An all-process drug intelligent research and development method based on a fifth-generation AIDD lighthouse medusa algorithm framework

Abstract

The application relates to the field of AI medicine, and aims to solve the problems of large deviation between prediction and wet experiments and difficulty in penetrating the whole process in the prior art. A dry-wet closed loop iteration engine is constructed based on a lighthouse jellyfish algorithm framework, the framework is combined with a geometric graph neural network G-GNN, a multi-scale protein pocket perception encoder, multi-agent reinforcement learning and a quantum chemistry-deep learning hybrid model, and a multi-task joint learning framework is formed. The process covers target identification, verification, lead compound discovery, optimization and PCC determination. The core lies in that the combination of the pocket perception encoder and the G-GNN enables the R² of target identification prediction and SPR / ITC results to be greater than or equal to 0.85; the multi-agent reinforcement learning and the quantum chemistry-deep learning model are coordinated, a "prediction-synthesis-feedback" closed loop iteration is carried out in the lead optimization stage, the probability of the comprehensive score of a synthesized molecule being greater than or equal to 7.5 / 10 is greater than 85%, and the IVIVC R² of the final PCC is greater than or equal to 0.80. The application realizes intelligent whole-process drug research and development.

Description

Technical Field

[0001] This invention belongs to the field of AI-Driven Drug Design (AIDD) technology, specifically involving a full-process intelligent drug development method and system based on the fifth-generation AIDD Turtle Jellyfish algorithm architecture. It is applicable to the full-process intelligent development of small molecule chemical drugs, including target identification, lead compound discovery, lead compound generation and optimization, and preclinical candidate compound (PCC) determination. Background Technology

[0002] Drug development follows a progressive process: target identification → lead compound discovery → lead compound optimization → preclinical candidate compound → clinical development. The traditional model takes an average of 5-7 years from target identification to clinical completion (PCC), with a success rate of less than 10%. In recent years, artificial intelligence (AI) technology has been gradually introduced into all stages of drug development, forming an AI-driven drug design (AIDD) paradigm characterized by data-driven and computationally intensive approaches. This paradigm can accelerate key stages such as target identification, candidate molecule screening, and pharmacological evaluation.

[0003] However, existing AI-driven drug development technologies still face the following key bottlenecks: First, most AI models only cover a single stage of drug development, with loosely coupled connections between stages, lacking the ability to connect the entire process from target identification to PCC. Second, there is a systematic bias between AI model predictions and wet experimental data. In the target identification and validation stage, the Pearson correlation coefficient between the protein-ligand binding affinity predicted by existing deep learning models (such as virtual screening methods based on AlphaFold structure prediction) and the measured SPR / ITC values ​​is typically only 0.5-0.6. In the lead compound optimization stage, the in vitro-in vivo correlation (IVIVC) between AI-predicted ADMET properties and in vivo measured data is low, resulting in approximately 60% of AI-optimized molecules failing after wet experimental synthesis due to insufficient activity or poor ADMET properties. In the lead compound discovery stage, the hit rate of AI-predicted lead compounds is typically below 30% (wet experimental confirmation rate). Third, the model lacks the ability to continuously learn from wet experiment feedback and cannot form a closed-loop iterative mechanism of "prediction-synthesis-feedback", which means that each round of optimization needs to start from scratch.

[0004] Currently, there is no technical solution in the world that can fully cover the above five stages, have a high degree of statistical consistency between the prediction results of each stage and the wet experimental data, and have the ability to learn in a closed loop. Summary of the Invention

[0005] Technical problems to be solved

[0006] The purpose of this invention is to overcome the above-mentioned defects of the prior art and provide a whole-process intelligent drug development method and system based on the fifth-generation AIDD lighthouse jellyfish algorithm architecture to solve the following technical problems: (1) how to achieve seamless connection from target identification to PCC; (2) how to make the prediction results of each link of the algorithm and the wet experimental measured data achieve a high degree of consistency in statistical significance (R²≥0.80); (3) how to construct a closed-loop iterative mechanism that can continuously learn from the feedback of wet experiments.

[0007] Technical solution

[0008] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0009] Firstly, a full-process intelligent drug development method based on the fifth-generation AIDD (Lighthouse Jellyfish) algorithm architecture includes: S1 target identification and verification step, S2 lead compound discovery step, S3 lead compound generation step, S4 lead compound optimization step, and S5 preclinical candidate compound PCC determination step. (For details of each step, please refer to the claims and the detailed embodiments below.)

[0010] Secondly, an intelligent R&D system implementing the above method includes a target identification and verification module, a lead compound discovery module, a lead compound generation module, a lead compound optimization module, and a PCC determination module.

[0011] Thirdly, a computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the above-described method.

[0012] Description of the invention title

[0013] "Turritopsis dohrnii" is the internal code name for this invention, derived from the biologically significant self-renewal ability of the jellyfish, metaphorically representing the core innovation of this algorithm architecture—achieving continuous self-updating and iterative evolution of the model through a dry-wet closed-loop feedback mechanism. The "Fifth Generation AIDD" in the invention title is a generational division based on domain technology evolution: the first generation was based on classic CADD such as molecular docking and molecular dynamics; the second generation introduced machine learning QSAR models; the third generation introduced deep learning single-point prediction; the fourth generation introduced generative AI and multi-task learning; the core features of the fifth generation are fully automated process, multi-agent reinforcement learning collaborative optimization, quantum chemistry-deep learning hybrid modeling, and dry-wet closed-loop iterative learning. This generational positioning is used to distinguish the essential differences between this technology and existing technologies and should not be construed as limiting the scope of protection of this invention.

[0014] Beneficial effects

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0016] (1) Seamless Integration: For the first time, the five key steps from target identification to PCC are fully integrated under a unified algorithm framework, overcoming the information gaps and efficiency losses caused by the loose coupling of each step in existing schemes. Thanks to this unified framework, the target identification and verification steps can provide higher-precision protein pocket parameters for the lead compound discovery step, directly improving the hit rate of lead compounds; the lead compound generation and optimization steps are seamlessly connected through a multi-agent reinforcement learning framework; the PCC determination step utilizes wet experimental feedback data accumulated in the preceding optimization steps to achieve precision in the in vivo prediction model.

[0017] (2) High prediction accuracy: Through the innovative combination of protein pocket morphology sensing encoder, G-GNN, quantum chemistry-deep learning hybrid model, multi-agent reinforcement learning and attention fusion gating mechanism, the following results are achieved: the determination coefficient R² of the binding affinity of target identification prediction and the measured value of SPR / ITC is ≥0.85; the comprehensive score (covering efficacy, pharmacokinetics and safety) of the output molecule in the lead compound optimization stage is ≥85% after wet experiment verification (≥7.5 / 10); in the PCC determination stage, the R² of the in vitro-in vivo correlation IVIVC of the final output PCC is ≥0.80.

[0018] (3) Closed-loop self-evolution: The dry and wet closed-loop iterative engine and incremental learning module enable the model to autonomously correct its parameters based on feedback from wet experiments after each iteration, driving a steady improvement in subsequent prediction accuracy. The incremental learning strategy employing elastic weight consolidation (EWC) regularization effectively solves the catastrophic forgetting problem, ensuring that the model retains learned knowledge while absorbing new experimental data. Simulation results show that after three rounds of closed-loop iteration, the root mean square error (RMSE) of the predicted values ​​decreases by approximately 40%-50% compared to the initial model. Attached Figure Description

[0019] Figure 1 This is the overall flowchart of the intelligent drug development method of the present invention.

[0020] Figure 2 This is a system module structure diagram.

[0021] Figure 3 This is a structural diagram of a protein pocket morphology sensing encoder.

[0022] Figure 4 It is a message passing graph of a geometric graph neural network (G-GNN).

[0023] Figure 5 It is a multi-agent reinforcement learning framework.

[0024] Figure 6This is the flowchart of the dry and wet closed-loop iterative engine (S4 step).

[0025] Figure 7 This is a structural diagram of a quantum chemistry-deep learning hybrid model.

[0026] Figure 8 It is a multi-omics data fusion and heterogeneous relationship map. Detailed Implementation

[0027] The present invention will be further described in detail below through specific embodiments. However, those skilled in the art will understand that the following embodiments are for illustrative purposes only and should not be considered as limiting the scope of the invention.

[0028] Example 1: Target Identification and Verification (Specific Implementation of Step S1)

[0029] 1.1 Data Acquisition and Preprocessing

[0030] Multi-omics data for the target disease were obtained from public databases such as TCGA (Cancer Genome Atlas), GTEx (Genome-Tissue Expression), UniProt, and ClinVar. These data included: genomic data (somatic mutation information from whole-exome sequencing), transcriptomics data (gene expression levels from RNA-seq), proteomics data (protein abundance quantified by mass spectrometry), and clinical phenotypic data (ICD coding and disease staging information).

[0031] 1.2 Construction of Heterogeneous Relationship Map

[0032] A three-layer heterogeneous relationship map G=(V,E,R) is constructed for gene-protein-disease, where the node set V contains three categories: gene nodes, protein nodes, and disease nodes; the edge set E contains gene-protein coding relationships, protein-protein interaction relationships, gene-disease association relationships, and protein-disease association relationships; and R is the set of relationship types. Pre-trained BioBERT embedding vectors are used to initialize the features of each node.

[0033] 1.3 Geometric Graph Neural Network (G-GNN) Encoding

[0034] This invention employs a unique Geometric Graph Neural Network (G-GNN) to encode heterogeneous relational graphs. Its core is a message-passing mechanism based on geometric attention. Message passing occurs on the heterogeneous relational graph, considering both the biological characteristics of nodes and the semantic relationship types of edges. After message passing, global context-enhanced embeddings of gene and protein nodes are obtained through graph readout operations, used for downstream target ranking.

[0035] 1.4 Protein Pocket Morphological Sensing Encoder

[0036] The protein pocket morphology sensing encoder of this invention is one of the key innovations that distinguishes it from the prior art. The specific construction process is as follows: (a) the protein-ligand cocrystal structure is obtained from the PDB database, the solvent-accessible surface is calculated using DSSP software, and triangular meshing is performed using the Marching Cubes algorithm; (b) at each mesh vertex v_i, differential geometric features (mean curvature H, Gaussian curvature K, shape index SI) and physicochemical features (electrostatic potential φ, obtained by solving the Poisson-Boltzmann equation using APBS software) are calculated, and a multi-channel surface feature map F_i = [H_i, K_i, SI_i, φ_i] is constructed; (c) the SE(3)-equivariant graph neural network is used for message passing, so that the output features are equivariant to the three-dimensional rotation and translation of the protein; (d) supervised pre-training is performed using the PDBbind and sc-PDB datasets.

[0037] 1.5 Target Ranking and Output

[0038] The protein embedding vector output by the G-GNN is concatenated with the pocket features output by the pocket encoder, and then input into the multilayer perceptron (MLP) to predict the druggability score D_score and disease association confidence score Assoc_conf of the target. The final output is a list of candidate targets sorted in descending order of comprehensive score (Comprehensive Score = 0.6×D_score + 0.4×Assoc_conf).

[0039] 1.6 Experimental Verification

[0040] In the target identification test for non-small cell lung cancer (NSCLC), the top 5 candidate targets identified by the method in this embodiment included EGFR, KRAS G12C, ALK, ROS1, and MET, all of which are consistent with clinically known driver genes in NSCLC. The R² of the predicted binding affinity of the EGFR target determined by SPR was 0.87, significantly better than the R² of 0.58 achieved using only AlphaFold structure prediction plus standard virtual screening.

[0041] Example 2: Discovery of Leading Compounds (Specific Implementation of Step S2)

[0042] 2.1 Drugpable pocket identification and dynamic pharmacophore generation

[0043] The druggable pockets of the target protein EGFR (PDB ID: 1M17) were identified using the protein pocket morphology-aware encoder constructed in Example 1. The encoder outputs the coordinates of the pocket center and the pocket morphology feature vector. Based on the identified pockets, a 200 ns molecular dynamics simulation was performed, extracting a protein conformation frame every 10 ns from the trajectory. Pharmacophore features were automatically generated for each conformation frame using Pocket v.3 software, and the pharmacophore features of all frames were merged to obtain a dynamic pharmacophore model.

[0044] 2.2 Flexible Matching Virtual Screening

[0045] The virtual screening library uses a diverse subset of Enamine REAL (approximately 4.8 million compounds). Up to 100 low-energy conformations are generated for each molecule in the library, and each conformation is flexibly matched against a dynamic pharmacophore model. Molecules that pass the matching are scored using a quantum chemistry-deep learning hybrid model.

[0046] 2.3 Quantum Chemistry-Deep Learning Hybrid Model Scoring

[0047] This invention employs a unique quantum chemistry-deep learning hybrid model to score matching molecules. The semi-empirical quantum chemistry method GFN2-xTB is used to calculate the electronic structure descriptors of the molecules (HOMO level, LUMO level, dipole moment, polarizability), which are then concatenated with the hidden layer features of the molecular fingerprint extracted based on a Graph Isomorphism Network. These features are then input into a multi-task prediction head to predict the binding free energy ΔG_bind, ligand efficiency LE, and drug-likeness QED. The overall score is calculated as: Score_hit = -ΔG_bind + 5×LE + 2×QED.

[0048] 2.4 Experimental Verification

[0049] In the screening of lead compounds targeting EGFR, 500 candidate lead compounds were initially selected from a library of 4.8 million compounds. After ranking based on comprehensive scores, the top 50 were selected for wet assay validation (SPR to determine binding affinity). Of the 50 compounds, 23 showed binding activity with Kd < 10 μM (46% hit rate), and 7 out of the top 10 compounds showed a hit rate (70% hit rate), significantly higher than the hit rate of traditional high-throughput screening (typically 0.1%-1%).

[0050] Example 3: Generation of lead compound (specific implementation of step S3)

[0051] 3.1 Multi-agent reinforcement learning framework

[0052] Using the lead compound Molecule_A (pIC50=6.2, MW=385, clogP=3.1) from Example 2 as a template, molecular generation was performed using the multi-agent reinforcement learning framework unique to this invention. The framework comprises three specialized agents: a pharmacodynamic agent (Agent_Pharm), a pharmacokinetic agent (Agent_ADME), and a safety agent (Agent_Tox). Each agent has an independent policy network but shares the same molecular graph encoder.

[0053] 3.2 Attention Fusion Gating Mechanism

[0054] The gradient signals of the three agents are weighted and fused through an attention fusion gating mechanism. For each atom node v on the molecular graph, the gating weights of each agent are first calculated through the MLP_gate network, and the gradient signal G(v) obtained by weighted fusion of the three agents is used to guide the molecular graph mutation operation (atomic replacement, bond addition / deletion, looping / opening).

[0055] 3.3 Generation Results

[0056] After 5000 rounds of generation, a total of 3200 effective molecules were produced. The top 20 were selected as a candidate set of lead compounds based on their multi-objective Pareto front ranking. Wet experiments verified that 4 out of the top 5 lead compounds achieved pIC50 > 7.5, and their metabolic stability (liver microsomal half-life t_{1 / 2}) was greater than 60 min.

[0057] Example 4: Lead compound optimization and dry-wet closed-loop iteration (specific implementation of step S4)

[0058] 4.1 Batch Prediction Using Hybrid Models

[0059] For the 20 lead compound candidate molecules output from Example 3, a quantum chemistry-deep learning hybrid model was used for multi-attribute spectrum prediction, covering eight dimensions (target pIC50, selectivity index, liver microsome t_{1 / 2}, Caco-2Papp, plasma protein binding rate, hERG IC50, AMES mutagenicity classification, and solubility logS). The overall score was calculated as a weighted average of the eight dimensions after normalization. With K=5, the five molecules with the highest overall scores were selected to proceed to the next sub-step.

[0060] 4.2 Wet test verification

[0061] Five candidate molecules were synthesized in parallel (average yield ≥40%), and the following parameters were determined: SPR target affinity (Kd), HepG2 liver microsomal metabolic stability (t_{1 / 2}), apparent permeability coefficient of Caco-2 cell monolayer membrane (Papp), hERG channel binding inhibition rate (patch clamp method), and mutagenicity in the AMES fluctuation assay. The measured data were compiled into a structured training sample.

[0062] 4.3 Incremental Learning

[0063] Elastic Weight Consolidation (EWC) regularization is used to incrementally learn and update the hybrid model. EWC estimates the importance of each parameter to historical tasks through the Fisher information matrix and imposes constraints on important parameters when updating them, thereby absorbing new experimental feedback without forgetting previously learned knowledge.

[0064] 4.4 Iteration Results

[0065] After three rounds of closed-loop iterative optimization, the overall score of the five candidate molecules improved from the initial 5.8-7.2 to 7.5-8.6, all reaching the preset standard of ≥7.5 / 10. The RMSE between the wet experimental values ​​and the model predictions decreased from 0.45 in the first round to 0.22 in the third round, a decrease of approximately 51%.

[0066] Example 5: PCC determination of preclinical candidate compounds (specific implementation of step S5)

[0067] 5.1 PBPK Model Prediction

[0068] For the optimized lead compound Optimized_Lead_A (pIC50=8.3, MW=428, clogP=2.6) finally output from Example 4, a rat PBPK model was constructed using GastroPlus software. The input parameters included: measured Caco-2 Papp (used to predict the human oral absorption fraction Fa), measured plasma protein binding rate f_u, and measured hepatic microsomal clearance rate CL_int (used to predict hepatic clearance rate).

[0069] 5.2 In vivo efficacy verification

[0070] In the NSCLC nude mouse xenograft model (H1975 cell line, EGFR L858R / T790M mutation), after oral administration of Optimized_Lead_A at 30 mg / kg once daily for 14 days, the tumor growth inhibition rate (TGI) reached 78%, with a deviation of only 4.8 percentage points (R²=0.83) from the TGI=82% predicted by the PBPK model.

[0071] 5.3 PCC Determination

[0072] Based on the comprehensive efficacy (TGI≥60%), pharmacokinetics (oral bioavailability F≥30%), safety window (therapeutic index≥10 times), and synthetic scalability (linear synthesis steps≤8 steps, total yield≥15%), Optimized_Lead_A is determined to meet all the criteria, and the output is the PCC.

[0073] Example 6: Comparison and Validation with Other AI Methods

[0074] To objectively evaluate the overall performance of the technical solution of this invention, a systematic comparative verification was conducted. The baseline method adopted a mainstream AI platform in the industry (including a combined pipeline of AlphaFold structure prediction + AutoDock Vina virtual screening + RNN-based molecular generation + traditional machine learning ADMET prediction). Head-to-head comparative tests were performed on three targets: EGFR, BRAF V600E, and BTK (n=5 independent runs for each target), and the results are shown in the table below: Target identification affinity prediction R² 0.87±0.03 0.56±0.08 ＜0.001 Hit rate of lead compounds (wet assay confirmation rate, Kd < 10 μM) 46% 18% ＜0.001 The overall score compliance rate of the lead compound after optimization (≥7.5 / 10) 85%（17 / 20） 40%（8 / 20） ＜0.001 PCC-IVIVC's R² 0.83±0.04 0.41±0.12 ＜0.001

[0075] Example 7: Process Scale-up and Drugability Validation

[0076] The resulting PCC compound underwent process scale-up studies and a comprehensive drug-likeness assessment. The synthetic route consisted of seven linear steps. During the scale-up from gram to hundred-gram levels, the overall yield increased from 22% in the laboratory to 18% in the pilot-scale (a slight decrease, but within an acceptable range). The purity of each intermediate, as determined by HPLC, was ≥98.5%. Accelerated stability testing over six months at 40℃ / 75%RH showed a decrease in chemical purity of <0.5% and no crystal form transformation, demonstrating good molecular stability. In a 28-day repeated-dose toxicity study in rats, no significant adverse reactions were observed at a dose of 30 mg / kg, with a safety window (NOAEL / effective dose) of approximately 15 times.

[0077] Terminology Explanation

[0078] Fifth Generation AIDD: A technological generational division of Artificial Intelligence Driven Drug Design (AIDD). The fifth generation AIDD of this invention is characterized by: (a) fully automated process – seamless connection of the five stages from target identification to PCC under a unified algorithm framework; (b) multi-agent reinforcement learning collaborative optimization; (c) quantum chemistry-deep learning hybrid modeling; and (d) dry and wet closed-loop iterative learning mechanism.

[0079] The Turritopsis dohrnii algorithm is the core algorithm architecture of this invention. Its name is derived from the unique "rejuvenation" ability of the Turritopsis dohrnii jellyfish, which in biology can revert from a mature individual back to a juvenile stage. This metaphorically represents the core innovation of this algorithm architecture—achieving continuous self-updating and iterative evolution of the model through a dry-wet closed-loop feedback mechanism, with the model performance increasing rather than decreasing after each round of wet experiment feedback.

[0080] Dry-wet closed-loop iteration: refers to the closed-loop iterative process of "computation prediction → wet experiment verification → data feedback → model update → re-prediction", which is different from the one-time prediction mode of traditional AI drug development.

[0081] Protein pocket morphology sensing encoder: a deep learning network that encodes the three-dimensional geometric morphology and physicochemical properties of protein surfaces. Its core feature is the use of an equivariant graph neural network to achieve SE(3) equivariance.

[0082] Quantum chemistry-deep learning hybrid model: A hybrid prediction model that fuses electronic structure descriptors calculated by semi-empirical or first-principles quantum chemistry with hidden layer features extracted by deep learning molecular encoders. Its predictive power surpasses that of purely data-driven deep learning models.

[0083] Industrial applicability

[0084] The methods and systems of this invention can be widely applied to small molecule drug development pipelines in pharmaceutical companies, biotechnology companies, and contract research organizations (CROs), and are particularly suitable for the discovery and optimization of first-in-class and best-in-class drugs in therapeutic areas with well-defined targets but a lack of lead compounds, such as oncology, autoimmune diseases, and infectious diseases. This invention can run on ordinary servers (≥4 GPUs, ≥128 GB memory), requires no proprietary hardware, and has good deployability and scalability.

[0085] Although specific embodiments of the present invention have been described in detail, those skilled in the art can make various modifications and substitutions to the details of the technical solutions of the present invention based on all the teachings disclosed, and all such modifications and substitutions are within the scope of protection of the present invention. The full scope of the present invention is given by the appended claims and any equivalents thereof.

Claims

1. A full-process intelligent drug development method based on the fifth-generation AIDD (Lighthouse Jellyfish) algorithm architecture, characterized in that, Includes the following steps: S1 Target Identification and Validation Steps: Acquire multi-omics data of the target disease, including genomics, transcriptomics, proteomics, and clinical phenotypes. Input the multi-omics data into a pre-trained target identification model. The target identification model uses a geometric graph neural network (G-GNN) to encode the gene-protein-disease heterogeneity map and integrates a protein pocket morphology-aware encoder to extract the geometric and physicochemical features of the protein's three-dimensional surface structure. Output a candidate target ranking list and the druggability score and disease association confidence of each candidate target. S2 Lead Compound Discovery Steps: Based on the three-dimensional structure of the target protein determined in S1, the druggable pocket is identified using the protein pocket morphology-sensing encoder. Molecular dynamics simulations based on pharmacophore constraints are performed on the pocket to generate a dynamic pharmacophore model. Based on the dynamic pharmacophore model, flexible matching and screening are performed in a virtual compound library to output a list of lead compounds. The screening process uses a quantum chemistry-deep learning hybrid model to score the matching molecules. The scoring indicators include the predicted binding free energy, ligand efficiency, and drug-likeness. S3 Lead Compound Generation Step: Using the lead compound output from S2 as a template, a multi-agent reinforcement learning framework is used for molecule generation. The multi-agent reinforcement learning framework includes at least three specialized agents—pharmacological agent, pharmacokinetic agent, and safety agent. Each agent adopts a different reward function strategy, and the gradient signals of the multiple agents are weighted and fused through an attention fusion gating mechanism to guide the molecular graph structure variation operation, outputting a candidate set of lead compounds. S4 Lead Compound Optimization Steps: For the lead compound candidate set output from S3, perform a "prediction-synthesis-feedback" dry-wet closed-loop iterative optimization. Each iteration includes: using a quantum chemistry-deep learning hybrid model to predict the multi-attribute spectra of each candidate compound, including target affinity, selectivity, metabolic stability, membrane permeability, hERG inhibition risk, and AMES mutagenicity; selecting Top-K compounds for wet synthesis and in vitro efficacy and ADMET assays based on the prediction results; feeding back the wet experimental data into the incremental learning module of the hybrid model, correcting the model parameters, and performing the next round of prediction until the comprehensive score of the candidate compounds reaches a preset threshold, and outputting the optimized lead compound. S5 Preclinical candidate compound PCC determination steps: For the optimized lead compounds output from S4, the in vivo efficacy and pharmacokinetic behavior are predicted by combining the quantum chemistry-deep learning hybrid model with the physiological pharmacokinetic PBPK model. After verifying the efficacy and preliminary toxicology in animals through the dry-wet closed-loop iterative engine, the prediction model is corrected by feedback. The compound with the highest comprehensive score and that meets the preset PCC criteria is output as the PCC.

2. The method according to claim 1, characterized in that, The construction and training method of the protein pocket morphology-aware encoder includes: Protein-ligand cocrystal structure data were collected from the protein database PDB. Each protein surface was divided into triangular meshes, and the average curvature, Gaussian curvature, shape index, and electrostatic potential at each vertex were calculated to construct a multi-channel surface feature map. An isotropic graphical neural network was used to message-pass the multi-channel surface feature map, ensuring that the output representation is isotropic to three-dimensional rotation and translation. The encoder was pre-trained under supervision using labeled data of known druggable pockets and pocket-ligand binding affinity data. The loss function was a weighted sum of the pocket detection Dice loss and the affinity prediction mean square error loss.

3. The method according to claim 1, characterized in that, When the geometric graph neural network (G-GNN) transmits node features in a heterogeneous relational graph, it employs a message passing mechanism based on geometric attention. For an edge e_{ij} between node i and its neighbor node j, the calculation of the attention weight α_{ij} depends on both the node features and the type of biological relationship represented by the edge: α_{ij} = softmax( W_a · [h_i ∥ h_ j ∥ r_emb(e_{ij})] + b_a ) Where h_i and h_j are the feature vectors of nodes i and j, respectively, r_emb(e_{ij}) is the learnable embedding vector of relation type e_{ij}, ∥ represents the vector concatenation operation, and W_a and b_a are learnable parameters.

4. The method according to claim 1, characterized in that, In the aforementioned multi-agent reinforcement learning framework, the reward function for the pharmacological agent is: R_pharm = w_1 · ΔpIC50 + w_2 · Selectivity_Score - w_3 · Dist_2_Template Where ΔpIC50 is the predicted pIC50 increment of the generated molecule relative to the template molecule, Selectivity_Score is the selectivity score of the target point relative to the anti-target point, Dist_2_Template is the backbone similarity penalty term between the generated molecule and the template molecule, and w_1, w_2, and w_3 are weight coefficients with values ​​of 0.5, 0.3, and 0.2, respectively.

5. The method according to claim 1, characterized in that, The attention fusion gating mechanism is implemented as follows: Let g_k(v) be the gradient signal of the k-th agent to the molecular graph node v. Then the fused gradient signal G(v) is: G(v) = ∑ _k σ( MLP_gate([h_v ∥ g_k(v)]) ) ⊙ g_k(v) Where h_v is the hidden feature of node v, MLP_gate is the gated network, σ is the sigmoid activation function, and ⊙ represents element-wise multiplication.

6. The method according to claim 1, characterized in that, The quantum chemistry-deep learning hybrid model is constructed as follows: Semi-empirical quantum chemistry methods are used to calculate the electronic structure descriptor of molecules, including molecular orbital energy levels, dipole moments, and polarizability. The electronic structure descriptor is used as one of the initial features of molecular graph nodes. After being spliced ​​and fused with the hidden layer features extracted by a deep learning encoder based on molecular fingerprints, it is input into a multi-task prediction head to predict target affinity, metabolic stability, and toxicity endpoints, respectively.

7. The method according to any one of claims 1 to 6, characterized in that, In step S4, each closed-loop iteration includes: (a) Hybrid model batch prediction sub-step: Use a hybrid model to perform multi-attribute spectrum prediction on all molecules in the current candidate set, and select the Top-K molecules according to the comprehensive score, where K is 3~10; (b) Wet experimental verification sub-step: The Top-K molecules were chemically synthesized in parallel, and the in vitro target activity (IC50 / Kd), liver microsomal metabolic stability, Caco-2 membrane permeability, hERG channel binding inhibition rate and AMES mutagenicity were measured. (c) Incremental learning sub-step: The wet experimental data is used as new training samples, and the Elastic Weight Consolidation (EWC) regularization method is used to incrementally learn and update the hybrid model so as to absorb new experimental feedback without forgetting the learned knowledge.

8. A full-process intelligent drug development system based on the fifth-generation AIDD (Lighthouse Jellyfish) algorithm architecture, characterized in that: include: The target identification and verification module is used to perform step S1 as described in claim 1, and has a built-in geometric graph neural network G-GNN and a protein pocket morphology sensing encoder. A lead compound discovery module is used to perform step S2 as described in claim 1, and has a built-in dynamic pharmacophore generation engine and a quantum chemistry-deep learning hybrid scorer. A lead compound generation module is used to perform step S3 as described in claim 1, and incorporates a multi-agent reinforcement learning framework and an attention fusion gating network. The lead compound optimization module is used to execute step S4 as described in claim 1, and has a built-in dry-wet closed-loop iteration engine, incremental learning engine and wet experimental data interface. The PCC determination module is used to perform step S5 as described in claim 1, and has a built-in PBPK simulation engine and an in vivo-in vitro correlation IVIVC analyzer.

9. The system according to claim 8, characterized in that, The dry-wet closed-loop iterative engine includes: a candidate molecule priority sorter, an experimental design optimizer, and a data feedback path that connects to the wet experimental data interface.

10. A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method of any one of claims 1 to 7.

Images