Substrate specificity prediction method and model of UGT enzyme subtype

Abstract

The invention relates to a UGT enzyme subtype substrate specificity prediction method and model. On the basis of a directional message passing neural network, graph structure characterization of a small molecule compound and features of specific protein binding sites of UGT enzyme are deeply fused, a bimodal prediction normal form of'molecule + protein binding sites' is designed, a deep learning model is constructed, conversion from compound center prediction to molecule-enzyme binding site comprehensive prediction is achieved, and the prediction accuracy is improved. And accurate classification prediction can be carried out on UGT enzyme substrates and non-substrates.

Description

Technical Field

[0001] This invention belongs to the field of computational medicinal chemistry and bioinformatics technology, and relates to a substrate-specific prediction method and model for UGT enzyme subtypes. Background Technology

[0002] In modern drug development, the metabolic properties of candidate compounds are a key factor determining their successful drug development. UDP-glucuronosyltransferases (UGTs), as an important family of phase II metabolic enzymes, increase the water solubility of compounds and promote their excretion from the body by catalyzing glucuronidation, thereby regulating drug clearance, half-life, and toxicity. The UGT family includes multiple isoenzymes, such as UGT1A1, UGT1A9, and UGT2B7, each exhibiting significant substrate specificity. Therefore, accurately predicting which UGT enzyme isoenzyme a new compound is a substrate for is one of the core challenges in drug metabolism research.

[0003] To obtain information on the interaction between UGT enzymes and compounds, current technologies largely rely on traditional biological experimental methods. For example, they assess individual differences in UGT enzyme activity through gene polymorphism detection (e.g., CN117305460A, CN119876399A) or develop various specific probe substrates (e.g., CN102898295A, CN119751352A) to detect enzyme activity in vitro. While these biological experimental methods provide reliable data, they have inherent limitations. First, wet-lab experiments are complex, time-consuming, labor-intensive, and costly, making them unsuitable for the high-throughput, low-cost virtual screening of large-scale compound libraries needed in the early stages of drug discovery. Second, metabolic data for not all compounds can be easily obtained experimentally, especially in the early design stages of New Chemical Entities (NCEs), where data is severely scarce.

[0004] In recent years, with the development of artificial intelligence technology, some computational methods have been used to predict UGT-related metabolic properties. These methods can be mainly divided into two categories:

[0005] (1) Traditional machine learning methods based on molecular descriptors: Researchers use traditional machine learning models such as support vector machines (SVM) and partial least squares discriminant analysis (PLSDA) in combination with artificially designed molecular descriptors (such as quantum chemical descriptors and two-dimensional descriptors) to predict the metabolic or inhibitory activity of UGT enzymes. For example, WO2022018253A1 discloses a training method and model for predicting drug-metabolizing enzyme inhibitors. Its input features also include physicochemical descriptors and binding energies generated by molecular dynamics simulations, and it uses random forest (RF) or SVM as a classification model. The core idea of ​​this type of method is the "similarity principle", that is, molecules with similar structures should have similar biological activities. However, its fundamental limitation is that this type of model only models the characteristics of the compound itself, which is essentially a "compound-centered" prediction paradigm. This paradigm ignores the unique biological information of UGT enzyme subtypes (such as amino acid sequences, three-dimensional structures and binding site features), which makes the model unable to truly understand the fundamental biological question of "why a certain molecule is a substrate of UGT1A1 and not a substrate of UGT1A3". When faced with molecules that have similar structures but act on different UGT subtypes, their prediction accuracy and specificity are limited. (2) Metabolic site prediction methods based on graph neural networks (GNNs): such as the XenoSite model and the method for predicting P450 metabolic sites in CN110428875A, etc., regard the metabolic prediction task as an atomic-level classification problem, that is, predicting which atom in the molecule is most likely to undergo a metabolic reaction (Site of Metabolism, SOM). These models usually use GNNs or similar methods to represent molecules as graph structures and learn the local chemical environment of atoms in molecules.

[0006] In summary, existing computational models cannot effectively and explicitly integrate the biological information of UGT enzyme subtypes (especially protein binding site features) into the prediction framework, resulting in a superficial understanding of UGT enzyme-substrate interactions and limited prediction accuracy and specificity. Therefore, there is an urgent need to construct a deep learning framework capable of simultaneously processing heterogeneous data (small molecule diagram structures and UGT enzyme binding site features) and enabling the model to learn the joint characterization of synergistic interactions between molecules and enzymes, thereby achieving more accurate and biologically meaningful predictions. Summary of the Invention

[0007] To address the shortcomings of existing technologies and practical needs, this invention provides a substrate-specific prediction method and model for UGT enzyme subtypes, aiming to achieve accurate classification and prediction of UGT enzyme substrates and non-substrate components.

[0008] To achieve the above objectives, the present invention adopts the following technical solution:

[0009] In a first aspect, the present invention provides a substrate-specific prediction method for UGT enzyme subtypes, characterized in that the prediction method includes:

[0010] Compound and UGT enzyme data are acquired. Based on a directed message passing neural network, a molecular encoder is used to construct a compound molecular graph with atomic information as nodes and chemical bond information as edges. A protein encoder is used to construct a UGT enzyme binding site graph with amino acid residue information of the UGT enzyme binding site as nodes and chemical interaction information as edges. The compound characterization vector and the UGT enzyme binding site characterization vector are output. The compound characterization vector and the UGT enzyme binding site characterization vector are fused using a gating fusion module to obtain a joint characterization vector. The joint characterization vector is input into a convolutional neural network learning model, which outputs the substrate probability value corresponding to each UGT enzyme subtype. Based on the output substrate probability value, it is determined whether the compound is the substrate of the corresponding UGT enzyme subtype. The model is trained and optimized to obtain a UGT enzyme metabolism prediction model with fused protein binding site information.

[0011] Obtain the joint characterization vector of the compound to be analyzed and the UGT enzyme, input it into the UGT enzyme metabolism prediction model containing the binding site information of the fusion protein, and output the substrate probability value.

[0012] This invention applies a unified Directed Message Passing Neural Network (DMPNN) to the feature extraction of two fundamentally different biomolecules: molecules and proteins. By processing these two heterogeneous graphs using the same DMPNN architecture and message passing mechanism, it ensures that molecular and protein vectors are output into the same feature space. This achieves isomorphic representation of heterogeneous biomolecules and precise alignment of bimodal features, breaking through the single-modality limitation of traditional UGT metabolic prediction that is "based solely on compound structure." By utilizing a bimodal prediction paradigm of "molecule + protein binding site," it explicitly incorporates protein binding site information into the prediction model. Through a convolutional neural network (CNN) fusion module, it learns the molecule-enzyme interaction pattern, realizing a shift from compound center prediction to comprehensive molecule-enzyme binding site prediction.

[0013] Preferably, the acquisition of compound and UGT enzyme data includes:

[0014] (1-1) Data Collection

[0015] Structural information of compounds was obtained from public compound structure databases and represented in SMILES string format. Substrate and non-substrate tag information of different UGT enzyme subtypes was obtained; protein amino acid sequence information of UGT enzymes was obtained.

[0016] (1-2) Data Preprocessing

[0017] Remove compound records with incomplete structures, invalid SMILES, or duplicates; merge labels for multiple experimental records of the same compound and the same UGT subtype using consistency checks and majority voting; discard data records with missing labels.

[0018] The SMILES string is standardized and a molecular object is generated. Hydrogen atom completion, aromaticity identification, and bond type standardization are performed on the molecule.

[0019] Preferably, the UGT enzyme subtype includes at least one of UGT1A1, UGT1A4, UGT1A6, UGT1A8, UGT1A9, UGT1A10, UGT2B7, or UGT2B15.

[0020] Preferably, the construction of the compound molecular map includes:

[0021] Construct atomic node feature vectors, which include atomic feature vectors and chemical bond information including chemical bond feature vectors; the atomic feature vectors include atom type, atomicity, formal charge, hybridization type, and aromaticity;

[0022] Construct a chemical bond edge feature vector, wherein the chemical bond feature vector includes the bond type and whether it is in a ring;

[0023] Based on the molecular encoder, a compound molecular graph is constructed using atomic node feature vectors and chemical bond edge feature vectors.

[0024] Preferably, the construction of the UGT enzyme binding site map includes:

[0025] The amino acid residue information of the binding site of the UGT enzyme is extracted. The amino acid residue information of the binding site of the UGT enzyme includes the amino acid residue sequence and One-Hot coding information and structural environment feature information. The structural environment feature information includes the spatial coordinates of the amino acid residue, the solvent-accessible surface area, the secondary structure type, the physicochemical properties of the residue, the distance from the center of the binding site and the local atomic density and chemical environment descriptor.

[0026] Nodes are constructed using the amino acid residue information;

[0027] Edges are constructed using chemical interaction information, which includes interaction type, interaction strength, and spatial distance; the interaction type includes covalent bond connection, spatial distance connection, and chemical interaction connection.

[0028] A UGT enzyme binding site map is constructed using the nodes and edges built based on the protein encoder.

[0029] Preferably, the edges of the UGT enzyme binding site map include sequence-adjacent edges, i.e., the sequences themselves are connected, generally peptide bond edges; and spatially adjacent edges, which are edges with a spatial distance of less than 6. The first type of edge is generally considered to have some interaction; the third type of edge is an edge that does not exist in reality. The edge is defined as a virtual edge because the relative distance between binding sites is smaller. Virtual edges are defined to allow features to flow better to each other. However, for virtual edges, the flow intensity will be reduced accordingly (initially 0.2, learnable).

[0030] Preferably, the message passing mechanism of the molecular encoder includes:

[0031] ;

[0032] in, For the t-th round, follow the node Send to node The message vector (equivalent to the edge hidden state), For activation functions (such as ReLU). For the weight matrix of the message linear layer, put Shift to the message dimension;

[0033] ;

[0034] in, For nodes The results of neighbor message aggregation For nodes The set of adjacent nodes, For round t from Send to node The message vector (equivalent to the edge hidden state);

[0035] ;

[0036] in, Characterize node j of the read molecular graph. For splicing operations, This represents the node information of node j in the molecular graph. This represents the information of all neighboring edges of node j in the molecular graph.

[0037] Preferably, the message passing mechanism of the protein encoder includes:

[0038] ;

[0039] in, This is the message vector pointing from residue i to residue j at the (t+1)th iteration. It is a ReLU activation function. To create a learnable matrix, the concatenation of node and edge features is mapped to the dimensional space of the message vector. To concatenate node features and edge features, Given the set of neighbors of the original node i, excluding the target node j, To aggregate all incoming edge messages pointing from neighbor k to neighbor i, The weight coefficient (scalar) is used to adjust the contribution of each incoming edge message to the update. Real edges are initially set to 1, and virtual edges to 0.2. This refers to the incoming edge message from neighbor k to neighbor i.

[0040] Preferably, the virtual edge weights corresponding to UGT1A1 are 0.431, UGT1A4 are 0.129, UGT1A6 are 0.321, UGT1A8 are 0.371, UGT1A9 are 0.527, UGT1A10 are 0.415, UGT2B7 are 0.656, and UGT2B15 are 0.589.

[0041] Preferably, the gating fusion module includes a gating coefficient calculation module. and :

[0042] ;

[0043] ;

[0044] in, For compound characterization vectors, This is a vector representing the UGT enzyme binding site. For the Sigmoid function, and For the science department parameter matrix, , The dynamic gating parameter takes values ​​of [0,1] and controls the ratio of compound and UGT enzyme features to the final combined feature, respectively.

[0045] Preferably, the joint representation vector is defined as: ;

[0046] in, This indicates element-wise multiplication.

[0047] Preferably, the model training and optimization includes:

[0048] The model is trained using a binary cross-entropy loss function, with each UGT enzyme subtype treated as an independent task. An optimizer is used to train each task using gradient descent. Performance is monitored on the validation set, and an early stopping mechanism is employed to prevent overfitting and save the best-performing model.

[0049] Secondly, the present invention provides a UGT enzyme metabolism prediction model with fusion protein binding site information, wherein the prediction model is obtained by the method for constructing the UGT enzyme metabolism prediction model with fusion protein binding site information described in the first aspect.

[0050] This invention designs and constructs a deep learning model that deeply integrates the graphical structure representation of small molecule compounds with the features of specific protein binding sites of UGT enzymes to achieve accurate classification and prediction of UGT enzyme substrates and non-substrates.

[0051] Thirdly, the present invention provides a method for predicting the substrate of UGT enzyme, the method comprising making predictions using a UGT enzyme metabolism prediction model utilizing the fusion protein binding site information described in the second aspect.

[0052] The method described in this invention can serve as a highly efficient and low-cost virtual screening tool for assessing ADMET (absorption, distribution, metabolism, excretion, and toxicity) properties in the early stages of drug development. By providing accurate UGT metabolism predictions in the early stages of drug design, it aims to significantly reduce the risk of failure in later clinical trials, thereby accelerating the discovery and optimization of lead compounds, shortening the R&D cycle, and saving costs.

[0053] Compared with the prior art, the present invention has the following beneficial effects:

[0054] (1) This invention realizes unified modeling of molecular and protein heterogeneous data: By uniformly adopting the Directed Message Passing Neural Network (DMPNN) architecture on two types of heterogeneous data, molecular graphs and protein binding site graphs are homogeneously represented, so that the model can efficiently learn the interaction relationship between molecules and enzymes in a unified feature space, thereby improving the model's prediction accuracy and generalization.

[0055] (2) This invention is the first to realize substrate prediction modeling that integrates UGT enzyme binding site information: For the first time, the structural features of UGT enzyme binding sites are combined with small molecule graph structures for modeling, breaking through the traditional prediction paradigm based on structural similarity, enabling the model to explicitly capture the structural differences of different UGT subtypes and their impact on substrate selectivity, thereby significantly improving the accuracy of prediction. Attached Figure Description

[0056] Figure 1 This is a diagram of the predictive model architecture.

[0057] Figure 2 Build a flowchart for the prediction model (using UGT1A4 as an example). Detailed Implementation

[0058] The technical solution of the present invention will be further described below with reference to the accompanying drawings and specific embodiments. However, the following examples are merely simplified examples of the present invention and do not represent or limit the scope of protection of the present invention. The scope of protection of the present invention is determined by the claims.

[0059] Where specific techniques or conditions are not specified in the examples, they shall be performed in accordance with the techniques or conditions described in the literature in this field, or in accordance with the product instructions. Reagents or instruments whose manufacturers are not specified are all conventional products that can be purchased from legitimate channels.

[0060] This invention designs a metabolic prediction method based on a Directed Message Passing Neural Network (DMPNN) that integrates UGT enzyme binding site information. It applies DMPNN uniformly to the extraction of features from two fundamentally different biomolecules: molecules and proteins. A schematic diagram of the principle is shown below. Figure 1 As shown, the molecular encoder constructs small molecules into a graph structure with atoms as nodes (including features such as atom type, degree, formal charge, hybridization type, and aromaticity) and chemical bonds as edges (including bond type and ring state features); the protein encoder constructs the UGT enzyme binding site into a graph structure with amino acid residues as nodes (residue type and structural environment features) and chemical interactions as edges. Finally, the outputs of these two encoders are concatenated in a fusion layer for use in the downstream convolutional neural network (CNN) fusion and prediction module.

[0061] Its specific implementation process is as follows: Figure 2 As shown, it includes the following steps:

[0062] Step 1: Data Preparation and Feature Engineering

[0063] 1. Data Sources and Preprocessing

[0064] (1) Data collection

[0065] Compound structure data: The structural information of small molecule compounds is obtained from public databases (such as ChEMBL, PubChem, DrugBank, etc.) and represented in SMILES (Simplified Molecular Input Line Entry System) format.

[0066] UGT enzyme substrate tag data: Combining literature experimental data and database annotations, substrate / non-substrate tag information for different UGT enzyme subtypes (including UGT1A1, UGT1A4, UGT1A6, UGT1A8, UGT1A9, UGT1A10, UGT2B7 and UGT2B15) was obtained.

[0067] Protein sequence data: The protein amino acid sequences of the above-mentioned UGT enzyme subtypes were obtained from the UniProt database, and included the corresponding UniProt ID, sequence version and annotation information.

[0068] (2) Data cleaning

[0069] Remove compound records that are incomplete in structure, have invalid SMILES, or contain duplicates.

[0070] For multiple experimental records of the same compound and the same UGT enzyme subtype, the labels were merged using a consistency check and majority voting method.

[0071] Data records with missing labels are removed to ensure the completeness of the training data labeling.

[0072] (3) Standardization and format conversion

[0073] Use open-source toolkits (such as RDKit) to standardize the SMILES strings (Canonical SMILES) and generate molecule objects.

[0074] The molecules are subjected to hydrogen atom completion, aromaticity identification, and bond type standardization.

[0075] 2. Molecular map construction

[0076] Molecular graph features: For each molecule, generate atomic node feature vectors (such as atom type, atomicity, formal charge, hybridization type, etc.) and chemical bond edge feature vectors (such as bond type, whether conjugated, whether in a ring, etc.).

[0077] (1) Construction of atomic node eigenvectors

[0078] For each atom in the molecule, the following feature vector (partial) is generated:

[0079] Atom type (C, N, O, S, etc.) ----- One-Hot encoding;

[0080] Atomicity (the number of atoms connected to it);

[0081] Formal charge;

[0082] Hybridization types (sp, sp2, sp3, etc.);

[0083] Aromaticity (whether it is in an aromatic ring).

[0084] The atomic node characteristics of the molecular diagram are shown in Table 1.

[0085] Table 1

[0086]

[0087] (2) Construction of chemical bond edge eigenvectors

[0088] Key type (single bond, double bond, triple bond, aromatic bond) ------ One-Hot encoding;

[0089] Whether it is conjugate (Boolean value);

[0090] Whether it is in a ring structure (Boolean value);

[0091] Stereochemistry.

[0092] The chemical bond edge characteristics of the molecular diagram are shown in Table 2.

[0093] Table 2

[0094]

[0095] 3. Construction of UGT enzyme binding site map

[0096] UGT binding site characteristics: For each UGT subtype, its binding site sequence is converted into a graph structure, and residue characteristics (such as amino acid type, position, etc.) are generated.

[0097] (1) Extraction of binding site information

[0098] Obtain the three-dimensional structures of UGT enzymes (UGT1A1, UGT1A4, UGT1A6, UGT1A8, UGT1A9, UGT1A10, UGT2B7, UGT2B15) from crystal structure databases (such as PDB).

[0099] Use a pocket detection tool (such as Fpocket) to identify ligand binding sites and obtain the set of binding site residues and their spatial coordinates (x, y, z).

[0100] (2) Construction of protein (binding site) map

[0101] Node construction (amino acid information):

[0102] Residue type features: Twenty standard amino acids are one-hot encoded to generate a 20-dimensional feature vector;

[0103] Structural environmental characteristics include:

[0104] The three-dimensional coordinates (x, y, z) of the residue at the binding site;

[0105] Solvent-accessible surface area (SASA);

[0106] Secondary structure types (α-helix, β-fold, random coil, etc.);

[0107] Physicochemical properties of residues (hydrophobicity, charge, polarity, etc.);

[0108] Distance from the center of the binding site;

[0109] Local atomic density and chemical environment descriptor.

[0110] The node characteristics of the protein binding site map are shown in Table 3.

[0111] Table 3

[0112]

[0113] Edge construction (connection information) establishes connection relationships based on the types of chemical interactions between residues, including:

[0114] Covalent linkage: Peptide bonds connecting adjacent residues;

[0115] Spatial distance connection: Set a distance threshold (e.g., 8) Edge connections are established between residue pairs within the threshold.

[0116] Chemical interaction connections:

[0117] Hydrogen bond interactions (donor-acceptor residue pairs);

[0118] Hydrophobic interactions (between hydrophobic residues);

[0119] Electrostatic interaction (between charged residues);

[0120] π-π stacking interactions (between aromatic residues);

[0121] Edge feature encoding: Each edge contains feature information such as interaction type, interaction strength, and spatial distance;

[0122] Edge definition: We define edges in three ways: sequence-adjacent edges, i.e., edges where the sequences themselves are connected, generally peptide bonds; spatially adjacent edges, i.e., edges where the spatial distance is less than 6. The first type of edge is generally considered to have some interaction; the third type of edge is an edge that does not exist in reality. We define this type of edge as virtual edge because the relative distance between binding sites is smaller. By defining virtual edges, features can flow better to each other. However, for virtual edges, the flow intensity will be reduced accordingly (initially 0.2, learnable).

[0123] The edge features of the protein binding site map are shown in Table 4.

[0124] Table 4

[0125]

[0126] 4. Data partitioning

[0127] Based on the molecular scaffold, the scaffold is divided into segments, using the Murcko scaffold as a basis, to ensure that the training and test sets do not overlap in terms of chemical framework.

[0128] Ratio: 80% training set, 20% validation set.

[0129] This method can significantly reduce overfitting of the model to the known skeleton and improve generalization performance.

[0130] Step 2: Model Architecture

[0131] (1) Dual encoder module:

[0132] The DMPNN molecular encoder receives the molecular graph and iteratively updates the hidden states of atoms and bonds through message passing via a multi-layered DMPNN. Finally, a readout function aggregates all the hidden states of atoms into a vector representation representing the entire small molecule.

[0133] Message Passing Mechanism:

[0134] information:

[0135] ;

[0136] in, For the t-th round, follow the node Send to node The message vector (equivalent to the edge hidden state), For activation functions (such as ReLU). For the weight matrix of the message linear layer, put Shift to the message dimension;

[0137] polymerization:

[0138]

[0139] in, For nodes The results of neighbor message aggregation For nodes The set of adjacent nodes, For round t from Send to node The message vector (equivalent to the edge hidden state);

[0140] ;

[0141] in, Characterize node j of the read molecular graph. For splicing operations, This represents the node information of node j in the molecular graph. This represents the information of all neighboring edges of node j in the molecular graph.

[0142] DMPNN protein encoder:

[0143] Binding site map construction: First, the binding sites of the UGT enzyme are represented as a residue-level map. Node features not only include residue types (20-dimensional one-hot), but also incorporate structural environmental features related to the binding sites, such as three-dimensional coordinates, solvent accessible surface area (SASA), secondary structure, physicochemical properties (hydrophobicity, charge, polarity), distance from the pocket center, and local atomic density. These features can systematically characterize the spatial conformation and chemical environment of the binding pocket.

[0144] Edge Design and Virtual Edge Mechanism: In constructing the graph edges, we retain real interactions (covalent bonds, hydrogen bonds, hydrophobic interactions, electrostatic interactions, π–π stacking) and further introduce virtual edges to improve graph connectivity. The motivation for designing virtual edges is that binding site residues often have long-distance but functionally related synergistic effects, and relying solely on local interaction edges would lead to limited information transmission.

[0145] Specifically, in the protein DMPNN network, the message passing formula is modified as follows:

[0146] ;

[0147] in, This is the message vector pointing from residue i to residue j at the (t+1)th iteration. It is a ReLU activation function. To create a learnable matrix, the concatenation of node and edge features is mapped to the dimensional space of the message vector. This involves concatenating node features and edge features (equivalent to the sum of the original dimensions). Given the set of neighbors of the original node i, excluding the target node j, To aggregate all incoming edge messages pointing from neighbor k to neighbor i, The weight coefficient (scalar) is used to adjust the contribution of each incoming edge message to the update. Real edges are initially set to 1, and virtual edges to 0.2. This refers to the incoming edge message from neighbor k to neighbor i.

[0148] Learnability of virtual edge weights: The initial weight of each virtual edge is set to 0.2, significantly less than that of real edges (1.0), to ensure its "weak connection" nature. Based on this, the virtual edge weights are defined as learnable parameters, allowing the model to automatically fine-tune them according to task requirements during training. In this way, the model can dynamically discover potential long-range dependencies without compromising the dominance of real interactions.

[0149] Specifically, the final virtual edge weights for each UGT model are shown in Table 5.

[0150] Table 5

[0151]

[0152] (2) Fusion and Prediction Module:

[0153] Gated Fusion Module: A gating mechanism-based fusion method is proposed for the joint process of molecular characterization and protein binding site characterization. Specifically, two learnable gating coefficients are calculated separately:

[0154] ;

[0155] ;

[0156] in, For compound characterization vectors, This is a vector representing the UGT enzyme binding site. For the Sigmoid function, and For the science department parameter matrix, , The dynamic gating parameter takes values ​​of [0,1] and controls the ratio of compound and UGT enzyme features to the final combined feature, respectively.

[0157] The final joint representation vector is defined as:

[0158] in, This indicates element-wise multiplication. This mechanism allows the model to dynamically adjust the fusion strength of two types of modal information based on the characteristics of specific molecule-enzyme pairs, rather than simply splicing them together, thus capturing interaction patterns more flexibly.

[0159] CNN Fusion and Prediction: The joint representation vector z is input into a CNN module consisting of multiple convolutional, activation, and pooling layers to further extract deep features of molecule-enzyme interactions. The CNN output is finally passed through a fully connected layer with a sigmoid activation function, outputting the substrate probability value corresponding to each UGT subtype. Based on the output substrate probability value, it is determined whether the molecule or compound is a substrate of the corresponding UGT subtype. The threshold is set to 0.5. When the output substrate probability value is higher than 0.5, it is judged as a positive sample (substrate); otherwise, it is a negative sample (non-substrate).

[0160] Step 3: Model Training and Evaluation

[0161] Loss function: The binary cross-entropy loss function is used. This loss function is used for training for each UGT subtype task.

[0162] Training process: A separate model instance is trained for each UGT subtype (UGT1A1, UGT1A4, UGT1A6, UGT1A8, UGT1A9, UGT1A10, UGT2B7, UGT2B15) prediction task. Gradient descent training is performed using an optimizer (such as Adam). Early stopping is employed to prevent overfitting by monitoring performance on the validation set, and the best-performing model is saved.

[0163] Performance evaluation: The model's performance was evaluated on an independent test set using accuracy (ACC) and area under the curve (AUC).

[0164] Accuracy (ACC) reflects the proportion of correctly classified samples in all predicted samples and is a fundamental metric for evaluating the performance of binary classification models. In model evaluation, we used a threshold of 0.5 for positive and negative samples. When the substrate probability value output by the model is higher than 0.5, it is judged as a positive sample (substrate), that is, the molecular compound is judged as the substrate of the corresponding UGT subtype; otherwise, it is a negative sample (non-substrate).

[0165] The area under the curve (AUC) is the area under the receiver operating characteristic (ROC) curve. The ROC curve is plotted with the false positive rate (FPR) on the horizontal axis and the true positive rate (TPR) on the vertical axis. A UC value closer to 1 indicates better classification performance; an AUC value of 0.5 indicates performance comparable to random guessing. The AUC metric is robust to imbalanced samples and is particularly suitable for evaluating bioinformatics tasks with positive-negative sample imbalance, such as UGT substrate prediction.

[0166] Example 1

[0167] This embodiment provides the size of the UGT subtype substrate prediction dataset for training and evaluating the model proposed in this invention. The data are sourced from publicly available bioactivity databases (such as ChEMBL), ADMET benchmark datasets (such as TDC), and literature data. The training and validation sets will be partitioned in an 8:2 ratio using a scaffold-split-based partitioning strategy to ensure significant differences in the chemical backbones of the training and validation sets.

[0168] Data preprocessing:

[0169] Molecular pretreatment: The SMILES string is normalized using an open-source toolkit to remove salts and solvents and neutralize molecular charges.

[0170] Protein pretreatment: Local sequences near the binding sites are extracted from the protein sequence of the UGT enzyme and normalized.

[0171] The dataset size for the targeted UGT subtype is shown in Table 6.

[0172] Table 6

[0173]

[0174] Example 2

[0175] This embodiment is based on the prediction of UGT enzyme substrates using a fusion model.

[0176] This embodiment provides a complete and reproducible process for training and evaluating the fusion model proposed in this invention.

[0177] Technology stack:

[0178] Programming language: Python 3.9;

[0179] Deep learning framework: PyTorch 1.10;

[0180] Graph neural network library: PyTorch Geometric (PyG);

[0181] Cheminformatics database: RDKit;

[0182] Protein informatics library: Biopython.

[0183] 1. Model Architecture and Parameter Configuration

[0184] The fusion model in this embodiment uses the following hyperparameter configuration:

[0185] DMPNN molecular encoder: The number of layers is set to 4, and the hidden layer dimension is 512.

[0186] DMPNN protein encoder: The number of layers is set to 4, and the hidden layer dimension is 512.

[0187] CNN fusion module: Composed of a multi-layer 1D convolutional network.

[0188] Input layer: Receives the spliced ​​molecular-protein joint feature vector (dimension: 1024, from the output of two 512-dimensional DMPNN encoders).

[0189] First convolutional layer group:

[0190] 1D convolutional layer: kernel size 3, output channels 512, stride 1, padding 1;

[0191] Batch Normalization Layer (BatchNorm1D);

[0192] ReLU activation function;

[0193] Max pooling layer: pooling window size 2, stride 2;

[0194] Output dimension: 512;

[0195] Second convolutional layer group:

[0196] 1D convolutional layer: kernel size 5, output channels 256, stride 1, padding 2;

[0197] Batch Normalization Layer (BatchNorm1D);

[0198] ReLU activation function;

[0199] Max pooling layer: pooling window size 2, stride 2;

[0200] Output dimension: 256;

[0201] Third convolutional layer group:

[0202] 1D convolutional layer: kernel size 7, output channels 256, stride 1, padding 3;

[0203] Batch Normalization Layer (BatchNorm1D);

[0204] ReLU activation function;

[0205] Global AvgPool1D;

[0206] Output dimension: 256;

[0207] Dropout layer: Dropout rate 0.1 to prevent overfitting;

[0208] Detailed structure of the prediction head:

[0209] Fully Connected Network (FFN):

[0210] First fully connected layer: Input dimension 256, output dimension 128

[0211] ReLU activation function

[0212] Dropout layer (rate 0.1)

[0213] Second fully connected layer: Input dimension 128, output dimension 64

[0214] ReLU activation function

[0215] Output layer: Input dimension 64, output dimension 1

[0216] Sigmoid activation function (outputs substrate probability value)

[0217] Training hyperparameters: Learning rate 3×10 −4 Batch size 64, Adam optimizer, Dropout rate 0.1.

[0218] 2. Training and Assessment

[0219] Training: A separate model instance is trained for each UGT subtype task. End-to-end training is performed on the training set, and early stopping is performed using the validation set.

[0220] Evaluation: Calculate the ACC and AUC values ​​for each UGT subtype task on a separate test set.

[0221] The results are shown in Table 7, indicating that the constructed prediction model has high accuracy.

[0222] Table 7

[0223]

[0224] In summary, this invention explicitly constructs a residue-level graph structure from the characteristics of UGT enzyme binding sites, including residue type, three-dimensional coordinates, solvent accessibility, secondary structure, and physicochemical properties, and encodes it using DMPNN. Experimental results show that on multiple UGT enzyme subtype substrate prediction tasks, the average AUC is greater than 0.72 on the test set (e.g., UGT1A1 AUC=0.7623, UGT2B7 AUC=0.7489), significantly improving prediction accuracy compared to single-molecule modality methods. In constructing the protein binding site graph, this invention, in addition to retaining real chemical interaction edges, further introduces virtual edges to enhance graph connectivity. The initial weight of the virtual edges is set to 0.2 and can be learned and adjusted during training, ensuring its weak connectivity nature. At the same time, it allows the model to utilize potential connections between distant residues when needed. This mechanism effectively alleviates the shortcomings of traditional GNN models in modeling long-distance dependencies of binding sites, enabling protein characterization to exhibit stable prediction performance in different UGT subtype tasks (ACC for each subtype is between 0.62 and 0.72). To achieve accurate classification and prediction of UGT enzyme substrates and non-substrates.

[0225] The applicant declares that the above description is only a specific embodiment of the present invention, but the protection scope of the present invention is not limited thereto. Those skilled in the art should understand that any changes or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention fall within the protection and disclosure scope of the present invention.

Claims

1. A method for predicting substrate specificity of UGT enzyme subtypes, characterized by, The prediction method includes: The system acquires compound and UGT enzyme data. Based on a directed message passing neural network, a molecular encoder is used to construct a compound molecular graph with atomic information as nodes and chemical bond information as edges. A protein encoder is used to construct a UGT enzyme binding site graph with amino acid residue information of the binding site of the UGT enzyme as nodes and chemical interaction information as edges. The system outputs compound characterization vectors and UGT enzyme binding site characterization vectors. The compound characterization vector and the UGT enzyme binding site characterization vector are fused using a gated fusion module to obtain a joint characterization vector. The joint characterization vector is then input into a convolutional neural network learning model, which outputs the substrate probability value corresponding to each UGT enzyme subtype. Based on the output substrate probability value, it is determined whether the compound is the substrate of the corresponding UGT enzyme subtype. The model is then trained and optimized to obtain a UGT enzyme metabolism prediction model with fusion protein binding site information. Obtain the joint characterization vector of the compound to be analyzed and the UGT enzyme, input it into the UGT enzyme metabolism prediction model containing the binding site information of the fusion protein, and output the substrate probability value.

2. The method of predicting substrate specificity of UGT enzyme subtypes according to claim 1, wherein, The acquisition of compound and UGT enzyme data includes: (1-1) Data Collection Structural information of compounds was obtained from public compound structure databases and represented in SMILES string format. Substrate and non-substrate tag information of different UGT enzyme subtypes was obtained; protein amino acid sequence information of UGT enzymes was obtained. (1-2) Data Preprocessing Remove compound records with incomplete structures, invalid SMILES, or duplicates; merge labels for multiple experimental records of the same compound and the same UGT enzyme subtype using consistency checks and majority voting; discard data records with missing labels. The SMILES string is standardized and a molecular object is generated. Hydrogen atom completion, aromaticity identification and bond type standardization are performed on the molecule. Preferably, the UGT enzyme subtype includes at least one of UGT1A1, UGT1A4, UGT1A6, UGT1A8, UGT1A9, UGT1A10, UGT2B7, or UGT2B15.

3. The method of predicting substrate specificity of UGT enzyme subtypes according to claim 1 or 2, characterized in that, The constructed compound molecular map includes: Construct atomic node feature vectors, which include atomic type, atomicity, number of explicit hydrogen atoms, number of implicit hydrogen atoms, hybridization type, formal charge, aromaticity, whether it is in a ring, and chiral label; Construct a chemical bond edge feature vector, wherein the chemical bond feature vector includes bond type, whether it is conjugated, whether it is in a ring, and stereochemistry; Based on the molecular encoder, a compound molecular graph is constructed using atomic node feature vectors and chemical bond edge feature vectors.

4. The method of predicting substrate specificity of UGT enzyme subtypes according to any one of claims 1 to 3, characterized in that, The constructed UGT enzyme binding site map includes: The amino acid residue information of the binding site of the UGT enzyme is extracted. The amino acid residue information of the binding site of the UGT enzyme includes amino acid residue type and structural environment feature information. The structural environment feature information includes the spatial coordinates of the amino acid residue, solvent-accessible surface area, secondary structure type, physicochemical properties of the residue, distance from the center of the binding site, local atomic density, and chemical environment descriptor. Nodes are constructed using the amino acid residue information; Edges are constructed using chemical interaction information of amino acid residues, including interaction type, interaction strength, and spatial distance; the interaction type includes covalent bond connection, spatial distance connection, and chemical interaction connection. Based on a protein encoder, a map of UGT enzyme binding sites is constructed using the constructed nodes and edges.

5. The method of predicting substrate specificity of UGT enzyme subtypes according to any one of claims 1 to 4, characterized in that, The message passing mechanism of the molecular encoder includes: ； wherein, is the t-th round from node to node message vector, is an activation function, is a weight matrix of the message linear layer, transforming to the message dimension; ； wherein, is the aggregated result of the neighbor messages of the node , is the set of nodes adjacent to the node , is the message vector from the node to the node in the t-th round. ； wherein, is a read-out molecular graph node j representation, is a splicing operation, is a node information representation of a node j of a molecular graph, is a representation of all adjacency information of a node j of a molecular graph.

6. The method for constructing a UGT enzyme metabolism prediction model based on the fusion protein binding site information according to any one of claims 1-5, characterized in that, The message passing mechanism of the protein encoder includes: ； in, This is the message vector pointing from residue i to residue j at the (t+1)th iteration. It is a ReLU activation function. To create a learnable matrix, the concatenation of node and edge features is mapped to the dimensional space of the message vector. To concatenate node features and edge features, Given the set of neighbors of the original node i, excluding the target node j, To aggregate all incoming edge messages pointing from neighbor k to neighbor i, The weight coefficients for directed edges are scalars that adjust the contribution of each incoming edge message to the update; real edges are initially set to 1, and virtual edges to 0.

2. This refers to the incoming edge message from neighbor k to neighbor i; Preferably, the virtual edge weights corresponding to UGT1A1 are 0.431, UGT1A4 are 0.129, UGT1A6 are 0.321, UGT1A8 are 0.371, UGT1A9 are 0.527, UGT1A10 are 0.415, UGT2B7 are 0.656, and UGT2B15 are 0.

589.

7. The substrate-specific prediction method for UGT enzyme subtypes according to any one of claims 1-6, characterized in that, The gating fusion module includes a gating coefficient calculation module. and : ； ； in, For compound characterization vectors, This is a vector representing the UGT enzyme binding site. For the Sigmoid function, and For the science department parameter matrix, , is a dynamic gating parameter, taking values ​​[0,1], which controls the ratio of compound and UGT enzyme features to the final combined feature, respectively; Preferably, the joint representation vector is defined as: ; in, This indicates element-wise multiplication.

8. The method for constructing a UGT enzyme metabolism prediction model based on the fusion protein binding site information according to any one of claims 1-5, characterized in that, The model training and optimization include: The model is trained using a binary cross-entropy loss function, with each UGT enzyme subtype treated as an independent task. An optimizer is used to train each task using gradient descent. Performance is monitored on the validation set, and an early stopping mechanism is employed to prevent overfitting. The model with the best performance is saved. The UGT enzyme subtypes include UGT1A1, UGT1A4, UGT1A6, UGT1A8, UGT1A9, UGT1A10, UGT2B7, or UGT2B15.

9. A UGT enzyme metabolism prediction model with information on fusion protein binding sites, characterized in that, The prediction model includes the UGT enzyme metabolism prediction model based on the fusion protein binding site information as described in any one of claims 1-8.

Images