, range: 1-180).

[0033] Characterization and detection data include: transmission electron microscopy (TEM) images (resolution: 1000-100000x), scanning electron microscopy (SEM) images (resolution: 1000-100000x), dynamic light scattering (DLS) patterns, X-ray diffraction (XRD) patterns, and Fourier transform infrared (FT-IR) patterns.

[0034] In this embodiment, obtaining the molecular structure embedding vector, environmental feature vector, and image feature vector of the target traditional Chinese medicine monomer based on the multidimensional feature data includes: obtaining the expression of the traditional Chinese medicine monomer in the multidimensional feature data; converting the expression of the traditional Chinese medicine monomer into a molecular graph structure; constructing graph data with atoms as nodes and chemical bonds as edges; learning the topological relationship between atoms through the graph convolutional layer of a graph neural network; and outputting the molecular structure embedding vector of a predetermined dimension through a global average pooling layer; obtaining environmental parameters of various types in the multidimensional feature data; encoding the environmental parameters of various types based on the pre-constructed correspondence between types and encoding methods to obtain an environmental feature vector of a predetermined dimension; obtaining transmission electron microscopy (TEM) images in the multidimensional feature data; preprocessing the TEM images; extracting deep texture and morphological features from the preprocessed TEM images; and outputting the image feature vector of a predetermined dimension through global average pooling.

[0035] Specifically, the SMILES expression of Chinese herbal monomers is converted into a molecular graph structure. Graph data is constructed with atoms as nodes (recording atom type, charge, and valence attributes) and chemical bonds as edges (recording bond type and bond length attributes). The topological relationships between atoms are learned through two graph convolutional layers of a GNN (graph convolutional network) (with 128 and 256 convolutional kernels, respectively). A molecular structure embedding vector with a dimension of 256 is output through a global average pooling layer.

[0036] One-hot encoding was used for categorized environmental parameters (solvent type, herbal origin, chemical structure type), such as encoding "aqueous phase" as [1,0,0,0] and "ethanol phase" as [0,1,0,0]. Numerical environmental parameters (concentration, pH, temperature, etc.) were standardized using Z-score. The concatenated environmental parameters were then transformed into a 256-dimensional environmental feature vector via a linear projection layer (weight matrix dimension: n×256, where n is the initial feature length). The formula for Z-score standardization is: ,in, This represents the standardized numerical environmental parameters. This represents the original numerical environmental parameters. This represents the mean of a numerical environmental parameter. The standard deviation of a numerical environmental parameter.

[0037] TEM (Transmission Electron Microscopy) images were selected as the core data. TEM images are the most intuitive way to characterize particle size, and most characterization methods for self-assembled nanoparticles involve TEM images, which provide direct evidence of self-assembly. Preprocessing of the TEM images included size unification (224×224 pixels), normalization ([0,1] interval), random flipping, and Gaussian blur enhancement. The first 10 layers of ResNet50 (including convolutional layers, batch normalization layers, ReLU activation layers, and pooling layers) were used to extract deep texture and morphological features, and global average pooling was used to output a 256-dimensional image feature vector.

[0038] like Figure 1 As shown, the artificial intelligence prediction method for the self-assembly behavior of traditional Chinese medicine monomer components further includes the following steps: Step S200: The molecular structure embedding vector, environmental feature vector, and image feature vector are fused to obtain a fused feature vector.

[0039] In this embodiment of the application, step S200 specifically includes: Step S210: Calculate the first attention weight corresponding to the molecular structure embedding vector, the second attention weight corresponding to the environmental feature vector, and the third attention weight corresponding to the image feature vector using the pre-constructed modal attention weight calculation function; Step S220: Based on the first attention weight, the second attention weight, and the third attention weight, the molecular structure embedding vector, the environmental feature vector, and the image feature vector are weighted and fused to obtain a fusion vector of a predetermined dimension; Step S230: Concatenate the fusion vector with the molecular structure embedding vector, environmental feature vector and image feature vector to obtain the fusion feature vector.

[0040] Specifically, the pre-constructed modal attention weight calculation function is as follows: ,in, This represents the attention weights corresponding to the feature vectors of each modality. This represents the feature vector for each modality. When m=1, it corresponds to the molecular structure embedding vector; when m=2, it corresponds to the environmental feature vector; and when m=3, it corresponds to the image feature vector. This represents the learnable weight matrix (256×256). The bias term (256×1) is represented by sigma. The activation function is used to output weights in the range [0,1].

[0041] When weighted and fused three-modal features, the expression is: ; in, Represents the fusion vector. This represents the first attention weight corresponding to the molecular structure embedding vector. Represents the molecular structure embedding vector. This represents the second attention weight corresponding to the environmental feature vector. Represents the environmental feature vector. This represents the third attention weight corresponding to the image feature vector. This represents the image feature vector. The fusion vector is 256-dimensional. The fusion vector is then concatenated with the original features of each modality to form a final 768-dimensional fusion feature vector.

[0042] like Figure 1 As shown, the artificial intelligence prediction method for the self-assembly behavior of traditional Chinese medicine monomer components further includes the following steps: Step S300: Perform multi-task prediction based on the fused feature vector and output the self-assembly behavior prediction result.

[0043] In this embodiment of the application, step S300 specifically includes: Step S310: Perform position encoding on the fused feature vector, perform multi-task prediction on the position-encoded fused feature vector, and output the self-assembly behavior prediction result. Step S320: The self-assembly behavior prediction results include: self-assembly occurrence probability, product particle size, zeta potential, morphology type, and in vitro stability period.

[0044] Specifically, an improved sine-cosine function is adopted, and adaptive parameters are introduced. For example, the adaptive parameter for the molecular structure embedding vector is 0.1, the adaptive parameter for the environmental feature vector is 0.2, and the adaptive parameter for the image feature vector is 0.3. The formula is: Even-numbered dimensions: ; Odd-numbered dimensions: ; Where pos is the position index (0), i is the dimension index (0≤i<384), and d is the fusion feature dimension (768), generating the input sequence. .

[0045] In this embodiment of the application, the training steps of the prediction model include: Step S10: Construct a multi-dimensional database, which includes several Chinese herbal medicine monomers and their corresponding multi-dimensional feature data. Step S20: Preprocess the multidimensional feature data in the multidimensional database, and divide the preprocessed multidimensional feature data into molecular structure features, environmental parameter features and characterization image features; Step S30: Perform feature encoding and extraction on molecular structure features, environmental parameter features and characterization image features respectively to obtain molecular structure embedding vector, environmental feature vector and image feature vector corresponding to the Chinese herbal medicine monomers in the multi-dimensional database; Step S40: The molecular structure embedding vector, environmental feature vector and image feature vector corresponding to the Chinese herbal medicine monomers in the multi-dimensional database are fused to obtain the fused feature vector. The fused feature vector is then positionally encoded to generate the input sequence. Step S50: Divide the input sequence into a training set, a validation set, and a test set. Use a transfer learning strategy to pre-train and fine-tune the constructed prediction model. Optimize the hyperparameters using the validation set to obtain the trained prediction model. Use the test set to evaluate the trained prediction model.

[0046] The specific steps include: (1) Construct a multi-dimensional database of the self-assembly behavior of Chinese herbal monomers. The multi-dimensional feature data in the multi-dimensional database includes: basic information of Chinese herbal monomers, physicochemical property data, self-assembly environmental condition data, self-assembly product characteristic data and characterization and detection data; (2) Preprocess the multidimensional feature data, including data cleaning, outlier removal, and missing value filling. Divide the preprocessed data features into three categories: molecular structure features, environmental parameter features, and characterization image features, and determine the variable range of each type of feature. (3) The three types of features are encoded and extracted respectively. The molecular structure features are encoded using graph neural network (GNN) topological structure encoding. The environmental parameter features are encoded using a combination of one-hot encoding (categorical) and standardization (numerical). The image features are characterized by deep visual features extracted using convolutional neural network (CNN). (4) Use attention mechanism to fuse three-modal features, construct a feature vector of unified dimension, perform position encoding on the fused feature vector, and generate input sequence; (5) Divide the input sequence into training set, validation set and test set (ratio of 7:1.5:1.5), use transfer learning strategy to pre-train and fine-tune the pre-built CNN-GNN-Transformer hybrid artificial intelligence model, optimize the hyperparameters through the validation set, and use the test set to evaluate the model's generalization ability.

[0047] In one specific embodiment, the data structure of the multi-dimensional database is shown in Table 1: Table 1

[0048] The database uses MySQL for storage and is protected by AES-256 encryption. It supports data querying, updating, and exporting, and allows setting administrator (data management) and regular user (query and export) permissions to ensure data security.

[0049] When preprocessing multidimensional feature data in a multidimensional database, ensure that the data quality meets the requirements for model training. The training requirements include: (1) Data cleaning: use box plot method to identify outliers (data exceeding 1.5 times the interquartile range is judged as outliers), and remove obvious outlier data (such as particle size <10nm or >500nm); (2) Missing value filling: use KNN algorithm (K=5) to fill numerical missing data, use mode filling for categorical missing data, and directly remove samples with a missing rate >20%; (3) Feature classification: divide the preprocessed data into molecular structure features (SMILES, MW, logP, etc.), environmental parameter features (concentration, pH, temperature, etc.), and characterization image features (TEM / SEM images); (4) Data standardization: numerical features are standardized by Z-score. This ensures that the data are on the same scale and avoids model weight shifts.

[0050] like Figure 2 As shown, Figure 2 This is a schematic diagram of the three-modal feature fusion process, which shows the fusion process of molecular structure features (GNN encoding), environmental parameter features (one-hot + normalization), and representation image features (CNN extraction). It includes attention weight calculation (sigmoid function), weighted fusion (vector multiplication and addition), feature concatenation (256-dimensional × 3 → 768-dimensional), and position encoding (sine-cosine function), which finally generates the model input sequence.

[0051] In one embodiment of this application, the prediction model includes: a feature encoding layer, a modality fusion layer, a multi-task prediction layer, and a loss function optimization layer; the feature encoding layer includes: a convolutional neural network sublayer, a graph neural network sublayer, and a fully connected encoding layer.

[0052] Specifically, the prediction model is a CNN-GNN-Transformer hybrid model, which can achieve simultaneous prediction across multiple tasks. The model structure is as follows: Figure 3 As shown, Figure 3 The model's hierarchical structure is shown from bottom to top as follows: feature encoding layer (GNN sub-layer, CNN sub-layer, fully connected encoding layer), modality fusion layer (attention weight calculation + feature concatenation), Transformer encoder layer (4 stacked encoder blocks), multi-task prediction layer (4 parallel branches), and loss function optimization layer. The core parameters of each layer (such as the number of convolutional kernels, number of neurons, and number of heads) and output dimensions (such as 256-dimensional or 768-dimensional) are labeled. This includes: (1) Feature encoding layer: A dedicated encoding path is designed for the three-modal features of “molecular structure + environmental parameters + image” to ensure that the core information of each feature is not lost. GNN captures molecular topological relationships, fully connected layer processes structured environmental parameters, CNN extracts microscopic morphological features of image, and finally outputs a 256-dimensional unified dimension vector.

[0053] Specifically, the feature encoding layer contains three parallel sub-layers: the GNN sub-layer contains two graph convolutional layers (128 and 256 kernels) + one pooling layer to process molecular structure data; the CNN (convolutional neural network) sub-layer contains the first 10 layers of ResNet50 to process image representation data; and the fully connected encoding layer (two fully connected layers with 128 and 256 neurons) processes environmental parameters.

[0054] (2) Modal fusion layer: an attention mechanism is introduced to dynamically allocate modal weights (e.g., when the molecular structure feature weight is higher, the model will prioritize relying on it to judge the possibility of self-assembly), and then the information density is improved by "weighted fusion + original feature splicing". Finally, position encoding is added to provide sequence position information for Transformer.

[0055] Specifically, the attention weight calculation unit of the modality fusion layer contains a 2-layer fully connected network (128 hidden neurons) to calculate the weights of each modality; the feature concatenation unit concatenates the original features of each modality after weighted fusion and outputs a 768-dimensional vector.

[0056] (3) Transformer encoder layer: Four encoder blocks are stacked to enhance feature interaction capabilities. The multi-head self-attention layer captures the correlation of different feature dimensions simultaneously (such as "the synergistic effect of particle size and temperature"). Residual connections and layer normalization avoid the vanishing of training gradients and ensure stable convergence of deep networks.

[0057] Specifically, the Transformer encoder layer includes four stacked encoder blocks, each containing an eight-head multi-head self-attention layer with a dimension of 96, capturing long-range dependencies of features; the feedforward neural network contains 2048 hidden neurons, enabling non-linear transformation of features; layer normalization and residual connections improve training stability and avoid gradient vanishing.

[0058] (4) Multi-task prediction layer: Parallel branch design realizes simultaneous prediction of "classification + regression" multi-task, avoiding training bias of single task. Each branch deepens feature expression through 2 fully connected layers, and the activation function is selected as needed (Sigmoid adapts to probability, Softmax adapts to multi-class, linear activation adapts to continuous parameters).

[0059] Specifically, the multi-task prediction layer includes four parallel branches: the self-assembly probability branch contains two fully connected layers (512 and 256 neurons) + sigmoid, outputting probabilities (0~1); the product structure parameter branch contains two fully connected layers (512 and 256 neurons) + linear activation, outputting particle size (nm) and zeta potential (mV); the morphology type branch contains two fully connected layers (512 and 256 neurons) + softmax, outputting the morphology probability distribution; and the stability period branch contains two fully connected layers (512 and 256 neurons) + linear activation, outputting the stability period (d).

[0060] (5) Loss function optimization layer: The weighted multi-task loss function expression is: ; in, For binary cross-entropy, , All are mean square errors. To balance the training priority of each task through multi-class cross-entropy, for The corresponding weight can take a value of 0.2; for The corresponding weight can take a value of 0.3. for The corresponding weight can take a value of 0.25. for The corresponding weight can be 0.25.

[0061] The weighted loss function balances the priorities of each task (e.g., the prediction of product structure parameters has the highest weight, since particle size / zeta potential directly affects the efficacy of the formulation), guiding the model to optimize all prediction objectives at the same time and improve the overall prediction accuracy.

[0062] Step S30 specifically includes: Step S31: Perform feature encoding and extraction on molecular structure features, environmental parameter features, and characterization image features respectively to obtain the molecular structure embedding vector, environmental feature vector, and image feature vector corresponding to the traditional Chinese medicine monomers in the multi-dimensional database, including: Step S32: Encode the molecular structure features using the sub-layer of the convolutional neural network to obtain the molecular structure embedding vector; Step S33: Encode the image features using the graph neural network sublayer to obtain the image feature vector; Step S34: Use the fully connected coding layer to perform deep feature extraction on the environmental parameter features to obtain the environmental feature vector.

[0063] Specifically, differential coding and extraction methods are used to target molecular structure features, environmental parameter features, and characterization image features.

[0064] When encoding molecular structure features, the SMILES expression is converted into a molecular graph structure (atoms are nodes, chemical bonds are edges); the atomic topological relationships are learned through two graph convolutional layers of GNN (128 and 256 kernels), and a global average pooling layer outputs a 256-dimensional molecular structure embedding vector. For example, the first 10 dimensions of the encoded vector of curcumin SMILES (CC(=O)OC1=CC(=C(C=C1)O)C2=CC=C(C=C2)O) are [0.125, 0.089, 0.156, 0.078, 0.210, 0.132, 0.095, 0.168, 0.102, 0.087].

[0065] When encoding environmental parameters, categorical parameters (solvent type, structure type) are encoded using one-hot encoding, such as "aqueous phase" being encoded as [1,0,0,0]; numerical parameters (concentration, pH, etc.) are standardized by Z-score and then concatenated with the one-hot results; the concatenated data is then converted into a 256-dimensional environmental feature vector through a linear projection layer (n×256 weight matrix).

[0066] When performing deep feature extraction on the characterization image, the TEM image is preprocessed as follows: the size is uniformly set to 224×224 pixels, normalized to [0,1], randomly flipped and enhanced with Gaussian blur; the first 10 layers of ResNet50 are used to extract deep features, and a 256-dimensional image feature vector is output after global average pooling to reflect the microscopic morphology information of the product.

[0067] In this embodiment of the application, step S40 specifically includes: Step S41: Calculate the attention weights corresponding to the molecular structure embedding vector, environmental feature vector, and image feature vector of the Chinese herbal medicine monomers in the multi-dimensional database using the pre-constructed modal attention weight calculation function; Step S42: Based on the attention weights corresponding to the molecular structure embedding vector, environmental parameter features, and characterization image features, perform weighted fusion of the molecular structure embedding vector, environmental feature vector, and image feature vector; Step S43: Concatenate the weighted fused feature vector with the corresponding molecular structure embedding vector, environmental feature vector and image feature vector to obtain the fused feature vector; Step S44: Perform positional encoding on the fused feature vector to generate the input sequence.

[0068] Specifically, the modal attention weight calculation function is as follows: The modal attention weight calculation function is used to dynamically allocate the weights of each modality (molecular structure, environmental parameters, and representation image weight range [0,1]).

[0069] The feature-weighted fusion formula is: This yields a 256-dimensional fusion vector, which is then concatenated with the original features of each modality to form a final 768-dimensional fusion vector.

[0070] An improved sine-cosine function (with adaptive parameters) is adopted. This generates a positional encoding vector, which is then concatenated with the final fused vector to form the model input sequence. .

[0071] In one specific embodiment, the transfer learning strategy in step S50 includes: (1) Divide the dataset into a training set (8400 records), a validation set (1800 records), and a test set (1800 records) in a ratio of 7:1.5:1.5. (2) Pre-training stage: 200,000 small molecule self-assembly data from the PubChem database were selected; the Adam optimizer was used for parameter optimization, with a learning rate of 0.001, a batch size of 64, and 100 training rounds; the basic network parameters (GNN, CNN, Transformer encoder) were trained. (2) Fine-tuning stage: The initial weights are the parameters of the pre-trained model; 70% of the parameters of the basic network are frozen, and only the fusion layer, prediction layer and part of the encoding layer are trained; the AdamW optimizer weights decay by 0.01, the learning rate is 0.0003, the batch size is 32, and the training is carried out for 50 rounds; if the validation set loss does not decrease for 10 consecutive rounds, the training stops. (3) Hyperparameter optimization: The Bayesian optimization algorithm was selected, and the optimization parameters were learning rate (0.0001-0.001), batch size (16-64), number of Transformer layers (3-6), and number of GNN convolution kernels (64-256). The goal was to minimize the total loss of the validation set. (4) Model Evaluation: The evaluation metrics for the test set include self-assembly probability prediction (accuracy, precision, recall, F1), structural parameter prediction (RMSE, MAE, R²), morphology prediction (accuracy, Kappa coefficient), and stability prediction (RMSE, MAE, R²). The test set evaluation results require an accuracy ≥ 92.3%, RMSE ≤ 5.8, and R² ≥ 0.87.

[0072] In one embodiment, the data output from the self-assembly behavior prediction results include: (1) the probability of self-assembly (e.g., 0.95) and the most likely product morphology (e.g., micelles, probability 0.82); (2) detailed parameters: particle size (e.g., 85.6 nm), PDI (0.23), zeta potential (-28.5 mV), CMC (0.05 mg / mL), and stability period (60 days). (3) optimization suggestions: adjust pH to 6.5 and concentration to 5 mg / mL to improve stability. The output formats include: visualization charts (morphology probability distribution, parameter error graph), and structured reports (Word / PDF).

[0073] This application also provides an artificial intelligence system for predicting the self-assembly behavior of monomeric components in traditional Chinese medicine. The system is built based on the above method, adopts a B / S architecture, and the modules interact through a communication bus. The structure is as follows: Figure 4 As shown, Figure 4 The system's eight modules are illustrated in a flowchart: database construction, data preprocessing, feature encoding and extraction, modality fusion and sequence generation, hybrid model construction, model training and optimization, self-assembly behavior prediction, and result visualization and output. These include: (1) Database construction module: collect and integrate basic information, physicochemical properties, self-assembly environment conditions, product characteristics and characterization data of Chinese herbal monomers, construct a multi-dimensional database, support data update, query (by monomer name / structure type / product type) and export (CSV / Excel / JSON), AES-256 encryption and user permission management; (2) Data preprocessing module: Cleaning the raw data (removing outliers using box plot method), filling missing values ​​(KNN / mode), feature classification and Z-score standardization, and generating a data quality analysis report (proportion of missing values, number of outliers, distribution histogram); (3) Feature encoding and extraction module: includes molecular structure encoding unit (SMILES to molecular map + GNN encoding), environmental parameter encoding unit (one-hot + normalization + linear projection), and image feature extraction unit (TEM preprocessing + ResNet50 extraction), outputting three types of feature vectors; (4) Modality fusion and sequence generation module: The attention weight calculation unit dynamically allocates modality weights, the feature fusion unit performs weighted concatenation, and the position encoding unit generates input sequences with position information; (5) Hybrid model building module: Visual configuration of model parameters (number of Transformer layers, number of GNN convolution kernels, etc.), supports model saving (.pth format) and loading; (6) Model training and optimization module: transfer learning training (pre-training + fine-tuning), Bayesian optimization of hyperparameters, visualization of the training process (loss curve, accuracy curve), and generation of model performance evaluation report; (7) Self-assembly behavior prediction module: Receives target single-unit feature data (SMILES, environmental parameters, image optional), calls the trained model to calculate (≤10s / sample), and caches historical prediction records; (8) Results visualization and output module: Generates product morphology probability distribution map and parameter prediction error map, outputs structured report (Word / PDF), and provides environmental condition optimization suggestions.

[0074] The database construction module is used to collect, integrate, store, and manage data related to the self-assembly of Chinese herbal monomers. Data collection includes: (1) Literature crawling: connecting to CNKI, PubMed, and Web of Science API; (2) Manual entry: entering experimental data through a visual interface; (3) Third-party import: supporting DrugBank, TCMSP, and PubChem data (CSV / Excel). Data management includes: (1) Query: querying by monomer name / structure type / product type; (2) Update: batch uploading and updating data; (3) Export: supporting CSV / Excel / JSON formats; (4) Data security: AES-256 encryption, user permission management (administrator / ordinary user), and operation log recording.

[0075] The data preprocessing module is used for cleaning, imputing, classifying, and standardizing data. Core algorithms include: outlier detection, such as box plots; missing value imputation, such as KNN (for numerical data) and mode (for categorical data); and standardization, such as Z-score. Visualization includes: generating data quality reports, such as the proportion of missing values, the number of outliers, and distribution histograms.

[0076] The feature encoding and extraction module includes: (1) Molecular structure encoding unit: SMILES to molecular map + GNN encoding, outputting a 256-dimensional vector; (2) Environmental parameter encoding unit: one-hot + normalization + linear projection, outputting a 256-dimensional vector; (3) Image feature extraction unit: TEM preprocessing + ResNet50 extraction, outputting a 256-dimensional vector; (4) Output: three types of feature encoding vectors.

[0077] The modality fusion and sequence generation module includes: (1) Attention weight calculation unit: dynamically calculates modality weights; (2) Feature fusion unit: weighted concatenation generates a 768-dimensional fusion vector; (3) Position encoding unit: generates position encoding and concatenates it into an input sequence; (4) Output: model input sequence (768-dimensional).

[0078] The hybrid model building module is used to build, initialize, save, and load hybrid models. Visual configuration includes: adjusting parameters such as the number of Transformer layers and the number of GNN convolutional kernels through the interface; model management includes: supporting saving and loading in .pth format and reusing trained models.

[0079] The model training and optimization module is used for: (1) training control: setting training rounds, learning rate, batch size, and visualizing the training process (loss curve, accuracy curve); (2) hyperparameter optimization: Bayesian optimization automatically searches for the optimal parameters; (3) performance evaluation: calculating evaluation indicators and generating model performance reports.

[0080] The self-assembly behavior prediction module includes: (1) Input interface: input SMILES, environmental parameters, and upload TEM image (optional); (2) Prediction calculation: call the model to calculate, with a time consumption of ≤10s / sample; (3) History record: cache recent prediction results and support viewing.

[0081] The results visualization and output module includes: (1) Visual charts: product morphology probability distribution chart, parameter prediction error chart; (2) Report generation: structured report (Word / PDF), including prediction results and optimization suggestions; (3) Interactive functions: adjust chart styles and download reports.

[0082] The system operating environment includes: (1) Hardware: server (Intel Core i9-13900K, 64GB memory, RTX4090), client (Intel Core i5, 16GB memory); (2) Software: server (Ubuntu 20.04, Python 3.9, PyTorch 1.13, MySQL 8.0), client (Windows 10 / 11, Chrome 90.0+); (3) Deployment: B / S architecture, users access the preset website through the browser without installing the client.

[0083] The database construction module of this application supports importing data from multiple sources: API crawling of literature databases (CNKI, PubMed, Web of Science), visualization and entry of experimental data, and import of CSV / Excel data from third-party databases (DrugBank, TCMSP, PubChem); data operation logs are recorded to prevent unauthorized modification.

[0084] This application's method constructs a multi-dimensional database covering the physicochemical properties, self-assembly environmental conditions, and product characterization data of traditional Chinese medicine (TCM) monomers. It innovatively employs a three-modal feature fusion strategy—"molecular structure features + environmental parameter features + characterization image features"—and designs a CNN-GNN-Transformer hybrid artificial intelligence model to achieve simultaneous multi-task prediction of the probability of TCM monomer self-assembly, product structural parameters (particle size, zeta potential, morphology), and stability. The system addresses the scarcity of TCM monomer data through transfer learning and introduces an attention mechanism to strengthen the weights of key features. Multiple validation experiments demonstrate a prediction accuracy exceeding 92.3% and a stability error of less than 5%. Therefore, this invention overcomes the limitations of traditional experimental trial-and-error methods, significantly shortens the development cycle of TCM self-assembled formulations, reduces development costs, and provides a novel technical tool for the efficient development of TCM monomer self-assembled drugs, possessing significant theoretical value and promising industrial application prospects.

[0085] The following are specific examples for illustration.

[0086] Example 1: This embodiment describes the construction of a multi-dimensional database of the self-assembly behavior of Chinese medicinal monomers. The database contains 500 kinds of Chinese medicinal monomers, and the steps are as follows: Step A1: Data Acquisition; Monomer screening: 500 kinds of Chinese medicinal monomers were selected, covering flavonoids (120 kinds), alkaloids (100 kinds), terpenes (80 kinds), phenolic acids (90 kinds), and saponins (110 kinds). Basic information: Obtain the monomer name, source, structure type, CAS number, and SMILES from TCMSP; Physicochemical properties: MW, logP, etc. were obtained from DrugBank and PubChem. Missing data were supplemented through experiments (e.g., solubility was measured by HPLC). Environmental and product data: 12,000 experimental records were extracted from PubMed and Web of Science literature. Characterization data: 8000 TEM / SEM images, 12000 DLS spectra, 8000 XRD spectra, and 10000 FT-IR spectra were extracted; Atlas: Uniform image size 224×224 pixels, atlas baseline correction.

[0087] Step A2: Data storage; A relational database was built using MySQL, and the following data tables were designed: basic information table of Chinese medicinal monomers, physicochemical property data table, environmental condition table, product characteristic table, and characterization image table. Tables are linked by a "single ID" to ensure data consistency; AES-256 encryption, with administrator (data management) and regular user (query and export) permissions set.

[0088] Step A3: Database verification; 1000 records from 50 different monomers were randomly selected to check data integrity (missing rate ≤5%) and accuracy (consistency with the original literature ≥98%). Test the query response time (≤1s) to verify the multi-condition query function.

[0089] The final database contains 500 individual entities, 12,000 experimental records, 8,000 images and corresponding atlases, meeting the needs of model training.

[0090] Example 2: This embodiment describes data preprocessing and feature encoding extraction. Taking curcumin (flavonoids), salvianolic acid B (phenolic acids), and ginsenoside Rg3 (saponins) as examples, the processing procedure is illustrated: Step B1: Data preprocessing; Data cleaning: Remove 12 abnormal data entries from 200 records (such as curcumin particles with a diameter of 600nm). Missing value filling: 5 missing values ​​for the solubility of salvianolic acid B (KNN filling, error ≤3%), 3 missing values ​​for the solvent type of ginsenoside Rg3 (mode filling is "aqueous phase"); Feature classification: molecular structure features (SMILES, MW, etc.), environmental parameters (concentration 5mg / mL, pH 6.5, etc.), characterization images (TEM); Standardization: Z-score standardization for numerical features (mean 0, standard deviation 1).

[0091] Step B2, Feature Encoding and Extraction; Molecular structure encoding: Curcumin SMILES are converted into a molecular diagram, and the GNN outputs a 256-dimensional vector (first 10 dimensions: [0.125, 0.089, 0.156, 0.078, 0.210, 0.132, 0.095, 0.168, 0.102, 0.087]). Environmental parameter encoding: "solvent = aqueous phase" one-hot encoding [1,0,0,0], concentration and pH are standardized and then concatenated, linearly projected as a 256-dimensional vector; Image extraction: After TEM preprocessing of curcumin, ResNet50 outputs a 256-dimensional vector to reflect the micelle morphology features.

[0092] Step B3, Feature Quality Assessment; Random forest feature importance: molecular structure (45%), environmental parameters (35%), and image (20%) all significantly affect self-assembly; The variance explained is ≥90%, preserving the original information.

[0093] Example 3: This embodiment describes the construction and training of a hybrid model. Based on the database from Embodiment 1, this embodiment constructs and trains a CNN-GNN-Transformer hybrid model.

[0094] Step C1: Model parameter configuration; Feature encoding layers: GNN (128, 256 kernels), CNN (first 10 layers of ResNet50), fully connected (128, 256 neurons); Modal fusion layer: attention computation (128 hidden neurons), 768-dimensional output; Transformer: 4 encoder blocks, 8 attention heads, 2048 feedforward neurons.

[0095] Loss function: .

[0096] Step C2: Model training; Dataset partitioning: 8400 training records, 1800 validation records, and 1800 test records; Pre-training: PubChem 200,000 data points, Adam (0.001), batch size 64, training 100 epochs, validation loss 0.32; Fine-tuning: Freeze 70% of the basic parameters, AdamW (0.0003, weight decay 0.01), batch size 32, training for 50 epochs, early stopping (no decrease in 10 epochs), validation loss 0.17; Hyperparameter optimization: Bayesian optimization of optimal parameters: learning rate 0.00025, batch size 32, Transformer 4 layers, GNN 256 cores.

[0097] Step C3: Model evaluation; The test set results are shown in Table 2: Table 2

[0098] Comparative experiments: Compared with single Transformer / GNN / CNN models, the self-assembly accuracy of the model of this invention is 8.5%-12.3% higher, and the RMSE is reduced by 15%-22%.

[0099] Example 4: This embodiment demonstrates model prediction and application validation. Using puerarin (flavonoids), baicalin (flavonoids), and forsythoside (flavonoids) as targets, the predictive performance is validated. (1) Prediction process: Input data: SMILES, environmental parameters (concentration 5 mg / mL, pH 6.8, 37℃, aqueous phase), TEM image (optional); Model calculation: The prediction results are output within 10 seconds, as shown in Table 3.

[0100] (2) Experimental verification: Self-assembly experiments were conducted under the predicted conditions. Morphology was observed using TEM, particle size / zeta potential was measured using DLS, and the cycle was measured using accelerated stability testing (40℃, RH75%). The experimental results are compared with the predictions in Table 4.

[0101] Table 3

[0102] Table 4

[0103] Verification conclusion: All errors are ≤5%, the prediction accuracy is high, and it can accurately reflect the self-assembly behavior.

[0104] Example 5: This embodiment describes the system deployment and usage process.

[0105] (1) System deployment: Server: Intel Core i9-13900K, 64GB RAM, 2TB SSD, RTX 4090, Ubuntu 20.04, Python 3.9, PyTorch 1.13, MySQL 8.0, Nginx; Clients: Windows 11, macOS 12.0+, Chrome 90.0+ browser; Access method: B / S architecture, http: / / www.tcm-selfassembly-prediction.ihbc.info (default URL), no client installation required.

[0106] (2) User flow: Step 1: Registration and Login: Register with your mobile phone number / email, and log in after administrator approval; Step 2: Data Query (Optional): Query existing data for the target entity; Step 3: Input parameters: Input SMILES, environmental conditions, and upload TEM image (optional); Step 4: Start Prediction: Click the button, and the calculation will be completed within 10 seconds; Step 5: Results Viewing: View visualizations and reports; Step 6: Report Export: Download the Word / PDF report and adjust the conditions according to the optimization suggestions.

[0107] This application achieves the following beneficial effects: First, high prediction accuracy: trimodal feature fusion + hybrid model, self-assembly accuracy ≥92.3%, parameter error ≤5%, which is better than existing single models; Second, it is highly targeted: it is designed based on the characteristics of Chinese herbal medicine monomers and a database of 500 monomers is built to solve the problem of poor adaptability; Third, high R&D efficiency: shortening the R&D cycle from several months to several days, improving efficiency by more than 60%; Fourth, low R&D costs: Reduced input of reagents, instruments, and manpower, resulting in a cost reduction of over 60%; Fifth, comprehensive functionality: multi-task synchronous prediction + optimization suggestions, covering the entire R&D process; Sixth, high flexibility: supports importing data from multiple sources and customizing parameters to adapt to diverse needs; Seventh, high data utilization: The unified database enables data reuse, providing support for subsequent research.

[0108] This invention breaks through the limitations of traditional experimental trial-and-error methods, providing a brand-new technical tool for the efficient development of self-assembled Chinese medicine preparations, promoting the modernization of Chinese medicine, and has significant theoretical value, industrial application prospects, and social benefits.

[0109] In one embodiment, such as Figure 5 As shown, based on the above-mentioned artificial intelligence prediction method for the self-assembly behavior of traditional Chinese medicine monomer components, the present invention also provides an artificial intelligence prediction system for the self-assembly behavior of traditional Chinese medicine monomer components, comprising: The input module 100 is used to acquire multidimensional feature data of the target Chinese medicine monomer, input the multidimensional feature data into the trained prediction model, and obtain the molecular structure embedding vector, environmental feature vector and image feature vector of the target Chinese medicine monomer based on the multidimensional feature data. The fusion module 200 is used to fuse the molecular structure embedding vector, the environmental feature vector, and the image feature vector to obtain a fused feature vector; The prediction module 300 is used to perform multi-task prediction based on the fused feature vector and output the self-assembly behavior prediction result.

[0110] It should be noted that the explanation of the above-mentioned embodiment of the artificial intelligence prediction method for the self-assembly behavior of traditional Chinese medicine monomer components also applies to the artificial intelligence prediction system for the self-assembly behavior of traditional Chinese medicine monomer components in this embodiment, and will not be repeated here.

[0111] This invention discloses an artificial intelligence prediction system for the self-assembly behavior of traditional Chinese medicine (TCM) monomer components. By acquiring multidimensional feature data of the target TCM monomer and inputting this data into a trained prediction model, the system obtains the molecular structure embedding vector, environmental feature vector, and image feature vector of the target TCM monomer based on the multidimensional feature data. The system then fuses these vectors to obtain a fused feature vector. Based on this fused feature vector, multi-task prediction is performed, and the self-assembly behavior prediction result is output. Employing a three-modal feature fusion strategy combining molecular structure features, environmental parameter features, and characterizing image features, the system achieves simultaneous multi-task prediction of TCM monomers. This overcomes the limitations of traditional experimental trial-and-error methods, significantly shortens the development cycle of TCM self-assembly formulations, reduces development costs, and enables the efficient development of TCM monomer self-assembly drugs.

[0112] Figure 6 A schematic diagram of the structure of a terminal provided in an embodiment of this application. The terminal may include: The memory 501, the processor 502, and the computer program stored on the memory 501 and capable of running on the processor 502.

[0113] When the processor 502 executes the program, it implements the artificial intelligence prediction method for the self-assembly behavior of traditional Chinese medicine monomer components provided in the above embodiments.

[0114] Furthermore, the terminal also includes: Communication interface 503 is used for communication between memory 501 and processor 502.

[0115] The memory 501 is used to store computer programs that can run on the processor 502.

[0116] The memory 501 may include high-speed RAM memory, and may also include non-volatile memory, such as at least one disk storage device.

[0117] If the memory 501, processor 502, and communication interface 503 are implemented independently, they can be interconnected via a bus to communicate with each other. The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of representation, only one line is used in the diagram, but this does not imply that there is only one bus or one type of bus.

[0118] Optionally, in a specific implementation, if the memory 501, processor 502, and communication interface 503 are integrated on a single chip, then the memory 501, processor 502, and communication interface 503 can communicate with each other through an internal interface.

[0119] Processor 502 may be a central processing unit (CPU), an application specific integrated circuit (ASIC), or one or more integrated circuits configured to implement the embodiments of this application.

[0120] This embodiment also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-mentioned artificial intelligence prediction method for the self-assembly behavior of traditional Chinese medicine monomer components.

[0121] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0122] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of that feature. In the description of this application, "N" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0123] Any process or method description in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or N executable instructions for implementing custom logic functions or processes, and the scope of the preferred embodiments of this application includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as should be understood by those skilled in the art to which embodiments of this application pertain.

[0124] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a ordered list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can read and execute instructions from or in conjunction with such an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by or in conjunction with an instruction execution system, apparatus, or device. More specific examples (a non-exhaustive list) of computer-readable media include: an electrical connection having one or more wires (electronic device), a portable computer disk drive (magnetic device), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). In addition, computer-readable media can even be paper or other suitable media on which programs can be printed, because programs can be obtained electronically by optically scanning paper or other media, followed by editing, interpreting or otherwise processing as necessary, and then stored in computer memory.

[0125] It should be understood that the various parts of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, the N steps or methods can be implemented using software or firmware stored in memory and executed by a suitable instruction execution system. If implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0126] Those skilled in the art will understand that all or part of the steps of the methods described in the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, it includes one or a combination of the steps of the method embodiments.

[0127] Furthermore, the functional units in the various embodiments of this application can be integrated into a processing module, or each unit can exist physically separately, or two or more units can be integrated into a module. The integrated module can be implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium.

[0128] The storage medium mentioned above can be a read-only memory, a disk, or an optical disk, etc. Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions, and variations to the above embodiments within the scope of this application.

Claims

1. An artificial intelligence prediction method for the self-assembly behavior of monomeric components in traditional Chinese medicine, characterized in that, The method includes: The multidimensional feature data of the target Chinese herbal medicine monomer is obtained, and the multidimensional feature data is input into the trained prediction model. Based on the multidimensional feature data, the molecular structure embedding vector, environmental feature vector and image feature vector of the target Chinese herbal medicine monomer are obtained. The molecular structure embedding vector, environmental feature vector, and image feature vector are fused to obtain a fused feature vector; Multi-task prediction is performed based on the fused feature vector, and the self-assembly behavior prediction result is output.

2. The artificial intelligence prediction method for the self-assembly behavior of traditional Chinese medicine monomer components according to claim 1, characterized in that, Based on the multidimensional feature data, the molecular structure embedding vector, environmental feature vector, and image feature vector of the target traditional Chinese medicine monomer are obtained, including: Obtain the expression of the Chinese herbal monomer in the multidimensional feature data, convert the expression of the Chinese herbal monomer into a molecular graph structure, construct graph data with atoms as nodes and chemical bonds as edges, learn the topological relationship between atoms through the graph convolutional layer of the graph neural network, and output the molecular structure embedding vector of a predetermined dimension through the global average pooling layer. Obtain environmental parameters of various types from the multidimensional feature data, and encode each type of environmental parameter based on the pre-constructed correspondence between types and encoding methods to obtain an environmental feature vector of a predetermined dimension; The transmission electron microscope (TEM) image in the multidimensional feature data is acquired, the TEM image is preprocessed, deep texture and morphological features are extracted from the preprocessed TEM image, and a predetermined dimension image feature vector is output by global average pooling.

3. The artificial intelligence prediction method for the self-assembly behavior of traditional Chinese medicine monomer components according to claim 1, characterized in that, The molecular structure embedding vector, environmental feature vector, and image feature vector are fused to obtain a fused feature vector, including: The first attention weight corresponding to the molecular structure embedding vector, the second attention weight corresponding to the environmental feature vector, and the third attention weight corresponding to the image feature vector are calculated using a pre-constructed modal attention weight calculation function. Based on the first attention weight, the second attention weight, and the third attention weight, the molecular structure embedding vector, the environmental feature vector, and the image feature vector are weighted and fused to obtain a fusion vector of a predetermined dimension. The fusion vector is concatenated with the molecular structure embedding vector, the environmental feature vector, and the image feature vector to obtain the fusion feature vector.

4. The artificial intelligence prediction method for the self-assembly behavior of traditional Chinese medicine monomer components according to claim 1, characterized in that, Based on the fused feature vector, multi-task prediction is performed, and the self-assembly behavior prediction result is output, including: The fused feature vector is positionally encoded, and the positionally encoded fused feature vector is used for multi-task prediction to output the self-assembly behavior prediction result. The self-assembly behavior prediction results include: self-assembly occurrence probability, product particle size, zeta potential, morphology type, and in vitro stability period.

5. The artificial intelligence prediction method for the self-assembly behavior of traditional Chinese medicine monomer components according to claim 1, characterized in that, The training steps for the prediction model include: Construct a multi-dimensional database, which includes several Chinese herbal medicine monomers and their corresponding multi-dimensional feature data; The multidimensional feature data in the multidimensional database is preprocessed, and the preprocessed multidimensional feature data is divided into molecular structure features, environmental parameter features and characterization image features. Feature encoding and extraction were performed on molecular structure features, environmental parameter features, and characterization image features respectively to obtain molecular structure embedding vectors, environmental feature vectors, and image feature vectors corresponding to Chinese herbal monomers in a multi-dimensional database. The molecular structure embedding vector, environmental feature vector, and image feature vector corresponding to the Chinese herbal monomers in the multi-dimensional database are fused to obtain the fused feature vector. The fused feature vector is then positionally encoded to generate the input sequence. The input sequence is divided into a training set, a validation set, and a test set. The constructed prediction model is pre-trained and fine-tuned using a transfer learning strategy. The hyperparameters are optimized using the validation set to obtain the trained prediction model. The trained prediction model is then evaluated using the test set.

6. The artificial intelligence prediction method for the self-assembly behavior of traditional Chinese medicine monomer components according to claim 5, characterized in that, The prediction model includes: a feature encoding layer, a modality fusion layer, a multi-task prediction layer, and a loss function optimization layer; the feature encoding layer includes: a convolutional neural network sub-layer, a graph neural network sub-layer, and a fully connected encoding layer. Feature encoding and extraction were performed on molecular structure features, environmental parameter features, and characterization image features, respectively, to obtain molecular structure embedding vectors, environmental feature vectors, and image feature vectors corresponding to traditional Chinese medicine monomers in a multi-dimensional database, including: The molecular structure features are encoded using the sublayers of the convolutional neural network to obtain the molecular structure embedding vector; The image feature vector is obtained by encoding the image features using the graph neural network sublayer. The fully connected coding layer is used to perform deep feature extraction on environmental parameter features to obtain an environmental feature vector.

7. The artificial intelligence prediction method for the self-assembly behavior of traditional Chinese medicine monomer components according to claim 5, characterized in that, The molecular structure embedding vector, environmental feature vector, and image feature vector corresponding to traditional Chinese medicine monomers in a multi-dimensional database are fused to obtain a fused feature vector. The fused feature vector is then positionally encoded to generate an input sequence, including: The attention weights corresponding to the molecular structure embedding vector, environmental feature vector, and image feature vector of traditional Chinese medicine monomers in the multi-dimensional database are calculated using a pre-constructed modal attention weight calculation function. Based on the attention weights corresponding to molecular structure embedding vector, environmental parameter features, and image feature features, the molecular structure embedding vector, environmental feature vector, and image feature vector are weighted and fused. The weighted and fused feature vector is concatenated with the corresponding molecular structure embedding vector, environmental feature vector, and image feature vector to obtain the fused feature vector. Position encoding is performed on the fused feature vectors to generate the input sequence.

8. An artificial intelligence prediction system for the self-assembly behavior of monomeric components in traditional Chinese medicine, characterized in that, The system includes: The input module is used to acquire multidimensional feature data of the target Chinese medicine monomer, input the multidimensional feature data into the trained prediction model, and obtain the molecular structure embedding vector, environmental feature vector and image feature vector of the target Chinese medicine monomer based on the multidimensional feature data. The fusion module is used to fuse the molecular structure embedding vector, environmental feature vector, and image feature vector to obtain a fused feature vector; The prediction module is used to perform multi-task prediction based on the fused feature vector and output the self-assembly behavior prediction result.

9. A terminal, characterized in that, include: The system includes a memory, a processor, and an artificial intelligence prediction program for the self-assembly behavior of traditional Chinese medicine monomer components stored in the memory and executable on the processor. When the artificial intelligence prediction program for the self-assembly behavior of traditional Chinese medicine monomer components is executed by the processor, it implements the steps of the artificial intelligence prediction method for the self-assembly behavior of traditional Chinese medicine monomer components as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that can be executed to implement the steps of the artificial intelligence prediction method for the self-assembly behavior of traditional Chinese medicine monomer components as described in any one of claims 1 to 7.

Images