Multi-feature fusion oral bioavailability prediction method, device, equipment and medium

Abstract

The application provides a multi-feature fusion oral bioavailability prediction method, device, equipment and medium. Four-dimensional features of a candidate drug, including a graph structure feature, a fingerprint feature, a physicochemical descriptor and a pharmacophore feature, are fused to comprehensively depict drug molecular characteristics, so that the representation capability is stronger, and the prediction accuracy is improved. In addition, the fused features are processed by a plurality of different expert processing models corresponding to a plurality of preset different result types, so that different processing strategies can be adopted for different chemical structure types, and different application scenarios can be adapted.

Description

Technical Field

[0001] The embodiments of this invention relate to the field of computer information processing technology, specifically to a method, apparatus, device, and medium for predicting oral bioavailability through multi-feature fusion. Background Technology

[0002] Oral bioavailability (OB) is a crucial pharmacokinetic parameter that measures the extent and rate at which a drug enters the systemic circulation after oral administration. For traditional Chinese medicine (TCM) research, assessing the oral bioavailability of its chemical components is essential for screening active ingredients, optimizing administration routes, and predicting in vivo efficacy. However, experimental determination of oral bioavailability requires animal experiments or human clinical trials, which are costly, time-consuming, and subject to numerous ethical restrictions, making it difficult to comprehensively evaluate the thousands of chemical components in TCM. Therefore, the development of computational prediction methods is imperative.

[0003] Existing methods for predicting oral bioavailability (OB) are mainly based on empirical rules such as the Lipinski five-rule or machine learning models based on molecular descriptors. The Lipinski rule uses four simple indicators—molecular weight, lipid-water partition coefficient (LogP), number of hydrogen bond donors, and number of hydrogen bond acceptors—for screening. While computationally simple, it is too coarse to accurately distinguish the differences in bioavailability among different compounds and has poor applicability to natural products. Machine learning-based methods use more molecular descriptors as features to train regression or classification models to predict OB values. The main problems with these methods include: First, the selection of molecular descriptors relies on human experience, and different studies use different sets of descriptors, making it difficult to establish a unified standard. Second, traditional machine learning models struggle to capture complex nonlinear relationships and long-range dependencies in molecular structures. Third, existing models typically treat OB prediction as a single regression task, performing poorly when dealing with complex drug molecules and failing to meet the high-precision prediction requirements of modern drug discovery. Summary of the Invention

[0004] In view of this, the present invention provides a method, apparatus, device, and computer-readable storage medium for predicting oral bioavailability with high prediction accuracy through multi-feature fusion.

[0005] In a first aspect, some embodiments of the present invention provide a multi-feature fusion oral bioavailability prediction method, applied to a multi-feature fusion oral bioavailability prediction model, the multi-feature fusion oral bioavailability prediction method comprising:

[0006] Retrieve the molecular SMILES string corresponding to the candidate drug;

[0007] The SMILES string is parsed into a structured molecular object, and multi-dimensional features are extracted from the molecular object to obtain multi-dimensional features; wherein, the multi-dimensional features include graph structure features, fingerprint features, physicochemical descriptors and pharmacophore features;

[0008] The features in the multi-dimensional features are fused to obtain the fused features;

[0009] Multiple expert processing models corresponding to multiple different preset structure types are used to process the fusion features respectively. The probability that the molecular object belongs to multiple different preset structure types is determined according to the graph structure features. The output features of each expert processing model are weighted with the probability as the corresponding weight to obtain the expert processing features.

[0010] Based on the expert processing characteristics, ADME prediction is performed on the candidate drug, and the ADME prediction result is obtained based on the ADME prediction to obtain the target oral bioavailability score of the candidate drug.

[0011] In some embodiments, the multi-feature fusion oral bioavailability prediction method further includes:

[0012] Output the oral bioavailability score, and analyze the contribution of each sub-step parameter in the ADME prediction result to the oral bioavailability score;

[0013] The sub-step parameters include at least one of the following: absorption fraction, hepatic clearance, apparent volume of distribution, and free fraction of the candidate drug after oral administration.

[0014] In some embodiments, the multi-dimensional feature extraction of the molecular object includes:

[0015] Using atoms of the molecular object as nodes and chemical bonds as edges, the molecular object is represented as a graph structure, and the initialization features of each node and edge in the graph structure are determined.

[0016] The graph structure is encoded by a message passing network that controls message flow through a multi-layer gated loop unit; wherein, the encoding process includes: updating the hidden state of each node at each network layer of the message passing network by aggregating information of the neighboring nodes and connected edges of each node in the graph structure;

[0017] The final hidden states of each node after being encoded through multiple network layers are aggregated to obtain the structural graph features.

[0018] In some embodiments, the multi-dimensional feature extraction of the molecular object includes:

[0019] Calculate the extended connectivity fingerprint and bond fingerprint of the molecular object;

[0020] The extended connection fingerprint and the key fingerprint are concatenated to obtain a concatenated fingerprint. The concatenated fingerprint is then subjected to a linear transformation and GELU activation function processing to obtain the fingerprint features.

[0021] In some embodiments, the multi-dimensional feature extraction of the molecular object includes:

[0022] Calculate multiple primitive physicochemical descriptors characterizing the physicochemical properties of the molecular object;

[0023] Weights representing the importance of each of the original physicochemical descriptors are obtained through attention mechanisms;

[0024] Based on the weights corresponding to each of the original physicochemical descriptors, the original physicochemical descriptors are weighted and summed to obtain the physicochemical descriptors in the multi-dimensional features.

[0025] In some embodiments, the multi-dimensional feature extraction of the molecular object includes:

[0026] Key features characterizing the pharmacophore were identified from the molecular object;

[0027] The key features are mapped into a three-dimensional space, and the spatial distribution and interaction patterns of the key features in the three-dimensional space are extracted through a three-dimensional CNN network to generate pharmacophore features in the multi-dimensional feature extraction.

[0028] In some embodiments, fusing the features in the multi-dimensional features to obtain fused features includes:

[0029] The features in the multi-dimensional features are concatenated to obtain the concatenated features;

[0030] The fusion weight of each feature in the splicing features is calculated using multiple different gating mechanisms.

[0031] The fused features are obtained by weighting each feature in the multi-dimensional features based on the corresponding fusion weights.

[0032] In some embodiments, performing ADME prediction on the candidate drug includes:

[0033] Based on the expert processing features, the passive diffusion probability, P-glycoprotein efflux probability, and uptake transporter-mediated absorption probability of the candidate drug after oral administration are predicted using the corresponding prediction model.

[0034] According to the passive diffusion probability, the The glycoprotein efflux probability and the uptake transporter-mediated absorption probability are used to determine the absorption fraction used to assess the extent of absorption of the candidate drug by the subject after oral administration; wherein, the sub-parameters in the ADME prediction results include the absorption fraction, the absorption fraction, the passive diffusion probability, and the... The relationship between the glycoprotein efflux probability and the uptake transporter-mediated absorption probability satisfies:

[0035] ;

[0036] Among them, the The value representing the absorption fraction, The value representing the passive diffusion probability, the Characterizing the The value of the glycoprotein efflux probability, the The value characterizing the probability of absorption mediated by the uptake transporter.

[0037] In some embodiments, performing ADME prediction on the candidate drug includes:

[0038] For each cytochrome enzyme in a set of preset cytochrome enzymes associated with the metabolism of the candidate drug after oral administration, the substrate probability of the candidate drug as its substrate and the metabolic clearance rate mediated by it are predicted respectively. Using the substrate probability corresponding to each cytochrome enzyme in the set of preset cytochrome enzymes as a weighting weight, the metabolic clearance rates corresponding to each cytochrome enzyme are weighted and summed to obtain the total liver clearance rate, which characterizes the degree of metabolism of the candidate drug by the drug user. The sub-parameters in the ADME prediction results include the total liver clearance rate.

[0039] In some embodiments, predicting the oral bioavailability of the candidate drug based on the ADME prediction results to obtain an oral bioavailability score for the candidate drug includes:

[0040] The liver extraction rate is calculated based on the total liver clearance rate and the preset liver blood flow rate, and the liver availability is calculated based on the liver extraction rate.

[0041] The initial oral bioavailability score of the candidate drug is obtained based on the liver availability and the absorption fraction; wherein, the relationship between the liver availability, the absorption fraction, and the initial oral bioavailability score is as follows: The Characterizing the initial oral bioavailability score, the Characterizing the absorption fraction, the The liver availability is characterized by the regulating factor, which is an adjustable parameter of the oral bioavailability prediction model.

[0042] The initial oral bioavailability score was cropped to obtain... Target oral bioavailability score within the interval.

[0043] In some embodiments, the multi-feature fusion oral bioavailability prediction method further includes training the multi-feature fusion oral bioavailability prediction model using a two-stage transfer learning strategy; wherein the two-stage transfer learning strategy includes a first-stage training and a second-stage training.

[0044] The first stage of training includes: pre-training the multi-feature fusion oral bioavailability prediction model on a general molecular dataset whose dataset size meets preset conditions. The pre-training objective includes data-augmented contrastive learning loss. And the multi-task prediction loss for molecular properties;

[0045] The second stage of training includes: training the multi-feature fusion oral bioavailability prediction model on the target oral bioavailability dataset using a progressive parameter unfreezing strategy to adjust the model parameters. The adjustment range of the model parameters in the second stage of training is smaller than that in the first stage of training.

[0046] In some embodiments, the second phase of training further includes:

[0047] Oral bioavailability prediction of unlabeled samples is performed using the current multi-feature fusion oral bioavailability prediction model.

[0048] The prediction uncertainty is estimated based on Monte Carlo estimation;

[0049] High-confidence prediction results are selected as pseudo-labels and added to the training set for iterative optimization.

[0050] Secondly, some embodiments of the present invention provide a multi-feature fusion oral bioavailability prediction device, used to predict multi-feature fusion oral bioavailability using a multi-feature fusion oral bioavailability prediction model, the multi-feature fusion oral bioavailability prediction device comprising:

[0051] The acquisition module is used to obtain the molecular SMILES string corresponding to the candidate drug;

[0052] The multidimensional feature extraction module is used to parse the SMILES string into a structured molecular object and extract multidimensional features from the molecular object to obtain multidimensional features; wherein, the multidimensional features include graph structure features, fingerprint features, physicochemical descriptors and pharmacophore features;

[0053] The feature fusion module is used to fuse the features in the multi-dimensional features to obtain fused features;

[0054] The expert processing module is used to process the fusion features by employing multiple expert processing models corresponding to multiple different preset structure types, and to determine the probability that the molecular object belongs to multiple different preset structure types based on the graph structure features, and to weight the output features of each expert processing model with the probability as the corresponding weight to obtain the expert processing features.

[0055] The prediction module is used to perform ADME prediction on the candidate drug based on the expert processing features, and obtain the ADME prediction result based on the ADME prediction to obtain the target oral bioavailability score of the candidate drug.

[0056] Thirdly, some embodiments of the present invention provide a multi-feature fusion oral bioavailability prediction device, including a processor and a memory, wherein the memory is used to store programs, instructions or code, and the processor is used to execute the programs, instructions or code in the memory to complete the multi-feature fusion oral bioavailability prediction method as described in any of the foregoing claims.

[0057] Fourthly, some embodiments of the present invention provide a computer-readable storage medium storing a computer program, which is loaded by a processor to execute the multi-feature fusion oral bioavailability prediction method as described in any of the preceding claims.

[0058] In several embodiments provided by this invention, by fusing features from four dimensions—graph structure features, fingerprint features, physicochemical descriptors, and pharmacophore features—of the candidate drug's molecular object, the characteristics of the drug molecule can be comprehensively characterized, resulting in stronger characterization capabilities and improved prediction accuracy. Furthermore, by employing multiple different expert processing models corresponding to multiple preset result types to process the fused features, differentiated processing strategies can be adopted for different chemical structure types, making the prediction method provided by this invention particularly suitable for scenarios with high diversity in the structures of traditional Chinese medicine components. Attached Figure Description

[0059] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0060] Figure 1 This is a schematic diagram of the process flow corresponding to the multi-feature fusion oral bioavailability prediction method provided according to some embodiments of the present invention.

[0061] Figure 2 This is a flowchart illustrating the method for extracting structural graph features from molecular objects.

[0062] Figure 3 This is a flowchart illustrating the method for extracting fingerprint features from molecular objects.

[0063] Figure 4 This is a flowchart illustrating the method for extracting physicochemical descriptors from molecular objects.

[0064] Figure 5 This is a flowchart illustrating the method for extracting pharmacophore features from molecular objects.

[0065] Figure 6 This is a flowchart illustrating a method for fusing features from multiple dimensions according to some embodiments of the present invention.

[0066] Figure 7 This is a flowchart illustrating the method for predicting the ADME of the candidate drug according to some embodiments of the present invention.

[0067] Figure 8 This is a flowchart illustrating the method for predicting the ADME of the candidate drug according to some embodiments of the present invention.

[0068] Figure 9 This is a schematic diagram of a method for predicting the oral bioavailability score of a candidate drug based on the ADME prediction results, according to some embodiments of the present invention.

[0069] Figure 10 This is a schematic diagram of the system architecture corresponding to the multi-dimensional feature fusion oral bioavailability prediction method provided in the embodiments of the present invention.

[0070] Figure 11 Radar contribution map predicted by ADEM.

[0071] Figure 12 This is a schematic diagram of the expert network gating distribution.

[0072] Figure 13 This is a schematic diagram showing the structural type distribution of orally bioavailable components.

[0073] Figure 14 This diagram illustrates a comparison of experimental results between our method (a multi-feature fusion method for predicting oral bioavailability) and various baseline methods based on a regression task.

[0074] Figure 15 This diagram illustrates the comparison of the experimental results of our method (multi-feature fusion oral bioavailability prediction method) with those of various baseline methods in a binary classification task (OB≥30% is considered high bioavailability).

[0075] Figure 16 A schematic diagram showing the comparison of ablation test effects for multi-feature fusion oral bioavailability prediction provided by the present invention.

[0076] Figure 17 A diagram showing the performance comparison of fine-tuning for different pre-training configurations.

[0077] Figure 18 This is a schematic diagram of the structure of a multi-feature fusion oral bioavailability prediction device provided according to some embodiments of the present invention.

[0078] Figure 19 This is a schematic diagram of the structure of a multi-feature fusion oral bioavailability prediction device provided according to some embodiments of the present invention. Detailed Implementation

[0079] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0080] Please see Figure 1 The diagram shown is a flowchart illustrating the multi-feature fusion oral bioavailability prediction method provided according to some embodiments of the present invention. The multi-feature fusion oral bioavailability prediction method provided in these embodiments is applied to a multi-feature fusion oral bioavailability prediction model; that is, the multi-feature fusion oral bioavailability prediction model executes the multi-feature fusion oral bioavailability prediction method to obtain the corresponding prediction results. The multi-feature fusion oral bioavailability prediction method provided in these embodiments includes steps S01, S03, S05, S07, and S09, each described below.

[0081] S01: Obtain the SMILES string corresponding to the candidate drug.

[0082] Candidate drugs refer to drugs for which oral bioavailability prediction is required. SMILES (Simplified Molecular Input Line Entry System) is a specification that uses ASCII strings to explicitly describe molecular structures. It uses simple character combinations to represent information such as the types of atoms, their connections, bond types (single, double, triple), ring structures, and branches, enabling efficient computer recognition and processing of chemical structures. The corresponding molecular SMILES string can be obtained through chemical structure file conversion, or by looking up the compound name in a database.

[0083] S03: Parse the SMILES string into a structured molecular object, and extract multi-dimensional features from the molecular object to obtain multi-dimensional features; wherein, the multi-dimensional features include graph structure features, fingerprint features, physicochemical descriptors and pharmacophore features.

[0084] Parsing a molecular SMILES string into a structured molecular object refers to converting molecular structural information (such as CCO representing ethanol) into a structured data object that a computer can understand and manipulate. The process typically includes string parsing, constructing a molecular graph, and generating the molecular object. String parsing involves reading the SMILES string and identifying chemical information such as atoms, bonds, rings, and branches. Constructing a molecular graph involves converting this information into a graph structure, where nodes represent atoms and edges represent chemical bonds. Generating the molecular object specifically involves encapsulating this graph structure into a molecular object within a cheminformatics library (such as RDKit or Open Babel). This object contains atomic properties (such as element type and charge), bond properties (such as bond order), and topological information (such as ring structure).

[0085] Graph structural features are a mathematical representation of molecular structures in cheminformatics, used to describe the topological connections of molecules. Graph structural features mainly consist of nodes and edges. Nodes represent atoms in the molecule (hydrogen atoms are usually ignored), and each node can carry attribute labels such as atom type, charge, and hybridization state. Edges represent chemical bonds between atoms, including single, double, triple, or aromatic bonds, and can carry bond order information.

[0086] Fingerprint features are a mathematical representation method in cheminformatics that encodes molecular structure as a fixed-length binary vector. A physicochemical descriptor is a parameter used to quantitatively characterize the physicochemical properties of a compound; it converts molecular structural information into calculable numerical values, thereby establishing the relationship between structure and properties. Pharmacophore features are an abstract set of three-dimensional spatial and electronic characteristics of a drug molecule that play a crucial role in producing specific biological effects at biological targets (such as receptors and enzymes); they are the most basic structural requirements necessary for a drug to exert its efficacy.

[0087] S05: Fuse the features in the multi-dimensional features to obtain the fused features.

[0088] Feature fusion, in cheminformatics, drug design, or artificial intelligence modeling, refers to integrating molecular features from different representations—such as graph structural features, fingerprint features, physicochemical descriptors, and pharmacophore features—into a unified input framework to more comprehensively and accurately describe the properties and behavior of molecules. This fusion can overcome the limitations of single feature types and improve the predictive power and generalization performance of the model.

[0089] S07: Multiple expert processing models corresponding to multiple different preset structure types are used to process the fusion features respectively, and the probability that the molecular object belongs to multiple different preset structure types is determined according to the graph structure features. The output features of each expert processing model are weighted according to the probability as the corresponding weight to obtain the expert processing features.

[0090] Different preset structure types correspond to different expert processing models. After determining the preset structure type, the corresponding expert processing model can be determined based on the correspondence between the preset structure type and the expert processing model. The expert processing model is a neural network model that relies on the empirical rules of domain experts to construct the decision logic.

[0091] After being input into the corresponding expert processing model, the fused features are essentially processed and filtered by an expert. That is, the output features of the expert processing model are those processed and filtered by an "expert" to better predict oral bioavailability. In real-world scenarios, different experts may have different opinions on drug data. Therefore, the oral bioavailability prediction method provided in this embodiment uses multiple expert processing models corresponding to different preset structural types to process the fused features, resulting in different output features. The probability of a molecule belonging to different preset structural types is used as the weighting weight for the corresponding output features. The output features of each different expert processing model are then weighted and fused to obtain expert-processed features that incorporate the opinions of multiple experts. Therefore, the oral bioavailability predicted based on the expert-processed features provided in this embodiment is more accurate.

[0092] S09: Based on the expert processing characteristics, perform ADME prediction on the candidate drug, and obtain the ADME prediction result based on the ADME prediction to obtain the target oral bioavailability score of the candidate drug.

[0093] ADME prediction refers to the preliminary assessment of four key pharmacokinetic processes of a drug in the body: absorption, distribution, metabolism, and excretion, to determine its behavior and potential effects in the human body. This prediction is usually carried out in the early stages of drug development to screen candidate compounds with favorable pharmacokinetic properties and reduce the risk of failure in later clinical trials.

[0094] In the oral drug bioavailability prediction method provided in this embodiment of the invention, the ADME prediction result is first obtained based on the expert processing characteristics and the ADME prediction model. Then, the corresponding oral bioavailability score is predicted based on the ADME prediction result. That is, the oral bioavailability score obtained in this embodiment of the invention can be decomposed into predictions of biological links such as absorption and metabolism. The output includes not only the final score, but also the contribution analysis of each link, which helps researchers understand the causes of the oral bioavailability characteristics of compounds (candidate drugs), has biological interpretability, and facilitates the optimization of drug structures.

[0095] The multi-feature fusion oral bioavailability prediction method provided in this invention integrates four dimensions of features from candidate drug molecules: graph structure features, fingerprint features, physicochemical descriptors, and pharmacophore features. This comprehensive approach comprehensively characterizes drug molecular properties, resulting in stronger characterization capabilities and improved prediction accuracy. Furthermore, by employing multiple expert processing models corresponding to different preset result types to process the fused features, differentiated processing strategies can be adopted for different chemical structure types (such as alkaloids, flavonoids, and terpenes). This makes the prediction method provided in this invention particularly suitable for scenarios with high diversity in the structures of traditional Chinese medicine components.

[0096] In addition, the embodiments of the present invention decompose the bioavailability of oral drugs into biological processes such as absorption and metabolism, and predict them separately. The output includes not only the final score, but also the contribution analysis of each process, which helps researchers understand the causes of the bioavailability characteristics of oral drugs and has biological interpretability, which is beneficial to guiding structural optimization.

[0097] This invention provides multiple expert processing models that perceive the structure type of multiple preset molecular objects. These multiple expert processing models constitute a Mixture of Experts (MoE) model, using different expert networks for different chemical structural frameworks.

[0098] First, molecules are divided into several structural categories (such as alkaloids, flavonoids, terpenes, etc.) through unsupervised clustering. The expression for the clustering is as follows:

[0099] ;

[0100] The identifier representing the clustering result indicates the cluster category number to which each molecule is assigned. Each molecule is assigned to a specific cluster based on its structural characteristics, corresponding to a unique category ID. The name of the clustering algorithm, i.e. Means clustering algorithm. This is an unsupervised learning method that iteratively assigns data points to clustering algorithms. Within each cluster, the data points are made as similar as possible, while the data points between different clusters are significantly different. A graph representation of a molecule is typically a high-dimensional feature vector or graph structure. In cheminformatics, this can be the molecular topology, chemical fingerprint, molecular descriptor, etc., used to characterize the structural features of a molecule. This represents the desired number of clusters, i.e., the predefined number of structural categories. For example, if we want to classify molecules into structural categories such as alkaloids, flavonoids, and terpenes, then... It should be set to the number of categories (e.g., 3 or more).

[0101] Then, a dedicated expert processing model is trained for each type of structure. The corresponding expert processing model expression is:

[0102] ;

[0103] Indicates the first The output of an expert network, Indicates the expert network's ID. This represents the input feature vector. Each expert network specializes in processing a specific type of molecular structure. Indicates the first One feedforward neural network ( The output of ) is used to process the input features. . Typically composed of multiple fully connected layers, it is the basic building block of neural networks.

[0104] Finally, experts are dynamically selected through a gating network. The function expression for this dynamic selection is:

[0105] ;

[0106] ;

[0107] Let represent the gating weight vector, and let represent the activation weight of each expert network. All weights sum to 1. This represents the weight matrix of the gating network, used to learn how to assign expert weights based on input features. The molecular graph feature vector is used as the input to the gating network. The final output of the expert processing model is the fusion result, which is the weighted sum of the expert outputs, i.e., the expert processing features. Indicates the first The gating weight of the nth expert represents the gating weight of the nth expert. The contribution of each expert to the final output. Indicates the first A network of experts on fusion characteristics The output of . The fused feature vector typically contains feature information from different sources. This embodiment dynamically calculates the weights of each expert network using a gating network, and then performs a weighted sum of the expert network outputs to automatically select the most suitable expert processing model for prediction based on the input features.

[0108] In some embodiments, the gated output is soft-assigned weights, enabling the model to handle molecules with ambiguous boundaries (such as compounds with both flavonoid and glycoside structures). To avoid expert collapse (where all samples are processed by a few experts), a load balancing loss is introduced.

[0109] ;

[0110] This represents the load balancing loss value, used to penalize uneven distribution of expert loads and ensure that all experts can be effectively utilized. The coefficient of variation measures the dispersion of the load distribution. The smaller the coefficient of variation, the more evenly the load is distributed among the experts. For the first Number of samples processed by each expert .

[0111] In some embodiments, the multi-feature fusion oral bioavailability prediction method provided by the present invention further includes: outputting the oral bioavailability score, and analyzing the contribution of each sub-step parameter in the ADME prediction result to the oral bioavailability score. The sub-step parameters include at least one of the following: absorption fraction, hepatic clearance, apparent volume of distribution, and free fraction of the candidate drug after oral administration.

[0112] In this embodiment, not only is the predicted oral bioavailability score output so that users can more intuitively understand the prediction results, but the prediction results of the ADME sub-parameters that contributed to the prediction results are also output, thereby making the output oral bioavailability score more biologically interpretable and improving the reliability of the prediction results.

[0113] In some embodiments, the multi-dimensional feature extraction of the molecular object includes extracting structural graph features from the molecular object. A flowchart illustrating the method for extracting structural graph features from the molecular object is shown below. Figure 2 As shown, it includes S0311, S0312 and S0313.

[0114] S0311: Using the atoms of the molecular object as nodes and chemical bonds as edges, represent the molecular object as a graph structure, and determine the initialization features of each node and edge in the graph structure.

[0115] S0312: The graph structure is encoded by a message passing network that controls message flow through a multi-layer gated loop unit; wherein the encoding process includes: updating the hidden state of each node by aggregating the information of the neighboring nodes and connected edges of each node in the graph structure at each network layer of the message passing network.

[0116] S0313: Aggregate the final hidden states of each node after encoding through multiple network layers to obtain the structure graph features.

[0117] The structural graph feature extraction method provided in this application treats molecules as graphs rather than simple sequences or vectors, and utilizes the message passing mechanism of graph neural networks for encoding. In the field of computational chemistry for drug development, this represents a significant paradigm shift from artificial feature engineering to data-driven representation learning. By preserving complete structural information, enabling end-to-end learning, providing good generalization ability, and opening up interpretability pathways, it provides a powerful and advanced technical foundation for building more accurate and reliable molecular property prediction models (especially ADME prediction).

[0118] In some embodiments, the structural graph features of the molecule are extracted using a modified message-passing neural network (MPNN). The graph structure expression of the molecule is: Where V is the set of atomic nodes and E is the set of chemical bond edges, the initial atomic features include atomic type, hybridization type, formal charge, aromaticity, etc., and the initial bond features include bond type, conjugation, ring properties, etc. The message passing formula of the message passing neural network (MPNN) is:

[0119] ;

[0120] in, Represents a node The aggregated message received at layer t, Represents a node The set of all neighboring nodes, Indicates the first The message functions of a layer define how messages are generated between nodes. Represents a node In the previous layer ( The hidden state of the layer. Representing neighboring nodes In the hidden state of the previous layer, Indicates the connection node and The characteristics of the edges (chemical bonds).

[0121] The message function M_t of the Message Passing Neural Network (MPNN) employs a side-aware gating design, and its corresponding expression is:

[0122] ;

[0123] in, It is the sigmoid activation function. , and Each can learn a matrix. This indicates element-wise multiplication. This represents the concatenation of node state and edge features. This indicates that the gating signal is calculated based on edge characteristics. Using the method provided in this embodiment of the invention enables message passing to adaptively adjust according to the chemical bond type.

[0124] Furthermore, in this embodiment, node updates adopt a GRU structure, and the corresponding expression is:

[0125] ;

[0126] in, This represents the updated hidden state of node v at level t. This represents a gated loop unit used to process sequence information.

[0127] Based on the transformation expressions above, the final molecular graph result can be obtained through the Set2Set readout mechanism, and the resulting molecular graph structure expression is:

[0128] ;

[0129] in, This is the final diagrammatic representation of the entire molecule. For network layers, This is a sequence-to-sequence readout mechanism that aggregates the final states of all atomic nodes to generate a fixed-dimensional molecular representation.

[0130] This embodiment is designed to enable the model to effectively capture global structural information of molecules while maintaining sensitivity to local chemical environments, providing high-quality molecular characterization for tasks such as ADME property prediction.

[0131] In some embodiments, the multi-dimensional feature extraction of the molecular object includes extracting fingerprint features from the molecular object. A flowchart illustrating the method for extracting fingerprint features from the molecular object is shown below. Figure 3 As shown, it includes S0321 and S0322.

[0132] S0321: Calculate the extended connection fingerprint and key fingerprint of the molecular object.

[0133] S0322: The extended connection fingerprint and the key fingerprint are concatenated to obtain a concatenated fingerprint, and the concatenated fingerprint is processed by linear transformation and GELU activation function to obtain the fingerprint features.

[0134] Molecular fingerprint features utilize a combination of extended connected fingerprint (ECFP) and functional group fingerprint (MACCS). The high-dimensional sparse fingerprint is mapped to a low-dimensional dense space via a multilayer perceptron.

[0135] ;

[0136] The fingerprint feature vector is a dense molecular feature representation after nonlinear transformation. Indicates learnable weights, This represents the learnable bias vector.

[0137] This embodiment combines traditional fingerprinting technology with modern deep learning activation functions, maintaining the interpretability of chemical information while achieving powerful nonlinear modeling capabilities, which helps improve prediction accuracy.

[0138] In some embodiments, the multi-dimensional feature extraction of the molecular object includes extracting physicochemical descriptors from the molecular object. A flowchart illustrating the method for extracting physicochemical descriptors from the molecular object is shown below. Figure 4 As shown, it includes S0331, S0332 and S0333.

[0139] S0331: Calculate multiple primitive physicochemical descriptors characterizing the physicochemical properties of the said molecular object.

[0140] S0332: The weights representing the importance of each of the original physicochemical descriptors are obtained through attention mechanism learning.

[0141] S0333: Based on the weights corresponding to each of the original physicochemical descriptors, the original physicochemical descriptors are weighted and summed to obtain the physicochemical descriptors in the multi-dimensional features.

[0142] In some embodiments, the physicochemical descriptor features include traditional medicinal chemistry indicators such as molecular weight, LogP, topological polar surface area (TPSA), and number of rotatable bonds. The extraction of these features employs an attention-based adaptive selection mechanism:

[0143] ;

[0144] ;

[0145] in, Each descriptor can be independently transformed and then weighted and fused. This represents the attention weight vector, which represents the importance score for each input descriptor (or feature). For activation function, This is the original descriptor (feature) matrix, which contains multiple physicochemical or structural descriptors. This represents the learnable weight matrix. This represents the feature vector of the final physicochemical descriptor after weighted fusion. Indicates index Iterate through all descriptors and perform a summation operation. Indicates the first The importance weights of each descriptor are usually calculated using an attention mechanism. This is a function that performs a nonlinear transformation on the original descriptor. Indicates the first A primitive physicochemical descriptor, that is, a physicochemical property index calculated from the molecular structure.

[0146] This embodiment of the physicochemical descriptor extraction method based on the attention mechanism achieves an important transformation from "computing descriptors" to "learning descriptors" through intelligent weight learning and weighted fusion. It not only improves the quality and task relevance of feature representation, but also provides a more flexible, accurate and interpretable feature engineering scheme for molecular property prediction in drug development.

[0147] In some embodiments, the multi-dimensional feature extraction of the molecular object includes extracting pharmacophore features from the molecular object. A flowchart illustrating the method for extracting pharmacophore features from the molecular object is shown below. Figure 5 As shown, it includes S0341 and S0342.

[0148] S0341: Identify key features characterizing the pharmacophore from the molecular object.

[0149] S0342: Map the key features to three-dimensional space, and then pass through three-dimensional... The network extracts the spatial distribution and interaction patterns of the key features in the three-dimensional space to generate pharmacophore features in the multi-dimensional feature extraction.

[0150] Pharmacophore features are obtained by detecting pharmacophore patterns (hydrogen bond donors / acceptors, hydrophobic centers, aromatic rings, etc.) in the molecule. In some embodiments, the spatial arrangement of pharmacophores can be encoded as a three-dimensional feature vector, the expression of which is as follows:

[0151] ;

[0152] in, This represents the final generated three-dimensional feature vector of the pharmacophore. Represents a three-dimensional convolutional neural network ( ), Represents a pharmacodynamic map ( ).

[0153] This embodiment identifies key pharmacophore features in molecules and maps them into three-dimensional space. It then uses a three-dimensional convolutional neural network to extract spatial distribution and interaction patterns, thereby obtaining a more accurate and bioactively relevant molecular feature representation. This effectively captures three-dimensional interaction information between molecules and targets, improving the accuracy of drug design and property prediction.

[0154] Please see Figure 6 As shown, it is a flowchart of a method for fusing features in a multi-dimensional feature according to some embodiments of the present invention. In some embodiments, S05 includes S051 and S052.

[0155] S051: The features in the multi-dimensional features are concatenated to obtain the concatenated features.

[0156] S052: Calculate the fusion weight of each feature in the splicing feature through multiple different gating mechanisms, and weight each feature in the multi-dimensional feature according to the corresponding fusion weight to obtain the fused feature.

[0157] This embodiment of the multi-feature fusion method based on gated weights achieves effective integration and optimized utilization of multi-dimensional features through differentiated dynamic weight allocation. It not only retains the complete information of each feature, but also adaptively adjusts the importance of features according to specific tasks. While improving prediction performance, it enhances the interpretability of the model and provides a powerful feature fusion framework for molecular property prediction and drug design.

[0158] Please see Figure 7 As shown, it is a flowchart corresponding to the method for predicting ADME of the candidate drug according to some embodiments of the present invention. The steps for predicting ADME of the candidate drug provided in some embodiments of the present invention include S091 and S092.

[0159] S091: Using the corresponding prediction model and based on the expert processing features, predict the passive diffusion probability of the candidate drug after oral administration. The probability of glycoprotein efflux and the probability of uptake transporter-mediated absorption.

[0160] S092: Based on the passive diffusion probability, the... The efflux probability of glycoproteins and the uptake transporter-mediated absorption probability are used to determine the absorption fraction used to assess the extent to which the candidate drug is absorbed by the subject after oral administration.

[0161] The sub-parameters in the ADME prediction results include the absorption fraction, the passive diffusion probability, and the... The relationship between the glycoprotein efflux probability and the uptake transporter-mediated absorption probability satisfies:

[0162] ;

[0163] Among them, the The value representing the absorption fraction, The value representing the passive diffusion probability, the Characterizing the The value of the glycoprotein efflux probability, the The value characterizing the probability of absorption mediated by the uptake transporter.

[0164] In this embodiment, ADME prediction includes intestinal absorption rate prediction, and the ADME prediction module includes an absorption submodule, which is used to predict intestinal absorption rate. Intestinal absorption prediction mainly considers two mechanisms: passive diffusion and active transport, so as to accurately predict intestinal absorption rate.

[0165] Furthermore, in some embodiments, ADME prediction also includes the distribution volume of the candidate drug after oral administration and its plasma protein binding rate with the target drug, i.e., the distribution prediction of the candidate drug. Therefore, the ADME prediction module further includes a distribution submodule to predict the distribution volume and plasma protein binding rate. The prediction expressions corresponding to the candidate drug distribution volume and plasma binding rate are as follows:

[0166] ;

[0167] ;

[0168] in, The apparent volume of drug distribution is the volume of distribution of the drug within the body. This represents the weight parameter vector used to calculate the distribution volume. This represents the comprehensive feature representation after multi-feature fusion processing, indicating the free fraction of the candidate drug in plasma, i.e., the proportion of drug not bound to plasma proteins. This represents the weight parameter vector used to calculate plasma binding rate.

[0169] Please see Figure 8 As shown, it is a flowchart corresponding to the method for predicting ADME of the candidate drug according to some embodiments of the present invention. The steps for predicting ADME of the candidate drug provided in some embodiments of the present invention further include S093 and S094.

[0170] S093: For each of a predefined set of cytochrome enzymes associated with the metabolism of the candidate drug after oral administration, predict the substrate probability of the candidate drug as its substrate and the metabolic clearance rate mediated by it.

[0171] S094: Using the substrate probability corresponding to each cytochrome enzyme in the preset set of cytochrome enzymes as a weighting factor, the metabolic clearance rates corresponding to each cytochrome enzyme are weighted and summed to obtain the total hepatic clearance rate, which characterizes the extent to which the candidate drug is metabolized by the user. The sub-parameter in the ADME prediction result includes the total hepatic clearance rate.

[0172] In this embodiment, ADME prediction also includes metabolic prediction, therefore the ADME prediction module further includes a metabolic submodule. The metabolic submodule is used to predict first-pass metabolic clearance rate by integrating substrates of cytochrome P450 enzymes (CYP450). The expression corresponding to this prediction is:

[0173] ;

[0174] This represents the rate constant for the clearance of the candidate drug through hepatic metabolism. This indicates a summation calculation for all CYP450 isoenzymes. Candidate drugs as the first Substrate probability of CYP450 enzyme substrates Indicates the first The maximum reaction rate catalyzed by CYP450 enzymes. Indicates the first Michaelis constants of CYP450 enzymes. This indicates the drug concentration in the liver. Furthermore, ADEM prediction includes excretion prediction, which primarily includes renal and bile clearance prediction. The ADEM prediction module also includes an excretion submodule for predicting renal and bile clearance.

[0175] This embodiment simultaneously considers the metabolic activities of a pre-defined group of cytochrome enzymes, avoiding the limitations of prediction based on a single enzyme system. It provides quantitative predictions of substrate probability and metabolic clearance rate, enhancing the reliability of the prediction results. The metabolic clearance rates of different enzyme systems are weighted based on substrate probability, reflecting the actual contribution of each enzyme system. The total hepatic clearance rate is obtained through weighted summation, more accurately reflecting the overall degree of drug metabolism. Therefore, this embodiment, by integrating metabolic information from multiple enzyme systems, achieves a comprehensive assessment of the hepatic metabolic behavior of candidate drugs, providing a scientifically reliable computational framework for predicting ADME properties in drug development.

[0176] Please see Figure 9As shown, it is a schematic diagram of the method flow for predicting the oral bioavailability score of the candidate drug based on the ADME prediction result according to some embodiments of the present invention. In some embodiments, the step of predicting the oral bioavailability score based on the ADME prediction result includes S095, S096 and S097.

[0177] S095: Calculate the liver extraction rate based on the total liver clearance rate and the preset liver blood flow, and calculate the liver availability based on the liver extraction rate.

[0178] S096: Based on the liver availability and the absorption fraction, obtain the initial oral bioavailability score of the candidate drug.

[0179] The relationship between the liver availability, the absorption fraction, and the initial oral bioavailability score is as follows: The Characterizing the initial oral bioavailability score, the Characterizing the absorption fraction, the The liver availability is characterized by the regulating factor, which is an adjustable parameter of the oral bioavailability prediction model.

[0180] S097: The initial oral bioavailability score is trimmed to obtain... Target oral bioavailability score within the interval.

[0181] The oral bioavailability (ADME) scoring method provided in this embodiment offers a comprehensive and interpretable predictive framework by integrating drug metabolic clearance, hepatic extraction, and absorption characteristics. Its core advantages lie in its multi-factor comprehensive assessment, strong physiological relevance, good interpretability, and guiding nature. It can comprehensively evaluate the oral absorption behavior of drugs from multiple dimensions, such as hepatic metabolic clearance rate, hepatic availability, and intestinal absorption fraction, providing a scientific and systematic computational framework for predicting ADME properties in drug development, thus helping to improve drug development efficiency and success rate.

[0182] In some embodiments, the multi-feature fusion oral bioavailability prediction method further includes training the multi-feature fusion oral bioavailability prediction model using a two-stage transfer learning strategy; wherein the two-stage transfer learning strategy includes a first-stage training and a second-stage training. The first-stage training includes: pre-training the multi-feature fusion oral bioavailability prediction model on a general molecular dataset with a dataset size that meets preset conditions, wherein the pre-training objective includes a data-augmented contrastive learning loss. And the molecular property multi-task prediction loss. The second stage of training includes: training the multi-feature fusion oral bioavailability prediction model on the target oral bioavailability dataset using a progressive parameter unfreezing strategy to adjust the model parameters. The adjustment range of the model parameters in the second stage of training is smaller than that in the first stage of training.

[0183] This invention employs a two-stage training strategy to address the scarcity of OB data. The first stage involves pre-training on a large-scale general molecular dataset. Using a contrastive learning framework, data augmentation (atom masking, edge deletion, subgraph sampling) is performed on the molecules, and the augmentation-invariant molecular representation is learned as follows:

[0184] ;

[0185] in, and This represents two enhancements to the same molecule. For the representation of other molecules, This refers to the temperature parameter.

[0186] To compare the learning loss value, which is used to measure the quality of molecular representation learning, the goal is to minimize this loss value. It is the representation vector of the original molecule, which is the feature representation extracted from the original molecular structure (atoms, chemical bonds, etc.) by a molecular encoder. The enhanced representation vector of the original molecule is the enhanced version of the molecule obtained by transforming the original molecule through data augmentation techniques (such as atom masking, edge deletion, and subgraph sampling). The vector represents the molecules sampled from the negative samples, which are compared with the original molecules. The cosine similarity function is used to calculate the similarity between two vectors, and its range is within the range of... between. The temperature parameter controls the distribution range of similarity scores in contrastive learning. It is usually a positive scalar value that affects the smoothness of the gradient. This is used to sum over all negative sample representation vectors and to calculate the normalization term in the denominator.

[0187] In some embodiments, the above-mentioned training pre-tasks include OB-related auxiliary tasks such as LogP prediction, solubility prediction, and ADMET classification. The specific training method can be to simultaneously perform multi-task pre-training on molecular properties, and the corresponding training expression is:

[0188] ;

[0189] This represents the total pre-training loss, which is a weighted combination of the contrastive learning loss and the losses for each task, used to guide the joint training of the model. To contrast the learning loss, as previously described, a reinforcement-invariant molecular representation is used. Summing over all predefined tasks indicates simultaneous optimization of multiple prediction tasks. For the first The loss weighting coefficient for each task is used to adjust the contribution of different tasks to the total loss. For the first The loss function value for each molecular property prediction task is used, such as the mean squared error in regression or the cross-entropy loss in classification. This training expression, through joint optimization of contrastive learning and multi-task prediction loss, enables the model to learn a general molecular representation while also optimizing for specific property prediction tasks, thereby improving the model's generalization ability and prediction accuracy.

[0190] In some embodiments, the second stage of training further includes: predicting the oral bioavailability of unlabeled samples using the current multi-feature fusion oral bioavailability prediction model; estimating the prediction uncertainty based on Monte Carlo estimation; and selecting high-confidence prediction results as pseudo-labels to be added to the training set for iterative optimization.

[0191] In some embodiments, the second stage involves fine-tuning on the OB dataset of traditional Chinese medicine components. A progressive unfreezing strategy is adopted, which involves: initially updating only the prediction head parameters, unfreezing the shallow feature extraction layer in the middle stage, and unfreezing all parameters but using a smaller learning rate in the later stage. To further alleviate data shortage, this embodiment also introduces uncertainty-aware self-training. Self-training mainly includes: using the current model to predict OB for unlabeled traditional Chinese medicine components, selecting samples with high prediction confidence as pseudo-labels, and adding the pseudo-label samples to the training set for iterative optimization. Samples with high confidence (low uncertainty) are more likely to have accurate pseudo-labels. During training, the confidence is estimated by MC Dropout, and the corresponding confidence estimation expression is:

[0192] ;

[0193] in, The model represents the input The larger the value, the more uncertain the model's prediction of that input. This represents the variance calculated using the dropout mechanism. Dropout is a regularization technique that randomly "drops out" some neurons during training (setting their output to zero) to prevent overfitting. During the prediction phase, by enabling dropout multiple times and sampling different subsets of neurons, multiple predictions can be obtained from the model's output, and the variance of these predictions can then be calculated. The parameter is The neural network model for input The predicted output. These are the weight parameters of the model.

[0194] By running the model with dropout multiple times (each dropout randomly discards different neurons), a series of predicted values ​​are obtained. Then, the variance of these predictions is calculated. The larger the variance, the greater the difference in the model's predictions under different dropout stochastic states, meaning the higher the uncertainty of the model regarding that input; conversely, the smaller the variance, the more stable the prediction and the lower the uncertainty.

[0195] Figure 10 This is a schematic diagram of the system architecture corresponding to the multidimensional feature fusion oral bioavailability prediction method provided in this embodiment of the invention. The system architecture includes a molecular SMILES module for acquiring molecular SMILES, various feature extraction modules connected to the molecular SMILES module, a gate multidimensional feature fusion module connected to each feature extraction module, a hybrid network module (corresponding to multiple expert processing models) connecting the structural graph feature module and the gate multidimensional feature fusion module among the multiple feature extraction modules, and a module including ADEM sub-stage prediction and OB scoring. Among these, the multiple feature extraction modules include, in addition to the structural graph feature extraction module, a molecular fingerprint extraction module, a physicochemical description adaptive selection module, and a pharmacophore feature module. ADEM sub-stage prediction includes absorption prediction, metabolism and excretion prediction, as well as distribution prediction, which includes the distribution volume of the candidate drug and protein binding.

[0196] Taking quercetin as an example, the radar contribution map of its corresponding ADEM prediction is as follows: Figure 11 As shown in the figure, the predicted size of the OB and the contribution of each sub-link can be seen intuitively, making the OB prediction more biologically interpretable.

[0197] Taking flavonoids as an example of candidate drugs, the expert network gating distribution diagram corresponding to the prediction of their oral bioavailability using the method of this invention is shown in the figure below. Figure 12 As shown in the diagram, the structural types of flavonoids, as highly orally bioavailable drugs, are distributed as follows: Figure 13 As shown, the candidate drugs correspond to multiple different structural types, namely flavonoids, alkaloids, terpenes, phenolic acids, saponins, and others. Multiple expert networks of different types include alkaloid experts, flavonoid experts, terpene experts, saponin experts, and other experts. The opinions of each type of expert have different gating weights. Based on these weights, the opinions of each type of expert are weighted, resulting in high accuracy of the OB prediction results provided in this application embodiment.

[0198] The multi-feature fusion oral bioavailability prediction method provided in this invention can achieve at least the following technical effects:

[0199] First, by comprehensively characterizing the molecular properties of candidate drugs from four dimensions—graph structure features, molecular fingerprint features, physicochemical descriptors, and pharmacophore features—and adaptively integrating them through a gating fusion mechanism, this method has stronger characterization capabilities compared to methods based on a single feature source.

[0200] Second, the hybrid expert architecture enables differentiated processing strategies for different chemical structure types (such as alkaloids, flavonoids, and terpenes), and is adaptive to structural types, making it particularly suitable for scenarios with high diversity of Chinese medicine component structures.

[0201] Third, biological interpretability. This invention breaks down OB into biological processes such as absorption and metabolism for prediction. The output includes not only the final score but also the contribution analysis of each process, which helps researchers understand the causes of the compound's OB properties and guides structural optimization.

[0202] Fourth, this invention effectively alleviates the problem of scarce OB data for traditional Chinese medicine through strategies such as large-scale pre-training, progressive fine-tuning, and uncertainty-aware self-training. It can still obtain reliable predictions in small sample scenarios and has small sample adaptability.

[0203] Fifth, each module in the prediction model of this invention is a differentiable design, supporting end-to-end training and optimization, and has a stronger learning ability compared with traditional methods based on handcrafted features.

[0204] To more clearly illustrate the technical effects of the multi-feature fusion oral bioavailability prediction method provided in this embodiment, the following description, in conjunction with comparative experiments and specific drug examples, explains the multi-feature fusion oral bioavailability prediction method provided by this invention.

[0205] The training dataset for the prediction model used to implement the multi-feature fusion oral bioavailability prediction method provided in this embodiment of the invention comes from 1 million molecules in the ZINC database, and the dataset for fine-tuning training comes from 2,042 compounds with OB experimental values ​​in the TCMSP database. The OB data distribution of the dataset is as follows: OB is less than... The occupation 0B in and Between OB is greater than or equal to The occupation The dataset is divided into... The model is divided into training, validation, and test sets to train, validate, and test the prediction model, resulting in a multi-feature fusion oral bioavailability prediction model.

[0206] Figure 14 The diagram illustrates a comparison of experimental results between this method (a multi-feature fusion method for predicting oral bioavailability) and various baseline methods based on a regression task. Clearly, this invention achieves optimal performance across all evaluation metrics. Improvement compared to the best baseline method .

[0207] Figure 15 The diagram illustrates the performance of this method (a multi-feature fusion oral bioavailability prediction method) in a binary classification task compared to various baseline methods. The diagram shows a comparison of experimental results for high bioavailability. Clearly, the accuracy and F1 score of the method of this invention are significantly higher than those of other methods.

[0208] Figure 16 This diagram illustrates a comparison of the ablation test results of the multi-feature fusion oral bioavailability prediction provided by the present invention. Through this diagram, the contribution of each module of the prediction model can be analyzed. By comparison, it can be demonstrated that multi-dimensional feature fusion and pre-training contribute the most to the prediction performance.

[0209] Figure 17 The diagram illustrates the performance comparison of fine-tuning with different pre-training configurations. Through comparison, it is shown that the combination of contrastive learning and multi-task pre-training achieves the best results.

[0210] In summary, it can be demonstrated that the multidimensional feature fusion oral bioavailability prediction method provided in the embodiments of the present invention can achieve better prediction results. When applied to screen active ingredients of traditional Chinese medicine with high oral bioavailability, OB prediction values ​​were screened from 26,348 ingredients. The components. Screening results: 3,421 components ( High OB component coverage validated by literature: Newly discovered potential high-OB components: approximately 1,200.

[0211] The above experiments demonstrate that the multi-dimensional feature fusion oral bioavailability prediction method proposed in this invention exhibits significant advantages in prediction accuracy, interpretability, and practicality.

[0212] Please see Figure 18The diagram shown is a structural schematic of a multi-feature fusion oral bioavailability prediction device 180 provided according to some embodiments of the present invention. The multi-feature fusion oral bioavailability prediction device 180 is used to predict multi-feature fusion oral bioavailability using a multi-feature fusion oral bioavailability prediction model. It includes an acquisition module 181, a multi-dimensional feature extraction module 182, a feature fusion module 183, an expert processing module 184, and a prediction module 185.

[0213] The acquisition module 181 is used to acquire the molecular SMILES string corresponding to the candidate drug. The multidimensional feature extraction module 182 is used to parse the molecular SMILES string into a structured molecular object, and to extract multidimensional features from the molecular object to obtain multidimensional features. The multidimensional features include graph structure features, fingerprint features, physicochemical descriptors, and pharmacophore features. The feature fusion module 183 is used to fuse the features in the multidimensional features to obtain fused features. The expert processing module 184 is used to process the fused features using multiple expert processing models corresponding to multiple different preset structure types, and to determine the probability that the molecular object belongs to multiple different preset structure types based on the graph structure features. Furthermore, the output features of each expert processing model are weighted according to the probability as a corresponding weight to obtain expert processing features. The prediction module 185 is used to perform ADME prediction on the candidate drug based on the expert processing features, and to obtain the target oral bioavailability score of the candidate drug based on the ADME prediction result.

[0214] Please continue reading. Figure 18 As shown, in some embodiments, the multi-feature fusion oral bioavailability prediction device 180 further includes an output module 186. The output module 186 is used to output the oral bioavailability score, and an analysis of the contribution of each sub-parameter in the ADME prediction result to the oral bioavailability score. The sub-parameter includes at least one of the following: absorption fraction, hepatic clearance, apparent volume of distribution, and free fraction of the candidate drug after oral administration.

[0215] The multi-feature fusion oral bioavailability prediction device 180 provided in this embodiment of the invention is specifically used to implement the corresponding steps in the multi-feature fusion oral bioavailability prediction method provided in any embodiment of the invention, which will not be repeated here.

[0216] The multi-feature fusion oral bioavailability prediction device provided in this embodiment of the invention and the multi-feature fusion oral bioavailability prediction method provided in this embodiment of the invention can achieve essentially the same technical effect, and will not be described again here.

[0217] Please see Figure 19The diagram shows a schematic of a multi-feature fusion oral bioavailability prediction device 190 provided according to some embodiments of the present invention. The multi-feature fusion oral bioavailability prediction device 190 includes a processor 191 and a memory 192. The memory 192 stores programs, instructions, or code, and the processor 191 executes the programs, instructions, or code in the memory to perform the multi-feature fusion oral bioavailability prediction method as described in the foregoing embodiments. The memory 192 can be random access memory (RAM), flash memory, read-only memory (ROM), EPROM, non-volatile read-only memory (EEPROM), registers, hard disk, removable disk, etc.

[0218] The multi-feature fusion oral bioavailability prediction device 190 provided according to the embodiments of the present invention can achieve essentially the same technical effect as the multi-feature fusion oral bioavailability prediction method provided according to the embodiments of the present invention, and will not be described in detail here.

[0219] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. A computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the flow or function according to the embodiments of the present invention is generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape) or a semiconductor medium (e.g., solid-state disk (SSD)).

[0220] This invention also provides a readable storage medium for storing the multi-feature fusion oral bioavailability prediction method provided in the above embodiments. Examples include random access memory (RAM), flash memory, read-only memory (ROM), EPROM, non-volatile read-only memory (EEPROM), registers, hard disks, removable disks, or any other form of storage medium in the art.

[0221] It is understood that the term "connection" in the embodiments of the present invention can be interpreted as "electrical connection," "communication connection," etc., if the connected circuits, units, etc. can transmit electrical signals or data to each other.

[0222] It is understood that the specific examples in this document are only intended to help those skilled in the art better understand the embodiments of the present invention, and are not intended to limit the scope of the present invention.

[0223] It is understood that in the various embodiments of the present invention, the sequence number of each process does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present invention.

[0224] It is understood that the various embodiments described in this invention can be implemented individually or in combination, and the embodiments of this invention are not limited in this respect.

[0225] Unless otherwise stated, all technical and scientific terms used in the embodiments of this invention have the same meaning as commonly understood by one of ordinary skill in the art. The terminology used in this invention is for the purpose of describing particular embodiments only and is not intended to limit the scope of the invention. The term "and / or" as used in this invention includes any and all combinations of one or more of the associated listed items. The singular forms "a," "the," and "the" as used in the embodiments of this invention and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise.

[0226] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this invention.

[0227] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A multi-fingerprint fusion oral bioavailability prediction method, characterized in that, An application to a multi-feature fusion oral bioavailability prediction model, the multi-feature fusion oral bioavailability prediction method includes: Obtain the molecular SMILES string corresponding to the candidate drug; the candidate drug is a traditional Chinese medicine. The SMILES string is parsed into a structured molecular object, and multi-dimensional features are extracted from the molecular object to obtain multi-dimensional features; wherein, the multi-dimensional features include graph structure features, fingerprint features, physicochemical descriptors and pharmacophore features; The features in the multi-dimensional features are fused to obtain the fused features; Multiple expert processing models corresponding to multiple different preset structure types are used to process the fused features respectively. The probability that the molecular object belongs to multiple different preset structure types is determined according to the graph structure features. The output features of each expert processing model are weighted with the probability as the corresponding weight to obtain the expert processing features. The multiple different preset structure types include alkaloids, flavonoids and terpenes. Based on the expert processing characteristics, ADME prediction is performed on the candidate drug, and the ADME prediction result is obtained based on the ADME prediction to obtain the target oral bioavailability score of the candidate drug; Output the oral bioavailability score, and the contribution analysis of each sub-parameter in the ADME prediction result to the oral bioavailability score; wherein, the sub-parameter includes at least one of the following: absorption fraction, hepatic clearance, apparent volume of distribution, and free fraction of the candidate drug after oral administration; The ADME prediction results include an absorption fraction used to assess the extent to which the candidate drug is absorbed by the subject after oral administration, and a total hepatic clearance rate used to characterize the extent to which the candidate drug is metabolized by the subject. The step of predicting the oral bioavailability of the candidate drug based on the ADME prediction results to obtain an oral bioavailability score for the candidate drug includes: The liver extraction rate is calculated based on the total liver clearance rate and the preset liver blood flow rate, and the liver availability is calculated based on the liver extraction rate. According to the liver availability and the absorption fraction, an initial oral bioavailability score of the candidate drug is obtained; wherein the relationship among the liver availability, the absorption fraction and the initial oral bioavailability score is: The The initial oral bioavailability score is characterized, the The absorption fraction is characterized, the The liver availability is characterized, and the adjustment factor is an adjustable parameter of the oral bioavailability prediction model; The initial oral bioavailability score was cropped to obtain... Target oral bioavailability score within the interval.

2. The multi-feature fusion oral bioavailability prediction method according to claim 1, characterized in that, The ADME prediction of the candidate drug includes: Using the corresponding prediction model and based on the expert processing features, the passive diffusion probability of the candidate drug after oral administration is predicted. The probability of glycoprotein efflux and the probability of uptake transporter-mediated absorption; According to the passive diffusion probability, the The glycoprotein efflux probability and the uptake transporter-mediated absorption probability are used to determine the absorption fraction; wherein, the sub-parameters in the ADME prediction result include the absorption fraction, the absorption fraction, the passive diffusion probability, and the... The relationship between the glycoprotein efflux probability and the uptake transporter-mediated absorption probability satisfies: Among them, the The value representing the absorption fraction, The value representing the passive diffusion probability, the Characterizing the The value of the glycoprotein efflux probability, the The value characterizing the probability of absorption mediated by the uptake transporter.

3. The multi-feature fusion oral bioavailability prediction method according to claim 2, characterized in that, The ADME prediction of the candidate drug includes: For each cytochrome enzyme in a set of preset cytochrome enzymes associated with the metabolism of the candidate drug after oral administration, the substrate probability of the candidate drug as its substrate and the metabolic clearance rate mediated by it are predicted respectively; the metabolic clearance rates corresponding to each cytochrome enzyme in the set of preset cytochrome enzymes are weighted and summed to obtain the total liver clearance rate; wherein, the sub-step parameters in the ADME prediction result include the total liver clearance rate.

4. The multi-feature fusion oral bioavailability prediction method according to any one of claims 1 to 3, characterized in that, It also includes using a two-stage transfer learning strategy to train the multi-feature fusion oral bioavailability prediction model; wherein the two-stage transfer learning strategy includes a first-stage training and a second-stage training. The first stage of training includes: pre-training the multi-feature fusion oral bioavailability prediction model on a general molecular dataset whose dataset size meets preset conditions. The pre-training objective includes data-augmented contrastive learning loss. And molecular property multi-task prediction loss; the second stage of training includes: training the multi-feature fusion oral bioavailability prediction model on the target oral bioavailability dataset using a progressive parameter unfreezing strategy to adjust the model parameters, wherein the adjustment range of the model parameters in the second stage of training is smaller than the adjustment range of the model parameters in the first stage of training.

5. The multi-feature fusion oral bioavailability prediction method according to claim 4, characterized in that, The second phase of training also includes: Oral bioavailability prediction of unlabeled samples is performed using the current multi-feature fusion oral bioavailability prediction model. The prediction uncertainty is estimated based on Monte Carlo estimation; High-confidence prediction results are selected as pseudo-labels and added to the training set for iterative optimization.

6. A multi-feature fusion oral bioavailability prediction device, characterized in that, This device is used to predict the bioavailability of a multi-feature fusion oral bioavailability using a multi-feature fusion oral bioavailability prediction model. The multi-feature fusion oral bioavailability prediction device includes: an acquisition module for acquiring the molecular SMILES string corresponding to a candidate drug; the candidate drug is a traditional Chinese medicine. A multi-dimensional feature extraction module is used to parse the SMILES string into a structured molecular object and extract multi-dimensional features from the molecular object to obtain multi-dimensional features; wherein, the multi-dimensional features include graph structure features, fingerprint features, physicochemical descriptors and pharmacophore features; a feature fusion module is used to fuse the features in the multi-dimensional features to obtain fused features; An expert processing module is used to process the fused features using multiple expert processing models corresponding to multiple different preset structure types, and to determine the probability that the molecular object belongs to multiple different preset structure types based on the graph structure features. The module then uses the probabilities as weights to weight the output features of each expert processing model to obtain expert processed features. The multiple different preset structure types include alkaloids, flavonoids, and terpenes. The prediction module is used to perform ADME prediction on the candidate drug based on the expert processing features, and obtain the ADME prediction result based on the ADME prediction to obtain the target oral bioavailability score of the candidate drug. The output module outputs the oral bioavailability score and the contribution analysis of each sub-parameter in the ADME prediction result to the oral bioavailability score; wherein, the sub-parameter includes at least one of the following: absorption fraction, hepatic clearance, apparent volume of distribution, and free fraction of the candidate drug after oral administration; The ADME prediction results include an absorption fraction used to assess the extent to which the candidate drug is absorbed by the subject after oral administration, and a total hepatic clearance rate used to characterize the extent to which the candidate drug is metabolized by the subject. The step of predicting the oral bioavailability of the candidate drug based on the ADME prediction results to obtain an oral bioavailability score for the candidate drug includes: The liver extraction rate is calculated based on the total liver clearance rate and the preset liver blood flow rate, and the liver availability is calculated based on the liver extraction rate. The initial oral bioavailability score of the candidate drug is obtained based on the liver availability and the absorption fraction; wherein, the relationship between the liver availability, the absorption fraction, and the initial oral bioavailability score is as follows: The Characterizing the initial oral bioavailability score, the Characterizing the absorption fraction, the The liver availability is characterized by the regulating factor, which is an adjustable parameter of the oral bioavailability prediction model. The initial oral bioavailability score was cropped to obtain... Target oral bioavailability score within the interval.

7. A multi-feature fusion oral bioavailability prediction device, characterized in that, The method includes a processor and a memory, the memory being used to store programs, instructions, or code, and the processor being used to execute the programs, instructions, or code in the memory to perform the multi-feature fusion oral bioavailability prediction method as described in any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that, The system contains a computer program that is loaded by a processor to execute the multi-feature fusion oral bioavailability prediction method as described in any one of claims 1 to 5.

Images