Literature-driven drug-loading corneal contact lens candidate molecule proxy type automatic research and development method and system

Abstract

The invention provides a literature-driven drug-loading corneal contact lens candidate molecule proxy type automatic research and development method and system, and the method comprises the steps: S1, executing semantic retrieval and analysis through a literature semantic analysis engine, and outputting a structured molecule feature vector set; s2, inputting the feature vector set into a molecule generation model through a generation strategy controller, and generating a candidate molecule set; s3, performing parallel property prediction on the candidate molecules through a multi-dimensional property evaluator to generate a multi-dimensional property score vector; and S4, screening candidate molecules based on the score vector, feeding back the candidate molecules to a generation strategy controller, and constructing a research and development trajectory database. By constructing a closed-loop system architecture of four functional modules including a literature semantic analysis engine, a generation strategy controller, a multi-dimensional property evaluator and a dynamic feedback optimizer, full-process automatic research and development from scientific literatures to candidate molecules are realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of artificial intelligence, in particular to a literature-driven candidate molecule agent-based automatic research and development method and system for drug-loaded contact lenses. BACKGROUND

[0002] Drug-loaded contact lenses (a new type of ocular drug delivery system combining drugs with contact lenses) can achieve long-acting and precise release of drugs on the corneal surface, avoiding the defects of low bioavailability and frequent administration of traditional eye drops. However, drug-loaded contact lenses need to deliver what drugs (i.e. effective drug molecules), which requires pre-discovery and screening of candidate molecules.

[0003] The traditional candidate molecule research and development process highly depends on manual intervention and discrete tool chains, and its core operation mode is based on static task allocation and independent software calls. However, in real research and development scenarios, there are complex links such as massive literature evidence extraction, multi-dimensional molecular attribute prediction, cross-platform tool scheduling and dynamic feedback optimization. The data of each stage is in serious isolation, lacks unified semantic context and execution state memory mechanism, resulting in broken task connection, untraceable verification results and high repetition rate of generated solutions. Especially in the face of novel structure or harshly constrained target molecule design, the existing system often falls into local optimal solution or invalid cycle due to its inability to autonomously decompose the target, dynamically select tool combinations and iteratively optimize the path in a closed loop, greatly restricting the conversion efficiency and success rate from literature insight to synthesizable candidates.

[0004] Among them, the agent-based automatic research and development system as a new technology direction aims to realize autonomous planning and execution of research and development tasks through agent architecture. This type of system usually takes task goals as the driving force to call retrieval, docking, ADMET prediction and other sub-modules to complete the stage output, which can be used for effective drug molecule screening in drug-loaded contact lenses. However, most of the current implementations still remain in the linear mode of "single-point trigger-single-response", lack the ability to model the context of historical interactions, have no decision mechanism for dynamically adapting tool strategies for different sub-tasks, and have not built a closed-loop pipeline of generation-verification-feedback, resulting in structural defects such as response lag, path rigidity and resource mismatch when the system faces multi-constrained, multi-target and long-chain research and development tasks.

[0005] The existing technologies generally have problems such as extensive task scheduling, context fault, lack of verification concurrency ability, weak feedback learning mechanism, etc. Especially in complex research and development scenarios that require high-frequency iteration and multi-tool collaboration, the system is difficult to maintain task state consistency, ensure candidate solution diversity and improve verification throughput efficiency, ultimately leading to low overall link efficiency and uncontrollability from literature input to high-quality candidate output, and an automatic research and development system architecture with agent scheduling, memory enhancement, closed-loop verification and adaptive learning ability is urgently needed.

[0006] Patent document CN114996462B discloses a scientific literature professional knowledge candidate mining method and system based on human-computer cooperation, but it mainly optimizes data mining and does not solve the above problems. SUMMARY

[0007] In view of the defects in the prior art, the purpose of the present application is to provide a literature-driven drug-loaded contact lens candidate molecule agent-type automated research and development method and system.

[0008] According to the literature-driven drug-loaded contact lens candidate molecule agent-type automated research and development method provided by the present application, the following steps are included: Step S1: receiving the target point name or disease field keyword input by the literature semantic analysis engine, performing multiple rounds of semantic expansion retrieval on multiple scientific literature databases, extracting known active molecule structures, mechanism descriptions, experimental condition parameters and biological activity values related to the target point, and outputting a structured molecular feature vector set; Step S2: inputting the structured molecular feature vector set as a conditional constraint into a molecular generation model based on a graph neural network through a generation strategy controller to generate a candidate molecule set; Step S3: performing parallelized property prediction on the candidate molecule set through a multi-dimensional property evaluator to output a multi-dimensional property score vector; Step S4: performing a Pareto frontier screening according to the multi-dimensional property score vector, retaining candidate molecules that meet the preset threshold at the same time, feeding back the screening results to the generation strategy controller as a new round of generation of positive sample guide signals, and recording the intermediate data of each generation and evaluation to construct a traceable research and development track database.

[0009] Preferably, the step S1 includes: Performing character encoding unification, paragraph boundary identification, punctuation standardization, chemical named entity recognition and molecular structure image extraction on the original literature text; using a bidirectional long short-term memory network based on an attention mechanism to model the context of the preprocessed text sequence, extract molecular structure features, synthesis constraints, active target point associations and physicochemical property boundaries; encoding the extraction results into a structured feature vector containing molecular skeleton descriptors, functional group distribution matrix, synthesis feasibility score, target affinity tendency value and physicochemical property constraint interval; performing integrity check and numerical range verification on the structured feature vector and packaging it as a standardized data package output to the generation strategy controller.

[0010] Preferably, the process of inputting the structured molecular feature vector set as a conditional constraint into a molecular generation model based on a graph neural network through a generation strategy controller includes: mapping the molecular skeleton descriptor, functional group distribution matrix and physicochemical property constraint interval in the structured feature vector into a conditional vector in the latent space through the conditional constraint encoder, limiting the boundary of the chemical space of molecule generation; performing a random sampling operation based on a variational autoencoder in the limited latent space region, and the sampling process is controlled by a historical optimization strategy parameter; sending the latent variable vector obtained by sampling to a molecular structure decoder of a graph neural network architecture, and predicting the atom type and chemical bond type node by node using an autoregressive graph generation strategy; performing ring structure closure detection and stereochemical configuration annotation to output a standard molecular structure graph and transmit it to a multi-dimensional property evaluator.

[0011] Preferably, the step S3 comprises: making the physicochemical property prediction module predict solubility, melting point, boiling point, polar surface area and partition coefficient based on a hybrid model of quantum chemical calculation and empirical formula; making the biological activity prediction module predict the binding free energy change value, agonistic or inhibitory effect intensity and selectivity index using a deep neural network under a transfer learning framework; making the synthesis feasibility prediction module evaluate the number of synthesis steps, stability of key intermediates and availability of raw materials based on a reaction rule database and a retrosynthesis path search algorithm; making the toxicity risk prediction module integrate multi-dimensional prediction models of liver toxicity, cardiac toxicity, genetic toxicity and skin irritation to output a comprehensive toxicity score; making the target affinity prediction module input the three-dimensional conformation of the candidate molecule and the pocket structure of the target protein, and output the binding affinity estimate value; After each prediction module runs independently, it enters the score normalization unit, which maps the original prediction value of each dimension to the interval of zero to one using the min-max normalization method and generates a comprehensive evaluation score by weighted summation according to the preset weight.

[0012] Preferably, the performing of the Pareto frontier screening according to the multi-dimensional property score vector comprises: making the performance gradient calculation unit perform time series analysis on the comprehensive evaluation score to calculate the performance improvement gradient of the current batch of molecules relative to the historical batch; making the constraint boundary adjustment unit dynamically adjust the lower limit of solubility, the lower limit of affinity, the upper limit of synthesis steps and the upper limit of toxicity according to the direction of the performance gradient, and the adjustment amplitude is proportional to the absolute value of the gradient; making the sampling distribution update unit calculate the skewness and kurtosis parameters by counting the distribution histogram of each dimension of the normalized score, adjusting the probability density function offset of the latent space sampler according to the skewness direction, and adjusting the sampling variance according to the kurtosis value; the updated constraint boundary value and sampling distribution parameter are packaged as a strategy update package and returned to the generation strategy controller to complete the closed-loop iteration.

[0013] Preferably, the molecule generation model can support multiple generation mode switching: In the exploration mode, the hidden space sampling temperature parameter is set to 1.5, and candidate molecules with novel structure novelty meeting the preset value are preferentially generated; In the optimization mode, the hidden space sampling temperature parameter is set to 0.8, and candidate molecules with a similarity to known active molecules reaching a preset value are preferentially generated; Mode switching is automatically triggered by a dynamic feedback optimizer according to the current round screening result, and the system automatically switches to the exploration mode when no candidate molecules meeting all preset thresholds are obtained for three consecutive rounds.

[0014] Preferably, the target point affinity prediction module adopts a hybrid prediction architecture of molecular docking and machine learning, first performs fast rigid docking to screen out candidate molecules with binding conformation energy below a threshold, performs flexible side chain optimization calculation on the screened molecules to calculate the binding free energy, inputs the optimized conformation into a graph convolutional neural network to extract the binding interface features and outputs the final affinity prediction value, and the single molecule evaluation time is controlled within 30 seconds.

[0015] Preferably, the research and development trajectory database adopts a time series graph database architecture, each candidate molecule is taken as a graph node, the node attributes include a molecular structure string, a generation timestamp and an evaluation score vector, the edge relationship between nodes represents the generation path dependence, literature reference association and property similarity, and multi-dimensional queries are supported.

[0016] According to the literature-driven drug-loaded contact lens candidate molecule agent-type automated research and development system provided by the application, the system comprises: The literature semantic analysis engine is used for receiving a target point name or disease field keyword input by a user, performing multi-round semantic expansion retrieval on multiple scientific literature databases, and outputting a structured molecular feature vector set; The generation strategy controller is used for inputting the structured molecular feature vector set as a conditional constraint into a graph neural network-based molecule generation model to generate a candidate molecule set; The multi-dimensional property evaluator is used for performing parallelized property prediction on the candidate molecule set output by the generation strategy controller, and outputting a multi-dimensional property score vector; The dynamic feedback optimizer is used for receiving the multi-dimensional property score vector to perform a Pareto frontier screening, retaining candidate molecules meeting preset thresholds at the same time, feeding back the screening result to the generation strategy controller as a positive sample guide signal for a new round of generation, and recording intermediate data of each round of generation and evaluation to construct a traceable research and development trajectory database.

[0017] Preferably, the system further comprises: The intellectual property pre-screening module is used for automatically performing structural novelty retrieval before outputting a final candidate molecule list, comparing a global public patent database with a marketed drug structure library, marking candidate molecules with structural similarity of more than 90% as potential infringement risks, and automatically generating a technical disclosure document framework for candidate molecules with structural novelty scores higher than a threshold value.

[0018] Compared with the prior art, the present application has the following beneficial effects: 1、The present application realizes full-process automation of research and development from scientific literature to candidate molecules by constructing a closed-loop system architecture of four functional modules of literature semantic analysis engine, generation strategy controller, multi-dimensional property evaluator and dynamic feedback optimizer, and provides feasibility for the automation of screening and optimization of effective drug molecules in drug-loaded contact lenses.

[0019] 2、The system provided by the present application eliminates the manual switching and tool fragmentation of the retrieval, generation and verification links in traditional research and development, shortens the average candidate molecule output period from 14 days of the traditional method to 72 hours or less, realizes full-link tracing of any candidate molecule through the research and development track database constructed by the dynamic feedback optimizer, supports reverse tracing from the final molecular structure to the original literature paragraph, the generation model hidden space coordinates and the property prediction intermediate layer activation value, and solves the technical defect of untraceable research and development process.

[0020] 3、The abnormal processing module and the incremental learning mechanism built-in the system can ensure long-term operation stability and performance continuous improvement, the multi-dimensional property evaluator adopts a hybrid prediction architecture to control the single molecule evaluation time consumption while ensuring the prediction accuracy, and meets the high-throughput research and development demand; the introduction of the intellectual property pre-screening module prepositions the patent risk identification and technical disclosure document generation to the end of the research and development process, and significantly improves the subsequent intellectual property layout efficiency. BRIEF DESCRIPTION OF DRAWINGS

[0021] Other features, objects and advantages of the present application will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 is the overall technical scheme architecture schematic diagram of the literature-to-candidate agent-type automated research and development system proposed by the present application.

[0022] Figure 2 is the core principle framework schematic diagram of the closed-loop research and development mechanism driven by the dynamic feedback optimizer in the present application.

[0023] Figure 3 is the logic flow framework diagram of the multi-source literature feature extraction and structured output executed by the literature semantic analysis engine in the present application.

[0024] Figure 4Is the molecular generation logic framework in the application that the policy controller generates based on conditional constraints and implicit space sampling.

[0025] Figure 5 Is the evaluation logic framework in the application that the multi-dimensional property evaluator performs parallelized hybrid prediction and score normalization.

[0026] Figure 6 Is the multi-level interaction relationship and data flow diagram of the functional modules of the system and the man-machine cooperation, intellectual property pre-screening module in the application. DETAILED DESCRIPTION

[0027] The application will be described in detail below with specific embodiments. The following examples will help those skilled in the art to further understand the application, but do not limit the application in any form. It should be noted that for those skilled in the art, without departing from the concept of the application, a number of changes and improvements can be made. These are within the scope of the application.

[0028] The literature-driven drug-loaded contact lens candidate molecule agent-type automated research and development method proposed in the application aims to screen effective drug molecules (candidate molecules) suitable for drug-loaded contact lenses from scientific literature through an automated process, to shorten the output cycle of candidate molecules, and to improve the intelligence and scientific nature of candidate molecule screening.

[0029] It should be noted that the literature-driven drug-loaded contact lens candidate molecule agent-type automated research and development method proposed in the application can not only be used for the automated research of effective drug molecules (candidate molecules) in drug-loaded contact lenses, but also includes but is not limited to the automated research of high-throughput candidate molecule screening requirements in the fields of drug research and development, material science, and chemical engineering.

[0030] A literature-driven drug-loaded contact lens candidate molecule agent-type automated research and development method, comprising: Step S1: receiving the input target name or disease field keyword through the literature semantic analysis engine, performing multiple rounds of semantic expansion retrieval on multiple scientific literature databases, extracting known active molecule structures, mechanism descriptions, experimental condition parameters, and biological activity values related to the target, and outputting a set of structured molecular feature vectors; The set of structured molecular feature vectors includes molecular skeleton types, substitution group distribution patterns, pharmacophore space arrangement parameters, and activity value intervals; Step S2: inputting the set of structured molecular feature vectors to the graph neural network-based molecule generation model as conditional constraints through the generation strategy controller to generate a set of candidate molecules; The molecule generation model adopts a variational autoencoder architecture, with a 512-dimensional latent space, and uses an atomic-level graph expansion strategy to generate molecular graph structures node by node. During the generation process, the pharmacophore spatial arrangement parameters are dynamically called to impose spatial geometric constraints on the generated path, ensuring that the generated molecules maintain topological similarity with known active molecules at the three-dimensional conformation level. Additionally, by adjusting the latent space sampling temperature parameter, the generated molecules cover a wider chemical space. Step S3: Perform parallel property prediction on the candidate molecule set through a multi-dimensional property evaluator, outputting a multi-dimensional property score vector. The parallel property prediction includes physicochemical property prediction, pharmacokinetic property prediction, and target affinity prediction. The physicochemical property prediction module outputs molecular weight, octanol-water partition coefficient, hydrogen bond donor and acceptor number, and polar surface area value. The pharmacokinetic property prediction module outputs oral bioavailability score, blood-brain barrier penetration probability, and liver metabolic stability index. The target affinity prediction module inputs the three-dimensional conformation of the candidate molecule and the target protein pocket structure to output the binding affinity estimate value. After normalization, all prediction results form a multi-dimensional property score vector. Step S4: Perform Pareto frontier screening based on the multi-dimensional property score vector, retaining candidate molecules that simultaneously satisfy the preset threshold values, including a binding affinity estimate value lower than -8.0 kcal / mol, an oral bioavailability score higher than 0.7, and a liver metabolic stability index higher than 0.85. The screened molecule set is returned to the generation strategy controller as the positive sample guide signal for the next round of generation. Meanwhile, the intermediate data of each generation and evaluation are recorded to construct a traceable research and development trajectory database, supporting full-link traceability of any candidate molecule.

[0031] The target name or disease field keyword input by the user is received through the literature semantic analysis engine, which interfaces with multiple scientific literature databases to perform multiple rounds of semantic expansion retrieval. The known active molecule structures, mechanism descriptions, experimental condition parameters, and biological activity values related to the target are extracted, and a structured molecular feature vector set is output, including: Perform character encoding unification, paragraph boundary identification, punctuation symbol standardization, chemical named entity recognition, and molecular structure image extraction on the original literature text; Use a bidirectional long short-term memory network based on attention mechanism to model the context of the preprocessed text sequence, extract molecular structure features, synthesis constraints, active target associations, and physicochemical property boundaries; Encode the extraction results into a structured feature vector containing molecular skeleton descriptors, functional group distribution matrix, synthesis feasibility score, target affinity tendency value, and physicochemical property constraint interval; Performing integrity check and numerical range verification on the structured feature vector, and then encapsulating it into a standardized data package for output to the generation policy controller.

[0032] The structured molecular feature vector set is input as a conditional constraint to a graph neural network-based molecular generation model, including: The molecular skeleton descriptor, functional group distribution matrix, and physicochemical property constraint interval in the structured feature vector are mapped to a conditional vector in a hidden space by a conditional constraint encoder, limiting the chemical space boundary of molecular generation; Performing random sampling operations based on variational autoencoders within the limited hidden space region, and the sampling process is controlled by historical optimization strategy parameters to ensure that the sampling point distribution is biased towards the high evaluation score region; The sampled hidden variable vector is sent to the molecular structure decoder of the graph neural network architecture, which uses an autoregressive graph generation strategy to predict atom types and bond types node by node, and is constrained by atomic valence rules to ensure chemical reasonableness; After performing ring structure closure detection and stereochemical configuration annotation, output the standard molecular structure graph and transmit it to the multi-dimensional property evaluator.

[0033] As shown in Figure 5 The candidate molecule set output by the generation policy controller is subjected to parallelized property prediction, including: The physical and chemical property prediction module predicts solubility, melting point, boiling point, polar surface area, and partition coefficient based on a hybrid model of quantum chemical calculations and empirical formulas; The biological activity prediction module uses a deep neural network under a transfer learning framework to predict binding free energy change, agonistic or inhibitory effect strength, and selectivity index; The synthesis feasibility prediction module evaluates the number of synthesis steps, the stability of key intermediates, and the availability of raw materials based on a reaction rule database and a retrosynthesis path search algorithm; The toxicity risk prediction module integrates multidimensional prediction models for liver toxicity, cardiac toxicity, genetic toxicity, and skin irritation to output a comprehensive toxicity score; After each prediction module runs independently, it enters the score normalization unit, which uses the min-max normalization method to map each dimension of the original prediction value to the interval of zero to one and weightedly sums the results according to the preset weight to generate a comprehensive evaluation score.

[0034] As shown in Figure 2 The multi-dimensional property score vector is subjected to Pareto frontier screening, including: The performance gradient calculation unit performs time series analysis on the comprehensive evaluation score to calculate the performance improvement gradient of the current batch of molecules relative to historical batches; The constraint boundary adjustment unit adjusts the solubility lower limit, affinity lower limit, synthesis step upper limit and toxicity upper limit according to the performance gradient direction, and the adjustment range is proportional to the absolute value of the gradient; The sampling distribution updating unit calculates the skewness and kurtosis parameters based on the distribution histogram of the normalized scores of each dimension, adjusts the probability density function offset of the latent space sampler according to the skewness direction, and adjusts the sampling variance according to the kurtosis value. The updated constraint boundary values and sampling distribution parameters are packaged into a strategy update package and returned to the generated strategy controller to complete the closed-loop iteration.

[0035] The scientific literature database connected by the literature semantic analysis engine includes CNKI scientific journal library, Wanfang data medical literature library, VIP Chinese scientific and technological journal database, PubMed biomedical literature library, ScienceDirect engineering and technology literature library, IEEE Xplore electronic engineering literature library, and the engine has a built-in multi-source heterogeneous data fusion module to perform field alignment, unit unification and term standardization processing on literature records returned by different databases.

[0036] The molecular generation model supports multiple generation mode switching. In the exploration mode, the latent space sampling temperature parameter is set to 1.5 to preferentially generate candidate molecules with high novelty of structure. In the optimization mode, the latent space sampling temperature parameter is set to 0.8 to preferentially generate candidate molecules with high similarity to known active molecules. Mode switching is automatically triggered by the dynamic feedback optimizer according to the current round screening results. When no candidate molecules meeting all preset thresholds are obtained for three consecutive rounds, the system automatically switches to the exploration mode.

[0037] The target affinity prediction module adopts a hybrid prediction architecture of molecular docking and machine learning. First, fast rigid docking is performed to screen out candidate molecules with binding conformation energy below a threshold value. Then, flexible side chain optimization is performed on the screened molecules to calculate the binding free energy. Finally, the optimized conformation is input into a graph convolutional neural network to extract binding interface features and output the final affinity prediction value. The single molecule evaluation time is controlled within 30 seconds.

[0038] The R&D trajectory database adopts a time series graph database architecture. Each candidate molecule is a graph node, and the node attributes include the molecular structure string, generation timestamp, and evaluation score vector. The edge relationship between nodes represents the generation path dependence, literature citation association, and property similarity, supporting multi-dimensional queries based on substructure matching, property similarity threshold, and literature source tracing.

[0039] As shown in Figure 1 and Figure 6 A literature-driven drug-loaded contact lens candidate molecule agent-type automated R&D system, comprising: A literature semantic analysis engine is configured to receive a target point name or disease field keyword input by a user, perform multi-round semantic expansion retrieval on multiple scientific literature databases, extract known active molecule structures, mechanism of action descriptions, experimental condition parameters, and biological activity values related to the target point, and output a structured molecule feature vector set containing a molecular skeleton type, a substituent distribution pattern, a pharmacophore spatial arrangement parameter, and an activity value interval. A generation strategy controller is configured to input the structured molecule feature vector set as a conditional constraint into a graph neural network-based molecule generation model; the molecule generation model adopts a variational autoencoder architecture with a 512-dimensional hidden space, and the decoder part adopts an atomic-level graph expansion strategy to generate a molecular graph structure node by node, dynamically calls the pharmacophore spatial arrangement parameter to impose spatial geometric constraints on the generation path during the generation process, ensures that the generated molecule maintains topological similarity with the known active molecule at the three-dimensional conformation level, and guides the generated molecule to cover a wider chemical space by adjusting the hidden space sampling temperature parameter. A multi-dimensional property evaluator is configured to perform parallelized property prediction on the candidate molecule set output by the generation strategy controller; the parallelized property prediction includes physicochemical property prediction, pharmacokinetic property prediction, and target point affinity prediction, the physicochemical property prediction module outputs molecular weight, octanol-water partition coefficient, hydrogen bond donor-acceptor number, and polar surface area value, the pharmacokinetic property prediction module outputs oral bioavailability score, blood-brain barrier penetration probability, and liver metabolic stability index, the target point affinity prediction module inputs the three-dimensional conformation of the candidate molecule and the target protein pocket structure to output the binding affinity estimate value, and all prediction results are normalized to form a multi-dimensional property score vector. A dynamic feedback optimizer is configured to receive the multi-dimensional property score vector to perform a Pareto frontier screening, retain candidate molecules that simultaneously satisfy preset threshold values, the preset threshold values include a binding affinity estimate value lower than -8.0 kcal / mol, an oral bioavailability score higher than 0.7, and a liver metabolic stability index higher than 0.85, and the screened molecule set is fed back to the generation strategy controller as a positive sample guide signal for a new round of generation, while recording the intermediate data of each generation and evaluation to construct a traceable research and development track database, supporting full-link traceability of any candidate molecule.

[0040] The literature semantic analysis engine is used to perform character encoding unification, paragraph boundary identification, punctuation standardization, chemical named entity recognition and molecular structure image extraction on the original literature text; a bidirectional long short-term memory network based on attention mechanism is used to model the context of the preprocessed text sequence, extract molecular structure features, synthesis constraints, active target point association and physicochemical property boundaries; the extracted results are encoded into a structured feature vector containing molecular skeleton descriptors, functional group distribution matrix, synthesis feasibility score, target point affinity tendency value and physicochemical property constraint interval; after performing integrity check and numerical range verification on the structured feature vector, it is packaged as a standardized data packet and output to the generation strategy controller.

[0041] The generation strategy controller is used to map the molecular skeleton descriptor, functional group distribution matrix and physicochemical property constraint interval in the structured feature vector into a conditional vector in the hidden space through the conditional constraint encoder, to limit the chemical space boundary of molecular generation; perform random sampling operation based on variational autoencoder in the limited hidden space region, the sampling process is controlled by the historical optimization strategy parameters, to ensure that the sampling point distribution is biased towards the high evaluation score region; the sampled hidden variable vector is sent to the molecular structure decoder of the graph neural network architecture, and the autoregressive graph generation strategy is used to predict the atom type and chemical bond type node by node, which is constrained by the atomic valence rule to ensure chemical rationality; after performing ring structure closure detection and stereochemical configuration annotation, output the standard molecular structure graph and transmit it to the multi-dimensional property evaluator.

[0042] The multi-dimensional property evaluator is used for the physical and chemical property prediction module to predict solubility, melting point, boiling point, polar surface area and partition coefficient based on a hybrid model of quantum chemical calculation and empirical formula; the biological activity prediction module uses a deep neural network under a transfer learning framework to predict the binding free energy change value, agonistic or inhibitory effect intensity and selectivity index; the synthesis feasibility prediction module evaluates the number of synthesis steps, stability of key intermediates and availability of raw materials based on a reaction rule database and retrosynthesis path search algorithm; the toxicity risk prediction module integrates multidimensional prediction models of hepatotoxicity, cardiotoxicity, genotoxicity and skin irritation to output a comprehensive toxicity score; After each prediction module runs independently, it enters the score normalization unit, which maps each dimension of the original prediction value to the interval of zero to one using the min-max normalization method and generates a comprehensive evaluation score by weighted summation according to the preset weight.

[0043] The dynamic feedback optimizer is used for time series analysis of the comprehensive evaluation score by the performance gradient calculation unit to calculate the performance improvement gradient of the current batch of molecules relative to historical batches; the constraint boundary adjustment unit dynamically adjusts the solubility lower limit, affinity lower limit, synthesis step upper limit and toxicity upper limit according to the performance gradient direction, and the adjustment range is proportional to the absolute value of the gradient; the sampling distribution update unit calculates the skewness and kurtosis parameters by counting the distribution histogram of the normalized score in each dimension, adjusts the probability density function offset of the latent space sampler according to the skewness direction, and adjusts the sampling variance according to the kurtosis value; the updated constraint boundary value and sampling distribution parameter are packaged as a strategy update package and returned to the generated strategy controller to complete the closed-loop iteration.

[0044] The system is provided with an abnormality processing module, which automatically switches to a backup database interface when the literature semantic analysis engine fails to search continuously for five times, triggers a structure correction subroutine to automatically adjust the atomic connection mode when the generated strategy controller outputs a molecule that violates the chemical valence rule, and starts a model recalibration process to fine-tune the prediction model parameters using the latest evaluation data when the result variance predicted by the multi-dimensional property evaluator exceeds a preset threshold.

[0045] The system is deployed in a distributed computing cluster, the literature semantic analysis engine is deployed in a storage optimization node equipped with a solid state disk array, the generated strategy controller is deployed in a computing optimization node equipped with a graphics processor acceleration card, the multi-dimensional property evaluator is deployed in a general-purpose computing node equipped with a central processing unit multi-core parallel computing unit, and the dynamic feedback optimizer serves as a central scheduler to dynamically allocate computing resources according to task load to ensure that the overall throughput of the system does not fall below 2,000 candidate molecules processed per hour.

[0046] The system is provided with a man-machine cooperation interface, which allows researchers to manually label key literature paragraphs to strengthen feature extraction weights during the literature semantic analysis stage, specify that certain functional groups must be included or excluded during the molecule generation stage, and adjust the weight coefficients of each property dimension during the property evaluation stage. All manual intervention operations are recorded in the research and development trajectory database.

[0047] The system supports incremental learning mechanism, automatically extracts high-score candidate molecules generated in the current round, their corresponding literature features, generation path and evaluation data to construct an incremental training set every time a complete research and development cycle is completed, and performs model parameter update once a month to fine-tune the literature semantic analysis model, molecule generation model and property prediction model jointly using the incremental training set.

[0048] The system is also provided with a knowledge property pre-screening module, which automatically performs structure novelty search and compares global public patent database and marketed drug structure library before outputting the final candidate molecule list, labels candidate molecules with more than 90% structural similarity as potential infringement risk, and automatically generates a technical disclosure framework for candidate molecules with a structure novelty score higher than the threshold.

[0049] The application discloses an agent-type automatic research and development system from literature to candidate, aiming to solve the technical problems of fragmented links, multiple manual interventions, untraceable process and long cycle in traditional drug research and development. The system comprises four modules of literature semantic analysis engine, generation strategy controller, multi-dimensional property evaluator and dynamic feedback optimizer, realizes seamless connection of data and control flow through standardized interface: the literature engine analyzes the molecular characteristics related to the target; the generation controller generates candidate molecules by combining graph neural network with spatial constraints; the evaluator predicts the physicochemical, pharmacokinetic and affinity properties in parallel; the optimizer performs Pareto screening and feeds back to guide the next generation, while constructing a traceable time sequence database.

[0050] Through the above closed-loop architecture, the application realizes the full-process automation from literature to candidate molecules within 72 hours, supports full-link traceability, abnormal self-healing, incremental learning and patent pre-screening, and significantly improves the research and development efficiency, traceability and intellectual property layout speed.

[0051] Embodiment 1 The application proposes an agent-type automatic research and development system from literature to candidate, the core goal of which is to break through the fragmented state of the three links of retrieval, generation and verification in the effective drug molecule research and development link of drug-loaded contact lenses, and to construct an automatic research and development agent system with closed-loop feedback, semantic driving and multi-dimensional evaluation capability. In the traditional research and development process, literature retrieval relies on manual keyword matching, the generation link lacks structured constraint guidance, and the verification process is isolated from the generation logic, and there is no data closed loop and dynamic optimization mechanism among the three, resulting in low research and development efficiency, poor result traceability and uncontrollable innovation path. The application forms an end-to-end automatic research and development closed loop from original literature input to candidate molecule output by constructing four core functional modules of literature semantic analysis engine, generation strategy controller, multi-dimensional property evaluator and dynamic feedback optimizer, and realizes the intelligentization, structurization and iterative optimization of the research and development process.

[0052] In the embodiment, the agent-type automatic research and development system from literature to candidate comprises a literature semantic analysis engine, a generation strategy controller, a multi-dimensional property evaluator and a dynamic feedback optimizer. The literature semantic analysis engine is used to receive original literature data, perform multi-source literature feature extraction and structured output; the generation strategy controller is used to receive structured literature features and historical generation feedback, perform molecule generation based on conditional constraints and implicit space sampling; the multi-dimensional property evaluator is used to receive generated molecule structures, perform parallelized hybrid prediction and score normalization processing; and the dynamic feedback optimizer is used to receive evaluation results, perform closed-loop optimization strategy generation and return to the generation strategy controller to form a research and development closed loop.

[0053] The literature semantic parsing engine receives raw text data from scientific literature databases, patent literature libraries, experimental report libraries, and technical white paper libraries. The raw text data includes but is not limited to compound structure descriptions, synthesis path instructions, biological activity data, physicochemical property parameters, experimental condition records, and molecular action mechanism descriptions.

[0054] The literature semantic parsing engine first performs preprocessing operations on the raw text, including character encoding unification, paragraph boundary identification, punctuation symbol standardization, chemical named entity recognition, and molecular structure image extraction. The preprocessed text data enters the semantic parsing module, which uses a bidirectional long short-term memory network based on attention mechanism to model the context of the text sequence, extract the implicit molecular structure features, synthesis constraint conditions, active target point correlation, and physicochemical property boundaries in the literature. The semantic parsing module outputs a structured feature vector, which includes molecular skeleton descriptors, functional group distribution matrices, synthesis feasibility scores, target point affinity tendency values, and physicochemical property constraint intervals. The structured feature vector is encapsulated as a standardized data packet and transmitted to the generation strategy controller through the internal data bus.

[0055] As shown in Figure 4 The generation strategy controller receives the structured feature vector from the literature semantic parsing engine and synchronously receives the historical optimization strategy parameters from the dynamic feedback optimizer. The generation strategy controller internally includes three sub-modules: conditional constraint encoder, latent space sampler, and molecular structure decoder. The conditional constraint encoder maps the molecular skeleton descriptors, functional group distribution matrices, and physicochemical property constraint intervals in the structured feature vector into a conditional vector in the latent space, which is used to define the chemical space boundary for subsequent molecular generation. The latent space sampler performs random sampling operations based on variational autoencoder within the defined latent space region, and the sampling process is controlled by the historical optimization strategy parameters to ensure that the sampling point distribution is biased towards the high evaluation score region. The latent space sampler outputs a latent variable vector, which is sent to the molecular structure decoder. The molecular structure decoder uses a graph neural network architecture to gradually decode the latent variable vector into an atomic node sequence and a chemical bond edge sequence, and finally outputs a molecular structure graph that conforms to the chemical valence rule. The molecular structure graph is output in a standard molecular file format and is simultaneously transmitted to the multi-dimensional property evaluator.

[0056] The multi-dimensional property evaluator receives the molecular structure graph from the generation strategy controller and performs parallelized hybrid prediction and score normalization processing. The multi-dimensional property evaluator internally includes a physical property prediction module, a biological activity prediction module, a synthesis feasibility prediction module, and a toxicity risk prediction module.

[0057] The physical property prediction module is based on a hybrid model of quantum chemical calculations and empirical formulas to predict the solubility, melting point, boiling point, polar surface area, and partition coefficient of the molecule.

[0058] The bioactivity prediction module uses a deep neural network under the transfer learning framework to predict the binding affinity of molecules to a specified target, the intensity of agonistic or inhibitory effects, and the selectivity index.

[0059] The synthesis feasibility prediction module, based on a reaction rule database and a retrosynthesis path search algorithm, evaluates the number of synthesis steps, the stability of key intermediates, and the availability of raw materials for molecular structures.

[0060] The toxicity risk prediction module integrates multi-dimensional prediction models for hepatotoxicity, cardiotoxicity, genotoxicity, and skin irritation, and outputs a comprehensive toxicity score.

[0061] Each prediction module operates independently, outputting raw predicted values, which are then fed into the scoring normalization unit. The scoring normalization unit uses a min-max normalization method to map the raw predicted values ​​for each dimension to the zero-to-one interval, and performs a weighted summation based on preset weight coefficients to generate a comprehensive evaluation score. The comprehensive evaluation score and the raw predicted values ​​for each dimension are encapsulated together into an evaluation result data packet, which is transmitted to the dynamic feedback optimizer via an internal data bus.

[0062] The dynamic feedback optimizer receives evaluation result data packets from the multidimensional property evaluator and executes closed-loop optimization strategy generation. Internally, the dynamic feedback optimizer includes a performance gradient calculation unit, a constraint boundary adjustment unit, and a sampling distribution update unit. The performance gradient calculation unit performs time-series analysis on the comprehensive evaluation score to calculate the performance improvement gradient of the current batch of molecules relative to historical batches. The constraint boundary adjustment unit dynamically shrinks or expands the physicochemical property constraint interval output by the literature semantic parsing engine based on the performance gradient direction; for example, when solubility prediction values ​​are generally low, the upper limit constraint of polar surface area is automatically relaxed. The sampling distribution update unit adjusts the sampling probability density function of the latent space sampler based on the distribution characteristics of the original prediction values ​​for each dimension; for example, when bioactivity prediction values ​​show a right-skewed distribution, the sampling weight of high-affinity regions is increased. The dynamic feedback optimizer outputs optimization strategy parameters, including updated constraint boundary values ​​and sampling distribution parameters, which are fed back to the generation strategy controller through a feedback channel, completing one closed-loop iteration of the R&D process.

[0063] like Figure 3 As shown in this embodiment, the specific process of the document semantic parsing engine performing multi-source document feature extraction and structured output is as follows: The original literature data is first sent to the text cleaning module to remove the header and footer, table title, reference list, and non-text paragraphs. The cleaned text enters the named entity recognition module, which uses a conditional random field model to identify chemical substance names, experimental parameter values, biological target codes, and instrument model identifiers. The recognition results are sent to the dependency syntax analysis module to build a sentence-level semantic dependency tree and extract the core semantic relationships in the subject-predicate-object structure, such as "compound A inhibits the activity of protein B" is parsed as an inhibition relationship triple. The dependency syntax analysis results and named entity recognition results are jointly input into the semantic role labeling module to label the semantic roles of each entity in the sentence, such as "inhibitor", "substrate", "condition parameter", etc. The semantic role labeling results are sent to the feature vectorization module, which uses a pre-trained chemical language model to encode the semantic role sequence into a fixed-length feature vector. The feature vector output by the feature vectorization module enters the structured packaging module, which is packaged into a structured data package containing the molecule description field, constraint condition field, activity association field, and property boundary field according to the predefined data pattern. The structured data package passes through the data verification module to perform integrity check and value range verification to ensure that there are no missing fields and illegal values, and then outputs to the generation strategy controller.

[0064] In this embodiment, the generation strategy controller is based on conditional constraints and implicit space sampling molecular generation logic as follows: The conditional constraint encoder receives the molecule description field and property boundary field in the structured data package and maps it to the hyperplane constraint condition in the implicit space. The implicit space sampler performs importance sampling under the hyperplane constraint condition, and the sampling point distribution is controlled by the sampling distribution parameters provided by the dynamic feedback optimizer. The implicit variable vector output by the implicit space sampler is sent to the molecular structure decoder, which uses an autoregressive graph generation strategy to predict atom types and bond types node by node. The decoding process is constrained by the atomic valence rule to ensure the chemical reasonableness of the generated structure. After the decoder outputs the atom node sequence and edge sequence, it performs ring structure closure detection and stereochemical conformation annotation, and finally outputs the standard molecular structure graph. The molecular structure graph is packaged into a molecule data file containing atomic coordinates, bond types, charge distribution, and stereochemical conformation information, and output to the multi-dimensional property evaluator.

[0065] In this embodiment, the multi-dimensional property evaluator performs parallelized mixed prediction and score normalization evaluation logic as follows: The physical property prediction module receives the molecular data file, first performs molecular force field parameterization to generate an interatomic potential function. Then it performs Monte Carlo simulation to calculate solubility, performs density functional theory approximation to calculate polar surface area, and performs empirical formula interpolation to calculate partition coefficient. The biological activity prediction module loads the pre-trained target binding model, encodes the molecular structure graph into a graph embedding vector, and inputs it into a fully connected neural network to predict the binding free energy change value. The synthesis feasibility prediction module performs retrosynthesis path search, matches feasible synthesis paths based on a reaction rule base, and calculates the path length and key step risk coefficient. The toxicity risk prediction module calls multiple toxicity prediction sub-models to calculate liver toxicity probability, cardiac toxicity risk value, genetic toxicity score, and skin irritation index, and takes the maximum value as the comprehensive toxicity score. The prediction results of each module are sent to the score normalization unit, and the normalization formula is as follows: One, the solubility normalization value is equal to the solubility prediction value minus the minimum solubility threshold, and then divided by the difference between the maximum solubility threshold and the minimum solubility threshold. Two, the affinity normalization value is equal to the binding free energy change value minus the minimum affinity threshold, and then divided by the difference between the maximum affinity threshold and the minimum affinity threshold. Three, the synthesis score normalization value is equal to one minus the number of synthesis steps divided by the maximum allowed number of steps. Four, the toxicity normalization value is equal to one minus the comprehensive toxicity score.

[0066] After normalization, each dimension score is weighted and summed according to the preset weight. The weight distribution is solubility weight 0.2, affinity weight 0.4, synthesis score weight 0.3, and toxicity weight 0.1, to generate a comprehensive evaluation score.

[0067] In this embodiment, the dynamic feedback optimizer-driven closed-loop R&D mechanism is as follows: The performance gradient calculation unit receives the current batch of comprehensive evaluation scores and the sequence of historical batch scores, and calculates the average score change rate in the sliding window. If the change rate is positive and greater than a preset threshold, it is determined that the performance is improving, and the constraint boundary expansion operation is triggered; if the change rate is negative, the constraint boundary contraction operation is triggered. The constraint boundary adjustment unit adjusts the solubility lower limit, the affinity lower limit, the synthesis step upper limit, and the toxicity upper limit dynamically according to the performance gradient direction, and the adjustment amplitude is proportional to the absolute value of the gradient. The sampling distribution update unit calculates the skewness and kurtosis parameters by counting the distribution histogram of each dimension normalization score, adjusts the probability density function offset of the hidden space sampler according to the skewness direction, and adjusts the sampling variance according to the kurtosis value. The optimization strategy parameters are packaged into a strategy update package, including the boundary adjustment vector and the distribution parameter vector, which are returned to the generation strategy controller through the feedback channel to complete the closed-loop iteration.

[0068] The system described in this embodiment also includes a human-computer collaborative interface module and a knowledge property pre-screening module. The human-computer collaborative interface module provides a graphical operation interface, allowing R&D personnel to manually input priority constraints, pause or restart a generation batch, view intermediate evaluation results, and export a candidate molecule list. The knowledge property pre-screening module performs preliminary screening of patent novelty immediately after molecule generation, compares compound structures in a global patent database, marks molecules with potential infringement risks, and generates avoidance design recommendations. The human-computer collaborative interface module and the knowledge property pre-screening module interact with the main system through an event-driven mechanism, ensuring that manual intervention and legal compliance checks are seamlessly embedded in the automated R&D process.

[0069] In actual operation, the system described in this embodiment starts with a literature semantic analysis engine, receives a user-specified literature set or automatically crawls the latest literature in the relevant field. After the literature is analyzed, a structured feature vector is generated, driving the generation strategy controller to produce the first batch of candidate molecules. After the candidate molecules are scored by the multi-dimensional property evaluator, the dynamic feedback optimizer adjusts the generation strategy based on the score distribution, starting the next batch of molecule generation. This process is iterated until the comprehensive evaluation score converges or the preset iteration upper limit is reached. The final output candidate molecule list includes molecular structure, dimension prediction values, comprehensive score, and intellectual property risk level, for use by R&D personnel in decision-making.

[0070] The system described in this embodiment effectively solves the problems of low efficiency and poor traceability caused by the fragmentation of the R&D chain by building an automated R&D agent with literature-driven, closed-loop optimization, and multi-dimensional evaluation. The system has semantic understanding capabilities, can automatically extract R&D constraints from unstructured literature; has dynamic optimization capabilities, can automatically adjust the generation strategy based on evaluation feedback; has multi-dimensional evaluation capabilities, can predict physical, biological, synthetic, and toxicity properties in parallel; and has human-computer collaboration and legal compliance capabilities, ensuring that the R&D process is controllable and legal. The candidate molecules output by the system have high innovation, high feasibility, and high compliance, significantly improving R&D efficiency and success rate.

[0071] Those skilled in the art know that, in addition to implementing the system provided by the present application and each device, module, and unit thereof in pure computer-readable program code, the same functions can be achieved by logically programming the method steps in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers. Therefore, the system provided by the present application and each device, module, and unit thereof can be considered as a hardware component, and the devices, modules, and units included therein for achieving various functions can also be considered as structures within the hardware component. The devices, modules, and units for achieving various functions can also be considered as both software modules implementing methods and structures within hardware components.

[0072] The specific embodiments of the present application are described above. It needs to be understood that the present application is not limited to the specific embodiments described above, and various changes or modifications can be made by those skilled in the art within the scope of the claims, which does not affect the essential content of the present application. The embodiments of the present application and the features in the embodiments can be combined with each other at will without conflict.

Claims

1. A literature-driven automated development method for drug-eluting corneal contact lens candidate molecular surrogates, characterized in that, include: Step S1: Receive the target name or disease domain keywords through the literature semantic parsing engine, connect to multiple scientific literature databases to perform multiple rounds of semantic expansion retrieval, extract the known active molecular structures, mechanism descriptions, experimental condition parameters and biological activity values ​​related to the target, and output a set of structured molecular feature vectors. Step S2: The structured molecular feature vector set is input as a conditional constraint into the molecular generation model based on graph neural network through the generation strategy controller to generate a set of candidate molecules; Step S3: Perform parallel property prediction on the candidate molecule set using a multidimensional property evaluator, and output a multidimensional property score vector; Step S4: Perform Pareto front screening based on the multidimensional property scoring vector, retain candidate molecules that simultaneously meet the preset threshold, feed back the screening results to the generation strategy controller and use them as a positive sample guidance signal for the next round of generation, and record the intermediate data of each round of generation and evaluation to build a traceable R&D trajectory database.

2. The literature-driven automated development method for drug-eluting corneal contact lens candidate surrogate molecules according to claim 1, characterized in that, Step S1 includes: The original literature text undergoes character encoding standardization, paragraph boundary recognition, punctuation standardization, chemical nominate entity recognition, and molecular structure image extraction. A bidirectional long short-term memory network based on an attention mechanism is used to perform contextual modeling on the preprocessed text sequence, extracting molecular structure features, synthesis constraints, active target associations, and physicochemical property boundaries. The extracted results are encoded into structured feature vectors containing molecular skeleton descriptors, functional group distribution matrices, synthesis feasibility scores, target affinity propensity values, and physicochemical property constraint intervals. After performing integrity checks and numerical range verification on the structured feature vectors, they are encapsulated into standardized data packets and output to the generation strategy controller.

3. The literature-driven automated development method for drug-eluting corneal contact lens candidate surrogate molecules according to claim 2, characterized in that, The process of inputting the structured molecular feature vector set as a conditional constraint into the graph neural network-based molecular generation model through a generation strategy controller includes: The molecular skeleton descriptor, functional group distribution matrix, and physicochemical property constraint intervals in the structured feature vector are mapped to conditional vectors in the latent space through a conditional constraint encoder, thus defining the chemical space boundary for molecular generation. Random sampling based on a variational autoencoder is performed within the defined latent space region, and the sampling process is controlled by historical optimization strategy parameters. The sampled latent variable vectors are fed into a molecular structure decoder based on a graph neural network architecture, and an autoregressive graph generation strategy is used to predict the atom type and chemical bond type node by node. After performing ring structure closure detection and stereochemical configuration annotation, a standard molecular structure diagram is output and transmitted to a multidimensional property evaluator.

4. The literature-driven automated development method for drug-eluting corneal contact lens candidate surrogate molecules according to claim 1, characterized in that, Step S3 includes: The physicochemical property prediction module predicts solubility, melting point, boiling point, polar surface area, and partition coefficient based on a hybrid model of quantum chemical calculations and empirical formulas. The bioactivity prediction module uses a deep neural network under the transfer learning framework to predict the combination of free energy change, agonist or inhibitory effect intensity, and selectivity index. The synthesis feasibility prediction module evaluates the number of synthesis steps, the stability of key intermediates, and the availability of raw materials based on a reaction rule database and a retrosynthesis path search algorithm. The toxicity risk prediction module integrates multi-dimensional prediction models of hepatotoxicity, cardiotoxicity, genotoxicity, and skin irritation to output a comprehensive toxicity score; The target affinity prediction module takes the three-dimensional conformation of the candidate molecule and the pocket structure of the target protein as input and outputs the estimated binding affinity. After each prediction module runs independently, it enters the scoring normalization unit. The min-max standardization method is used to map the original predicted values ​​of each dimension to the zero-to-one interval and then sums them according to the preset weights to generate a comprehensive evaluation score.

5. The literature-driven automated development method for drug-eluting corneal contact lens candidate surrogate molecules according to claim 4, characterized in that, The Pareto front screening based on the multidimensional property scoring vector includes: The performance gradient calculation unit performs time-series analysis on the comprehensive evaluation score to calculate the performance improvement gradient of the current batch of molecules relative to historical batches. The constraint boundary adjustment unit dynamically adjusts the lower limit of solubility, lower limit of affinity, upper limit of synthesis steps, and upper limit of toxicity according to the performance gradient direction, and the adjustment magnitude is proportional to the absolute value of the gradient. The sampling distribution update unit calculates the distribution histogram of the normalized scores for each dimension, calculates the skewness and kurtosis parameters, adjusts the probability density function offset of the latent space sampler according to the skewness direction, and adjusts the sampling variance according to the kurtosis value. The updated constraint boundary values ​​and sampling distribution parameters are encapsulated into a policy update package and sent back to the generation policy controller to complete the closed-loop iteration.

6. The literature-driven automated development method for drug-eluting corneal contact lens candidate surrogate molecules according to claim 4, characterized in that, The molecular generation model can support switching between multiple generation modes: In exploration mode, the hidden space sampling temperature parameter is set to 1.5, and candidate molecules whose structural novelty meets the preset value are generated first. In optimization mode, the hidden space sampling temperature parameter is set to 0.8, and candidate molecules with a structural similarity to known active molecules that reach a preset value are generated first. The mode switching is automatically triggered by the dynamic feedback optimizer based on the screening results of the current round. When no candidate molecules that meet all preset thresholds are obtained in three consecutive rounds, the system automatically switches to exploration mode.

7. The literature-driven automated development method for drug-eluting corneal contact lens candidate molecular surrogates according to claim 4, characterized in that, The target affinity prediction module adopts a hybrid prediction architecture of molecular docking and machine learning. First, it performs rapid rigid docking to screen candidate molecules with binding conformation energy below a threshold. Then, it performs flexible side chain optimization to calculate the binding free energy of the screened molecules. Finally, it uses a convolutional neural network to extract the binding interface features from the optimized conformation input graph and outputs the final affinity prediction value. The single-molecule evaluation time is controlled within 30 seconds.

8. The literature-driven automated development method for drug-eluting corneal contact lens candidate surrogate molecules according to claim 1, characterized in that, The R&D trajectory database adopts a time-series graph database architecture, with each candidate molecule as a graph node. The node attributes include the molecular structure string, generation timestamp, and evaluation score vector. The edge relationships between nodes represent generation path dependencies, literature citation associations, and property similarities, and support multi-dimensional queries.

9. A literature-driven automated research and development system for drug-eluting corneal contact lens candidate molecules, characterized in that, include: The literature semantic parsing engine is used to receive target names or disease domain keywords input by users, connect to multiple scientific literature databases to perform multi-round semantic expansion retrieval, and output a structured molecular feature vector set; A generation strategy controller is used to input the structured molecular feature vector set as a conditional constraint into a graph neural network-based molecular generation model to generate a set of candidate molecules; A multidimensional property evaluator is used to perform parallel property predictions on the candidate molecule set output by the generation strategy controller, and output a multidimensional property score vector. A dynamic feedback optimizer is used to receive the multidimensional property score vector to perform Pareto front screening, retain candidate molecules that simultaneously meet the preset threshold, feed the screening results back to the generation strategy controller and use them as a positive sample guidance signal for the next round of generation, and record the intermediate data of each round of generation and evaluation to build a traceable R&D trajectory database.

10. The literature-driven automated research and development system for drug-eluting corneal contact lens candidate molecules according to claim 9, characterized in that, Also includes: The intellectual property pre-screening module is used to automatically perform structural novelty searches before outputting the final candidate molecule list. It compares the results with the global publicly available patent database and the structure database of marketed drugs, marks the potential infringement risks of candidate molecules with a structural similarity of more than 90%, and automatically generates a technical disclosure framework for candidate molecules with a structural novelty score higher than the threshold.

Images