Molecular generation and optimization method based on protein large language model

Abstract

The invention relates to the field of artificial intelligence assisted drug discovery, in particular to a protein large language model-based molecule generation and optimization method, which comprises the following steps of: acquiring amino acid sequence information and three-dimensional structure information of a target protein pocket; encoding the amino acid sequence of the protein pocket by using a protein encoder constructed based on a protein large language model to obtain a protein pocket feature vector; using a context encoder module to encode the context information according to a preset molecule generation mode (de novo generation or optimization based on a seed compound) to obtain a latent vector; and fusing the protein pocket feature vector with the latent vector. According to the method, accurate representation of the protein pocket is realized by utilizing the protein large language model, and a generation-screening-optimization iterative drug design strategy is developed by supporting a unified framework of two generation modes, so that the targeting specificity of generated molecules and the overall efficiency of drug design are improved.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence-assisted drug discovery technology, and more specifically to a molecular generation and optimization method based on a large protein language model. Background Technology

[0002] Drug discovery has long been considered a high-investment, high-risk, and lengthy systematic project. With the rapid development of artificial intelligence, deep learning-based molecular generative models are beginning to emerge in the field of drug design. Among them, protein pocket-based molecular generative models, which can directly utilize target structural information to guide drug molecule design, show great application potential.

[0003] Currently, several molecular generation models based on protein pockets have been developed, such as TargetDiff, which employs an isovariant diffusion model; Pocket2Mol, which utilizes efficient sampling from 3D protein pockets; and PocketFlow, an autoregressive flow model driven by both data and knowledge. However, existing molecular generation methods based on protein pockets still face several technical bottlenecks:

[0004] (1) The characterization accuracy of protein pockets needs to be further improved. Protein pockets have diverse physicochemical properties and complex three-dimensional geometry. Existing models may fail to capture these key information fully and accurately. This is mainly due to the insufficient ability of protein encoders to capture long-range dependencies and fine structural features, which affects the matching degree and binding specificity between the generated molecules and the pocket.

[0005] (2) There is a gap between the molecular generation process and the actual drug optimization process. Most existing models focus on de novo molecule generation, while in drug development practice, researchers often need to modify and optimize the structure of known active molecules (i.e., seed compounds) in order to obtain candidate drugs with better performance. Existing models do not adequately support this iterative drug design process of "generation-screening-optimization".

[0006] (3) The evaluation system for generated molecules still needs improvement. Currently, the evaluation indicators mainly rely on property calculations and molecular docking scores, and the systematic evaluation of pharmacokinetic properties such as ADMET has not been fully incorporated into the research framework. The lack of ADMET property evaluation may lead researchers to overestimate the clinical translation potential of generated molecules.

[0007] Therefore, how to more accurately understand and utilize protein pocket information, effectively integrate de novo generation and iterative optimization as two generation modes, and at the same time construct a multi-dimensional evaluation and screening system including ADMET are problems that molecular generation models urgently need to solve in the field of drug design. Summary of the Invention

[0008] To overcome the aforementioned deficiencies of the prior art, this invention provides a molecular generation and optimization method based on a protein large language model to solve the problems existing in the background art.

[0009] In view of the above, the purpose of this invention is to provide a molecular generation and optimization method based on a protein big language model. This method can more accurately utilize protein pocket information through a deeply pre-trained protein big language model, and combine it with an iterative drug design strategy of "generation-screening-optimization" that includes a multi-dimensional molecular evaluation system. According to different generation requirements (de novo generation or optimization based on seed compounds), contextual information can be flexibly introduced to generate high-quality new molecules with target specificity or optimize existing molecules. It is particularly suitable for drug development scenarios with high requirements for ADMET properties.

[0010] On one hand, this invention provides a molecular generation and optimization method based on a protein big language model, comprising: acquiring the amino acid sequence information and three-dimensional structural information of a target protein pocket; using a protein encoder constructed based on the protein big language model to encode the amino acid sequence information of the protein pocket to obtain a protein pocket feature vector; using a context encoder module to encode context information to obtain a latent vector according to a preset molecular generation mode (de novo generation or optimization based on seed compounds); fusing the feature vector of the protein pocket with the latent vector to obtain a fused feature; and using a compound decoder to generate a new molecular structure based on the fused feature.

[0011] Furthermore, the protein big language model is a pre-trained ESM Cambrian (ESM C) model or other ESM series protein big language models.

[0012] Furthermore, the context encoder module includes: when the molecule generation mode is de novo generation, the module samples the latent vector from a preset prior distribution (such as a standard normal distribution); when the molecule generation mode is seed compound-based optimization, the module encodes the input seed compound information and protein pocket information based on a variational autoencoder architecture, and obtains the latent vector through a reparameterization method.

[0013] Furthermore, the fusion method between the feature vector of the protein pocket and the latent vector includes vector concatenation or element-wise summation.

[0014] Furthermore, the compound decoder is based on a pre-trained GPT model and generates the new molecular structure represented as a string of SMILES in an autoregressive manner according to the fusion features.

[0015] Secondly, this invention also provides an iterative drug design strategy of "generation-screening-optimization", including: in the de novo generation stage, only the target protein pocket information is input to generate molecules and obtain an initial compound library; a multi-dimensional molecular evaluation system is constructed to comprehensively evaluate the molecules in the initial compound library, and one or more seed compounds are selected from the initial compound library according to the evaluation indicators; in the seed compound-based optimization stage, the target protein pocket information and seed compounds are input into the model to generate molecules and obtain an optimized compound library; the molecules in the optimized compound library are evaluated in the same way as the initial compound library, and one or more seed compounds are selected from the optimized compound library according to the evaluation indicators; the seed compound-based optimization and evaluation screening process is repeated until potential candidate compounds with high binding affinity, excellent drug properties and easy synthesis are obtained.

[0016] Furthermore, the multidimensional molecular evaluation system includes binding affinity, drug-likeness, synthetic feasibility, lipid-water partition coefficient, compliance with Lipinski's rule, and pharmacokinetic properties such as absorption, distribution, metabolism, excretion, and toxicity (ADMET).

[0017] The technical effects and advantages of this invention are as follows:

[0018] The core of this invention lies in using a large protein language model to accurately extract protein pocket features and constructing a unified framework that can flexibly switch between "de novo generation" and "seed compound-based optimization" modes. Combined with a multi-dimensional molecular evaluation system including ADMET properties, this invention designs an iterative drug design strategy of "generation-screening-optimization": first, seed compounds are generated and screened in the de novo generation mode, and then the seed compounds are iteratively optimized and screened in the optimization mode until candidate molecules with high binding affinity and excellent drug properties are obtained, thereby significantly improving the target specificity of generated molecules and the efficiency of drug discovery. Attached Figure Description

[0019] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0020] Figure 1 This is a schematic flowchart of the molecular generation and optimization method based on a protein-based large language model provided in an embodiment of the present invention.

[0021] Figure 2This is a schematic diagram of the molecular generation and optimization method based on a protein-based large language model provided in an embodiment of the present invention.

[0022] Figure 3 This is a flowchart illustrating the "generation-screening-optimization" iterative drug design strategy provided in this embodiment of the invention. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0024] It is understood that the terms “first,” “second,” etc., used in this application may be used herein to describe various elements, but unless otherwise stated, these elements are not limited by these terms. These terms are used only to distinguish one element from another.

[0025] To enable those skilled in the art to better understand the technical solutions in this application, the technical solutions in the embodiments of this application will be clearly and completely described below. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0026] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0027] It should be noted that existing protein pocket-based molecular generation models have limited ability to extract protein pocket features, resulting in low adaptability between generated molecules and target pockets. Furthermore, most models only support de novo molecular generation, failing to meet the needs of structural optimization of known active molecules in actual drug development. In addition, existing evaluation systems generally neglect the assessment of ADMET properties of molecules. For example, traditional convolutional neural network (CNN)-based protein encoders can only capture local geometric features, failing to capture features that are more than a certain distance from the target pocket. The synergistic effects between amino acid residues are difficult to model, leading to frequent mismatches between hydrophobic groups and pocket polar regions in the generated molecules. Furthermore, about 60% of the molecules generated by existing models are eliminated in the early screening because they fail to pass hERG toxicity prediction, highlighting the necessity of ADMET evaluation.

[0028] In view of this, the present invention proposes a molecular generation and optimization method based on a protein-based large language model, specifically, Figure 1 This is a schematic flowchart of the molecular generation and optimization method based on a protein-based large language model provided in an embodiment of the present invention.

[0029] like Figure 1 As shown, the method includes steps S110-S150, and steps S110-S150 and related steps will be described in detail below.

[0030] S110, obtains the amino acid sequence information and three-dimensional structure information of the target protein pocket.

[0031] Protein pockets are cavities on or inside proteins that are suitable for binding to ligands (such as small molecules, drugs, or other biomolecules). Their amino acid sequence and three-dimensional structural information are crucial for structured drug design. Specifically, pocket regions can be automatically identified using the "FindPocket" function in PyMOL software, or a radius can be defined centered on a known ligand. The scope of the pocket is defined; for proteins without ligand structures, molecular dynamics simulations can be used to predict the dynamic pocket conformation. For example, in pocket recognition of the EGFR kinase domain, a pocket sequence consisting of 120 amino acids containing key residues such as Lys721 and Cys797 is extracted by referring to the coordinates of the ATP binding site (PDBID: 1M17).

[0032] The straightforward approach is to download the amino acid sequence and structure file of a protein from a protein database, and then identify pockets within the protein structure based on reference ligand coordinates, thus obtaining the amino acid sequence and three-dimensional structure information of the target protein pocket. AlphaFold3 is preferably used to predict the three-dimensional structure of protein pockets without experimental structures, achieving a prediction accuracy of up to C0. α atom Approaching the level of X-ray crystal diffraction.

[0033] Based on obtaining the amino acid sequence information and three-dimensional structure information of the target protein pocket in step S110, step S120 is further performed.

[0034] S120, using a protein encoder constructed based on a protein big language model, the amino acid sequence information of the protein pocket is encoded to obtain the protein pocket feature vector h. pocket .

[0035] Protein big language models, such as ESM Cambrian (ESM C) or other ESM series models, are pre-trained on massive protein sequence data using a masked language model paradigm to grasp protein sequence patterns and potential structural and functional characteristics. Taking the ESM C model as an example, its training data contains 230 million UniRef50 protein sequences. By predicting the types of masked amino acids (a total of 20 natural amino acids), it learns the co-evolutionary relationships of amino acids in the sequence. For example, when a position in the sequence contains a hydrophobic amino acid, the probability that its downstream components tend to contain nonpolar amino acids can be quantitatively represented by the model's attention weights.

[0036] In this embodiment, the protein encoder uses the amino acid sequence information of the protein pocket as input, enabling it to simultaneously capture the physicochemical properties of the amino acids, long-range interaction information, and potential local geometric features of the protein pocket. This information is crucial for understanding molecular recognition and binding. The encoder output is the feature vector h of the protein pocket. pocket This vector can provide a deep and accurate description of the pocket's environment.

[0037] In step S120, the protein pocket feature vector h is obtained. pocket Based on this, further, step S130 is performed.

[0038] S130 utilizes the context encoder module to encode context information to obtain the latent vector z according to the preset molecule generation mode (de novo generation or optimization based on seed compounds).

[0039] The latent vector z serves as a control signal here, guiding the subsequent molecule generation process. This embodiment supports two main molecule generation modes:

[0040] (1) De novo molecule generation mode: In this mode, there is no specific seed compound as input. The latent vector z is randomly sampled from a pre-defined, molecule-independent prior probability distribution (usually a standard normal distribution N(0, I)). This latent vector sampled from the prior distribution provides an unbiased, exploratory contextual guide, allowing the model to freely explore the chemical space within the constraints of the protein pocket. During the de novo generation stage, the chemical space diversity of the generated molecules can be controlled by adjusting the sampling temperature parameter (e.g., from 0.5 to 1.2). The higher the temperature, the higher the structural novelty of the generated molecules, which is suitable for lead compound discovery scenarios.

[0041] (2) Seed Compound-Based Molecular Optimization Model: Here, one or more known seed compounds exist, and the goal is to generate new molecules with similar structures but potentially superior properties based on these seed compounds. This embodiment introduces a context encoder based on a variational autoencoder architecture. The context encoder simultaneously encodes the input seed compound information and protein pocket information, and its output is a continuous latent vector space, including the mean (μ) and standard deviation (σ) of the latent vectors. The latent vector z is obtained through a reparameterization method, i.e., z = μ + ∈ · exp(logσ), where ∈ ~ N(0, I). In this way, the obtained latent vector z can capture the characteristics of the seed compound in a specific protein pocket environment, rather than just the isolated characteristics of the seed compound itself. The input form of the seed compound is the string SMILES, which is input into the context encoder along with the protein pocket information. For example, when optimizing the known EGFR inhibitor gefitinib, the model can search in the latent vector space in the direction of "reducing CYP3A4 metabolic rate," resulting in a derivative with 1.5 times improved metabolic stability.

[0042] Based on obtaining the latent vector z in step S130, step S140 is further executed.

[0043] S140, the protein pocket h pocket The feature vector is fused with the latent vector z to obtain the fused feature h. fusion .

[0044] Specifically, the protein pocket feature vector h obtained in S120 pocket The latent vector z obtained in S130 is then fused. This fusion can be achieved through vector concatenation or element-wise summation, i.e., h... fusion =[h pocket ;z] or h fusion =h pocket +z. The obtained fusion feature h fusion This will serve as a condition for subsequent molecular decoding.

[0045] The fusion feature h is obtained in step S140. fusion Based on this, further, step S150 is performed.

[0046] S150, using a compound decoder, a new molecular structure is generated based on the fusion features.

[0047] The compound decoder is based on a pre-trained GPT architecture. It is first pre-trained on 10 million randomly selected SMILES from PubChem to learn common chemical language rules. After pre-training, a cross-attention module is introduced, using the fused features obtained in S140 as conditional input. The compound decoder then generates SMILES characters that constitute new molecules one by one in an autoregressive manner, ultimately outputting a complete molecular structure.

[0048] Figure 2 This is a schematic diagram of a molecular generation and optimization method based on a protein-based large language model provided in an embodiment of this disclosure.

[0049] like Figure 2 As shown, the molecular generative model based on the protein big language model mainly includes: a pre-trained protein big language model, a context encoder, and a pre-trained compound decoder.

[0050] A pre-trained protein big language model is used to convert the amino acid sequence information of protein pockets into feature vectors for those pockets. Specifically, the pre-trained protein big language model is an ESM-C model with 300 million parameters. This ESM-C model is composed of stacked identical modified Transformer modules, employing a 30-layer structure with a hidden dimension of 960. Within each module, the input is first normalized once, followed by multi-head self-attention computation, which uses rotational position embedding instead of traditional absolute position encoding. The attention output is then concatenated with the original input via residuals and merged with a fixed scaling factor to enhance deep gradient stability. The module then applies layer normalization again and passes through a two-layer feedforward neural network using the SwiGLU activation function. Bias terms are disabled in this feedforward neural network, all linear mappings, and layer normalization to reduce redundancy and further stabilize training. After 30 stacked Transformer modules, the output of the protein big language model is the feature vector h of the protein pocket. pocket Rotational position embedding uses trigonometric functions to define positional encoding, avoiding the positional information confusion problem that occurs with absolute positional encoding in long sequences.

[0051] A context encoder, used to convert reference ligand (or seed compound) and protein pocket information into latent vectors, is activated only during training and seed compound-based molecular optimization modes to enhance the model's control over the compound generation process. Specifically, the context encoder is based on a variational autoencoder architecture, employing a 4-layer structure of the standard Transformer module with a hidden dimension of 960. In both training and seed compound-based molecular optimization modes, the context encoder first determines the mean (μ) and standard deviation (σ) for each given pair of compound and protein pocket features. Then, it samples the latent vector z from the distribution consisting of μ and σ and adds it to the output of the protein encoder to obtain the fused feature h. fusion The input compounds are then passed to the compound decoder. During training, the model aims to reconstruct the input compounds; however, in practical applications, once seed compounds are provided, the context encoder can assist in further optimization of the compounds. During training, the KL divergence weight β (e.g., gradually increasing from 0.1 to 1) can balance the information retention capability of the latent vectors with the alignment with the prior distribution.

[0052] If the preset molecule generation mode is de novo, the context encoder will not be activated, and the latent vector z is directly sampled randomly from the standard normal distribution N(0, I). Similarly, the latent vector z is added to the output of the protein encoder to obtain the fused feature h. fusion Then it is passed to the compound decoder.

[0053] A pre-trained compound decoder is used to fuse features h fusion The SMILES string is converted into the new molecule. Specifically, the pre-trained compound decoder is composed of stacked identical standard Transformer modules in a 12-layer structure with a hidden dimension of 768.

[0054] This embodiment uses the CrossDocked2020 dataset, which contains 22.5 million docking protein-ligand pairs of varying quality. First, pairs with a root mean square bias (RMSD) of binding conformation greater than [value missing] are removed. The data was used to obtain a refined subset containing 184,057 data points. Then, the data was clustered using the mmseqs2 tool with a 30% sequence identity threshold. Finally, 100,000 protein-ligand pairs were randomly selected as the training set, and 100 proteins from the remaining clusters were selected as the test set.

[0055] During the training phase, the overall training objective of the model is to minimize the following formula:

[0056]

[0057] Where x is the protein pocket; y is the reference ligand; D is the training corpus, i.e., the set of (protein pocket, reference ligand) pairs; z in logP(...) is sampled from q(z|x,y); β is a hyperparameter; p(z) represents the standard Gaussian distribution; D kl Let θ represent the KL divergence; θ is the model parameter. The first term in the formula teaches the model how to generate reasonable ligands based on the input pocket, while the second term is the KL divergence constraint, which ensures that the learned latent space is similar to the prior distribution p(z), thus promoting smooth interpolation and good generalization ability. This constraint regularizes the encoder by approximating the posterior distribution (i.e., the encoder's output) to the prior distribution (usually a simple Gaussian distribution), helping to prevent overfitting and ensuring meaningful latent vectors are obtained.

[0058] The following comparative experiments, based on the molecular generation and optimization model of the protein big language model provided in the embodiments of the present invention, are used to fully illustrate the effectiveness of the scheme.

[0059] For 100 target proteins in the CrossDocked2020 test set, 100 compounds were generated using both the baseline model (BaseMol) and the protein big language-based model provided in this embodiment (PLLMol). Unlike PLLMol, which uses a pre-trained protein big language model, BaseMol's protein pocket encoding module is composed of stacked standard Transformer modules, employing a 4-layer structure with a hidden dimension of 256; the context encoder and compound decoder structures are identical to those of PLLMol. Ultimately, BaseMol generated 9936 effective molecules, while PLLMol generated 9911 effective molecules. Table 1 summarizes the mean properties of the effective molecules generated by the two models.

[0060] Table 1. Mean properties of molecules generated by the baseline model and examples provided.

[0061]

[0062] The docking score reflects the binding free energy between the molecule and the target protein. According to thermodynamic principles, a lower docking score indicates a more stable binding conformation between the molecule and the target, and a stronger affinity between them. This experiment used QuickVina2 for molecular docking calculations. The center coordinates of the docking box were set to the center coordinates of the reference ligand, and the size of the docking box was set to... The search intensity is set to 16, and the other parameters remain at their default values.

[0063] The Quantitative Estimate of Drug-likeness (QED) comprehensively considers eight parameters: molecular weight, lipid-water partition coefficient, number of hydrogen bond donors, number of hydrogen bond acceptors, topological polar surface area, number of rotatable bonds, number of aromatic rings, and structural warning signs. It calculates a score from 0 (non-drug-like) to 1 (highly drug-like) using a weighted geometric mean to assess the overall drug-likeness of a molecule.

[0064] The Synthetic Accessibility Score (SAS) quantifies the difficulty of molecule synthesis by analyzing the complexity of molecular fragments and ring systems, combined with predefined fragment contribution values. The score ranges from 1 (easy to synthesize) to 10 (difficult to synthesize). It is worth noting that the Normalized SAS used in this study ranges from 0 (difficult to synthesize) to 1 (easy to synthesize), and its calculation formula is as follows:

[0065]

[0066] Based on statistical analysis of marketed oral drugs, Lipinski's five rules propose four empirical thresholds that small molecule drugs should meet: molecular weight ≤ 500 Da, lipid-water partition coefficient ≤ 5, number of hydrogen bond donors ≤ 5, and number of hydrogen bond acceptors ≤ 10. Molecules that meet these criteria are more likely to have good oral bioavailability.

[0067] Experimental results show that, compared to the baseline model BaseMol, the molecules generated by PLLMol exhibit superior performance in key metrics: the average docking score decreased from -7.098 to -7.557 (an improvement of 6.5%), the drug-likeness score increased from 0.573 to 0.635 (an improvement of 10.8%), and the proportion of molecules satisfying Lipinski rules increased from 93.82% to 95.28%, while maintaining the stability of the synthetic feasibility score. These experimental results validate the advantages of the protein-based large language model method provided in this embodiment for target-specific molecule generation.

[0068] Corresponding to the molecular generation and optimization methods based on protein big language models described in the above embodiments, this invention also provides an iterative drug design strategy of "generation-screening-optimization".

[0069] Specifically, Figure 3 This is a flowchart illustrating the "generation-screening-optimization" iterative drug design strategy provided in this embodiment of the invention.

[0070] like Figure 3As shown, firstly, in the de novo generation stage, only the target protein pocket information is input to generate molecules, obtaining an initial compound library; secondly, the molecules in the initial compound library are evaluated in multiple dimensions, and one or more seed compounds are selected from the initial compound library based on the evaluation indicators; next, in the seed compound-based optimization stage, the protein pocket information and seed compounds are input together into the model to generate molecules, obtaining an optimized compound library; then, the molecules in the optimized compound library are evaluated in multiple dimensions, and one or more seed compounds are selected from the optimized compound library based on the evaluation indicators; finally, the seed compound-based optimization and evaluation screening process is repeated until potential candidate compounds with high binding affinity, excellent drug properties, and easy synthesis are obtained.

[0071] It should be noted that the multi-dimensional assessment includes binding affinity, drug-likeness, synthetic feasibility, lipid-water partition coefficient, compliance with Lipinski's rule, and pharmacokinetic properties such as absorption, distribution, metabolism, excretion, and toxicity (ADMET).

[0072] Alternatively, in addition to selecting from the generated molecules, seed compounds can also be derived from known active compounds to further improve the quality of the generated molecules in the optimization stage.

[0073] Corresponding to the molecular generation and optimization methods based on protein big language models described in the above embodiments, this invention also provides an integrated AI drug development system. This system encapsulates molecular generation and optimization methods and iterative design strategies into interactive engineering modules, realizing full-process automation from target input to candidate molecule output.

[0074] The system comprises four main functional modules, which can be interconnected via API interfaces using the Python Flask framework and deployed on a high-performance computing cluster (CPU+GPU hybrid architecture):

[0075] 1. Protein Information Acquisition Module

[0076] Functionality implementation:

[0077] It supports directly downloading target protein structure files (such as PDB ID: 1M17) from protein databases (PDB), or importing protein sequences and structures by uploading local files (.pdb, .fasta format) by the user;

[0078] The PyMOL plugin is used to automatically identify protein pockets: these pockets are defined centered on known ligands. Radius range, extract pocket amino acid sequence (e.g., EGFR kinase pocket contains 120 residues such as Lys721 and Cys797);

[0079] For proteins without ligand structures, the AlphaFold3API was used to predict the three-dimensional pocket model. The prediction results were then optimized using Rosetta software. α atom

[0080] Output: Standardized pocket amino acid sequence file (.fasta) and three-dimensional structure file (.pdb), with annotations of key pocket residues (such as catalytic sites and hydrogen bond acceptors).

[0081] 2. Multi-mode generation module

[0082] Generate from scratch mode:

[0083] Input: Protein pocket sequence and 3D structure (preprocessed by the protein information acquisition module);

[0084] Execution flow: The protein encoder (ESM C model) encodes the pocket sequence into a 960-dimensional feature vector h. pocket ;

[0085] The context encoder samples the latent vector z from a standard normal distribution N(0, I), and then compares it with h. pocket The fused feature h is generated by vector concatenation. fusion

[0086] Compound decoder (12-layer GPT architecture) with h fusion As a condition, the autoregressive generation of SMILES strings is performed, and the size of a single generation is configurable (e.g., 10,000-50,000 molecules).

[0087] Application scenario: Discovery of seed compounds targeting novel targets, resulting in molecules with a novelty index (FNI) ≥ 0.7.

[0088] Seed compound optimization mode:

[0089] Input: Protein pocket information + seed compound (SMILES string or molecular structure file, such as SMILES for gefitinib: COc1cc2c(c(c1)O)NC(=O)C(N3CCN(CC3)C)C2=O);

[0090] Execution process:

[0091] Seed compound SMILES and protein pocket information are input into the context encoder (VAE architecture);

[0092] The context encoder outputs the latent vector mean μ and standard deviation σ, which are used to generate z through a reparameterization method, guiding the compound decoder to generate derivatives in the latent vector neighborhood of the seed compound (e.g., σ = 0.5).

[0093] Application scenarios: Structural optimization of known active molecules, such as reducing metabolic rate and improving target selectivity, to generate molecules and seed compounds with a Tanimoto coefficient ≥ 0.6.

[0094] Multi-dimensional evaluation module

[0095] Evaluation parameters and tools:

[0096]

[0097] Execution logic:

[0098] Parallel computing: Utilizing the Dask distributed framework, it supports an evaluation speed of 10,000 molecules / hour;

[0099] Weight configuration: Users can customize the weight of indicators (e.g., 50% weight for the docking score in the early discovery phase, and 60% weight for ADMET in the optimization phase);

[0100] Output report: Generates an interactive HTML report.

[0101] 4. Iterative Optimization Module

[0102] Automated processes:

[0103] Initial screening: The top 100 molecules in the de novo library were screened based on docking score ≤ -7.0 and QED ≥ 0.6, and 5-10 representative seed compounds were selected through cluster analysis;

[0104] Iterative loop:

[0105] Each round of optimization generates 2000 derivatives, and after multi-dimensional evaluation, the top 5% of molecules are retained as seeds for the next round.

[0106] The optimization direction is guided by a reinforcement learning reward function, such as

[0107] R = 10 × (Δ docking score) + 5 × (ΔQED) - 3 × (ΔSAS)

[0108] Termination conditions: No molecule with an improvement of ≥5% in indicators appears in 2 consecutive rounds of optimization, or the preset limit of 5 iterations is reached;

[0109] Manual intervention interface:

[0110] Supports manual replacement of seed compounds (e.g., importing clinically failed drugs for toxicity optimization);

[0111] Adjust assessment thresholds in real time (e.g., tighten ADMET compliance standards);

[0112] Output result:

[0113] Final candidate molecule list (including SMILES, three-dimensional structure and scores for each dimension).

[0114] Finally, let's take the discovery of LRRK2 inhibitors as an example.

[0115] 1. Target preparation and pocket feature extraction

[0116] Target selection: The key target of Parkinson's disease, LRRK2 protein (PDB ID: 7LI4), was selected, and the protein was extracted with reference ligand ATP as the foci. Amino acid residues within the radius serve as protein pockets.

[0117] Feature Encoding: The pocket sequence was encoded using an ESM C protein encoder to generate a 960-dimensional feature vector h. pocket The model captured the synergistic effect of the hydrophobic clusters (Leu1949, Val1950) and the key polar site (Asp2017) within the ATP binding pocket. Attention weights showed that amino acids 1900-1920 contributed a relatively large proportion to the binding energy.

[0118] 2. Two-stage molecular generation process

[0119] 2.1 De novo generation stage (first stage) model configuration: Set gen_vae = False, the context encoder samples the latent vector z from the standard normal distribution N(0, I), and... pocket After splicing, the molecules are input into a 12-layer GPT compound decoder to generate 620 initial molecules (compound library 1).

[0120] Multi-dimensional filtering:

[0121] Binding affinity: The docking fraction was calculated using QuickVina2, with a threshold set to ≤-7.7 kcal / mol (83 molecules retained);

[0122] Drug-like properties: QED≥0.68, SAS≥0.78 (29 molecules retained);

[0123] Blood-brain barrier penetration: 29 molecules were screened out based on a BBBP probability ≥ 0.6 predicted using ADMETlab 3.0.

[0124] Skeleton clustering: Based on the maximum common substructure (MCS) clustering, with a distance threshold set to 0.4, 3 clusters and 3 outliers were obtained. One representative molecule was selected from each cluster (e.g., SMILES for Cluster1: O=C(NCc1ccc(S(=O)(=O)c2cccc(C(F)(F)F)c2)cc1)n1ccnc1), and combined with 2 active molecules reported in the literature, a total of 5 seed compounds were obtained.

[0125] 2.2 Seed Compound Optimization Stage (Second Stage)

[0126] Model configuration: Set gen_vae=True. The context encoder is based on the VAE architecture. The seed compound SMILES and protein pocket information are input into the context encoder to generate the latent vector z=μ+∈·exp(logσ), where ∈~N(0,I).

[0127] Directed generation: The beam search parameter was set to beam=20, generating 8423 optimized molecules (compound library 2).

[0128] Iterative filtering:

[0129] Improved affinity: The average docking score decreased from -7.729 to -8.246 kcal / mol, and the screening threshold tightened to ≤-8.2;

[0130] ADMET optimization: Molecules with BBBP≥0.6 and LogP between 2 and 4 were preferentially retained, ultimately yielding 186 potential lead compounds.

[0131] 3. Comparison of key properties of molecules generated in the two stages

[0132]

[0133] Note: The blood-brain barrier penetration of the molecules was significantly improved during the optimization stage, and the proportion of molecules with BBBP≥0.6 increased by 25.5%, while maintaining stable synthesis difficulty, which verified the effectiveness of the "generation-screening-optimization" strategy in the design of drugs for the central nervous system.

[0134] Model generalization: t-SNE visualization shows that the optimized molecule retains the core skeleton of the seed compound (such as pyridine ring and sulfonamide group) in chemical space, while also expanding the diversity of side chains (such as fluoroalkyl and piperazine ring), demonstrating the protein big language model's ability to accurately model target-ligand interactions.

[0135] in conclusion

[0136] This embodiment successfully constructed a library of LRRK2 inhibitors with high blood-brain barrier penetration by integrating a large protein language model and an iterative optimization strategy, demonstrating the practicality of the patented method in drug design for complex targets. This process can be seamlessly transferred to drug discovery for other central nervous system targets (such as Aβ and α-synuclein), significantly improving the efficiency and clinical translation potential of AI-driven drug development.

[0137] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

[0138] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0139] The above embodiments merely illustrate several implementation methods of the present invention, and their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the protection scope of the present invention. Therefore, the protection scope of this patent should be determined by the appended claims.

[0140] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A molecular generation and optimization method based on a protein-based large language model, characterized in that, include: Obtain the amino acid sequence and three-dimensional structure information of the target protein pocket; The amino acid sequence information is encoded using a protein encoder built based on a pre-trained protein large language model to obtain a 960-dimensional protein pocket feature vector h. pocket The protein big language model mentioned therein is an ESM Cambrian (ESM C) model or other ESM series models, which adopts a 30-layer Transformer architecture and includes rotational position embedding and SwiGLU activation function; Using the context encoder module, context information is encoded to obtain the latent vector z according to the preset molecule generation mode (de novo generation or optimization based on seed compound); The protein pocket feature vector h pocket The latent vector z is fused with the latent vector through vector concatenation or element-wise summation to obtain the fused feature h. pocket ; Using a compound decoder based on a pre-trained GPT model, based on the fusion feature h pocket The new molecular structure in the form of SMILES strings is generated in an autoregressive manner. The compound decoder adopts a 12-layer Transformer architecture with a hidden dimension of 768.

2. The molecular generation and optimization method based on a protein-based large language model according to claim 1, characterized in that, The context encoder module includes: When the molecule generation mode is de novo generation, the latent vector z is sampled from the standard normal distribution N(0, I); When the molecular generation mode is based on seed compound optimization, the input seed compound information (SMILES) and protein pocket information are encoded based on the variational autoencoder (VAE) architecture, and the latent vector z is generated by the reparameterization method = μ + ∈·exp(logσ), where μ and σ are the mean and standard deviation of the encoded data, ∈ ~ N(0, I).

3. The molecular generation and optimization method based on a protein-based large language model according to claim 1, characterized in that, It also includes an iterative drug design strategy of "generation-screening-optimization", the specific steps of which are as follows: De novo generation stage: Only target protein pocket information is input to generate an initial compound library; Multidimensional evaluation: Seed compounds are screened from the initial compound library based on binding affinity (docking score), quantitative estimate of drug-likeness (QED), synthetic accessibility score (SAS), lipid-water partition coefficient (LogP), Lipinski rule compliance, and pharmacokinetic properties such as absorption, distribution, metabolism, excretion, and toxicity (ADMET). Optimization phase: Input the protein pocket information and seed compounds into the model to generate an optimized compound library; Iterative screening: Repeat the optimization and evaluation process until candidate compounds with high binding affinity (docking score ≤ -7.5), drug-like properties (QED ≥ 0.6), and synthetic feasibility (SAS ≥ 0.7) are obtained.

4. The molecular generation and optimization method based on a protein-based large language model according to claim 4, characterized in that, The pharmacokinetic properties of ADMET include absorption (Caco-2 permeability), volume of distribution (Vd), blood-brain barrier penetration (BBBP), and metabolic stability (CL). int ), excretion rate (renal clearance rate) and toxicity (hERG inhibitory activity), etc.

5. The molecular generation and optimization method based on a protein-based large language model according to claim 1, characterized in that, The training data for the protein encoder includes the CrossDocked2020 dataset, which, after RMSD filtering and sequence clustering, yields 100,000 protein-ligand pairs. The training objective is to minimize the following formula: Where x is the protein pocket; y is the reference ligand; D is the training corpus, i.e., the set of (protein pocket, reference ligand) pairs; z in logP(...) is sampled from q(z|x,y); β is a hyperparameter; p(z) represents the standard Gaussian distribution; D kl θ represents the KL divergence; θ is the model parameter.

6. An AI drug discovery system comprising the method of any one of claims 1-6, characterized in that, include: Protein information acquisition module: used to extract the sequence and three-dimensional structure of protein pockets; Multi-mode generation module: supports switching between de novo generation and seed compound optimization modes; Multi-dimensional evaluation module: integrates docking calculation, QED scoring, SAS calculation and ADMET prediction tools; Iterative optimization module: Automatically executes the "generate-screen-optimize" cycle and outputs a candidate compound report.

Images