Lead compound optimization method and system based on skeleton constraint and training enhancement

Abstract

The invention discloses a lead compound optimization method and system based on skeleton constraint and training enhancement, and belongs to the field of artificial intelligence drug discovery. The method comprises the following steps: performing structural treatment on an initial lead compound, extracting a growth skeleton structure, and pairing a protein pocket structure; inputting information of the skeleton structure and the protein pocket structure into a coding model based on a three-dimensional graph neural network for joint coding to obtain structure coding representation; on the basis of a Delete model, an atomic distance-based resampling mechanism and a hydrophobic group mask are introduced, and a molecular generation model is constructed; training and tuning a molecule generation model through staged pre-training and a knowledge enhancement fine tuning strategy so as to improve the structural accuracy, the binding activity and the druggability of generated molecules; and inputting the structure coding expression into a trained and optimized molecular generation model PocketGrow to generate optimized molecules, screening the generated optimized molecules, and outputting a lead compound with development potential.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of artificial intelligence drug discovery, and particularly relates to a lead compound optimization method and system based on skeleton constraint and training enhancement. BACKGROUND

[0002] In modern drug research and development, the optimization of lead compounds is an important link connecting the initial hit compounds and clinical drug candidates. The goal is to improve the drug properties such as efficacy, safety, metabolic stability, and synthetic feasibility of the hit molecules by structural modification, while retaining the active skeleton. Traditional molecular optimization methods mainly rely on expert experience, rule-based fragment replacement, or heuristic algorithms based on local search. Although these methods are effective in some cases, they often lack consideration of the target structure environment, resulting in generated molecules that may not match the spatial structure of the protein binding pocket, thus limiting their real biological activity.

[0003] With the development of deep learning, neural network-based molecular generation methods have gradually become mainstream, such as recursive neural networks (RNN) and Transformer models based on SMILES sequences, GraphVAE, JT-VAE, MolGAN, GraphAF, and other methods based on graph structures. These methods can generate compounds with specific properties from scratch. These methods have made some progress in improving molecular diversity and generation efficiency, but their generation targets often deviate from specific pharmacophore skeletons, making it difficult to achieve targeted optimization. In addition, most models ignore the guiding role of protein structure and generate new structures based only on internal molecular features, lacking spatial perception of the shape, electrical properties, and hydrophobic distribution of the binding pocket.

[0004] To solve the above problems, some works have begun to incorporate protein structure information into molecular generation models. For example, based on docking scoring or interaction maps as auxiliary scoring items, the generation process is guided to converge towards high-affinity structures, or the three-dimensional structure of protein-ligand complexes is directly introduced for comparison and optimization. However, most of these methods still remain in the evaluation or screening stage and fail to incorporate target structure information into the core mechanism of structure generation. In addition, due to the complexity of three-dimensional spatial data, there are still few methods that can directly model the protein pocket and the molecule together and generate reasonable structures.

[0005] In recent years, three-dimensional graph neural networks (3D GNN) have received widespread attention in structural biology and computational drug discovery. Representative methods such as SchNet, EGNN, and GVP-GNN have the ability to perceive and model spatial geometric relationships such as atomic distances and directions, and can capture the microstructure characteristics between protein-ligand. However, existing research has mostly focused on prediction tasks such as structure scoring and virtual screening, and there is still a lack of systematic application in molecular optimization and generation tasks.

[0006] At the same time, the non-covalent interaction between protein-ligand (such as hydrophobic interaction, hydrogen bond, pi-stacking, etc.) in the role of drug binding is repeatedly verified, and the existing molecular generation model often lacks clear modeling of such mechanism, resulting in the generated structure lacking biophysical rationality. Especially in the fragment growing task, ignoring the spatial position control and functional mask mechanism will seriously reduce the optimization efficiency and the quality of the generated molecules.

[0007] In addition, in terms of training strategy, most of the current models rely on single supervised learning process, lack of pre-training, multi-task cooperation and data disturbance enhancement mechanism, which is not conducive to the improvement of model generalization ability. The lack of systematic training process for mask prediction, atom connection decision, and spatial coordination modeling in the fragment growing scene also makes it difficult for the model to adapt to the structural variation between different compounds and targets.

[0008] Therefore, how to construct a structure-guided molecular optimization method that integrates skeleton constraint, protein pocket structure information, spatial resampling mechanism, hydrophobic feature modeling and training enhancement strategy is still an important challenge in the field of computational molecular design, and is also the key to efficient lead compound optimization. SUMMARY

[0009] In view of the shortcomings of the prior art, the purpose of the present application is to provide a lead compound optimization method and system based on skeleton constraint and training enhancement, which solves the problems in the prior art.

[0010] The purpose of the present application can be realized by the following technical solutions:

[0011] The lead compound optimization method based on skeleton constraint and training enhancement comprises the following steps:

[0012] Performing structure processing on the initial lead compound, extracting the growable skeleton structure, and pairing the protein pocket structure;

[0013] Inputting the skeleton structure and protein pocket structure information into a coding model based on a three-dimensional graph neural network for joint coding to obtain a structure coding representation;

[0014] On the basis of the Delete model, an atom distance-based resampling mechanism and a hydrophobic group mask are introduced to construct a molecular generation model;

[0015] Through a phased pre-training and knowledge enhancement fine-tuning strategy, the molecular generation model is trained and optimized to improve the structure accuracy, binding activity and drugability of the generated molecules;

[0016] Input the structure coding representation into the training-optimized molecular generation model PocketGrow to generate optimized molecules, and screen the generated optimized molecules to output lead compounds with development potential.

[0017] Further, the step of extracting the skeleton structure and pairing the protein pocket structure comprises:

[0018] S101, processing the initial lead compound to extract the skeleton structure retaining the key pharmacophore, and preferentially selecting the non-cyclic single bond in the non-functional group region as the breaking point;

[0019] S102, obtaining the three-dimensional structure of the target protein corresponding to the skeleton, and locating the binding pocket region thereof;

[0020] S103, pairing the skeleton structure with the protein pocket structure and performing standardization processing.

[0021] Further, the joint coding step comprises:

[0022] S201, inputting the skeleton structure and the protein pocket structure corresponding thereto into a coding model based on a three-dimensional graph neural network, wherein the node features of the model include the type, spatial coordinates and charge information of atoms, and the edge features include the spatial distance and bond type between atoms;

[0023] S202, adopting a dual-graph coding mechanism in the three-dimensional graph neural network to code the skeleton graph and the protein pocket graph respectively, and then fusing the coding results of the two through a cross-attention module to output the fused dual-graph feature representation;

[0024] S203, using a geometric vector perception network to take the fused dual-graph feature representation output by S202 as input, and processing the scalar features and directional vector features contained therein to obtain a unified structure coding representation.

[0025] Further, the contents of the resampling mechanism comprise:

[0026] In the process of generating a new atom, a local coordinate system is constructed based on the current focus atom, and a hybrid density network is used to predict the spatial offset vector of the new atom relative to the focus atom;

[0027] After predicting the coordinates of each atom, a resampling mechanism is introduced, and when the Euclidean distance between the generated atom and the focus atom exceeds a threshold r max , the position is resampled until the spatial constraint is met;

[0028] The contents of the hydrophobic group mask include:

[0029] 1) Based on the protein-ligand complex structure, a non-covalent interaction analysis tool is used to identify a set of small molecule atoms that undergo hydrophobic interaction therein;

[0030] 2) Identify the complete group in which the hydrophobic atom is located using the SMARTS mode, obtain the topologically connected hydrophobic fragment structure, and generate the corresponding hydrophobic mask area;

[0031] 3) In the model training stage, a mask learning strategy is adopted to shield the hydrophobic mask area, and the model is set to predict the structure and atom type of the target task.

[0032] Further, the total loss function L of the molecular generation model during training all is:

[0033] L all = L focal + L pos + L atom + L bond

[0034] Wherein, L focal is the focus atom prediction loss, using binary cross-entropy loss function; L pos is the atom coordinate prediction loss, using negative log-likelihood loss function; L atom is the atom type prediction loss, using negative log-likelihood loss function; L bond is the bond type prediction loss, using negative log-likelihood form.

[0035] Further, the step of training the molecular generation model comprises:

[0036] Using the docking data set for pre-training, using the mask reconstruction target task to recover the learning of atoms, bonds and coordinate information at different positions;

[0037] On the basis of pre-training, introduce the complex data set for fine-tuning, so that the model focuses on the real protein-ligand binding mode and molecular conformation, and performs knowledge enhancement fine-tuning guided by drugability;

[0038] Further integrate the hydrophobic group mask information during fine-tuning, and combine the sparse area optimization mechanism.

[0039] The lead compound optimization system based on skeleton constraint and training enhancement comprises:

[0040] Data extraction module: structure processing is performed on the initial lead compound, and the growable skeleton structure is extracted, and the protein pocket structure is paired;

[0041] Data processing module: input the skeleton structure and protein pocket structure information into the encoding model based on three-dimensional graph neural network for joint encoding to obtain the structure encoding representation;

[0042] Model construction module: on the basis of the Delete model, a resampling mechanism based on atomic distance and a hydrophobic group mask are introduced to construct a molecular generation model PocketGrow;

[0043] Model training module: through a phased pre-training and knowledge enhancement fine-tuning strategy, the molecular generation model PocketGrow is trained and optimized to improve the structural accuracy, binding activity and drugability of the generated molecules;

[0044] Molecular generation and screening module: input the structural code representation into the trained and optimized molecular generation model PocketGrow to generate optimized molecules, and screen the generated optimized molecules to output lead compounds with development potential.

[0045] A computer storage medium stores a readable program, when the program runs, the program can instruct a computing device to execute the lead compound optimization method based on skeleton constraint and training enhancement as described above.

[0046] An electronic device includes a processor, a memory, a communication interface and a communication bus, the processor, the memory and the communication interface complete communication with each other through the communication bus;

[0047] The memory is used to store at least one executable instruction, and the executable instruction makes the processor execute the operations corresponding to the lead compound optimization method based on skeleton constraint and training enhancement as described above.

[0048] A computer program product includes computer instructions, the computer instructions instruct a computing device to execute the operations corresponding to the lead compound optimization method based on skeleton constraint and training enhancement as described above.

[0049] The beneficial effects of the present application are:

[0050] The present application innovatively integrates protein structure guidance and chemical knowledge driving, through the synergistic framework of structure perception generation, hydrophobic mask training and phased optimization, while ensuring the rationality of the molecular space, significantly improving the biological activity and drugability. Compared with the prior art, this method realizes breakthroughs in computing efficiency, activity-drugability synergistic optimization and other aspects, providing a new path for efficient and reliable targeted optimization for targeted drug research and development. BRIEF DESCRIPTION OF DRAWINGS

[0051] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments or prior art description will be briefly introduced below, and obviously, other drawings can also be obtained by those skilled in the art without creative labor.

[0052] Figure 1 is a flow chart of the lead compound optimization method of the present application;

[0053] Figure 2 is a flow chart of the atom generation and resampling mechanism of the present application;

[0054] Figure 3 is a flow chart of the hydrophobic mask training of the present application. DETAILED DESCRIPTION

[0055] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of the present application.

[0056] Embodiment 1

[0057] As shown in the following table, the lead compound optimization method based on skeleton constraint and training enhancement includes the following steps: Figure 1

[0058] Step S100, structure processing is performed on the initial lead compound, the growable skeleton structure is extracted, and the protein pocket structure is paired;

[0059] The process of extracting the skeleton structure and the protein pocket structure includes:

[0060] Step S101, the initial lead compound is processed, the skeleton structure retaining the key pharmacophore is extracted, and the breaking point is preferentially selected as the non-cyclic single bond in the non-functional group region. Because the non-functional group region usually does not directly participate in the specific interaction between the drug and the target, breaking at this point can maximize the retention of the key pharmacophoric characteristics of the lead compound, ensuring that the subsequent skeleton-based derivative compounds still have the ability to bind to the target. In addition, compared with the single bond or double bond within the ring, the chemical activity of the non-cyclic single bond is relatively low and the bond energy distribution is relatively uniform, and the breaking process is easier to control, which can effectively reduce the occurrence of side reactions and improve the selectivity and yield of skeleton extraction.

[0061] Step S102, the three-dimensional structure of the target protein corresponding to the skeleton is obtained, and the binding pocket region is located. The binding source can be a crystal structure, an AlphaFold (a protein structure prediction model developed by DeepMind, a company under Google, which can quickly and accurately predict the three-dimensional structure of a protein based on the amino acid sequence of the protein through a deep learning algorithm) predicted structure or a docking model;

[0062] Step S103, the skeleton structure is paired with the protein pocket structure and standardized processing is performed to ensure that there is a uniform spatial reference system when inputting the model.​

[0063] The standardized processing adopts the z-score standardization formula, that is:

[0064]

[0065] Wherein, z is the standardized data, X is the original data point, μ is the average value of the data set, and σ is the standard deviation of the data set. Through the formula, the structural coordinate data of different sources and scales is converted into standard normal distribution data with mean value of 0 and standard deviation of 1, eliminating the interference of data scale difference on subsequent model analysis, so that the model can more accurately learn and capture the interaction characteristics between the skeleton and the protein pocket.

[0066] Step S200, inputting the skeleton structure and the protein pocket structure information into the encoding model based on three-dimensional graph neural network for joint encoding to obtain the structure encoding representation;

[0067] The step of joint encoding includes:

[0068] Step S201: inputting the skeleton structure and its corresponding protein pocket structure into the encoding model based on three-dimensional graph neural network, wherein the node features of the model include the type, spatial coordinates and charge of atoms, and the edge features include the spatial distance and bond type between atoms;

[0069] S202, adopting a double-graph encoding mechanism in the three-dimensional graph neural network, wherein the skeleton graph and the protein pocket graph are subjected to local feature extraction through a graph convolution layer respectively to obtain respective graph encoding features, and then the two kinds of graph encoding features are fused through a cross-attention module to output a joint feature matrix of fused double-graph interaction information, realizing the structure information coupling and representation alignment between the skeleton and the protein pocket, and directly serving as the input data of step S203;

[0070] S203, using a geometric vector perceptron (GVP) to take the fused double-graph feature representation output by S202 as input, and processing the scalar feature s∈R n and the directional vector feature contained therein, so as to obtain a unified structure encoding representation

[0071] Step S300, introducing an atomic distance-based resampling mechanism and a hydrophobic group mask on the basis of the Delete model to construct a molecule generation model PocketGrow (pocket-constrained molecule growth model);

[0072] The Delete model is a "pocket-aware" three-dimensional molecular generation model designed for lead compound optimization. It analyzes the geometric and energy characteristics of protein pockets based on isometric geometric networks, designs molecules through physical constraints rather than large amounts of data, and directly generates natural conformations that match pockets, meets rotational and translational invariance to truly reflect three-dimensional spatial interactions. With a unified masking strategy, the model integrates four optimization tasks: fragment expansion, linker design, backbone transition, and side chain modification, to achieve comprehensive optimization of lead compounds and meet the needs of new target drug design.

[0073] (1) Resampling mechanism

[0074] Iteratively generate new atoms and introduce an atom distance-based resampling mechanism during fragment growth to ensure the structural rationality of the generated molecules.

[0075] As shown in Figure 2 , the contents of the resampling mechanism include:

[0076] 1) During the generation of new atoms, a local coordinate system is constructed based on the current focus atom, and a spatial offset vector of the new atom relative to the focus atom is predicted by a mixture density network (MDN).

[0077] 2) To ensure geometric rationality, a resampling mechanism is introduced after each atomic coordinate prediction. When the Euclidean distance d = ||x gen -x focus || (x gen is the coordinate of the generated atom, and x focus is the coordinate of the focus atom) between the generated atom and the focus atom exceeds the threshold r max (e.g. ), the position is resampled until the spatial constraint is met.

[0078] 3) The model PocketGrow training target is a weighted combination of multi-task loss functions, considering multiple sub-tasks such as focus atom prediction, atomic three-dimensional coordinate prediction, atomic type prediction, and bond type prediction. The total loss function L all is defined as follows:

[0079] L all = L focal + L pos + L atom + L bond

[0080] Where: L focal is the focus atom prediction loss, using a binary cross-entropy loss function (Binary CrossEntropy) to determine the current atom node to be expanded; L posFor the atomic coordinate prediction loss, a negative log likelihood loss function (NLL) is used to evaluate the model's prediction accuracy for the masked atoms; L atom For the atomic type prediction loss, a negative log likelihood loss function is also used to measure the model's ability to reconstruct the element type of the masked atoms; L bond For the bond type prediction loss, the model's prediction quality for the connection between the newly generated atoms and the existing structure is evaluated, also in the form of a negative log likelihood.

[0081] (2) Hydrophobic group mask

[0082] Using the information of hydrophobic interactions in protein-ligand complexes, a hydrophobic group mask is constructed to guide the model to preferentially generate fragments with hydrophobic characteristics;

[0083] As shown in Figure 3 , the content of the hydrophobic group mask includes:

[0084] 1) Based on the structure of protein-ligand complexes, the small molecule atom set that undergoes hydrophobic interaction is identified through non-covalent interaction analysis tools;

[0085] 2) The complete group in which the hydrophobic atom is located is identified using SMARTS patterns, obtaining topologically connected hydrophobic fragment structures and generating corresponding hydrophobic mask regions;

[0086] 3) In the training phase of the model PocketGrow, a mask learning strategy is adopted to shield the hydrophobic mask region and set the model's target task to predict its structure and atomic type, thereby guiding the model to strengthen its learning and reconstruction ability for key hydrophobic regions.

[0087] Step S400, train and optimize the molecular generation model PocketGrow through a phased pre-training and knowledge enhancement fine-tuning strategy to improve the structural accuracy, binding activity, and drugability of the generated molecules;

[0088] The steps for training and optimizing the molecular generation model PocketGrow include:

[0089] Step S401: Use large-scale docking datasets such as CrossDock for pre-training, adopt mask reconstruction target tasks to recover and learn the atomic, bond, and coordinate information in different positions, and improve the spatial generalization ability of the model;

[0090] Step S402: Based on pre-training, introduce high-quality crystal structure complex datasets (such as BindingMOAD) for fine-tuning, so that the model focuses on real protein-ligand binding patterns and molecular conformations;

[0091] The high-quality crystal structure referred to in this embodiment has a resolution of and an R-factor (an index for measuring the quality of crystal structure refinement) of no more than 0.25. By using such a data set, the model focuses on the real protein-ligand binding mode and molecular conformation, and the model's ability to simulate the actual biological system is enhanced;

[0092] Step S403: Further introduce a complex data set with high drugability characteristics. The data set needs to meet the following strict screening criteria:

[0093] Lipinski rule: molecular weight less than 500 daltons, number of hydrogen bond donors (such as hydroxyl, amino, etc.) no more than 5, number of hydrogen bond acceptors no more than 10, and log value of octanol-water partition coefficient (logP) less than or equal to 5;

[0094] SA (synthetic accessibility): The SA score needs to be less than or equal to 5 points (based on the commonly used Ertl and Schuffenhauer scoring system, the score range is 1-10 points, and the lower the score, the lower the difficulty of synthesis).

[0095] QED (quantitative estimate of drug similarity): QED value greater than or equal to 0.65 (QED value range 0-1, closer to 1 indicates that the compound has better drug similarity).

[0096] Step S404: Further integrate the hydrophobic group mask information in the above fine-tuning process, and combine the sparse area optimization mechanism to make the model generate a result that balances between structural rationality, biological activity and pharmacological properties.

[0097] Step S500, input the structure code representation obtained in S200 into the trained and optimized molecular generation model PocketGrow to generate optimized molecules, and screen the generated optimized molecules to output lead compounds with development potential;

[0098] The screening step of the optimized molecules is:

[0099] Step S501: Screen the generated optimized molecules according to multiple indicators, including but not limited to drugability (QED value), synthetic feasibility (SA score), and protein binding activity (molecular docking score);

[0100] Step S502: On the basis of meeting the above threshold conditions, perform structure similarity clustering on the molecule population with the same skeleton source, and retain the representative molecule with the best performance in each class to improve the structural diversity and practical value of the final result;

[0101] Step S503: Output the optimized lead compounds that meet the above comprehensive screening criteria for subsequent experimental verification and drug development.

[0102] Based on the similar inventive concept, the embodiment of the present application also provides a computer storage medium, which stores a readable program, when the program is run by a processor, the method for optimizing a lead compound based on skeleton constraint and training enhancement can be executed.

[0103] Based on the similar inventive concept, the embodiment of the present application provides an electronic device, comprising: a processor, a memory, a communication interface and a communication bus, the processor, the memory and the communication interface complete the communication among each other through the communication bus;

[0104] The memory is used to store at least one executable instruction, and the executable instruction makes the processor execute the operation corresponding to the method for optimizing a lead compound based on skeleton constraint and training enhancement.

[0105] Based on the similar inventive concept, the embodiment of the present application also provides a computer program product, comprising computer instructions, the computer instructions instruct a computing device to execute the operation corresponding to the method for optimizing a lead compound based on skeleton constraint and training enhancement.

[0106] Embodiment 2

[0107] In this embodiment, the novel coronavirus main protease (Nsp5) target is selected as an example target of the method to specifically introduce the optimization process of the lead compound.

[0108] As shown in the method for optimizing a lead compound based on skeleton constraint and training enhancement, the following steps are included: Figure 1

[0109] Step S100, performing structure processing on the initial lead compound, extracting a growable skeleton structure, and pairing a protein pocket structure;

[0110] Step S200, inputting the skeleton structure and the protein pocket structure information into a coding model based on a three-dimensional graph neural network for joint coding to obtain a structure coding representation;

[0111] Step S300, introducing an atomic distance-based resampling mechanism and a hydrophobic group mask on the basis of the Delete model to construct a molecular generation model PocketGrow;

[0112] Step S400, training and tuning the molecular generation model PocketGrow through a phased pre-training and knowledge enhancement fine-tuning strategy to improve the structure accuracy, binding activity and drugability of the generated molecules;

[0113] ​Step S500, input the structure code obtained in S200 into the trained and optimized molecular generation model PocketGrow to generate optimized molecules, and screen the generated optimized molecules to output lead compounds with development potential.

[0114] The following are the detailed execution steps of the above steps.

[0115] Step S100 is executed, including the following specific steps:

[0116] Step S101: Skeleton extraction, initial lead compound: SMILES expression CC(C)N(Cc1ccccc1)C(=O)Cn1nnc2ccccc21 (molecular weight 308.39, LogP = 2.87, SA = 1.99, QED = 0.73). Manual clipping is performed using the PyMOL tool to obtain two fragments: Frag1: break the single bond between the benzene ring and the nitrogen atom, retain the right heterocyclic structure, Frag2: break the single bond between the carbonyl group and the nitrogen atom, retain the left benzene ring structure.

[0117] Step S102: Protein pocket acquisition, target: SARS-CoV-2 main protease Nsp5 (PDB ID: 6LU7), centered on the original ligand binding site in the crystal structure, radius Region as pocket region

[0118] Step S103: Data standardization, align the skeleton and the protein pocket to the same coordinate system, and normalize the atomic coordinates to zero mean unit variance.

[0119] Step 200 is executed, including the following specific steps:

[0120] Step S201: Three-dimensional graph neural network architecture, input features: atom types (12 elements such as C / N / O) are encoded using One-hot, geometric features such as coordinates (x, y, z), charge, and van der Waals radius, edge features such as bond types (single / double / triple bond) and spatial distance;

[0121] Step S202: In the dual-graph fusion mechanism, first encode the skeleton graph and the protein pocket graph in parallel: the skeleton graph takes the atoms of Frag1 and Frag2 as nodes, node features include One-hot encoded atom types (such as C / N / O), charge, and van der Waals radius, edge features include bond types (single bond / double bond) and three-dimensional spatial distance between atoms; the protein pocket graph takes the amino acid residues (such as His41, Cys145, etc. Key catalytic residues) in the Nsp5 binding pocket as nodes, node features include amino acid type, side chain charge, and coordinate information, and edge features are the spatial distance and relative orientation between residues.

[0122] Step S203: The encoding module is:

[0123]

[0124] wherein scalar_out is a scalar feature transformation, and vector_out is a vector feature rotation equivariant.

[0125] Step 300 is performed, including the following specific steps:

[0126] (1) Iteratively generate new atoms, and introduce an atomic distance-based resampling mechanism in the fragment growth process to ensure the structural rationality of the generated molecules; as shown in Figure 2 The content of the resampling mechanism includes:

[0127] 1) Atomic generation, first select atoms with Sigmoid output > 0.7 as growth points, then use the MDN network to predict coordinates and output relative offset delta_r;

[0128]

[0129] wherein delta_r is a predicted offset vector, and dist is a bond length constraint;

[0130] 2) Introduce a resampling mechanism after each atomic coordinate prediction:

[0131] 3) The model training target is a weighted combination of multi-task loss functions, with weights set as: λ_focal=0.3, λ_pos=0.5, λ_atom=0.1, λ_bond=0.1, and the total loss decreases after 200 epoch pre-training.

[0132] (2) Use the hydrophobic interaction information in the protein-ligand complex to construct a hydrophobic group mask to guide the model to preferentially generate fragments with hydrophobic characteristics; as shown in Figure 3 The content of the hydrophobic group mask includes:

[0133] S1, input the protein-ligand complex, analyze it by PLIP, and identify the hydrophobic interaction (Phe140 in Nsp5 interacts with the benzene ring of the ligand)

[0134] S2, perform SMARTS matching on the hydrophobic atoms, with the matching pattern being: [#6]~[#6](~[#6])~[#6] (alkyl / aromatic ring pattern), and the matching pattern is explained as follows:

[0135] [#6] represents a carbon atom (because carbon has an atomic number of 6 in the periodic table of elements); ~ represents any type of connection (bond), which is more general than specific single bond-, double bond=, etc.

[0136] S3, in the model training stage, a mask learning strategy is adopted, and the mask label is that the methyl atoms {C1, C2, C3} in Frag1 are marked, and the fine-tuning effect is shown in Table 1 as follows:

[0137] Table 1 Influence of hydrophobic group mask on model performance

[0138] Model version Average number of hydrophobic groups Vina docking score PocketGrow-CM-K 5.81±0.1 -6.44±0.25 PocketGrow-CM-KH 6.56±0.4 -6.71±0.42

[0139] Among them, PocketGrow-CM-K is a model without using hydrophobic mask; PocketGrow-CM-KH is a model after introducing hydrophobic mask.

[0140] The step S400 includes the following specific steps:

[0141] Step S401: Pre-training is performed using large-scale docking data sets such as CrossDock, and mask reconstruction target task is used to recover and learn the atomic, bond and coordinate information at different positions, so as to improve the spatial generalization ability of the model;

[0142] Step S402: On the basis of pre-training, introduce high-quality crystal structure complex data set (such as BindingMOAD) for fine-tuning, so that the model focuses on the real protein-ligand binding mode and molecular conformation;

[0143] Step S403: Further introduce complex data with high drugability characteristics (meet the screening standards such as Lipinski rule, SA, QED, etc.), and perform drugability-oriented knowledge enhancement fine-tuning;

[0144] Step S404: Further integrate hydrophobic group mask information in the above fine-tuning process, and combine the sparse area optimization mechanism, so that the model generates results balanced between structural rationality, biological activity and pharmacological properties.

[0145] The data set configuration in the training process is as shown in Table 2:

[0146] Table 2 Data set configuration of each stage of model training

[0147] Stage Dataset Number of samples Pre-training (S501) CrossDock 100000 Conformational fine-tuning (S502) BindingMOAD 35516 Drugability fine-tuning (S503) High drugability subset 7064

[0148] The step S500 includes the following specific steps:

[0149] Step S501: Multi-index screening, threshold setting: QED≥0.65, SA≤5.0, vina score≤6.5kcal / mol, meet the Lipinski rule, and the optimization result example is as follows:

[0150] Optimized molecule:

[0151]

[0152] Optimized molecules:

[0153]

[0154] Step S502: Structure clustering, using Butina clustering (similarity threshold of 0.7), retaining 5 structure cluster representative molecules;

[0155] Step S503: Comprehensive output of the following results, as shown in Table 3 below:

[0156] Table 3 Comparison of key properties of optimized molecules and initial compounds

[0157]

[0158] Example 3

[0159] In this embodiment, a lead compound optimization system based on skeleton constraint and training enhancement is proposed, which specifically includes:

[0160] Data extraction module: structure processing is performed on the initial lead compound, the skeleton structure that can be grown is extracted, and the protein pocket structure is paired;

[0161] Data processing module: input the skeleton structure and protein pocket structure information into the encoding model based on three-dimensional graph neural network for joint encoding to obtain structure encoding representation;

[0162] Model construction module: on the basis of Delete model, introduce resampling mechanism based on atomic distance and hydrophobic group mask to construct molecular generation model PocketGrow;

[0163] Model training module: train and optimize the molecular generation model PocketGrow through the phased pre-training and knowledge enhancement fine-tuning strategy to improve the structure accuracy, binding activity and drugability of the generated molecules;

[0164] Molecule generation and screening module: input the structure encoding representation into the molecular generation model PocketGrow trained and optimized to generate optimized molecules, and screen the generated optimized molecules to output lead compounds with development potential.

[0165] The methods of the present invention can be implemented in hardware, firmware, or as software or computer code that can be stored in a recording medium (such as a CD-ROM, RAM, floppy disk, hard disk, or magneto-optical disk), or as computer code originally stored on a remote recording medium or a non-transitory machine-readable medium and subsequently stored on a local recording medium, downloaded via a network. Thus, the methods described herein can be processed by software stored on a recording medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware (such as an ASIC or FPGA). It is understood that the computer, processor, microprocessor controller, or programmable hardware includes storage components (e.g., RAM, ROM, flash memory, etc.) capable of storing or receiving software or computer code that, when accessed and executed by the computer, processor, or hardware, implements the methods described herein. Furthermore, when a general-purpose computer accesses the code used to implement the methods shown herein, the execution of the code transforms the general-purpose computer into a dedicated computer for performing the methods shown herein.

[0166] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention.

Claims

1. A lead compound optimization method based on skeleton constraints and training enhancement, characterized in that, Includes the following steps: The initial lead compound was structurally processed to extract a growable backbone structure and pair it with protein pocket structures. The backbone structure and protein pocket structure information are input into a coding model based on a three-dimensional graph neural network for joint encoding to obtain a structural coding representation; Based on the Delete model, a resampling mechanism based on atomic distance and a hydrophobic group mask are introduced to construct a molecular generation model; By employing a phased pre-training and knowledge-enhanced fine-tuning strategy, the molecular generation model is trained and optimized to improve the structural accuracy, binding activity, and drug-likeness of the generated molecules. The structure encoding is input into the trained and optimized molecular generation model PocketGrow to generate optimized molecules. The generated optimized molecules are then screened to output lead compounds with development potential.

2. The lead compound optimization method based on skeleton constraints and training enhancement according to claim 1, characterized in that, The steps for extracting the backbone structure and pairing protein pocket structures include: S101, the initial lead compound is processed to extract and retain the skeletal structure of the key pharmacophores, and the break point is preferentially selected from the non-cyclic single bonds in the non-functional group region. S102, obtain the three-dimensional structure of the target protein corresponding to the backbone and locate its binding pocket region; S103 pairs the scaffold structure with the protein pocket structure and normalizes the process.

3. The lead compound optimization method based on skeleton constraints and training enhancement according to claim 1, characterized in that, The steps of the joint encoding include: S201, the backbone structure and its corresponding protein pocket structure are input together into a coding model based on a three-dimensional graph neural network. The node features of the model include the type of atom, spatial coordinates and charge information, and the edge features include the spatial distance between atoms and the bond type. S202 employs a dual-graph encoding mechanism in a 3D graph neural network, encoding the skeleton graph and the protein pocket graph separately, and then fusing the encoding results of the two through a cross-attention module to output the fused dual-graph feature representation; S203 utilizes a geometric vector perception network, taking the fused dual-image feature representation output by S202 as input, and simultaneously processes the scalar features and direction vector features contained therein to obtain a unified structural encoding representation.

4. The lead compound optimization method based on skeleton constraints and training enhancement according to claim 1, characterized in that, The resampling mechanism includes: During the generation of new atoms, a local coordinate system is constructed based on the current focal atom, and the spatial offset vector of the new atom relative to the focal atom is predicted by a hybrid density network. A resampling mechanism is introduced after each atomic coordinate prediction. When the Euclidean distance between the generated atom and the focal atom exceeds the threshold r, a resampling mechanism is implemented. max When this happens, the location is resampled until the spatial constraints are met; The content of the hydrophobic group mask includes: 1) Based on the structure of protein-ligand complexes, identify sets of small molecular atoms that undergo hydrophobic interactions using non-covalent interaction analysis tools; 2) Use SMARTS mode to identify the intact groups in which hydrophobic atoms are located, obtain the topologically connected hydrophobic fragment structure, and generate the corresponding hydrophobic mask region; 3) During the model training phase, a mask learning strategy is adopted to shield the hydrophobic mask region and set the target task of predicting its structure and atom type.

5. The lead compound optimization method based on skeleton constraints and training enhancement according to claim 4, characterized in that, The total loss function L during the training of the molecular generation model all for: L all =L focal +L pos +L atom +L bond Among them, L focal For the focus atom prediction loss, a binary cross-entropy loss function is used; L pos For the loss prediction of atomic coordinates, the negative log-likelihood loss function is used; L atom For predicting loss based on atomic type, the negative log-likelihood loss function is used; L bond Losses are predicted for chemical bond types using negative log-likelihood.

6. The lead compound optimization method based on skeleton constraints and training enhancement according to claim 4, characterized in that, The steps for training the molecular generation model include: Pre-training was performed using the docking dataset, and the target reconstruction task using masking was used to recover and learn the atomic, bond, and coordinate information at different positions. Based on pre-training, a complex dataset is introduced for fine-tuning, so that the model focuses on the real protein-ligand binding patterns and molecular conformations, and performs drug-oriented knowledge-enhanced fine-tuning. During the fine-tuning process, hydrophobic group mask information is further integrated, and a sparse region optimization mechanism is combined.

7. A lead compound optimization system based on skeleton constraints and training enhancement, characterized in that, include: Data extraction module: Performs structural processing on the initial lead compound, extracts the growable backbone structure, and pairs it with protein pocket structures; Data processing module: Inputs the backbone structure and protein pocket structure information into a coding model based on a three-dimensional graph neural network for joint encoding to obtain a structural coding representation; Model building module: Based on the Delete model, an atomic distance-based resampling mechanism and hydrophobic group mask are introduced to build the molecular generation model PocketGrow; Model training module: Through phased pre-training and knowledge-enhanced fine-tuning strategies, the PocketGrow molecular generation model is trained and optimized to improve the structural accuracy, binding activity, and drug-likeness of generated molecules. Molecular generation and screening module: The structure encoding representation is input into the trained and optimized molecular generation model PocketGrow to generate optimized molecules, and the generated optimized molecules are screened to output lead compounds with development potential.

8. A computer storage medium storing a readable program, characterized in that, When the program runs, it can instruct the computing device to perform the lead compound optimization method based on skeleton constraints and training enhancement as described in any one of claims 1-6.

9. An electronic device, characterized in that, include: The processor, memory, communication interface, and communication bus are provided, wherein the processor, memory, and communication interface communicate with each other via the communication bus. The memory is used to store at least one executable instruction that causes the processor to perform the operation corresponding to the lead compound optimization method based on skeleton constraints and training enhancement as described in any one of claims 1-6.

10. A computer program product comprising computer instructions, characterized in that, The computer instructions instruct the computing device to perform the operations corresponding to the lead compound optimization method based on skeleton constraints and training enhancement as described in any one of claims 1-6.

Images