Medicinal preparation AI intelligent optimization method and system

Abstract

The invention relates to the technical field of artificial intelligence and biological medicine crossing, discloses a pharmaceutical preparation AI intelligent optimization method and system, and aims to solve the problems of difficult multi-source data integration, limited feature expression, optimization closed loop deficiency and insufficient model generalization ability in the prior art. The method comprises the following steps: collecting multi-source pharmaceutical data and carrying out normalization processing; constructing a knowledge graph covering a whole chain of medicine, auxiliary materials, process, performance and safety; receiving a design target and a constraint condition input by a user; and outputting a candidate preparation formula by using a graph attention guided sequence generation model. According to the scheme, intelligent generation and dynamic optimization of the preparation formula are achieved, the research and development efficiency and the formula success rate are remarkably improved, the development period is shortened to be within 45 days, the key performance prediction accuracy is improved, the invalid experiment cost is reduced, and deep application of AI in the field of high-end pharmacy is promoted.

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of artificial intelligence and biomedicine, and specifically relates to an AI-based intelligent optimization method and system for drug formulation. Background Technology

[0002] With the rapid development of artificial intelligence technology, drug formulation research and development is gradually evolving towards intelligence and precision. As the core link in drug development, drug formulation directly affects the stability, bioavailability and clinical efficacy of drugs. Its formulation design and process optimization have long relied on experimental trial and error and expert experience, which is time-consuming, costly and has a low success rate.

[0003] Machine learning-based intelligent optimization methods have been introduced into the pharmaceutical field in recent years, aiming to accelerate the formulation screening process and improve R&D efficiency through data-driven approaches. This method typically targets key quality attributes, combining physicochemical performance parameters and process conditions to construct predictive models and guide optimization paths. However, existing technologies still face significant challenges in practical applications: First, their ability to integrate multi-source data is weak; heterogeneous data such as formulation components, process parameters, and in vitro release curves lack a unified characterization mechanism, resulting in incomplete model input information. Second, feature engineering relies on prior human knowledge, making it difficult to automatically uncover nonlinear relationships and potential interaction effects between variables. Third, the optimization process lacks a closed-loop feedback mechanism, leading to a disconnect between model predictions and experimental verification, resulting in low iteration efficiency. Finally, existing systems have poor generalization capabilities; remodeling is required for new drug categories or dosage forms, hindering cross-project knowledge transfer and severely restricting the deep application of AI technology in formulation development.

[0004] Current AI-powered formulation platforms (such as Insilico Medicine and BenchSci) have three main limitations:

[0005] (1) Knowledge fragmentation: Only integrates a single source of literature or patent, lacking the full chain of "drug-excipient-process-performance-safety" association; (2) Generation-evaluation disconnect: The generation model and performance prediction module are separated, and it is impossible to dynamically avoid high-risk combinations during the generation process; (3) Static optimization: The model is fixed after it is launched, and it is impossible to automatically update the confidence of the knowledge graph and the generation strategy based on new experimental data.

[0006] To address the challenges of data integration, limited feature expression, lack of optimization loops, and insufficient model generalization ability in existing AI optimization technologies for drug formulations, there is an urgent need to build an intelligent system with self-learning capabilities that supports multimodal data fusion and dynamic iterative optimization, in order to break through the bottlenecks of traditional R&D models. Summary of the Invention

[0007] The purpose of this invention is to overcome the shortcomings of existing technologies by providing an AI-powered intelligent optimization method and system for pharmaceutical formulations, which can effectively solve the problems mentioned in the background. Currently, in the process of pharmaceutical formulation development, traditional formulation design relies on experience-based trial and error and offline experimental iteration, resulting in long development cycles, high costs, and low success rates. Simultaneously, there are highly nonlinear coupling relationships among multidimensional influencing factors (such as the physicochemical properties of raw materials, excipient ratios, process parameters, and environmental conditions), making it difficult to accurately model using a single model. Existing auxiliary design tools are mostly limited to static rule matching or shallow statistical analysis, lacking a deep understanding of complex pharmaceutical knowledge systems and dynamic reasoning mechanisms, leading to poor adaptability, weak interpretability, and difficulty in meeting personalized treatment needs in recommended solutions. This invention constructs a knowledge-enhanced artificial intelligence architecture that integrates multi-source heterogeneous pharmaceutical data, achieving end-to-end intelligent generation and optimization from raw data to formulation, significantly improving the efficiency, stability, and clinical applicability of pharmaceutical formulation development.

[0008] To achieve the above objectives, the present invention provides the following technical solution: In one aspect, a pharmaceutical formulation AI intelligent optimization system, comprising the following components:

[0009] The data integration and preprocessing module is used to uniformly access and standardize the processing of multi-source pharmaceutical data from literature databases, patent databases, experimental record systems, drug regulatory platforms, and clinical feedback systems. The multi-source pharmaceutical data covers the structure information, solubility, crystal form, pKa value, stability parameters, excipient compatibility data, known formulations, production process parameters, quality control indicators, and adverse reaction reports of active pharmaceutical ingredients (APIs).

[0010] The knowledge graph construction module is used to automatically extract entities, attributes and semantic relationships based on preprocessed structured and semi-structured data, and construct a dedicated knowledge graph for the pharmaceutical formulation field covering the entire chain of "drug-excipient-process-performance-safety". The knowledge graph uses a graph neural network embeddable representation to store node vectors and edge weights, and supports dynamic updates and incremental learning.

[0011] The formulation generation engine module receives user input regarding the target formulation type, route of administration, release characteristics, target specifications, and constraints. It combines the current knowledge graph state with a sequence generation model guided by graph attention to output a set of candidate formulations. Each candidate formulation includes the dosage of the active ingredient, the type of excipient, the mass ratio range, the suggested preparation method, and key control points.

[0012] The performance prediction and screening module is used to perform multi-dimensional virtual evaluation of each item in the candidate formulation set, including dissolution behavior simulation, physical stability prediction, bioavailability estimation, production feasibility scoring and potential interaction warning, and output a comprehensive score ranking list.

[0013] The closed-loop feedback optimization module is used to receive external validation result data, including laboratory test data, pilot batch quality reports or clinical trial feedback, analyze the deviation between actual performance and predicted results, and drive the joint reverse update of knowledge graph and generation model parameters to form a continuously evolving intelligent optimization closed loop.

[0014] Preferably, when performing data cleaning, the data integration and preprocessing module adopts a method combining rule-based filtering and outlier detection to mark and verify records whose auxiliary material usage ratio deviates from the industry standard average by more than 3 times the standard deviation, and to initiate a cross-database association and completion process for entries with missing key parameters, so as to ensure that the integrity of the input data is not less than 98%.

[0015] Furthermore, in the entity linking stage, the knowledge graph construction module introduces a chemical structure fingerprint hash index mechanism to perform topological encoding on the SMILES string, achieving accurate matching of the same compound in different databases, with an entity normalization accuracy of over 99.2%.

[0016] In addition, during the relation extraction process, the knowledge graph construction module adopts a deep learning model combining bidirectional long short-term memory network and conditional random field (BiLSTM-CRF) to automatically identify implicit semantic relations such as "API-excipient incompatibility" and "high temperature leads to degradation" from scientific literature abstracts. The sensitivity of new relation discovery is 42 percentage points better than the traditional keyword matching method.

[0017] Preferably, the sequence generation model used by the recipe generation engine module is based on an improved Transformer architecture, with its decoder part embedded in a graph attention layer, which can dynamically focus on the subgraph region in the knowledge graph that is most relevant to the current context in each generation operation. The number of attention heads is set to 8, and the maximum sequence length is limited to 128 tokens.

[0018] Furthermore, the formulation generation engine module implements a hard constraint injection mechanism during the generation process, which blocks excipient combination options that violate pharmacopoeia regulations at the vocabulary level, and forces the user to meet special kinetic requirements such as zero-order release or pulse release through gating logic, with a constraint satisfaction rate of up to 100%.

[0019] Furthermore, the dissolution behavior simulation unit in the performance prediction and screening module employs a Noyes-Whitney equation discretization and deduction framework based on physical mechanisms. It combines actual dissolution curves of historically similar formulations as prior distributions, performs Bayesian correction, and outputs a predicted curve with a prediction time resolution better than 5 minutes. R0 2 The average goodness of fit reached 0.91;

[0020] Preferably, the bioavailability estimation unit in the performance prediction and screening module integrates the BCS classification system judgment logic and the PBPK model simplification algorithm, quickly classifies the APIs according to their permeability and solubility characteristics, and maps them to a preset absorption efficiency range, with the estimation error controlled within ±15%.

[0021] Furthermore, after receiving new experimental verification data, the closed-loop feedback optimization module triggers a two-stage update strategy: the first stage performs a knowledge graph triple confidence reassessment and lowers the weights of relevant edges in the predicted failed paths; the second stage uses the gradient-pruned AdamW optimizer to fine-tune the generated model, with a learning rate set to 2e-5, a batch size of 16, and a single iteration time not exceeding 8 seconds.

[0022] In addition, the system also includes a human-machine collaborative decision-making interface module, which is used to present the Top-5 candidate formulas in the form of a visual comparison matrix, listing the scores of each key indicator, risk warnings and knowledge traceability paths, supporting expert manual intervention to adjust weights or exclude specific solutions, and finally confirming the command to be sent back to the formula generation engine to complete the locking.

[0023] Preferably, the human-machine collaborative decision-making interface module has a built-in interpretability analysis component, which can trace the knowledge support link of each recommendation reason in reverse, show the complete reasoning path from the original literature evidence to the final conclusion, and the traceability response delay is less than 300 milliseconds;

[0024] On the other hand, a drug formulation AI intelligent optimization method includes the following specific steps:

[0025] Step S110: Collect multi-source pharmaceutical data and perform format normalization and quality verification. The multi-source pharmaceutical data includes physicochemical parameters of active pharmaceutical ingredients, database of excipient functional properties, existing marketed formulation information, pharmaceutical process specification documents, and clinical efficacy and safety monitoring data.

[0026] Step S120: Construct a drug formulation knowledge graph based on the normalized data. The knowledge graph contains entity nodes and semantic relationship edges. The entity nodes at least cover drugs, excipients, equipment, processes, dosage forms, performance indicators, and indication categories.

[0027] Step S130: Receive the formulation design goals and boundary conditions input by the user. The design goals include administration method, release mode, dose intensity and storage stability requirements. The boundary conditions include a list of prohibited ingredients, cost limits and production equipment compatibility restrictions.

[0028] Step S140: The knowledge graph context is parsed using a graph-enhanced sequence generation model to generate candidate formulation sequences that meet the constraints. The generation process adopts an autoregressive sampling strategy, outputting one formulation component and its dosage range each time.

[0029] Step S150: Perform multidimensional virtual performance evaluation on the generated candidate formulations, calculate their comprehensive scores in terms of dissolution, stability, manufacturability and safety, and sort them in descending order of total score to form a recommendation list;

[0030] Step S160: Output the recommended list and receive confirmation instructions or correction feedback through the human-computer interaction interface. Store the final selected formula in the enterprise knowledge base and simultaneously start the closed-loop feedback optimization process to update the model parameters.

[0031] Preferably, in step S110, when performing deduplication on the collected data, a clustering algorithm based on Jaccard similarity is used to merge formulas with a difference of less than 5% in the proportion of auxiliary materials and the same process steps as duplicates, and retain the latest version data.

[0032] Furthermore, in step S120, when constructing the knowledge graph, an initial confidence score is assigned to each semantic relation edge, with a value ranging from 0.6 to 1.0, determined by weighting based on the authority of the data source, and subsequently dynamically adjusted through closed-loop feedback;

[0033] In addition, in step S130, the release mode input by the user includes sustained release, controlled release, enteric coating, immediate release or targeted release, and the system automatically converts it into the corresponding mathematical representation template and embeds it into the prompt space of the generated model.

[0034] Preferably, in step S140, the graph-enhanced sequence generation model introduces a Monte Carlo Tree Search (MCTS) strategy in the decoding stage to explore potentially high-value recipe paths that are not covered by high-frequency training samples. The exploration coefficient is set to 1.4, which effectively improves the probability of discovering innovative recipes.

[0035] Furthermore, in step S150, the multidimensional virtual performance evaluation adopts a weighted summation model. The weights of each dimension are pre-configured by the user or automatically generated by the system based on the clustering of historical successful cases. The default weight allocation is as follows: dissolution performance 30%, physical stability 25%, production feasibility 20%, safety 15%, and cost percentage 10%.

[0036] In addition, in step S160, the closed-loop feedback optimization process performs A / B testing verification before the model is updated, and the old and new versions of the model are evaluated against each other on the same set of blind test tasks. The new version is only allowed to be deployed online when the Top-3 hit rate is improved by more than 8 percentage points.

[0037] Compared with the prior art, the present invention has the following beneficial effects:

[0038] Full-chain knowledge graph: The first construction of a five-dimensional entity network covering everything from API physicochemical properties to clinical safety;

[0039] Graph-sequence tightly coupled generation: In each step of Transformer decoding, the most relevant subgraph is dynamically retrieved through graph attention to achieve "generation while reasoning";

[0040] Closed-loop dynamic optimization: An A / B testing verification mechanism is introduced, and the knowledge graph and the generative model are jointly updated only when the Top-3 hit rate of the new model increases by more than 8%, so as to ensure evolutionary stability.

[0041] By constructing a knowledge graph specifically for pharmaceutical preparations and deeply integrating it with a deep generation model, a systematic expression and dynamic utilization of complex pharmaceutical knowledge systems were achieved, overcoming the problems of knowledge fragmentation and delayed updates in traditional methods, and increasing the knowledge utilization rate to over 90%.

[0042] By employing a graph attention-guided sequence generation mechanism, highly feasible novel formulations can be generated while ensuring regulatory compliance. The development cycle has been shortened from an average of 18 months to less than 45 days, and the first-time approval rate of the formulations has been increased to 76%.

[0043] By introducing a multi-dimensional virtual performance prediction and closed-loop feedback optimization mechanism, the system is equipped with self-evolution capabilities. After running continuously for 6 months, the prediction accuracy of key performance indicators has been improved by a cumulative 23%, significantly reducing the investment in ineffective experiments.

[0044] It integrates human-machine collaborative decision support functions, provides a complete recommendation basis traceability path, enhances the transparency and credibility of AI decision-making, and achieves an expert adoption rate of 89%, effectively promoting the application of artificial intelligence technology in the high-end pharmaceutical field. Attached Figure Description

[0045] Figure 1 This is a schematic diagram of the overall technical architecture of the AI-based intelligent optimization method and system for drug formulation proposed in this invention. Detailed Implementation

[0046] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0047] Example 1

[0048] Please refer to Figure 1This embodiment uses the development of an oral sustained-release tablet for the treatment of type 2 diabetes as a specific application scenario to fully demonstrate the entire process of the AI-powered intelligent optimization system and method for drug formulation described in this invention. The sustained-release tablet must meet the technical requirements of once-daily administration, maintaining stable blood drug concentration, and avoiding postprandial hyperglycemia, and must comply with the Chinese Pharmacopoeia's quality standards for oral solid dosage forms, including dissolution rate, content uniformity, and stability. The user-input design objectives include: oral administration, zero-order sustained-release release, a dose intensity of 50 mg glimepiride, storage conditions of room temperature and protection from light, cost control below RMB 80 per 1000 tablets, and prohibition of lactose as an excipient (due to lactose intolerance in some patients). The system initiates the intelligent optimization process based on the above constraints.

[0049] The first step is S110, which involves the acquisition and preprocessing of multi-source pharmaceutical data. The system synchronously connects to real-time data streams from PubChem, DrugBank, the FDA Orange Book, the China Center for Drug Evaluation (CDE) database, the European Pharmacopoeia (EP), pharmaceutical companies' internal LIMS laboratory management systems, and the ADR-Net clinical adverse reaction monitoring platform via standardized API interfaces. The collected data items include, but are not limited to: the molecular structure SMILES string of the active pharmaceutical ingredient glimepiride (CCS(=O)(=O)N(Cc1ccc(cc1)C(C)C)c2ccc(cc2)C(=O)Nc3ccccc3), logP value (3.82), pKa (5.4), crystal form distribution (mainly type I), thermodynamic solubility (0.012 mg / mL), kinetic dissolution rate (t50% = 18 min in pH 6.8 medium), viscosity grades (K4M, K15M, K100M) of hydroxypropyl methylcellulose (HPMC) from the excipient functional properties database, film-forming parameters of povidone K30, compression molding index of microcrystalline cellulose PH101, and existing marketed sustained-release formulations such as Glucotrol. The publicly available formulation information of XL (including HPMC backbone material content of 35%-42%), the wet granulation parameter range in the production process specification document (stirring speed 80-120rpm, granulation time 3-5min), drying temperature 50-60℃, and efficacy and safety monitoring data from real-world studies, such as the incidence of gastrointestinal bloating caused by formulations containing lactose excipients being 17.3%. After all raw data entered the data integration and preprocessing module, a unified format normalization operation was performed: dosage units from different sources were converted to the International System of Units (SI) (mg / g / L), pH values ​​were uniformly calibrated using the three-point calibration method, process parameter timestamps were accurate to the second, and text description fields were standardized and mapped using the BERT-Chem model. Following quality verification, a combination of rule-based filtering and outlier detection was used to identify a formula in a third-party database that showed "HPMC usage at 95%". This deviation from the industry average (average usage 40.2%, standard deviation ±8.7%) exceeded three standard deviations (threshold >66.3%). The system automatically flagged this entry and triggered a cross-database completion process, cross-validating with DrugBank and the FDA Orange Book to confirm the correct value was 45%, and the correction was completed. For missing key parameters, such as the water absorption swelling rate of a novel disintegrant, the system activated a knowledge graph-assisted reasoning mechanism to retrieve historical data of similar excipients with similar chemical structures (such as croscarmellose sodium), and used K-nearest neighbor interpolation to fill the gaps, ensuring overall data integrity reached 98.7%.The deduplication stage uses a clustering algorithm based on Jaccard similarity to calculate the overlap of component sets between each pair of formulations. The threshold is set to consider duplicates as having an excipient ratio difference of less than 5% and identical core process steps. A total of 12 highly similar entries are merged, and the latest version data is retained for the next stage.

[0050] Next, step S120 is executed to construct a knowledge graph specific to drug formulations. Preprocessed structured and semi-structured data are fed into the knowledge graph construction module. In the entity extraction stage, the system uses a Named Entity Recognition (NER) model to extract key concepts such as "glimepiride," "HPMC K4M," "wet granulation," "delayed gastric emptying," and "crystallization inhibition" from literature abstracts. Normalization is then performed using a chemical structure fingerprint hash index mechanism: the SMILES string is topologically encoded to generate a fixed-length Morgan fingerprint (radius = 2, length = 2048), and an inverted index table is established to achieve precise matching of the same compound in different databases, with an entity normalization accuracy of 99.4%. Semantic relation extraction employs a deep learning model combining a bidirectional long short-term memory network and a conditional random field (BiLSTM-CRF), trained on 120,000 pharmaceutical literature abstracts indexed in PubMed. It can identify implicit triplet relationships such as "glimepiride—and—magnesium stearate—existence—decreased compatibility," "high-temperature drying—leads to—glimepiride—occurrence—oxidative degradation," and "HPMC concentration >30%—can—achieve—zero-order release." The sensitivity for new relation discovery reaches 89.7%, a 43.2 percentage point improvement over traditional keyword matching. Each extracted relation edge is assigned an initial confidence score ranging from 0.6 to 1.0, weighted according to the authority of the data source: 1.0 for data from FDA approval documents, 0.9 for peer-reviewed journal articles, 0.8 for patent documents, 0.7 for internal non-validation data, and 0.6 for social media reports. The final knowledge graph contains a total of 1,247,358 nodes, including 86,421 drug entity nodes, 12,893 excipient nodes, 5,672 process nodes, 3,210 equipment nodes, 1,890 performance index nodes, and 1,543 indication category nodes; it has a total of 3,892,105 edges, covering various semantic types such as "drug-excipient compatibility," "process-parameter dependence," "dosage form-release behavior mapping," "API-BCS classification," and "excipient-functional attribute association." The graph is stored in the Neo4j graph database, and node vectors are embedded into a 128-dimensional continuous space using the TransE algorithm, supporting fast subgraph retrieval and similarity calculation with a response latency of less than 150 milliseconds.

[0051] Knowledge graph node / edge definition: Explicit definition:

[0052] • Node type:

[0053] - Drug (API): Contains SMILES, logP, pKa, BCS classification, and crystal form

[0054] -Excipients: including functional category (filler / disintegrant / sustaining release material), viscosity grade, and compatibility label.

[0055] -Processes: Wet granulation, dry tableting, fluidized bed coating, etc.

[0056] -Performance indicators: f2 similarity factor, dissolution t50%, content uniformity RSD, accelerated stability Δimpurity

[0057] • Edge type:

[0058] - "API-Excipients": Compatibility (positive / negative), Interaction Strength (0~1)

[0059] - "Additives-Process": Suitability score (e.g., HPMC is suitable for wet granulation)

[0060] - "Process-Performance": Influence function (e.g., drying temperature ↑ → moisture ↓, degradation ↑).

[0061] Normalization algorithm; structured data: unified units (mg / g / L), pH calibration (three-point buffer calibration).

[0062] • Semi-structured text: Using the BERT-Chem terminology normalization model (fine-tuned on the PubChem+Patent corpus), “hydroxypropyl methylcellulose K4M”, “HPMC K4M”, and “Methocel K4M” are mapped to the same standard entity.

[0063] • Missing value imputation: For missing excipient functional parameters, K-nearest neighbor interpolation (K=5, based on Morgan fingerprint similarity) is used.

[0064] In step S130, the system receives the user's input of the formulation design objectives and boundary conditions. The human-machine collaborative decision-making interface module provides a graphical input interface. The user sequentially selects "oral tablets" as the dosage form, "sustained release" as the release mode, "50mg" as the dose intensity, and "room temperature storage" as the stability requirement, and checks three constraints: "lactose prohibited," "cost cap of 80 yuan / 1000 tablets," and "only using existing GZP-600 compatible tablet press molds." The system automatically converts these natural language instructions into a structured parameter set. Specifically, the release mode "sustained release" is parsed by the system into a mathematical representation template: the target dissolution curve must satisfy an f2 similarity factor ≥ 50, and the cumulative dissolution in pH 6.8 buffer ≤ 30% after 2 hours and ≥ 80% after 8 hours. This template is encoded as a soft constraint vector in the prompt space of the generative model, with a dimension of 64, and injected into the initial state of the Transformer decoder through a differentiable mapping function. Production equipment compatibility limitations were translated into physical constraints: the die punch diameter must be 12mm, the tablet weight range is limited to 300-400mg, and the tablet thickness does not exceed 5.2mm. All boundary conditions constitute a Boolean logic expression for hard screening in the subsequent generation process.

[0065] Execute step S140 to start the recipe generation engine module. This module adopts a graph-enhanced sequence generation model based on an improved Transformer architecture. The decoder's l-th layer embeds a graph attention sub-layer, whose input is a local subgraph G of the knowledge graph. sub ={(v i ,r ij ,v j Attention calculation formula:

[0066]

[0067] Where, α ij The attention weight of node i to node j, [Wh i ||Wh i The splicing operation, eakyReLU(x) is the leakage correction linear unit activation function, exp(·) is the exponential function, and ∑ k The summation of the denominator of exp(·) over all possible neighbor nodes k achieves softmax normalization. i h is a node vector. j It is also a node vector. For learnable weights, For the attention vector, the graph attention output and the self-attention output are fused through gating: H out =σ(G)·H graph+(1 - σ(G))·H self , H out The finally output hidden state, H graph The output of the graph attention sublayer, H self The output of the self-attention sublayer, the G gating signal, the σ(G) Sigmoid function. Its decoder part embeds a graph attention layer (Graph Attention Layer), with a total of 8 attention heads set, and the maximum sequence length is limited to 128 tokens. The generation process adopts an autoregressive sampling strategy, outputting one formulation component and its dosage range each time. When the model is initialized, the latest version of the pre-trained weights is loaded. The input is the local subgraph (with a radius of 3 hops) related to "glimepiride" in the current knowledge graph, and the design target vector encoded in step S130. At each generation step t, the model not only focuses on the previously generated token sequence X<t, but also dynamically focuses on the most relevant subset of nodes in the knowledge graph through the graph attention mechanism. For example, when generating the first excipient, the attention weights are concentrated on the polymer material nodes strongly related to the "sustained release" function, such as HPMC, ethyl cellulose, carbomer, etc. Among them, the HPMC K series obtains the highest attention score due to having the strongest historical evidence chain of "zero-order release". A hard constraint injection mechanism is implemented during the generation process: all candidate tokens containing "lactose" are masked at the vocabulary level; the release kinetics requirements are forced to be met through gating logic - when it is detected that the current generation path may lead to the risk of immediate release, the proportion of the sustained release matrix material is automatically increased. In addition, the Monte Carlo tree search (MCTS) strategy is introduced for exploration, and the exploration coefficient is set to 1.4, opening up potential high-value paths in addition to the regular greedy sampling. For example, trying to combine HPMC with a small amount of sodium alginate to form a double-network controlled release system. Although the frequency of this scheme in historical data is less than 2%, it is theoretically predicted to have better pH-responsive characteristics. After 23 rounds of iterative sampling, the system generates a set of 200 candidate formulations, and each formulation consists of the following fields: active ingredient name, dosage, list of excipient types (in order), mass ratio range of each excipient, recommended preparation method (such as wet granulation - drying - sizing - tabletting), key control points (such as upper limit of drying temperature, lower limit of mixing time).

[0068] Execute step S150 to perform multi-dimensional virtual performance evaluation on the candidate formulations. The performance prediction and screening module runs five evaluation units in parallel. The dissolution behavior simulation unit adopts a discretized推演 framework based on the physical mechanism of the Noyes-Whitney equation:

[0069]

[0070] Where D is the diffusion coefficient (determined by looking up API logP and medium pH), A is the particle surface area (calculated from the statistical moments of particle size distribution), h is the diffusion layer thickness (empirically taken as 80 μm), and C... s For saturated solubility, C b The concentration of the main body fluid is represented by the equation. The equation is numerically integrated over a time axis with 5-minute intervals. Using historical dissolution curves of similar formulations (such as glipizide sustained-release tablets) as prior distributions, a Bayesian correction method is employed to adjust model parameters, ultimately outputting the predicted dissolution curve. For candidate formulation #17 (containing 38% HPMC K15M, 45% microcrystalline cellulose pH101, and 1% magnesium stearate), it is predicted to dissolve 28.3% in pH 6.8 medium after 2 hours and 82.1% after 8 hours, with an f² value of 53.7, meeting the sustained-release requirements. The physical stability prediction unit extrapolates accelerated test results based on the Arrhenius equation, calculating the degree of aging after 3 months at 40℃ / 75%RH. A stability score is given by comprehensively considering three indicators: crystal form transformation tendency, oxidative degradation rate, and hygroscopic weight gain. The production feasibility scoring unit interfaces with the MES (Manufacturing Execution System) to verify whether each material in the formulation is in the current GMP workshop inventory list, check whether the recommended process parameters are within the allowable range of the GZP-600 tablet press operating procedures, and assess whether the mixing uniformity RSD is <5%. The safety warning unit scans for known interactions between excipient combinations. For example, if a candidate formulation uses sodium dodecyl sulfate and a protein API simultaneously, it immediately issues a "may cause denaturation" warning. The bioavailability estimation unit integrates the BCS classification system's judgment logic with a simplified PBPK model algorithm: Glimepiride belongs to BCS II (low solubility, high permeability), and the system maps it to a preset absorption efficiency range of 60%-75%. Combining the predicted dissolution curve to correct for the first-pass effect, the estimated oral bioavailability is approximately 68.4% ± 10.2%. After summing the evaluation results, a weighted summation model is used to calculate the overall score. The default weight allocation is: dissolution performance 30%, physical stability 25%, production feasibility 20%, safety 15%, and cost 10%. Cost calculations are based on the current bulk purchase price list, accurate to the price per gram of raw material, and finally generate a Top-5 recommended list, with recipe #17 ranking first with a total score of 91.6.

[0071] Step S160 is executed, outputting a recommendation list and receiving confirmation instructions via a human-computer interaction interface. The Top-5 candidate formulations are presented in a visual comparison matrix, with the horizontal axis representing five key indicators and the vertical axis representing each candidate scheme. A radar chart and heat map are overlaid to visually reflect advantages and disadvantages. Although formulation #17 scored low in cost (78 points), its scores in the other four categories all exceeded 90 points, making it the best overall. Experts can manually adjust the weights through the interface, for example, increasing the weight of "production feasibility" to 30%, and it still maintains its top position after re-sorting. Experts decide to adopt this scheme and click the "Confirm Lock" button. The system stores the final selected formulation in the enterprise knowledge base, generating a unique identifier FID-2024-GLU-0017, and simultaneously initiates a closed-loop feedback optimization process. This process performs A / B testing verification before model updates: 20 sustained-release tablet R&D tasks completed in the past 6 months are extracted as a blind test set. The old and new versions of the system independently generate Top-3 recommended schemes, which are then submitted to a third-party laboratory for actual preparation and testing. Evaluation results show that the Top-3 hit rate (i.e., at least one recommended solution passed the initial verification) of the new system increased from 65% to 74.2%, an increase of 9.2 percentage points, exceeding the threshold of 8 percentage points, thus allowing for deployment. The closed-loop feedback optimization module then triggered a two-stage update strategy: In the first stage, the confidence of the knowledge graph triplet was re-evaluated. To address the bias of underestimating the ability of the high viscosity grade of HPMC K15M to maintain the plateau period in the later stage of dissolution, the confidence of the "HPMC viscosity - and - release time" relationship edge was lowered by 0.05 units. In the second stage, the AdamW optimizer with gradient pruning was used to fine-tune the generated model. The learning rate was set to 2e to the power of negative 5, the batch size was 16, and the input was the newly added experimental verification data pair (predicted formulation vs. measured performance). Each iteration took 7.8 seconds. After completing one round of parameter updates, a new model snapshot V2.1.3 was saved.

[0072] Throughout the entire process, the interpretability analysis component built into the human-machine collaborative decision-making interface module ran continuously. When an expert asked "Why recommend HPMC instead of ethyl cellulose?", the system traced the reasoning path backward within 300 milliseconds, demonstrating the complete knowledge support chain: the original evidence came from a 2021 paper in the *Journal of Controlled Release*, which stated that "HPMC exhibits more stable swelling behavior under changes in gastrointestinal pH." This conclusion was extracted by the BiLSTM-CRF model as a "HPMC—superior—EC—for—oral sustained release" relationship with a confidence level of 0.92, receiving the highest attention weight in this generation. The traceability path was presented in the form of a directed graph, including the literature source, experimental conditions, sample size, and statistical significance indicators, greatly enhancing the credibility of the AI ​​decision.

[0073] Example 2

[0074] This embodiment focuses on the formulation optimization of injectable nanosuspensions, highlighting the technical advantages of this invention in handling highly nonlinear coupled variables and extreme constraints. The goal is to develop a paclitaxel nanoinjection for targeted tumor therapy, requiring a particle size distribution concentrated in the 100-150 nm range, an absolute Zeta potential >30 mV to ensure colloidal stability, an encapsulation efficiency >95%, sterility and pyrogen-free properties, and the ability to achieve tumor tissue enrichment in mice that is more than four times higher than that of free drug. Compared to oral solid dosage forms, nanoformulations involve more synergistic regulation of physicochemical parameters, and the preparation process is more sensitive; even minor changes can lead to product failure.

[0075] The core difference in this embodiment lies in the different performance prediction modeling approaches employed. In Embodiment 1, the dissolution behavior simulation is based on the classical Noyes-Whitney equation, which is suitable for macroscopic dissolution processes. However, in this embodiment, facing complex fluid dynamics and interface phenomena at the nanoscale, the system switches to a machine learning-driven hybrid modeling paradigm. It no longer relies on a single physical equation but integrates molecular dynamics simulation features with deep neural networks for end-to-end prediction.

[0076] In step S110, in addition to routine pharmaceutical data, the system additionally accesses the Materials Project database, the NIST nanomaterials characterization library, the cryo-electron microscopy (Cryo-EM) image archiving system, and real-time sensor data from the microfluidic chip experimental platform. Key parameters acquired include: Henry's constant for paclitaxel in different solvents, the critical micelle concentration (CMC) of the surfactant Tween 80 and lecithin, the glass transition temperature (Tg) of PLGA polymers, video frames of the Brownian motion trajectories of nanoparticles, the raw correlation function curve of dynamic light scattering (DLS), and the time series of zeta potential measurements. The data preprocessing module addresses the noise problem unique to nanoscale data by introducing a wavelet denoising algorithm (Daubechies db4 wavelet basis, decomposition level 5) to process the DLS signal, improving particle size inversion accuracy; and automatically identifying particle contours in the Cryo-EM images using a U-Net segmentation network to extract equivalent diameters and morphological factors. During the deduplication process, because nano-formulations are highly sensitive to the preparation path, even if the excipient ratio is the same, different microjet pressures or homogenization times are considered different formulations. Therefore, the Jaccard similarity threshold is tightened to only merge formulations if the difference in excipients is <2% and the process parameters are completely identical.

[0077] In step S120, the knowledge graph construction module expands by adding a sub-graph branch for "nanomaterials-surface modification-targeting ligands," introducing specific relationships such as "PEGylation degree—impact—circulatory half-life," "particle size—determines—EPR effect intensity," and "surface charge—correlation—protein crown formation." During the fine-tuning phase, the relationship extraction model incorporates full-text training sets from journals such as Nature Nanotechnology to enhance its understanding of the semantics of the "nano-biological interface." Specifically, a higher initial confidence level (0.95) is assigned to negative relationships such as "ultrasonic dispersion time—leads to—PLGA degradation," as violations of these relationships directly result in product failure.

[0078] In step S130, the user-inputted "tumor-targeted release" is parsed by the system into multiple mathematical constraints: target organ uptake rate ≥ 4 × control group, blood clearance half-life extended to more than 6 hours, and liver and spleen accumulation ratio < 15%. These indicators cannot be directly mapped to simple formulas, so the system encodes them into a multi-objective optimization vector with a dimension of 128, containing the expected pharmacokinetic parameter distribution.

[0079] In step S140, the formulation generation engine still uses the graph-enhanced Transformer model, but a dedicated sub-decoder is enabled in the "surface modification layer" generation stage, forcing the output of structured fragments containing target ligands (such as folic acid, RGD peptides) and their coupling chemical bond types. The MCTS exploration coefficient is increased to 1.8, encouraging the exploration of unconventional combinations, such as using novel amphiphilic Janus particles as stabilizers, a scheme that has only been seen in three cutting-edge papers in historical data.

[0080] In step S150, the performance prediction and screening module adopts a novel modeling paradigm. The dissolution behavior simulation unit is replaced by a nanoparticle dispersion stability prediction unit, whose internal structure is as follows: First, short-period simulations (50 ns) are performed using LAMMPS molecular dynamics software to extract three features: interparticle van der Waals forces, electrostatic repulsion energy, and solvation free energy. Then, these high-dimensional features are input into a pre-trained graph convolutional network (GCN), which models the nanoparticle cluster as a spatial graph. The node features are the size, charge, and surface modification state of each particle, while the edge features are Euclidean distance and interaction potential energy, outputting an overall aggregation trend score. Simultaneously, another parallel 1D-CNN network processes the original DLS correlation function curve to identify anomalous fluctuation patterns in the polydispersity index (PDI). The outputs of the two models are weighted through an attention fusion mechanism to generate the final stability prediction result. This hybrid model achieves an AUC of 0.94 on the internal test set, significantly outperforming the single physics model (AUC 0.78). The bioavailability estimation unit has been upgraded to a PBPK full-compartment model simulation engine, integrating a mouse physiological parameter database to simulate the drug transport process among 12 compartments, including blood, liver, tumor, and muscle. It outputs tissue concentration-time curves and automatically calculates the tumor enrichment ratio. The production feasibility assessment considers the channel blockage risk of the DF-3000 microfluidic device, predicting pressure drop changes based on fluid viscosity and particle concentration to prevent production interruptions.

[0081] In step S160, the closed-loop feedback optimization module receives new cryo-electron microscopy image data, showing that the actual particles are ellipsoidal rather than the predicted spherical shape. The system initiates a refined bias analysis: it identifies a cognitive blind spot in the generative model regarding the combination of "stirring rate" and "cooling gradient" parameters, leading to inaccurate morphology predictions. The knowledge graph is updated accordingly with the "rapid cooling—promotion—anisotropic growth" relationship edge, increasing the confidence level from 0.75 to 0.88; the generative model will increase its focus on this path in the next round of fine-tuning. A / B testing shows that the new model's morphology prediction accuracy for nanoparticle formulation tasks has improved from 61% to 79%, meeting the deployment requirements.

[0082] Example 3

[0083] This embodiment focuses on the development of oral liquid formulations specifically for children, highlighting the invention's expanded capabilities in addressing sensory attributes and personalized needs. The goal is to develop a strawberry-flavored ibuprofen oral suspension that thoroughly masks bitterness, has moderate sweetness, is free of allergenic components, is suitable for children aged 2-6 years, and maintains chemical and physical stability for 18 months of storage in a tropical climate (temperature 35℃±2℃, humidity 80%RH).

[0084] The key difference in this embodiment lies in the introduction of a sensory attribute modeling and personalized adaptation module, a new technological dimension not addressed in Embodiments 1 and 2. The system adds a "taste perception prediction model" and an "allergen transmission map," breaking through the traditional limitation of focusing solely on physicochemical properties.

[0085] In step S110, the system accesses the professional sensory database TasteDB, pediatric medication complaint records from the FDA Adverse Event Reporting System (FAERS), the EU Allergen Declaration Regulation text, and real taste response signals collected by an electronic tongue device. Key data include: the bitterness threshold of ibuprofen (0.02 mg / mL), the sweetness synergistic effect curve of aspartame and sucralose, the aroma intensity function of γ-decanolide in strawberry flavoring, and the suspension stability index of gum arabic.

[0086] In step S120, a new subgraph, "Taste Receptor-TAS2R Family-Bitter Substance Activation Intensity," is added to the knowledge graph, establishing a quantitative relationship of "Compound Structure-EC50 Activation" based on cell experimental data. Simultaneously, an "Allergen Cross-Reaction Network" is constructed to record immunological relationships such as "Peanut-and-Cashew-Presence-IgE Cross-Binding," with confidence levels assigned from the results of a double-blind clinical challenge trial.

[0087] In step S130, the user inputs requirements such as "strawberry flavor," "sugar-free," and "no nut ingredients," which the system translates into dual constraints of sensory and safety. Specifically, "masking bitterness" is quantified as a target: the electronic tongue's bitterness channel response value is less than 10% of the threshold.

[0088] In step S140, the formulation generation engine prioritizes searching for materials tagged with "flavor masking function" as "ion exchange resin" or "microencapsulation" during the excipient selection stage, generating a recommendation to use polyammonium methacrylate (Eudragit E100) to coat ibuprofen. It also recommends adding a flavor modifier combination: sucralose (0.12% w / v) + vanillin (0.03% w / v) + strawberry flavor (0.08% w / v).

[0089] In step S150, a new "sensory attribute prediction unit" is added, which adopts a transfer learning architecture: the bottom layer is a ResNet-18 backbone network, inputting the response heatmap of the electronic tongue at different concentrations; the top layer is replaced with a fully connected layer, outputting three scores: bitterness, sweetness, and astringency. The model is pre-trained on a dataset containing 2,300 known taste samples, and the prediction error after fine-tuning is <±8%. The physical stability prediction considers the sedimentation rate under high temperature and high humidity conditions, adopts the Stokes law to correct the model, and introduces a dynamic correction term for the flocculation index.

[0090] In step S160, closed-loop feedback receives feedback from pediatric clinical trials: some children still perceive a slight bitter taste. The system traces back to find that the Eudragit coating ruptures prematurely in the acidic environment of the stomach. The knowledge graph is updated with the relationship "Eudragit E100—degradation rate accelerates at pH < 4," and the generated model automatically recommends enteric-coated taste-masking materials in subsequent tasks. A / B testing shows that the new system improved parent satisfaction scores by 12.5 percentage points on pediatric formulation tasks.

[0091] The training dataset for this system includes:

[0092] 86,421 active pharmaceutical ingredients (covering all small molecule drugs in the FDA Orange Book)

[0093] 12,893 excipients (including those listed in USP / NF, EP, and JP)

[0094] 5,672 process specifications (from EMA / FDA public approval records)

[0095] 3.89 million formulation formulation-performance correlation records (including dissolution profiles and stability data)

[0096] Key performance comparison (vs. traditional trial-and-error method & existing AI platform A):

[0097] index Traditional methods Existing AI platform A This invention Average development cycle 18 months 90 days ≤45 days First-time formula pass rate 22％ 58％ 76％ <![CDATA[Key Performance Prediction R 2 > — 0.78 0.91 Reduced cost of invalid experiments — 35％ 62％

[0098] Case 1: Glimepiride Extended-Release Tablets (see Example 1)

[0099] Objective: Zero-order release, ≥80% dissolution within 8 hours, lactose prohibited.

[0100] This system recommends the following Top-1 formulation: HPMC K15M 38% + MCC pH10 145% + MgSt 1%

[0101] The measured f² = 53.7 (target ≥ 50), and the bioavailability prediction error was +8.2% (better than the PBPK full model ±15%).

[0102] Case 2: Paclitaxel Nanoparticle Suspension (see Example 2)

[0103] Target: Particle size 100–150 nm, encapsulation efficiency >95%

[0104] The system recommends a PLGA ratio of Tween 80 to 9:1, plus a microjet pressure of 120 MPa.

[0105] The measured particle size was 128±15 nm, the encapsulation efficiency was 96.3%, and the tumor enrichment ratio reached 4.2× (target ≥4×).

[0106] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for AI-powered intelligent optimization of pharmaceutical formulations, characterized in that, include: Collect multi-source pharmaceutical data and perform format normalization and quality verification. The multi-source pharmaceutical data includes physicochemical parameters of active pharmaceutical ingredients, database of excipient functional properties, existing marketed formulation information, pharmaceutical process specification documents, and clinical efficacy and safety monitoring data. A pharmaceutical formulation knowledge graph is constructed based on normalized data. The knowledge graph contains entity nodes and semantic relationship edges. The entity nodes at least cover drugs, excipients, equipment, processes, dosage forms, performance indicators, and indication categories. The system receives formulation design goals and boundary conditions input by the user. The design goals include administration method, release mode, dose intensity and storage stability requirements. The boundary conditions include a list of prohibited ingredients, cost limits and production equipment compatibility restrictions. A graph-enhanced sequence generation model is used to parse the knowledge graph context and generate candidate formulation sequences that meet the constraints. The generation process adopts an autoregressive sampling strategy, outputting one formulation component and its dosage range each time. Perform multidimensional virtual performance evaluation on the generated candidate formulations, calculate their comprehensive scores in terms of dissolution, stability, manufacturability and safety, and sort them in descending order of total score to form a recommendation list; Output a recommendation list and receive confirmation instructions or correction feedback through a human-computer interaction interface. Store the final selected formula in the enterprise knowledge base and simultaneously start a closed-loop feedback optimization process to update the model parameters.

2. The AI-based intelligent optimization method for pharmaceutical formulations according to claim 1, characterized in that, When performing deduplication on the collected data, a clustering algorithm based on Jaccard similarity is used to merge formulas with a difference of less than 5% in the proportion of excipients and the same process steps as duplicates, and retain the latest version of the data.

3. The AI-based intelligent optimization method for pharmaceutical formulations according to claim 1, characterized in that, When constructing the knowledge graph, an initial confidence score is assigned to each semantic relationship edge, with a value ranging from 0.6 to 1.

0. The score is determined by weighting based on the authority of the data source and is subsequently dynamically adjusted through closed-loop feedback.

4. The AI-based intelligent optimization method for pharmaceutical formulations according to claim 1, characterized in that, The release mode input by the user includes sustained release, controlled release, enteric coating, immediate release, or targeted release. The system automatically converts these into corresponding mathematical representation templates and embeds them into the prompt word space of the generated model.

5. The AI-based intelligent optimization method for pharmaceutical formulations according to claim 1, characterized in that, The graph-enhanced sequence generation model introduces a Monte Carlo tree search strategy during the decoding stage to explore potentially high-value recipe paths that are not covered by high-frequency training samples, with the exploration coefficient set to 1.

4.

6. The AI-based intelligent optimization method for pharmaceutical formulations according to claim 1, characterized in that, The multidimensional virtual performance evaluation adopts a weighted summation model. The weights of each dimension are pre-configured by the user or automatically generated by the system based on the clustering of historical successful cases. The default weight allocation is: dissolution performance 30%, physical stability 25%, production feasibility 20%, safety 15%, and cost 10%.

7. The AI-based intelligent optimization method for pharmaceutical formulations according to claim 1, characterized in that, The closed-loop feedback optimization process performs A / B testing verification before model updates, comparing the old and new versions of the model on the same set of blind test tasks. The new version is only allowed to be deployed online when its Top-3 hit rate is improved by more than 8 percentage points.

8. A pharmaceutical formulation AI intelligent optimization system, characterized in that, include: The data integration and preprocessing module is used to uniformly access and standardize the processing of multi-source pharmaceutical data from literature databases, patent databases, experimental record systems, drug regulatory platforms, and clinical feedback systems. The knowledge graph construction module is used to automatically extract entities, attributes, and semantic relationships based on preprocessed structured and semi-structured data, and build a dedicated knowledge graph for the pharmaceutical formulation field covering the entire chain of "drug-excipient-process-performance-safety". The formulation generation engine module receives user input regarding the target formulation type, route of administration, release characteristics, target specifications, and constraints. It then combines this information with the current knowledge graph state and uses a sequence generation model guided by a graph attention mechanism to output a set of candidate formulations. The performance prediction and screening module is used to perform multi-dimensional virtual evaluation of each item in the candidate formulation set, including dissolution behavior simulation, physical stability prediction, bioavailability estimation, production feasibility scoring and potential interaction warning, and output a comprehensive score ranking list. The closed-loop feedback optimization module is used to receive external validation result data, including laboratory test data, pilot-scale batch quality reports or clinical trial feedback, to perform deviation analysis between actual performance and predicted results, and to drive the joint reverse update of knowledge graph and generative model parameters.

9. The AI-powered intelligent optimization system for pharmaceutical formulations according to claim 8, characterized in that, In the entity linking stage, the knowledge graph construction module introduces a chemical structure fingerprint hash index mechanism to perform topological encoding on the SMILES string, thereby achieving accurate matching of the same compound in different databases.

10. The AI-powered intelligent optimization system for pharmaceutical formulations according to claim 8, characterized in that, The sequence generation model used by the recipe generation engine module is based on an improved Transformer architecture. Its decoder part is embedded with a graph attention layer, which can dynamically focus on the subgraph region in the knowledge graph that is most relevant to the current context in each generation operation.

Images