Medicinal preparation attribute prediction method and system based on multi-task learning

Abstract

The invention discloses a pharmaceutical preparation attribute prediction method and system based on multi-task learning. The method comprises the following steps: acquiring first information and second information from pharmaceutical preparation data; according to the first information, constructing a hierarchical diagram based on the spatial characteristics of the drug atoms; constructing a hierarchical graph convolutional network based on the hierarchical graph; a hierarchical graph convolutional network is adopted, and medicine structure features are extracted through bottom-up and top-down bidirectional information aggregation; according to the second information, auxiliary material and solvent characteristics based on the types of the medicine auxiliary materials are constructed; the auxiliary material and the solvent are characterized by one-dimensional triple feature vectors comprising a surfactant, a latent solvent and an oil phase; constructing a medicament component matrix and a medicament component interaction matrix according to the structural characteristics of the medicament and the characteristics of auxiliary materials and solvents; the two are fused through a cross attention mechanism to obtain a medicament feature matrix; and drug attribute prediction is carried out based on multi-task learning. The method can accurately predict the drug attributes.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of drug formulation property prediction, and in particular to a drug formulation property prediction method and system based on multi-task learning. BACKGROUND

[0002] In the process of forming a drug formulation (hereinafter referred to as "drug") by a drug, a combination of excipients is often needed to change the physical, pharmacological and pharmacokinetic behavior of the drug, which in turn affects the quality, safety and stability of the drug, and the drug property is a quantitative description of these properties. Usually, researchers evaluate the effectiveness of the drug dosage form formula based on the drug property. Traditional formula research still strongly relies on trial and error experiments and personal experience, which leads to labor consumption, tediousness and high cost. At the same time, different properties often need to be based on different methods, which requires repeated experiments, which is time-consuming and laborious. For example, particle size and PDI (Polydispersity Index) need to be obtained by laser diffraction method, and drug release rate needs to be obtained by dissolution experiment. At present, although high-throughput screening technology can speed up the screening speed, it still cannot solve the problem of cost and point-to-point experimental method limiting the efficiency of research and development.

[0003] In the work of Vu et al., the authors took the ratio of the drug rosuvastatin and three different excipients as the input of the ANN model to predict the particle size and distribution of the self-emulsifying formulation. In the work of Dong et al., the researchers believed that the composition of the excipient is complex, with a mixture of small molecules and large molecules, etc. For small molecule excipients, a method similar to the drug was used to represent the structural information based on ECFP (Extended Connectivity Fingerprints) or the physical information of the molecule fingerprint. For large molecule mixtures, the physical characteristics or molecular structure of the monomer were used to represent them. After that, the drug-exipient pair was constructed, and the features of the two were spliced with the drug-exipient ratio and input into the machine learning model to predict the properties of the drug.

[0004] The prior art document with publication number CN117223061A provides a machine learning model for predicting the properties of a chemical formulation. The invention extracts information contained in the SMILES formula of the drug through a recurrent neural network (RNN) to obtain the structural information of the drug. Then, the structural information of the drug and the information of other components are input into the machine learning model to predict the characteristics of the formulation composed of multiple components.

[0005] However, the current scheme has the following problems: first, the introduction of two-dimensional molecular fingerprints as structural features ignores the influence of the three-dimensional chemical characteristics of the drug on its properties, thereby affecting the attribute information of the pharmaceutical. Second, the method of directly combining the drug and excipient into a drug-excipient pair to represent the pharmaceutical in the previous scheme has the following main problems: first, it fails to fully consider the influence of different excipient categories on the properties of the pharmaceutical; second, it does not assess the potential influence of the solvent on the excipient, and the choice of solvent not only affects the solubility of the excipient, but also can change its physical state, thereby affecting the dispersibility and stability of the drug; third, the method of simply combining the drug and excipient cannot capture the influence of the interaction between the components of the drug on the properties of the pharmaceutical, which is crucial for understanding the drug release mechanism and optimizing the pharmaceutical formulation. In addition, the current scheme predicts various properties separately without considering the inherent relationship between the properties of the pharmaceutical (such as drug particle size and PDI), for example, the substructure in the drug will simultaneously affect the particle size and PDI. Finally, for time-related features such as drug release rate, the current scheme only considers the influence of the properties of the pharmaceutical on the drug release rate, without considering the influence of the drug concentration in the pharmaceutical on the drug release rate.

[0006] Therefore, the present application is proposed. SUMMARY

[0007] The present application aims to provide a drug formulation attribute prediction method and system based on multi-task learning. First, spatial hierarchical graph convolution network is used to extract drug information, including spatial information and atomic attribute information of the drug, for drug representation. Second, for pharmaceutical ingredients, a pharmaceutical ingredient matrix is constructed by fusing ingredients of the same type to represent the influence of ingredients on the pharmaceutical. Then, a pharmaceutical interaction matrix is constructed to capture the influence of ingredient interactions on pharmaceutical properties. Finally, a multi-task learning framework is introduced to combine Neural ODE to learn the time series characteristics of drug release rate and the relationship between pharmaceutical characteristics, for accurate prediction of pharmaceutical properties.

[0008] The present application is implemented by the following technical solutions:

[0009] In a first aspect, the present application provides a drug formulation attribute prediction method based on multi-task learning, which comprises:

[0010] Obtaining first information and second information from drug formulation data; the first information is drug information, and the second information is excipient and solvent information;

[0011] According to the first information, a hierarchical graph based on the spatial characteristics of drug atoms is constructed; the hierarchical graph includes atom group layer, bond layer, atom layer, and the relationship between the atom group layer and the bond layer, the relationship between the bond layer and the atom layer;

[0012] Based on the hierarchical graph, a hierarchical graph convolutional network is constructed. Using the hierarchical graph convolutional network, drug structural features are extracted through bidirectional information aggregation from bottom to top and top to bottom. The drug structural features are the three-dimensional multi-scale structural features of drug molecules.

[0013] Based on the second information, excipient and solvent features are constructed based on the type of drug excipient; the excipient and solvent features are one-dimensional tripartite feature vectors containing surfactants, cosolvents, and oil phases.

[0014] Based on the structural characteristics of the drug and the characteristics of the excipients and solvents, a drug component matrix and a drug component interaction matrix are constructed; and the drug component matrix and the drug component interaction matrix are fused through a cross-attention mechanism to obtain the drug feature matrix;

[0015] Based on the drug feature matrix, drug attribute prediction is performed using multi-task learning.

[0016] Furthermore, the steps for constructing the atomic cluster layer include:

[0017] Aggregate atomic nodes and their adjacent nodes to construct atomic cluster nodes;

[0018] Based on the atomic cluster nodes, establish the connections, i.e., edges, between atomic cluster nodes;

[0019] Merge duplicate cluster nodes and remove redundant edges to obtain the cluster graph and use it as the cluster layer.

[0020] The steps for constructing the mapping between bond layers, atomic layers, and atomic group layers, and between bond layers and the mapping between bond layers and atomic layers, include:

[0021] Initialize the bond graph and simultaneously establish an empty mapping set; the bond graph is used to store bond nodes and their connection relationships; the mapping set is used to record the mapping relationship from atomic groups to bonds;

[0022] Traverse all atomic group nodes, extract atomic group bonds and create a set of bond nodes, and record the correspondence between the atomic group and the bond nodes in the mapping set;

[0023] Iterate through the set of bond nodes for each atomic group. If two bond nodes share an atom, add an edge to the bond graph.

[0024] After traversing all atomic groups, output the bond diagram and cross-layer mapping, and use the bond diagram as the bond layer; the nodes of the bond diagram contain the chemical bonds inside all atomic groups, and the edges represent the bond connections between shared atoms; the cross-layer mapping records the many-to-many associations from atomic groups to bonds;

[0025] And based on the bond layers, a mapping relationship between bond layers and atomic layers is established.

[0026] Furthermore, based on the hierarchical graph, a hierarchical graph convolutional network is constructed; using the hierarchical graph convolutional network, drug structural features are extracted through bidirectional information aggregation from bottom to top and top to bottom, including:

[0027] Based on hierarchical graphs, a hierarchical graph convolutional network is constructed using the MPNN framework;

[0028] A hierarchical graph convolutional network is adopted, which aggregates information from the bottom up according to the atomic layer, bond layer and atomic group layer;

[0029] A hierarchical graph convolutional network is adopted, which aggregates information from top to bottom in sequence according to the atomic group layer, bond layer and atomic layer;

[0030] By combining the above information, the structural features of the drug are extracted.

[0031] Furthermore, the drug characteristics in the preparation are obtained by extracting the three-dimensional multi-scale structural features and atomic information of the drug molecules;

[0032] The chemical characteristics of surfactants, co-solvents, and oil phases in the excipient and solvent profiles are initially embedded from the database.

[0033] Furthermore, based on the drug's structural characteristics and the characteristics of its excipients and solvents, a drug component matrix and a drug component interaction matrix are constructed, including:

[0034] Based on the structural characteristics of the drug and the characteristics of the excipients and solvents, a drug component matrix is ​​constructed by direct splicing.

[0035] Based on the structural characteristics of the drug and the characteristics of the excipients and solvents, an interaction matrix of drug components is constructed through a fusion approach.

[0036] Among them, the direct splicing method refers to first splicing by column, splicing the characteristics of each component in the drug preparation and the content of that component into a vector, and then splicing by horizontally, that is, splicing the feature vectors of different drugs into a drug component matrix;

[0037] The fusion approach refers to the modeling of pairwise interaction matrices between individual drug components and excipients, as well as the interaction matrices between components in a drug formulation, using learnable weight matrices.

[0038] Furthermore, the drug component interaction matrix includes a drug binary interaction feature matrix and a drug multivariate interaction feature matrix.

[0039] Furthermore, drug property prediction based on multi-task learning includes:

[0040] For classification tasks, the probability of the target label is predicted based on the first model; the first model uses a multilayer perceptron (MLP) combined with a softmax function; the formula for the first model is: , Representing the The sample at the th The corresponding drug feature vector in each classification task, i.e., the drug feature matrix; Indicates the first The sample at the th Predicted values ​​of drug attributes in a classification task;

[0041] For the regression task, the target value is predicted based on the second model; the second model uses a multilayer perceptron (MLP); the formula for the second model is: , Representing the The sample at the th The corresponding drug feature vector in each regression task; Indicates the first The sample at the th Predicted values ​​of drug attributes in a classification task;

[0042] For time-series tasks, the target value is predicted based on the third model. The third model is an improvement on the traditional NODE by introducing a drug feature matrix and combining it with concentration state variables to design novel differential equations that accurately characterize the dynamic drug release process. The formula for the third model is: ,in, Representing the One sample in In each time-series task rate of time, Representing the One sample in In each time-series task The cumulative value at each moment, This represents the learnable parameters in the neural network within Neural ODE. Representing the The sample at the th The drug feature vector corresponding to each time-series task, considering that it represents the initial characteristics of the drug, does not change with time; This represents the partial differential equation to be trained.

[0043] Secondly, the present invention provides a drug formulation attribute prediction system based on multi-task learning, the system comprising:

[0044] The pharmaceutical information acquisition unit is used to acquire first information and second information from pharmaceutical preparation data; the first information is drug information, and the second information is excipient and solvent information;

[0045] The hierarchical diagram construction unit is used to construct a hierarchical diagram based on the spatial characteristics of drug atoms according to the first information; the hierarchical diagram includes atomic group layers, bond layers, atomic layers, and the relationships between atomic group layers and bond layers, and between bond layers and atomic layers;

[0046] The drug structure feature extraction unit is used to construct a hierarchical graph convolutional network based on the hierarchical graph. The hierarchical graph convolutional network is used to extract drug structure features through bidirectional information aggregation from bottom to top and top to bottom. The drug structure features are the three-dimensional multi-scale structural features of drug molecules.

[0047] The excipient and solvent feature extraction unit is used to construct excipient and solvent features based on the type of drug excipient according to the second information; the excipient and solvent features are one-dimensional tripartite feature vectors containing surfactant, co-solvent, and oil phase;

[0048] The overall drug feature construction unit is used to construct the drug component matrix and the drug component interaction matrix based on the drug structural features and the excipient and solvent features; and the drug component matrix and the drug component interaction matrix are fused through a cross attention mechanism to obtain the drug feature matrix;

[0049] The multi-task learning unit is used to predict drug attributes based on the drug feature matrix and multi-task learning.

[0050] Furthermore, the execution process of the drug structure feature extraction unit is as follows:

[0051] Based on hierarchical graphs, a hierarchical graph convolutional network is constructed using the MPNN framework;

[0052] A hierarchical graph convolutional network is adopted, which aggregates information from the bottom up according to the atomic layer, bond layer and atomic group layer;

[0053] A hierarchical graph convolutional network is adopted, which aggregates information from top to bottom in sequence according to the atomic group layer, bond layer and atomic layer;

[0054] By combining the above information, the structural features of the drug are extracted.

[0055] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0056] 1. This invention discloses a method and system for predicting drug formulation properties based on multi-task learning. First, a hierarchical graph convolutional network is used to extract drug information, including spatial and atomic property information, for drug characterization. Second, for drug components, a drug component matrix is ​​constructed by fusing components of the same type to characterize the influence of components on the drug. Next, a drug-component interaction matrix is ​​constructed to capture the influence of interactions between components on drug properties. Finally, a multi-task learning framework is introduced, combining Neural ODE to learn the temporal features of drug release rate and the relationship between drug characteristics for accurate prediction of drug properties.

[0057] 2. This invention constructs a hierarchical graph and extracts multi-scale three-dimensional structural features of drug molecules at the atomic, bond, and group levels using HG-GCN, overcoming the limitations of traditional two-dimensional fingerprints and enhancing the model's ability to deeply represent drug spatial information. It constructs a ternary feature matrix of drug components through concentration-weighted aggregation, standardizes features according to excipient and solvent types (surfactants, cosolvents, oil phase), and unifies the dimension by concentration, solving the problem of characterizing excipient mixtures and enhancing the standardization of component features. It utilizes binary and multi-element interaction matrices to quantify pairwise and higher-order interactions between drugs and excipients, compensating for the shortcomings of traditional methods in capturing component relationships and improving the model's accuracy in modeling inter-component influences. It integrates component and interaction features using a cross-attention mechanism to achieve joint modeling of single component attributes and inter-component relationships, strengthening the understanding of dynamic interactions between features. It employs a multi-task learning framework, reducing redundant calculations by sharing parameters and adjusting parameters for different attributes to capture intrinsic relationships (such as particle size and PDI), improving model prediction efficiency and attribute relationship modeling capabilities. For time-series tasks, this invention introduces Neural... ODE combines drug characteristics with concentration status to model the dynamic release process, solving the problem that traditional methods cannot capture time-dependent factors and improving the accuracy of predicting dynamic attributes such as drug release rate. Attached Figure Description

[0058] The accompanying drawings, which are included to provide a further understanding of embodiments of the invention and form part of this application, do not constitute a limitation thereof. In the drawings:

[0059] Figure 1 This is a flowchart of a drug formulation attribute prediction method based on multi-task learning according to the present invention.

[0060] Figure 2 This is a flowchart illustrating the construction process of the hierarchical diagram of the present invention.

[0061] Figure 3 This invention is based on the fusion of excipient and solvent characteristics according to the category of pharmaceutical ingredients (taking self-surfactants as an example).

[0062] Figure 4This is a flowchart illustrating the construction of the binary interaction feature matrix of the pharmaceutical agents in this invention.

[0063] Figure 5 This is a flowchart illustrating the construction of the pharmaceutical component interaction matrix of the present invention;

[0064] Figure 6 This is a schematic diagram of the drug property prediction based on multi-task learning according to the present invention;

[0065] Figure 7 This is a block diagram of a drug formulation attribute prediction system based on multi-task learning according to the present invention. Detailed Implementation

[0066] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the embodiments and accompanying drawings. The illustrative embodiments and descriptions of the present invention are only used to explain the present invention and are not intended to limit the present invention.

[0067] This invention designs a method and system for predicting drug formulation properties based on multi-task learning. First, a hierarchical graph convolutional network is used to extract drug information, including spatial and atomic property information, for drug characterization. Second, for drug components, a drug component matrix is ​​constructed by fusing components of the same type to characterize the influence of components on the drug. Next, a drug-component interaction matrix is ​​constructed to capture the influence of inter-component interactions on drug properties. Finally, a multi-task learning framework is introduced, combining Neural ODE to learn the temporal features of drug release rate and the relationship between drug characteristics for accurate prediction of drug properties.

[0068] The key technical points of this invention are as follows:

[0069] (1) This invention constructs a three-dimensional hierarchical graph to capture the multi-scale structural features of drugs (i.e., drug structural features), and designs a hierarchical graph convolutional network HG-GCN to extract the structural features at the atomic, bond, and group levels, thereby realizing the deep characterization of the three-dimensional spatial information of drugs.

[0070] (2) To address the diversity of pharmaceutical components, a concentration-weighted polymerization strategy was adopted to construct a ternary feature matrix (surfactant, cosolvent, oil phase), and the component interaction matrix was used to quantify the component relationships. A drug component interaction matrix was specially designed to focus on capturing the interactions between the drug and excipients and solvents.

[0071] (3) A multi-task learning framework is proposed, which extracts general drug features by sharing parameters and combines task-specific learnable parameters and offsets to achieve accurate modeling of differences in drug properties. This mechanism enhances task adaptability while maintaining feature sharing.

[0072] (4) Neural ODE is used to model drug release dynamics, with drug properties embedded as control variables in the differential equation, and the concentration change trajectory is solved by the ODE solver. Combined with adjoint sensitivity analysis, the parameter learning efficiency is optimized to capture the temporal characteristics of drug release rate.

[0073] Example 1

[0074] like Figure 1 As shown, this invention provides a method for predicting drug formulation attributes based on multi-task learning, the method comprising:

[0075] Step 1: Obtain first information and second information from the drug formulation data; the first information is drug information, and the second information is excipient and solvent information.

[0076] In this embodiment, the drug formulation (hereinafter referred to as drug) data comes from multiple sources. One part of the data comes from long-acting drug data provided in the literature "Machine learning models to accelerate the design of polymeric long-actinginjectables"; another part comes from self-emulsifying formulation data in the literature "A dataset of formulation compositions for self-emulsifying drug delivery systems." This data is stored in CSV format on the Open Science Framework (OSF) platform. We downloaded the CSV format drug data and used the Pandas library for reading and preprocessing. The data file is "sedds_dataset_full.csv," and each data entry contains two parts of the drug's attributes: one part relates to the drug's components, including the characteristics and amounts of the drug, excipients, and solvents; the other part relates to the properties of the drug itself. If any drug attributes are missing, the data is discarded.

[0077] In addition, we obtained the corresponding SDF files of drug molecules from the PubChem database based on their names, and used the RDKit tool to generate a geometric undirected graph of the drug molecules, denoted as G = (V, E), where V represents the set of nodes, E represents the set of edges, and |V| represents the total number of nodes.

[0078] Step 2: Based on the first information, construct a hierarchical diagram based on the spatial characteristics of drug atoms; the hierarchical diagram includes atomic cluster layers, bond layers, atomic layers, and the relationships between atomic cluster layers and bond layers, as well as the relationships between bond layers and atomic layers;

[0079] In this embodiment, to effectively capture molecular spatial features, this invention proposes a hierarchical graph construction method. Addressing the limitation of traditional GCNs that only focus on node aggregation, this method extracts multi-dimensional spatial information through hierarchical modeling: First, it aggregates atoms and their adjacent nodes to construct an atomic cluster layer, capturing local structural features; second, it constructs a bond layer based on the bond information between atomic clusters, extracting bond-level structural features and establishing a mapping relationship between the atomic cluster layer and the bond layer; finally, through the mapping relationship between the bond layer and the atomic layer, it integrates multi-granularity features to achieve a multi-level representation of the molecular three-dimensional structure. The specific construction process is as follows:

[0080] First, the steps for constructing the atomic group layer include:

[0081] Aggregate atomic nodes and their adjacent nodes to construct atomic cluster nodes;

[0082] Based on the atomic cluster nodes, establish the connections, i.e., edges, between atomic cluster nodes;

[0083] Merge duplicate cluster nodes and remove redundant edges to obtain the cluster graph and use it as the cluster layer.

[0084] In order to capture different local structural information of molecules, the present invention identifies drug molecule maps. The nodes of the atomic clusters are constructed from the edge nodes and their adjacent nodes. The chemical bonds between these atomic clusters are considered as edges, thus constructing the atomic cluster graph. The specific construction process of the atomic cluster graph is detailed below:

[0085] (1) Initialization and Traversal Preparation

[0086] First, initialize the molecular diagram. An array of access markers for all nodes, used to record whether a node has been processed. Iterate through each node. Unmarked nodes are processed sequentially to ensure that each node participates in the construction of atom clusters only once.

[0087] (2) Degree screening

[0088] For the current node v, check its degree d(v): if d(v) < 2, skip the node; only process nodes with d(v) ≥ 2. Nodes with a degree ≥ 2 are the core of the molecular skeleton, containing local structural information such as branches and rings, and are the basis for building atomic groups.

[0089] (3) Constructing atomic cluster nodes

[0090] With atomic nodes At the core, it is merged with all adjacent atomic nodes, retaining all chemical bonds between these atoms, thus forming an atomic cluster node. This atomic group node fully characterizes the molecular structure. The local chemical structural unit in which an atom is located.

[0091] (4) Establishing connections between atomic groups

[0092] Traverse all atomic cluster nodes ,examine nodes If u belongs to another atomic group Then add edges to the atomic group diagram. .

[0093] (5) Deduplication and output (generating a cluster diagram)

[0094] Merge duplicate cluster nodes, remove redundant edges, and finally output the cluster graph. In this diagram, each node represents a local structural unit, and edges represent connections between units.

[0095] Second, the construction steps for the relationships between bond layers, atomic layers, and atomic group layers, as well as the relationships between bond layers and atomic layers, include:

[0096] Initialize the bond graph, traverse all group nodes, extract group bonds, and create a set of bond nodes;

[0097] Iterate through the set of bond nodes for each atomic group. If two bond nodes share an atom, add an edge to the bond graph.

[0098] After traversing all atomic groups, output the bond diagram and cross-layer mapping, and use the bond diagram as the bond layer; the nodes of the bond diagram contain the chemical bonds inside all atomic groups, and the edges represent the bond connections between shared atoms; the cross-layer mapping records the many-to-many associations from atomic groups to bonds;

[0099] And based on the bond layers, a mapping relationship between bond layers and atomic layers is established.

[0100] The above technical solution aims to extract the topological correlations of bonds within atomic groups and cross-group interactions. This invention constructs a bond layer based on the adjacency relationship of atomic groups. Its function is to acquire bond-level structural features, overcoming the limitations of atomic group layers in capturing bond-level structural features and the limitations of atomic layers in modeling higher-order bond relationships. Each bond node corresponds to a chemical bond within an atomic group. If two bonds share an atom, an inter-bond connection is established. Simultaneously, based on the inclusion relationship between atomic groups and bonds, a bidirectional mapping is established—bond nodes are mapped to their respective atomic groups. If a bond is shared by multiple atomic groups, it is simultaneously associated with all relevant atomic groups, achieving cross-group feature interaction. The specific steps of the bond layer are as follows:

[0101] (1) Initialize the key diagram

[0102] Create an empty key graph It is used to store key nodes (chemical bonds) and their connections, and to simultaneously establish an empty mapping set. Used to record the mapping relationship between atomic groups and bonds.

[0103] (2) Extraction of group bonds and construction of bond nodes

[0104] Traverse atomic cluster nodes Extract its atomic set Screening the original molecular graph where both ends belong to The key. For each selected key, create a unique key node. and in the mapping set The correspondence between the atomic group and the bond node is recorded in the database.

[0105] (3) Establish key connections

[0106] Traverse each atomic group The set of key nodes of a point set If two key nodes If atoms are shared, edges are added to the bond diagram. Edges represent the spatial adjacency of bonds (such as conjugate bonds and bond angle relationships), and local bond networks are constructed (such as connections between bonds within a ring).

[0107] (4) Output bond graph (i.e., bond layer) and mapping

[0108] After traversing all atomic groups, output the bond diagram. and cross-layer mapping Key graph nodes Includes chemical bonds within all atomic groups, edge Represents the bond connections between shared atoms; mapping Record the many-to-many relationship between atomic groups and bonds, that is, represent the mapping relationship between atomic group layers and bond layers.

[0109] After constructing the bond layers, a mapping relationship is established between the bond layers and atomic layers based on the inclusion relationship between them, which is used for learning between the bond layers and atomic layers. The specific construction process is as follows:

[0110] (1) Initialize the atomic diagram

[0111] Creating an atomic graph It is used to store the connection relationships between atomic nodes and atoms, where .

[0112] (2) Establish the mapping relationship between bond layers and atomic layers:

[0113] For each key node in the key graph Establish mapping , It represents the mapping relationship between bond layers and atomic layers; that is, each bond node uniquely corresponds to the two atomic nodes it is connected to.

[0114] Step 3: Construct a hierarchical graph convolutional network based on the hierarchical graph; using the hierarchical graph convolutional network, extract drug structural features through bidirectional information aggregation from bottom to top and top to bottom; the drug structural features are the three-dimensional multi-scale structural features of drug molecules.

[0115] In this embodiment, step 3 specifically includes:

[0116] Based on hierarchical graphs, a hierarchical graph convolutional network is constructed using the MPNN framework;

[0117] A hierarchical graph convolutional network is adopted, which aggregates information from the bottom up according to the atomic layer, bond layer and atomic group layer;

[0118] A hierarchical graph convolutional network is adopted, which aggregates information from top to bottom in sequence according to the atomic group layer, bond layer and atomic layer;

[0119] By combining the above information, the structural features of the drug are extracted.

[0120] The above technical solutions, addressing the multi-granularity characteristics of molecular hierarchical graphs, present invention design a Hierarchical Graph Convolutional Neural Network (HG-GCN) to capture multi-scale structural features through a hierarchical modeling mechanism. Hierarchical graph features are divided into intra-hierarchical features and inter-hierarchical features: intra-hierarchical features encode local structural information at the atomic layer, bond layer, and group layer, respectively; inter-hierarchical features achieve cross-layer feature interaction through the inclusion relationship between atoms, bonds, and groups. The construction process of the hierarchical graph is as follows: Figure 2 As shown.

[0121] This invention employs the MPNN framework to construct a hierarchical graph convolutional network. Through bidirectional information aggregation—both bottom-up (atom → bond → group of atoms) and top-down (group of atoms → bond → atom)—it extracts multi-scale structural features of drug molecules. The hierarchical graph convolutional network model comprises three layers: atomic layers... , bond layer atomic clusters Information exchange between layers is achieved through explicit mapping (bond-atom, group-bond).

[0122] Specifically, the bottom-up information aggregation (bottom layer → top layer) is as follows:

[0123] (1) Atomic layer :

[0124] The atomic layer, as the bottom layer, only contains information on the chemical bonds between atoms in the same layer (original molecular diagram). of ):

[0125] (1)

[0126] in, Nodes representing atomic layers In the Structural features of a wheel. Nodes representing atomic layers In the The structural features of the wheel after it aggregates from bottom to top. Indicates the atomic layer and A set of directly connected atoms, | indicates the atomic layer and The number of directly connected atoms. Indicates the first Layer Atom Learnable weights. This represents the activation function.

[0127] (2) Bond layer :

[0128] When constructing a hierarchical model, starting from the bond layer, each node not only aggregates information from other nodes in the same layer, but also considers information from nodes in the next layer to supplement the nodes with lower-granularity atomic information. For example, for nodes in the bond layer, it is necessary to aggregate not only the information inside the bond layer, but also the information from the atomic layer nodes adjacent to the bond layer nodes. The specific formula is as follows:

[0129] (2)

[0130] (3)

[0131] (4)

[0132] (5)

[0133] in, Represents a key layer node. Represents atomic layer nodes. This represents the mapping function between bond layer nodes and atomic layer nodes. Indicates the first Wheel key layer nodes The characteristic of bottom-up polymerization from atomic layers. Represents key layer nodes In the The wheel exhibits a bottom-up, layer-by-layer aggregation characteristic. Represents key layer nodes In the Structural characteristics of layers. Represents key layer nodes In the The structural characteristics of layers that aggregate from bottom to top. Indicates the bond layer and The set of bond neighbors that share atoms. Indicates the bond layer and The number of bond neighbors of a shared atom. Indicates the first Layer key node Learnable weights. This represents a multilayer perceptron. This represents the activation function.

[0134] (3) Atomic group layer:

[0135] For nodes in the atomic cluster layer, in order to supplement information at a lower granularity, it is necessary not only to aggregate the information adjacent to the node in the atomic cluster layer, but also to aggregate the information of the bond nodes adjacent to it. The specific formula is as follows:

[0136] The specific formula is as follows:

[0137] (6)

[0138] (7)

[0139] (8)

[0140] in, Represents the atomic group layer node, Represents a key layer node. This represents the mapping function between atomic group layer nodes and bond layer nodes. Represents the nodes of the atomic group layer Through mapping The number of key layer nodes obtained. Indicates the first Wheel atomic cluster layer Features that aggregate from the bond layer from the bottom up. Indicates atomic group layer In the The wheel exhibits a bottom-up, layer-by-layer aggregation characteristic. Represents atomic group layer nodes In the Structural characteristics after wheel aggregation. Indicating the atomic group layer with The set of neighbors for a shared key. Indicates the first Layer atomic cluster layer nodes Learnable weights. This represents a multilayer perceptron. This represents the activation function.

[0141] Specifically, the top-down information aggregation is as follows:

[0142] The top-down process involves passing complex set information from higher levels to lower levels, thus expanding the representational capabilities of lower-level nodes. The process is as follows:

[0143] (9)

[0144] (10)

[0145] in, Represents atomic layer nodes In the Structural features of wheels Represents key layer nodes In the Structural features of wheels Represents atomic group layer nodes In the Structural characteristics after wheel aggregation. Indicates containing key A collection of atomic groups, Indicates containing key The number of atomic groups. Represents key layer nodes In the The structural characteristics of layers that aggregate from bottom to top. Indicates connecting atoms The set of keys, Indicates connection of atomic nodes The number of keys. Nodes representing atomic layers In the The structural features of the wheel after it aggregates from bottom to top.

[0146] Finally, the drug structural features are obtained by reading the atomic layer information that incorporates the upper-level information:

[0147] (11)

[0148] in, This means that the drug structural features are obtained by reading atomic layer information. This represents the structural features of atomic layer nodes after learning by a hierarchical graph convolutional neural network, n=| | indicates the number of nodes in the atomic layer.

[0149] Step 4: Based on the second information, construct excipient and solvent features based on the type of drug excipient; the excipient and solvent features are one-dimensional tripartite feature vectors containing surfactants, cosolvents, and oil phases;

[0150] In this embodiment, considering the multi-component characteristics of self-emulsifiers, the present invention constructs a ternary feature representation method: the chemical features of surfactants, cosolvents, and oil phases are initially embedded from a database. Considering that excipients are mostly mixtures, a concentration-weighted polymerization strategy is adopted to unify the dimensions, ultimately forming a ternary feature vector containing surfactants, cosolvents, and oil phases.

[0151] (12)

[0152] (13)

[0153] in, These represent the initial characteristics of the surfactant, co-solvent, and oil phase in the pharmaceutical preparation, respectively. They represent the first Initial characteristics of surfactants, co-solvents, and the oil phase. They represent the first The content of various surfactants, cosolvents, and oil phases in the pharmaceutical preparation. These represent the contents of surfactants, co-solvents, and the oil phase in the pharmaceutical preparation, respectively. These represent the quantities of surfactant, cosolvent, and oil phase in the pharmaceutical preparation, respectively. This represents a multilayer perceptron.

[0154] like Figure 3 As shown, Figure 3 To integrate the characteristics of excipients and solvents based on the category of pharmaceutical ingredients, let's take surfactants as an example.

[0155] Step 5: Based on the drug structural characteristics and the characteristics of excipients and solvents, construct the drug component matrix and the drug component interaction matrix; and fuse the drug component matrix and the drug component interaction matrix through a cross-attention mechanism to obtain the drug feature matrix;

[0156] In this embodiment, the drug component interaction matrix includes a drug binary interaction feature matrix and a drug multivariate interaction feature matrix, specifically:

[0157] (1) Construction of the characteristic matrix of drug-pharmaceutical binary interaction

[0158] First, a multilayer perceptron is used to align the dimensions of the multi-source features to support subsequent matrix fusion analysis.

[0159] (14)

[0160] in, Indicates The feature embedding vectors of drug, surfactant, cosolvent, and oil phase in the mapped drug.

[0161] Considering that the drug is the core component that performs the function in the pharmaceutical preparation, an interaction matrix was constructed between the drug and surfactant, the drug and cosolvent, and the drug and oil phase to reflect the influence of binary interactions on the characteristics of the pharmaceutical preparation.

[0162] (15)

[0163] in, This represents the interaction matrix between the drug and surfactant, the drug and cosolvent, and the drug and oil phase. Indicates the content of the drug in the preparation. These represent the contents of surfactant, co-solvent, and oil phase in the pharmaceutical preparation, respectively.

[0164] like Figure 4 As shown, Figure 4 This is a flowchart illustrating the construction of the characteristic matrix of drug-pharmaceutical binary interactions.

[0165] (2) Construction of the characteristic matrix of drug-pharmaceutical multivariate interactions

[0166] Considering that the reaction matrix of pharmaceutical components becomes extremely complex as the number of components in a drug increases, this invention integrates pharmaceutical component categories of the same broad class to construct a ternary reaction matrix of drug, excipient, and oil phase. The reaction matrices of drug and excipient, drug and oil phase, and excipient and oil phase are constructed through 1*1 convolution, and then obtained through matrix weighting. The specific construction process is as follows:

[0167] (16)

[0168] (17)

[0169] (18)

[0170] (19)

[0171] (20)

[0172] (twenty one)

[0173] (twenty two)

[0174] (twenty three)

[0175] in, Represents the drug response tensor. Represents the reaction tensor of pharmaceutical excipients. This indicates the embedding of excipient characteristics in a pharmaceutical preparation. This indicates the content of excipients in the drug, and Conv represents the convolution function. These are the reaction matrices between the drug and excipients considering the properties of the oil phase, the reaction matrices between the excipients and the oil phase considering the properties of the drug, and the reaction matrices between the drug and the oil phase considering the properties of the excipients. represent The corresponding weight contribution; These represent the learnable weights for surfactant characteristics and cosolvent characteristics, respectively. The feature embedding dimension represents the drug characteristics, oil phase characteristics, and excipient characteristics of the formulation; This represents the characteristic matrix of multi-component interactions.

[0176] like Figure 5 As shown, Figure 5 This is a flowchart illustrating the construction of the drug component interaction matrix.

[0177] In this embodiment, the drug component matrix and the drug component interaction matrix are fused: First, a drug component interaction matrix is ​​constructed to quantify the impact of reactions between drug components on drug properties. This embodiment also constructs a drug component feature matrix by concatenating the feature vectors and content information of each component in the drug to assess the independent impact of a single component and its content on formulation properties. Furthermore, through an attention mechanism, the drug component feature matrix and the drug component interaction feature matrix are fused to generate a total drug feature matrix, achieving joint modeling of the inherent properties of components and interactions between components.

[0178] Specifically, based on the structural characteristics of the drug and the characteristics of the excipients and solvents, a drug component matrix is ​​constructed by direct splicing; based on the structural characteristics of the drug and the characteristics of the excipients and solvents, a drug component interaction matrix is ​​constructed by fusion.

[0179] Specifically, the formula for constructing the pharmaceutical component feature matrix is ​​as follows:

[0180] (twenty four)

[0181] in, These respectively represent the characteristics of the drug, surfactant, cosolvent, and oil phase, as well as their content in the solution. This indicates that the data is concatenated column-wise, combining the characteristics of each component of the drug and its concentration into a single vector. This indicates that the data is spliced ​​horizontally, that is, the feature vectors of different drugs are spliced ​​together to form a drug composition matrix. The component feature matrix represents the drug, with each row representing the feature information of a class of components.

[0182] The influence of various reactions on drug characteristics is learned through a learnable weight matrix. This is used to model the pairwise interaction matrices between a single drug component and excipients such as surfactants, cosolvents, and oil phases, as well as the interaction matrices between individual components in the drug.

[0183] (25)

[0184] in, This represents the weights of the pairwise interaction matrix between a single drug component and excipients such as surfactants, cosolvents, and oil phases. This represents the weights corresponding to the multi-element interaction feature matrix; These represent the interaction matrices between a single drug component and surfactants, cosolvents, and the oil phase, respectively. This represents the characteristic matrix of multi-component interactions. The overall characteristics of drug-pharmaceutical interactions.

[0185] By fusing the drug component matrix and the drug component interaction matrix through a cross-attention mechanism, the final characterization of the drug is obtained, which is used to evaluate the influence of drug components and the interactions between drug components on drug properties.

[0186] (26)

[0187] (27)

[0188] in, Representative drug component characteristics and overall drug-drug interaction characteristics. It indicates the overall characteristics of the drug. represent The feature dimensions.

[0189] Step 6: Based on the drug feature matrix, predict drug attributes using multi-task learning.

[0190] In this embodiment, during the drug feature modeling process, feature mapping can be achieved through a fully connected layer to capture the influence of components and reactions on drug properties. Specifically, the weight matrix and bias terms of the fully connected layer are used to transform the comprehensive features of the drug, thereby learning key feature representations relevant to a specific task. The formula is as follows:

[0191] (28)

[0192] in, This represents the drug feature vector corresponding to the v-th task. This indicates the operation of the fully connected layer. For learnable matrices and offsets, Characteristics of representative pharmaceutical preparations.

[0193] After extracting the features corresponding to each task, this model matches different processing models according to the differences in task type in order to achieve accurate prediction of task attributes.

[0194] For classification tasks, the probability of the target label is predicted based on the first model. The first model uses a multilayer perceptron (MLP) combined with a softmax function. During the training process of the classification task, the binary cross-entropy loss function is used to measure the difference between the predicted result and the true label. The specific formula of the first model is as follows:

[0195] (29)

[0196] (30)

[0197] in, Representing the The sample at the th The corresponding drug feature vector in each classification task, i.e., the drug feature matrix; Indicates the first The sample at the th Predicted values ​​in a classification task, This represents the true label of the nth sample in the cth classification task; This represents the loss for the c-th classification task. This represents the number of samples used for the c-th classification task.

[0198] For regression tasks, the target value is predicted based on a second model; the second model employs a multilayer perceptron (MLP); during the training process of regression tasks, the mean squared error loss function is typically used to measure the difference between the predicted and true values. The specific formula for the second model is as follows:

[0199] (31)

[0200] (32)

[0201] in, Representing the The sample at the th The corresponding drug feature vector in each regression task, i.e., the drug feature matrix; Indicates the first The sample at the th Predicted values ​​of drug properties in a regression task. It means that the nth sample is in the nth order. Predicted values ​​of drug properties in a regression task; Representing the The loss of a regression task. This represents the total number of samples used for the r-th regression task.

[0202] In time-series tasks, the release and dissolution process of a drug in a liquid environment exhibits dynamic characteristics. Its rate depends not only on the inherent characteristics of the drug itself but also on its current concentration state. Traditional multilayer perceptrons (MLPs), limited by their static mapping characteristics, cannot effectively capture the dynamic relationship between rate and concentration evolving over time. While neural network-based differential equations (NODEs) can model dynamic processes through differential equations, traditional NODEs only describe the instantaneous relationship between release rate and concentration at the current moment, failing to comprehensively consider the dual influence of drug characteristics and concentration state. To address this issue, this invention improves upon traditional NODEs by introducing a drug feature matrix and combining it with concentration state variables to design novel differential equations that accurately characterize the dynamic drug release process. Specifically, for time-series tasks, this invention predicts the target value based on a third model; the specific formula for the third model is as follows:

[0203] (33)

[0204] in, Representing the One sample in In each time-series task rate of time, Representing the One sample in In each time-series task The cumulative value at each moment, This represents the learnable parameters in the neural network within Neural ODE. Representing the The sample at the th The drug feature vector corresponding to each time-series task, considering that it represents the initial characteristics of the drug, does not change with time; This represents the partial differential equation to be trained.

[0205] The above ODE function can be solved using an ODE solver, and the solution process can be expressed as follows:

[0206] (34)

[0207] in, Representing the One sample in In each time-series task The predicted value of the cumulative value at time step. Representing the One sample in The amount of drug released in the initial state of a time-series task.

[0208] The prediction accuracy is measured by the cumulative error of time-series tasks.

[0209] (35)

[0210] in, This represents the predicted cumulative value of the nth sample at time s in the t-th time series task. That is the corresponding actual value. This represents the total number of samples used for the t-th time-series task. Let be the number of values ​​to be predicted for the nth sample in the t-th time series task, and It represents the total number of values ​​to be predicted for all samples in the t-th time series task.

[0211] To train the model, this invention defines an overall loss and updates the model's parameters through backpropagation:

[0212] (36)

[0213] in, Representing the The loss of each classification task, Representing the The loss of a regression task Representing the The loss of each sequential task; The number of classification tasks. Represents the number of regression tasks. This represents the number of sequential tasks.

[0214] like Figure 6 As shown, Figure 6 This is a schematic diagram illustrating drug feature fusion and drug attribute prediction based on multi-task learning.

[0215] This invention overcomes the limitations of traditional two-dimensional fingerprints by constructing hierarchical graphs and extracting multi-scale three-dimensional structural features of drug molecules at the atomic, bond, and group levels using HG-GCN, thus enhancing the model's ability to deeply represent drug spatial information. It constructs a triplet feature matrix of drug components through concentration-weighted aggregation, standardizes features according to excipient type, and unifies dimensions by concentration, solving the challenge of representing excipient mixtures and enhancing the standardization of component features. It quantifies pairwise and higher-order interactions between drugs and excipients using binary and multi-element interaction matrices, compensating for the shortcomings of traditional methods in capturing component relationships and improving the model's accuracy in modeling inter-component influences. It integrates component and interaction features using a cross-attention mechanism to achieve joint modeling of single component attributes and inter-component relationships, strengthening the understanding of dynamic interactions between features. It employs a multi-task learning framework, reducing redundant computations by sharing parameters and adjusting parameters for different attributes to capture intrinsic relationships (such as particle size and PDI), improving model prediction efficiency and attribute relationship modeling capabilities. For time-series tasks, this invention introduces Neural... ODE combines drug characteristics with concentration status to model the dynamic release process, solving the problem that traditional methods cannot capture time-dependent factors and improving the accuracy of predicting dynamic attributes such as drug release rate.

[0216] Example 2

[0217] like Figure 7 As shown, the difference between this embodiment and Embodiment 1 is that this embodiment provides a drug formulation attribute prediction system based on multi-task learning, which corresponds one-to-one with the drug formulation attribute prediction method based on multi-task learning in Embodiment 1; the system includes:

[0218] The pharmaceutical information acquisition unit is used to acquire first information and second information from pharmaceutical preparation data; the first information is drug information, and the second information is excipient and solvent information;

[0219] The hierarchical diagram construction unit is used to construct a hierarchical diagram based on the spatial characteristics of drug atoms according to the first information; the hierarchical diagram includes atomic group layers, bond layers, atomic layers, and the relationships between atomic group layers and bond layers, and between bond layers and atomic layers;

[0220] The drug structure feature extraction unit is used to construct a hierarchical graph convolutional network based on the hierarchical graph. The hierarchical graph convolutional network is used to extract drug structure features through bidirectional information aggregation from bottom to top and top to bottom. The drug structure features are the three-dimensional multi-scale structural features of drug molecules.

[0221] The excipient and solvent feature extraction unit is used to construct excipient and solvent features based on the type of drug excipient according to the second information; the excipient and solvent features are one-dimensional tripartite feature vectors containing surfactant, co-solvent, and oil phase;

[0222] The overall drug feature construction unit is used to construct the drug component matrix and the drug component interaction matrix based on the drug structural features and the characteristics of excipients and solvents; and to fuse the drug component matrix and the drug component interaction matrix through a cross attention mechanism to obtain the drug feature matrix;

[0223] The multi-task learning unit is used to predict drug attributes based on the drug feature matrix and multi-task learning.

[0224] As a further implementation, the execution process of the drug structure feature extraction unit is as follows:

[0225] Based on hierarchical graphs, a hierarchical graph convolutional network is constructed using the MPNN framework;

[0226] A hierarchical graph convolutional network is adopted, which aggregates information from the bottom up according to the atomic layer, bond layer and atomic group layer;

[0227] A hierarchical graph convolutional network is adopted, which aggregates information from top to bottom in sequence according to the atomic group layer, bond layer and atomic layer;

[0228] By combining the above information, the structural features of the drug are extracted.

[0229] The execution process of each unit can be carried out according to the steps of the drug formulation attribute prediction method based on multi-task learning in Example 1, and will not be described in detail in this example.

[0230] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0231] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0232] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0233] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0234] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for predicting attributes of a pharmaceutical preparation based on multi-task learning, characterized by, The method comprises: obtaining first information and second information from drug preparation data; the first information is drug information, and the second information is excipient and solvent information; constructing a hierarchical graph based on the atomic space characteristics of the drug according to the first information; the hierarchical graph comprises an atomic group layer, a bond layer, an atomic layer, and the relationships between the atomic group layer and the bond layer and the relationships between the bond layer and the atomic layer; constructing a hierarchical graph convolution network based on the hierarchical graph; using the hierarchical graph convolution network, drug structure features are extracted through bidirectional information aggregation from bottom to top and from top to bottom; the drug structure features are three-dimensional multi-scale structure features of the drug molecules; constructing excipient and solvent features based on the types of the drug excipients according to the second information; the excipient and solvent features are one-dimensional three-tuple feature vectors containing surfactants, potential solvents, and oil phases; constructing a drug component matrix and a drug component interaction matrix according to the drug structure features and the excipient and solvent features; and fusing the drug component matrix and the drug component interaction matrix through a cross-attention mechanism to obtain a drug feature matrix; performing drug attribute prediction based on multi-task learning according to the drug feature matrix. 2.The method of claim 1, wherein, The construction of the atomic group layer comprises: aggregating atomic nodes and their adjacent nodes to construct atomic group nodes; establishing connections between the atomic group nodes, i.e., edges, according to the atomic group nodes; merging repeated atomic group nodes and removing redundant edges to obtain an atomic group graph and take the atomic group graph as the atomic group layer; The construction of the bond layer, the atomic layer, and the relationships between the atomic group layer and the bond layer and the relationships between the bond layer and the atomic layer comprises: initializing a bond graph and synchronously establishing an empty mapping set; the bond graph is used to store bond nodes and their connection relationships; the mapping set is used to record the mapping relationship between atomic groups and bonds; traversing all atomic group nodes, extracting atomic group bonds, creating a bond node set, and recording the correspondence between the atomic group and the bond node in the mapping set; traversing the bond node set of each atomic group, and adding an edge in the bond graph if two bond nodes share an atom; after traversing all atomic groups, outputting the bond graph and the cross-layer mapping, and taking the bond graph as the bond layer; and establishing the relationship mapping between the bond layer and the atomic layer according to the bond layer. 3.The method of claim 1, wherein, Based on the hierarchical graph, a hierarchical graph convolution network is constructed; using the hierarchical graph convolution network, drug structure features are extracted through bidirectional information aggregation from bottom to top and from top to bottom, comprising: based on the hierarchical graph, a hierarchical graph convolution network is constructed using an MPNN framework; using the hierarchical graph convolution network, information is aggregated from bottom to top in the order of the atomic layer, the bond layer, and the atomic group layer; using the hierarchical graph convolution network, information is aggregated from top to bottom in the order of the atomic group layer, the bond layer, and the atomic layer; combined with the above information aggregation, drug structure features are extracted. 4.The method of claim 1, wherein, The drug features in the drug feature matrix are extracted from the three-dimensional multi-scale structure features of the drug molecules and atomic information; the chemical features of the surfactants, potential solvents, and oil phases in the excipient and solvent features are obtained from the initial embedding in the database. 5.The method of claim 1, wherein, According to the drug structure characteristics and the auxiliary material and solvent characteristics, a drug component matrix and a drug component interaction matrix are constructed, including: According to the drug structure characteristics and the auxiliary material and solvent characteristics, a drug component matrix and a drug component interaction matrix are constructed by direct splicing; According to the drug structure characteristics and the auxiliary material and solvent characteristics, a drug component interaction matrix is constructed by fusion; The direct splicing method refers to splicing the characteristics and content of each component in the drug preparation into a vector first, and then splicing horizontally, that is, splicing different drug feature vectors into a drug component matrix. The fusion method refers to fusing the two-by-two interaction matrix between a single drug component and each auxiliary material in the drug preparation through a learnable weight matrix modeling.

6. The multi-task learning-based pharmaceutical preparation attribute prediction method according to claim 5, wherein, The drug component interaction matrix includes a drug binary interaction feature matrix and a drug multi-component interaction feature matrix.

7. The method for predicting drug formulation attributes based on multi-task learning according to claim 1, characterized in that, Based on multi-task learning, the drug attribute prediction is performed, including: For classification tasks, the probability of the target label is predicted based on the first model; the first model uses a multilayer perceptron (MLP) combined with a softmax function; the formula for the first model is: , Representing the The sample at the th The corresponding drug feature vector in each classification task, i.e., the drug feature matrix; Indicates the first The sample at the th Predicted values ​​of drug attributes in a classification task; For the regression task, the target value is predicted based on the second model; the second model uses a multilayer perceptron (MLP); the formula for the second model is: , Representing the The sample at the th The corresponding drug feature vector in each regression task; Indicates the first The sample at the th Predicted values ​​of drug attributes in a classification task; For the time series task, a target value is predicted based on a third model; the third model is improved from a traditional NODE, a new type of differential equation is designed to depict the dynamic release process of the drug by introducing a drug feature matrix and combining the drug feature matrix with the concentration state variable; the formula of the third model is: wherein, represents the rate of the i th sample at the j th time point in the k th time series task, represents the cumulative value of the i th sample at the j th time point in the k th time series task, represents a learnable parameter in a neural network in the Neural ODE, a drug feature vector corresponding to the i th sample in the k th time series task, considering that it represents the initial characteristics of the drug, it does not change with time; indicates a partial differential equation to be trained.​​​​​​​​ 8.A drug formulation attribute prediction system based on multi-task learning, characterized by, The system includes: A drug information acquisition unit is used to acquire first information and second information from drug preparation data; the first information is drug information, and the second information is auxiliary material and solvent information; A hierarchical graph construction unit is used to construct a hierarchical graph based on drug atomic space characteristics according to the first information; the hierarchical graph includes an atomic group layer, a bond layer, an atomic layer, and the relationship between the atomic group layer and the bond layer, the relationship between the bond layer and the atomic layer; A drug structure feature extraction unit is used to construct a hierarchical graph convolution network based on the hierarchical graph; the hierarchical graph convolution network is used to extract drug structure features through bottom-up and top-down bidirectional information aggregation; the drug structure features are three-dimensional multi-scale structure features of drug molecules; An auxiliary material and solvent feature extraction unit is used to construct auxiliary material and solvent features based on drug auxiliary material types according to the second information; the auxiliary material and solvent features are one-dimensional three-tuple feature vectors containing surfactants, latent solvents, and oil phases; A drug feature matrix construction unit is used to construct a drug component matrix and a drug component interaction matrix according to the drug structure characteristics and the auxiliary material and solvent characteristics; and the drug component matrix and the drug component interaction matrix are fused through a cross-attention mechanism to obtain a drug feature matrix; A multi-task learning unit is used to perform drug attribute prediction based on multi-task learning according to the drug feature matrix. 9.The multi-task learning based pharmaceutical formulation attribute prediction system of claim 8, wherein, The execution process of the drug structure feature extraction unit is as follows: Based on the hierarchical graph, a hierarchical graph convolution network is constructed using an MPNN framework; Using the hierarchical graph convolution network, information is aggregated from bottom to top in the order of atomic layer, bond layer, and atomic group layer; Using the hierarchical graph convolution network, information is aggregated from top to bottom in the order of atomic group layer, bond layer, and atomic layer; Based on the above information aggregation, drug structure features are extracted.

Images