A method for screening α-glucosidase-specific inhibitors based on multimodal closed-loop active learning
By employing a multimodal closed-loop active learning method, combining chemical language models and physicochemical descriptors, and using an adaptive smooth weighted pseudo-Huber loss function and XGBoost model, the low efficiency and insufficient selectivity of α-glucosidase inhibitor screening in existing technologies were addressed. This led to the discovery of highly selective inhibitors with low side effects, achieving efficient drug discovery.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHINA AGRI UNIV
- Filing Date
- 2026-04-22
- Publication Date
- 2026-07-14
AI Technical Summary
Existing deep learning screening methods suffer from limitations in feature representation, imbalanced data distribution, failure of traditional active learning strategies, and insufficient adaptability of loss functions in the screening of α-glucosidase inhibitors, resulting in low screening efficiency and low selectivity.
A multimodal closed-loop active learning method is adopted. By combining chemical language models and physicochemical descriptors to generate high-dimensional semantic vectors, an adaptive smooth weighted pseudo-Huber loss function and an XGBoost model are used for active learning sampling optimization to screen out highly active compounds. The specificity is verified by secondary molecular docking.
This significantly improved the screening efficiency and accuracy of α-glucosidase-specific inhibitors, leading to the discovery of highly selective inhibitors with low side effects, and reducing research and development costs and time.
Smart Images
Figure CN122392705A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of computer-aided drug design technology, specifically relating to a method for screening α-glucosidase-specific inhibitors based on multimodal closed-loop active learning. Background Technology
[0002] Alpha-glucosidase is a key enzyme in the human body responsible for breaking down polysaccharides such as starch into glucose, and its inhibitors are important targets for regulating postprandial blood glucose. Currently used alpha-glucosidase inhibitors (such as acarbose and miglitol) effectively lower blood glucose, but often result in large amounts of undigested starch entering the colon due to their strong simultaneous inhibition of alpha-amylase. This starch, after microbial fermentation, causes gastrointestinal side effects such as bloating and diarrhea, severely impacting patient adherence. Therefore, developing specific inhibitors with high selectivity for alpha-glucosidase and very weak inhibitory activity against alpha-amylase is an effective strategy to address these clinical challenges.
[0003] Virtual screening, as one of the core technologies of modern drug discovery, can rapidly enrich potentially active molecules from massive compound libraries, significantly reducing R&D costs and time. Traditional virtual screening relies heavily on molecular docking and classical scoring functions, but its prediction accuracy is limited by the precision of the scoring functions, and it requires high computational resources, making it inefficient when processing ultra-large compound libraries.
[0004] In recent years, machine learning (ML), especially deep learning (DL) techniques, has been widely used to improve the performance of virtual screening. However, existing deep learning screening methods still face the following key challenges: Limitations of feature representation: Traditional molecular fingerprints (such as ECFP4) are difficult to capture deep semantic context information of molecules, while simple end-to-end deep learning models often lack explicit physicochemical prior knowledge (such as LogP, molecular weight, etc.), which limits the model's generalization ability in small samples.
[0005] Extreme imbalance in data distribution: In real-world drug screening scenarios, highly active compounds are extremely rare (distributed in a long tail). Conventional machine learning models (such as random forests and SVMs) are easily dominated by a large number of inactive samples during training, resulting in extremely low recall rates for the few highly active samples.
[0006] The failure of traditional active learning: Although active learning can reduce labeling costs through iterative sampling, existing strategies are mostly based on uncertainty sampling, that is, prioritizing the samples whose model predictions are most uncertain (usually located near the decision boundary). When active molecules are extremely sparse, this strategy often fails to effectively explore truly high-activity regions.
[0007] Insufficient adaptability of loss functions: Standard loss functions (such as mean squared error (MSE) or mean absolute error (MAE) usually treat all samples equally or are too sensitive to outliers, failing to force the model to focus on a very small number of highly active samples while ensuring training stability.
[0008] Therefore, there is an urgent need for a closed-loop screening method that can integrate deep semantic and physicochemical features and accurately locate highly active compounds through an adaptive weighting mechanism, so as to improve the discovery efficiency of α-glucosidase selective inhibitors. Summary of the Invention
[0009] The purpose of this invention is to overcome the shortcomings of the prior art and provide a computational method based on multimodal closed-loop active learning that has high screening efficiency, good accuracy, and can effectively identify α-glucosidase specific inhibitors.
[0010] To achieve the above objectives, the first aspect of this invention provides a method for screening α-glucosidase-specific inhibitors based on closed-loop active learning, the core process of which includes: 1) Data preparation and preprocessing: Obtain the compound library, screen the compounds for drug-likeness, and use molecular docking software to dock them with the α-glucosidase protein target to obtain docking scores as initial labels; 2) Molecular representation and feature extraction: The SMILES sequences of compounds are encoded into fixed-dimensional molecular embedding vectors using a pre-trained chemical language model; 3) Machine learning model construction and training: Using the molecular embedding vector as input features and docking score or activity label derived therefrom as output, train a machine learning classification or regression model; 4) Active learning sampling optimization: An active learning strategy is adopted to iteratively select the samples with the most information from the candidate compound pool and add them to the training set in order to optimize the model's ability to identify highly active compounds; 5) Virtual screening and specificity verification: The trained model is used to perform initial screening on a large-scale compound library. Then, the potential active compounds obtained from the initial screening are subjected to secondary molecular docking to verify their specificity for α-glucosidase.
[0011] Furthermore, the compound library in step 1) is the commercial compound library ChemDiv, and the drug-likeness screening is performed according to the Lipinski Five Rules of Drug-likeness; the molecular docking software is AutoDock Vina.
[0012] Furthermore, in step 2), a multimodal feature fusion strategy is adopted: firstly, a pre-trained model (such as ChemBERTa) is used to encode the SMILES sequence, and a dual pooling strategy combining mean pooling and max pooling is used to generate a high-dimensional semantic vector (such as 1536 dimensions); at the same time, cheminformatics tools (such as RDKit) are used to calculate the physicochemical descriptor of the molecule (including molecular weight, number of rotatable bonds, LogP, topological polar surface area, etc.), and the semantic vector and the physicochemical descriptor are concatenated as the model input.
[0013] Furthermore, in step 3), a smooth nonlinear weighted pseudo-Huber loss function was designed. This loss function combines the advantages of mean squared error (MSE) and mean absolute error (MAE), and by introducing adaptive weights based on the sigmoid function, significantly higher weights are assigned to highly active samples, thereby accurately optimizing the prediction results of key highly active regions while maintaining training stability.
[0014] Furthermore, the machine learning model in step 3) is an XGBoost model, and the DART boosting strategy is adopted; cross-validation is used for hyperparameter optimization during model training.
[0015] Furthermore, the active learning strategy in step 4) is a smooth weighted sampling strategy. Unlike random sampling or uncertain sampling, this strategy calculates sampling weights based on the current model's prediction scores for candidate pool samples. Samples with stronger predicted binding energy (lower scores) have a higher probability of being selected, thereby guiding the model to quickly focus on highly active regions in the chemical space.
[0016] Furthermore, a custom adaptive smooth weighted pseudo-Huber loss function is employed during model training. This function employs a non-linear smooth weighting mechanism based on the Sigmoid function, and its expression is as follows:
[0017] in, For predicted values, For the true value, For smoothing coefficients, The adaptive sample weights are calculated using the following formula:
[0018] Furthermore, in the specificity verification step of step 5), α-amylase protein is used as a control target. Potentially active compounds obtained from the initial screening are molecularly docked with α-amylase to screen out compounds that bind weakly to α-amylase as specific inhibitors of α-glucosidase.
[0019] A second aspect of the present invention is to provide the application of the method described in the first aspect in screening hypoglycemic drugs. Attached Figure Description
[0020] Figure 1 This is an overall flowchart of the method described in this invention.
[0021] Figure 2 Performance comparison chart before and after model optimization.
[0022] Figure 3 The inhibitory effect is achieved through competitive binding to the substrate.
[0023] Figure 4 A comparison of the fluorescence quenching effect of B19 on α-glucosidase with that of acarbose.
[0024] Figure 5 The specific interactions between B19 and the α-glucosidase active site (A, C) and its comparison with acarbose (B, D)
[0025] Figure 6 Effects of the compound on type 2 diabetic mice: A. Fasting blood glucose, B. Oral glucose. Detailed Implementation
[0026] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments.
[0027] Example: Method Implementation Process
[0028] like Figure 1 As shown.
[0029] 1. Data Preparation: Approximately 100,000 compounds were downloaded from the ChemDiv database. Small molecules were preprocessed using tools such as OpenBabel to generate and optimize their three-dimensional structures. Crystal structures of α-glucosidase (2QMJ) and α-amylase (1XCW) were obtained from the PDB database, and water molecules were removed, hydrogen was added, and energy was optimized.
[0030] 2. Molecular docking: Using AutoDock Vina software, a docking box was set with the original ligand location as the center. All compounds were docked with α-glucosidase, and the score was output as the regression target for the subsequent model.
[0031] 3. Multimodal Feature Generation: A multimodal feature fusion strategy is employed. a) A pre-trained model (such as ChemBERTa-77M) is used to convert the compound SMILES sequences into semantic embedding vectors. A dual pooling strategy (mean pooling + max pooling) is used to generate 1536-dimensional semantic features. b) The RDKit tool is used to calculate the physicochemical descriptors of the molecules (such as molecular weight, LogP, TPSA, etc.) to generate physicochemical features. c) The above two sets of features are concatenated to form the final model input vector.
[0032] 4. Model Training and Active Learning: a) Initial training: Randomly select 20% of the samples (approximately 20,000) from the total dataset to form the initial training set.
[0033] b) Model settings: Use XGBoost as the base model and set core parameters such as booster='dart', max_depth=6, learning_rate=0.1.
[0034] c) Initiate the active learning loop: (1) Sampling: The current model is used to predict the remaining unlabeled samples in the candidate pool. A smooth weighted sampling strategy is adopted, and the sampling weight is calculated based on the prediction score (the sample with the stronger affinity / lower score has a higher weight), and 2,000 samples with the highest information content and the highest potential activity are selected from them.
[0035] (2) Labeling: Calculate the true docking score of these 2000 samples through molecular docking (i.e. label them with true labels).
[0036] (3) Update: Add the newly labeled samples to the training set and remove them from the candidate pool.
[0037] (4) Retraining: The XGBoost model is retrained using the expanded training set. During training, an adaptive smooth nonlinear weighted pseudo-Huber loss function is used to enhance the model's ability to fit highly active samples.
[0038] (5) Termination condition: Repeat the above loop until the number of training set samples reaches the upper limit of the total number of candidate pools (i.e., 20,000), and complete the model training.
[0039] 5. Virtual screening and verification: a) Use the finally trained model to predict the entire compound library, and select the approximately 1% of compounds with the best (lowest) prediction scores, totaling 14,354 compounds.
[0040] b) These compounds were docked with α-amylase (1XCW) to screen for compounds with scores higher than -6.0, ultimately yielding 32 highly specific candidate inhibitors.
[0041] c) Through in vitro enzyme activity assays, several highly active and selective inhibitors were identified, with compound B19 showing the best IC50 against α-glucosidase. 50 The concentration was 58.43 ± 3.54 µM, while the inhibition rate against α-amylase remained below 50% even at a concentration of 500 µM, demonstrating excellent selectivity (Table 1). Figure 2 ).
[0042] Table 1. IC50 values of five candidate compounds against two amylases.
[0043] d) Further investigation into the inhibition kinetics of compound B19 revealed that B19, as a competitive inhibitor of α-glucosidase, has an action site that overlaps with the active site of the enzyme's hydrolysis of oligosaccharides, achieving its inhibitory effect through competitive binding to the substrate. Figure 3 The inhibition constant (Kic) of B19 with α-glucosidase was significantly lower than that of acarbose, indicating that B19 has a stronger affinity for the enzyme (Table 2).
[0044] Table 2
[0045] e) Subsequently, the affinity of compound B19 for α-glucosidase was determined by fluorescence quenching experiments. It was found that B19 exhibited a significantly stronger fluorescence quenching effect on α-glucosidase than acarbose, accompanied by a redshift of the maximum emission wavelength (λem). This suggests that B19 interacts with the aromatic rings and hydrophobic groups of the TRP and TYR residues in the enzyme molecule, leading to local conformational changes and partial structural unfolding of the enzyme protein. Figure 4 (where A is compound B19 and B is the control acarbose). Parameter analysis indicates that B19 may bind to α-glucosidase in a 1:1 molar ratio, and moderate heating favors the binding process (Table 3). Thermodynamic parameter analysis further reveals that the binding of B19 to α-glucosidase is a spontaneous exothermic process, with van der Waals forces and hydrogen bonds being the main driving forces (Table 4).
[0046] Table 3
[0047] Table 4
[0048] f) Further investigation was conducted using molecular docking to explore the specific interaction mode between B19 and the α-glucosidase active site. The results showed that B19 forms a hydrogen bond network with catalytic residues ASP542, ASP443, and ARG202, while also forming hydrophobic interactions with hydrophobic residues such as TRP406, PHE450, ARG202, ARG526, ASP542, and PHE575, and exhibiting π-π stacking interactions with PHE450. Figure 5 Where A and C are compounds B19, and B and D are control acarbose. The docking results of B19 with α-amylase showed that it failed to completely occupy the enzyme's active pocket, suggesting that it may bind to α-amylase in a non-competitive manner. Subsequently, we optimized the structure of compound B19, and the optimized compound showed significant efficacy in the treatment of type 2 diabetic mice. Figure 6 AC).
[0049] The embodiments described above are only some embodiments of the present invention, not all embodiments. Other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are all within the scope of protection of the present invention.
Claims
1. A method for screening α-glucosidase-specific inhibitors based on multimodal closed-loop active learning, characterized in that, Includes the following steps: 1) Data preparation and preprocessing: Obtain the compound library, screen the compounds for drug-likeness, and use molecular docking software to dock them with the α-glucosidase protein target to obtain docking scores as initial labels; 2) Molecular representation and feature extraction: The SMILES sequences of compounds are encoded into fixed-dimensional molecular embedding vectors using a pre-trained chemical language model; 3) Machine learning model construction and training: Using the molecular embedding vector as input features and docking score or activity label derived therefrom as output, train a machine learning classification or regression model. 4) Active learning sampling optimization: An active learning strategy is adopted to iteratively select the samples with the most information from the candidate compound pool and add them to the training set in order to optimize the model's ability to identify highly active compounds; 5) Virtual screening and specificity verification: The trained model is used to perform initial screening on a large-scale compound library. Then, the potential active compounds obtained from the initial screening are subjected to secondary molecular docking to verify their specificity for α-glucosidase.
2. The method according to claim 1, characterized in that, The compound library in step 1) is the commercial compound library ChemDiv, and the drug-likeness screening is performed according to the Lipinski five rules of drug-likeness; the molecular docking software is AutoDock Vina.
3. The method according to claim 1, characterized in that, In step 2), the molecular representation adopts a multimodal feature fusion strategy: the SMILES sequence is encoded using a pre-trained chemical language model, and a high-dimensional semantic vector is generated by a dual pooling strategy combining mean pooling and max pooling; at the same time, the physicochemical descriptor of the molecule is calculated, and the two are concatenated to form the final molecular representation vector.
4. The method as described in claim 1, characterized in that, The machine learning model in step 3) is the XGBoost model, and the DART boosting strategy is adopted; cross-validation is used for hyperparameter optimization during model training.
5. The method according to claim 1, characterized in that, In step 3) model construction, a smooth nonlinear weighted pseudo-Huber loss function is adopted. This loss function combines the advantages of mean squared error (MSE) and mean absolute error (MAE), and by introducing adaptive weights based on the sigmoid function, highly active samples are given significantly higher weights, thereby accurately optimizing the prediction results of key highly active regions while maintaining training stability.
6. The method as described in claim 1, characterized in that, The active learning strategy in step 4) is a smooth weighted sampling strategy; the strategy calculates the sampling probability based on the model prediction value of the compound, and assigns higher sampling weights to samples with higher predicted activity in order to guide the model to focus on high-activity regions.
7. The method as described in claim 1 or 4, characterized in that, A custom smooth-weighted pseudo-Huber loss function is used in model training, and its expression is as follows: in, For predicted values, For the true value, For smoothing coefficients, For adaptive sample weights.
8. The method according to claim 7, characterized in that, in The calculation formula is: 。 9. The method as described in claim 1, characterized in that, In the specificity verification step of step 5), α-amylase protein is used as a control target. Potentially active compounds obtained from the initial screening are molecularly docked with α-amylase to screen out compounds that bind weakly to α-amylase as specific inhibitors of α-glucosidase.
10. The application of the method according to any one of claims 1 to 9 in screening hypoglycemic drugs.