Method and system for predicting admet properties of drug-like small molecules

Abstract

The present application relates to the technical field of drug research and development, in particular to a drug small molecule ADMET property prediction method and system, the drug small molecule ADMET property prediction method constructs a molecular graph by processing SMILES string, extracts atomic and bond features and calculates molecular descriptors;Using the directed message passing mechanism, while excluding the interference of reverse edges, combining the learnable gating weight guided by atomic chemical categories to realize the residual enhanced message update;Introducing the position coding of fusion laplacian feature vector and topological distance, modeling the distance constraint attention of atomic hidden vector to capture long-range dependence;After aggregation, the molecular descriptors are spliced to form a composite representation, and the ADMET prediction result is output by the feedforward network with residual connection. The method effectively improves the prediction accuracy and generalization ability, and is suitable for early screening in drug research and development.

Description

Technical Field

[0001] This invention relates to the field of drug development technology, specifically to a method and system for predicting the properties of small drug molecules, ADMET. Background Technology

[0002] In the early stages of drug development, accurately predicting the absorption, distribution, metabolism, excretion, and toxicity (ADMET) properties of candidate small molecules is crucial for reducing the risk of clinical failure, shortening the development cycle, and controlling costs. In recent years, with the development of deep learning technology, molecular representation learning methods based on graph neural networks (GNNs) have been widely applied to ADMET property prediction tasks. These methods typically represent molecules as graph structures, with atoms as nodes and chemical bonds as edges. They aggregate neighbor information through message passing mechanisms to update atomic representations, and finally obtain molecular-level features through global pooling for downstream prediction. For example, CN120072102A proposes a joint molecular characterization strategy, fusing molecular graphs, SMILES sequences, and molecular fingerprints to improve the stability of ADMET predictions; CN120823910A constructs a multimodal, multi-scale feature fusion framework, combining graph convolutional networks (GCNs), gated recurrent units (GRUs), and attention mechanisms to achieve high-precision predictions on 73 ADMET properties. Furthermore, CN120183541A enhances the interpretability and predictive reliability of the model by integrating one-dimensional sequence features and two-dimensional graph features and introducing visualized attention weights; while CN121905339A effectively improves the generalization ability of specific ADMET properties such as solubility by combining graph-gated attention networks with meta-learning. Other studies, such as CN113707235B, utilize a self-supervised pre-training strategy to enhance the semantic consistency between molecular graphs and SMILES sequences through contrastive learning, thereby improving the prediction performance of downstream properties.

[0003] However, existing message-passing-based deep learning models still have significant shortcomings in ADMET prediction.

[0004] On the one hand, as the number of message passing layers increases, after multiple rounds of nonlinear transformation, the original key chemical structural information (such as pharmacophore skeleton, aromatic ring system, hydrogen bond donor / acceptor sites, etc.) is easily diluted or even covered layer by layer, resulting in the loss of local structural features with clear chemical significance in deep characterization and weakening the model's sensitivity to functional groups.

[0005] On the other hand, in multi-task ADMET prediction, different properties (such as blood-brain barrier penetration, hERG inhibitory toxicity, and liver microsomal metabolic stability) depend on the isomerization of molecular structure response mechanisms. However, existing models generally employ a uniform message passing and aggregation strategy, lacking the ability to dynamically adapt to task-specific structural patterns. This easily leads to information interference between tasks or main task shift, thereby reducing the overall accuracy and robustness of the prediction. These problems severely restrict the application effectiveness of deep learning models in practical drug discovery scenarios such as high-throughput virtual screening and lead compound optimization. Summary of the Invention

[0006] To address the shortcomings of existing deep learning models in predicting molecular ADMET properties—namely, the loss of key chemical structural information due to increasing network depth and the lack of dynamic adaptation to task-specific structural response mechanisms in multi-task scenarios—leading to insufficient prediction accuracy and robustness, this paper proposes a method for predicting small molecule ADMET properties. This method integrates chemoselectivity residual gating and distance-aware self-attention mechanisms, using a Directed Message Passing Neural Network (DMPNN) as its core framework. It integrates attention mechanisms, molecular representation learning, and deep residual network structures, and fuses these with traditional molecular descriptors. Through multi-dimensional feature extraction and information complementarity, the method significantly improves the model's predictive performance and generalization ability, thereby achieving comprehensive and accurate prediction of small molecule ADMET properties.

[0007] This invention provides a method for predicting the ADMET properties of small drug molecules, comprising: The process involves obtaining the original SMILES string of a small molecule drug sample and processing it. A molecular graph structure is constructed based on the processed SMILES string, extracting atomic features (nodes) and bond features (edges), and calculating the corresponding molecular descriptor vectors. The atomic and bond features are processed using a directed message passing mechanism, undergoing linear transformation to obtain initial messages. Messages are passed along directed edges in each round of message passing, excluding reverse edges during neighbor message aggregation. In each round of message passing, based on the atomic chemical categories determined by the atomic features, learnable gating weights are used to weight and fuse the previous round's input message with the current aggregated message to obtain residual enhanced messages. Finally, based on the last round's update message, various... The method involves: latent vector representation of atoms; attention-weighted modeling of the latent vector representation of each atom based on position encoding and molecular topological distance constraints to characterize long-range dependencies within the molecule; aggregation of the attention-weighted modeled atom latent vector representations to obtain molecular-level representation vectors; concatenation of the molecular-level representation vectors with the molecular descriptor vectors to obtain composite molecular representation vectors; training of an ADMET property prediction model composed of the directed message passing mechanism, the attention-weighted modeling, and the feedforward network structure based on the composite molecular representation vectors of tagged drug small molecule samples; mapping of the composite molecular representation vectors of the test molecule using the trained ADMET property prediction model; and outputting the ADMET property prediction results of the test molecule by combining residual connections.

[0008] Furthermore, the processing of the original SMILES string includes: sequentially performing charge removal, small fragment removal, chiral information elimination, and tautomerization on the original SMILES string to obtain a standardized SMILES string; and using the standardized SMILES string as a deduplication matching key to filter missing values ​​and remove duplicates to obtain a processed SMILES string.

[0009] Furthermore, the atomic features are represented by one-hot encoding to indicate element type, atomic charge, atomic aromaticity, and the number of attached hydrogen atoms; the bond features include bond types such as single bonds, double bonds, triple bonds, and aromatic bonds; and the molecular descriptor vector includes a physicochemical descriptor vector composed of molecular weight, lipid-water partition coefficient, topological polar surface area, and the number of hydrogen bond donors and acceptors.

[0010] Furthermore, in each round of message passing, based on the atomic chemical category determined by atomic features, the previous round input message and the current aggregated message are weighted and fused using learnable gating weights to obtain a residual enhanced message. The latent vector representation of each atom is generated based on the last round update message, including: determining the atomic chemical category based on the element type, atomic charge, and atomic aromaticity represented by the atomic features; calculating the gating weights using the Sigmoid activation function, wherein the network parameters used to calculate the gating weights are initialized using a Xavier normal distribution to maintain the inter-layer input-output direction during network transmission. The message is consistent with the previous round's input message; the retention ratio of the previous round's input message is determined according to the gating weight, and weighted and added to the current aggregated message to obtain the message with residual enhancement; the message with residual enhancement is processed by linear transformation and added to the original input message of the current message passing layer, and the message update of the current round is completed by activation function and Dropout regularization; for each atom in the molecule, the messages updated in the last round of all directed edges connected to it are aggregated to obtain the atomic-level total message, and the atomic features of the atom are concatenated with the atomic-level total message, and the latent vector representation of each atom is generated after mapping processing.

[0011] Furthermore, the attention-weighted modeling of the latent vector representations of each atom based on position encoding and molecular topological distance constraints includes: transforming the molecular graph structure into position encoding features to characterize each atom, wherein the position encoding features consist of a Laplacian feature vector capturing the global structure and a center distance feature capturing the local position, and the center distance feature is the topological distance from the atom to the molecular center; generating a query matrix, a bond matrix, and a value matrix through linear mapping based on the atomic latent vector representation containing the position encoding features, and splitting them into multiple attention heads for parallel computation; calculating self-attention weights based on the query matrix and the bond matrix, and applying a decay penalty constraint to the self-attention weights according to the molecular topological distance to suppress the allocation of attention to chemically unrelated long-range atom pairs; and performing a weighted summation of the value matrix according to the distance-constrained self-attention weights to obtain the attention-enhanced atomic latent vector representation.

[0012] Furthermore, the aggregation of the attention-weighted modeled atomic latent vector representations to obtain a molecular-level representation vector, and the concatenation of the molecular-level representation vector with the molecular descriptor vector to obtain a composite molecular representation vector, includes: aggregating the latent vector representations of all atoms within the same molecule using a preset summation, averaging, or normalization aggregation method to obtain the molecular-level representation vector with a fixed dimension; concatenating the molecular-level representation vector with the molecular descriptor vector, and performing linear mapping processing on the concatenated vector to complete feature fusion, obtaining the molecular representation vector of the fused descriptor; if the current drug small molecule sample contains dosage information, then extracting a dosage feature vector, and concatenating the dosage feature vector as an additional scalar feature with the molecular representation vector of the fused descriptor to obtain the composite molecular representation vector.

[0013] Furthermore, the mapping process of the composite molecular representation vector of the molecule to be tested using the trained ADMET property prediction model includes: sequentially performing Dropout regularization and linear mapping on the composite molecular representation vector to obtain initial hidden features; performing at least one residual connection process on the initial hidden features, each residual connection process sequentially including layer normalization, linear mapping, activation function, Dropout regularization, and linear mapping, and directly adding the output of the next linear mapping to the features before performing the residual connection process, maintaining the complete transmission of the molecular-level representation vector without introducing atom-type gating; sequentially performing layer normalization, activation function, Dropout regularization, and linear mapping on the features after the residual connection process to obtain the hidden feature vector used to output the prediction result.

[0014] Furthermore, the step of outputting the ADMET property prediction result of the molecule to be tested includes: performing a final linear mapping operation on the hidden feature vector to map the high-dimensional hidden features to the output dimension of the target task, generating a predicted value; in response to the prediction task of the molecule to be tested being a classification task, performing a nonlinear activation operation on the predicted value, and outputting the ADMET property prediction result representing the classification probability; in response to the prediction task of the molecule to be tested being a regression task, outputting the predicted value as a continuous ADMET property prediction result.

[0015] This invention also provides a drug small molecule ADMET property prediction system, comprising: The data acquisition and preprocessing unit is used to acquire the original SMILES string of the drug small molecule sample and process the original SMILES string. The graph construction and descriptor calculation unit is used to construct a molecular graph structure based on the processed SMILES string, extract atomic features with atoms as nodes and bond features with chemical bonds as edges, and calculate the corresponding molecular descriptor vectors. The directed message passing unit is used to process the atomic features and bond features based on the directed message passing mechanism, perform linear transformation on the atomic features and bond features to obtain the initial message, pass the message along the directed edge in each round of message passing, and exclude the passing of the reverse edge when the neighbor message is aggregated. The chemically selective gating fusion unit is used to perform weighted fusion of the previous round input message and the current aggregated message in each round of message passing, based on the atomic chemical category determined by atomic features and using learnable gating weights, to obtain the residual enhancement message, and to generate the latent vector representation of each atom based on the last round update message; Topologically constrained attention units are used to perform attention-weighted modeling of the latent vector representations of each atom based on position encoding and molecular topological distance constraints, in order to characterize long-range dependencies within the molecule; The aggregation and feature fusion unit is used to aggregate the atomic latent vector representation after attention-weighted modeling to obtain a molecular-level representation vector, and to concatenate the molecular-level representation vector with the molecular descriptor vector to obtain a composite molecular characterization vector, and to further concatenate the dose feature vector when dose information is present. The model training and prediction unit is used to train an ADMET property prediction model based on labeled drug small molecule samples, and to use the trained ADMET property prediction model to map the composite molecular characterization vector of the molecule to be tested, and to output the ADMET property prediction result of the molecule to be tested by combining the residual connection.

[0016] The present invention also provides a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor implements the steps of the method described above when executing the program.

[0017] This invention has at least the following technical effects: 1. This invention features targeted optimizations in its model architecture. A chemically selective residual gating module is introduced into the message passing layer. As the network depth increases, although each layer performs nonlinear transformations on atomic features, residual connections allow the network to directly retain the original atomic feature information, effectively avoiding the loss of key features caused by multiple layers of nonlinear transformations. The residual module of this invention does not directly add the input from the previous layer to the aggregated message. Instead, it dynamically adjusts the fusion ratio of the two based on the chemical category of the atom through learnable gating weights. This allows the model to dynamically determine which layer's "original structural information" (key skeleton atoms) to retain based on the chemical category of the atom, thereby significantly improving the completeness and effectiveness of feature transmission. Furthermore, residual connections are also introduced into the readout layer to ensure that key information at the atomic level is not completely covered by the nonlinear transformations of higher-level networks. Simultaneously, the readout layer integrates a distance-aware self-attention mechanism and introduces a position encoding module. It captures the global structure through Laplacian feature vectors and captures local positions through distance to the center; these two complement each other, giving the model both permutation invariance and structure awareness. This mechanism constrains attention weights to molecular topological distances, preventing attention from being diverted to chemically unrelated long-range atom pairs, thereby more accurately capturing pharmacophore interactions related to ADMET properties and enabling precise discovery of the correlation between molecular structure and specific properties. 2. This invention ensures data quality during the dataset processing stage. A dataset containing nearly 70,000 labeled data points, covering 55 ADMET property prediction tasks, was constructed and systematically and thoroughly cleaned. The data cleaning process includes, but is not limited to, removing duplicate data, handling missing values, and resolving outliers to ensure data quality and accuracy. This dataset comes from the largest and most comprehensive publicly available ADMET database, ensuring the high quality of the dataset used in this invention and providing a solid data foundation for stable model training and reliable prediction. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart of a method for predicting the properties of small drug molecules (ADMET) according to an embodiment of the present invention.

[0020] Figure 2 This is a schematic diagram of the architecture of a drug small molecule ADMET property prediction model in one embodiment of the present invention.

[0021] Figure 3 This is a block diagram of a drug small molecule ADMET property prediction system according to an embodiment of the present invention.

[0022] Figure 4 This is a comparison chart of the prediction results of classification tasks in embodiments of the present invention.

[0023] Figure 5 This is a comparison chart of the regression task prediction results in embodiments of the present invention. Detailed Implementation

[0024] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application. Furthermore, it should be understood that the specific embodiments described herein are only for illustration and explanation of this application and are not intended to limit this application.

[0025] It should be noted that the order of description of the following embodiments is not intended to limit the preferred order of the embodiments of this application. Furthermore, the descriptions of each embodiment in the following embodiments have their own emphasis; for parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0026] While some deep learning models have shown success in predicting drug ADMET properties, their application still faces several challenges. Each layer of a deep learning model performs a non-linear transformation on the input atomic features. As the network deepens, the original atomic features undergo multiple transformations, potentially diluting or even completely covering early key information layer by layer. This leads to the loss of chemically meaningful original structural information in the deep representation. Furthermore, ADMET properties involve multiple dimensions such as absorption, distribution, metabolism, excretion, and toxicity. Different properties exhibit different chemical structural response mechanisms, placing higher demands on multi-task learning. Current models struggle to simultaneously accommodate the heterogeneous feature requirements of various properties when performing multi-task predictions, easily leading to information interference or task-specific shifts, thus affecting the overall accuracy and robustness of the predictions. Therefore, designing a deep learning architecture that can preserve original key information and efficiently balance multi-task learning remains a crucial problem to be solved in this field.

[0027] Example 1.

[0028] like Figure 1 and Figure 2 As shown, a method for predicting the properties of small drug molecules ADMET includes the following steps.

[0029] S1: Data Acquisition and SMILES String Processing: Datasets containing drug small molecule SMILES strings and corresponding ADMET experimental labels were collected from the TDC, SIDER, and ChEMBL databases, covering 55 prediction tasks, including AMES mutagenicity, BBB_Martins blood-brain barrier penetration, hERG cardiotoxicity, CYP450 enzyme inhibition / substrate activity, multiple nuclear receptor pathway activities, skin reactions, liver injury, nephrotoxicity, hematologic toxicity, phototoxicity, and neurotoxicity. Neurotoxicity data included dose information in mg / kg. The original SMILES strings were sequentially subjected to charge removal, fragment removal, chiral information removal, and tautomerism normalization to obtain normalized SMILES strings. These normalized SMILES strings were then used as deduplication matching keys for missing value filtering and duplicate removal to obtain processed SMILES strings. Specifically, the RDKit method was used to normalize the SMILES strings, and duplicate data in the normalized SMILES strings were removed to obtain 55 ADMET datasets.

[0030] S2: Molecular Graph Construction and Feature Extraction: Based on the processed SMILES string, a molecular graph structure is constructed using RDKit, where atoms are nodes and chemical bonds are edges, corresponding to nodes and edges in a graph neural network, respectively. Atomic features are represented by one-hot encoding, indicating element type, atomic charge, aromaticity, and the number of attached hydrogen atoms, resulting in four one-hot vectors representing the four atomic properties. Bond features include bond types such as single, double, triple, and aromatic bonds. Simultaneously, RDKit is used to calculate molecular descriptor vectors for physicochemical properties. These vectors consist of physicochemical descriptor vectors composed of molecular weight, lipid-water partition coefficient, topological polar surface area, number of hydrogen bond donors, and number of hydrogen bond acceptors.

[0031] S3: Directed Message Passing and Residual Augmentation: Linear transformations are performed on atomic and bond features to generate initial messages. In each round of message passing, messages are passed along directed edges, and reverse edge passing is excluded during neighbor message aggregation to avoid information oscillations between chemical bonds. Atom chemical categories are determined based on element type, atomic charge, and atomic aromaticity represented by atomic features. Learnable gating weights are calculated using the Sigmoid activation function, with network parameters initialized using a Xavier normal distribution to maintain consistent input-output variance between layers. The input messages from the previous round and the current aggregated messages are weighted and fused according to these gating weights to obtain residual augmented messages. These residual augmented messages are linearly transformed and added to the original input messages of the current layer, and message updates are completed using activation functions and Dropout regularization. After the final round of message passing, update messages from all connected directed edges of each atom are aggregated to obtain the atomic-level total message. The original atomic features of the atom are concatenated with the atomic-level total message, and a latent vector representation for each atom is generated through mapping.

[0032] S4: Attention-Weighted Modeling of Long-Range Dependencies: The molecular graph structure is transformed into position-encoded features, which consist of Laplacian feature vectors capturing the global structure and center distance features capturing local positions, where the center distance feature is the topological distance from the atom to the molecular center. Based on the atomic latent vector representation containing position-encoded features, a query matrix, bond matrix, and value matrix are generated through linear mapping and split into multiple attention heads for parallel computation. Self-attention weights are calculated based on the query matrix and bond matrix, and a decay penalty constraint is applied to these self-attention weights according to the molecular topological distance to suppress the allocation of attention to chemically unrelated long-range atom pairs. The value matrix is ​​weighted and summed according to the distance-constrained self-attention weights to obtain the attention-enhanced atomic latent vector representation.

[0033] S5: Molecular-level representation generation and composite vector construction: Using a summation aggregation method, all attention-weighted modeled atomic latent vector representations within the same molecule are aggregated to obtain a fixed-dimensional molecular-level representation vector; wherein, the molecular-level representation vector is also called the molecular representation tensor; this molecular-level representation vector is concatenated with the aforementioned molecular descriptor vector, and the concatenated vector is linearly mapped to complete feature fusion, resulting in the molecular representation vector of the fused descriptor; for neurotoxicity samples containing dose information, the dose value in the original data is extracted separately as a new column of indicators, i.e., the dose feature vector is extracted, and this dose feature vector is concatenated with the molecular representation vector of the fused descriptor as an additional scalar feature to obtain the composite molecular representation vector.

[0034] S6: ADMET Property Prediction Model Training and Inference: Based on the composite molecular characterization vector of labeled drug small molecule samples, the ADMET property prediction model, which consists of a directed message passing mechanism, attention-weighted modeling, and a feedforward network structure, is trained end-to-end.

[0035] Specifically, during training, the composite molecular representation vector is sequentially subjected to Dropout regularization and linear mapping to obtain initial hidden features. These initial hidden features are then subjected to two residual connection processes. Each residual connection process sequentially includes layer normalization, linear mapping, activation function, Dropout regularization, and linear mapping. The output of the second linear mapping is directly added to the unprocessed features, maintaining the complete transmission of the molecular-level representation vector without introducing atom-type gating. The features after residual connection processing are then sequentially subjected to layer normalization, activation function, Dropout regularization, and linear mapping to obtain the hidden feature vector used to output the prediction results. For classification tasks, a binary cross-entropy loss function is used; for regression tasks, a mean squared error loss function is used.

[0036] S7: Prediction Result Output: Perform the final linear mapping operation on the hidden feature vector to map it to the output dimension of the target task and generate the predicted value; if the prediction task is a classification task, perform a Sigmoid non-linear activation operation on the predicted value and output the ADMET property prediction result representing the probability of toxicity or non-toxicity; if the prediction task is a regression task, directly output the predicted value as a continuous ADMET property prediction result, such as quantitative indicators such as IC50, TD50 or clearance rate.

[0037] In a preferred embodiment, the polymerization method may also employ averaging or normalization instead of summation to accommodate the molecular representation requirements at different scales.

[0038] In a preferred embodiment, the dose feature is only included in the composite molecular characterization vector in the neurotoxicity prediction task because it has a sufficient sample size to support dose-toxicity relationship modeling, while other toxicity tasks do not introduce this feature due to the lack of reliable dose data.

[0039] In a preferred embodiment, the gating weights are initialized using a Xavier normal distribution. Compared to Kaiming normal initialization or orthogonal initialization, this method can more effectively avoid Sigmoid gating saturation, maintain the stability of the forward activation value distribution and the reverse gradient flow, and thus improve the ability to retain the atomic structure information of key pharmacophores.

[0040] In a preferred embodiment, the topological distance decay penalty constraint introduced in the self-attention mechanism significantly suppresses spurious attention assignments between chemically unrelated long-range atom pairs, making the model more focused on local pharmacophores and mid-range interactions related to ADMET properties, thereby improving prediction specificity.

[0041] Example 2.

[0042] like Figure 3 As shown, a drug small molecule ADMET property prediction system includes a data acquisition and preprocessing unit, a graph construction and descriptor calculation unit, a targeted message passing unit, a chemical selectivity gating fusion unit, a topological constraint attention unit, an aggregation and feature fusion unit, and a model training and prediction unit.

[0043] The data acquisition and preprocessing unit retrieves the raw SMILES strings of drug small molecule samples from the TDC, SIDER, and ChEMBL databases. The TDC database provides strings including AMES, BBB_Martins, Bioavailability_Ma, CYP series metabolic enzyme inhibition / substrate prediction, Carcinogens_Lagunin, ClinTox, DILI, HIA_Hou, NR series nuclear receptor pathway activity, PAMPA_NCATS, Pgp_Broccatelli, SR series stress response pathway activity, Skin_Reaction, hERG, Caco2_Wang, and Clearance_He. This unit utilizes 41 ADMET task datasets, including patocyte_AZ, Clearance_Microsome_AZ, Half_Life_Obach, HydrationFreeEnergy_FreeSolv, LD50_Zhu, Lipophilicity_AstraZeneca, PPBR_AZ, Solubility_AqSolDB, and VDss_Lombardo. The SIDER and ChEMBL databases supplement this with experimental data for four toxicity indicators: Hematotoxicity, Nephrotoxicity, Phototoxicity, and Neurotoxicity. This unit uses RDKit to standardize the original SMILES strings, including removing charges, small fragments, chiral information, and performing tautomerism standardization. Missing value filtering and duplicate removal are also performed, ultimately constructing an ADMET dataset containing approximately 70,000 samples across 55 categories. For the neurotoxicity task, this unit also extracts dose information in mg / kg from animal experimental descriptions and retains it as an independent feature.

[0044] The graph construction and descriptor calculation unit receives the standardized SMILES string, calls RDKit to parse and generate a molecular graph structure, where atoms are nodes and chemical bonds are edges. It also generates an adjacency matrix and an edge list to fully represent the molecular topological connections. This unit further extracts the atomic features of each atom (including atom type, hybridization state, formal charge, and whether it is an aromatic atom) and the bond features of each chemical bond (including bond type, whether it is a conjugated bond, and whether it is an aromatic bond). Based on RDKit, it calculates a physicochemical molecular descriptor vector, including molecular weight, logP, number of hydrogen bond donors / acceptors, number of rotatable bonds, and topological polar surface area, totaling 200 dimensions. If dosage information exists, the dosage value is stored separately as a scalar feature.

[0045] The directed message passing unit is based on a standard directed message passing neural network architecture. First, it performs linear transformations on atomic and bond features and generates initial messages using the ReLU activation function. In each round of message passing, messages are propagated along directed edges. That is, each undirected chemical bond is split into two directed edges in opposite directions, and messages are only passed along the specified direction. When aggregating neighbor messages, messages from the opposite edges are explicitly excluded to avoid information oscillation at both ends of the bond, which would cause redundancy. This mechanism ensures that each round of message passing focuses on the unidirectional information flow of the local chemical environment, improving the chemical rationality and stability of message passing.

[0046] The chemically selective gating fusion unit introduces a gating mechanism based on atomic chemical categories in each round of message passing: First, the chemical category of an atom is determined according to its type (such as carbon, nitrogen, oxygen, sulfur, halogen, etc.). Then, a learnable gating network generates weight coefficients between 0 and 1. These weights are used to weight and fuse the previous round's input message with the current aggregated message to form a residual enhanced message. The gating network adopts a Xavier normal initialization strategy, combined with a sigmoid activation function, to avoid gradient vanishing while maintaining the stability of the forward propagation variance. This allows the model to dynamically adjust the information retention ratio of different atom types—giving higher weights to key atoms in the pharmacophore (such as nitrogen on aromatic rings and carbon on metabolic sites) to retain their structural sensitivity, while allowing more aggregated information updates for non-key framework atoms. Finally, based on the message updated in the last round, combined with the original atomic features, a latent vector representation of each atom is generated through a linear layer and an activation function.

[0047] The topologically constrained attention unit applies a position-aware self-attention mechanism to the latent vector representations of all atoms: First, it captures the global topological structure of the molecule through Laplacian eigenvector encoding, and simultaneously combines the Euclidean distance from each atom to the molecular centroid to encode local position information. The two are concatenated and injected into the attention calculation as position encoding. During the multi-head self-attention process, the query, bond, and value matrices are generated by linear projection of the same latent vector, but the attention weights are constrained by the shortest path distance between atoms within the molecule—atom pairs that are too far apart (e.g., more than 6 hops) or chemically unrelated are subject to high penalty terms, thereby suppressing ineffective long-range interactions. This mechanism enables the model to accurately focus on long-range pharmacophore interactions related to ADMET properties (such as cation-π interactions in hERG channel blocking and synergistic effects of lipophilic regions in blood-brain barrier penetration), while maintaining invariance to the molecular arrangement order.

[0048] The aggregation and feature fusion unit first uses a summation aggregation method on the attention-weighted atomic latent vectors to generate a fixed-dimensional molecular-level representation vector; then it concatenates this vector with the aforementioned molecular descriptor vector to form a composite molecular representation vector; when the input sample contains dose information, the dose feature vector (scalar) is further concatenated to the end of the composite molecular representation vector to form the final enhanced representation used for prediction.

[0049] The model training and prediction unit trains an end-to-end prediction model based on labeled ADMET samples: for binary classification tasks (e.g., toxic / non-toxic), a binary cross-entropy loss function is used; for regression tasks (e.g., IC50, clearance rate), a mean squared error loss function is used; the model readout part adopts a feedforward network structure containing layer normalization, linear transformation, ReLU activation, Dropout, and residual connections to ensure that the global molecular features are not diluted during deep transmission; after training, the composite molecular representation vector of the test molecule is mapped and the ADMET property prediction results are output through the output layer combined with residual connections, including toxicity probability, metabolic stability score, permeability index, or continuous dose-effect value; in the neurotoxicity prediction task, due to the introduction of dose features, the model can output toxicity risk curves at different dose levels, realizing refined dose-toxicity relationship modeling.

[0050] For example, those skilled in the art can use the collected and processed dataset to train the model, with the Adam optimizer, the cross-entropy loss function for the classification task, a training epoch of 150, a maximum learning rate of 0.005, and a minimum learning rate of 0.0001, and the mean squared error (MSE) loss function for the regression task, with a training epoch of 150, a maximum learning rate of 0.005, and a minimum learning rate of 0.0001, to predict the ADMET properties of small molecules.

[0051] Specifically, TDC is a standardized platform focused on computational tasks in drug discovery and development, providing a variety of end-to-end datasets, including ADMET, whose data represents compound structures in SMILES string format; SIDER is a database that collects adverse reaction information for marketed drugs, focusing on drug toxicity labeling, with compounds provided in drug name and corresponding SMILES format; and ChEMBL is a large-scale bioactivity database containing rich data on drug metabolism and toxicity experiments, with compounds also stored in SMILES or standard chemical structure format. The acquired compound data includes compound SMILES and experimental values. Of the 55 datasets used in this invention, 41 (10 regression tasks and 31 classification tasks) are from the TDC database, covering predictive properties including: AMES (mutagenicity), BBB_Martins (blood-brain barrier penetration), Bioavailability_Ma (bioavailability), CYP1A2_Veith (CYP1A2 inhibition / substrate prediction), CYP2C19_Veith (CYP2C19 inhibition / substrate prediction), and CYP2C9_Substrate_CarbonMangels (CYP2C9 substrate prediction).CYP2C9_Veith (CYP2C9 inhibition / substrate prediction), CYP2D6_Substrate_CarbonMangels (CYP2D6 substrate prediction), CYP2D6_Veith (CYP2D6 inhibition / substrate prediction), CYP3A4_Substrate_CarbonMangels (CYP3A4 substrate prediction), CYP3A4_Veith (CYP3A4... Inhibition / Substrate Prediction), Carcinogens_Lagunin (Carcinogenicity), ClinTox (Clinical Toxicity), DILI (Drug-Induced Liver Injury), HIA_Hou (Human Intestinal Absorption), NR-AR-LBD (Androgen Receptor-Ligand Binding Domain), NR-AR (Androgen Receptor), NR-AhR (Aromatase), NR-Aromatase (Aromatase), NR-ER-LBD (Estrogen Receptor-Ligand Binding Domain), NR-ER (Estrogen Receptor), NR-PPAR-gamma (Nuclear Receptor Pathway), PAMPA_NCATS (Artificial Membrane Permeability), Pgp_Broccatelli (P-Glycoprotein), SR-ARE (Antioxidant Response Element Activity Prediction), SR-ATAD5 (ATAD5 Detection Pathway Activity Prediction), SR-HSE (Heat Shock Element Activity Prediction), SR-MMP (Mitochondrial Membrane Potential Prediction), SR-p53 (p53 Signaling Pathway Activity Prediction), Skin_Reaction (Skin The study examined various toxicological properties, including skin reaction, hERG (cardiotoxicity), Caco2_Wang (cell permeability), Clearance_Hepatocyte_AZ (hepatocyte clearance), Clearance_Microsome_AZ (microsome clearance), Half_Life_Obach (half-life), HydrationFreeEnergy_FreeSolv (hydration free energy), LD50_Zhu (median lethal dose), Lipophilicity_AstraZeneca (lipophilicity), PPBR_AZ (plasma protein binding), Solubility_AqSolDB (water solubility), VDss_Lombardo (volume of distribution), and four key toxicological properties verified in this invention: Hematotoxicity, Nephrotoxicity, Phototoxicity, and Neurotoxicity.

[0052] Furthermore, experimental data for four toxicity indicators—nephrotoxicity, hematologic toxicity, phototoxicity, and neurotoxicity—were collected from the SIDER and ChEMBL databases. All data were organized according to experimental type (animal experiments, cell experiments), species (human, mouse, rat), route of administration (oral, intravenous, intraperitoneal injection), and assay method.

[0053] Because each of the four toxicity indicators has both quantitative and qualitative experimental types, datasets for different tasks were constructed based on the data volume of each type of experiment. For each indicator, a quantitative prediction dataset (predicting continuous activity values ​​such as IC50 and TD50) and a qualitative classification dataset (predicting whether a compound is toxic or non-toxic) were constructed. Dosage information was extracted from all experimental data descriptions, mainly reflecting the dosage administered to animals and the observation of whether there was a clear toxic effect; the dosage unit was mg / kg. Phototoxicity data lacked dose-related data. Animal experimental data for toxicity, hematologic toxicity, and neurotoxicity contained dose information, but after collection, only a large amount of neurotoxicity data was sufficient to support model training. Therefore, only neurotoxicity data supported dose information input, and its use as a feature input was explored to improve the model's precision in predicting dose-toxicity relationships.

[0054] The original SMILES strings were standardized using RDKit, including charge removal, small fragment removal, chiral information removal, and tautomerism standardization. Based on this, the standardized SMILES strings were further processed by missing value filtering and duplicate removal. After these operations, an ADMET dataset containing approximately 70,000 data points across 55 categories was constructed, which will be used as the dataset for the ADMET property prediction model.

[0055] The graph structure features of the molecules were extracted from the original SMILES string using RDKit. In this molecular graph, atoms are defined as nodes, and the chemical bonds between atoms are defined as edges. To fully describe the molecular topology, both adjacency matrices and edge lists are used to represent the connections between atoms.

[0056] Based on this, the constructed molecular graph structure and its associated adjacency matrix and atomic attribute information are input into the target graph neural network model. Through the information transmission and aggregation mechanism of the graph neural network, a corresponding hidden vector representation is learned for each node (i.e., atom) in the molecular graph. The physicochemical molecular descriptor of each molecule, calculated using RDKit, serves as additional feature input. For dose features, the dose values ​​in the original data are extracted separately as a new column of indicators.

[0057] The model architecture of this invention comprises two modules: a feature extraction module and a readout module.

[0058] The feature extraction module employs a residual-enhanced directed message-passing neural network, and its core process includes: 1. Linear transformation of input features: Perform linear transformation on atomic and bond features, and obtain the initial message through an activation function; 2. In each round of message passing, perform the following operations: Neighbor message aggregation: Calculates the total message received by each atom from its neighboring bonds based on the atom-to-bond mapping. D-MPNN propagates messages along directed edges, and the aggregation of each edge explicitly excludes reverse edges, preventing information from oscillating back and forth between the two ends of the chemical bond. Chemoselective residual connections: ADMET properties (such as blood-brain barrier penetration and hERG toxicity) are often determined by specific pharmacophores or long-range interactions. General residual connections cannot distinguish key structural atoms, leading to the dilution of crucial structural information during deep-layer propagation. To overcome the lack of chemoselectivity in general residual connections, this invention introduces a chemoselective residual gating module. Instead of directly adding the input from the previous layer to the aggregation message using a general residual method, the fusion ratio is dynamically adjusted based on the chemical category of the atoms through learnable gating weights. This gating mechanism dynamically determines how much of the "original structural information" (backbone atoms crucial to the properties) from the previous layer is retained, directly corresponding to the mechanism specificity of structure-property associations in ADMET.

[0059] Specifically, in the standard D-MPNN, messages are propagated along directed edges, and message aggregation for each edge explicitly excludes reverse edges to avoid redundancy caused by information oscillating back and forth between the two ends of the chemical bond.

[0060] Based on this, the present invention introduces a selective information path based on atomic type in residual gating: the input of the previous layer and the current aggregated message are weighted and fused through gating.

[0061] Improper weight initialization can lead to gating saturation and variance imbalance. Therefore, this invention selected three initialization schemes (Kaiming normal initialization, orthogonal initialization, and Xavier normal initialization) for ablation experiments for comparison. The results of the ablation experiments are shown in Table 1. D-MPNN message aggregation effectively suppresses variance inflation during message propagation by excluding backward edges. Based on this, chemoselectivity gating uses the sigmoid activation function. Improper weight initialization can easily lead to gating saturation (output approaching 0 or 1), thereby causing gradient vanishing or forward propagation variance imbalance. Xavier normal initialization maintains consistent input-output variance between layers, precisely balancing the forward activation value distribution and backward gradient flow under sigmoid nonlinearity. This allows the gating weights to dynamically adjust the information retention ratio according to the atomic chemical type, maximizing the retention of key atomic structure information for ADMET property prediction.

[0062] Message update: The message after residual enhancement is input into the linear layer, added to the original input, passed through the activation function, and then regularized using Dropout.

[0063] Atomic latent vector generation: Aggregate all the final messages connected to each atom to obtain the atomic-level total message; after concatenating the original atomic features with the aggregated message, the final atomic latent vector is generated through the output layer and activation function.

[0064] Molecular-level readout: Based on a preset aggregation method (such as summation, averaging, or normalization), the latent vectors of all atoms within the same molecule are aggregated into a fixed-dimensional molecular-level representation vector for use by downstream prediction tasks.

[0065] Descriptor fusion: The molecular-level representation vector is concatenated with the traditional physicochemical molecular descriptor vector calculated by RDKit, and then fused through a descriptor layer. The dose feature is appended to the end of the composite molecular representation vector as an additional scalar feature input for subsequent prediction.

[0066] The readout module employs a residual attention-enhanced feedforward neural network, and its core process is as follows.

[0067] Specifically, for the input layer: the molecular graph vector output by the encoder is first regularized by Dropout, and then the input dimension is mapped to the hidden layer dimension through the first linear transformation layer.

[0068] The following is the iteration process of the intermediate hidden layer.

[0069] Specifically, for chemoselective residual connection blocks: each residual block sequentially includes layer normalization (LayerNorm), linear transformation (dimensionality invariant), activation function, Dropout, linear transformation (dimensionality invariant), and finally, the sub-layer output is added to the original input. The identity mapping preserves the molecular representation after the previous layer's nonlinear transformation, preventing the global molecular features in deep networks from being diluted layer by layer. Unlike the chemoselective residual gating in the feature extraction module, the residual connections in the readout module do not introduce atom-type gating, but rather maintain the complete transmission of molecular-level vectors, enabling subsequent self-attention layers to reweight the long-range dependencies between different atomic positions while preserving the original global information.

[0070] Specifically, for the standard layer: layer normalization, activation function, Dropout, and linear transformation are applied in sequence, with no residual connections.

[0071] Specifically, for distance-aware self-attention enhancement: In the ADMET property, distance constraints prevent attention from being scattered to chemically unrelated atom pairs. Self-attention generates a QKV matrix through a single linear layer, which is then split into multiple attention heads to compute attention weights in parallel, enabling the model to capture long-range dependencies between different atoms within the molecule. This invention introduces a position encoding module into standard multi-head self-attention, transforming the molecular graph structure into positional features available to each atom. This allows the neural network to "see" the topological role of atoms throughout the molecule, regardless of the arbitrary order of atom inputs. It captures the global structure through Laplacian eigenvectors and captures local positions through distance to the center; these two methods complement each other, giving the model both permutation invariance and structure awareness. This mechanism constrains attention weights to molecular topological distances, preventing attention from being scattered to chemically unrelated distant atom pairs, thereby more accurately capturing pharmacophore interactions associated with the ADMET property.

[0072] Specifically, for the output layer: after completing the iteration of all intermediate layers, the activation function, Dropout, and the last linear transformation layer are applied in sequence to map the hidden layer dimension to the target output dimension.

[0073] Specifically, for the loss function: the binary cross-entropy loss function is used, and the formula is: ; Where: N is the total number of samples. ∈{0,1} represents the true label of the i-th sample (1 indicates toxic, 0 indicates non-toxic). ∈(0,1) represents the probability that the i-th sample is toxic as predicted by the model.

[0074] For regression tasks, the mean squared error loss function is used, and the formula is: Where: N is the total number of samples. ∈R represents the true value of the i-th sample. ∈R represents the model's predicted value.

[0075] In the classification experiment, PRC-AUC (area under the precision-recall curve) and MCC (Matthews correlation coefficient) were used as evaluation metrics. PRC-AUC can effectively evaluate the model's predictive performance on positive samples, and is especially suitable for classification tasks with imbalanced positive and negative sample distributions. First, the formulas for precision and recall are defined as follows: .

[0076] TP (True Positive) represents the number of samples correctly predicted as positive, and FP (False Positive) represents the number of samples incorrectly predicted as positive. The formula is: ; Here, FN (False Negative) represents the number of false negatives, incorrectly predicted as negative. The recall rate is plotted on the x-axis, and the precision rate on the y-axis. By adjusting the classification threshold, a PR curve is obtained. PRC-AUC is the area under this curve.

[0077] MCC is a balanced metric that comprehensively considers true positives, true negatives, false positives, and false negatives. Its value ranges from -1 to 1. A value closer to 1 indicates a higher consistency between the model's predictions and the actual situation. The formula is: ; Where MCC=1 indicates a completely correct prediction, MCC=0 indicates a prediction result that is comparable to a random guess, and MCC=-1 indicates a completely incorrect prediction.

[0078] In the regression experiment, MAE (mean absolute error) and R0 were used. 2 The coefficient of determination (COD) is used as an evaluation metric. MAE measures the average level of the absolute difference between predicted and actual values; a smaller value indicates higher model prediction accuracy. The formula is: ; Where N is the total number of samples, Let i be the true value of the i-th sample. Let be the predicted value for the i-th sample.

[0079] R 2 This reflects the model's ability to explain the variability of the target variable, and its value ranges from (-∞, 1]. The closer the value is to 1, the better the model fits. The formula is: ; Among them, R 2 =1: The model perfectly fits the data, R0 2 =0: The model's predictive performance is comparable to simply using the mean; R0 2 <0: The model's prediction performance is worse than using the mean.

[0080] This invention provides a molecular ADMET property prediction model based on a message passing neural network with residual attention enhancement. The ADMET property prediction model is input with the SMILES string of the molecule and the ADMET properties of the small molecule are calculated.

[0081] Table 1 presents the ablation experimental results of three weight initialization schemes (Kaiming normal initialization, orthogonal initialization, and Xavier normal initialization) on four toxicity datasets (Hematotoxicity, Nephrotoxicity, Phototoxicity, and Neurotoxicity), with the Matthews correlation coefficient (MCC) used as the evaluation metric. The experimental results show that Xavier normal initialization is significantly superior to the other two schemes.

[0082] Table 1: .

[0083] Table 2 compares the enhancement mechanisms used by different models in each module. Model 1 uses distance-aware self-attention enhancement only in the readout module; Model 2 uses chemically selective residual enhancement in both the feature extraction and readout modules; this invention uses both chemically selective residual enhancement and distance-aware self-attention enhancement in both modules.

[0084] Table 2: .

[0085] This invention uses four toxicity datasets (Hematotoxicity, Nephrotoxicity, Phototoxicity, and Neurotoxicity) to verify the performance difference between a message passing network with a residual attention enhancement mechanism and a traditional message passing network. In this embodiment, the data can be divided into three types according to experimental type: in vivo animal experiments, human cell experiments, and human clinical experiments. In vivo animal experiments and human clinical experiments are classification experiments, while human cell experiments are regression experiments. In classification experiments, PRC-AUC and MCC are used as evaluation indicators; in regression experiments, MAE and R... 2 As an evaluation indicator.

[0086] As can be seen from the ablation experiment results in Tables 3-6, this invention, by introducing a unique chemically selective residual attention-enhanced network architecture and fully utilizing a rich toxicity dataset covering multiple sources and types of data, can more comprehensively and deeply capture the structural features and intrinsic relationships of compounds. Compared to traditional message passing networks, this invention achieves superior prediction performance in both classification and regression tasks, significantly improving the accuracy and generalization ability of compound property prediction. This advancement not only helps accelerate drug design and discovery but also brings tangible benefits and technological advancements to the field of compound property prediction.

[0087] Table 3: .

[0088] Table 4: .

[0089] Table 5: .

[0090] Table 6: .

[0091] Figure 4 and Figure 5 This is a comparison chart of the prediction performance of the present invention and the basic model in an embodiment of the invention. The present invention utilizes a unique network architecture to build a model on 55 ADMET property prediction tasks. The present invention surpasses the basic model in terms of model architecture novelty, data volume, and prediction performance. The present invention can more comprehensively predict the ADMET properties of molecules, helping to exclude compounds with poor ADMET properties at an early stage. It can also assist in evaluating the effect of structure optimization, effectively avoiding the risk of failure in drug clinical trials, increasing the success rate of research and development, and reducing experimental costs.

[0092] Example 3.

[0093] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps of the ADMET property prediction method for small molecule drugs according to any one of claims 1 to 8. The overall architecture of the computer device is built on a general-purpose computing platform. The memory stores a standardized SMILES string dataset, molecular graph structure, atomic features, bond features, molecular descriptor vectors, model parameters, and intermediate variables generated during training and inference. The processor is configured to call the program code in the memory and sequentially execute operations such as data preprocessing, molecular graph construction, directed message passing, attention-weighted modeling, molecular-level characterization generation, and ADMET property prediction.

[0094] In the specific operation, the processor first obtains a dataset containing drug small molecule SMILES strings and corresponding ADMET experimental labels from the TDC, SIDER, and ChEMBL databases. This dataset covers 55 prediction tasks, including AMES mutagenicity, BBB_Martins blood-brain barrier penetration, hERG cardiotoxicity, CYP450 enzyme inhibition / substrate activity, multiple nuclear receptor pathway activity, skin reaction, liver injury, nephrotoxicity, hematologic toxicity, phototoxicity, and neurotoxicity. The neurotoxicity data includes dose information in mg / kg. Subsequently, the original SMILES strings are subjected to charge removal, small fragment removal, chiral information removal, and tautomerism normalization to obtain a normalized SMILES string. This normalized string is then used as a deduplication matching key for missing value filtering and duplicate removal to obtain a processed SMILES string. This operation is consistent with step S1 in Example 1 and will not be detailed here.

[0095] Next, the processor calls RDKit to construct a molecular graph structure based on the processed SMILES string, with atoms as nodes and chemical bonds as edges. Atomic features are represented by one-hot encoding to indicate element type, atomic charge, atomic aromaticity, and the number of attached hydrogen atoms. Bond features include single, double, triple, and aromatic bond types. At the same time, RDKit is used to calculate a molecular descriptor vector consisting of molecular weight, lipid-water partition coefficient, topological polar surface area, number of hydrogen bond donors, and number of hydrogen bond acceptors. This part is the same as step S2 in Example 1 and is only referenced here.

[0096] In the directed message passing phase, the processor performs linear transformations on atomic features and bond features to generate initial messages. Messages are passed along directed edges in each round of message passing, and reverse edges are excluded during neighbor aggregation. Based on the atomic chemical categories determined by the atomic features, learnable gating weights are calculated using the Sigmoid activation function. The network parameters of these weights are initialized using a Xavier normal distribution to maintain consistent variance between layers. Based on these gating weights, the input messages from the previous round and the current aggregated messages are weighted and fused to obtain residual enhanced messages. After linear transformation, these messages are added to the original inputs and updated using activation functions and Dropout. After the final round, the original atomic features are concatenated and mapped with their aggregated messages to generate latent vector representations for each atom. This process is consistent with step S3 in Example 1 and is only briefly referenced.

[0097] Subsequently, the processor transforms the molecular graph into position-encoded features, which consist of Laplacian feature vectors and topological distances from atoms to the molecular center. Based on the position-encoded atomic latent vectors, a query matrix, a bond matrix, and a value matrix are generated through linear mapping, and the process is split into multiple attention heads for parallel computation. Self-attention weights are calculated based on the queries and bonds, and a decay penalty constraint based on molecular topological distance is applied to suppress attention allocation to chemically unrelated long-range atom pairs. Finally, the value matrices are weighted and summed to obtain the attention-enhanced atomic latent vector representation. This part corresponds to step S4 of Embodiment 1 and is cited only.

[0098] In the molecular-level characterization generation stage, the processor aggregates all attention-enhanced atomic latent vectors using a summation method to obtain a fixed-dimensional molecular-level representation vector. This vector is then concatenated with the molecular descriptor vector and fused through a linear mapping to obtain the molecular characterization vector of the fused descriptor. For neurotoxic samples, an additional dose feature vector is extracted and concatenated as a scalar feature to the end of the vector to form a composite molecular characterization vector. This operation is the same as step S5 in Example 1 and is only briefly described.

[0099] Finally, the processor performs end-to-end training on the ADMET property prediction model based on the composite molecular representation vector of labeled samples, and performs inference on the test molecules: the composite molecular representation vector is sequentially subjected to Dropout regularization and linear mapping to obtain initial hidden features; two residual connection processes are performed, each including layer normalization, linear mapping, activation function, Dropout, and linear mapping, and the output is directly added to the input to maintain the complete transmission of molecular-level representation; then, the hidden feature vector is obtained through layer normalization, activation function, Dropout, and linear mapping; for classification tasks, the vector is subjected to final linear mapping and then Sigmoid activation is applied to output the probability of toxicity or non-toxicity; for regression tasks, continuous prediction values ​​such as IC50, TD50, or clearance rate are directly output. This process is consistent with steps S6 and S7 of Example 1 and is only cited here.

[0100] In a preferred embodiment, the polymerization method may also use averaging or normalization instead of summation to accommodate the representation needs of molecules at different scales.

[0101] In a preferred embodiment, the dose feature is only included in the composite molecular characterization vector in the neurotoxicity prediction task because it has a sufficient sample size to support dose-toxicity relationship modeling, while other toxicity tasks do not introduce this feature due to the lack of reliable dose data.

[0102] In a preferred embodiment, the gating weights are initialized using a Xavier normal distribution. Compared to Kaiming normal initialization or orthogonal initialization, this method can more effectively avoid Sigmoid gating saturation, maintain the stability of the forward activation value distribution and the reverse gradient flow, and thus improve the ability to retain the atomic structure information of key pharmacophores.

[0103] In a preferred embodiment, the topological distance decay penalty constraint introduced in the self-attention mechanism significantly suppresses spurious attention assignments between chemically unrelated long-range atom pairs, making the model more focused on local pharmacophores and mid-range interactions related to ADMET properties, thereby improving prediction specificity.

[0104] Any process or method description in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or more executable instructions for implementing a particular logical function or process, and the scope of the preferred embodiments of the invention includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the functions involved, as will be understood by those skilled in the art to which embodiments of the invention pertain.

[0105] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus or device (such as a computer-based system, a system including a processing module or other system that can fetch and execute instructions from, an instruction execution system, apparatus or device).

[0106] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting the properties of small drug molecules using ADMET, characterized in that, include: Obtain the original SMILES string of the drug small molecule sample and process the original SMILES string; A molecular graph structure is constructed based on the processed SMILES string. Atomic features with atoms as nodes and bond features with chemical bonds as edges are extracted, and the corresponding molecular descriptor vectors are calculated. The atomic and bond features are processed based on a directed message passing mechanism. The atomic and bond features are linearly transformed to obtain the initial message. In each round of message passing, the message is passed along the directed edge, and the passing of the reverse edge is excluded when the neighbor message is aggregated. In each round of message passing, based on the atomic chemical category determined by atomic features, the previous round input message and the current aggregated message are weighted and fused using learnable gating weights to obtain the residual enhancement message, and the latent vector representation of each atom is generated based on the last round update message; Based on position encoding and molecular topological distance constraints, attention-weighted modeling is performed on the latent vector representations of each atom to characterize long-range dependencies within the molecule. The attention-weighted modeled atomic latent vector representations are aggregated to obtain molecular-level representation vectors, and the molecular-level representation vectors are concatenated with the molecular descriptor vectors to obtain composite molecular characterization vectors. Based on the composite molecular characterization vector of the tagged drug small molecule sample, the ADMET property prediction model, which consists of the directed message passing mechanism, the attention weighted modeling and the feedforward network structure, is trained. The trained ADMET property prediction model is then used to map the composite molecular characterization vector of the molecule to be tested, and the ADMET property prediction result of the molecule to be tested is output by combining the residual connection.

2. The method for predicting the properties of small molecule drugs using ADMET according to claim 1, characterized in that, The processing of the original SMILES string includes: The original SMILES string is subjected to charge removal, small fragment removal, chiral information elimination, and tautomerization normalization in sequence to obtain the normalized SMILES string; Using the standardized SMILES string as the deduplication matching key, missing value filtering and duplicate item removal are performed to obtain the processed SMILES string.

3. The method for predicting the properties of small molecule drugs using ADMET according to claim 1, characterized in that, The atomic features are represented by one-hot encoding to indicate element type, atomic charge, atomic aromaticity, and the number of attached hydrogen atoms; the bond features include bond types such as single bond, double bond, triple bond, and aromatic bond; the molecular descriptor vector includes a physicochemical descriptor vector composed of molecular weight, lipid-water partition coefficient, topological polar surface area, and the number of hydrogen bond donors and acceptors.

4. The method for predicting the properties of small drug molecules using ADMET according to claim 1, characterized in that, In each round of message passing, based on the atomic chemical categories determined by atomic features, the previous round's input message and the current aggregated message are weighted and fused using learnable gating weights to obtain a residual enhanced message. Furthermore, a latent vector representation of each atom is generated based on the last round's update message, including: Atomic chemical categories are determined based on the element type, atomic charge, and atomic aromaticity characterized by atomic features; The gating weights are calculated using the Sigmoid activation function. The network parameters used to calculate the gating weights are initialized using the Xavier normal distribution to maintain consistent input-output variance between layers during network propagation. The retention ratio of the previous round of input messages is determined based on the gating weights, and then weighted and added to the current aggregated message to obtain the message with residual enhancement. The message with the residual enhancement is processed by linear transformation and then added to the original input message of the current message passing layer. The message update for the current round is completed by activation function and Dropout regularization. For each atom within a molecule, the messages updated in the last round of all directed edges connected to it are aggregated to obtain the atomic-level total message. The atomic features of the atom are then concatenated with the atomic-level total message, and after mapping processing, the implicit vector representation of each atom is generated.

5. The method for predicting the properties of small drug molecules using ADMET according to claim 1, characterized in that, The attention-weighted modeling of the latent vector representations of each atom based on position encoding and molecular topological distance constraints includes: The molecular graph structure is transformed into a positional encoding feature to characterize each atom. The positional encoding feature is composed of a Laplacian feature vector that captures the global structure and a center distance feature that captures the local position. The center distance feature is the topological distance from the atom to the center of the molecule. Based on the atomic latent vector representation containing the location encoding features, a query matrix, a key matrix, and a value matrix are generated through linear mapping, and then split into multiple attention heads for parallel computation. Self-attention weights are calculated based on the query matrix and bond matrix, and a decay penalty constraint is imposed on the self-attention weights according to the molecular topological distance to suppress the allocation of attention to chemically unrelated long-range atom pairs. The value matrix is ​​weighted and summed based on the distance-constrained self-attention weights to obtain the attention-enhanced atomic latent vector representation.

6. The method for predicting the properties of small drug molecules using ADMET according to claim 1, characterized in that, The aggregation of the attention-weighted modeled atomic latent vector representations to obtain molecular-level representation vectors, and the concatenation of the molecular-level representation vectors with the molecular descriptor vectors to obtain composite molecular characterization vectors, includes: By using a preset summation, averaging, or normalization aggregation method, the implicit vector representations of all atoms within the same molecule are aggregated to obtain the molecular-level representation vector with a fixed dimension. The molecular-level representation vector is concatenated with the molecular descriptor vector, and the concatenated vector is linearly mapped to complete feature fusion, resulting in a molecular representation vector of the fused descriptor. If the current drug small molecule sample contains dosage information, then the dosage feature vector is extracted, and the dosage feature vector is concatenated with the molecular representation vector of the fusion descriptor as an additional scalar feature to obtain the composite molecular representation vector.

7. The method for predicting the properties of small drug molecules using ADMET according to claim 1, characterized in that, The mapping process using the trained ADMET property prediction model to the composite molecular characterization vector of the molecule under test includes: The composite molecule representation vector is sequentially subjected to Dropout regularization and linear mapping to obtain the initial hidden features; At least one residual connection process is performed on the initial hidden features. Each residual connection process includes layer normalization, linear mapping, activation function, Dropout regularization and linear mapping in sequence. The output of the linear mapping is then directly added to the features before the residual connection process, so as to maintain the complete transmission of molecular-level representation vectors without introducing atomic-type gating. After processing the residual connections, the features are sequentially subjected to layer normalization, activation function, Dropout regularization, and linear mapping to obtain the hidden feature vector used to output the prediction results.

8. The method for predicting the properties of small drug molecules using ADMET according to claim 7, characterized in that, The output of the ADMET property prediction results of the molecule to be tested includes: The hidden feature vectors are then subjected to a final linear mapping operation to map the high-dimensional hidden features to the output dimension of the target task, thereby generating predicted values. In response to the fact that the prediction task for the molecule to be tested is a classification task, a nonlinear activation operation is performed on the predicted value, and the ADMET property prediction result representing the classification probability is output. In response to the fact that the prediction task for the molecule to be tested is a regression task, the predicted value is output as a continuous ADMET property prediction result.

9. A drug small molecule ADMET property prediction system, characterized in that, include: The data acquisition and preprocessing unit is used to acquire the original SMILES string of the drug small molecule sample and process the original SMILES string. The graph construction and descriptor calculation unit is used to construct a molecular graph structure based on the processed SMILES string, extract atomic features with atoms as nodes and bond features with chemical bonds as edges, and calculate the corresponding molecular descriptor vectors. The directed message passing unit is used to process the atomic features and bond features based on the directed message passing mechanism, perform linear transformation on the atomic features and bond features to obtain the initial message, pass the message along the directed edge in each round of message passing, and exclude the passing of the reverse edge when the neighbor message is aggregated. The chemically selective gating fusion unit is used to perform weighted fusion of the previous round input message and the current aggregated message in each round of message passing, based on the atomic chemical category determined by atomic features and using learnable gating weights, to obtain the residual enhancement message, and to generate the latent vector representation of each atom based on the last round update message; Topologically constrained attention units are used to perform attention-weighted modeling of the latent vector representations of each atom based on position encoding and molecular topological distance constraints, in order to characterize long-range dependencies within the molecule; The aggregation and feature fusion unit is used to aggregate the atomic latent vector representation after attention-weighted modeling to obtain a molecular-level representation vector, and to concatenate the molecular-level representation vector with the molecular descriptor vector to obtain a composite molecular characterization vector, and to further concatenate the dose feature vector when dose information is present. The model training and prediction unit is used to train an ADMET property prediction model based on labeled drug small molecule samples, and to use the trained ADMET property prediction model to map the composite molecular characterization vector of the molecule to be tested, and to output the ADMET property prediction result of the molecule to be tested by combining the residual connection.

10. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the method according to any one of claims 1 to 8.

Images