Iterative drug screening method and system based on active learning, computer equipment and medium

Abstract

The invention relates to the technical field of machine learning, in particular to an iterative drug screening method and system based on active learning, computer equipment and a medium. The method comprises the following steps: obtaining a representation vector of a candidate drug set, wherein the vector is a low-dimensional dense vector extracted from a knowledge graph; taking the representation vector as input of a preset drug prediction model and a key attribute prediction model, and executing multiple rounds of active learning operation; in response to the condition that the number of iterations of the current round meets a preset iteration termination condition, taking the screening drug set of the current round as a target screening drug set and outputting the target screening drug set; the high-uncertainty characteristics of the key attributes are used for representing the confidence degree of the model for predicting the attributes. In this way, the technical problem that an existing intelligent medicine screening technology is insufficient in screening efficiency and reliability in a low-data and high-cost scene is solved, and the accuracy, novelty and resource utilization efficiency of candidate medicine screening are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of machine learning, in particular to an active learning-based iterative drug screening method and system, a computer device and a medium. BACKGROUND

[0002] With the rapid development of the biomedical field, the demand for intelligent screening and verification strategies in the process of new drug discovery and drug repositioning is increasingly urgent. The core of drug research and development is to efficiently identify molecules with potential therapeutic effects from a large number of candidate drugs, and to verify their effectiveness and safety through biomedical experiments to ensure the feasibility of clinical translation. However, the drug screening process faces multiple challenges such as severe data loss, high experimental cost, and insufficient reliability of prediction models. In particular, in low-data scenarios such as rare diseases, traditional methods are difficult to balance screening accuracy and efficiency, and innovative technical means are needed to break through the bottleneck.

[0003] In the field of intelligent drug screening, existing technologies have made some progress. Prediction models based on artificial intelligence and machine learning (such as "dry experiments") are widely used in candidate drug efficacy prediction to reduce dependence on high-cost biomedical experiments (i.e. "wet experiments"). Current mainstream methods rely on knowledge graphs, molecular representation learning, and other technologies to complete and reason about drug-target-disease relationships, thereby initially screening high-probability effective candidate drugs. However, the drug screening environment has characteristics such as data sparsity and mechanism complexity, and existing technologies still have obvious limitations in dealing with these challenges. Comparative Document 1 (GB2600154A) discloses a drug optimization method based on active learning, which defines a compound population, trains a machine learning model, and selects untested compounds for synthesis and testing based on model uncertainty or acquisition functions, achieving iterative optimization. However, this method is mainly aimed at regular drug design scenarios, and the candidate selection strategy of this method does not fully integrate mechanism analysis and target-disease network relationships, making it difficult to adapt to the high-efficiency screening needs of rare diseases and other data-scarce backgrounds. Specifically, there are two major problems with existing technologies: on the one hand, the prediction model lacks sufficient uncertainty assessment of key biological attributes (such as drug action mechanisms), resulting in high-cost experimental resources often being concentrated on samples with high prediction scores but low confidence, leading to low verification efficiency; on the other hand, the screening process lacks systematic consideration of sample diversity, which easily leads to the selection of highly similar drugs in structure or mechanism, making it difficult to break through existing cognitive boundaries and discover novel candidate drugs. These problems overlap, making it difficult for existing technologies to achieve a balance between drug screening accuracy and resource efficiency under the constraints of data loss and high experimental costs.

[0004] In addition, other prior art solutions such as machine learning systems based on differential geometric modeling (US20210027862A1) or Bayesian statistics-driven iterative optimization methods (US20240029834A1) have improved prediction accuracy in specific tasks, but have not formed a complete "prediction-verification-feedback" closed-loop system, and lack the unified integration of model uncertainty priority exploration and mechanism diversity. Especially in rare disease drug discovery, data gaps are serious, experimental resources are limited, and existing methods are difficult to dynamically optimize screening strategies, making it difficult to effectively balance short-term efficacy prediction and long-term mechanism discovery goals.

[0005] Therefore, the existing drug intelligent screening technology has the technical problems of insufficient integration of model uncertainty, lack of consideration of sample diversity, and disconnection between prediction and verification, which makes the existing drug intelligent screening technology inefficient and unreliable in low-data and high-cost scenarios. SUMMARY

[0006] The present application provides an iterative drug screening method and system based on active learning, which can solve the technical problems of inefficient and unreliable screening of existing drug intelligent screening technology in low-data and high-cost scenarios.

[0007] The present application provides an iterative drug screening method based on active learning, comprising: Obtain the representation vector of the candidate drug set, which is a low-dimensional dense vector extracted from a pre-constructed knowledge graph.

[0008] Use the representation vector as the input of the preset drug prediction model and the key attribute prediction model, and perform multiple rounds of active learning operations.

[0009] In response to the current round of iteration satisfying the preset iteration termination condition, the current round of screening drug set is taken as the target screening drug set and output.

[0010] Each round of active learning operation includes: Perform a preliminary screening operation to screen a preliminary drug set from the candidate drug set based on the effective probability, key attribute high uncertainty features and diversity indicators obtained by clustering algorithm extracted from the input representation vector; update the association relationship in the knowledge graph based on the key attribute biological experiment results of the preliminary drug set; and update the drug prediction model and the key attribute prediction model using the updated knowledge graph to update the corresponding model parameters.

[0011] The screening operation is performed, and the initial screening drug set is used as the current candidate drug set. The efficacy probability of each drug is calculated by the drug prediction model after updating the model parameters. The highest-ranked drugs from the ranking results of efficacy probabilities are selected to form the screening drug set. The efficacy evaluation results of the corresponding drugs are determined based on the biological experimental results of the two key attributes fed back by the screening drug set. The screening drug set of the current round is removed from the candidate drug set, and the remaining drug set is used as the input for the next round of operation.

[0012] The high uncertainty feature of key attributes refers to the score obtained by quantifying the uncertainty of key attributes through the key attribute prediction model, which is used to characterize the confidence level of the model in predicting the attribute.

[0013] According to a second aspect, this invention provides an iterative drug screening system based on active learning, comprising: The drug feature collection module is used to obtain the representation vector of the candidate drug set. The representation vector is a low-dimensional dense vector extracted from a pre-constructed knowledge graph.

[0014] The active learning screening module is used to perform multiple rounds of active learning operations, using representation vectors as inputs to preset drug prediction models and key attribute prediction models.

[0015] The target drug output module is used to output the current round of screened drugs as the target screened drug set when the number of iterations in the current round meets the preset iteration termination condition.

[0016] The active learning screening module includes a drug initial screening module and a drug screening module.

[0017] The drug screening module is used to select a preliminary drug set from candidate drugs based on the efficacy probability extracted from the input representation vector, the high uncertainty features of key attributes, and the diversity index obtained through clustering algorithms; it updates the associations in the knowledge graph based on the biological experimental results of a class of key attributes fed back from the preliminary drug set; and it uses the updated knowledge graph to train the drug prediction model and the key attribute prediction model to update the corresponding model parameters.

[0018] The drug screening module is used to take the initial screening drug set as the current candidate drug set, calculate the efficacy probability of each drug through the drug prediction model after updating the model parameters, select the preset number of drugs with the highest efficacy probabilities from the ranking results to form the screening drug set; determine the efficacy evaluation results of the corresponding drugs based on the biological experimental results of the two key attributes fed back by the screening drug set; remove the screening drug set of the current round from the candidate drug set, and use the remaining drug set as the input for the next round of operation.

[0019] The high uncertainty feature of key attributes refers to the score obtained by quantifying the uncertainty of key attributes through the key attribute prediction model, which is used to characterize the confidence level of the model in predicting the attribute.

[0020] According to a third aspect, the present invention provides a computer device comprising: At least one processor; and a memory communicatively connected to the at least one processor; The memory stores instructions that can be executed by the at least one processor, which enables the at least one processor to perform any of the active learning-based iterative drug screening methods in the embodiments of the present invention.

[0021] According to another aspect of the present invention, a non-transitory computer-readable storage medium storing computer instructions is provided, wherein the computer instructions are used to cause a computer to execute any of the active learning-based iterative drug screening methods in the embodiments of the present invention.

[0022] The present invention provides an iterative drug screening method based on active learning, which is achieved through five core steps: representation vector acquisition, dual-model collaborative screening, knowledge feedback update, model iterative training, and loop termination judgment. Specifically, acquiring the representation vectors of the candidate drug set is used to extract low-dimensional dense vectors of drugs from a pre-constructed knowledge graph, capturing the essential characteristics of drugs in a structured manner. These representation vectors are then used as input to both the drug prediction model and the key attribute prediction model, performing multiple rounds of active learning to achieve intelligent screening and iterative optimization. Finally, in response to the current iteration number satisfying a preset iteration termination condition, the target drug set is output to ensure the convergence and efficiency of the screening process under limited resources.

[0023] In the entire technical solution, each round of active learning is executed collaboratively through two stages: initial screening and final screening. The initial screening stage selects a preliminary drug set based on the efficacy probability, high uncertainty features of key attributes, and diversity indicators extracted from the representation vector. The high uncertainty features of key attributes are obtained by quantifying the uncertainty of key attributes through a key attribute prediction model, which is used to characterize the confidence level of the model's prediction of that attribute. This addresses the problem of insufficient uncertainty assessment in model predictions mentioned in the background technology by explicitly quantifying the prediction confidence level, providing key information for screening decisions and solving the defect of low screening reliability caused by ignoring uncertainty in existing technologies. Furthermore, by using efficacy probability, high uncertainty features of key attributes, and diversity indicators together as the initial screening criteria, an evaluation dimension that takes into account therapeutic potential, exploration breadth, and mechanistic diversity is constructed. This effectively avoids the drawback of traditional methods getting stuck in local optima and improves the discovery capability of novel candidate drugs, addressing the problem of lack of consideration for the diversity of screening samples in the background technology. Subsequently, the knowledge graph associations are updated based on the results of biological experiments on one key attribute of the initial drug screening set. The updated knowledge graph is then used to retrain the dual model. Addressing the disconnect between computational prediction and experimental verification in the background technology, a two-way feedback loop between "dry experiments" and "wet experiments" is established, achieving dynamic fusion of experimental knowledge into the computational model. This overcomes the technical limitations of isolated prediction and verification, and the inability to iteratively optimize the screening strategy. In the screening stage, the updated model calculates the probability of effectiveness and selects the highest-ranking drugs to form the screening drug set. Drug effectiveness is then evaluated based on the results of biological experiments on two key attributes. The current round of screening drugs is removed from the candidate pool, and the remaining drug set serves as input for the next round. Iterative termination conditions control multiple rounds of operations. Addressing the low screening efficiency in low-data, high-cost scenarios in the background technology, this invention achieves automated and targeted exploration under limited experimental resource constraints, avoiding resource waste and ineffective iterations. Therefore, the technical solution of this invention solves the technical problems of insufficient screening efficiency and reliability in existing intelligent drug screening technologies under low-data, high-cost scenarios, improving the accuracy, novelty, and resource utilization efficiency of candidate drug screening. Attached Figure Description

[0024] Figure 1 This is a flowchart of an embodiment of the iterative drug screening method based on active learning according to the present invention; Figure 2 This diagram illustrates an iterative intelligent drug screening and validation framework based on active learning, according to other embodiments of the present invention. Figure 3 This is a structural block diagram of an iterative drug screening system based on active learning, according to some embodiments of the present invention. Figure 4 This is a block diagram of a computer device for implementing embodiments of the present invention. Detailed Implementation

[0025] The following description, in conjunction with the accompanying drawings, illustrates exemplary embodiments of the present invention, including various details to aid understanding. These details should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope of the invention. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.

[0026] During the development of this invention, the inventors, through extensive experiments and data analysis, revealed the intrinsic link between high treatment probability and high predictive uncertainty: a single screening strategy not only struggles to effectively assess the reliability of model predictions but also misses high-potential candidate drugs due to cognitive blind spots that ignore key attributes. Based on this relationship, the inventors innovatively proposed this technical solution, utilizing an active learning mechanism to achieve simultaneous improvement in screening accuracy and model generalization ability through dual-model collaborative prediction combined with experimental verification feedback, embodying the core concept of "learning through verification and iterative optimization."

[0027] Specifically, through comparative experiments, the invention team discovered that traditional screening methods based on a single efficacy prediction score have three technical flaws: (1) Lack of uncertainty assessment: Relying solely on prediction scores cannot identify the cognitive boundaries of the model, resulting in insufficient ability to explore potential new mechanism drugs; (2) Insufficient sample diversity: Continuous screening is prone to producing a candidate set with homogeneous structures, which weakens the coverage of the chemical space; (3) Model iteration lag: The prediction and verification links are separated, and a closed loop of knowledge accumulation cannot be formed.

[0028] These technical limitations lead to local optima in the screening process and hinder effective exploration under high-cost experimental constraints. However, the "probability-uncertainty" dual-dimensional evaluation method proposed in this invention can improve the information gain of a single round of screening; through real-time updates of the knowledge graph and model retraining, dynamic optimization of the screening strategy and continuous convergence of cognitive blind spots can be achieved. Experimental data show that this approach significantly improves the discovery rate of novel candidate drugs compared to traditional methods in rare disease drug screening.

[0029] Therefore, according to the first aspect, this invention provides an iterative drug screening method based on active learning, which can be applied to an intelligent drug discovery platform system (hereinafter referred to as the "system"). This system can be implemented via cloud-based Software as a Service (SaaS). This system can be deployed in a distributed computing environment, either locally or privately, to automate the entire screening process from candidate drug prediction to experimental validation feedback. Specifically, the system can be deployed in various hardware environments, including but not limited to: cloud computing clusters, high-performance computing (HPC) platforms, and local servers equipped with graphics processing units (GPUs). This flexible deployment architecture allows the system to meet the high-concurrency computing needs of large-scale virtual screening in pharmaceutical companies while also accommodating the privacy protection requirements of small-scale, precision research in research institutions.

[0030] like Figure 1 As shown, the method may include: Step S110: Obtain the representation vector of the candidate drug set. The representation vector is a low-dimensional dense vector extracted from a pre-constructed knowledge graph.

[0031] Among them, the representation vector refers to a fixed-dimensional numerical vector formed by embedding multi-source association information (such as molecular structure, target interaction, biological pathway, etc.) of drug entities in the knowledge graph through graph representation learning algorithms; low-dimensional dense vectors specifically refer to floating-point arrays whose dimensions are usually set between 128 and 512, and whose vast majority of dimensions contain non-zero effective information.

[0032] Specifically, the system can utilize a Graph Attention Network (GAT) or a Graph Convolutional Network (GCN). Deep learning models such as Generic Networks (GCN) perform multi-level aggregation of the neighborhood information of nodes in a knowledge graph to generate vectors with semantic representation capabilities.

[0033] For example, the system generates a 384-dimensional floating-point vector for each drug from a knowledge graph containing 120,000 drug entities and 450,000 relationships. The vector format is as follows: array (i.e.) (format), a single vector is approximately 3 kilobytes (KB) in size.

[0034] Step S120: Using the representation vector as input to the preset drug prediction model and key attribute prediction model, perform multiple rounds of active learning operations.

[0035] Among them, the high uncertainty feature of key attributes refers to the score obtained by quantifying the uncertainty of key attributes through the key attribute prediction model, which is used to characterize the confidence level of the model in predicting the attribute.

[0036] Furthermore, high uncertainty in key attributes refers to a quantitative assessment result with low prediction confidence obtained after forward reasoning and quantitative analysis of the key attributes of a drug (i.e., pharmacobiological characteristics that can be verified through specific biological experiments and have a significant impact on efficacy) using a key attribute prediction model (configured with uncertainty estimation units and adversarial training perturbation units). This characteristic is specifically manifested in the high dispersion and large information entropy of the model's predicted output probability distribution for a certain drug's key attribute, indicating a significant blind spot in the model's understanding of this sample. In this invention, high uncertainty in key attributes serves as the core driving indicator for active learning, guiding the system to prioritize the selection of such high-uncertainty drugs for experimental verification. This maximizes the model's information acquisition and generalization ability by minimizing experimental costs, ultimately solving the problems of low exploration efficiency and insufficient reliability caused by neglecting prediction uncertainty in traditional screening methods.

[0037] Each round of active learning operations includes: The initial screening process is performed, and a preliminary screening drug set is obtained from candidate drugs based on the efficacy probability extracted from the input representation vector, the high uncertainty features of key attributes, and the diversity index obtained through clustering algorithms. The relationships in the knowledge graph are updated based on the biological experimental results of a class of key attributes fed back by the preliminary screening drug set. The updated knowledge graph is then used to train the drug prediction model and the key attribute prediction model to update the corresponding model parameters.

[0038] One type of key attribute biological experiment refers to experimental methods that are low-cost and can quickly verify key attributes, such as cell bioluminescence detection experiments (which can verify the rhythm regulation attributes of 128 drugs within 96 hours).

[0039] Specifically, the system can calculate drug priority using a weighted fusion algorithm (such as the formula: comprehensive score = 0.7 × probability of effectiveness + 0.3 × key attribute high uncertainty feature), and then combine it with the K-means clustering algorithm (a classic unsupervised machine learning algorithm whose core goal is to automatically divide a set of data points into K clusters, so that the sum of the squared distances between each data point and the center of its cluster is minimized) to ensure the diversity of screening results.

[0040] For example, the system performs an initial screening of 5,000 candidate drugs, selects the top 500 drugs based on their comprehensive scores, and then selects 100 drugs representing different mechanisms of action from these 500 drugs through cluster analysis to form the initial screening drug set.

[0041] The screening operation is performed, and the initial screening drug set is used as the current candidate drug set. The efficacy probability of each drug is calculated by the drug prediction model after updating the model parameters. The highest-ranked drugs from the ranking results of efficacy probabilities are selected to form the screening drug set. The efficacy evaluation results of the corresponding drugs are determined based on the biological experimental results of the two key attributes fed back by the screening drug set. The screening drug set of the current round is removed from the candidate drug set, and the remaining drug set is used as the input for the next round of operation.

[0042] Among them, Category II key attribute biological experiments refer to experimental methods that are costly and used for final validity verification (such as animal model experiments or preclinical trials).

[0043] Specifically, the system can sort the efficacy probabilities in descending order using a sorting algorithm (such as quicksort) and select the top 10 drugs to enter the high-cost experimental verification stage.

[0044] For example, the system selects the 10 drugs with the highest probability of effectiveness from 100 initial screening drugs for animal experiment verification, and removes the drugs that pass the verification in this round from the candidate pool.

[0045] Step S130: In response to the current iteration number satisfying the preset iteration termination condition, the current round of screened drug set is used as the target screened drug set and output.

[0046] The iteration termination conditions include reaching a preset upper limit for the number of iterations (e.g., 5 rounds) or the change in the model performance evaluation index (e.g., the screening hit rate) being less than a preset threshold (e.g., the difference in hit rates between two consecutive rounds is less than 2%).

[0047] Specifically, the system can monitor the current round using a counter and compare it with a preset threshold. If the condition is met, the iteration will automatically terminate and the final result will be output.

[0048] For example, after completing the 5th round of screening, the system detects that the hit rate change is only 1.5%, which is lower than the 2% threshold. It then terminates the iteration and outputs the drug set of this round of screening as the final target drug set.

[0049] In some embodiments, the system can predict the probability of a drug possessing a specific key attribute from drug characterization using the following formula (a): (a) Formula (a) is the key attribute prediction formula, used to calculate the probability that candidate drug d possesses a certain key attribute (such as regulating cell rhythm). Wherein, This represents the predicted probability value of drug d possessing a key attribute, and its value ranges from [value missing]. ; This represents the low-dimensional dense representation vector of drug d extracted from the knowledge graph; This represents a pre-trained multilayer perceptron ( The Model for Multi-Level Processing (MLP) is responsible for processing representation vectors. Mapped to probability values.

[0050] Next, in this embodiment, the system can also calculate a more robust prediction probability distribution by combining model ensemble and adversarial training perturbation using formula (b): (b) Formula (b) is the formula for calculating the ensemble prediction distribution, used to fuse the prediction results of multiple models to estimate uncertainty. Wherein, This represents the final predicted probability distribution of drug d; This represents the set of E different model parameters obtained through Monte Carlo Dropout sampling; This indicates that the parameter θ is in the labeled training set. The posterior probability under this ensemble framework is often assumed to be uniformly distributed; This indicates the use of a model with parameter θ to characterize the drug after applying a perturbation δ. The predicted distribution obtained by forward inference; δ is the value applied to the original representation. The adversarial perturbation vector on.

[0051] Furthermore, in this embodiment, the system can also generate the adversarial training perturbation δ using formula (c): (c) Equation (c) is the perturbation-resistant generation formula, employing the Fast Gradient Sign Method (FGSM). Wherein, The disturbance intensity coefficient is a preset small positive scalar used to control the magnitude of the disturbance. Represents the model prediction loss function Relative to input drug characterization The gradient; The model's predicted output for drug d is y, where y is the true label (which can be predicted using the current model when the perturbation is generated). As an alternative); This is the sign function, used to determine the direction of the gradient.

[0052] Furthermore, in this embodiment, the system can also calculate the uncertainty score based on adversarial training Bayesian estimation using formula (d): (d) in, Calculated by the following formula: ; and Formula (d) is the normalized uncertainty scoring formula. Wherein, Indicates in the labeled dataset The above assumes that the expected information gain from labeling drug d is measured by the difference in the log probability score G of the predicted distribution before and after labeling. This represents the expectation of the possible true label y of drug d; It is a log-probability score function under a strict and appropriate scoring rule, used to evaluate the determinism of the predicted distribution; It is The absolute value is normalized to The final uncertainty score after the interval facilitates comparisons between different drugs.

[0053] Therefore, by combining the above formulas (a) to (d), the system can first make basic predictions on the key attributes of drugs, then obtain a more robust prediction distribution through integration and adversarial training techniques, and finally quantify the uncertainty score of the model's prediction results for each drug, providing core quantitative basis for subsequent diversity screening based on active learning.

[0054] In some embodiments, to improve the diversity of drug screening, the system employs a K-Means-based diversity selection algorithm. The system can calculate the uncertainty change vector of drug d using the following formula (e): (e) Formula (e) is used to quantify the local influence pattern of drug d on the model's predicted distribution. The vector representing the uncertainty change of drug d is an N-dimensional vector; Indicates in the labeled dataset Above, assuming a reference drug is additionally listed. Afterwards, the uncertainty score of drug d The change value reflects the difference between drug d and the reference drug. Interactions affecting the model; It is a set of N reference drugs randomly selected from a large pool of unlabeled drugs to construct a unified evaluation benchmark. The value of N can be, for example, 100.

[0055] Next, in this embodiment, the system can also calculate the uncertainty change distance between any two drugs d and d′ using formula (f): (f) Formula (f) is the distance calculation formula based on vector norm. Wherein, This represents the distance of uncertain change between drug d and d′; The L2 norm of a vector, also known as the Euclidean distance, is used to measure the distance between two vectors with uncertain changes. and The straight-line distance in vector space. The larger this distance value, the greater the difference in the influence patterns of the two drugs on the model's predicted behavior.

[0056] Furthermore, in this embodiment, the system can also calculate the drug characterization distance between drug d and d′ using formula (g): (g) Formula (g) is the distance calculation formula based on cosine similarity. Wherein, This represents the drug characterization distance between drug d and d′; and Let d and d′ represent the final representation vectors output by the Lth layer of the graph neural network, respectively. This represents the cosine similarity between two representation vectors. Representing vectors The distance measures the similarity of drugs in terms of their essential characteristics such as chemical structure, target, and function; the larger the distance value, the less similar the drugs are.

[0057] Furthermore, in this embodiment, the system can also perform linear weighted fusion of the two distances mentioned above using formula (h) to obtain the comprehensive distance between drugs: ;(h) Formula (h) is the weighted fusion formula. Wherein, This represents the final combined distance between drugs d and d′; and These are the drug characterization distance and uncertainty variation distance after minimum-maximum normalization, respectively. Normalization aims to eliminate the influence of different distance dimensions. The weighting coefficients are preset and satisfy the following conditions: This is used to balance the importance of the two distances in the overall metric. The higher the value, the more emphasis is placed on the inherent differences in drug characteristics during screening; The smaller the value, the more emphasis is placed on the differences in the drug's effect on the model. Weighting coefficient The value can be determined using a grid search method, with a range of 0.3 to 0.7, and an optimal value of 0.5.

[0058] Furthermore, in this embodiment, the system can also calculate the preliminary screening score for each drug using formula (i): (i) Formula (i) is the drug scoring formula. Wherein, This indicates the initial selection score for drug d; This represents the probability of drug d taking effect as output by the drug prediction model; This indicates that the key attribute of drug d is characterized by high uncertainty; the λ here is the same weighting coefficient as λ in formula (h), ensuring consistency in the evaluation strategy. This score integrates the drug's "therapeutic potential" (high... ) and "information value" (high) ).

[0059] Finally, in this embodiment, the system performs K-Means clustering based on the comprehensive distance matrix calculated by formula (j) using formula (k): (j) Formula (j) describes the clustering process. Wherein, This represents the set of B cluster centroids obtained after performing the K-Means clustering algorithm; is the inter-drug distance matrix calculated using formula (h); B is the preset final number of drugs to be selected, i.e., the number of clusters. After clustering, the system selects the drug closest to the cluster center from each cluster as the representative of that class. , where c is the cluster center.

[0060] Therefore, through the combination and synergy of the above formulas (e) to (j), the system can first quantify the dual differences of drugs in influencing model behavior and their own characteristics, and then perform diversity clustering by fusing distance. Finally, it can select a representative set of drugs from high-scoring drugs that have both high potential and high uncertainty, and also have diversity in characteristics and influence patterns, thereby greatly optimizing the experimental efficiency and information gain of the active learning loop.

[0061] In other embodiments, an iterative intelligent drug screening and validation framework based on active learning is used, such as... Figure 2 As shown. The iterative intelligent drug screening and validation framework based on active learning is as follows: Figure 2As shown, this framework constructs a complete iterative optimization closed loop through the tight coupling and bidirectional feedback of "dry experiment" computational prediction and "wet experiment" biological verification. Specifically, the dry experiment part on the left side of the framework uses the "biomedical knowledge graph" as the basic data source, extracts the representation vector of each drug in the "candidate drug set D", and inputs it into the parallel "drug recommendation model" and "key attribute prediction model". The drug recommendation model is responsible for predicting the "effective probability" of each drug for the target disease; the key attribute prediction model quantifies the "uncertainty" of the model's prediction of the drug's "key attributes" through uncertainty estimation methods such as "adversarial Bayes", which must meet the two conditions of "low biological detection cost" and "significant influence on drug effectiveness". Subsequently, the system enters the "K-Means-based drug multi-selection recommendation" stage, which selects the "initial drug set H" from the candidate set based on the "evaluation" criteria - that is, simultaneously satisfying "high effective probability", "high uncertainty of key attributes" and "high diversity" (the "drug representation distance" and "change distance" calculated by the latter two are used as the clustering basis). This set is not directly output but immediately fed into the wet experiment section on the right for "biological experimental validation (low-cost detection; determination of key attributes)". This experiment aims to rapidly determine the true values ​​of key attributes for each drug in the initial drug pool at low cost, completing the step of "key attributes being determined". The results determined by the experiment are fed back to the dry experiment as "result feedback": on the one hand, "updating and supplementing" the biomedical knowledge graph to enrich its semantic associations; on the other hand, it is used for "model optimization", that is, retraining the drug recommendation model and the key attribute prediction model to achieve "enhancement". The updated model will "re-predict" the "effective probability" of the "initial drug pool H" and sort and screen it only according to the criterion of "high effective probability" to obtain the "screened drug pool H". This high-confidence set will then undergo more in-depth "biomedical experimental validation (target exploration)" and even "clinical experimental validation" for final effectiveness confirmation and "mechanism analysis", with the ultimate goal of "discovering new drugs". After one round of validation, the "screened drug pool H" will be "removed" from the original "candidate drug pool D", and the remaining drugs will enter the next iteration. This cycle continues until the preset iteration termination condition is met. This framework improves the efficiency and reliability of drug screening with limited experimental resources by combining priority experimental validation of high-uncertainty samples with iterative model optimization.

[0062] For example, consider the search for neuroinflammatory inhibitors targeting Alzheimer's disease. The system extracts representation vectors for 5000 candidate compounds from a knowledge graph integrating relationships between drug targets, inflammatory pathways, and blood-brain barrier permeability. In the first round of dry experiments, a drug recommendation model predicts the "effective probability" of each compound's anti-neuroinflammatory activity, while a key attribute prediction model quantifies the uncertainty in predicting the key attribute of "inhibiting specific microglial activation pathways." The system, considering high probability, high uncertainty, and structural diversity, selects 200 compounds as an initial drug pool using K-Means clustering. A wet experiment is then conducted: using low-cost, high-content cell imaging technology, the inhibitory effect of these 200 compounds on the target pathway is measured within 96 hours, thus determining the true values ​​of the key attribute. The experimental results are fed back to the system to update the knowledge graph (e.g., establishing a new association between compound A and the pathway) and optimize the two prediction models. Based on the optimized models, the system recalculates the effective probability of these 200 compounds, selecting only the 20 with the highest probabilities for the final drug pool. Next, these 20 lead compounds will undergo high-cost animal model experiments (wet assay phase 2) to verify their effectiveness in improving cognitive function in model mice. After this round, these 20 compounds will be removed from the original pool of 5000 candidates. The remaining 4980 compounds will be screened again in the next iteration using an enhanced knowledge graph and a more precise model. This process will continue until, after five iterations, two highly effective lead compounds with novel mechanisms of action are successfully discovered and validated.

[0063] Therefore, according to the above implementation method, the system achieves its goals through five core steps: representation vector acquisition, dual-model collaborative screening, knowledge feedback update, model iterative training, and loop termination judgment. Specifically, acquiring the representation vectors of the candidate drug set is used to extract low-dimensional dense vectors of drugs from a pre-constructed knowledge graph, capturing the essential characteristics of drugs in a structured manner. These representation vectors are then used as input to the drug prediction model and the key attribute prediction model, performing multiple rounds of active learning to achieve intelligent screening and iterative optimization. Finally, in response to the current iteration number satisfying a preset iteration termination condition, the system outputs the target drug set for screening, ensuring the convergence and efficiency of the screening process under limited resources.

[0064] Specifically, each round of active learning is executed collaboratively through a two-stage process of initial screening and final selection. The initial screening stage selects a preliminary drug set based on the efficacy probability extracted from the representation vector, high uncertainty features of key attributes, and diversity indicators. The high uncertainty features of key attributes are obtained by quantifying the uncertainty of key attributes using a key attribute prediction model, which characterizes the confidence level of the model's prediction for that attribute. This addresses the problem of insufficient uncertainty assessment in model predictions described in the background art by explicitly quantifying prediction confidence, providing crucial information for screening decisions and resolving the low screening reliability caused by ignoring uncertainty in existing technologies. Furthermore, by using efficacy probability, high uncertainty features of key attributes, and diversity indicators together as initial screening criteria, an evaluation dimension that considers therapeutic potential, exploration breadth, and mechanistic diversity is constructed. This effectively avoids the drawback of traditional methods getting trapped in local optima, addressing the lack of consideration for sample diversity in the background art, and improving the discovery capability of novel candidate drugs. Subsequently, the knowledge graph associations are updated based on the results of biological experiments on one key attribute of the initial drug screening set. The updated knowledge graph is then used to retrain the dual model. Addressing the disconnect between computational prediction and experimental verification in the background technology, a two-way feedback loop between "dry experiments" and "wet experiments" is established, achieving dynamic fusion of experimental knowledge into the computational model. This overcomes the technical limitations of isolated prediction and verification, which prevent iterative optimization of the screening strategy. During the screening phase, the updated model calculates the probability of effectiveness and selects a preset number of drugs with the highest ranking to form the screening drug set. Drug effectiveness is then evaluated based on the results of biological experiments on two key attributes. The current round of screening drugs is removed from the candidate pool, and the remaining drug set serves as input for the next round. Iterative termination conditions control multiple rounds of operations. Addressing the low screening efficiency in low-data, high-cost scenarios in the background technology, this solution achieves automated and targeted exploration under limited experimental resource constraints, avoiding resource waste and ineffective iterations. Therefore, this implementation solves the technical problems of insufficient screening efficiency and reliability in existing intelligent drug screening technologies under low-data, high-cost scenarios, improving the accuracy, novelty, and resource utilization efficiency of candidate drug screening.

[0065] In some embodiments, the drug prediction model is configured with a node representation learning unit and a probability prediction unit; the step of extracting the efficacy probability from the input representation vector includes: The node representation learning unit extracts the representation vectors of each drug in the candidate drug set from the knowledge graph. The representation vectors are obtained by aggregating the feature information of adjacent nodes in the knowledge graph.

[0066] Among them, the aggregation operation refers to the computational process of weighted fusion of the feature information of the target node's directly connected neighboring nodes through the message passing mechanism in the graph neural network.

[0067] Specifically, the system employs a multi-head graph attention mechanism. First, it calculates the attention coefficients between the target node and each of its neighboring nodes. Then, it performs a weighted summation of the neighboring node features based on these coefficients. Finally, it concatenates the results from multiple attention heads and performs a linear transformation to obtain the updated node representation. For example, the system uses a graph attention network (GAT) with eight attention heads to aggregate the features of an average of 15 neighboring nodes for each drug, generating a feature vector with a dimension of 256. The computation time for the representation of a single drug is approximately 3 milliseconds (ms).

[0068] The extracted representation vector is input into the probability prediction unit, which performs nonlinear transformation and mapping on the representation vector to calculate the efficacy probability value of each drug.

[0069] Nonlinear transformation and mapping refers to the process of converting a high-dimensional input feature vector into a scalar output representing probability values ​​through a neural network containing multiple fully connected layers and activation functions.

[0070] Specifically, the system uses a three-layer fully connected network structure, where the hidden layer contains 128 neurons and employs... (Rectified Linear Unit, a commonly used nonlinear function that sets input values ​​less than zero to zero and leaves values ​​greater than or equal to zero unchanged) activation function; the output layer contains one neuron and uses... (A method for mapping arbitrary real number inputs to) The activation function (a nonlinear function over an interval) ultimately compresses the output value to between 0 and 1 as the probability of effectiveness. For example, the probability prediction unit performs parallel computation on the representation vectors of a batch of 256 drugs, completing all inference in 8 milliseconds on an NVIDIA T4 graphics processor (GPU), with the output probability value retained to four decimal places.

[0071] Therefore, according to the above implementation method, the system can efficiently learn the complex relationship features of drug nodes from the knowledge graph and accurately predict their bioactivity probability through differentiable neural network transformation, providing a reliable quantitative basis for subsequent active learning screening.

[0072] In some embodiments, the system can extract the textual feature vector of the disease using the model shown in formula (k) below: ; (k) Wherein, formula (k) is the formula for extracting disease text features. Indicated by large language model Diseases extracted The text feature embedding vector has a preset dimension of 128. Indicates about disease Structured or unstructured textual descriptions are used as input text for the model.

[0073] Next, in this embodiment, the system can also calculate the molecular structure representation embedding of the drug using formula (l): ;(l) Formula (l) is the formula for drug substructure embedding and polymerization. The molecular structure of drug d is represented by an embedding vector; This represents the unique set of substructures obtained after decomposing drug molecule d using the BRICS method (Breaking of Retrosynthetically Interesting Chemical Substructures, a computational chemistry algorithm that breaks down complex drug molecules into smaller, chemically synthesizable substructures). The embedding vector corresponding to substructure s is derived from a learnable global substructure embedding matrix. ; This represents the number of substructures contained in drug d, and aggregation is achieved by summing and averaging.

[0074] Furthermore, in this embodiment, the system can also define and initialize the feature vectors of various types of nodes in the heterogeneous graph using formula (m): ;(m) Where formula (m) is the formula for defining the initial features of a node. H0 represents the set of initial feature vectors for all nodes in the graph; These represent the initial feature vectors of drug node d, disease node x, and target node p, respectively. These are attribute feature vectors of drugs, diseases, and targets obtained through one-hot encoding, respectively; "" indicates a vector concatenation or element-wise addition operation, used to fuse multi-source features.

[0075] Then, in this embodiment, the system employs a Graph Attention Network (GAT) for multi-level information propagation. After L layers of propagation, the system fuses the multi-view representations of each node using formula (n) to obtain its final representation: ; (n) Wherein, formula (n) is the final representation fusion formula for nodes. This represents the final representation vector of any type of node (drug, disease, or target) after L-layer GAT learning (a deep learning method that operates on graph-structured data, uses an attention mechanism to assign differential weights to different neighbor nodes, thereby aggregating information to learn node representations). and These represent the view-specific representation vectors learned by the node in the drug-disease view and the drug-target-disease ternary interactive view, respectively; "+" indicates vector addition, used to integrate semantic information from different views.

[0076] Subsequently, in this embodiment, the system predicts the probability of effective treatment of disease x by drug d using formula (o): ;(o) Here, formula (o) is a formula for predicting the probability of treatment based on cosine similarity. Let represent the predicted treatment probability of drug d for disease x, with a range of . After the follow-up The function can be mapped to standard probability values. ; and These are the final representation vectors of drug d and disease x, respectively; The L2 norm of a vector is a mathematical operation used to calculate the square root of the sum of the squares of the elements in a vector, which measures the "length" or "modulus" of the vector.

[0077] Finally, in this embodiment, the system predicts the correlation probability between drug d and target p using formula (p): (p) Wherein, formula (p) is the formula for predicting the probability of drug-target binding. This represents the predicted probability of drug d being associated with target p; This is the final representation vector of the target point p.

[0078] Therefore, through the combination and synergy of the above formulas (k) to (p), the system can effectively extract and fuse multimodal features (including text, molecular structure and attributes) of drugs, diseases and targets from the knowledge graph, model the high-order interaction relationships between them through graph attention mechanism, and finally achieve high-precision prediction of drug efficacy and target binding activity in a unified embedding space, laying a reliable computational foundation for subsequent active learning iterative screening.

[0079] In some embodiments, the key attribute prediction model is configured with an uncertainty estimation unit and an adversarial training perturbation unit; the step of extracting high uncertainty features of key attributes from the input representation vector includes: By using uncertainty estimation units to perform multiple forward inferences on the representation vectors of each drug in the candidate drug set, the predicted probability distribution of key attributes is obtained.

[0080] Among them, multiple forward inference refers to a sampling method that randomly disables some neural network neurons during the model inference stage by using Monte Carlo Dropout (a technique that randomly disables some neurons in the inference stage of a deep neural network and performs multiple forward propagations to approximate the model's prediction of the posterior distribution and perform uncertainty estimation), thereby obtaining multiple differential prediction results; the predicted probability distribution refers to the discrete probability distribution of the presence or absence of key attributes obtained by statistically analyzing the results of multiple inferences.

[0081] Specifically, the system keeps the discard layer active during inference, independently repeating inference on the same drug representation vector, recording the binary classification probability value (i.e., the probability of the key attribute being present) of each output, and finally forming a set of probability values. For example, the system performs 100 repeated inferences on each drug, obtaining a set of 100 probability values. ,in The statistical characteristics of this set constitute the predicted probability distribution.

[0082] By training the perturbation unit adversarially, a preset perturbation is applied to the representation vector, and forward inference is performed multiple times on the perturbated representation vector.

[0083] The preset perturbation refers to a targeted noise vector calculated based on the Fast Gradient Sign Method (FGSM), which is an attack and defense technique that efficiently generates adversarial examples by calculating the gradient direction of the model loss function with respect to the input data and adding a small perturbation along this direction. Its direction is the gradient direction that maximizes the model's prediction error.

[0084] Specifically, the system first calculates the gradient of the representation vector with respect to the model loss function, and then applies the gradient at a fixed step size ( A perturbation vector is generated along the direction of the gradient sign. This perturbation is then superimposed onto the original representation vector before re-inference sampling. For example, the system applies a perturbation vector to a 256-dimensional drug representation vector. Adversarial perturbations with a norm constraint of 0.01 result in changes in the vector representation after perturbation controlled within a certain range. Within the range, another 100 Monte Carlo samples were then taken.

[0085] Based on multiple predicted probability distributions obtained before and after the perturbation, the high uncertainty characteristics of the key attributes of each drug are obtained by information entropy quantification.

[0086] Among them, information entropy quantification calculation refers to the use of the Shannon entropy formula. A mathematical method for measuring the uncertainty of a probability distribution; the larger the calculated value, the more dispersed the distribution and the higher the uncertainty.

[0087] Specifically, the system calculates the information entropy values ​​of the original probability distribution and the perturbated probability distribution separately, and takes the maximum of the two as the final uncertainty score for the drug. For example, if the original probability distribution entropy value of a drug is 0.52 bits and the perturbated distribution entropy value is 0.61 bits, then the system takes the maximum value of 0.61 as the uncertainty score, and the score range is normalized to [range missing]. Interval.

[0088] Therefore, according to the above implementation method, the system can achieve a dual uncertainty assessment mechanism ( This effectively identifies the confidence level of the model in predicting key attributes, providing a quantitative basis for sample selection in active learning.

[0089] In some embodiments, a preliminary drug set is obtained from candidate drugs based on the efficacy probability extracted from the input representation vector, high uncertainty features of key attributes, and diversity indicators obtained through clustering algorithms, including: Based on a preset weighted fusion rule, the probability of effectiveness and the high uncertainty features of key attributes are fused to generate a comprehensive score for each drug.

[0090] The weighted fusion rule refers to the mathematical expression that combines two evaluation indicators with different dimensions into a single score using a linear weighting method. Its general formula is: ,in and Preset weighting coefficients and satisfying .

[0091] Specifically, the system sets configuration parameters. We perform a weighted calculation on the two indicators for each drug and normalize the results to an integer range of 0 to 100.

[0092] For example, if a drug has a probability of effectiveness of 0.85 and an uncertainty score of 0.72, then its... The score is 81 after rounding.

[0093] The drugs in the candidate drug set are ranked according to the comprehensive score, and the top-ranked drugs are selected to form a candidate subset.

[0094] Among them, "ranked high" refers to drugs that are in the top 20% after being sorted in descending order of comprehensive score.

[0095] Specifically, the system uses a quicksort algorithm to sort the drugs by their overall scores and selects the top 20% of drugs to proceed to the next stage of screening. For example, from a set of 1,000 candidate drugs, the system selects the top 200 drugs by overall score to form a candidate subset, a process that takes approximately 50 milliseconds on the central processing unit (CPU).

[0096] Based on the characterization vectors and uncertainty features of each drug in the candidate subset, a diversity clustering algorithm is used to group the candidate subset.

[0097] Among them, the diversity clustering algorithm specifically refers to the improved grouping algorithm based on K-means clustering (an unsupervised machine learning algorithm that divides data points into K clusters through iterative optimization, minimizing the sum of squared distances between each point and the center of its cluster). This algorithm considers both the Euclidean distance of drug characterization vectors and the cosine similarity of uncertainty features.

[0098] Specifically, the system first performs principal component analysis (PCA) on the 200-dimensional drug representation vectors to reduce the dimensionality to 50 dimensions. PCA is a classic statistical method that uses orthogonal transformations to linearly map the original features onto a new set of orthogonal coordinate axes to achieve dimensionality reduction, eliminate correlations between features, and retain the main variation information of the data. Then, it performs cluster analysis using a weighted distance metric, with the number of clusters, k, automatically determined based on the sample size. For example, the system divides 200 candidate drugs into 15 clusters, each containing 8-20 drugs. The silhouette coefficient of the clustering process reaches 0.68, indicating good grouping performance.

[0099] Representative drugs are selected from each group and combined to form the initial screening drug set.

[0100] Among them, the representative drug refers to the drug that is closest to the cluster center and has the most representative uncertainty characteristics in its group.

[0101] Specifically, the system calculates the weighted distance from each drug within a group to the cluster center and selects the two drugs with the smallest distance as representatives of that group. For example, by selecting two representative drugs from each of the 15 groups, a preliminary screening drug set of 30 drugs is finally formed, covering 10 different mechanisms of action.

[0102] Therefore, according to the above implementation method, the system can effectively improve the mechanism diversity and structural representativeness of the screening results while ensuring the therapeutic potential of the drug through a multi-stage screening strategy, providing a high-quality candidate set for subsequent experimental verification.

[0103] In some embodiments, the knowledge graph is configured with drug entities, target entities, specific biomedical concept entities, and relationships between these entities from multiple biomedical databases; the relationships in the knowledge graph are updated based on biological experimental results of a class of key attributes fed back from the initial drug screening set, including: The true values ​​of key attributes in the experimental verification results are transformed into standardized knowledge triples.

[0104] In this context, a standardized knowledge triple refers to a data structure that follows the Resource Description Framework (RDF) specification and uses the "Subject-Predicate-Object" form to represent the relationships between entities.

[0105] Specifically, the system converts records in the experimental data table into RDF triples using data mapping rules. The subject is the drug entity identifier, the predicate is the key attribute relation predicate, and the object is the attribute value obtained from experimental verification. For example, if cell experiments confirm that the drug "DB00126" has the property of regulating circadian rhythms (p-value < 0.05), the system generates a normalized triple: .

[0106] Add knowledge triples to the knowledge graph to supplement or correct the association between drug entities and key attributes.

[0107] The addition operation includes incremental updates to the knowledge graph. When triples with the same subject and predicate exist, an overwrite update is performed; otherwise, an insertion operation is performed.

[0108] Specifically, the system executes the MERGE statement using the Cypher query language of the graph database to achieve idempotent writing of triples, ensuring data consistency. For example, in the Neo4j graph database, the following is executed: MERGE (d:Drug {id:"DB00126"}) MERGE (p:Property {name:"hasCircadianEffect"}) MERGE (d)-[r:HAS_PROPERTY]->(p) SET r.value ="True".

[0109] The representation vectors of drug entities, target entities, and specific biomedical concept entities are recalculated based on the updated knowledge graph data.

[0110] Among them, recalculation refers to generating representation vectors for all nodes of the updated knowledge graph based on the dynamic graph neural network algorithm.

[0111] Specifically, the system employs a graph convolutional network (PCNN). The Generic Networking (GCN) algorithm is a deep learning method that operates on graph-structured data and learns effective node representations end-to-end by aggregating feature information from the node itself and its neighbors. It uses 64 hidden layer neurons with a learning rate of 0.001 and performs 50 training iterations to generate 128-dimensional node representation vectors. For example, after adding 500 triples to the knowledge graph, the system takes approximately 15 minutes on an NVIDIA T4 GPU to update the representation vectors of all 120,000 nodes in the graph, with each node vector requiring approximately 2KB of storage.

[0112] The true values ​​of key attributes are objective measurement results or qualitative judgments obtained through specific biological experiments. Specific biological experiments include, but are not limited to, standardized experimental methods such as cell viability testing, Western blotting, and enzyme-linked immunosorbent assay (ELISA).

[0113] Specifically, for quantitative measurement results, the system records the specific values ​​and confidence intervals; for qualitative judgments, the system uses standardized terminology for description. For example, in quantitative measurement: the diurnal rhythm regulation intensity value of drug "DB00126" is... (Mean ± Standard Deviation); Qualitative judgment: The binding activity of drug "DB00345" to target "P12345" is "strong binding".

[0114] Therefore, according to the above implementation method, the system can realize the automatic conversion and integration of experimental data into knowledge graph, and maintain the timeliness of knowledge representation by dynamically updating node representation vectors, providing an accurate data foundation for subsequent iterative screening.

[0115] In some embodiments, the updated knowledge graph is used to train the drug prediction model and the key attribute prediction model to update the corresponding model parameters, including: Representation vectors for drug entities, target entities, and specific biomedical concept entities were re-extracted based on the updated knowledge graph data.

[0116] Here, "re-extraction" refers to regenerating the vector representation of all nodes based on the updated graph topology and node attributes using a graph neural network algorithm.

[0117] Specifically, the system employs the Graph Attention Network (GAT) algorithm, setting up two graph attention layers, each containing four attention heads, with a hidden layer dimension of 256, to learn representations for all nodes in the graph. For example, the system extracts representations from an updated knowledge graph containing 150,000 entity nodes and 500,000 relationships, completing the full graph computation in approximately 5 minutes on an NVIDIA A100 GPU, generating representation vectors for all nodes with a dimension of 128.

[0118] A new training sample set is constructed using the re-extracted representation vectors and their corresponding relationships.

[0119] The new training sample set refers to the set of positive and negative samples required for model training based on the updated map data, including drug-disease associations, drug-target effects, and other relationship pairs.

[0120] Specifically, the system employs a random sampling strategy, selecting positive samples from known associations and randomly sampling negative samples from unknown associations, maintaining a 1:3 ratio of positive to negative samples. For instance, the system constructs a dataset containing 100,000 training samples from the updated knowledge graph, including 25,000 positive samples and 75,000 negative samples. The file is stored in a formatted format and is approximately 25MB in size.

[0121] A new training sample set is used to perform end-to-end joint training on the drug prediction model in order to update the model parameters of the drug prediction model.

[0122] End-to-end joint training refers to simultaneously optimizing the model parameters of the node representation learning unit and the probability prediction unit, enabling the entire network to learn collaboratively.

[0123] Specifically, the system uses the Adam optimizer, with an initial learning rate of 0.001, a batch size of 256, and 100 training epochs. Every 20 epochs, the learning rate decays to 0.5 times its original value. For example, after end-to-end training, the drug prediction model's AUC on the validation set improved from 0.82 to 0.89, with model convergence taking approximately 45 minutes.

[0124] The key attribute prediction model is retrained using a new training sample set to update the model parameters.

[0125] Retraining refers to fine-tuning the model parameters using new data based on the pre-trained model to adapt to the updated data distribution of the knowledge graph.

[0126] Specifically, the system employs a transfer learning strategy, fixing the underlying parameters of the model and fine-tuning only the top-level classifier, training for 50 epochs with a small learning rate (0.0001). For example, after retraining, the F1 score of the key attribute prediction model on the test set improved from 0.76 to 0.83, and the model fine-tuning process took approximately 20 minutes.

[0127] Therefore, according to the above implementation method, the system can effectively perform incremental training on the prediction model based on real-time updated data of the knowledge graph, ensuring that the model always maintains the optimal performance state and providing accurate prediction capabilities for iterative screening.

[0128] In some embodiments, when performing active learning operations, the above method further includes: After completing each round of drug screening, determine whether the current round has met the preset iteration termination condition.

[0129] The judgment operation refers to the analysis process in which the system automatically compares the iteration count and performance index changes of the current round with a preset threshold.

[0130] Specifically, the system maintains an iteration counter and calculates the rate of change of key performance indicators (such as the proportion of effective candidate drugs) relative to the previous round after each round. These two values ​​are then logically ANDed with a preset threshold. For example, the system sets a maximum of 5 iterations and a performance change rate threshold of 2%. If, after completing the third round of screening, the system detects that the proportion of effective drugs has changed from 28% in the second round to 29.5%, with a change rate of 1.5% and the number of iterations has not reached the upper limit, then the system determines that the termination condition has not been met.

[0131] In response to the failure to meet the iteration termination condition, the target drug set obtained in this round is used as the candidate drug set for the next round, and the drug screening operation continues; or, in response to the failure to meet the iteration termination condition, the screened drug set obtained in the current round is determined as the target screened drug set.

[0132] Among them, the response operation refers to the conditional branch control flow executed by the system based on the judgment result, so as to realize the dynamic path selection of the filtering process.

[0133] Specifically, the system uses if-else conditional statements to implement branching logic: if the termination condition is not met, the current drug set is renamed to the candidate set for the next round, and the screening process is restarted; if the termination condition is met, the loop terminates and the results are output. For example, when the system detects that the effective drug ratios in rounds 4 and 5 are 30.1% and 30.3% respectively, and the change rate of 0.2% is lower than the threshold, the iteration is automatically terminated, the result of round 5 is used as the final output, and a report file containing detailed data from all 5 rounds of screening is generated (PDF format, Portable Document Format, a cross-platform file format developed by Adobe Systems and becoming an international standard in 2008).

[0134] The iteration termination conditions include at least one of the following: reaching the preset upper limit of the number of iterations or the change in the model performance evaluation index being less than a preset threshold.

[0135] Specifically, performance evaluation metrics include, but are not limited to, quantitative indicators such as the proportion of effective candidate drugs, the average confidence level of model predictions, or the diversity index of screening results.

[0136] For example, the system is set with dual termination conditions: a maximum of 6 iterations and an absolute change in the proportion of effective drugs between two consecutive iterations of less than 1.5%. When the proportion is 32.4% in the 5th round and 33.2% in the 6th round, the change rate of 0.8% meets the threshold requirement, and the system automatically terminates after the 6th round.

[0137] Therefore, according to the above implementation method, the system can achieve fully automated intelligent iterative control, effectively avoid invalid iterations while ensuring screening quality, and improve the efficiency and resource utilization of the drug discovery process.

[0138] Figure 3 This is a structural block diagram of an iterative drug screening system based on active learning according to an embodiment of the present invention.

[0139] like Figure 3 As shown, this active learning-based iterative drug screening system includes: The drug feature collection module 210 is used to obtain the representation vector of the candidate drug set. The representation vector is a low-dimensional dense vector extracted from a pre-constructed knowledge graph.

[0140] The active learning screening module 220 is used to perform multiple rounds of active learning operations, using the representation vector as input to the preset drug prediction model and key attribute prediction model.

[0141] The target drug output module 230 is used to output the current round of screened drug set as the target screened drug set when the number of iterations in the current round meets the preset iteration termination condition.

[0142] The active learning screening module includes a drug initial screening module 221 and a drug screening module 222.

[0143] The drug screening module 221 is used to select a preliminary screening drug set from candidate drugs based on the efficacy probability extracted from the input representation vector, the high uncertainty features of key attributes, and the diversity index obtained by the clustering algorithm; update the association relationship in the knowledge graph based on the biological experimental results of a class of key attributes fed back by the preliminary screening drug set; and use the updated knowledge graph to train the drug prediction model and the key attribute prediction model to update the corresponding model parameters.

[0144] The drug screening module 222 is used to take the initial screening drug set as the current candidate drug set, calculate the efficacy probability of each drug through the drug prediction model after updating the model parameters, select the preset number of drugs with the highest ranking from the ranking results of each efficacy probability to form the screening drug set; determine the efficacy evaluation result of the corresponding drug based on the biological experimental results of the two key attributes fed back by the screening drug set; remove the screening drug set of the current round from the candidate drug set, and use the remaining drug set as the input for the next round of operation.

[0145] The high uncertainty feature of key attributes refers to the score obtained by quantifying the uncertainty of key attributes through the key attribute prediction model, which is used to characterize the confidence level of the model in predicting the attribute.

[0146] The specific functions and examples of each module and submodule of the device in this embodiment of the invention can be found in the relevant descriptions of the corresponding steps in the above method embodiments, and will not be repeated here.

[0147] According to embodiments of the present invention, the above-described method of the present invention can be applied to a computer device and a readable storage medium.

[0148] Figure 4 A schematic block diagram of an example computer device 600 that can be used to implement embodiments of the present invention is shown. The computer device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The computer device can also represent various forms of mobile devices, such as personal digital assistants, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the invention described and / or claimed herein.

[0149] like Figure 4 As shown, the computer device 600 includes a computing unit 601, which can perform various appropriate actions and processes based on a computer program stored in a read-only memory (ROM) 602 or a computer program loaded from a storage unit 608 into a random access memory (RAM) 603. The RAM 603 may also store various programs and data required for the operation of the computer device 600. The computing unit 601, ROM 602, and RAM 603 are interconnected via a bus 604. An input / output (I / O) interface 605 is also connected to the bus 604.

[0150] Multiple components in computer device 600 are connected to I / O interface 605, including: input unit 606, such as keyboard, mouse, etc.; output unit 607, such as various types of monitors, speakers, etc.; storage unit 608, such as disk, optical disk, etc.; and communication unit 609, such as network card, modem, wireless transceiver, etc. Communication unit 609 allows computer device 600 to exchange information / data with other devices through computer networks such as the Internet and / or various telecommunications networks.

[0151] The computing unit 601 can be a variety of general-purpose and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 601 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, a digital signal processor (DSP), and any suitable processor, controller, microcontroller, etc. The computing unit 601 performs the various methods and processes described above, such as an active learning-based iterative drug screening method. For example, in some embodiments, an active learning-based iterative drug screening method can be implemented as a computer software program tangibly contained in a machine-readable medium, such as storage unit 608. In some embodiments, part or all of the computer program can be loaded and / or installed on the computer device 600 via ROM 602 and / or communication unit 609. When the computer program is loaded into RAM 603 and executed by the computing unit 601, one or more steps of an active learning-based iterative drug screening method described above can be performed. Alternatively, in other embodiments, computing unit 601 may be configured by any other suitable means (e.g., by means of firmware) to perform an iterative drug screening method based on active learning.

[0152] Various embodiments of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), systems-on-a-chip (SoCs), payload-programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may include implementations in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which may be a dedicated or general-purpose programmable processor, capable of receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.

[0153] The program code used to implement the methods of the present invention can be written in any combination of one or more programming languages. This program code can be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing device, such that when executed by the processor or controller, the program code causes the functions / operations specified in the flowcharts and / or block diagrams to be implemented. The program code can be executed entirely on the machine, partially on the machine, as a standalone software package partially on the machine and partially on a remote machine, or entirely on a remote machine or server.

[0154] In the context of this invention, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. Machine-readable media can include, but are not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0155] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device for displaying information to the user (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor); and a keyboard and pointing device (e.g., a mouse or trackball) through which the user provides input to the computer. Other types of devices can also be used to provide interaction with the user; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including sound input, voice input, or tactile input).

[0156] The systems and technologies described herein can be implemented in computing systems that include backend components (e.g., as a data server), or computing systems that include middleware components (e.g., an application server), or computing systems that include frontend components (e.g., a user computer with a graphical user interface or web browser through which a user can interact with implementations of the systems and technologies described herein), or any combination of such backend, middleware, or frontend components. The components of the system can be interconnected via digital data communication of any form or medium (e.g., a communication network). Examples of communication networks include local area networks (LANs), wide area networks (WANs), and the Internet.

[0157] Computer systems can include clients and servers. Clients and servers are generally located far apart and typically interact via communication networks. Client-server relationships are created by computer programs running on the respective computers and having a client-server relationship with each other. Servers can be cloud servers, servers in distributed systems, or servers incorporating blockchain technology.

[0158] It should be understood that the various forms of processes shown above can be used to reorder, add, or delete steps. For example, the steps described in this invention can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution disclosed in this invention can be achieved, and this is not limited herein.

[0159] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the principles of this invention should be included within the scope of protection of this invention.

Claims

1. An iterative drug screening method based on active learning, characterized in that, The method comprises the following steps: obtain a set of candidate drugs, and the set of candidate drugs is a low-dimensional dense vector extracted from a pre-constructed knowledge graph; input the set of candidate drugs into a preset drug prediction model and a key attribute prediction model, and perform a plurality of rounds of active learning operations; if the number of iterations in the current round meets a preset iteration termination condition, output the set of candidate drugs in the current round as a target set of candidate drugs; wherein, the active learning operation in each round comprises: perform a preliminary screening operation to screen a set of preliminary drugs from the set of candidate drugs according to an effective probability, a key attribute high uncertainty feature and a diversity index obtained by a clustering algorithm extracted from the input set of candidate drugs; update the association in the knowledge graph according to a type of key attribute biological experimental result fed back by the set of preliminary drugs; and train the drug prediction model and the key attribute prediction model using the updated knowledge graph to update the corresponding model parameters; perform a screening operation, wherein the set of preliminary drugs is used as a current set of candidate drugs, the drug prediction model after updating the model parameters is used to calculate the effective probability of each drug, a preset number of drugs with the highest ranking are selected from the ranking results of the effective probability to form a set of screened drugs, the effectiveness evaluation result of the corresponding drug is determined according to a type of key attribute biological experimental result fed back by the set of screened drugs, and the set of screened drugs in the current round is removed from the set of candidate drugs, and the remaining set of drugs is used as the input of the next round of operation; the key attribute high uncertainty feature refers to a score obtained by quantifying the uncertainty of the key attribute by the key attribute prediction model, which is used to represent the confidence level of the model in predicting the attribute.

2. The method of claim 1, wherein, the drug prediction model is configured with a node representation learning unit and a probability prediction unit; the step of extracting the effective probability from the input set of candidate drugs comprises: extracting the set of candidate drugs from the knowledge graph using the node representation learning unit, wherein the set of candidate drugs is obtained by aggregating the feature information of adjacent nodes in the knowledge graph; inputting the extracted set of candidate drugs into the probability prediction unit, performing nonlinear transformation and mapping of the set of candidate drugs by the probability prediction unit, and calculating the effective probability value of each drug.

3. The method of claim 1, wherein, the key attribute prediction model is configured with an uncertainty estimation unit and an adversarial training perturbation unit; the step of extracting the key attribute high uncertainty feature from the input set of candidate drugs comprises: using the uncertainty estimation unit to perform multiple forward inferences on the set of candidate drugs, and obtaining the prediction probability distribution of the key attribute; applying a preset perturbation to the set of candidate drugs by the adversarial training perturbation unit, and performing the multiple forward inferences again on the perturbed set of candidate drugs; based on the plurality of prediction probability distributions obtained before and after the perturbation, calculating the key attribute high uncertainty feature of each drug by information entropy quantification.

4. The method of claim 1, wherein, The active probability, the key attribute high uncertainty feature and the diversity index obtained by the clustering algorithm are extracted from the input characteristic vector, and a preliminary screening drug set is screened from the candidate drugs, including: According to the preset weighted fusion rule, the active probability and the key attribute high uncertainty feature are fused to generate a comprehensive score of each drug; According to the comprehensive score, the drugs in the candidate drug set are sorted, and the top-ranked drugs are selected to form a candidate subset; Based on the characteristic vector and the uncertainty feature of each drug in the candidate subset, a diversity clustering algorithm is used to group the candidate subset; The representative drugs are selected from each group to form the preliminary screening drug set.

5. The method of claim 1, wherein, The knowledge graph is configured with drug entities, target point entities, specific biomedical concept entities and the association relationship between entities from multiple biomedical databases; the association relationship in the knowledge graph is updated according to the key attribute biological experimental results fed back by the preliminary screening drug set, including: The key attribute true value in the experimental verification result is converted into a standardized knowledge triple; The knowledge triple is added to the knowledge graph to supplement or modify the association relationship between the drug entity and the key attribute; Based on the updated knowledge graph data, the representation vector of the drug entity, the target point entity and the specific biomedical concept entity is recalculated; The key attribute true value is the objective measurement result or qualitative judgment conclusion about the key attribute obtained through a key attribute biological experiment.

6. The method of claim 1, wherein, The updated knowledge graph is used to train the drug prediction model and the key attribute prediction model to update the corresponding model parameters, including: The representation vector of the drug entity, the target point entity and the specific biomedical concept entity is recalculated based on the updated knowledge graph data; A new training sample set is constructed using the recalculated representation vector and the corresponding association relationship; The drug prediction model is trained end-to-end using the new training sample set to update the model parameters of the drug prediction model; The key attribute prediction model is retrained using the new training sample set to update the model parameters of the key attribute prediction model.

7. The method of claim 1, wherein, When performing the active learning operation, the method further includes: After completing the drug screening operation of each round, it is judged whether the current round reaches the preset iteration termination condition; In response to not reaching the iteration termination condition, the screening drug set obtained in this round is used as the candidate drug set in the next round, and the drug screening operation is continued; or, in response to reaching the iteration termination condition, the screening drug set obtained in the current round is determined as the target screening drug set; The iteration termination condition includes at least one of reaching a preset upper limit of iteration times or a change in model performance evaluation index being less than a preset threshold.

8. An active learning based iterative drug screening system, characterized in that, including: A drug feature collection module is configured to obtain a characteristic vector of a candidate drug set, the characteristic vector being a low-dimensional dense vector extracted from a pre-constructed knowledge graph; The active learning screening module is configured to perform multiple rounds of active learning operation by taking the feature vectors as inputs of preset drug prediction models and key attribute prediction models. The target drug output module is configured to output a target screening drug set as a screening drug set of a current round when a number of iterations of the current round meets a preset iteration termination condition. The active learning screening module further includes a drug preliminary screening module and a drug screening module. The drug preliminary screening module is configured to screen a preliminary screening drug set from the candidate drugs according to an effective probability, key attribute high uncertainty features and a diversity index obtained by a clustering algorithm extracted from the input feature vectors, update an association relationship in the knowledge graph according to a first type of key attribute biological experiment result fed back by the preliminary screening drug set, and train the drug prediction models and the key attribute prediction models by using the updated knowledge graph to update corresponding model parameters. The drug screening module is configured to take the preliminary screening drug set as a current candidate drug set, calculate an effective probability of each drug by using the drug prediction models after the model parameters are updated, select a preset number of drugs with the highest rankings from a ranking result of the effective probabilities to form a screening drug set, determine an effectiveness evaluation result of a corresponding drug according to a second type of key attribute biological experiment result fed back by the screening drug set, remove the screening drug set of the current round from the candidate drug set, and take a remaining drug set as an input of a next round of operation. The key attribute high uncertainty features refer to scores obtained by quantifying uncertainty of the key attributes by using the key attribute prediction models, and are used to represent confidence levels of the models in predicting the attributes.

9. A computer device, comprising: The computer instructions are configured to enable the computer to perform the method according to any one of claims 1-7. The computer instructions are configured to enable the computer to perform the method according to any one of claims 1-7. ​ ​ 10. A non-transitory computer-readable storage medium having stored thereon computer instructions, wherein the computer instructions, when executed by a processor, cause the processor to perform the method of any one of claims 1-9. ​ ​

Images