A general drug covalent target prediction method based on information passing neural network

Abstract

This invention provides a general method for predicting covalent drug targets based on information transmission neural networks. It is a deep learning-based covalent drug-target response type prediction model, including a data preparation module, a multi-level feature engineering template, and a deep learning prediction module. This invention improves the accuracy and reliability of target prediction; enhances target discovery efficiency; reduces R&D costs; and has good feasibility and application value.

Description

Technical Field

[0001] This invention relates to a general method for predicting covalent drug targets based on an information transmission neural network. Background Technology

[0002] Covalent modulators, due to their unique mechanism of action and significant pharmacodynamic advantages, have gradually become a research hotspot in molecular targeted therapy, precision drug design, and drug discovery and development. Compared with traditional non-covalent small molecule drugs, covalent modulators can achieve long-term target occupancy by forming stable covalent bonds with specific amino acid residues in target proteins, thus exhibiting higher binding specificity, stronger duration of action, and longer in vivo half-life. These drugs have significant advantages in reducing dosing frequency, overcoming endogenous ligand competition, and addressing drug resistance mutations, and have shown promising application prospects in multiple therapeutic areas, including oncology, autoimmune diseases, and neurological disorders.

[0003] In covalent drug design, the covalent "warhead" is a key structural unit that determines its reactivity and selectivity. Existing research indicates that covalent warheads can target a wide variety of protein amino acid residues, including but not limited to cysteine, lysine, serine, threonine, tyrosine, and histidine. Correspondingly, the types of covalent reactions are also diverse, including nucleophilic addition, nucleophilic substitution, Michael addition, Schiff base formation, and redox-related reactions. Different amino acid residues exhibit significant differences in spatial accessibility, pKa value, microenvironment polarity, and conformational flexibility, making the occurrence of covalent reactions highly dependent on the local chemical environment of the protein's active pocket.

[0004] Currently, drug target discovery still primarily relies on experimental screening, bioinformatics analysis, and literature mining. While these methods have advanced drug development to some extent, they still have significant limitations in practical applications. Experimental screening methods typically require substantial human and material resources, have long research cycles, and are costly, making them unsuitable for high-throughput and rapid iteration requirements. Traditional bioinformatics analysis methods, on the other hand, often rely on inferences based on single data dimensions or static correlations, resulting in limited predictive accuracy and difficulty in accurately characterizing the complex interactions between drug molecules and potential targets.

[0005] With the development of computer simulation technology, some studies have begun to introduce methods such as molecular docking, deep learning, and quantum chemical calculations for the prediction of drug covalent targets. Although quantum chemical calculations can finely characterize molecular interactions, they suffer from high computational complexity, high computational cost, limited system scale, and strong dependence on conformation and environment in large-scale drug target discovery and identification of novel binding types, making them difficult to directly apply to large-scale target screening.

[0006] The relevant literature currently reported includes: a prediction model for covalently reactive cysteine ​​(Cys) based on support vector machine: by calculating the pKa value of cysteine ​​and surrounding amino acid residues, the area of ​​the solvent exposure region, and other structural features, a support vector machine (SVM) prediction model is used to realize the automatic identification of cysteine ​​residues suitable for covalent ligand design, with a prediction accuracy of about 0.73 (Zhang W, Pei J, Lai L. Statistical Analysis and Prediction of Covalent Ligand Targeted Cysteine ​​Residues. J Chem Inf Model. 2017 Jun 26;57(6):1453-1460. doi: 10.1021 / acs.jcim.7b00163. Epub 2017 May 30. PMID:28510428.). DeepCoSI, a covalently reactive Cys site prediction model based on a neural network for information transfer, constructs a covalent interaction map by calculating binding pocket features and the interaction information between cysteine ​​residues and their surrounding microenvironment. This map is then used through a deep graph learning model to identify druggable covalently binding sites in proteins. Validation results on an external test set simulating real-world application scenarios demonstrate that DeepCoSI can effectively distinguish between covalently binding sites and non-binding sites (Hongyan Du, Dejun Jiang, Junbo Gao, Xujun Zhang, Lingxiao Jiang, Yundian Zeng, Zhenxing Wu, Chao Shen, Lei Xu, Dongsheng Cao, et al. Proteome-WideProfiling of the Covalent-Druggable Cysteines with a Structure-Based DeepGraph Learning Network. Research. 2022;2022:DOI:10.34133 / 2022 / 9873564). The CovCysPredictor is a reactive Cys prediction model based on an interpretable neural network model. It integrates structural data from the CovPDB and CovBinderInPDB databases to construct an interpretable machine learning model to predict potentially covalently modifiable cysteine ​​residues.This method comprehensively considers the physicochemical properties of cysteine, such as pKa, solvent exposure, and electrostatic characteristics, as well as the protein-ligand binding pocket descriptor. The resulting logistic regression model achieved a median F1 score of approximately 0.73 on the independent test set and demonstrated good site discrimination ability in whole-protein structure testing.

[0007] Existing methods can only predict reactivity for cysteine ​​residues and lack the ability to predict reactivity for other covalently reactive amino acids such as serine and lysine. The lack of modules for processing and integrating ligand structure data prevents virtual screening of specific covalent inhibitors.

[0008] One-Shot Rational Design of Covalent Drugs with CovalentLab, https: / / doi.org / 10.1021 / jacsau.5c01161 CovalentLab was introduced, an interactive computing platform integrating ligand-based and warhead-based strategies, for rational planning of the design of covalent ligands.

[0009] CovDocker: Benchmarking Covalent Drug Design with Tasks, Datasets, and Solutions (https: / / doi.org / 10.1145 / 3711896.3736896). This paper addresses the difficulty of existing molecular docking methods in handling covalent bond formation and related structural changes. It proposes the first comprehensive covalent docking benchmark, CovDocker, which innovatively decomposes the covalent docking process into three core tasks and provides corresponding datasets and evaluation metrics. Methodologically, it improves models such as Uni-Mol and Chemformer to handle the three tasks: reaction site prediction, covalent reaction prediction, and covalent docking. Experiments use datasets integrated from CovPDB and CovBinderInPDB, which, after preprocessing, form training, validation, and test sets containing various reaction mechanisms and target amino acids. Results show that the model based on this benchmark outperforms traditional methods in all tasks.

[0010] The shortcomings of the published literature and the solutions proposed in this invention mainly include the following aspects:

[0011] (1) Existing methods have the problem of insufficient characterization of covalent reaction sites: Existing software such as CovalentLab and CovDocker mainly adopts a ligand-first strategy for modeling, focusing on the analysis of ligand structural information, but lacks a systematic characterization of the interaction between the covalent warhead and the target amino acid reaction site, which may affect the accuracy of the judgment of covalent binding reaction. This invention introduces a covalent warhead structure identification and characterization module into the model structure to achieve a synergistic characterization of the interaction between the ligand reactive group and the covalent warhead, thereby more accurately characterizing the structural features related to covalent reactions and effectively improving the prediction accuracy of the model.

[0012] (2) Existing methods do not adequately consider the flexibility of protein conformation: Existing methods such as CovalentLab and CovDocker typically employ the rigid pocket assumption, which assumes that the protein-binding pocket maintains a static structure during ligand binding, thus ignoring the influence of protein conformational changes on the pocket's spatial shape and chemical environment. This invention constructs a network of contact relationships between residues to model the ligand-binding region graphically, thereby enhancing the model's ability to express the dynamic structural features of proteins and improving the generalization ability and prediction accuracy of the prediction model.

[0013] (3) Existing methods have the problem of limited representation of receptor pockets: Existing technologies such as CovalentLab and CovDocker typically use fpockets to model the binding region at the entire atomic scale. However, due to limitations in computational resources and model complexity, their description of receptor pockets is usually limited to atoms within a range of about 3–5 Å from the ligand binding site, and they do not systematically model the interactions between pocket residues. This invention, by providing a coarse-grained representation of the receptor structure, reduces computational complexity while extending the representation range of the receptor binding region to 5–10 Å, and further introduces the interaction relationships between pocket residues, thereby enhancing the model's ability to express the local structural environment of the receptor and improving prediction accuracy.

[0014] (4) Existing methods have insufficient predictive ability for covalent reactive residues in shallow or semi-open pockets: Existing methods perform well in predicting covalent reactive residues in closed pockets, but the prediction results are often concentrated on traditional targets such as kinases and nuclear receptors, and the ability to identify reactive residues in semi-open pockets or shallow pockets on the protein surface is weak. This invention introduces the TPS algorithm to enhance the expression of topological features between residues, which not only improves the efficiency of information transmission between amino acid residues, but also enhances the scalability of model information encoding, thereby significantly improving the model's predictive ability for covalent reactive residues in semi-open and shallow pockets.

[0015] While deep learning-based neural network models can be used to predict target reactivity, existing prediction models are limited to predicting the reactivity of cysteine ​​(Cys) residues in receptors. There are no reports of universal covalent target prediction tools that combine ligand structure, covalent warhead, and receptor structure. Summary of the Invention

[0016] To overcome the shortcomings of existing drug target prediction techniques in terms of model applicability, multi-source information integration, and result evaluation, this invention proposes a novel general scheme for predicting covalent drug targets based on information transfer neural networks. This scheme uses reactivity priority theory as one of its core theoretical foundations, incorporates ligand and receptor structural information, and combines a multi-level computational analysis strategy to systematically characterize the interaction features between drug molecules and potential targets.

[0017] This invention provides a general method for predicting covalent drug targets based on information transmission neural networks. It is a deep learning-based covalent drug-target response type prediction model, including a data preparation module, a multi-level feature engineering template, and a deep learning prediction module.

[0018] The data preparation module involves collecting and preprocessing protein structures, screening for covalently reactive small molecule complexes, constructing a dataset, and then dividing the database.

[0019] The multi-level feature engineering module adopts a multi-channel graph neural network input strategy to construct three complementary feature matrices respectively; the three feature matrices are the ligand feature matrix and adjacency matrix, the receptor feature matrix and adjacency matrix, and the covalent warhead and its surrounding residues feature matrix and adjacency matrix.

[0020] Before constructing the receptor feature matrix, the receptor is first characterized using coarse-grained receptor characterization.

[0021] The deep learning prediction module includes an information encoding and representation learning module, an information reading module, a feature concatenation and fully connected prediction module, and an output prediction result module.

[0022] The data preparation module includes the following steps:

[0023] S1. Protein structure collection and pretreatment:

[0024] Based on the PDBID in CovBinderInPDB, protein crystal structures from human and mouse sources are collected from the PDB database, a list of protein PDBIDs is established, and annotation information is added; based on the PDBIDs in the list, protein structures are downloaded in batches using the BioPython package, and then subunit separation, solution removal, and redundancy removal are performed to obtain a standardized monomer protein structure database in PDB format.

[0025] S2. Establish a covalent warhead identification module:

[0026] The CovBinderInPDB database is retrieved, and the adduct_smiles and binder_smiles items of each entry in the database are compared. Based on the differences in the SMILES descriptor, the SMARTS descriptor is established to define covalent warheads for the determination of covalent warheads in compounds. The difference patterns are converted into queryable structural patterns. The module contains 118 covalent reactive groups and the seven reaction types involved, including substitution reactions, Michael addition reactions, epoxide addition reactions, acetal reactions, disulfide bond formation, and aldehyde-amine condensation.

[0027] S3. Constructing the dataset:

[0028] For each covalent complex screened in S1, the protein and small molecule portions of the complex were separated using the Biopython package. The covalent warhead identification module established in S2 was used to label the covalently reactive structures (N1 in total) in the small molecules. Five classes of nucleophilic residues were labeled according to the amino acid residue names of the protein: Cys / Arg / Ser / Thr / His (N2 in total). These two variables were combined to generate candidate datasets (N1×N2 in total) using an exhaustive search method. The analyzed covalently reactive information and nucleophilic residues were then added to the candidate datasets. All structures in CovBinderInPDB were processed using the above method, resulting in a dataset containing 868,415 entries. The training set, internal test set, and external test set were split in a 90%:9%:1% ratio to obtain a three-level dataset with ground-truth labels.

[0029] The annotation information in step S1 includes HetID, UniprotID, protein name, and protein subunit; the solution removal includes deleting water, glycerol, and DMSO solvent molecules; and the redundancy removal includes removing duplicate protein structures.

[0030] The construction of the ligand feature matrix and adjacency matrix is ​​based on the graph neural network (GNN) paradigm, which transforms the ligand molecule into a topological graph structure, and includes the following steps:

[0031] a. Use the Chem.MolFromSmiles function in RDKit to convert the SMILES descriptor into the molecular structure and construct the planar structure of the compound;

[0032] b. Extract five types of parameters: atom type, atom valence state, number of hydrogen atoms, hybridization type, and atom charge, and establish an atom feature vector matrix;

[0033] c. Calculate two types of parameters based on the structure of the compound: chemical bond type and whether it is a conjugated structure, and establish a chemical bond feature vector matrix;

[0034] d. Construct the adjacency matrix of atoms based on chemical bonds;

[0035] e. Extract the atomic feature vector matrix, chemical bond feature vector matrix, and atomic adjacency matrix as inputs to the ligand characterization part of the model.

[0036] The method for constructing the receptor feature matrix and adjacency matrix is ​​to achieve a precise graph-structured representation of the microenvironment of covalent reaction sites through the topological path search (TPS) algorithm and multi-scale feature encoding.

[0037] Includes the following steps:

[0038] a. Using Biopython, the PDB file of the complex is converted into a stereostructure. Amino acids in the range of 7-10 Å are extracted with covalently reactive residues (R) as the center. An active pocket diagram is generated with 3.5 Å as the threshold for non-bonded interactions.

[0039] b. Using the amino acid closest to the covalently reactive residue (R) as the topological starting point, the TPS algorithm is used to find the one-way loop structure in the active pocket diagram; then, using the amino acid on the one-way loop as the starting point, the 3-5 residues closest to it are retained; finally, the one-way loop structure and its neighboring residue set are merged to form a coarse-grained structural representation of the receptor binding pocket.

[0040] c. Extract four types of information from the coarse-grained acceptor representation: amino acid residue type, number of hydrogen atoms, hydrophobicity, and residue charge, and establish an amino acid feature vector matrix.

[0041] d. Extract three types of properties from the non-covalent interactions between amino acid residues: whether each amino acid residue pair contains an aromatic ring and the distance between amino acid residues, and generate a non-covalent interaction feature vector matrix.

[0042] e. Construct an amino acid residue adjacency matrix based on the above non-covalent interactions;

[0043] f. Extract the amino acid feature vector matrix, the non-covalent interaction feature vector matrix, and the amino acid residue adjacency matrix as inputs to the receptor characterization part of the model.

[0044] The method for constructing the feature matrix and adjacency matrix of the covalent warhead and its surrounding residues includes the following steps:

[0045] a. Use the covalent warhead identification module established by S2 to identify the covalent warhead of the compound, extract the atoms of the covalent warhead, and realize the location of the warhead atoms;

[0046] b. Extract amino acid residues within 3-5 Å of the covalently reactive amino acid residues (R), and construct a microenvironment map of the covalent warhead site based on the non-covalent interactions between residues;

[0047] c. Connect the covalent reaction site to the surrounding amino acid residues to construct a virtual interface;

[0048] d. Extract the covalent warhead portion of the virtual interface and extract five types of parameters: atom type, atom valence state, number of hydrogen atoms, hybridization type, and atom charge. Establish the feature vector matrix of the covalent warhead portion.

[0049] e. Calculate the chemical bond type and whether it is a conjugated structure based on the chemical structure of the covalent warhead, and establish a chemical bond feature vector matrix;

[0050] f. Construct the adjacency matrix of the covalent warhead based on chemical bonds;

[0051] g. Extract the amino acid residue portion of the virtual interface and calculate four types of information: amino acid residue type, number of hydrogen atoms, hydrophobicity, and residue charge. Establish a feature vector matrix of the microenvironment of the covalent warhead.

[0052] h. Extract three types of properties: the type of non-covalent interaction between amino acid residues, whether all amino acid residue pairs contain aromatic rings, and the distance between amino acid residues, and generate a feature vector matrix of non-covalent interaction in the microenvironment of the covalent warhead.

[0053] i. Construct a partial amino acid residue adjacency matrix for the microenvironment based on the above non-covalent interactions;

[0054] j. Merge the feature vector matrix of the covalent warhead and the feature vector matrix of the microenvironment to construct the virtual interface feature vector matrix;

[0055] k. Combine the chemical bond eigenvector matrix of the covalent warhead part and the non-covalent interaction eigenvector matrix of the microenvironment part to construct the virtual interface interaction eigenvector matrix;

[0056] 1. Merge the adjacency matrix of the covalent warhead and the adjacency matrix of the amino acid residues to construct a virtual interface adjacency matrix;

[0057] m. The merged virtual interface feature vector matrix, virtual interface interaction feature vector matrix, and virtual interface adjacency matrix are used as inputs to the virtual interface representation part of the model.

[0058] The information encoding and representation learning module constructs three complementary feature matrices with topological information, propagates this information through multiple rounds of neighborhood information, and achieves local structure embedding updates and information modulation, including the following steps:

[0059] a. Edge-level message generation:

[0060] First, adjacency information is constructed based on edge connection relationships. The hidden states of adjacent nodes (including atoms or residues) are fused and mapped with the features of the edges (covalent or non-covalent interactions) to generate edge-level message vectors.

[0061] b. Node message aggregation:

[0062] The messages from the neighborhood are weighted and summed to form a candidate update representation for the node;

[0063] The characterization formula is as follows:

[0064] Let the molecular, residue, and interface node diagram be represented as follows: ,in Represents a set of nodes (atoms, amino acid residues, and denominator nodes). This represents a set of edges (chemical bonds, non-covalent interactions, or interactions between nodes). For any node... In its first The hidden state in the next message passing iteration is represented as: Corresponding edge The edge features are represented as ,node The set of neighboring nodes is denoted as ;

[0065] During the message passing phase, a neighborhood message is first constructed based on node and edge features, calculated using the following formula:

[0066] ;in:

[0067] · Indicates from neighboring nodes Passed to node The message vector;

[0068] · and The learnable parameter matrix;

[0069] · For bias terms;

[0070] · This represents a non-linear activation function.

[0071] Subsequently, for the nodes Aggregate all neighbor node messages to obtain a node-level message representation:

[0072] ;

[0073] c. GRU gated fusion:

[0074] During the node state update phase, a gated recurrent unit (GRU) is introduced to perform gated fusion of the current node's hidden state and aggregated messages. By controlling the ratio of historical information to new input information through update gates and reset gates, stable gradient propagation and long-term dependency modeling are achieved; each molecular graph outputs a 64-dimensional embedding vector.

[0075] Finally, the current node state and the aggregated message are merged and updated using a gated recurrent unit (GRU) to obtain the node representation for the next iteration:

[0076] ;

[0077] 8. The information readout module described above is an adaptive aggregation from atomic-level embedding to molecular-level global representation based on a self-attention mechanism, including the following steps:

[0078] This is used to transform the atomic-level embedding representation obtained in the message passing stage into a molecular-level global representation. Taking the atomic embedding vector of each molecule as input, the correlation weight between different nodes is calculated based on the self-attention mechanism. Through weighted aggregation, the adaptive enhanced expression of key structural units and potential reaction sites is achieved.

[0079] The specific algorithm and formula are as follows:

[0080] Suppose a certain graph contains Each node (a node in an atom, amino acid residue, or interface) is represented by its node embedding through a message passing network as follows:

[0081] ;

[0082] in For the first The feature vector of each node For the embedding dimension, the correlation weights between atoms are calculated using a multi-head self-attention mechanism, and the atom representation is updated. The calculation process can be uniformly represented as follows:

[0083] ;

[0084] in:

[0085] · This represents multi-head attention.

[0086] · This represents the transformation of a feedforward neural network;

[0087] · Layer Normalization;

[0088] · This indicates a global average pooling operation;

[0089] · This is the final molecular-level representation vector.

[0090] The multi-head attention calculation is as follows:

[0091] ; ;

[0092] in:

[0093] · , , For the first The query, key, and value matrix of each attention head;

[0094] · This is the learnable parameter matrix.

[0095] Through the above operations, the molecular-level embedding vector can be obtained:

[0096] ;in This represents the node representation after attention and feedforward network updates.

[0097] The feature splicing and fully connected prediction module is an end-to-end prediction module for the fusion of multi-scale molecular characterization and covalent reaction types, specifically including the following steps:

[0098] a. Feature concatenation and fully connected prediction: Embedding vectors between different molecular maps are concatenated using the `Concatenation` function (`tf.concat`), outputting a 192-dimensional fused vector.

[0099] b. Feedforward Neural Network (FNN) Architecture:

[0100] The feedforward neural network (FNN) is used to process the feature vector matrix of the complex to predict the covalent reactivity and reaction mode of small molecules and receptors. The FNN consists of three fully connected layers, each with 1024, 512, and 8 hidden units respectively. By performing multi-layer nonlinear mapping on the fused features, the FNN outputs multi-classification results of covalent bond types. Each fully connected layer is connected by a Dropout layer.

[0101] The output prediction module outputs eight types of covalent reactions: substitution reactions, Michael addition reactions, epoxide addition reactions, acetal reactions, disulfide bond formation, aldehyde-amine condensation, other addition reactions, and no reaction. This invention uses the flexible pocket assumption, replacing fixed coordinates with an adjacency matrix, resulting in stronger transfer learning capabilities. The model employs a coarse-grained method, incorporating a wider range of amino acid residues and improving the predictive ability for shallow pockets and surface residue reactivity in proteins. The "covalent warhead structure identification and characterization module" defines the binding of receptors and ligands in the covalent pocket region, further improving prediction accuracy. This invention significantly improves predictive capabilities, particularly the ability to predict amino acid covalent reactivity within protein pockets.

[0102] This invention identifies a target binding characteristic that differs from traditional conformational complementarity or empirical energy evaluation by constructing a computational model of the potential interaction region between drug molecules and target proteins. At the same time, by correlating the above molecular-level computational results with other computational analysis results, the invention achieves synergistic utilization of the advantages of different computational theories and avoids the problem of limited predictive ability of a single model.

[0103] Furthermore, this invention introduces a mixed-particle-size simulation method in the target prediction process, which simplifies the computational complexity without affecting prediction accuracy, enabling efficient screening of large compound libraries or protein databases. The technical solution features a clear computational flow and ease of implementation in a computer environment, effectively improving the accuracy and reliability of drug target prediction and possessing significant practical application value.

[0104] The beneficial effects of this invention are:

[0105] (1) Improve the accuracy and reliability of target prediction. This invention breaks through the limitations of traditional methods that rely solely on spatial conformation matching or empirical energy functions for judgment by using a hybrid granularity simulation method, and improves the model's compatibility with protein dynamic conformational changes, thereby effectively reducing the false positive rate and improving the accuracy and reliability of target prediction results.

[0106] (2) Improve target discovery efficiency and reduce R&D costs.

[0107] The technical solution described in this invention can complete the preliminary screening and optimization of drug targets in a computer environment, effectively reducing the reliance on large-scale experimental screening, thereby significantly shortening the research cycle, reducing human and material costs, and improving the overall efficiency of drug development.

[0108] (3) It has good feasibility and promotional application value.

[0109] The technical solution of this invention has a clear calculation process, is easy to implement through computer programs, can be combined with existing computing platforms and experimental verification processes, has strong engineering feasibility and practical application value, and is suitable for target discovery research in drug development and related fields. Attached Figure Description

[0110] Figure 1 For model flowchart;

[0111] Figure 2 Flowchart for the small molecule ligand structure characterization module;

[0112] Figure 3 A schematic diagram of the receptor coarse-grained module;

[0113] Figure 4 This is a flowchart of the receptor characterization section;

[0114] Figure 5 A schematic diagram representing a covalent warhead;

[0115] Figure 6 The SPR results of mouse-derived alpha7-nAChR and cinnamaldehyde are shown in the figure.

[0116] Figure 7 The SPR results for murine alpha7-nAChR mutants (C197A, C198A) and cinnamaldehyde are shown in the figure. Detailed Implementation

[0117] The following is a flowchart of the universal drug covalent target prediction method based on information transmission neural network of this invention. Figure 1 :

[0118] Example 1: A general drug covalent target prediction method based on information transmission neural networks according to the present invention.

[0119] S1. Protein Structure Collection and Preprocessing: Collect protein crystal structures from human and mouse sources from the PDB database, establish a PDBID list of proteins, and add annotation information such as HetID, UniprotID, protein name, and protein subunits; based on the PDBIDs in the list, use the BioPython package to batch download protein structures, split protein subunits, remove solvent molecules such as water, glycerol, and DMSO, remove duplicate protein structures (the protein and small molecule ligands are the same), and finally save each subunit as a separate PDB file to construct a protein structure database;

[0120] S2. Screening for protein structures containing covalent regulators: Based on the PDBID and HetID in CovBinderInPDB, the protein structure database established in S1 is screened to remove protein structures that do not contain covalent regulators; adduct_smiles and binder_smiles in the CovBinderInPDB database are compared, and the parts with differences are extracted to establish a "covalent warhead identification module based on SMARTS descriptors" and to label the types and reaction types of covalent warheads in the training set; record the PDBID of each protein, the amino acid sequence of the potential covalently reactive amino acid residues (cysteine, arginine, serine, threonine, and histidine), the sequence of the protein chain in which they are located, the SMILES descriptor of small molecules, and the potential reactive sites in small molecules to establish a dataset;

[0121] S3. Dataset Construction: Using the BioPython package, separate the protein and small molecule parts of the complex. Calculate the nearest distance between the small molecule ligand and the potentially covalently reactive amino acid residues (cysteine, arginine, serine, threonine, and histidine) in the protein. A nearest distance less than 1.8A indicates covalent bond formation, while a nearest distance greater than 1.8A indicates no covalent bond formation. Add the obtained covalent reactivity information to the dataset built in S2, and split the dataset into training set, internal test set, and external test set in a ratio of 100:1:0.1.

[0122] S4. Small Molecule Ligand Structure Characterization Module: Using the Chem.MolFromSmiles function in RDKit, SMILES descriptors are converted into molecular structures, and ligand maps are constructed based on the ligand molecule structures. L (V) L E L V in the diagram L Representing the set of all atoms in the ligand structure, five parameters are calculated for each atom: atom type, valence state, number of hydrogen atoms, hybridization type, and atomic charge, generating a 30-dimensional one-hot atomic feature vector matrix; E L Given a set of chemical bonds in a compound molecule, calculate two types of parameters: bond type and whether it is a conjugated structure, to generate a 6-dimensional one-hot chemical bond eigenvector matrix. Assume each molecule contains N... L There are 10 atoms, and the relationships between the atoms are represented by an N. L ×N L The adjacency matrix describes the structure of a small molecule. Ultimately, the structure of the small molecule is transformed into an atomic eigenvector matrix, a chemical bond eigenvector matrix, and an atomic adjacency matrix (see...). Figure 2 ).

[0123] S5. Receptor Coarse-grained Module: Biopython is used to parse the protein PDB file to reconstruct its three-dimensional spatial structure. A set of amino acid residues within a 6–10 Å range from a known covalently reactive amino acid residue (hotspot residue, denoted as R) is extracted to characterize the receptor's active binding pocket. To reduce model complexity and improve computational efficiency, the receptor structure is coarse-grained, retaining only the residue type, residue number, and spatial distance from the hotspot residue R as structural features. Based on this, to more accurately characterize the local chemical environment of the covalently reactive residues, a contact map of the active pocket is constructed. Using the amino acid residue closest to the covalently reactive residue as the starting node, the TPS (Topology Path Search) algorithm is used to identify unidirectional loop topologies (such as...) within the pocket map structure. Figure 3 (R1–R7). Then, using the amino acid residues in this ring structure as a new starting point, the 3–5 nearest neighboring residues are further screened to supplement local structural information. Finally, the unidirectional ring structure and its neighboring residue set are merged to form a coarse-grained structural representation of the receptor binding pocket.

[0124] S6. Coarse-grained receptor characterization module: Based on the coarse-grained active pocket structure representation, construct a receptor active pocket map (G... R (V) R E R V in the diagram R The set representing all amino acid residues in the active pocket is used to statistically analyze four types of information for each amino acid residue: amino acid type, number of hydrogen atoms, hydrophobicity, and residue charge, generating a 48-dimensional one-hot amino acid feature vector matrix. R This represents the set of interactions between amino acid residues in the active pocket, statistically analyzing non-covalent interactions within 3.5 Å of the amino acid spacing, and three other properties: whether the amino acid residues contain aromatic rings, and the distance between amino acid residues. Non-covalent interactions include electrostatic interactions (interactions between ARG, LYS, ASP, and GLU, categorized as attraction and repulsion), polar interactions (interactions between THR, SER, TYR, and HIS), and nonpolar interactions (interactions between other amino acid residues). Amino acid residues containing aromatic rings include HIS, TYR, and TRP. The interaction statistics generate a 15-dimensional one-hot non-covalent interaction matrix. Let N be the number of amino acid residues in each receptor active pocket. R One amino acid, with N-type amino acid residues forming N-type bonds. R ×N RThe adjacency matrix is ​​ultimately transformed into an amino acid residue feature vector matrix, a non-covalent interaction feature vector matrix, and an amino acid adjacency matrix (see [link to relevant documentation]). Figure 4 ).

[0125] S7. Covalent Warhead Structure Identification and Characterization Module: Using the covalent warhead identification module established in S2, atoms in the covalent warhead portion of the compound are located. Amino acid residues at a distance of 3-5 Å from the covalently reacting residues are extracted, and the residues and covalent warhead atoms are characterized using the coarse-grained processing method established in this method. The reaction sites in the covalent phase are linked to the covalently reacting amino acid residues to construct a virtual interface and an interface activity pocket diagram (G). I (V) I E I V I Given a set of virtual interface nodes, five types of parameters are calculated for each virtual interface node: node type, number of hydrogen atoms in the node, whether the node is aromatic, node hybridization type, and node charge. This generates a 30-dimensional one-hot node eigenvector matrix. I Given a set of interactions between nodes in a virtual interface, we calculate two types of parameters: interaction type and whether it involves conjugate relationships, generating a 6-dimensional one-hot node interaction feature vector matrix. Let each virtual interface contain N... I There are N nodes, and N interfaces are connected between them. I ×N I The adjacency matrix is ​​then used. Finally, the virtual interface part is transformed into an interface node feature vector matrix, an interface node interaction feature vector matrix, and an interface node adjacency matrix (see...). Figure 5 ).

[0126] S8, Information Encoding and Representation Learning Module: This module simultaneously receives topological information from three molecular graphs. Through multi-round neighborhood information propagation, it achieves local structure embedding updates and information modulation. First, it constructs adjacency information based on edge connections, fusing the hidden states of adjacent nodes (including atoms or residues) with edge features (covalent or non-covalent interactions) to generate edge-level message vectors. Then, it weights and sums the messages from the neighborhood to form candidate update representations for nodes. During the node state update stage, a gated recurrent unit (GRU) is introduced to perform gated fusion of the current node's hidden state and aggregated messages. The update and reset gates control the ratio of historical information to new input information, achieving stable gradient propagation and long-term dependency modeling. This structure is equivalent to constructing a recurrent neural network with a memory mechanism in the graph topological space, thereby enhancing the model's ability to express long-range electronic effects, spatial proximity relationships, and reactive site environments. Each molecular graph outputs a 64-dimensional embedding vector.

[0127] The information encoding and representation learning module constructs three complementary feature matrices with topological information, propagates this information through multiple rounds of neighborhood information, and achieves local structure embedding updates and information modulation, including the following steps:

[0128] a. Edge-level message generation:

[0129] First, adjacency information is constructed based on edge connections. The hidden states of adjacent nodes (including atoms or amino acid residues) are fused and mapped with the features of the edges (covalent or non-covalent interactions) to generate edge-level message vectors. Specifically, the graph of molecules, residues, and interface nodes is represented as follows: ,in Represents a set of nodes (atoms, amino acid residues, and interface nodes). This represents a set of edges (chemical bonds, non-covalent interactions, or interactions between nodes). For any node... In its first The hidden state in the next message passing iteration is represented as: Corresponding edge The edge features are represented as ,node The set of neighboring nodes is denoted as .

[0130] During the message passing phase, a neighborhood message is first constructed based on node and edge features, calculated using the following formula:

[0131] in:

[0132] · Indicates from neighboring nodes Passed to node The message vector;

[0133] · and The learnable parameter matrix;

[0134] · For bias terms;

[0135] · This represents a non-linear activation function.

[0136] b. Node message aggregation:

[0137] Subsequently, for the nodes The messages from all neighboring nodes are aggregated to obtain a node-level message representation, represented by the following formula:

[0138]

[0139] c. GRU gated fusion:

[0140] After obtaining the node aggregation message, the current node state and the aggregation message are merged and updated through a Gated Recurrent Unit (GRU) to obtain the node representation for the next iteration:

[0141] The GRU structure controls the information fusion ratio between historical node representations and new aggregated messages through update and reset gates, thereby achieving stable feature updates and long-range dependency information modeling.

[0142] S9, Information Readout Module: This module transforms the atomic-level embedding representations obtained during the message passing phase into molecular-level global representations. Taking the atomic embedding vector of each molecule as input, it calculates the correlation weights between different nodes based on a self-attention mechanism, and achieves adaptive enhanced representation of key structural units and potential reaction sites through weighted aggregation.

[0143] Suppose a molecule contains n atoms, its atom embedding representation obtained through a message passing network is as follows:

[0144] Suppose a certain graph contains Each node (a node in an atom, amino acid residue, or interface) is represented by its node embedding through a message passing network as follows:

[0145]

[0146] in For the first The feature vector of each node For the embedding dimension, the correlation weights between atoms are calculated using a multi-head self-attention mechanism, and the atom representation is updated. The calculation process can be uniformly represented as follows:

[0147]

[0148] in:

[0149] · This represents multi-head attention.

[0150] · This represents the transformation of a feedforward neural network;

[0151] · Layer Normalization;

[0152] · This indicates a global average pooling operation;

[0153] · This is the final molecular-level representation vector.

[0154] The multi-head attention calculation is as follows:

[0155]

[0156] in:

[0157] · , , For the first The query, key, and value matrix of each attention head;

[0158] · This is the learnable parameter matrix.

[0159] Through the above operations, the molecular-level embedding vector can be obtained:

[0160]

[0161] in This represents the node representation after updates via attention and feedforward networks. Since the weights between different atoms are adaptively learned through a self-attention mechanism during attention computation, atoms with important chemical environment information (such as potential reaction sites or key structural units) will receive a higher contribution in feature updates, thus achieving a stronger representation of key structural information.

[0162] S10. Feature Concatenation and Fully Connected Prediction: Embedding vectors between different molecular maps are concatenated using the `Concatenation` function (`tf.concat`), outputting a 192-dimensional fused vector. A feedforward neural network (FNN) is used to process the feature vector matrix of the complex, predicting the covalent reactivity and reaction mode between small molecules and receptors. The FNN consists of three fully connected layers, each with 1024, 512, and 8 hidden units respectively. Through multi-layer nonlinear mapping of the fused features, multi-classification results for covalent bond types are output. To avoid overfitting, each fully connected layer is connected by a Dropout layer.

[0163] S11. Output Results: The output covalent reaction types include: substitution reaction, Michael addition reaction, epoxide addition reaction, acetal reaction, disulfide bond formation, aldehyde-amine condensation, other addition reactions, and no reaction, totaling 8 types.

[0164] Example 2

[0165] The ligand characterization part of the model can be replaced by the MolGraphConvFeaturizer and DMPNNFeaturizer modules in the DeepChem package with little impact on prediction accuracy; the receptor characterization part can be characterized using the all-atomic method provided in DeepCosi and Deepcys, but the combined model has lower prediction accuracy for covalent reaction sites such as Lys and Ser.

[0166]

[0167] Example 3: The model predicted a novel acetylcholine receptor modulator: cinnamaldehyde.

[0168] (1) Input the SIMILES descriptor of cinnamaldehyde (C1=CC=C(C=C1)C=CC=O) into the ligand embedding characterization module to obtain the ligand molecule feature embedding characterization;

[0169] (2) Extract the mouse protein structure from the PDB database, perform coarse granulation on the protein structure, and use the active pocket embedding characterization module to process the protein structure to obtain pocket feature embedding characterization.

[0170] (3) Input SMILES and protein structure into the covalent warhead structure recognition and characterization module to obtain the covalent warhead structure feature characterization matrix.

[0171] (4) The predicted potential targets are as follows:

[0172]

[0173] MAPK, F2, and F11 are previously reported targets, with Neuronal acetylcholine receptor subunit alpha-7 being a novel target. SPR experimental results indicate that CHRNA7 is a covalent molecular target of cinnamaldehyde. Mutation of the cysteine ​​residue at the active site significantly reduced the affinity for cinnamaldehyde, further demonstrating the model's high predictive ability (see...). Figure 6 , Figure 7 ).

Claims

1. A general method for predicting covalent drug targets based on information transmission neural networks, characterized in that: It is a deep learning-based covalent drug-target response type prediction model, including a data preparation module, a multi-level feature engineering template, and a deep learning prediction module; The data preparation module involves collecting and preprocessing protein structures, screening for covalently reactive small molecule complexes, constructing a dataset, and then dividing the database. The multi-level feature engineering module uses a multi-channel graph neural network input strategy to construct three complementary feature matrices respectively. The three feature matrices are the ligand feature matrix and adjacency matrix, the receptor feature matrix and adjacency matrix, and the covalent warhead and its surrounding residues feature matrix and adjacency matrix. Before constructing the receptor feature matrix, the receptor is first characterized using coarse-grained receptor characterization. The deep learning prediction module includes an information encoding and representation learning module, an information reading module, a feature concatenation and fully connected prediction module, and an output prediction result module.

2. The general drug covalent target prediction method based on information transmission neural network according to claim 1, characterized in that: The data preparation module includes the following steps: S1. Protein structure collection and pretreatment: Based on the PDBID in CovBinderInPDB, protein crystal structures from human and mouse sources are collected from the PDB database, a list of protein PDBIDs is established, and annotation information is added; based on the PDBIDs in the list, protein structures are downloaded in batches using the BioPython package, and then subunit separation, solution removal, and redundancy removal are performed to obtain a standardized monomer protein structure database in PDB format. S2. Establish a covalent warhead identification module: The CovBinderInPDB database is retrieved, and the adduct_smiles and binder_smiles items of each entry in the database are compared. Based on the differences in the SMILES descriptor, the SMARTS descriptor is established to define covalent warheads for the determination of covalent warheads in compounds. The difference patterns are converted into queryable structural patterns. The module contains 118 covalent reactive groups and the seven reaction types involved, including substitution reactions, Michael addition reactions, epoxide addition reactions, acetal reactions, disulfide bond formation, and aldehyde-amine condensation. S3. Constructing the dataset: For each covalent complex screened in S1, the protein and small molecule parts of the complex were separated using the Biopython package. The covalent warhead identification module established in S2 was used to label the covalently reactive structures (N1 in total) in the small molecules. Five classes of nucleophilic residues were labeled according to the amino acid residue names of the protein, namely Cys / Arg / Ser / Thr / His (N2 in total). The above two variables were combined to generate candidate datasets (N1×N2 in total) by exhaustive search. The covalently reactive information and nucleophilic residues obtained from the analysis were added to the candidate datasets. All structures in CovBinderInPDB were processed using the above method, and a dataset containing 868,415 entries was finally obtained. The training set, internal test set and external test set were split according to the ratio of 90%:9%:1% to obtain a three-level dataset with ground-truth labels.

3. The general drug covalent target prediction method based on information transmission neural network according to claim 2, characterized in that: The annotation information in step S1 includes HetID, UniprotID, protein name, and protein subunit; the solution removal includes deleting water, glycerol, and DMSO solvent molecules; and the redundancy removal includes removing duplicate protein structures.

4. The general drug covalent target prediction method based on information transmission neural network according to claim 1, characterized in that: The construction of the ligand feature matrix and adjacency matrix is ​​based on the graph neural network (GNN) paradigm, which transforms the ligand molecule into a topological graph structure, and includes the following steps: a. Use the Chem.MolFromSmiles function in RDKit to convert the SMILES descriptor into the molecular structure and construct the planar structure of the compound; b. Extract five types of parameters: atom type, atom valence state, number of hydrogen atoms, hybridization type, and atom charge, and establish an atom feature vector matrix; c. Calculate two types of parameters based on the structure of the compound: chemical bond type and whether it is a conjugated structure, and establish a chemical bond feature vector matrix; d. Construct the adjacency matrix of atoms based on chemical bonds; e. Extract the atomic feature vector matrix, chemical bond feature vector matrix, and atomic adjacency matrix as inputs to the ligand characterization part of the model.

5. The general drug covalent target prediction method based on information transmission neural network according to claim 1, characterized in that: The method for constructing the receptor feature matrix and adjacency matrix is ​​to achieve a precise graph-structured representation of the microenvironment of covalent reaction sites through the topological path search (TPS) algorithm and multi-scale feature encoding. Includes the following steps: a. Using Biopython, the PDB file of the complex is converted into a stereostructure. Amino acids in the range of 7-10 Å are extracted with covalently reactive residues (R) as the center. An active pocket diagram is generated with 3.5 Å as the threshold for non-bonded interactions. b. Using the amino acid closest to the covalently reactive residue (R) as the topological starting point, the TPS algorithm is used to find the one-way loop structure in the active pocket diagram; then, using the amino acid on the one-way loop as the starting point, the 3-5 residues closest to it are retained; finally, the one-way loop structure and its neighboring residue set are merged to form a coarse-grained structural representation of the receptor binding pocket. c. Extract four types of information from the coarse-grained acceptor representation: amino acid residue type, number of hydrogen atoms, hydrophobicity, and residue charge, and establish an amino acid feature vector matrix. d. Extract three types of properties from the non-covalent interactions between amino acid residues: whether each amino acid residue pair contains an aromatic ring and the distance between amino acid residues, and generate a non-covalent interaction feature vector matrix. e. Construct an amino acid residue adjacency matrix based on the above non-covalent interactions; f. Extract the amino acid feature vector matrix, the non-covalent interaction feature vector matrix, and the amino acid residue adjacency matrix as inputs to the receptor characterization part of the model.

6. The general drug covalent target prediction method based on information transmission neural network according to claim 1, characterized in that: The method for constructing the feature matrix and adjacency matrix of the covalent warhead and its surrounding residues includes the following steps: a. Use the covalent warhead identification module established by S2 to identify the covalent warhead of the compound, extract the atoms of the covalent warhead, and realize the location of the warhead atoms; b. Extract amino acid residues within 3-5 Å of the covalently reactive amino acid residues (R), and construct a microenvironment map of the covalent warhead site based on the non-covalent interactions between residues; c. Connect the covalent reaction site to the surrounding amino acid residues to construct a virtual interface; d. Extract the covalent warhead portion of the virtual interface and extract five types of parameters: atom type, atom valence state, number of hydrogen atoms, hybridization type, and atom charge. Establish the feature vector matrix of the covalent warhead portion. e. Calculate the chemical bond type and whether it is a conjugated structure based on the chemical structure of the covalent warhead, and establish a chemical bond feature vector matrix; f. Construct the adjacency matrix of the covalent warhead based on chemical bonds; g. Extract the amino acid residue portion of the virtual interface and calculate four types of information: amino acid residue type, number of hydrogen atoms, hydrophobicity, and residue charge. Establish a feature vector matrix of the microenvironment of the covalent warhead. h. Extract three types of properties: the type of non-covalent interaction between amino acid residues, whether all amino acid residue pairs contain aromatic rings, and the distance between amino acid residues, and generate a feature vector matrix of non-covalent interaction in the microenvironment of the covalent warhead. i. Construct a partial amino acid residue adjacency matrix for the microenvironment based on the above non-covalent interactions; j. Merge the feature vector matrix of the covalent warhead and the feature vector matrix of the microenvironment to construct the virtual interface feature vector matrix; k. Combine the chemical bond eigenvector matrix of the covalent warhead part and the non-covalent interaction eigenvector matrix of the microenvironment part to construct the virtual interface interaction eigenvector matrix; 1. Merge the adjacency matrix of the covalent warhead and the adjacency matrix of the amino acid residues to construct a virtual interface adjacency matrix; m. The merged virtual interface feature vector matrix, virtual interface interaction feature vector matrix, and virtual interface adjacency matrix are used as inputs to the virtual interface representation part of the model.

7. The general drug covalent target prediction method based on information transmission neural network according to claim 1, characterized in that: The information encoding and representation learning module constructs three complementary feature matrices with topological information, propagates this information through multiple rounds of neighborhood information, and achieves local structure embedding updates and information modulation, including the following steps: a. Edge-level message generation: First, adjacency information is constructed based on edge connection relationships. The hidden states of adjacent nodes (including atoms or residues) are fused and mapped with the features of the edges (covalent or non-covalent interactions) to generate edge-level message vectors. b. Node message aggregation: The messages from the neighborhood are weighted and summed to form a candidate update representation for the node; The characterization formula is as follows: Let the molecular, residue, and interface node diagram be represented as follows: ,in Represents a set of nodes (atoms, amino acid residues, and denominator nodes). This represents a set of edges (chemical bonds, non-covalent interactions, or interactions between nodes). For any node... In its first The hidden state in the next message passing iteration is represented as: Corresponding edge The edge features are represented as ,node The set of neighboring nodes is denoted as ; During the message passing phase, a neighborhood message is first constructed based on node and edge features, calculated using the following formula: ;in: · Indicates from neighboring nodes Passed to node The message vector; · and The learnable parameter matrix; · For bias terms; · Represents a nonlinear activation function; Subsequently, for the nodes Aggregate all neighbor node messages to obtain a node-level message representation: ； c. GRU gated fusion: During the node state update phase, a gated recurrent unit (GRU) is introduced to perform gated fusion of the current node's hidden state and aggregated messages. By controlling the ratio of historical information to new input information through update gates and reset gates, stable gradient propagation and long-term dependency modeling are achieved; each molecular graph outputs a 64-dimensional embedding vector. Finally, the current node state and the aggregated message are merged and updated using a gated recurrent unit (GRU) to obtain the node representation for the next iteration: ； The information readout module is based on a self-attention mechanism to achieve adaptive aggregation from atomic-level embedding to molecular-level global representation, including the following steps: This is used to transform the atomic-level embedding representation obtained in the message passing stage into a molecular-level global representation. Taking the atomic embedding vector of each molecule as input, the correlation weight between different nodes is calculated based on the self-attention mechanism. Through weighted aggregation, the adaptive enhanced expression of key structural units and potential reaction sites is achieved. The specific algorithm and formula are as follows: Suppose a certain graph contains Each node (a node in an atom, amino acid residue, or interface) is represented by its node embedding through a message passing network as follows: ； in For the first The feature vector of each node For the embedding dimension, the correlation weights between atoms are calculated using a multi-head self-attention mechanism, and the atom representation is updated. The calculation process can be uniformly represented as follows: ； in: · This represents multi-head attention. · This represents the transformation of a feedforward neural network; · Layer Normalization; · This indicates a global average pooling operation; · This is the final molecular-level representation vector; The multi-head attention calculation is as follows: ； ； in: · , , For the first The query, key, and value matrix of each attention head; · The learnable parameter matrix; Through the above operations, the molecular-level embedding vector can be obtained: ； in This is the node representation after being updated by the attention and feedforward network; The feature splicing and fully connected prediction module is an end-to-end prediction module for the fusion of multi-scale molecular characterization and covalent reaction types, specifically including the following steps: a. Feature concatenation and fully connected prediction: Embedding vectors between different molecular maps are concatenated using the `Concatenation` function (`tf.concat`), outputting a 192-dimensional fused vector. b. Feedforward Neural Network (FNN) Architecture: The feedforward neural network (FNN) is used to process the feature vector matrix of the complex to predict the covalent reactivity and reaction mode of small molecules and receptors. The FNN consists of three fully connected layers, each with 1024, 512, and 8 hidden units respectively. By performing multi-layer nonlinear mapping on the fused features, the FNN outputs multi-classification results of covalent bond types. Each fully connected layer is connected by a Dropout layer. The output prediction results module outputs eight types of covalent reactions: substitution reaction, Michael addition reaction, epoxide addition reaction, acetal reaction, disulfide bond formation, aldehyde-amine condensation, other addition reactions, and no reaction.

Images