Method for predicting drug solubility in fasting state based on graph gating attention network of meta-learning and application of method for predicting drug solubility in fasting state based on graph gating attention network of meta-learning

Abstract

The invention relates to a method for predicting fasting state drug solubility through a graph gating attention network based on meta-learning and application of the method. The method comprises the steps that molecular characterization information is integrated to form a framework model through multi-modal mixed molecular feature engineering, and model-independent meta-learning sample training is adopted; the multi-modal mixed molecular feature engineering comprises the following steps: inputting an SMILES character string of a to-be-tested compound to obtain a molecular graph, a molecular fingerprint vector and a normalized descriptor vector, constructing the molecular graph into a molecular graph vector by adopting an attention gating graph network, splicing and fusing the molecular fingerprint vector and the normalized descriptor vector, and obtaining the multi-modal mixed molecular feature engineering of the to-be-tested compound. And fusing with a molecular graph vector, and inputting into a multi-layer perceptron to obtain a predicted value. According to the method, the solubility of the candidate drugs in various key biological related media can be rapidly predicted, the solubility characteristic of the drugs is accurately evaluated in the early stage of research and development, the research and development period of new drugs is shortened, and huge research and development cost is saved.

Description

Technical Field

[0001] This invention relates to the field of computer-aided drug design technology, and in particular to a method for predicting drug solubility in a fasting state using a graph-gated attention network based on meta-learning, and its application. Background Technology

[0002] In modern drug development, the solubility of a candidate drug is one of the key physicochemical properties determining its oral bioavailability. Low solubility often leads to incomplete drug absorption, low bioavailability, and significant individual variability, and is one of the main reasons for new drug development failures. Therefore, in the early stages of drug discovery, accurately predicting the solubility of candidate compounds under different physiological conditions is crucial for screening molecules with drug potential.

[0003] To simulate drug behavior in the human gastrointestinal tract, pharmaceutical science has developed various biomolecular mediators, among which FaSSIF, FeSSIF, and FaSSGF are the three most representative. FaSSIF (Fasted State Simulated Intestinal Fluid, simulating intestinal fluid in a fasting state); FeSSIF (Fed State Simulated Intestinal Fluid, simulating intestinal fluid in a fed state); and FaSSGF (Fasted State Simulated Gastric Fluid, simulating gastric fluid in a fasting state) are three more representative mediators. Accurate prediction of drug solubility in these mediators can provide crucial guidance for formulation design and clinical dosing regimens. Testing in these simulated bodily fluids yields data more relevant to in vivo pharmacokinetics (PK). Current research primarily focuses on solubility prediction for FaSSIF, but extending the prediction model to FaSSGF and FeSSIF to construct a complete absorption prediction model remains an urgent and underserved technological need.

[0004] With the development of artificial intelligence technology, deep learning models based on artificial neural networks (ANNs) have begun to be applied to solubility prediction. However, these standard deep learning methods also face a series of severe technical challenges. High-quality experimental data in specific media such as FaSSGF are often very scarce. In the case of sparse data, deep learning models are prone to overfitting, leading to a sharp decline in their predictive performance for new molecules. Deep learning models are often regarded as "black boxes," and their complex internal working mechanisms make it difficult for researchers to intuitively understand the basis for the model's specific predictions, i.e., it is difficult to establish a clear "structure-solubility" relationship. This greatly limits the application value of the models in the field of medicinal chemistry, because medicinal chemists not only need predictive results, but also interpretable guidance to optimize the structure of lead compounds.

[0005] Therefore, there is an urgent need for a data-efficient, adaptable, and highly accurate method for predicting drug solubility. Summary of the Invention

[0006] To address the aforementioned technical problems, this invention provides a method for predicting drug solubility in a fasting state using a graph-gated attention network based on meta-learning, and its application, offering a high-throughput, high-precision, and data-efficient in vitro (in-silico) screening platform. This platform can rapidly predict the solubility of candidate drugs in various key biologically relevant media, such as simulated gastric juice, simulated intestinal juice, and simulated intestinal juice after eating. By accurately assessing the solubility characteristics of drugs early in the research and development process, pharmaceutical companies can more effectively screen and optimize lead compounds, prioritizing candidates with good pharmacokinetic potential, thereby significantly reducing the risk of failure in later clinical trials, shortening the drug development cycle, and saving substantial R&D costs.

[0007] To achieve this objective, the present invention adopts the following technical solution:

[0008] In a first aspect, the present invention provides a method for predicting drug solubility in a fasting state using a graph-gated attention network based on meta-learning. The method includes integrating molecular characterization information into a model using multimodal hybrid molecular feature engineering and training with model-independent meta-learning samples.

[0009] (1) The multimodal hybrid molecular feature engineering includes: inputting the SMILES string of the compound to be tested, obtaining molecular graph, molecular fingerprint vector and normalized descriptor vector, constructing molecular graph vector from molecular graph using attention-gated graph network, splicing and fusing molecular fingerprint vector and normalized descriptor vector, and then fusing with molecular graph vector and inputting into multilayer perceptron to obtain predicted value.

[0010] (1.1) The attention-gated graph network architecture model includes: attention-based message aggregation, followed by gating state update and molecular-level readout;

[0011] The calculation process for the attention message aggregation includes alignment, weighting, and aggregation; the alignment calculation formula is as follows: ;in, For unnormalized attention scores, It is a learnable linear transformation layer. For target atoms The current state vector, Neighboring atoms The state vector, Neighboring atoms The edge;

[0012] The weighting calculation formula is as follows: ;in,

[0013] Attention weights;

[0014] The calculation formula for the aggregation is: ;in, For context vectors, For context vectors;

[0015] The formula for calculating the gate state update is: ; ;in, The target atom In the The state vector of the layer, It is a context vector aggregated from its neighbors. It is the first GRU cells that update the target atomic state at each layer. It is the first GRU cells that update edge states at each layer The target atom In the The updated state vector of the layer, Neighboring atoms In the The updated state vector of the layer, It is the edge In the The updated state vector of the layer;

[0016] (2) The model-independent meta-learning includes constructing a meta-task and then using a two-layer loop to optimize it to obtain the final prediction model;

[0017] The meta-task includes any one or a combination of at least two of the following: a fasting state simulating intestinal fluid task, a eating state simulating intestinal fluid task, or a fasting state simulating gastric fluid task.

[0018] This invention comprises three tightly coupled innovative modules: (1) multimodal hybrid molecular feature engineering; (2) a novel model architecture called "Attention Gated Graph Network (AGGN)"; and (3) a few-sample training strategy based on the Model-Agnostic Meta-Learning (MAML) framework. This invention transforms the problem of drug solubility prediction in different media from the traditional "single-task learning" paradigm to a "meta-learning" paradigm that better meets the needs of real-world scientific research. This paradigm shift combines innovative hybrid molecular characterization methods with an optimized graph neural network architecture.

[0019] This invention proposes a novel graph neural network architecture called "Attention-Gated Graph Network (AGGN)". This architecture deeply integrates graph attention mechanisms and gated recurrent units, aiming to achieve more efficient and discriminative molecular representation learning.

[0020] The first step is attention-based message aggregation: for any target atom in the graph... The model will assign each of its neighboring atoms Calculate an attention weight This weight dynamically measures the impact of factors from neighbors. The message is for updating the target atom The importance of this. The calculation process includes alignment, weighting, and aggregation; alignment involves aligning the target atoms... Current state vector With neighboring atoms state vector and edge The components are concatenated and passed through a learnable linear transformation layer. Then, the LeakyReLU nonlinear activation function is applied to obtain an unnormalized attention score. Weighting: Using the Softmax function on the target atoms. The attention scores of all neighbors are normalized to obtain the final attention weight. Aggregates into a state vector containing all neighboring atoms. After a linear transformation Then, based on its corresponding attention weight A weighted summation is performed to obtain the aggregated neighbor information, i.e., the context vector. Through this mechanism, the model can adaptively "focus" on the molecular local regions that are most important to the current prediction task.

[0021] In one specific embodiment of the invention, the meta-tasks are FaSSGF, FaSSIF, and FeSSIF. FaSSIF simulates the environment of the small intestine in a fasting state, which is crucial for assessing basal drug absorption. FeSSIF simulates the intestinal environment after eating and is used to study the effect of food on drug absorption. FaSSGF, on the other hand, simulates the gastric environment in a fasting state and is used to predict the initial dissolution behavior of drugs in the stomach.

[0022] Preferably, in step (1), the construction of the molecular graph includes converting the SMILES string into... Where V represents an atom, corresponding to a node in the graph, and E represents a chemical bond, corresponding to an edge in the graph.

[0023] This representation method preserves the molecular topology, enabling the model to directly learn the interatomic connections and local chemical environments. Using the open-source cheminformatics library RDKit, an initial feature vector is generated for each atomic node in the graph. This vector encodes the intrinsic properties of the atom and its local environment within the molecule. Edge (chemical bond) features: Similarly, a feature vector is generated for each edge in the graph to describe the type and properties of the chemical bond.

[0024] Preferably, in step (1), the construction of the molecular fingerprint vector includes construction using extended connectivity fingerprints.

[0025] In one specific embodiment of the invention, the Extended Connectivity Fingerprint (ECFP) used is ECFP4, which is encoded as a fixed-length (e.g., 1024-bit) binary vector. Each bit in the vector corresponds to the presence or absence of a specific circular chemical substructure (with a maximum diameter of 4 chemical bonds) in the molecule. ECFP provides a global, substructure-based perspective for examining molecules. It can efficiently encode key functional groups and structural fragments present in molecules, which are often important determinants of solubility.

[0026] Preferably, in step (1), the construction of the normalized descriptor vector includes obtaining it by normalizing one-dimensional and two-dimensional global molecular descriptors.

[0027] In this invention, the global molecular descriptor quantifies the overall physicochemical properties of the molecule. In a specific embodiment, key descriptors include descriptors such as molecular weight (MolWt), topological polar surface area (TPSA), and lipid-water partition coefficient (MolLogP), which provide macroscopic physicochemical background information for the model.

[0028] Preferably, in step (1), the fusion includes combining the molecular fingerprint vector and normalized descriptor vector The splicing and merging process is performed as follows:

[0029] ;

[0030] in To obtain learnable parameters, an attention-gated graph network is then used to construct molecular graph vectors from the molecular graph. The final molecular representation vector is obtained by fusion. The specific method is as follows:

[0031] ;

[0032] Where N is the total number of atoms in the molecule. Let be the state vector of the i-th atom in the molecule;

[0033] .

[0034] In this invention, the ECFP vector and the normalized physicochemical descriptor vector are concatenated. This fused non-graph feature vector is then fused again with the global graph representation vector learned from the molecular graph, and both are input into the final prediction layer. This multimodal fusion strategy ensures the comprehensiveness of the model's decisions.

[0035] Preferably, in step (2), the construction of the meta-task includes constructing the meta-task and dividing the data of the meta-task into a support set and a query set.

[0036] To address the issues of data sparsity and poor model adaptability, this invention employs the Model-Independent Meta-Learning (MAML) framework to train the AGGN model. The core objective of MAML is not to train a model that performs best on a specific task, but rather to learn a "general" initial model parameter set. This initial model can quickly adapt to any new, relevant task using only a few samples. The training process no longer treats the entire training set as a whole. Instead, it constructs the dataset into a series of "meta-tasks" simulating real-world small-sample scenarios. Each metatask These are all independent learning problems, such as "using 10 known samples (K-shot, K=10) to predict the solubility of 5 additional compounds in FaSSIF".

[0037] Preferably, the support set and the query set are the same number of samples but different samples.

[0038] Preferably, in step (2), the double-layer loop includes using an inner loop for rapid task adaptation and an outer loop for meta-knowledge learning.

[0039] In this invention, through MAML training, the model learns not just a simple "input-output" mapping, but the meta-capability of "how to learn solubility prediction." The final meta-model parameters are... While not optimized for any single medium, it occupies a central position in a "parameter space," being "very close" to the optimal parameters for various media (FaSSIF, FeSSIF, etc.). Therefore, when encountering a new medium with only a small amount of data, the model can start from this advantageous point and quickly converge to a high-performance, medium-specific model using just a few samples. This fundamentally solves the adaptability problem of existing technologies.

[0040] Preferably, the inner loop process includes replicating the meta-model of the meta-task, performing gradient descent updates 1-5 times using the support set of the task, and obtaining a set of parameters optimized for the specific task. And make fine adjustments, the specific calculation process is as follows: ;in, For the meta-task, For the task loss function on, This is the learning rate for the inner loop.

[0041] In this invention, for a sampling meta-task First, copy the current metamodel (with parameters). Using the support set for this task, a small number (e.g., 1-5) gradient descent updates are performed on the model copy to obtain a set of parameters optimized for this specific task. The goal of this step is to simulate the process of fine-tuning a model using a small amount of new data in a real-world application.

[0042] Preferably, the outer loop process includes using the adapted parameters obtained from the inner loop. In the mission The model performance is evaluated on the query set, and the loss is calculated. The meta-losses of all tasks in the entire meta-batch are aggregated, and this aggregated loss is calculated relative to the original meta-parameters. The gradient is calculated using the following formula:

[0043] in, The learning rate for the outer loop. For the distribution of tasks, For the updated meta-parameters, For query set.

[0044] In this invention, the adapted parameters obtained using the inner loop are used. In the mission The model performance is evaluated on a query set, and the loss is calculated. This process is repeated for a batch of different meta-tasks. The key step is to calculate the gradient of the total loss on the query set with respect to the original meta-model parameters θ. This gradient reflects how the initial parameters θ should be adjusted so that the model, after adaptation through the inner loop, performs better on the new task. The meta-model parameters θ are then updated based on this "meta-gradient".

[0045] In a second aspect, the electronic device includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores a computer program executable by the at least one processor, the computer program being executed by the at least one processor to enable the at least one processor to perform the method described in the first aspect for predicting drug solubility in a fasting state using a graph-gated attention network based on meta-learning.

[0046] Thirdly, the present invention provides a computer-readable storage medium according to the first aspect, the computer-readable storage medium storing computer instructions for causing a processor to execute and implement the method of the first aspect based on meta-learning graph gating attention network for predicting drug solubility in a fasting state.

[0047] Compared with the prior art, the present invention has at least the following beneficial effects:

[0048] 1. Through the AGGN architecture of multimodal feature fusion and optimization, this invention can deeply capture the complex internal and external molecular factors that affect solubility and has high prediction accuracy.

[0049] 2. Revolutionary Enhancement in Data Efficiency and Model Adaptability: The MAML training framework is one of the core innovations of this invention. It enables the model to possess powerful few-shot learning capabilities. For a new prediction task, thousands of experimental data points are no longer required; a small number of data points may suffice to fine-tune the model to a high accuracy, greatly saving experimental costs and time. Furthermore, the model can seamlessly and efficiently migrate from a data-rich domain (such as FaSSIF) to a data-sparse domain (such as FeSSIF), demonstrating flexibility and scalability unmatched by traditional models.

[0050] 3. Significant industrial application value: This invention provides a powerful, cost-effective, and highly adaptable computational tool. Using this method, at the very early stages of drug discovery, large-scale, high-precision solubility screening of millions of virtual compound libraries can be performed. This allows for the early identification and elimination of poorly soluble compounds, concentrating research resources on candidate molecules with greater drug potential, thereby improving the success rate of drug discovery. Furthermore, by replacing some expensive and time-consuming early physical experiments with computational simulation, it accelerates the process from lead compound discovery to preclinical candidates, shortening the research and development cycle. Attached Figure Description

[0051] Figure 1 A flowchart for the preparation and processing of the dataset.

[0052] Figure 2 This is a diagram of the model prediction architecture.

[0053] Figure 3 This is a flowchart of the MAML process.

[0054] Figure 4 This is a flowchart of the update process for the atomic and edge states based on AGGN.

[0055] Figure 5 This is a schematic diagram of the training process using FaSSIF as an example. Detailed Implementation

[0056] The technical solution of the present invention will be further described below with reference to the accompanying drawings and specific embodiments. However, the following examples are merely simplified examples of the present invention and do not represent or limit the scope of protection of the present invention. The scope of protection of the present invention is determined by the claims.

[0057] Example 1

[0058] This embodiment describes the acquisition, processing, and partitioning of the dataset.

[0059] The dataset used in this embodiment includes 1275 structurally diverse drug-like compounds and the original FaSSIF / FeSSIF / FaSSGF datasets. Data sources include public databases. These public databases were selected from large public cheminformatics databases such as ChEMBL and Drugbank, and literature data related to the solubility in biological media was screened and collected through keyword and structural searches.

[0060] The programming language used in this implementation is Python 3.9, the deep learning framework is PyTorch 1.10, the graph neural network library is PyTorch Geometric (PyG) 2.0, and the cheminformatics library is RDKit 2021.09.

[0061] The chemical structures of all compounds were stored in SMILES string format. The SMILES strings for each molecule were normalized using RDKit, including desalting, neutralization, and removal of solvent molecules, to generate unique, normalized SMILES representations and eliminate ambiguity in the structural representations. All solubility values ​​were converted to molar concentration (mol / L, M) and then the logarithm to base 10 was taken to obtain a uniform log value used as the model's prediction target.

[0062] The prepared compound dataset was randomly divided into a meta-training set and a meta-test set in an 8:2 ratio. The specific data is shown in Table 1. The dataset preparation and processing workflow is as follows: Figure 1 As shown

[0063] Meta-training set: Used for model training within the MAML framework. During training, this set is dynamically organized into a series of few-sample meta-tasks.

[0064] Meta-test set: This dataset is completely invisible to the model throughout the entire training process. It is only used to finally and unbiasedly evaluate the model's generalization performance after training is complete.

[0065] Table 1

[0066]

[0067] Example 2

[0068] This embodiment performs multimodal hybrid molecular feature engineering.

[0069] In this embodiment, all three characterizations of the molecules were automatically calculated and generated using the RDKit library. The specific process of multimodal hybrid molecular feature engineering is as follows: Figure 2 As shown.

[0070] (1) Molecular diagram characterization

[0071] The two-dimensional chemical structure of each drug molecule (usually represented by the string SMILES) is transformed into a mathematical diagram. The atoms correspond to the nodes (vertices) of the graph. ), the edges of the chemical bond graph ).

[0072] Atom (node) characterization includes atom number, number of atomic connections, atomic form charge, atomic hybridization, whether it is an aromatic bond, chiral label, number of hydrogen bonds, and relative atomic mass. Each atomic attribute is encoded as a one-hot vector and then concatenated. The initial node feature dimension is 134. The specific encoding method is as follows:

[0073] The atomic number is encoded as the position in the list of atomic numbers allowed to appear in the compound molecule, where the maximum allowed element number is 100, and elements with atomic numbers greater than 100 are encoded as 101;

[0074] The number of atomic connections, the number of atomic hydrogen bonds, the hidden valence of atoms, and the formal charge of atoms are encoded by the number itself.

[0075] The atomic hybridization mode is encoded as the position of the corresponding hybridization mode in the hybridization mode list, which includes 5 hybridization modes SP, SP2, SP3, SP3D, and SP3D2, encoded as 1, 2, 3, 4, and 5 respectively;

[0076] The atom is encoded as 0 or 1 depending on whether it is an aromatic bond; 1 if it is, 0 if it is not.

[0077] The code for the atom is 0 or 1 depending on whether it is chiral; 1 if it is chiral and 0 if it is not.

[0078] The relative atomic mass of an atom is encoded as 0.01 times its value.

[0079] Chemical bond (edge) characterization: This utilizes chemical bond attributes, including bond type, ring formation, conjugation state, and 3D spatial configuration. Chemical bond attributes are encoded into one-hot vectors according to rules and then concatenated. The initial feature length is 14.

[0080] Chemical bond valence states include single bonds, double bonds, triple bonds, aromatic bonds, etc., which are coded as 1, 2, 3, 4 respectively, while other types are coded as 0;

[0081] The code is 0 or 1 depending on whether the chemical bond forms a ring; 1 for yes and 0 for no.

[0082] The code is coded as 0 or 1 depending on whether the chemical bond is in a conjugated state; 1 for yes and 0 for no.

[0083] The stereochemical configurations of this chemical bond include no stereochemical configuration, unknown configuration, Z-type (STEREOZ), E-type, cis configuration, and trans configuration, coded as 1, 2, 3, 4, 5, and 6 respectively, with the others being 0.

[0084] (2) Attention-gated graph network architecture

[0085] Attention-based message aggregation, followed by gating state updates and then molecular-level reading, involves alignment, weighting, and aggregation. The alignment calculation formula is as follows: ;in, For unnormalized attention scores, It is a learnable linear transformation layer. For target atoms The current state vector, Neighboring atoms The state vector, Neighboring atoms The edges. The weighted calculation formula is: ;in, This represents the attention weight.

[0086] The formula for calculating aggregation is: ;in, For context vectors, This is the context vector.

[0087] The formula for calculating the gating state update is: ; ;in, The target atom In the The state vector of the layer, It is a context vector aggregated from its neighbors. It is the first GRU cells that update the target atomic state at each layer. It is the first GRU cells that update edge states at each layer The target atom In the The updated state vector of the layer, Neighboring atoms In the The updated state vector of the layer, It is the edge In the The updated state vector of the layer. Specific parameters are shown in Table 2. The attention-gated graph network architecture is as follows: Figure 3 As shown, the status update process is as follows: Figure 4 As shown.

[0088] Table 2

[0089]

[0090] (3) Normalized descriptor vector

[0091] This method uses descriptors calculated from the Rdkit library to represent the physicochemical characteristics of molecules. All descriptors for all molecules in the dataset are calculated, and then a correlation matrix is ​​calculated. Descriptors with a correlation greater than 0.9 are removed, resulting in 175 descriptors, which are then normalized. Some representative descriptors are as follows: molecular weight (MolWt), heavy atom molecular weight (excluding hydrogen), HeavyAtomMolWt, topological polar surface area (TPSA), number of hydrogen bond donors (NumHDonors), number of hydrogen bond acceptors (NumHAcceptors), octanol-water partition coefficient (lipid-water partition coefficient) (MolLogP), number of rotatable bonds (NumRotatableBonds), and number of rings in the molecule (RingCount).

[0092] (4) Molecular fingerprint vector

[0093] Extended-Connectivity Fingerprints (ECFPs) are circular topological fingerprints that can be used for molecular representation, similarity search, and structure-property relationship modeling. This method uses the default identifier for ECFPs, which contains highly specific atomic information and can represent a large number of substructure features. This method uses the GetMorganFingerprintAsBitVect method from Rdkit to calculate ECFP4, with the function parameters set to radius=2 and nBits=1024.

[0094] The molecular fingerprint vector and the normalized descriptor vector are concatenated and fused, then fused with the molecular graph vector and input into a multilayer perceptron to obtain the predicted value. The molecular fingerprint vector... and normalized descriptor vector The splicing and merging process is performed as follows:

[0095] ;

[0096] in To obtain learnable parameters, an attention-gated graph network is then used to construct molecular graph vectors from the molecular graph. The final molecular representation vector is obtained by fusion. The specific method is as follows:

[0097] ;

[0098] Where N is the total number of atoms in the molecule. Let be the state vector of the i-th atom in the molecule;

[0099] .

[0100] Example 3

[0101] This embodiment uses model-independent parameter learning samples for training.

[0102] The process of model-independent meta-learning is as follows: Figure 3 As shown in Table 1, a total of 1275 compounds and their solubility data in three different media were loaded. The dataset was strictly divided into a meta-training set and a meta-test set. Model and optimizer initialization: The AGGN model was constructed, and the meta-parameters were initialized. Initialize the meta-optimizer Adam and set the inner loop learning rate. outer loop learning rate .

[0103] Meta-tasks are constructed from the three media in the meta-training set: FaSSIF task FeSSIF task FaSSGF task .

[0104] Constructing Meta-batches: In each training iteration, a meta-batch of tasks is sampled from the task distribution described above. In FaSSIF model training, a meta-batch can consist of 8 tasks. The task composition, in the training of FeSSIF and FaSSGF, consists of 4 meta-batches each. Tasks and 4 Task composition. Each task consists of a support set (K=10) and a query set (10 samples) sampled from data in its corresponding medium.

[0105] First, train the FaSSIF model and perform a two-layer loop optimization: in the inner loop, for each meta-task in the meta-batch, copy the current meta-model. Using the support set data for this task, the model replica is updated three times with gradient descent to obtain a model that quickly adapts to this specific task. In the outer loop, the adapted models obtained in the inner loop are used. The performance of each task is evaluated on its respective query set, and the loss (meta-loss) is calculated. The meta-losses of all tasks in the entire meta-batch are aggregated, and this aggregated loss is calculated relative to the original meta-parameters. The gradient (meta-gradient) is used to update the global meta-model parameters using the meta-optimizer. After 30 epochs of training, the final model was obtained. The final prediction model for the FaSSIF training task .

[0106] Will The initialization parameters for the FeSSIF and FaSSGF training tasks are similarly trained using the methods described above, ultimately yielding the prediction models for FeSSIF and FaSSGF. and The parameters for model-independent meta-learning are shown in Table 3. The training process in this embodiment is performed on a high-performance computing server equipped with an NVIDIA 4090 24GB GPU. Under the above hyperparameter configuration, complete MAML training on the meta-training sets of FaSSIF, FeSSIF, and FaSSGF requires 30 epochs and takes approximately 6 hours in total. The specific process is as follows: Figure 5 As shown.

[0107] Table 3

[0108]

[0109] Finally, the performance of the prediction model was validated using the meta-test sets for each task, using RMSE (Root Mean Square Error) and R... 2 (R-squared, coefficient of determination) was used as the evaluation metric. Using the independent test set of 251 compounds listed in Table 1, the model prediction performance on the FaSSIF test set was RMSE and R... 2 The values ​​were 0.516 and 0.731 respectively; on the FeSSIF test set, the model's prediction performance was RMSE and R. 2 The values ​​were 0.583 and 0.697 respectively; on the FaSSGF test set, the model's prediction performance was RMSE and R. 2 The values ​​are 0.597 and 0.688, respectively. This demonstrates that the method of this invention profoundly captures the complex internal and external molecular factors affecting solubility, and has high prediction accuracy.

[0110] In summary, this invention provides a high-throughput, high-precision, and data-efficient in vitro (insilico) screening platform. This platform can rapidly predict the solubility of candidate drugs in various key biologically relevant media, such as simulated gastric juice, simulated intestinal juice, and simulated intestinal juice after a meal, with high predictive accuracy. By accurately assessing the solubility characteristics of drugs early in the research and development process, pharmaceutical companies can more effectively screen and optimize lead compounds, prioritize candidates with good pharmacokinetic potential, thereby significantly reducing the risk of failure in later clinical trials, shortening the drug development cycle, and saving substantial R&D costs.

[0111] The applicant declares that the above description is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Those skilled in the art should understand that any changes or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention fall within the protection and disclosure scope of the present invention.

Claims

1. A method for predicting drug solubility in a fasting state using a graph-gated attention network based on meta-learning, characterized in that, The method includes using multimodal hybrid molecular feature engineering to integrate molecular characterization information into an architecture model, and using model-independent meta-learning samples for training; (1) The multimodal hybrid molecular feature engineering includes: inputting the SMILES string of the compound to be tested to obtain a molecular graph, molecular fingerprint vector and normalized descriptor vector; using an attention-gated graph network to construct a molecular graph vector from the molecular graph; splicing and fusing the molecular fingerprint vector and the normalized descriptor vector; and then fusing them with the molecular graph vector and inputting them into a multilayer perceptron to obtain a predicted value. (1.1) The attention-gated graph network architecture model includes: attention-based message aggregation, followed by gating state update and molecular-level readout; The calculation process for the attention message aggregation includes alignment, weighting, and aggregation; the alignment calculation formula is as follows: ;in, For unnormalized attention scores, It is a learnable linear transformation layer. For target atoms The current state vector, Neighboring atoms The state vector, Neighboring atoms The edge; The weighting calculation formula is as follows: ;in, For attention weights, For nodes With nodes Unnormalized attention scores for connected edges; The calculation formula for the aggregation is: ;in, For context vectors, For context vectors; The formula for calculating the gate state update is: ; ;in, The target atom In the The state vector of the layer, It is a context vector aggregated from its neighbors. It is the first GRU cells that update the target atomic state at each layer. It is the first GRU cells that update edge states at each layer The target atom In the The updated state vector of the layer, Neighboring atoms In the The updated state vector of the layer, It is the edge In the The updated state vector of the layer; (2) The model-independent meta-learning includes constructing a meta-task and then using a two-layer loop to optimize it to obtain the final prediction model; The meta-task includes any one or a combination of at least two of the following: a fasting state simulating intestinal fluid task, a eating state simulating intestinal fluid task, or a fasting state simulating gastric fluid task.

2. The method for predicting drug solubility in a fasting state using a graph-gated attention network based on meta-learning according to claim 1, characterized in that, In step (1), the construction of the molecular graph includes converting the SMILES string into... Where V represents an atom, corresponding to a node in the graph, and E represents a chemical bond, corresponding to an edge in the graph.

3. The method for predicting drug solubility in a fasting state using a graph-gated attention network based on meta-learning, as described in claim 1 or 2, is characterized in that... In step (1), the construction of the molecular fingerprint vector includes construction using extended connectivity fingerprints; Preferably, in step (1), the construction of the normalized descriptor vector includes obtaining it by normalizing one-dimensional and two-dimensional global molecular descriptors.

4. The method for predicting drug solubility in a fasting state using a graph-gated attention network based on meta-learning according to any one of claims 1-3, characterized in that, In step (1), the fusion includes combining the molecular fingerprint vector and normalized descriptor vector The splicing and merging process is performed as follows: ； in To obtain learnable parameters, an attention-gated graph network is then used to construct molecular graph vectors from the molecular graph. The final molecular representation vector is obtained by fusion. The specific method is as follows: ； Where N is the total number of atoms in the molecule. Let be the state vector of the i-th atom in the molecule; 。 5. The method for predicting drug solubility in a fasting state using a graph-gated attention network based on meta-learning according to any one of claims 1-4, characterized in that, In step (2), the construction of the meta-task includes constructing the meta-task and dividing the data of the meta-task into a support set and a query set.

6. The method for predicting drug solubility in a fasting state using a graph-gated attention network based on meta-learning according to any one of claims 1-5, characterized in that, In step (2), the double loop includes using an inner loop for rapid task adaptation and an outer loop for meta-knowledge learning.

7. The method for predicting drug solubility in a fasting state using a graph-gated attention network based on meta-learning according to claim 6, characterized in that, The inner loop process includes replicating the meta-model of the meta-task, performing gradient descent updates 1-5 times using the support set of that task, and obtaining a set of parameters optimized for that specific task. And make fine adjustments, the specific calculation process is as follows: ;in, For the meta-task, For the task loss function on, For the inner loop learning rate, To support the set.

8. The method for predicting drug solubility in a fasting state using a graph-gated attention network based on meta-learning according to claim 6 or 7, characterized in that, The outer loop process includes using the adapted parameters obtained from the inner loop. In the mission The model performance is evaluated on the query set, and the loss is calculated. The meta-losses of all tasks in the entire meta-batch are aggregated, and this aggregated loss is calculated relative to the original meta-parameters. The gradient is calculated using the following formula: ;in, The learning rate for the outer loop. For the distribution of tasks, For the updated meta-parameters, For query set.

9. An electronic device, characterized in that, The electronic device includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores a computer program executable by the at least one processor, the computer program being executed by the at least one processor to enable the at least one processor to perform the method of predicting fasting drug solubility using a graph-gated attention network based on meta-learning as described in any one of claims 1-8.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions that, when executed by a processor, implement the method of predicting drug solubility in a fasting state using a graph-gated attention network based on meta-learning, as described in any one of claims 1-8.

Images