Deep learning-based fragment target interaction recognition and general compound library construction method

Abstract

The invention discloses a fragment target interaction recognition and general compound library construction method based on deep learning. The method comprises the following steps: extracting amino acid residues and complete ligand molecules around active binding sites in a protein crystal structure; constructing a compound skeleton fragment data set from a complete ligand molecule; performing interaction fingerprint coding based on the compound skeleton fragment and the active site amino acid residue, and constructing an interaction database; screening an active pocket of a target protein, cutting a ligand, and representing interaction; registering the interaction between the target protein and the molecular fragment with an interaction database; performing matching degree scoring on the registration result; screening potential target fragments which are highly consistent with the target protein and different molecular fragments in interaction scoring from the interaction database, and verifying the potential target fragments; and performing molecule generation and property evaluation based on the gated map neural network, thereby obtaining a final screening result. The present invention assists the discovery of lead compounds.

Description

Technical Field

[0001] The present invention belongs to the field of computer-aided drug design, and specifically relates to a method for fragment-target interaction identification and universal compound library construction based on deep learning. Background Art

[0002] In modern drug discovery, virtual screening uses advanced computer simulation technology and cheminformatics methods to rapidly predict and screen compounds with potential bioactivity from millions of compound libraries. The construction and optimization of virtual compound libraries are the foundation for effective virtual screening.

[0003] Fragment-based drug design approaches play a key role in this process. Starting with relatively simple but highly diverse small molecular fragments that cover a broad chemical space, high-throughput screening of these fragments allows the identification of those with excellent binding ability to specific biological targets. These fragments are then linked or chemically modified to design drug candidates with enhanced bioactivity and stronger affinity. Fragment-based drug design strategies offer a wide range of advantages over traditional drug design methods that rely on high-molecular-weight compounds. First, they enable researchers to explore a wider chemical space, increasing the chances of discovering innovative and highly effective drug molecules. Second, by precisely assembling the selected fragments, molecules with higher affinity for the target can be obtained, which is particularly effective for complex targets that are difficult to attack using traditional methods. Furthermore, although individual fragments may have low affinity for the target, through careful design and optimization, these fragments can be transformed into highly active potential drug molecules.

[0004] However, the current fragment-based drug design approach is more about starting from the fragments, considering the properties of the fragments themselves, and ignoring the interactions between the fragments and the target active sites, which often results in poor properties of the designed molecules. Summary of the Invention

[0005] To address the above-mentioned problems, one of the objectives of the present invention is to construct a library of backbone fragments based on an existing protein-ligand library using a molecular cleavage algorithm based on structure and chemical reactions. Furthermore, based on a triplet fingerprint encoding algorithm of interacting pseudoatoms, molecular fingerprints are used to characterize the interactions between the backbone fragments of compounds and the amino acid residues of proteins, thereby constructing a pharmacophore screening database covering a large number of molecular fragments. Another objective of the present invention is to propose a general, automatic, and unbiased method for constructing a compound library targeting a specific target, starting from the target protein-ligand complex, using theories related to protein-ligand interaction characterization. This method achieves the efficient generation of a large number of compound molecules focused on a specific target without the need for biological experiments and high-throughput screening.

[0006] To achieve the above object, the technical solution of the present invention is as follows:

[0007] A method for identifying fragment-target interactions and constructing a universal compound library based on deep learning, comprising the following steps:

[0008] Step 1: Extract the amino acid residues around the active binding site and the complete ligand molecule from the protein crystal structure.

[0009] Furthermore, the step 1 is specifically as follows: based on the protein-ligand complex crystal structure database, after protein-ligand complex structure screening, binding site identification, and structure correction, the protein-ligand complex crystal structure data containing at least one ligand are screened from the sc-PDB database. The crystal structure of the candidate protein-ligand complex must meet the following conditions: each complex structure contains at least one ligand molecule, the ligand molecule is not covalently bound to the surrounding protein amino acid residues, the surface area of the ligand molecule exposed to the solvent is less than 50%, and the X-ray / NMR analysis resolution is less than Only the amino acid residues near the active binding site in the screened protein ligand crystal structure and the complete ligand molecule are retained.

[0010] Step 2: Construct a compound skeleton fragment dataset based on the complete ligand molecule.

[0011] Furthermore, step 2 specifically involves fragmenting the intact ligand molecule in the protein-ligand complex structure using a molecular fragmentation algorithm. The algorithm uses the SMILES string of the ligand molecule as input, fragmenting the molecule while preserving the three-dimensional coordinates and atomic information of the fragments relative to the binding pocket. The algorithm then identifies ring structures, classifies atoms, processes substituents, and adds anchor atoms to obtain core backbone fragments of the ligand molecule. The resulting backbone fragments are then further screened to identify and retain fragments that exhibit at least one non-covalent interaction (hydrogen bond, van der Waals force, hydrophobic interaction, or electrostatic interaction) with the protein target. The broken bonds in the fragments (represented by the fictitious atom "Z") are then converted to hydrogen atoms, resulting in a molecular fragment containing a large number of amino acid residues from the active site of the protein target and exhibiting favorable interactions with these amino acid residues.

[0012] Step 3: Fingerprint encoding of interactions between the backbone fragments and the active site amino acid residues to construct an interaction database. Fingerprint encoding of the interactions between the backbone fragments of the compound separated from the intact ligand molecule in step 102 and the protein active site amino acid residues screened in step 101 is performed to construct a fragment-protein interaction database.

[0013] Furthermore, step 3 is specifically as follows: based on the interaction pseudo-atom triple fingerprint encoding algorithm, the atomic information and spatial coordinates of these interactions are captured and described, and the protein and ligand atoms are characterized according to the type of interaction they participate in (hydrophobic interaction, aromatic interaction, hydrogen bond donor, hydrogen bond acceptor, positive ionization, negative ionization, metal coordination), defining the interaction information between each molecular fragment and its corresponding protein active site and the surrounding environment. Six value ranges are further defined based on the interatomic interaction distance. Based on the interaction pseudo-atom representation, the interaction is fingerprinted using triples to form a 7×7×7×6×6×6-dimensional feature vector. After pruning, a 210-dimensional molecular fingerprint descriptor is obtained. This descriptor represents the interaction information between different molecular fragments and amino acid residues. This process is repeated to construct a fragment pharmacophore database.

[0014] Step 4: Screen the active pocket of the target protein, cut the ligand molecule, and characterize the interaction between the active site of the target protein and the different ligand molecule fragments obtained by cutting.

[0015] Furthermore, step 4 is specifically as follows: screening the crystal complex structure of the target to be studied from the PDB protein crystal database, processing the crystal structure, including removing multiple amino acid conformations, repairing missing residues, removing inactive ligands and metal ions, and removing crystal water molecules, and then adding hydrogen atoms to the protein structure and optimizing it based on the OPLS3 force field.

[0016] The ligand molecules in the target crystal structure are cut, and after ring structure recognition, atomic classification, linker retention, substituent processing, and anchor atom addition, the ligand molecules in the target binding pocket are split into different molecular fragments.

[0017] The interaction fingerprint encoding algorithm in step 3 is used to characterize the interactions between these molecular fragments and the active sites of the target protein.

[0018] Step 5: Interaction matching based on point cloud registration algorithm.

[0019] Furthermore, step 5 is specifically as follows: based on the point cloud registration algorithm in computer vision, the interaction fingerprints between different molecular fragments and active sites in the target protein constructed in step 4 are point cloud registered with the interaction fingerprint database in step 3.

[0020] The goal of point cloud registration is to register two point clouds P q and P k Alignment, where and are three-dimensional point sets, The registration process requires estimating a rigid transformation T = (R, t), where R∈SO(3) is the rotation matrix, is the translation vector so that the transformed P q With P k Align in space.

[0021] The first step of point cloud registration is to obtain the point cloud P q and P k The normal vector of each point is extracted and the geometric features are calculated based on the normal vector. The normal vector provides the local surface direction of each point in the point cloud and is the basis for subsequent feature description.

[0022] The Fast Point Feature Histogram (FPFH) is used to describe the geometric characteristics of each point in a point cloud. FPFH is an improvement on the Simplified Point Feature Histogram (SPFH), enhancing its descriptive power by considering the weighted relationships between neighboring points.

[0023] First, the target protein interaction features and the normal vectors and single-point feature histograms of different interactions in the interaction database are binned:

[0024]

[0025] Where: α ij =n i·n j : cosine of the angle between the normal vector and the vector; Projection of normal vector and connecting line; Angle deviation; Binning operation (number of bins n b =33).

[0026] Based on the following formula, the random sample consensus (RANSAC) method of fast point feature histogram (FPFH) is further used for preliminary coarse registration:

[0027] where ||P i -P q || indicates P i and P q The Euclidean distance of .

[0028] Furthermore, the color fast point feature histogram (c-FPFH) is introduced based on the following formula, which incorporates the geometric information of the feature descriptor and increases the dimension of the interaction type to improve the richness and accuracy of the feature description.

[0029] Among them, c-FPFH (P q ) represents point P q Color fast point feature histogram, combining geometric and color information. c-SPFH(P q ) is point P q The color single point feature histogram contains local geometric and color statistical features.

[0030] Step 6: Point cloud similarity scoring.

[0031] Furthermore, step 6 is specifically as follows: after completing the point cloud registration in step 5, based on the fitness (Fitness) and root mean square error (RMSE) scoring indicators, and the pharmacophore scoring function, the feature descriptors of different sub-pockets of the target protein are scored with the registered fragment library descriptors based on the geometric features and pharmacophore properties.

[0032]

[0033] ps∈P sourcept ∈Φ sreyer ; where N represents the total number of points in the point cloud, Ps i represents the i-th point in the source point cloud, Pt i Indicates the target point cloud with Ps i The corresponding i-th point, Φ sreyer Represents the transformation relationship of point cloud registration.

[0034] Step 7: Enumerate connectable fragments and atom pairs.

[0035] Furthermore, the step 7 is specifically as follows: after the matching in step 5 and the scoring in step 6, potential target fragments with highly consistent interaction scores with the target protein and different molecular fragments are screened from the fragment-protein interaction database, and then whether these fragments can be connected to generate complete compound molecules is verified.

[0036] The following steps are included: (a) performing a spatial combination traversal of the m substitution sites of the first fragment f1 and the n substitution sites of the second fragment f2 based on an enumeration algorithm to generate an m×n dimensional connection matrix; (b) for each atom pair (A i ,A j ), respectively establish the three-dimensional topological criterion of hydrogen bond connection: with heavy atom A i Bonded hydrogen atom H i The bond axis of the atom pair is used as a reference, and a conical space region with a half-angle of π / 4 is formed along the bond axis; (c) if and only if the two heavy atoms (Ai, Aj) in the atom pair are simultaneously located in the conical space region corresponding to the hydrogen bond axis of the other party, the atom pair is judged to form a valid hydrogen bond connection; (d) after the traversal is completed, all valid atom pairs that meet the three-dimensional topological constraints are screened out through the connection matrix to generate a combination set of connectable sites.

[0037] Step 8: Molecule generation and property evaluation based on gated graph neural network.

[0038] Furthermore, step 8 specifically involves using a breadth-first algorithm based on a gated graph neural network to iteratively select atoms from the initially defined "atom pool" and form bonds one by one until a new fragment is connected. The molecule generation process is divided into three parts: initialization, molecule construction, and generation termination.

[0039] The skeleton fragment is represented as a graph, where atoms are represented by nodes and bonds are represented by edges, where each node has a hidden state Z v and label L v The molecular graph is processed by a gated graph neural network to update the hidden state of the node to include its local environment information. The length threshold of the linker is set to a minimum of 3 heavy atoms to a maximum of 10 heavy atoms. The generated molecules are output as SMILES strings and further processed to eliminate the redundancy of each connectable atom pair. After this process, a large number of complete molecular SMILES characters are obtained.

[0040] The generated molecules are evaluated for synthesizability (SAscore), and molecules with higher synthesizability scores are screened based on Lipinski's five rules and quantitative estimation of drug similarity (QED). Several small molecule compounds with high rankings are selected as the final screening results.

[0041] It is to be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 A flowchart of a method for fragment-target interaction identification and universal compound library construction based on deep learning provided in an embodiment of the present invention.

[0043] Figure 2 A schematic diagram of a molecular cutting algorithm provided by an embodiment of the present invention.

[0044] Figure 3 A schematic diagram of a triple fingerprint encoding based on intermolecular interaction pseudo-atoms provided by an embodiment of the present invention.

[0045] Figure 4 A schematic diagram of enumerating connectable fragments and atom pairs provided in an embodiment of the present invention.

[0046] Figure 5 A schematic diagram of a molecule generation method based on a graph neural network provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0047] In order to more clearly understand the purpose, technical solutions and advantages of the present application, the technical solutions of the present invention are further explained below in conjunction with the accompanying drawings and embodiments. The described embodiments should not be regarded as limiting the present application. All other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of this application.

[0048] In the process of lead compound discovery, the affinity between protein and ligand mainly depends on the interaction between the key backbone fragments of the compound and the amino acid residues of the protein. By characterizing this interaction with feature descriptors, the binding situation between a molecular fragment and the surrounding protein active sites can be defined. At the same time, these feature descriptors are not unique. When two fragment-active sites have similar feature descriptors, exchanging their ligands will also produce similar interactions. Based on this design idea, the present invention starts from the interaction between fragments and targets, and develops a set of feature descriptors that characterize the interaction information of molecular fragments and surrounding active sites. Feature descriptors with similar interactions are then searched for in the constructed database, with the aim of identifying and obtaining potential fragments that can produce similar interactions. After obtaining a potential fragment library for a specific target active site, a deep generative model is used to generate linkers between the fragments to obtain specific molecules with high selectivity, thereby constructing a potential compound library for a specific target.

[0049] like Figure 1 As shown, the present invention provides a method for identifying fragment-target interactions and constructing a universal compound library based on deep learning, which includes the steps of:

[0050] In step 101, amino acid residues around the active binding site in the protein crystal structure are extracted.

[0051] Specifically, step 101 can be divided into screening out a protein-ligand complex crystal structure that meets the requirements from a protein crystal database, and extracting amino acid residues near the protein active site and bound ligand molecules in the protein-ligand complex crystal structure.

[0052] The protein-ligand complex crystal structure that meets the requirements is constructed on the basis of the protein-ligand complex crystal structure database, through a series of processes including protein-ligand complex structure screening, binding site identification, amino acid residue structure correction, ligand chemical structure correction, spatial coordinate annotation of ligand-protein interaction information, and precise position annotation of ligand in active pocket, to construct the original data set containing protein target and its corresponding ligand.

[0053] The protein-ligand complex crystal structures that meet the requirements are classified according to PDB ID, and each PDBID is ensured to contain the following files: protein.mol2, ligand.mol2&ligand.sdf, site.mol2, PDB_ID.mol2, ints_M.mol2.

[0054] The crystal structure of the candidate protein-ligand complex must meet the following conditions: each complex structure contains at least one ligand molecule, the ligand molecule is not covalently bound to the surrounding protein amino acid residues, the surface area of the ligand molecule exposed to the solvent is less than 50%, and the X-ray / NMR resolution is less than

[0055] Extraction of amino acid residues near the active site of the protein and the bound ligand molecule specifically refers to retaining only the amino acid residues near the active binding site and the complete ligand molecule in the screened protein ligand crystal structure.

[0056] In step 102, a compound skeleton fragment dataset is constructed starting from the complete ligand molecule.

[0057] Specifically, the construction of the backbone fragment dataset in step 102 involves segmenting the complete ligand molecule into molecular fragments using a molecular segmentation algorithm based on the amino acid residues near the active binding site in the different protein ligand crystal structures screened and obtained in step 101, and the complete ligand molecule. The molecular segmentation process is performed while retaining the core backbone fragments. The resulting backbone fragments are further screened to identify and retain fragments that have at least one non-covalent interaction (hydrogen bond, van der Waals force, hydrophobic interaction, electrostatic interaction) with the protein target. At the same time, the broken bonds of the fragments (represented by the fictitious atom "Z") are converted to hydrogen atoms, resulting in a molecular fragment that contains a large number of amino acid residues from the active site of the protein target and has good interactions with these amino acid residues.

[0058] Specifically, algorithms for molecular cutting, such as Figure 2 As shown, the SMILES string is used as input and the SMILES is parsed into a molecular graph G = (V, E), where V is a set of atoms (vertices) and E is a set of chemical bonds (edges). Breakable bonds are marked, and loop detection in the graph is used to identify ring structures in molecules, avoid breaking bonds within the ring, and generate a set of SMILES representations of fragments.

[0059] For the broken bond e=(v1,v2), in G i and G j Add anchor atoms to and

[0060] The segmentation process can be expressed as a recursive function:

[0061] where G\e represents the operation of removing edge e from G and adding an anchor atom.

[0062] In step 103, interaction fingerprint encoding is performed based on the backbone fragments and the active site amino acid residues to construct a fragment-protein interaction database.

[0063] Specifically, the interaction between the molecular fragments separated from the complete ligand molecule in step 102 and the amino acid residues of the protein active site screened in step 101 is characterized by using a triplet fingerprint encoding algorithm.

[0064] The specific process of triplet fingerprint encoding is as follows Figure 3 As shown in the figure, the interaction type and strength between molecular fragment atoms and amino acid residues are used as input. By capturing and describing the atomic information and spatial coordinates of these interactions, protein and ligand atoms are classified according to the types of interactions they participate in, such as hydrophobicity, aromaticity, hydrogen bond donors, hydrogen bond acceptors, ionization states, and metal complexes.

[0065] For each pair of interacting atoms (protein atoms P i and ligand atom L j ), first through its three-dimensional coordinates (x i ,y i ,z i ) and (x j ,y j ,z j ) Calculate the Euclidean distance And according to the distance threshold d threshold Determine whether there is a significant interaction. ij ≤d threshold , the interaction types are further classified according to the following consensus rules:

[0066] Hydrophobic interaction: When P i and L j When all atoms are non-polar (such as carbon atoms, hydrophilicity parameter logP < 0), it is determined to be a hydrophobic interaction;

[0067] Aromatic interactions: If P i and L j Located on the aromatic ring, and the ring spacing d ring exist to If the angle θ (calculated by the normal vector) satisfies |cosθ|>0.8, it is classified as π-π stacking or CH-π interaction;

[0068] Hydrogen bond: If P i or L j is a hydrogen bond donor (such as -OH, -NH), the other is an acceptor (such as C=O, N:), and the distance If the angle ∠DHA ≥ 120°, it is marked as a hydrogen bond interaction;

[0069] Ionic interaction: When P i and L j With opposite charges (such as -COO- and -NH3 + ), and the distance Calculating Coulomb interaction energy (where q i ,q j is the charge, ∈0 is the vacuum dielectric constant);

[0070] Metal complexation: If L j is a ligand atom (such as O, N, S) that interacts with the metal ion M in the protein k (such as Zn 2+ Mg 2+ ) coordination, coordination distance It is classified as metal complexation.

[0071] For each type of interaction identified, a pseudo-atom representation is generated, whose spatial coordinates are defined as the center of interaction geometry Alternatively, to emphasize the unilateral contribution of the interaction, the pseudo-atom position can be set to P i or L j 's coordinates.

[0072] The strength of the interaction is divided into six levels, which are encoded together with the interaction pseudo-atom representation into a triple and mapped to a predefined 210-bit array, where each bit represents a specific type and strength of interaction combination. Output binary fingerprint F = {0, 1} n .

[0073] By fingerprinting the interactions using triplets, a 7×7×7×6×6×6-dimensional feature vector is formed. After pruning, a 210-dimensional molecular fingerprint descriptor is obtained. This descriptor defines the type and strength of interaction between amino acid residues in a protein binding pocket and different molecular fragments. This process is repeated to characterize the interactions of different protein-ligand complexes screened in step 101, and a fragment-protein interaction database is constructed.

[0074] In step 104, the target site is identified, the active pocket is selected, the ligand is cleaved, and the interaction is characterized. This specifically involves selecting the optimal crystal conformation of the target protein, cleaving the ligand molecules bound to the target protein, and characterizing the interactions between the target protein active site and the different ligand fragments obtained by cleavage.

[0075] The optimal crystal conformation of the target protein to be studied is selected from the protein crystal database. First, steric conflict conformations are eliminated based on the van der Waals collision test. The calculation formula is: where r ij is the interatomic distance, ∈ ij and σ ij is the Lennard-Jones parameter of the atom pair, retaining E clash Minimal conformation. For missing amino acid residues, an insertion method based on main chain geometry constraints was used. Inactive ligands and metal ions were removed, as well as crystallographic water molecules. Subsequently, hydrogen atoms were added to the protein structure and optimized based on the OPLS3 force field. The energy function of the system was defined as:

[0076] E total =E bond +E angle +E torsin +E vdw +E elec

[0077] E bond =∑k b (r-r0) 2 : bond length stretching energy, k b is the force constant, r0 is the equilibrium bond length;

[0078] E angle =∑k θ (θ-θ0) 2 : bond angle bending energy, θ0 is the equilibrium angle;

[0079] E torsion =∑k φ [1+cos(nφ-φ0)]: torsional energy, n is periodicity;

[0080] van der Waals interactions;

[0081] Electrostatic interaction, q i is a partial charge.

[0082] The pre-selected protein structures are aligned to a unified coordinate system, the ligands and receptors are separated, and the optimal conformation is selected through cross-docking and superposition of protein crystal structures.

[0083] The ligand molecule bound to the target protein is cut. Specifically, with the help of the molecular cutting algorithm in step 102, the ligand molecule bound to the active site of the target protein is cut and split into different molecular fragments.

[0084] Characterize the interaction between the active site of the target protein and the different ligand molecular fragments obtained by cleavage. Specifically, use the triple fingerprint encoding algorithm of step 103 to encode the interaction between the active site of the target protein and the different molecular fragments using a 210-dimensional molecular fingerprint descriptor.

[0085] In step 105 , interaction matching is performed based on the point cloud registration algorithm. Specifically, the objects of interaction matching are the interaction between the target protein and the molecular fragment in step 104 and the interaction database constructed in step 103 .

[0086] The specific implementation process of the registration algorithm includes normal vector estimation and geometric feature extraction, calculation of FPFH features, coarse registration based on fast point feature histogram (FPFH), and fine registration using the iterative closest point (ICP) algorithm and the introduction of color fast point feature histogram (c-FPFH).

[0087] The generation of point cloud refers to converting the 210-dimensional triple interaction fingerprint encoding into a point cloud feature vector. Specifically, the interaction generation source point cloud S = {s i |s i ∈R 3} and target point cloud T = {t j |t j ∈R 3}, where each point s i =(x i ,y i ,z i ) represents the spatial coordinates, removes noise points (distance mean offset > 3σ) and downsamples.

[0088] Normal vector estimation and geometric feature extraction include normal vector estimation and geometric feature extraction.

[0089] Calculate the FPFH feature. Specifically, for each point s i Calculate the normal vector n i and single point feature histogram (SPFH). The normal vector is estimated by principal component analysis (PCA) based on the minimum eigenvector of the covariance matrix of the k-neighborhood (k=30). SPFH captures local geometric features and is defined as:

[0090]

[0091] in:

[0092] α ij =n i ·n j : cosine of the angle between the normal vector and the vector;

[0093] Projection of normal vector and connecting line;

[0094] Angle deviation;

[0095] Binning operation (number of bins n b =33).

[0096] Based on the following formula, the random sample consensus (RANSAC) method of fast point feature histogram (FPFH) is further used for preliminary coarse registration:

[0097] where ||P i -P q || indicates P i and P q The Euclidean distance of .

[0098] In point cloud registration, the quality of feature descriptors directly impacts both matching accuracy and registration robustness. Traditional fast point feature histograms (FPFHs) rely solely on geometric information (such as normal vectors and the relative positional relationships between points), which can lack sufficient discriminability in some scenarios, particularly when point clouds have similar geometric structures but different color distributions. To address this, we introduce color fast point feature histograms (c-FPFHs). By combining geometric and color information and increasing the dimension of interaction types, we improve the richness and accuracy of feature descriptions.

[0099] The calculation formula for c-FPFH is as follows:

[0100] Among them, c-SPFH (p q,i ) represents point p q,i Color feature histogram. k represents the neighborhood point set N(p q,i ) size. ‖p j -p q,i || represents point p j and p q,i The Euclidean distance between .

[0101] In step 106, the point cloud similarity is scored. Specifically, the interaction between the target protein and different ligand molecular fragments in step 105 and the fragment-protein interaction database after registration in step 103 are scored for matching.

[0102] By comprehensively considering the scoring method of geometric alignment and pharmacophore property matching, a detailed and comprehensive evaluation of the point cloud registration quality is achieved, so that fragments that can potentially interact well with the sub-pockets of the target protein are found in the fragment database. After point cloud registration and similarity scoring, the results are output in the form of key-value pairs of PDB ID-scoring results.

[0103] In step 107, linkable fragments and atom pairs are enumerated.

[0104] Potential target fragments with highly consistent interaction scores with target proteins and different molecular fragments were screened from the fragment-protein interaction database, and then it was verified whether these fragments could be connected to generate complete compound molecules.

[0105] Judgment basis Figure 4 :Define the substitution site set of f1 as S1={s 1,1 ,s 1,2 ,…,s 1,m}, the set of substitution sites of f2 is S2 = {s 2,1 ,s 2,2 ,…,s 2,n Each substitution site s i,j Contains a heavy atom A i,j (such as C, N, O) and its connected hydrogen atoms H i,j The algorithm generates all possible connection pairs (s 1,k ,s 2,l ), forming an m×n connection matrix C, where C k,l Indicates a potential connection.

[0106] To determine the connection pair (A 1,k ,A 2,l ) feasibility, the cone projection model is introduced. Let the bond axis vector v i,j =A i,j -H i,j , from H i,j Point to A i,j . With H i,j is the vertex, v i,j For the axis, define the half angle The cone space is:

[0107]

[0108] If the following conditions are met:

[0109] ·A 2,l ∈Cone(H 1,k ,v 1,k ,α),

[0110] ·A 1,k ∈Cone(H 2,l ,v 2,l ,α),

[0111] Then (A 1,k ,A 2,l ) are connectable atom pairs.

[0112] Step 108 involves molecule generation and property evaluation based on a gated graph neural network. Specifically, through the aforementioned processes of point cloud registration, similarity scoring, and connectivity screening, a series of fragment backbones targeting different subpockets of the target protein are selected. Subsequently, based on the graph neural network, new molecules are iteratively constructed from the initial sites of the given fragments through bond-by-bond formation, directly integrating 3D structural information into the molecular design process.

[0113] The algorithm implementation process of molecule generation is as follows Figure 5 Specifically, the algorithm accepts the SMILES representation of two molecular fragments, denoted as S1 and S2, and defines the initial atomic pool Each atom a i With type tag t i and hidden state h i , convert the molecular fragment into a graph G = (V, E), where node v i ∈V represents an atom, and edge e ij ∈E represents a chemical bond. Each node v i With hidden state h i and the atom type label t i , update the hidden state of the node through the gated graph neural network (Gated GNN),

[0114]

[0115] in, Represents node v i The neighbor set of f edge is the edge feature mapping function, l represents the number of GNN layers. Initialize the expansion node ε and sample the atom types, then gradually select edges and label the bond types, and output the generated molecule.

[0116] Based on the following formula, the generated molecules were evaluated for synthesizability (SAscore). Molecules with higher synthesizability scores were screened based on Lipinski's Rule of Five and Quantitative Estimate of Drug Similarity (QED). Several small molecule compounds with high rankings were selected as the final screening results.

[0117] The SAscore (Synthetic Accessibility Score) is used to assess the ease of synthesizing a molecule. The score typically ranges from 1 to 10, with lower scores indicating easier synthesis.

[0118] SAscore=fragmentScore-complexityPenalty

[0119] The fragment score is based on the frequency of common chemical fragments in a molecule. Common fragments (such as benzene rings and alkyl chains) are typically statistically analyzed from large chemical databases such as PubChem. The more common the fragment, the higher the score.

[0120] Complexity Penalty represents complexity penalty, which measures the complexity of the molecular structure

[0121] QED (Quantitative Estimate of Drug-likeness) is a method for quantitatively evaluating the similarity of molecular drugs.

[0122]

[0123] Where: d i Indicates the value of descriptor i (such as molecular weight, logP, etc.). i Denotes the weight of descriptor i. i,ref : represents the ideal value of descriptor i in the reference drug. n: represents the number of descriptors.

[0124] Lipinski's "rule of five" is a set of empirical rules used in medicinal chemistry to predict whether a compound has good oral bioavailability. Specifically, they are:

[0125] Molecular weight: less than 500 Daltons

[0126] Hydrophobicity (LogP): less than 5

[0127] Number of hydrogen bond donors: no more than 5

[0128] Number of hydrogen bond acceptors: no more than 10.

[0129] The above is a detailed introduction to a method for identifying fragment-target interactions and constructing a universal compound library based on deep learning provided by this application. The terms used in the examples of this application are only for the purpose of describing specific embodiments and are not intended to limit this application. The singular forms "a", "the" and "the" used in the examples of this application and the appended claims are also intended to include plural forms unless the context clearly indicates otherwise.

Claims

1. A method for fragment-target interaction identification and universal compound library construction based on deep learning, characterized in that: The following steps are involved: Step 1: extract the amino acid residues around the active binding site and the complete ligand molecule in the protein crystal structure; Step 2: construct a compound skeleton fragment dataset based on the complete ligand molecule; Step 3: Fingerprint encoding of interactions between compound backbone fragments and active site amino acid residues to construct a fragment-protein interaction database; Step 4: Screen the active pocket of the target protein, cut the ligand, and characterize the interaction between the target protein and the molecular fragment; Step 5: align the target protein-molecular fragment interaction obtained in step 4 with the fragment-protein interaction database in step 3; Step 6: Score the matching degree of the registration result of step 5; Step 7: Screen potential target fragments from the fragment-protein interaction database that have highly consistent interaction scores with the target protein-molecule fragments, and verify whether these fragments can be connected to generate complete compound molecules; Step 8: Molecular generation and property evaluation based on gated graph neural network to obtain the final screening results.

2. The method for fragment-target interaction identification and universal compound library construction based on deep learning according to claim 1, characterized in that: Step 1: Extracting protein crystal structures involves screening protein-ligand complex crystal structures from the protein database. The following conditions must be met: each complex structure contains at least one ligand molecule, the ligand molecule is not covalently bound to the surrounding protein amino acid residues, the surface area of the ligand molecule exposed to the solvent is less than 50%, and the X-ray / NMR resolution is less than 3Å. Amino acid residues outside the binding pocket in the protein are then removed.

3. The method for fragment-target interaction identification and universal compound library construction based on deep learning according to claim 1, characterized in that: The construction of the compound skeleton fragment data set in step 2 refers to using a molecular cutting algorithm to segment the complete ligand molecule obtained in step 1 to obtain molecular fragments.

4. The method for fragment-target interaction identification and universal compound library construction based on deep learning according to claim 1, characterized in that: The step 3 is to perform fingerprint coding on the interaction between the compound skeleton fragments separated from the complete ligand molecule in step 102 and the amino acid residues of the protein active site screened in step 101, and to construct a fragment-protein interaction database.

5. The method for fragment-target interaction identification and universal compound library construction based on deep learning according to claim 1, characterized in that: The step 4 is to first screen the active pocket of the target protein, and then use a molecular cutting algorithm to cut the ligand molecule to obtain the interaction fingerprint between different molecular fragments in the target protein and the active site, that is, the interaction between the target protein and the molecular fragment.

6. The method for fragment-target interaction identification and universal compound library construction based on deep learning according to claim 1, characterized in that: The step 5 is to perform point cloud registration on the interaction between the target protein and the molecular fragment in step 104 and the interaction database constructed in step 103 using a point cloud registration algorithm.

7. The method for fragment-target interaction identification and universal compound library construction based on deep learning according to claim 1, characterized in that: The step 6 is based on the fitness and root mean square error scoring indicators, as well as the pharmacophore scoring function, to perform similarity scoring between the feature descriptors of different sub-pockets of the target protein and the aligned fragment library descriptors from the perspective of geometric features and pharmacophore properties.

8. The method for fragment-target interaction identification and universal compound library construction based on deep learning according to claim 1, characterized in that: The step 8 is based on a graph neural network, which iteratively constructs new molecules from the initial site of a given fragment by forming bonds one by one, and directly integrates the three-dimensional structure information into the molecular design process. The synthesizability of the generated molecules is evaluated, and the molecules with higher synthesizability scores are screened based on Lipinski's five rules and quantitative estimation of drug similarity, and several small molecule compounds with high rankings are selected as the final screening results.

Images