A method for screening active molecules against carbapenem-resistant enterobacterium based on lightgbm algorithm, medium and equipment

Abstract

The present application belongs to the technical field of artificial intelligence assisted drug screening, and particularly relates to a method for screening active molecules against carbapenem-resistant enterobacteriaceae based on a LightGBM algorithm, a medium and an apparatus. The method constructs a LightGBM integrated learning framework, fuses ADME rules optimized for the outer membrane barrier characteristics of gram-negative bacteria and multi-target molecule docking verification, effectively solves the problem of data imbalance caused by the scarcity of active samples and the huge compound library, and significantly improves the hit rate of screening. The present application also provides a computer readable storage medium and an electronic device. By storing and executing the above program, the present application can quickly and standardizedly identify anti-CRE candidate molecules from a large number of compounds, greatly reducing the computing power cost and time cycle of new drug research and development. The present application is designed to overcome the bottleneck of gram-negative bacterial outer membrane permeation, and provides an efficient, accurate and intelligent screening tool for combating CRE super-bacterial infections.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence-assisted drug screening technology, specifically relating to a method, medium, and equipment for screening active molecules of carbapenem-resistant Enterobacteriaceae based on the LightGBM algorithm. Background Technology

[0002] Carbapenem-resistant Enterobacteriaceae (CRE) is a "superbug" of global concern, characterized by high pathogenicity, high transmission risk, and extremely limited clinical treatment options, making it a major pathogen threatening public health. Therefore, the development of highly effective and low-toxicity novel anti-CRE antibacterial drugs is urgently needed.

[0003] In the antimicrobial drug development process, lead compound screening is a crucial step determining drug discovery efficiency. Traditional high-throughput screening relies on experimentally validating the activity of massive numbers of compounds one by one, which has inherent drawbacks such as high experimental costs, long screening cycles, and huge consumption of human and material resources, making it difficult to meet the practical needs of rapid development of new antimicrobial drugs. In recent years, virtual screening technology based on machine learning has been widely used in the initial screening of antimicrobial molecules due to its advantages of low cost, high throughput, and high efficiency, significantly shortening the development cycle and reducing development costs.

[0004] However, applying machine learning virtual screening to the discovery of active molecules against CRE still faces several key technical bottlenecks, severely limiting screening accuracy and practicality: First, the problem of extreme data imbalance is prominent. Known active compounds against CRE are extremely scarce, while commercial compound libraries and virtual combination libraries can reach millions or even tens of millions in size. The ratio of active to inactive samples is extremely unbalanced, causing the model to easily favor the majority class, resulting in high false positives, low recall, and difficulty in accurately identifying potential active molecules. Second, conventional undersampling strategies have obvious defects. When balancing the dataset, they are prone to losing key structural and feature information in inactive samples, and they do not incorporate CRE. The use of specific screening rules based on pathogen characteristics and antimicrobial drug properties leads to a general lack of druggability, high off-target risk, and insufficient affinity for the target in the resulting candidate molecules. Thirdly, existing methods suffer from incomplete validation systems and insufficient interpretability. Most studies remain at the algorithm prediction level, lacking computational validation such as molecular docking and molecular dynamics simulations, and even more so, wet experimental confirmation such as in vitro antibacterial experiments. This makes it difficult to guarantee the reliability and scientific validity of the screening results. Furthermore, many models are "black box" models, failing to clearly define the structure-activity relationship between molecular structure, physicochemical properties, and anti-CRE activity, making it difficult to support subsequent molecular structure optimization.

[0005] In summary, given the current technical challenges in screening anti-CRE active molecules, such as data imbalance, crude sampling strategies, insufficient attention to drug-likeness, and lack of validation systems, there is an urgent need to develop a method that is suitable for CRE-resistant bacteria scenarios, can effectively alleviate extreme data imbalance, and possesses both high accuracy and high interpretability. This would provide intelligent and engineered technical support for the efficient discovery of novel anti-CRE antimicrobial drugs. Summary of the Invention

[0006] This invention aims to solve the above-mentioned technical problems and provides a method, medium, and device for screening active molecules of carbapenem-resistant Enterobacteriaceae based on the LightGBM algorithm, which significantly improves the screening hit rate and reliability.

[0007] To achieve this objective, the present invention adopts the following technical solution: A method for screening active molecules against carbapenem-resistant Enterobacteriaceae based on the LightGBM algorithm includes the following steps: S1. Construct a positive sample set P and a library of compounds to be screened U, and perform molecular standardization. S2. Extract features from the molecules obtained in step S1 to generate physicochemical property descriptors and molecular structure fingerprints, and perform feature fusion to obtain the numerical feature matrix X. S3. A training subset is formed by all positive samples and balanced negative samples randomly drawn from the compound library U to be screened. The base classifier is trained and iterated N times. The consensus score of the predicted probability output by each base classifier is obtained. LightGBM is selected as the base classifier. S4. Initial screening based on consensus score and applicable domain model; S5. The candidate molecules obtained in step S4 are screened for ADMET properties to optimize the outer membrane permeability of Gram-negative bacteria. The screening criteria include: molecular weight ≤ 600 Da; -5.0 ~ 1.5; TPSA ≤ 150 Ų; number of rotatable bonds ≤ 7; and molecules containing PAINS warning structures are excluded. S6. Perform multi-target molecular docking verification on the candidate molecules obtained in step S5, and screen molecules that have a strong binding affinity to at least one anti-CRE target to obtain the final screening results. Furthermore, the positive sample set P in step S1 consists of 61 compounds that have been shown to have anti-CRE activity in the literature; the compound library U to be screened is obtained from the coconut database; the ratio of positive samples to unlabeled samples is 1:10,000.

[0008] Furthermore, the physicochemical property descriptors mentioned in step S2 include: molecular weight, lipid-water partition coefficient, topological polar surface area, number of hydrogen bond donors, number of hydrogen bond acceptors, etc.; the molecular structure fingerprint is a MACCS Keys fingerprint.

[0009] Furthermore, in step S3, the number of iterations N is set to 50; In the i-th iteration, all 61 positive samples are retained, and 300 samples are randomly selected from the unlabeled pool U as negative samples for the current round to form the training set; After training the base classifier, predictions are made for all unlabeled samples; After the iteration is completed, the arithmetic mean of the predicted probabilities of each molecule in the 50 base classifiers is calculated as the final consensus score.

[0010] Furthermore, the consensus score threshold in step S4 is 0.8; the applicable domain filtering is as follows: calculate the Tanimoto similarity between the candidate molecule and 61 positive samples, and remove molecules with a maximum similarity of less than 0.6.

[0011] Furthermore, in the multi-target molecular docking verification described in step S6, the targets are selected from the following eight anti-CRE-related proteins: KPC, VIM, IMP, NDM, OXA-48, AmpC, OMP, and AcrAB-TolC; the docking tool is Autodock Vina; and the binding energy threshold is -7.0 kcal / mol.

[0012] Furthermore, the method verifies the effectiveness and robustness of the model through one or more of the following methods: Chemical spatial PCA distribution analysis was used to verify the distinguishability between positive and inactive samples; ROC curve and AUC value analysis showed that the average AUC reached 0.88; Analysis of PR curve and AP value showed an average AP of 0.62. Correlation analysis of the base classifiers showed correlation coefficients ranging from 0.52 to 0.71, demonstrating model diversity. SHAP interpretability analysis revealed that the model identified LogP (lipid solubility), MaxPartialCharge (charge), and PEOE_VSA (charge distribution) as core features, which confirms that this method can capture the specific physicochemical properties required for drugs from Gram-negative bacteria to pass through the outer membrane porins. The Y-permutation test showed that the average AUC dropped to 0.44 after shuffling the labels, proving that the model has good robustness.

[0013] Based on the same inventive concept, this invention also provides the application of the active molecules obtained by the method in the preparation of anti-CRE products, wherein the active molecules include compounds represented by the following CAS numbers: 92665-29-7, 72016-31-0, 68401-81-0, 65618-21-5, 65322-98-7, 65052-63-3, 61839-19-8, 61445-52-1, 60758-77-2, 52940-12-2, 528-58-5, 528-53-0, 52479 -85-3, 519-34-6, 50370-12-2, 34642-77-8, 2948-76-7, 24209-38-9, 193811-33-5, 15500-15-9, 144790-28-3, 134-04-3, 121412-77-9, 1154-78-5, 1151-98-0.

[0014] Based on the same inventive concept, the present invention also provides a computer-readable storage medium having a computer program stored thereon, wherein when the computer program is executed by a processor, the above-mentioned virtual screening and activity prediction method for the membrane permeability of carbapenem-resistant Enterobacteriaceae is implemented.

[0015] Based on the same inventive concept, the present invention also provides an electronic device, comprising: a processor and a memory; the memory for storing a computer program; and the processor for executing the computer program stored in the memory to cause the electronic device to perform the above-described method.

[0016] The present invention has the following beneficial effects: (1) Breaking through the bottleneck of extreme data imbalance Under extreme conditions of only 61 positive samples and 710,000 unlabeled samples (ratio 1:10,000), this invention achieves an average AUC of 0.88 and an average AP of 0.62 through an iterative random undersampling ensemble strategy, which is significantly better than traditional methods.

[0017] (2) Significantly improves screening hit rate and reduces wet experiment cost. Through consensus scoring and application domain filtering, only 1424 molecules were initially screened from 710,000 molecules; after optimization with ADMET, this was reduced to 214; and further refined to 26 through multi-target docking. This stepwise convergence strategy significantly reduced false positives and avoided unnecessary wet experiments.

[0018] (3) Overcoming the limitations of the "five principles of drug-like formulation" in antibiotic screening This invention addresses the outer membrane barrier characteristics of CRE by rationally setting parameters such as molecular weight (≤600 Da), LogP (-5~1.2), and polar surface area to avoid overlooking potential antibiotics with macrocyclic skeletons or moderate lipophilic properties, thereby improving the efficiency of antimicrobial drug discovery.

[0019] (4) Revealing the multi-target synergistic mechanism Through systematic docking of eight key CRE targets, it was found that candidate compounds (such as 144790-28-3) have strong binding affinity (<-8 kcal / mol) to both VIM and KPC, suggesting their potential for multi-target synergistic antibacterial activity.

[0020] (5) The model is highly interpretable and consistent with the understanding of medicinal chemistry. SHAP analysis revealed that the model identified LogP (lipid solubility), MaxPartialCharge (charge), and PEOE_VSA (charge distribution) as core features. This confirms that the method can capture the specific physicochemical properties required for drugs from Gram-negative bacteria to pass through outer membrane porins. It demonstrates that the model learns the true structure-activity relationship, rather than data noise.

[0021] (6) A complete multi-level verification system This invention constructs a full-chain screening and verification process of "artificial intelligence prediction → ADMET → molecular docking". Among the 26 candidate molecules, one has been confirmed by literature to have real antibacterial activity, and the technical solution is complete and reliable.

[0022] (7) The screening method was solidified into an executable program and device, providing dedicated hardware computing power support for high-throughput screening of anti-CRE drugs. This invention also provides a computer-readable storage medium and an electronic device. The storage medium solidifies the LightGBM model and multi-level screening rules, specifically optimized for the outer membrane barrier of CRE (Chronic Reproductive Respiratory Epidermal Genome), into a standard program. This not only enables rapid deployment and standardized reuse of the technical solution across different computing platforms but also effectively protects the core model assets built based on scarce positive samples. The electronic device utilizes its computing resources to fully leverage the high efficiency and low power consumption of the LightGBM algorithm when processing large-scale sparse data (710,000-level compounds), automatically executing the entire chain of tasks from 'AI intelligent prediction' to 'outer membrane permeability filtration' and 'multi-target enzyme binding verification'. This hardware-software collaborative approach shortens the screening cycle from weeks to hours and significantly reduces the investment in ineffective wet experiments through a precise step-by-step convergence strategy, providing an efficient, low-cost, and engineerable intelligent solution for addressing CRE, a global health threat. Attached Figure Description

[0023] Figure 1: Chemical spatial PCA distribution map, showing the clustering trend of positive samples; Figure 2 : ROC curve of the ensemble model, AUC=0.88; Figure 3 : PR curve of the ensemble model, AP=0.62; Figure 4 Correlation heatmap of base classifiers, correlation coefficient 0.52~0.71; Figure 5 SHAP Feature Importance Summary Diagram; Figure 6 Y-displacement test box plot; Figure 7 : Binding energy heatmap of docking between 26 candidate molecules and 10 target molecules. Detailed Implementation

[0024] The present invention will now be described in detail with reference to the accompanying drawings and specific embodiments, but this should not be construed as limiting the invention. Unless otherwise specified, the technical means used in the following embodiments are conventional means well known to those skilled in the art, and the materials, reagents, etc. used in the following embodiments are commercially available unless otherwise specified.

[0025] Example 1: A method for screening CRE-resistant active molecules based on the LightGBM algorithm.

[0026] The method flow of this embodiment includes the following steps: Step S1: Data Acquisition and Preprocessing Constructing the Positive Set: The inventors collected and organized data on 61 compounds with experimentally verified anti-CRE activity through extensive review of existing literature and public databases (Table 1). The data was cleaned to remove duplicates and invalid entries, and the SMILES (Simplified Molecular Linear Input Canonical) strings of the compounds were extracted to form the positive set P.

[0027] Table 1. CAS and SMILES of compounds with experimentally verified anti-CRE activity Constructing the Unlabeled Pool: The coconut database was selected as the screening source, and all small molecule compounds, totaling 715,838, were extracted. Their SMILES strings and unique identifiers (standard_inchi_key) were extracted to form a large-scale unlabeled sample set U. At this point, the ratio of positive samples to unlabeled samples was approximately 1:10,000, indicating a highly imbalanced data distribution.

[0028] Standardization: All the SMILES strings above are standardized using cheminformatics tools (RDKit is preferred in this embodiment), including desalting, neutralizing charges, and normalizing tautomers, to ensure the uniqueness of the molecular structure and machine readability.

[0029] Step S2: Molecular Feature Extraction and Vectorization In order to convert the chemical structure into a numerical vector that can be processed by a computer, this embodiment performs feature extraction on all molecules in step S1: Physicochemical property feature extraction: Calculate the physicochemical property descriptors of molecules, including molecular weight (MolWt), lipid-water partition coefficient (MolLogP), topological polar surface area (TPSA), number of hydrogen bond donors, number of hydrogen bond acceptors, etc., to characterize the pharmacokinetic properties of molecules.

[0030] Structural fingerprint extraction: Calculate the MACCS Keys fingerprint (167 bits) of the molecule, encoding the molecular structure into a binary feature vector to capture specific chemical substructure fragments.

[0031] Feature fusion: The physicochemical property descriptor vector and the structural fingerprint vector are concatenated to form the final numerical feature matrix X.

[0032] Step S3: Construct an ensemble screening model based on iterative undersampling To address the data imbalance problem described in step S1, this invention employs an integration strategy based on the Bagging PU (Positive-Unlabeled) concept: Initialization settings: Set the number of iterations for ensemble learning to N=50, and select LightGBM (ExtremeGradient Boosting) as the base classifier.

[0033] Iterative training: In the i-th iteration (i=1 to 50): Balanced sampling: All 61 positive samples are retained, and 300 samples are randomly selected from 715,838 unlabeled samples as "negative samples" to construct the balanced training set for the current round (361 samples in total).

[0034] Model training: Train the i-th LightGBM base classifier based on the balanced training set.

[0035] Full library prediction: The trained base classifier is used to predict the entire library of compounds to be screened (715,838 molecules), and the prediction probability is recorded.

[0036] Consensus Scoring: After the loop ends, the arithmetic mean of the predicted probabilities of each molecule in the 50 base classifiers is calculated as the final consensus score for that molecule.

[0037] Step S4: Primary Filtering and Applicable Domain Filtering High-confidence screening: Set the consensus score threshold to 0.8. Only retain molecules with a score > 0.8.

[0038] Application Domain (AD) Filtering: To ensure that candidate molecules are within the reliable prediction range of the model, the Tanimoto structural similarity between the candidate molecules and 61 positive samples is calculated. A maximum similarity threshold of 0.6 is set, and molecules with similarity values ​​lower than this value are removed.

[0039] Screening results: Through this step, 1,424 high-confidence preliminary screening compounds were obtained from 715,838 initial molecules.

[0040] Step S5: ADMET screening optimized for antimicrobial drugs To address the limitations of the traditional "five principles of drug-likeness" in antimicrobial drug screening, this invention establishes a specific ADMET screening criterion. The 1,424 compounds obtained in step S4 were subjected to computational screening.

[0041] Physicochemical property filtering: Set the following thresholds: Molecular weight ≤ 600 Da; nonspecific pore proteins of Gram-negative bacteria (such as OmpF / C) typically have small pore sizes, limiting the passage of molecules larger than 600 Da.

[0042] Lipid-water partition coefficient (LogP): -5.0 ~ 1.5; with a strict upper limit for LogP (<1.5). This is because highly lipid-soluble molecules are easily repelled by the hydrophobic outer membrane or recognized and expelled by efflux pumps, while hydrophilic molecules are more likely to pass through water-filled porin channels.

[0043] Topological polar surface area (TPSA) ≤ 150 Ų; a lower limit is set, not an upper limit. A higher polar surface area facilitates the interaction of molecules with polar residues within the channel as they pass through the porin, promoting transmembrane transport.

[0044] The number of rotatable bonds is ≤ 7; strict limits on molecular flexibility. Rigid molecules have lower entropy loss when passing through narrow porin channels and are more likely to accumulate in the cell (meeting the low flexibility requirement in the eNTRy rule).

[0045] PAINS Filtering: Utilizes a generalized assay interference compound (PAINS) filter to remove molecules containing structures with potential false positives or toxicity warnings.

[0046] Screening results: Through this step, 214 candidate compounds with good drug development potential were selected.

[0047] Step S6: Multi-target system molecular docking screening To further confirm the mechanism of action and binding affinity of the candidate molecules, systematic molecular docking was performed on the 214 compounds obtained in step S5: Target selection: Eight classic anti-CRE related target proteins were selected, specifically including: KPC, VIM, IMP, NDM, OXA-48, AmpC, OMP, and AcrAB-TolC; Docking implementation: A semi-flexible docking algorithm was used to simulate the binding of small molecules (Autodock vina) to the above-mentioned target sites.

[0048] Threshold screening: A binding affinity threshold of -7.0 kcal / mol is set. Only molecules with binding energies to at least one target site below this threshold are retained.

[0049] Screening results: 26 lead compounds with extremely strong theoretical binding affinity were finally selected.

[0050] The candidate molecules were confirmed through literature review, among which compound CNP0268126.1 possessed genuine antibacterial activity (see Table 2). This result indicates that the model is effective, the technical solution is complete and reliable, and the remaining compounds also have a high probability of possessing antibacterial activity, showing potential for preparation into anti-CRE products.

[0051] Table 2. Compounds with confirmed anti-CRE activity based on literature review. To verify the effectiveness of the method of the present invention, the inventors conducted experimental verification and recorded relevant data: Figure 1 The PCA analysis plot shows the chemical spatial distribution of the original training dataset. Red dots represent the positive sample set (known active molecules), and gray dots represent the library of compounds to be screened (background molecules), showing their distribution in the PCA dimensionality reduction space.

[0052] analyze: Figure 1Principal component analysis results showed that, although the library to be screened covered a wide chemical space, the known anti-CRE active molecules (red dots) were not uniformly distributed, but rather exhibited a clear clustering tendency, concentrated in specific chemical space regions.

[0053] Conclusion: This distribution characteristic indicates that anti-CRE active molecules possess specific structural or physicochemical property patterns, making them distinguishable from background molecules in the feature space. This verifies the effectiveness of the feature engineering method (physicochemical properties + fingerprint) selected in this invention, providing a solid data foundation for the subsequent construction of machine learning classification models.

[0054] Figure 2 The receiver operating characteristic (ROC) curve of the integrated model of this invention is shown. The solid red line in the figure represents the average ROC curve of the model after 50 iterations, the red shaded area represents the standard deviation range, and the gray dashed line is the random guess baseline.

[0055] Analysis: Experimental results show that the model's average AUC (area under the curve) is as high as 0.88. Meanwhile, the red shaded area is very narrow, indicating that the performance fluctuations of the 50 base classifiers are minimal, demonstrating a high degree of consistency.

[0056] Conclusion: This confirms that even with an extremely large ratio of positive to negative samples (1:10,000), the integrated undersampling strategy of this invention can still stably and accurately distinguish active molecules from background noise, and the model has extremely high classification robustness.

[0057] Figure 3 The precision-recall (PR) curves of the ensemble model of this invention are shown. The solid blue line in the figure represents the change in precision of the model under different recall rates.

[0058] Analysis: In extremely imbalanced datasets, the baseline precision-recall (PR) value is typically close to 0. However, this model achieved a mean precision-recall (AP) value of 0.62. This means that among the samples predicted as "active" by the model, a very high proportion are genuine active molecules, rather than false positives.

[0059] Conclusion: The results demonstrate the practical application value of the method of this invention in screening massive compound libraries, which can effectively reduce the "false positive rate" of subsequent wet experimental verification and significantly save research and development costs.

[0060] Figure 4 This displays a heatmap of Pearson correlations among the base classifiers in the ensemble model. The heatmap shows the correlation coefficients between the predictions of the top 10 base classifiers, with darker colors indicating higher correlations.

[0061] Analysis: Statistical analysis shows that the correlation coefficients between different base classifiers range from 0.52 to 0.71. This indicates that the base classifiers maintain consistency in the overall prediction direction (positive correlation) while also retaining significant differences (not perfect correlation).

[0062] Conclusion: This diversity stems from the "random undersampling" strategy of this invention, which allows each base classifier to learn features from different subsets of the library to be screened. This diversity is key to ensemble learning's ability to reduce variance and prevent overfitting.

[0063] Figure 5 This displays a summary plot of model feature importance based on SHAP values. Each row in the plot represents a feature, the color of the dots indicates the feature value (red for high, blue for low), and the position of the dots indicates the direction of their influence on the model output (right for positive contribution, left for negative contribution).

[0064] Analysis: The key features ranking highly include LogP (lipid solubility), MaxPartialCharge (charge), and PEOE_VSA (charge distribution) as core characteristics. This confirms that this method can capture the specific physicochemical properties required for drugs from Gram-negative bacteria to pass through the outer membrane porins.

[0065] Conclusion: This result is highly consistent with the pharmacological mechanism: anti-CRE drugs must possess specific lipophilic and polar characteristics to penetrate the thick bacterial cell wall. This proves that the model is not a "black box," but has successfully captured a structure-activity relationship (SAR) with physicochemical significance.

[0066] Figure 6 The results of the Y-Scrambling (Y permutation test) model robustness test are shown. The red box plot represents the performance distribution of the model trained with the true labels, and the gray box plot represents the performance distribution of the model after randomly shuffling the labels (Y-Scrambling).

[0067] Analysis: The AUC of the true label model remains stable at around 0.88, while after shuffling the labels, the model's AUC drops sharply to around 0.44 (i.e., at the level of random guessing). There is a huge performance gap between the two.

[0068] Conclusion: This comparison strongly demonstrates that the high performance of the model in this invention stems from learning the true relationship between molecular structure and activity, rather than randomly fitting the dataset noise, thus confirming the statistical significance of the screening results.

[0069] Figure 7This displays a heatmap of molecular docking binding energies between selected candidate molecules and multiple target sites. The vertical axis represents the IDs of the 26 selected candidate compounds, and the horizontal axis represents the 8 key anti-CRE targets. The color changes from red to blue, indicating a gradual decrease in binding energy (increased binding force).

[0070] Analysis: Several compounds exhibited broad-spectrum, strong binding affinity. In particular, for VIM (carbapenemase) and KPC (metallo-β-lactamase), the binding energies of some lead compounds (such as 144790-28-3) were significantly lower than -7.0 kcal / mol (the dark red area in the figure).

[0071] Conclusion: This not only validated the potential activity of the candidate molecules at the atomic level, but also revealed that these molecules may exert their anti-drug-resistant effects through a "multi-target synergistic mechanism," and have the potential to be developed into broad-spectrum anti-CRE drugs.

[0072] Example 2: Computer Program Product This embodiment provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described in Embodiment 1.

[0073] The computer-readable storage medium may be a storage medium such as ROM, RAM, disk, or optical disk.

[0074] Example 3: Electronic Equipment This embodiment provides an electronic device, including: a processor and a memory; the memory is used to store a computer program; the processor is used to execute the computer program stored in the memory to cause the electronic device to perform the method as described in Embodiment 1.

[0075] The electronic device may be a personal computer, server, workstation, high-performance computing cluster, or cloud computing platform.

[0076] It should be noted that when numerical ranges are mentioned in the claims of this invention, it should be understood that the two endpoints of each numerical range and any value between the two endpoints can be selected. To avoid redundancy, the present invention describes preferred embodiments.

[0077] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0078] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method, medium, and electronic device for screening active molecules of carbapenem-resistant Enterobacteriaceae based on the LightGBM algorithm, characterized in that, Includes the following steps: S1. Construct a positive sample set P and a library of compounds to be screened U, and perform molecular standardization. S2. Extract features from the molecules obtained in step S1 to generate physicochemical property descriptors and molecular structure fingerprints, and perform feature fusion to obtain the numerical feature matrix X. S3. A training subset is formed by all positive samples and balanced negative samples randomly drawn from the compound library U to be screened. The base classifier is trained and iterated N times. The consensus score of the predicted probability output by each base classifier is obtained. LightGBM is selected as the base classifier. S4. Based on the consensus score and the applicable domain model, a preliminary screening is conducted to select high-confidence candidate molecules. S5. The candidate molecules obtained in step S4 are screened for ADMET properties to optimize the outer membrane barrier of Gram-negative bacteria. The screening criteria include: molecular weight ≤ 600 Da; -5.0 ~ 1.5; TPSA ≤ 150 Ų; number of rotatable bonds ≤ 7; and molecules containing PAINS warning structures are excluded. S6. Perform multi-target molecular docking verification on the candidate molecules obtained in step S5, and screen molecules that have strong binding affinity to at least one anti-CRE target to obtain the final screening results.

2. The method according to claim 1, characterized in that, The positive sample set P in step S1 consists of 61 compounds that have been experimentally verified to have anti-CRE activity; the compound library U to be screened is obtained from the Coconut database; the ratio of positive samples to unlabeled samples is 1:10,000.

3. The method according to claim 1, characterized in that, The physicochemical property descriptors mentioned in step S2 include: molecular weight, lipid-water partition coefficient, topological polar surface area, number of hydrogen bond donors, number of hydrogen bond acceptors, etc.; the molecular structure fingerprint is a MACCS Keys fingerprint.

4. The method according to claim 1, characterized in that, In step S3, the number of iterations is set to N = 50; In the i-th iteration, all 64 positive samples are retained, and 300 samples are randomly selected from the compound library U to be screened as negative samples for the current round to form the training set. After training the base classifier, predictions are made for all unlabeled samples; After the iteration is completed, the arithmetic mean of the predicted probabilities of each molecule in the 50 base classifiers is calculated as the final consensus score.

5. The method according to claim 1, characterized in that, The consensus score threshold in step S4 is 0.8; the applicable domain filtering is: calculating the Tanimoto similarity between the candidate molecule and 61 positive samples, and removing molecules with a maximum similarity of less than 0.

6.

6. The method according to claim 1, characterized in that, In the multi-target molecular docking verification described in step S6, the targets were selected from the following eight anti-CRE-related proteins: KPC, VIM, IMP, NDM, OXA-48, AmpC, OMP, and AcrAB-TolC; the docking tool was Autodock Vina; and the binding energy threshold was -7.0 kcal / mol.

7. The method according to any one of claims 1 to 6, characterized in that, This method verifies the effectiveness and robustness of the model through one or more of the following methods: Chemical spatial PCA distribution analysis was used to verify the distinguishability between positive and inactive samples; ROC curve and AUC value analysis showed that the average AUC reached 0.88; Analysis of PR curve and AP value showed an average AP of 0.

62. Correlation analysis of the base classifiers showed correlation coefficients ranging from 0.52 to 0.71, demonstrating model diversity. SHAP interpretability analysis identified key features including LogP, MaxPartialCharge, and PEOE_VSA; The Y-permutation test showed that the average AUC dropped to 0.44 after shuffling the labels, proving that the model has good robustness.

8. The use of the active molecules screened by the method according to any one of claims 1 to 7 in the preparation of anti-CRE products, characterized in that, The active molecules include compounds with the following CAS numbers: 92665-29-7、72016-31-0、68401-81-0、65618-21-5、65322-98-7、65052-63-3、61839-19-8、61445-52-1、60758-77-2、52940-12-2、528-58-5、528-53-0、52479-85-3、519-34-6、50370-12-2、34642-77-8、2948-76-7、24209-38-9、193811-33-5、15500-15-9、144790-28-3、134-04-3、121412-77-9、1154-78-5、1151-98-0。 9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1-7.

10. An electronic device, characterized in that, include: Processor and memory; The memory is used to store computer programs; The processor is configured to execute a computer program stored in the memory to cause the electronic device to perform the method as described in any one of claims 1-7.

Images