A method for predicting admet properties of a drug compound molecule based on deep learning fusion of molecular graph and molecular point cloud
By using a multi-task graph neural network model that integrates molecular graphs and molecular point clouds through deep learning, the accuracy and precision issues of predicting ADMET properties of drug molecules were resolved, enabling more efficient drug design.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- CHINA PHARM UNIV
- Filing Date
- 2024-04-17
- Publication Date
- 2026-05-29
AI Technical Summary
Existing methods for predicting the ADMET properties of drug molecules suffer from low accuracy and poor precision. They ignore the fact that the ADMET properties of molecules are highly correlated, and traditional methods rely on artificial features, resulting in limited model adaptability.
A deep learning-based multi-task graph neural network model is adopted, which integrates molecular graphs and molecular point clouds. By extracting atomic physicochemical properties, atomic bond features and molecular position features, multi-task prediction is performed. Self-attention mechanism and multi-head attention mechanism are used to improve prediction accuracy and precision.
It significantly improves the accuracy and precision of predicting drug ADMET properties, enhances the model's generalization ability, reduces computational costs, and ensures the safety and stability of drug design.
Smart Images
Figure CN118280473B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of artificial intelligence drug development technology, specifically a method for predicting the ADMET properties of drug compound molecules based on deep learning fusion of molecular graphs and molecular point clouds. Background Technology
[0002] Drug discovery is a time-consuming, risky, costly, and complex process. During drug development, only approximately 10% of candidate drugs successfully complete clinical trials and gain market approval. Factors contributing to clinical trial failures are multifaceted, primarily including lack of clinical efficacy, uncontrollable toxicity and side effects, unfavorable pharmacokinetic characteristics, and strategic planning factors of the company. Pharmacokinetic characteristics refer to the processes of absorption, distribution, metabolism, and excretion of a drug in the human body, collectively known as ADMET properties along with toxicity. Defects in ADMET properties lead to 40-45% of clinical trial failures, becoming the leading cause of failure besides efficacy factors. Furthermore, one reason for the lack of clinical drug efficacy is unfavorable pharmacokinetic problems, which relate to whether the drug can reach the target protein in the body at an appropriate rate and sufficient concentration, and whether it can remain in the bloodstream for a certain period. ADMET properties are crucial factors affecting drug clinical efficacy and evaluating the druggability of compounds.
[0003] In traditional drug development, ADMET properties are typically evaluated only after the efficacy and selectivity of a candidate drug have been determined. Detecting defects in molecular ADMET properties at this stage usually necessitates a new round of molecular design and synthesis, and may even lead to project termination. Following the principle of "the earlier the failure, the smaller the loss," the drug discovery field has widely recognized the importance of accurately predicting molecular ADMET properties early in drug development. Early ADMET property evaluation can significantly reduce clinical trial failures due to ADMET property defects. However, biological experimental methods for evaluating molecular ADMET properties are not only time-consuming and costly, but also require extensive animal testing. In the early stages of drug development, there are often hundreds of candidate compounds; using this method would not only increase development costs but also lengthen the development cycle.
[0004] In the development of pharmaceutical informatics, experimental data on the molecular ADMET properties are continuously accumulated in major drug databases for public access. Utilizing this real and reliable data, along with various computer algorithms, drug developers can quickly and easily construct reliable ADMET property prediction models. Compared to traditional experimental methods, its advantages are: (1) High efficiency: It can quickly process a large number of compounds and predict their ADMET properties. (2) Low cost: It only requires the structure of the compound, and prediction can be made even if the compound has not been synthesized.
[0005] Currently, various methods have been developed for predicting the ADMET properties of compounds based on chemical structure. These mainly include: data-based methods, such as quantitative structure-activity relationship (QSAR), similarity search, and 3D QSAR; structure-based methods, such as ligand-protein docking and pharmacophore modeling; machine learning-based methods, such as random forests and support vector machines; and deep learning-based methods, such as graph convolutional neural networks and graph attention networks. Traditional molecular property prediction methods heavily rely on complex, handcrafted features such as molecular fingerprints and molecular descriptors, which require human expert knowledge. This leads to significant loss of molecular structural information and limited model adaptability. Deep learning methods, on the other hand, provide a data-driven approach that can automatically learn molecular characterizations from molecular structures. Compared to traditional methods, they can learn more molecular structural information, thus resulting in superior prediction performance. Therefore, the use of deep learning algorithms to predict molecular ADMET properties has attracted widespread attention.
[0006] Typically, deep learning models require large amounts of labeled data to achieve high efficiency and good generalization. However, compared to the vast amounts of data available in computer vision and natural language processing, data related to drug ADMET properties is quite limited. Furthermore, most current deep learning-based models for drug ADMET properties are single-task learning models, ignoring the fact that molecular ADMET properties are highly correlated. Summary of the Invention
[0007] Purpose of the Invention: To address the problems of low accuracy and poor precision in existing methods for predicting the ADMET properties of drug molecules, which ignore the fact that the ADMET properties of molecules are highly correlated, this invention proposes a method for predicting the ADMET properties of drug compound molecules based on deep learning and fusing molecular graphs and molecular point clouds. This method uses a multi-task graph neural network model to quickly and accurately predict the ADMET properties of drug molecules. By preprocessing the drug molecule data, relevant atomic physicochemical properties, atomic bond features, and atomic position features are extracted. Relying on multiple features for multi-task prediction significantly improves the accuracy and precision of the prediction, enhances the model's generalization ability, and reduces the overall computational cost.
[0008] Technical solution: A method for predicting the ADMET properties of drug compound molecules based on deep learning fusion of molecular graphs and molecular point clouds, comprising the following steps:
[0009] Step 1: Obtain the SMILES codes of the candidate drug molecules for which drug molecule properties need to be predicted;
[0010] Step 2: Process the SMILES code of the candidate drug molecule. Cut off the part that exceeds the maximum length of the SMILES code according to the set maximum length, and fill the insufficient part with 0 to the set maximum length.
[0011] Step 3: Generate a corresponding two-dimensional molecular diagram based on the number of atoms and chemical bonds in the SMILES code of the candidate drug molecule. Furthermore, extract atomic node features and atomic bond edge features from the SMILES code of the candidate drug molecule. Atomic node features include atom type, atom connector, number of free radical electrons, atom hybridization, aromaticity, chirality, and R / S configuration. Edge features include the type of atomic bond, whether it is a conjugated bond, whether it is in a ring, and the stereochemical form of the double bond: cis-trans. Add the atomic node features and atomic bond edge features from the drug molecule to the two-dimensional molecular diagram to obtain the drug molecule diagram.
[0012] Step 4: Extract the corresponding point cloud features of the candidate drug molecule based on the SMILES code of the candidate drug molecule;
[0013] Step 5: Use the preprocessed dataset to train the constructed deep learning model to obtain the trained deep learning model;
[0014] Step 6: Input the drug molecule map obtained in Step 3 and the corresponding point cloud features of the candidate drug molecules obtained in Step 4 into the trained deep learning model to obtain the ADMET property prediction results of the candidate drug molecules.
[0015] Furthermore, in step 4, the extraction of the corresponding point cloud features of the candidate drug molecule based on the SMILES code of the candidate drug molecule specifically includes the following operations:
[0016] Based on the SMILES code of the candidate drug molecule, each atom in the candidate drug molecule is traversed, the required H atoms are added to the candidate drug molecule, and it is transformed into a three-dimensional form to obtain the coordinate features of the candidate drug molecule.
[0017] The coordinate features of the candidate drug molecule are encoded into a two-dimensional array feature tensor of shape N*3, where N is the number of atoms in the candidate drug molecule and 3 is the three-dimensional coordinate (x,y,z) of the atom, thus obtaining the corresponding point cloud features of the candidate drug molecule.
[0018] Furthermore, the deep learning model includes a position feature encoder, an atom and atomic bond feature encoder, and a multi-task drug molecule property predictor;
[0019] The location spatial feature encoder is used to extract the location features of candidate drug molecules from the input two-dimensional array feature tensor using a self-attention mechanism.
[0020] The atom and atomic bond feature encoder is used to obtain global features in the input drug molecule diagram;
[0021] The multi-task drug molecule property predictor is used to fuse global features in the drug molecule graph with positional features of candidate drug molecules, obtain an attention weight matrix using continuous linear and sigmoid functions, assign different attention weights for different tasks, and predict the ADMET properties of output candidate drug molecules in a targeted manner.
[0022] Furthermore, in the location spatial feature encoder, the following steps are performed:
[0023] The input two-dimensional array feature tensor P∈R N×d Passing through a dimension d e Feature encoding yields feature vectors.
[0024] Use self-attention mechanism to process feature vectors Represented as:
[0025] (Q,K,V)=F e ×(W q W k W v )
[0026]
[0027] F out =F e +Relu(BatchNorm(Linear(F sa )))
[0028] In the formula, W q W k W v All are linear transformation parameters, d a F represents the dimension of the Q and K feature vectors. out These are the positional features of candidate drug molecules extracted via the attention mechanism.
[0029] Furthermore, in the atom and atomic bond feature encoder, the following steps are performed:
[0030] The different types of node and edge features in the drug molecule graph are aggregated, and the aggregated result is passed to the next hidden layer. The update mechanism is as follows:
[0031]
[0032] in, This represents the state vector of node i after l+1 iterations. Let c represent the set of neighboring nodes of edge r of node i. i,r c represents a regularization constant. i,r The value is This represents a linear transformation function that transforms neighboring nodes of the same type of edge. Using a linear transformation function Perform the transformation, where r∈R represents the type of edge;
[0033] A multi-head attention mechanism is used to update the node features in the drug molecule graph to obtain new node features. A set of node features is input. Where N is the number of nodes, R F Each node's features are tensors of dimension F; first, a linear transformation is performed on the node's feature vector h, and then attention coefficients are calculated for neighboring nodes, expressed as:
[0034]
[0035] Where || represents the concatenation operation, a T Let W be the weight vector, and W be a linear operation.
[0036] This allows us to obtain the global features of the input drug molecule diagram.
[0037] Furthermore, different property prediction tasks include classification tasks and regression tasks;
[0038] For classification tasks, the following loss function is used for training:
[0039] one_task = -[p c y n,c *logσ(x n,c )+(1-y n,c )*log(1-σ(x n,c ))]
[0040]
[0041] Where n represents the number of drug molecules in classification task c, p c y represents the weight of the positive samples in classification task c. n,c Let x represent the true value of the nth drug molecule in classification task c.n,c σ(·) represents the predicted value of the nth drug molecule in classification task c, σ(·) represents the Sigmoid function, and mean represents the average of the sub-task loss values in the multi-task process, which is used as the overall loss value of the model.
[0042] For regression tasks, the following loss function is used for training:
[0043]
[0044] Where n represents the number of drug molecules in regression task r, x n,r y represents the predicted value of the nth drug molecule in regression task r. n,c This represents the true value of the nth drug molecule in the regression task r.
[0045] This invention discloses a system for predicting the ADMET properties of drug compound molecules based on deep learning fusion of molecular graphs and molecular point clouds, comprising:
[0046] The SMILES code input module is used to obtain the SMILES codes of candidate drug molecules for which drug molecule properties need to be predicted.
[0047] The data cleaning module is used to process the SMILES codes of candidate drug molecules according to the set SMILES codes.
[0048] The code has a maximum length. Any part exceeding the maximum length will be truncated, and any insufficient part will be padded with 0s to the set maximum length.
[0049] The drug molecule graph construction module is used to generate corresponding two-dimensional molecular graphs based on the number of atoms and chemical bonds in the SMILES code of candidate drug molecules. Furthermore, it extracts atomic node features and atomic bond edge features from the SMILES code of candidate drug molecules. Atomic node features include atom type, atom connector, number of free radical electrons, atom hybridization, aromaticity, chirality, and R / S configuration. Edge features include the type of atomic bond, whether it is a conjugated bond, whether it is in a ring, and the stereochemical form of the double bond: cis-trans. The atomic node features and atomic bond edge features of the drug molecule are added to the two-dimensional molecular graph to obtain the drug molecule graph.
[0050] The point cloud feature extraction module is used to extract the corresponding point cloud features of candidate drug molecules based on the SMILES codes of candidate drug molecules.
[0051] The prediction module is used to input the drug molecule map and the corresponding point cloud features of the candidate drug molecule into the pre-trained deep learning model to obtain the ADMET property prediction results of the candidate drug molecule.
[0052] Furthermore, the extraction of point cloud features of candidate drug molecules based on SMILES codes specifically includes the following operations:
[0053] Based on the SMILES code of the candidate drug molecule, each atom in the candidate drug molecule is traversed, the required H atoms are added to the candidate drug molecule, and it is transformed into a three-dimensional form to obtain the coordinate features of the candidate drug molecule.
[0054] The coordinate features of the candidate drug molecule are encoded into a two-dimensional array feature tensor of shape N*3, where N is the number of atoms in the candidate drug molecule and 3 is the three-dimensional coordinate (x,y,z) of the atom, thus obtaining the corresponding point cloud features of the candidate drug molecule.
[0055] Furthermore, the deep learning model includes a position feature encoder, an atom and atomic bond feature encoder, and a multi-task drug molecule property predictor;
[0056] The location spatial feature encoder is used to extract the location features of candidate drug molecules from the input two-dimensional array feature tensor using a self-attention mechanism.
[0057] The atom and atomic bond feature encoder is used to obtain global features in the input drug molecule diagram;
[0058] The multi-task drug molecule property predictor is used to fuse global features in the drug molecule graph with positional features of candidate drug molecules, obtain an attention weight matrix using continuous linear and sigmoid functions, assign different attention weights for different tasks, and predict the ADMET properties of output candidate drug molecules in a targeted manner.
[0059] Furthermore, in the location spatial feature encoder, the following steps are performed:
[0060] The input two-dimensional array feature tensor P∈R N×d Passing through a dimension d e Feature encoding yields feature vectors.
[0061] Use self-attention mechanism to process feature vectors Represented as:
[0062] (Q,K,V)=F e ×(W q W k W v )
[0063]
[0064] F out =Fe +Relu(BatchNorm(Linear(F sa )))
[0065] In the formula, W q W k W v All are linear transformation parameters, d a F represents the dimension of the Q and K feature vectors. out Positional features of candidate drug molecules extracted via attention mechanisms;
[0066] In the atom and atomic bond feature encoder, perform the following steps:
[0067] The different types of node and edge features in the drug molecule graph are aggregated, and the aggregated result is passed to the next hidden layer. The update mechanism is as follows:
[0068]
[0069] in, This represents the state vector of node i after l+1 iterations. Let c represent the set of neighboring nodes of edge r of node i. ir c represents a regularization constant. ir The value is This represents a linear transformation function that transforms neighboring nodes of the same type of edge. Using a linear transformation function Perform the transformation, where r∈R represents the type of edge;
[0070] A multi-head attention mechanism is used to update the node features in the drug molecule graph to obtain new node features. A set of node features is input. Where N is the number of nodes, R F Each node's features are tensors of dimension F; first, a linear transformation is performed on the node's feature vector h, and then attention coefficients are calculated for neighboring nodes, expressed as:
[0071]
[0072] Where || represents the concatenation operation, a T Let W be the weight vector, and W be a linear operation.
[0073] This allows us to obtain the global features of the input drug molecule diagram.
[0074] Beneficial effects: Compared with the prior art, the present invention has the following advantages:
[0075] (1) The method of the present invention uses molecular SMILES codes to construct point clouds and molecular diagrams to characterize and describe compounds. The structural information and attribute information of compound molecules are represented in the form of numbers or characters and then input into a deep learning model, which greatly improves the accuracy and precision of predicting the properties of drug ADMET.
[0076] (2) The method of the present invention makes predictions based on the atomic physicochemical properties, atomic bond information and atomic position information in the drug molecule, which greatly improves the accuracy and precision of predicting the ADMET properties of the drug and ensures the safety and stability of the designed drug or candidate drug.
[0077] (3) The method of the present invention realizes multi-task learning, puts multiple related tasks together for learning, shares information during the learning process, and improves the generalization ability of the model. Attached Figure Description
[0078] Figure 1 This is a schematic diagram illustrating the basic principle of the present invention;
[0079] Figure 2 This is a schematic diagram of the data processing flow of the present invention;
[0080] Figure 3 This is a flowchart of the ADMET property prediction model for drug molecules of the present invention;
[0081] Figure 4 This is a framework diagram of the ADMET property prediction model for drug molecules in this invention. Detailed Implementation
[0082] This invention establishes a method for processing two-dimensional and three-dimensional structural data of drug molecules and predicting their ADMET properties using a deep learning model. By applying deep learning methods to the prediction of ADMET properties of drug molecules, the accuracy and precision of traditional detection methods based on chemical data structures and machine learning are significantly improved. The proposed method for predicting the ADMET properties of drug molecules can predict their ADMET properties based on the characteristics of atoms and atomic bonds in drug compound molecules, combined with the positional characteristics of the molecules in three-dimensional space.
[0083] The technical solution of the present invention will now be further described in conjunction with the accompanying drawings and embodiments.
[0084] like Figure 1 As shown, this embodiment proposes a method for predicting the ADMET properties of drug compound molecules based on deep learning fusion of molecular graphs and molecular point clouds, mainly including the following steps:
[0085] Step 1: Obtain the SMILES code data of the candidate drug molecules for which drug molecule properties need to be predicted;
[0086] Step 2: See Figure 2 The SMILES codes of candidate drug molecules are cleaned. Since the lengths of the SMILES codes vary, they are truncated to a set maximum length. Sequences exceeding the maximum length are truncated, while shorter sequences are padded with zeros to achieve a uniform length.
[0087] A two-dimensional molecular graph of a drug molecule is constructed. This graph represents the drug molecule structure using nodes and edges, where nodes represent atoms and edges represent chemical bonds connecting those atoms. Typically, the graph data is defined as a tuple G = (V, E), where V represents a non-empty finite set of atoms, denoted as V(G), and E represents a non-empty finite set of chemical bonds, denoted as E(G). The Deep Graph Library (DGL) is used to generate the corresponding molecular graph based on the number of atoms and chemical bonds in the SMILES code data. Each atom and chemical bond has its own characteristics, namely atomic node characteristics and atomic bond edge characteristics. Atom types include: ['B','C','N','O','F','Si','P','S','Cl','As','Se','Br','Te','I','At','other']. Atom types are encoded as binary vectors based on their atomic numbers. Atoms within the defined atomic types are represented as 1, and those outside are represented as 0. Atom node characteristics include: atom type, atom connector, number of free radical electrons, atom hybridization, aromaticity, chirality, and R / S configuration. Atom bond edge characteristics include: type of atomic bond, whether it is a conjugated bond, whether it is in a ring, and the stereochemical form of the double bond: cis-trans.
[0088] Atomic node features and atomic bond edge features of candidate drug molecules are extracted from the SMILES code structure data after cleaning. The extracted atomic node features and atomic bond edge features of drug molecules are added to a two-dimensional molecular graph to obtain a drug molecule graph.
[0089] The point cloud features of candidate drug molecules are extracted based on their SMILES codes. Specifically, a 3D point cloud is a collection of multiple three-dimensional coordinate points (X, Y, Z). Representing a molecule in point cloud form, where each atom corresponds to a point in space, captures not only the connectivity of covalent atoms but also the conformational preference information of the molecule. Therefore, firstly, based on the SMILES codes of the candidate drug molecules, the RDKit toolkit is used to traverse each atom in the candidate drug molecule, adding the required H atoms to each molecule and converting it into a three-dimensional form to obtain the coordinate features of the candidate drug molecule. Then, the coordinate features of the candidate drug molecule are encoded into a two-dimensional array feature tensor of shape N*3, where N is the number of atoms in the molecule and 3 is the three-dimensional coordinate (x, y, z) of the atom, thus obtaining the point cloud feature representation of the candidate drug molecule.
[0090] Step 3: Input the corresponding point cloud features and drug molecule map of the candidate drug molecule into the pre-trained deep learning model to predict the ADMET properties of the drug molecule;
[0091] The deep learning model used in this embodiment will now be further explained. For example... Figure 3 and Figure 4 As shown, the deep learning model used in this embodiment takes the point cloud features of drug molecules and drug molecule maps as input, and includes a position feature encoder, an atom and atomic bond feature encoder, and a multi-task drug molecule property predictor.
[0092] The positional feature encoder employs a point cloud description method, using a self-attention mechanism to extract point cloud features of candidate drug molecules from a two-dimensional array feature tensor. By representing drug molecules in the form of a point cloud, with each atom corresponding to a point in space, it can capture not only the connectivity of covalent atoms but also the conformational preference information of the drug molecule. The self-attention mechanism is used to obtain all the positional features information in the point cloud. Since there are interactions between atoms in the molecule, the self-attention mechanism is used to process the molecular point cloud data. Through comparison calculations between each atomic point and other atoms, the model can learn the correlations and connections between different atomic points, capturing the complex dependencies within the molecule. Specifically, the positional feature encoder performs the following steps:
[0093] The self-attention mechanism is used to process molecular point cloud data. By comparing each atomic point with other atoms, the model can learn the correlation and connection between different atomic points and capture the complex dependencies within the molecule.
[0094] The input two-dimensional array feature tensor P∈R N×d Passing through a dimension d e Feature encoding yields feature vectors.
[0095] Use self-attention mechanism to process feature vectors Represented as:
[0096] (Q,K,V)=F e ×(W q W k W v )
[0097]
[0098] F out =F e +Relu(BatchNorm(Linear(F sa )))
[0099] In the formula, W q W k W v All are linear transformation parameters, d a F represents the dimension of the Q and K feature vectors. out These are the positional features of candidate drug molecules extracted via the attention mechanism.
[0100] The atomic and atomic bond feature encoder is used to acquire global features from the input drug molecule graph. On one hand, a relational graph convolutional network is used to continuously aggregate information from adjacent nodes and edges in the molecule graph, thereby learning the relationships within the graph structure and information related to the ADMET properties of the drug molecule. On the other hand, to better capture the global information existing in the molecule graph, a graph attention mechanism is employed. By calculating the attention weights between each node and other nodes in the molecule graph, attention is focused on the parts of the input molecule graph most relevant to the target task, thus making accurate predictions.
[0101] Specifically, in the atom and atomic bond feature encoder, the following steps are performed:
[0102] The different types of node and edge features in the drug molecule graph are aggregated, and the aggregated result is passed to the next hidden layer. The update mechanism is as follows:
[0103]
[0104] in, This represents the state vector of node i after l+1 iterations. Let c represent the set of neighboring nodes of edge r of node i. i,r c represents a regularization constant. i,r The value is This represents a linear transformation function that transforms neighboring nodes of the same type of edge. Using a linear transformation function Perform the transformation, where r∈R represents the type of edge;
[0105] A multi-head attention mechanism is used to update the node features in the drug molecule graph to obtain new node features. A set of node features is input. Where N is the number of nodes, R F Each node's features are tensors of dimension F; first, a linear transformation is performed on the node's feature vector h, and then attention coefficients are calculated for neighboring nodes, expressed as:
[0106]
[0107] Where || represents the concatenation operation, a T Let W be the weight vector, and W be a linear operation. To ensure stable learning of more meaningful information, multi-head attention with 4 attention heads is used.
[0108] By stacking nodes to participate in the layer update calculation of their neighborhood features, different weights are assigned to different nodes in the neighborhood, thereby obtaining global features in the molecular graph.
[0109] The multi-task drug molecule property predictor fuses global and positional features in the molecular graph. For different property prediction tasks, it uses continuous linear and sigmoid functions to obtain attention weight matrices, applying different attention weights to different tasks. This allows for targeted prediction of the ADMET properties of the output drug molecule, enabling multi-task ADMET property prediction. Multi-task learning leverages the inherent correlation between multiple related tasks, improving the overall performance of each task, enhancing model generalization, and reducing overall computational cost.
[0110] Train the deep learning model using the following steps to obtain a pre-trained deep learning model:
[0111] A dataset of ADMET properties of drug molecules was constructed. Multiple ADMET datasets were collected from databases such as ChEMBL, ADMETNet, and Tox21, as well as from literature. The collected datasets were then processed as follows: (1) Data with the same attributes were integrated together according to the same standard; (2) Duplicate entries of molecules were removed; (3) A single dataset was randomly divided into training, validation, and test sets in a ratio of 8:1:1; (4) The datasets were integrated into five categories: absorption, distribution, metabolism, excretion, and toxicity, resulting in a total of 51 ADMET attributes, including 41 classification tasks and 10 regression tasks, totaling 288,728 data entries. Drug molecules were stored in SMILES code format. The ADMET property data of drug molecules were divided into five categories and stored in CSV files respectively. The first line of each file contained molecular attribute labels.
[0112] The ADMET property data of drug molecules stored in the CSV file were cleaned. Due to the inconsistent lengths of the SMILES codes for drug molecules, the length distribution of the collected SMILES codes was analyzed, and the maximum length of the SMILES codes was fixed at 120 based on this distribution. Sequences exceeding the maximum length were truncated, while shorter sequences were padded with 0s.
[0113] Molecular graph features were extracted from the cleaned drug molecules. The prepared CSV file of drug properties was further processed to construct a molecular graph, and atomic node features and atomic bond edge features of the drug molecules were extracted. Using the RDKit library, atomic feature matrices representing node feature information and chemical bond feature matrices representing edge feature information were constructed and stored in the molecular graph.
[0114] The training steps include:
[0115] S1: The molecular graph and atomic coordinate position tensor (120×3) are saved in the bin file. The information in the molecular graph includes atomic node features, atomic bond edge features and predicted label values.
[0116] S2: Transpose the atomic coordinate position tensor input from S1 to make its shape (3×120), then perform a 1D convolution operation with a stride of 1, padding=0, 64 kernels, and kernel size of 1, making the tensor shape (64×120). Then perform a BatchNorm1d normalization process and a ReLU activation operation, keeping the tensor shape unchanged.
[0117] S3: Perform a 1D convolution operation on the tensor output by S2 with a stride of 1, padding = 0, 128 kernels, and kernel size of 1. The tensor shape becomes (128×120). Then, it undergoes BatchNorm1d normalization and ReLU activation. The tensor shape remains unchanged.
[0118] S4: Perform a 1D convolution operation on the tensor output by S3 with a stride of 1, padding = 0, 128 kernels, and kernel size of 1. Repeat the above convolution activation operation twice to obtain a total of 3 identical tensors with a tensor shape of (128×120).
[0119] S5: Perform shape transformation on the three tensors obtained in S4 to make their shapes (120×4×32);
[0120] S6: Transpose the three tensors obtained in S5 to make their shapes (4×120×32), namely tensor q, tensor k, and tensor v;
[0121] S7: Transpose the k tensor obtained in S6 to make its shape (4×32×120);
[0122] S8: Multiply the q tensor obtained in S6 with the k tensor output in S7 by matrix multiplication to obtain a tensor of shape (4×120×120);
[0123] S9: Scale the tensor output by S8 by dividing the tensor by the logarithm of the vector dimension, while keeping the tensor shape unchanged;
[0124] S10: Perform a Softmax activation operation on the tensor output by S9, without changing the shape of the tensor;
[0125] S11: Multiply the tensor obtained in S10 with the v tensor obtained in S6 by matrix multiplication to obtain a tensor of shape (4×120×32);
[0126] S12: Perform shape transformation on the tensor output by S11, making its shape (120×128);
[0127] S13: Transpose the tensor output by S12 to make its shape (128×120);
[0128] S14: Perform a 1D convolution operation on the tensor obtained in S13 with a stride of 1, padding = 0, 128 kernels, and kernel size of 1. The tensor shape becomes (128×120). Then, it undergoes BatchNorm1d normalization and ReLU activation. The tensor shape remains unchanged.
[0129] S15: Add the tensor obtained in S14 to the tensor output in S3 to obtain a tensor with a shape of (128×120);
[0130] S16: Perform a 1D convolution operation on the tensor output by S15 with a stride of 1, padding = 0, 128 kernels, and kernel size of 1. Repeat the above convolution activation operation twice to obtain a total of 3 identical tensors with a tensor shape of (128×120).
[0131] S17: Perform shape transformation on the three tensors obtained in S16 to make their shapes (120×4×32);
[0132] S18: Transpose the three tensors obtained in S17 to make their shapes (4×120×32), namely tensor q, tensor k, and tensor v;
[0133] S19: Transpose the k tensor obtained in S18 to make its shape (4×32×120);
[0134] S20: Multiply the q tensor obtained in S18 with the k tensor output in S19 by matrix multiplication to obtain a tensor of shape (4×120×120);
[0135] S21: Scale the tensor output by S20 by dividing the tensor by the logarithm of the vector dimension, while keeping the tensor shape unchanged;
[0136] S22: Perform a Softmax activation operation on the tensor output by S21, without changing the shape of the tensor;
[0137] S23: Multiply the tensor obtained in S22 with the v tensor obtained in S18 by matrix multiplication to obtain a tensor of shape (4×120×32);
[0138] S24: Perform shape transformation on the tensor output by S23, making its shape (120×128);
[0139] S25: Transpose the tensor output by S24 to make its shape (128×120);
[0140] S26: Perform a 1D convolution operation on the tensor obtained in S25 with a stride of 1, padding = 0, 128 kernels, and kernel size of 1. The tensor shape becomes (128×120). Then, it undergoes BatchNorm1d normalization and ReLU activation. The tensor shape remains unchanged.
[0141] S27: Add the tensor obtained in S26 to the tensor output in S15 to obtain a tensor with a shape of (128×120);
[0142] S28: Perform a 1D convolution operation on the tensor output by S27 with a stride of 1, padding = 0, a number of kernels of 128, and a kernel size of 1. Repeat the above convolution activation operation twice to obtain a total of 3 identical tensors with a tensor shape of (128×120).
[0143] S29: Perform shape transformation on the three tensors obtained in S28 to make their shapes (120×4×32);
[0144] S30: Transpose the three tensors obtained in S29 to make them (4×120×32), which are tensors q, k, and v respectively.
[0145] S31: Transpose the k tensor obtained in S30 to make its shape (4×32×120);
[0146] S32: Multiply the q tensor obtained from S30 with the k tensor output from S31 by matrix multiplication to obtain a tensor of shape (4×120×120);
[0147] S33: Scale the tensor output by S32 by dividing the tensor by the logarithm of the vector dimension, while keeping the tensor shape unchanged;
[0148] S34: Perform a Softmax activation operation on the tensor output by S33, without changing the shape of the tensor;
[0149] S35: Multiply the tensor obtained in S34 with the v tensor obtained in S30 by matrix multiplication to obtain a tensor of shape (4×120×32);
[0150] S36: Perform shape transformation on the tensor output by S36, making its shape (120×128);
[0151] S37: Transpose the tensor output by S37 to make its shape (128×120);
[0152] S38: Perform a 1D convolution operation on the tensor obtained from S38 with a stride of 1, padding = 0, 128 kernels, and kernel size of 1. The tensor shape becomes (128×120). Then, it undergoes BatchNorm1d normalization and ReLU activation. The tensor shape remains unchanged.
[0153] S39: Add the tensor obtained in S38 to the tensor output in S27 to obtain a tensor with a shape of (128×120);
[0154] S40: Perform a 1D convolution operation on the tensor output by S39 with a stride of 1, padding = 0, a number of kernels of 128, and a kernel size of 1. Repeat the above convolution activation operation twice to obtain a total of 3 identical tensors with a tensor shape of (128×120).
[0155] S41: Perform shape transformation on the three tensors obtained in S40 to make their shapes (120×4×32);
[0156] S42: Transpose the three tensors obtained in S41 to make their shapes (4×120×32), which are tensor q, tensor k, and tensor v respectively;
[0157] S43: Transpose the k tensor obtained in S42 to make its shape (4×32×120);
[0158] S44: Multiply the q tensor obtained in S42 with the k tensor output in S43 by matrix multiplication to obtain a tensor of shape (4×120×120).
[0159] S45: Scale the tensor output by S44 by dividing the tensor by the logarithm of the vector dimension, while keeping the tensor shape unchanged.
[0160] S46: Perform a Softmax activation operation on the tensor output by S45, without changing the shape of the tensor;
[0161] S47: Multiply the tensor obtained in S46 with the v tensor obtained in S42 by matrix multiplication to obtain a tensor of shape (4×120×32);
[0162] S48: Perform shape transformation on the tensor output by S47, making its shape (120×128);
[0163] S49: Transpose the tensor output by S48 to make its shape (128×120);
[0164] S50: Perform a 1D convolution operation on the tensor obtained from S49 with a stride of 1, padding = 0, 128 kernels, and kernel size of 1. The tensor shape becomes (128×120). Then, it undergoes BatchNorm1d normalization and ReLU activation, while the tensor shape remains unchanged.
[0165] S51: Add the tensor obtained from S50 to the tensor output from S39 to obtain a tensor with a shape of (128×120);
[0166] S52: Concatenate the tensors obtained from S15, S27, S39, and S51 to obtain a tensor of shape (512×120);
[0167] S53: Perform a 1D convolution operation on S52 with a stride of 1, padding = 0, 256 kernels, and kernel size of 1. Then, perform a BatchNorm1d normalization process and a LeakyReLU activation operation. The tensor shape remains unchanged.
[0168] S54: Perform a 1D convolution operation on S53 with a stride of 1, padding = 0, 128 kernels, and kernel size of 1. The tensor shape becomes (128×120). Then, it undergoes BatchNorm1d normalization and LeakyReLU activation. The tensor shape remains unchanged.
[0169] S55: Perform a 1D convolution operation on S54 with a stride of 1, padding = 0, 64 kernels, and kernel size of 1. The tensor shape becomes (64×120). Then, it undergoes BatchNorm1d normalization and LeakyReLU activation. The tensor shape remains unchanged. S56: Perform a relation graph convolution operation on the molecular graph input from S1 with an output feature number of 64. The number of relations is 64*21. Use the basis weight regularizer to add a bias. Its shape is (64).
[0170] S57: Perform a linear fully connected operation on the features output by S56 with a feature number of 64, and the tensor shape becomes (64). Then perform a ReLU activation operation, and the feature shape remains unchanged.
[0171] S58: Add the features obtained in S57 to the features obtained in S56 to obtain features with shape (64), and then perform BatchNorm1d normalization processing to keep the feature shape unchanged.
[0172] S59: Perform a graph convolution operation on the features obtained in S58 with an output feature number of 128. The number of relations is 64*21. Use the basis weight regularizer to add a bias so that its shape is (128).
[0173] S60: Perform a linear fully connected operation on the features output by S59 with a feature count of 128, and the tensor shape becomes (128). Then perform a ReLU activation operation, and the feature shape remains unchanged.
[0174] S61: Add the features obtained in S60 to the features obtained in S59 to obtain features with shape (128), and then perform BatchNorm1d normalization processing to keep the feature shape unchanged; S62: Perform a convolutional neural network GATConv based on graph attention mechanism on the molecular graph input in S1 and the node feature tensor in the molecular graph, and its shape becomes (48).
[0175] S63: Perform a graph attention-based convolutional neural network GATConv on the feature tensor output by S62, and its shape becomes (128).
[0176] S64: Perform a concatenate operation on the feature tensor output by S63 and the feature tensor output by S61, and its shape becomes (256).
[0177] S65: For the graph features output by S64, perform a linear fully connected operation with an output feature number of 1 on the node tensor for different tasks, and then perform a sigmoid activation operation to calculate the weights of different tasks. Multiply the graph features output by S64 by their respective weights for different tasks to obtain a new feature tensor with a shape of (128×256).
[0178] S66: Perform shape transformation on the feature tensor output by S59, then concatenate it with the feature tensor output by S65, then perform a linear fully connected operation with 1024 output features, and the tensor shape becomes (128×1024), then perform a linear fully connected operation with 512 output features, and the tensor shape becomes (128×512).
[0179] S67: Perform a Dropout regularization operation with a loss rate of 0.2 on the feature tensor output by S66, then perform a linear fully connected operation with 64 output features, and the tensor shape becomes (128×64). Then perform a ReLU activation operation and a BatchNorm1d normalization process, and the tensor shape remains unchanged.
[0180] S68: Perform a Dropout regularization operation with a loss rate of 0.2 on the feature tensor output by S67, then perform a linear fully connected operation with 64 output features, and the tensor shape becomes (128×64). Then perform a ReLU activation operation and a BatchNorm1d normalization process, and the tensor shape remains unchanged.
[0181] S69: Perform a Dropout regularization operation with a loss rate of 0.2 on the feature tensor output by S68, then perform a linear fully connected operation with 64 output features, and the tensor shape becomes (128×64). Then perform a ReLU activation operation and a BatchNorm1d normalization process, and the tensor shape remains unchanged.
[0182] S70: Perform a linear full-connection operation on the feature tensor output by S69 with a feature number of 1, and the tensor shape becomes (128×1);
[0183] S71: Concatenate the tensors of all tasks output by S70 to synthesize an overall prediction result;
[0184] S72: Calculate the loss function based on the predicted value output from S71 and the label value input from S1. If it does not meet the requirements, perform backpropagation to update the parameter weights of each layer of the model. Repeat the above steps and iterate continuously until the model fits.
[0185] To obtain a robust and accurate model during training, the five datasets were randomly split and trained ten times to avoid problems such as model overfitting and waste of computational resources.
[0186] Each epoch is set to 500. Due to the large amount of data, the DataLoader class is used to input the data into the model for training in batches, with a batch size of 128. To avoid problems such as model overfitting and wasted computational resources, an EarlyStopping mechanism is used. Training stops when the optimal performance score on the validation set (the average of the sum of the performance evaluation values of each attribute) remains unchanged for 50 consecutive epochs. Adam (ptive moment estimation) is used as an optimization for the gradient backpropagation algorithm. Adam uses an adaptive learning rate to accelerate convergence. At the end of training, the model with the highest optimal performance score on the validation set from ten training iterations is saved in the pth file. The deep learning model used in this embodiment is a multi-task output prediction model, using different loss functions for different output types.
[0187] For classification tasks, which are all binary classifications, the model predicts a value between 0 and 1 for each classification task. The loss function used is BCEWithLogitsLoss, which is expressed as:
[0188] one_task = -[p c y n,c *logσ(x n,c )+(1-y n,c)*log(1-σ(x n,c ))]
[0189]
[0190] Where n represents the number of drug molecules in classification task c, p c y represents the weight of the positive samples in classification task c. n,c Let x represent the true value of the nth drug molecule in classification task c. n,c Let σ(x) represent the predicted value of the nth drug molecule in classification task c, σ(x) represent the Sigmoid function, and mean represents the average of the sub-task loss values in the multi-task classification, which is used as the overall loss value of the model.
[0191] For regression tasks, the model predicts a specific numerical value. The loss function used is MSE (Mean Squared Error), which can be expressed as:
[0192]
[0193] Where n represents the number of drug molecules in regression task r, x n,r y represents the predicted value of the nth drug molecule in regression task r. n,c This represents the true value of the nth drug molecule in the regression task r.
Claims
1. A method for predicting the ADMET properties of drug compound molecules based on deep learning fusion of molecular graphs and molecular point clouds, characterized in that: Includes the following steps: Step 1: Obtain the SMILES codes of the candidate drug molecules for which drug molecule properties need to be predicted; Step 2: Process the SMILES code of the candidate drug molecule. Cut off the part that exceeds the maximum length of the SMILES code according to the set maximum length, and fill the insufficient part with 0 to the set maximum length. Step 3: Generate a corresponding two-dimensional molecular graph based on the number of atoms and chemical bonds in the SMILES code of the candidate drug molecule. Furthermore, extract the atomic node features and edge features of the atomic bonds from the SMILES code of the candidate drug molecule. Node features include atom type, atom connector, number of free radical electrons, atom hybridization mode, aromaticity, chirality, and R / S type of atom; edge features include type of atomic bond, whether it is a conjugated bond, whether it is in a ring, and stereochemical form of double bond; by adding the atomic node features and atomic bond edge features of the drug molecule to the two-dimensional molecular diagram, a drug molecule diagram is obtained. Step 4: Extract the corresponding point cloud features of the candidate drug molecule based on the SMILES code of the candidate drug molecule; Step 5: Use the preprocessed dataset to train the constructed deep learning model to obtain the trained deep learning model; Step 6: Input the drug molecule map obtained in Step 3 and the corresponding point cloud features of the candidate drug molecules obtained in Step 4 into the trained deep learning model to obtain the ADMET property prediction results of the candidate drug molecules.
2. The method for predicting the ADMET properties of drug compound molecules based on deep learning fusion of molecular graphs and molecular point clouds according to claim 1, characterized in that: In step 4, the extraction of point cloud features of candidate drug molecules based on SMILES codes specifically includes the following operations: Based on the SMILES code of the candidate drug molecule, each atom in the candidate drug molecule is traversed, the required H atoms are added to the candidate drug molecule, and it is transformed into a three-dimensional form to obtain the coordinate features of the candidate drug molecule. The coordinate features of the candidate drug molecule are encoded into a two-dimensional array feature tensor of shape N*3, where N is the number of atoms in the candidate drug molecule and 3 is the three-dimensional coordinate (x,y,z) of the atom, thus obtaining the corresponding point cloud features of the candidate drug molecule.
3. The method for predicting the ADMET properties of drug compound molecules based on deep learning fusion of molecular graphs and molecular point clouds according to claim 1, characterized in that: The deep learning model includes a location-space feature encoder, an atom and atomic bond feature encoder, and a multi-task drug molecule property predictor. The location spatial feature encoder is used to extract the location features of candidate drug molecules from the input two-dimensional array feature tensor using a self-attention mechanism. The atom and atomic bond feature encoder is used to obtain global features in the input drug molecule diagram; The multi-task drug molecule property predictor is used to fuse global features in the drug molecule graph with positional features of candidate drug molecules, obtain an attention weight matrix using continuous linear and sigmoid functions, assign different attention weights for different tasks, and predict the ADMET properties of output candidate drug molecules in a targeted manner.
4. The method for predicting the ADMET properties of drug compound molecules based on deep learning fusion of molecular graphs and molecular point clouds according to claim 3, characterized in that: In the location-space feature encoder, the following steps are performed: The input two-dimensional array feature tensor Encode the data to obtain a set of dimensions. eigenvectors ; Use self-attention mechanism to process feature vectors , represented as: ； ； ； In the formula, All are linear transformation parameters. Let Q and K be the dimensions of the feature vectors. These are the positional features of candidate drug molecules extracted via the attention mechanism.
5. The method for predicting the ADMET properties of drug compound molecules based on deep learning fusion of molecular graphs and molecular point clouds according to claim 3, characterized in that: In the atom and atomic bond feature encoder, perform the following steps: The different types of node and edge features in the drug molecule graph are aggregated, and the aggregated result is passed to the next hidden layer. The update mechanism is as follows: ； in, Represents a node go through The state vector after +1 iterations Represents a node edge The set of neighboring nodes, This represents a regularization constant. The value is , This represents a linear transformation function that transforms neighboring nodes of the same type of edge. Using a linear transformation function Transformation, Indicates the type of edge; A multi-head attention mechanism is used to update the node features in the drug molecule graph to obtain new node features. A set of node features is input. , Where N is the number of nodes. Let h be a tensor of dimension F; first, perform a linear transformation on the node feature vector h, then calculate the attention coefficients for neighboring nodes, expressed as: ； Here, || represents the concatenation operation. Let W be the weight vector, and W be a linear operation. This allows us to obtain the global features of the input drug molecule diagram.
6. The method for predicting the ADMET properties of drug compound molecules based on deep learning fusion of molecular graphs and molecular point clouds according to claim 5, characterized in that: Different property prediction tasks include classification tasks and regression tasks; For classification tasks, the following loss function is used for training: ； ； in, Representing classification tasks The number of drug molecules in the middle Representing classification tasks The weights of neutral samples, Representing classification tasks No. The true value of a drug molecule Representing classification tasks No. Predicted value for each drug molecule, Represents the Sigmoid function; This means averaging the loss values of the subtasks in a multi-task scenario, and using this average as the overall loss value of the model. For regression tasks, the following loss function is used for training: ； in, Indicates the return task The number of drug molecules in the middle Indicates the return task The Middle Predicted value for each drug molecule, Indicates the return task The Middle The true value of a drug molecule.
7. A system for predicting the ADMET properties of drug compound molecules based on deep learning fusion of molecular graphs and molecular point clouds, characterized in that: include: The SMILES code input module is used to obtain the SMILES codes of candidate drug molecules for which drug molecule properties need to be predicted. The data cleaning module is used to process the SMILES codes of candidate drug molecules. It truncates the parts that exceed the maximum length of the SMILES code according to the set maximum length, and fills the insufficient parts with 0 to the set maximum length. The drug molecule graph construction module generates a corresponding two-dimensional molecular graph based on the number of atoms and chemical bonds in the SMILES code of candidate drug molecules. Furthermore, it extracts atomic node features and edge features of atomic bonds from the SMILES code of candidate drug molecules. Node features include atom type, atom connector, number of free radical electrons, atom hybridization mode, aromaticity, chirality, and R / S type of atom; edge features include type of atomic bond, whether it is a conjugated bond, whether it is in a ring, and stereochemical form of double bond; by adding the atomic node features and atomic bond edge features of the drug molecule to the two-dimensional molecular diagram, a drug molecule diagram is obtained. The point cloud feature extraction module is used to extract the corresponding point cloud features of candidate drug molecules based on the SMILES codes of candidate drug molecules. The prediction module is used to input the drug molecule map and the corresponding point cloud features of the candidate drug molecule into the pre-trained deep learning model to obtain the ADMET property prediction results of the candidate drug molecule.
8. The system for predicting the ADMET properties of drug compound molecules based on deep learning fusion of molecular graphs and molecular point clouds according to claim 7, characterized in that: The specific operations for extracting the corresponding point cloud features of candidate drug molecules based on SMILES codes include: Based on the SMILES code of the candidate drug molecule, each atom in the candidate drug molecule is traversed, the required H atoms are added to the candidate drug molecule, and it is transformed into a three-dimensional form to obtain the coordinate features of the candidate drug molecule. The coordinate features of the candidate drug molecule are encoded into a two-dimensional array feature tensor of shape N*3, where N is the number of atoms in the candidate drug molecule and 3 is the three-dimensional coordinate (x,y,z) of the atom, thus obtaining the corresponding point cloud features of the candidate drug molecule.
9. The system for predicting the ADMET properties of drug compound molecules based on deep learning fusion of molecular graphs and molecular point clouds according to claim 8, characterized in that: The deep learning model includes a location-space feature encoder, an atom and atomic bond feature encoder, and a multi-task drug molecule property predictor. The location spatial feature encoder is used to extract the location features of candidate drug molecules from the input two-dimensional array feature tensor using a self-attention mechanism. The atom and atomic bond feature encoder is used to obtain global features in the input drug molecule diagram; The multi-task drug molecule property predictor is used to fuse global features in the drug molecule graph with positional features of candidate drug molecules, obtain an attention weight matrix using continuous linear and sigmoid functions, assign different attention weights for different tasks, and predict the ADMET properties of output candidate drug molecules in a targeted manner.
10. The system for predicting the ADMET properties of drug compound molecules based on deep learning fusion of molecular graphs and molecular point clouds according to claim 9, characterized in that: In the location-space feature encoder, the following steps are performed: The input two-dimensional array feature tensor Encode the data to obtain a set of dimensions. eigenvectors ; Use self-attention mechanism to process feature vectors , is represented as: ； ； ； In the formula, All are linear transformation parameters. Let Q and K be the dimensions of the feature vectors. Positional features of candidate drug molecules extracted via attention mechanisms; In the atom and atomic bond feature encoder, perform the following steps: The different types of node and edge features in the drug molecule graph are aggregated, and the aggregated result is passed to the next hidden layer. The update mechanism is as follows: ； in, Represents a node go through The state vector after +1 iterations Represents a node edge The set of neighboring nodes, This represents a regularization constant. The value is , This represents a linear transformation function that transforms neighboring nodes of the same type of edge. Using a linear transformation function Transformation, Indicates the type of edge; A multi-head attention mechanism is used to update the node features in the drug molecule graph to obtain new node features. A set of node features is input. , Where N is the number of nodes. Let h be a tensor of dimension F; first, perform a linear transformation on the node feature vector h, then calculate the attention coefficients for neighboring nodes, expressed as: ； Here, || represents the concatenation operation. Let W be the weight vector, and W be a linear operation. This allows us to obtain the global features of the input drug molecule diagram.