Anti-endoplasmic reticulum stress glaucoma intraocular pressure lowering drug discovery method based on machine learning

Abstract

The present invention discloses a method for discovering anti-ER stress glaucoma intraocular pressure-lowering drugs based on machine learning, by constructing an anti-ER stress compound library: constructing an anti-ER stress compound library through a target-independent strategy; machine learning model training and drug screening: extracting multiple drug physicochemical characteristics and structural characteristics for machine learning model training, computer screening of multiple drugs that have been marketed and preclinical, and selecting anti-ER stress compounds; glaucoma disease phenotype verification: verifying the screened anti-ER stress compounds in a preset glaucoma cell model and animal model. The present invention will explore new glaucoma intraocular pressure-lowering drugs, expand the mechanism of action of existing drugs, establish a target-independent drug discovery model, and fully explore glaucoma drug discovery strategies based on disease phenotypes.

Description

Technical Field

[0001] The present invention relates to the technical field of biological information processing, and in particular discloses a method for discovering anti-endoplasmic reticulum stress glaucoma intraocular pressure lowering drugs based on machine learning. Background Art

[0002] Glaucoma is the world's leading irreversible blinding eye disease, with an incidence of about 3.05% in people aged 40-80 years worldwide. With the aging of the population, it is estimated that the number of glaucoma patients will reach nearly 111.8 million by 2040. Primary open angle glaucoma (POAG) is the most common type of glaucoma. The trabecular meshwork (TM) is the site with the greatest resistance in the aqueous humor outflow tract, and is therefore the main link in the regulation of intraocular pressure. TM dysfunction is an important factor in the pathogenesis of POAG, including a decrease in the number of TM cells, reduced phagocytic function, TM tissue stiffness caused by changes in the intracellular agonist protein network, and extracellular matrix (ECM) accumulation. Controlling intraocular pressure to slow down further damage to the optic nerve is currently the only strategy for the treatment of glaucoma in clinical practice. However, there is still a large proportion of patients whose intraocular pressure is poorly controlled by drugs, and these drugs mainly reduce aqueous humor production or increase aqueous humor outflow. Only Rho kinase inhibitors directly target the trabecular meshwork. However, a 2022 Cochrane systematic review found that the Rho kinase inhibitor Netarsudil was less effective than temarol or latanoprost in lowering intraocular pressure. Finding intraocular pressure-lowering drugs that directly target trabecular meshwork lesions is an urgent problem to be solved in the treatment of glaucoma.

[0003] Different risk factors in glaucoma, such as age, glucocorticoids, ischemia, and deleterious mutations, have been found to cause chronic ER stress, which may be a common feature in glaucoma. The endoplasmic reticulum (ER) is a quality control organelle for proteins. The protein quality control system is responsible for maintaining protein homeostasis. Misfolded proteins are generated under stress conditions and transferred from the ER to the cytoplasm to be degraded in the ubiquitin-proteasome via the ER-associated degradation (ERAD) pathway. However, the continued accumulation of misfolded proteins leads to ER stress, which may trigger an adaptive response called the unfolded protein stress response (UPR). Gene mutations are closely associated with chronic ER stress in POAG, leading to TM cell death. The most common POAG mutation is located in myocilin. Its mutation causes about 4% of POAG cases, causing myocilin protein to accumulate in TM cells and cannot be secreted outside the cell. The subsequent amyloidosis is the main trigger of ER stress in the development of POAG, triggering programmed cell death and impaired autophagy. Similarly, the use of glucocorticoids will lead to an overload of myocilin levels in TM cells, inducing chronic ER stress, leading to TM dysfunction and increased IOP.

[0004] Currently, there are chemical, natural compounds, gene therapy, and stem cell therapies aimed at addressing ER stress. In TM cells with MYOC mutations, Grp94 recognition leads to the accumulation of myocilin, triggering ER stress and TM cell function or loss. Inhibition of Grp94 can help mutant or misfolded myocilin to be degraded through efficient autophagy, rather than through ERAD, reducing abnormal accumulation of myocilin. The 4-Br-BnIm compound can selectively inhibit Grp94, thereby promoting the degradation of mutant or misfolded myocilin in TM cells. The PERK inhibitor LDN-0060609 can rescue TM cells from ER stress. 4-phenylbutyric acid (4-PBA) can increase the efflux of mutant myocilin and also degrade ECM by activating MMP9. In addition, in glucocorticoid-induced glaucoma, 4-PBA alleviated ER stress and reduced intraocular pressure. Astragaloside IV, which has been used in kidney and heart diseases, is effective in preventing myocilin deposition. Astragaloside IV can regulate ECM deposition and ER stress in TM by interacting with the MMP3 and MMP9 systems, thus rescuing TGF-β2-induced increased intraocular pressure. Although cell and animal experiments have confirmed that compounds that relieve ER stress are potential glaucoma treatments, there is still a long way to go before drug development. Currently, ER stress drugs registered on ClinicalTrials.org are mainly targeted at diseases of the cardiovascular, liver, diabetes, and adrenal systems, and no glaucoma drugs targeting ER stress have been transformed.

[0005] The main challenges of drug transformation at present are: (1) high cost (US$2.6 million is required for FDA approval); (2) long animal experiments and RCT experimental cycles; (3) unknown safety issues are exposed after the drug is launched on the market. In order to solve the above problems in drug development, the field of drug research and development has combined artificial intelligence and machine learning methods in recent years to develop models such as bioactivity prediction, target identification, virtual drug screening and new uses of old drugs. These methods usually use molecular dynamics simulation and complex computational processes to predict the spatial structure of drug candidates. For example, in 2022, Zeng et al. predicted 13 targets and corresponding drugs for the new coronavirus SARS-CoV-2. In 2023, Chinese scholars Zhu et al. found that AI can predict the three-dimensional protein structure of BDNF and TrkB and the interaction between them, and further simulate small molecule drugs for BDNF-TrkB interaction, thereby identifying 7,8-dihydroxyflavone (DHF) as a TrkB agonist. Basic experiments verified the neuroprotective effect of DHF in an acute glaucoma model in vivo.

[0006] Recent studies have developed machine learning models based on molecular fingerprints or learned chemical structure representations, some of which break away from traditional target-oriented drug discovery methods and adopt target-agnostic strategies to train models through phenotypic readouts. This target-agnostic approach provides a new way to expand the chemical starting point of the early stages of the drug discovery process and is particularly suitable for drug discovery with limited understanding of the molecular pathways that control cellular disease phenotypes. The MIT team designed the molecule halicin, which was confirmed in mouse models as a broad-spectrum antibiotic that can effectively treat Clostridium difficile and pan-resistant Acinetobacter baumannii infections. In addition, 8 antibacterial compounds that are structurally far from known antibiotics were identified from 23 discrete empirical test predictions of 107 million molecules in the ZINC15 database. The unique prospect of the target-agnostic strategy is that it can use disease phenotypes to discover new treatments for diseases with unknown root causes, complex or multifactorial diseases, and scientific understanding is insufficient to provide effective molecular targets. Just as ER stress has no known enzymatic activity and is difficult to target for drug binding, target-independent drug screening strategies can circumvent this limitation.

[0007] Therefore, how to explore new glaucoma IOP-lowering drugs through target-independent strategies, expand the mechanism of action of existing drugs, establish target-independent drug discovery models, and fully explore glaucoma drug discovery strategies based on disease phenotypes are technical issues that need to be solved urgently. Summary of the invention

[0008] The present invention provides a method for discovering anti-endoplasmic reticulum stress glaucoma intraocular pressure-lowering drugs based on machine learning, which aims to explore new glaucoma intraocular pressure-lowering drugs through target-independent strategies, expand the mechanism of action of existing drugs, establish a target-independent drug discovery model, and fully explore glaucoma drug discovery strategies based on disease phenotypes.

[0009] The present invention relates to a method for discovering anti-endoplasmic reticulum stress glaucoma intraocular pressure lowering drugs based on machine learning, comprising the following steps:

[0010] Construction of anti-ER stress compound library: Construction of anti-ER stress compound library through target-independent strategy;

[0011] Machine learning model training and drug screening: extract multiple drug physical and chemical characteristics and structural characteristics for machine learning model training, perform computer screening on multiple marketed and preclinical drugs, and select anti-endoplasmic reticulum stress compounds;

[0012] Glaucoma disease phenotype validation: The screened anti-ER stress compounds are validated in preset glaucoma cell models and animal models.

[0013] Furthermore, the steps of constructing an anti-endoplasmic reticulum stress compound library include:

[0014] Compound collection: Determine endpoint indicators through existing basic experiments and clinical research data, and search for effective anti-ER stress compounds;

[0015] Heterogeneity evaluation: The anti-ER stress compounds were clustered using the K-means algorithm; the Tanimoto distance was used to evaluate the similarity and heterogeneity of the clustered anti-ER stress compounds.

[0016] Furthermore, the step of constructing an anti-endoplasmic reticulum stress compound library also includes:

[0017] Training set construction: construct ER sress(+) compound training set and ER stress(-) compound training set;

[0018] Proportion of sample training set: The proportion of ER sress(+) compound samples is unbalanced. In the training set, ER sress(+) compounds account for about 7.46%.

[0019] Furthermore, the step of constructing an anti-endoplasmic reticulum stress compound library also includes:

[0020] RDKit feature extraction: Use Python and RDKit package to extract compound features through SMILES symbols. Compound features include physical and chemical features and structural features.

[0021] Compound label marking: Using ER sress(+) compounds and ER stress(-) compounds as labels, the ER sress(+) compounds and ERstress(-) compounds in the ERsress(+) compound training set and the ER stress(-) compound training set were annotated respectively.

[0022] Furthermore, the steps of machine learning model training and drug screening include:

[0023] Compound feature optimization: feature selection for the complete data set; sorting the extracted features;

[0024] Model training and optimization: Use support vector machine and XGBoost to train the sorted features to form a machine learning model, and use corresponding indicators to optimize the machine learning model.

[0025] Further, the step of optimizing compound characteristics includes:

[0026] Use random forest method to evaluate impurity metrics in classification tasks;

[0027] Sort the extracted features and delete those with a score of 0.

[0028] Furthermore, in the steps of model training and optimization, three indicators, precision, recall and F1 score, are used to evaluate the machine learning model. The precision, recall and F1 score are:

[0029] precision=TP / (TP+FP)

[0030] recall=TP / (TP+FN)

[0031] F1 score = TP / (TP+1 / 2(FP+FN))

[0032] Among them, precision represents accuracy, TP represents true positive rate, FP represents false positive rate, recall represents recall rate, FN represents false negative rate, and F1 score represents F1 value.

[0033] Furthermore, the steps of machine learning model training and drug screening include:

[0034] Compound library construction: Establish L3800 Selleck compound library;

[0035] Test set screening: The optimal machine learning model was used to screen the test set mainly based on the L3800 Selleck compound library, and the predicted probabilities were ranked using the z-score to screen out multiple compounds with high rankings.

[0036] Further, the steps of glaucoma disease phenotype validation include:

[0037] High-throughput screening and rescreening: Perform high-throughput screening on the top-ranked compounds to screen out the top three compounds; rescreen the high-throughput screened compounds, draw dose-response curves, and calculate IC50 and anti-ERstress index;

[0038] Trabecular meshwork cell model validation: The rescreened compounds were validated using the trabecular meshwork cell model to detect TM cell activity and ER stress levels.

[0039] Furthermore, the step of validating the glaucoma disease phenotype also includes:

[0040] Validation in chronic mouse ocular hypertension model: The rescreened compounds were validated using a chronic mouse ocular hypertension model to detect TM cell apoptosis, ER stress levels, and disease phenotypes, including IOP changes and retinal RGC survival.

[0041] The beneficial effects achieved by the present invention are:

[0042] The present invention discloses a method for discovering anti-endoplasmic reticulum stress glaucoma intraocular pressure-lowering drugs based on machine learning, by constructing an anti-endoplasmic reticulum stress compound library: constructing an anti-endoplasmic reticulum stress compound library through a target-independent strategy; machine learning model training and drug screening: extracting multiple drug physicochemical characteristics and structural characteristics for machine learning model model training, computer screening of multiple drugs that have been marketed and preclinical, and selecting anti-endoplasmic reticulum stress compounds; glaucoma disease phenotype verification: verifying the screened anti-endoplasmic reticulum stress compounds in a preset glaucoma cell model and animal model. The anti-endoplasmic reticulum stress glaucoma intraocular pressure-lowering drug discovery method based on machine learning provided by the present invention will explore new glaucoma intraocular pressure-lowering drugs, expand the mechanism of action of existing drugs, establish a target-independent drug discovery model, and fully explore glaucoma drug discovery strategies based on disease phenotypes. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 The figure is a schematic diagram of the process of the method for discovering anti-endoplasmic reticulum stress glaucoma intraocular pressure lowering drugs based on machine learning of the present invention;

[0044] Figure 2 This is a schematic diagram of the detailed process of the method for discovering anti-endoplasmic reticulum stress glaucoma intraocular pressure-lowering drugs based on machine learning of the present invention. DETAILED DESCRIPTION

[0045] In order to better understand the above technical solution, the above technical solution will be described in detail below in conjunction with the accompanying drawings and specific implementation methods.

[0046] like Figure 1 and Figure 2 As shown, the first embodiment of the present invention proposes a method for discovering anti-endoplasmic reticulum stress glaucoma intraocular pressure-lowering drugs based on machine learning, comprising the following steps:

[0047] Step S100, constructing an anti-endoplasmic reticulum stress compound library: constructing an anti-endoplasmic reticulum stress compound library through a target-independent strategy.

[0048] It is planned to construct a library of anti-endoplasmic reticulum stress compounds through a target-independent strategy.

[0049] Step S200, machine learning model training and drug screening: extract multiple drug physicochemical characteristics and structural characteristics for machine learning model training, perform computer screening on multiple marketed and preclinical drugs, and select anti-endoplasmic reticulum stress compounds.

[0050] More than 200 physical and chemical characteristics and structural features of drugs were extracted for model training, and multiple marketed and preclinical drugs were screened by computer to select anti-endoplasmic reticulum stress compounds.

[0051] Step S300, glaucoma disease phenotype verification: the screened anti-ER stress compounds are verified in a preset glaucoma cell model and animal model.

[0052] The screened anti-ER stress compounds are verified in glaucoma cell models and animal models. The glaucoma cell model can be a Dex-TM cell model, a TUN-TM cell model, a BreA-TM cell model, or a Thap-TM cell model. The animal model can be a chronic mouse ocular hypertension model.

[0053] Furthermore, the machine learning method for discovering intraocular pressure-lowering drugs for glaucoma induced by endoplasmic reticulum stress provided in this embodiment comprises step S100:

[0054] Step S110, compound collection: determine the endpoint indicators through existing basic experiments and clinical research data, and find effective anti-endoplasmic reticulum stress compounds.

[0055] Through the search of literature, patents and clinical studies, we are looking for effective anti-ER stress drugs, with the endpoint indicators being: ① clear improvement in apoptosis; ② clear changes in ER stress markers.

[0056] Step S120, heterogeneity evaluation: clustering the anti-ER stress compounds by using the K-means algorithm; and evaluating the similarity and heterogeneity of the clustered anti-ER stress compounds by using the Tanimoto distance.

[0057] Since anti-ER stress compounds come from different types of cells, different disease models, and different stimulus factors, quality control is required. Clustering is first performed using the K-means algorithm, and then the similarity and heterogeneity of the compounds are evaluated using the Tanimoto distance.

[0058] Preferably, the method for discovering anti-ER stress glaucoma intraocular pressure-lowering drugs based on machine learning provided in this embodiment, step S100 comprises:

[0059] Step S130, training set construction: constructing an ER stress (+) compound training set and an ER stress (-) compound training set.

[0060] The training set consisted of 198 ER stress(+) compounds and 2465 ER stress(-) compounds from databases such as the LOPAC library and the Prestwick chemical library.

[0061] Step S140, sample training set ratio: The ratio of ER sress (+) compound samples is unbalanced. In the training set, ER sress (+) compounds account for about 7.46%.

[0062] The proportion of anti-ER sress(+) compound samples was unbalanced, with ER sress(+) compounds accounting for approximately 7.46%.

[0063] Furthermore, the method for discovering anti-ER stress glaucoma intraocular pressure-lowering drugs based on machine learning provided in this embodiment, step S100 also includes:

[0064] Step S150, RDKit feature extraction: using Python and RDKit package, compound features are extracted through SMILES symbols, and the compound features include physical and chemical features and structural features.

[0065] Using Python and RDKit package, 210 compound features were extracted through SMILES (Simplified molecular input line entry system), which can be divided into two categories: ① physicochemical features, including maximum partial charge, molecular weight and number of valence electrons; ② structural features, including molecular connectivity Chi indexes, E-State topological parameters and Kappa shape indexes.

[0066] Step S160, compound labeling: using ER sress(+) compounds and ER stress(-) compounds as labels, annotating the ER sress(+) compounds and ER stress(-) compounds in the ER sress(+) compound training set and the ER stress(-) compound training set, respectively.

[0067] Furthermore, the machine learning method for discovering intraocular pressure-lowering drugs for glaucoma induced by endoplasmic reticulum stress provided in this embodiment comprises step S200:

[0068] Step S210, compound feature optimization: feature selection is performed on the complete data set; the extracted features are sorted.

[0069] First, perform feature selection on the complete dataset to reduce the number of features used for training. Sort the extracted features: Use the random forest method with the Mean Decrease Gini metric to evaluate the impurity metric in the classification task. Sort the features and delete those with a score of 0 to further streamline the features. Perform principal component analysis and calculate the cumulative variance.

[0070] Step S220, model training and optimization: train the sorted features with support vector machine and XGBoost to form a machine learning model, and use corresponding indicators to optimize the machine learning model.

[0071] Ten models, mainly support vector machine (SVM), random forest (RF) and XGBoost, were trained: K-nearest neighbors (K=5), Gaussian Bayes, Logistic regression (LR), decision tree Random forest (RF), Support Vector Machine (SVM), XGBoost, SMOTE+Tomek links with SVM, Stacking ensemble (Estimators: RF, SVM, XGBoost; Final estimator: LR). The model performance was evaluated based on the three indicators of precision, recall and F1 score, and the precision, recall and F1 score were as follows:

[0072] precision=TP / (TP+FP) (1)

[0073] recall=TP / (TP+FN) (2)

[0074] F1 score=TP / (TP+1 / 2(FP+FN)) (3)

[0075] In formulas (1) to (3), precision represents accuracy, TP represents true positive rate, FP represents false positive rate, recall represents recall rate, FN represents false negative rate, and F1 score represents F1 value.

[0076] Preferably, in the method for discovering anti-ER stress glaucoma intraocular pressure-lowering drugs based on machine learning provided in this embodiment, step S200 further includes:

[0077] Step S230, compound library construction: establishing an L3800 Selleck compound library.

[0078] Step S240, test set screening: using the preferred machine learning model to screen the test set mainly based on the L3800Selleck compound library, and using the z-score to sort the predicted probabilities, and screen out multiple compounds with high rankings.

[0079] The test set mainly based on the L3800 Selleck compound library was screened by the optimal model, and the 20 highest-ranked compounds were selected by ranking based on the predicted probability with z-score>8.

[0080] Furthermore, the machine learning method for discovering intraocular pressure-lowering drugs for glaucoma induced by endoplasmic reticulum stress provided in this embodiment comprises step S300:

[0081] Step S310, high-throughput screening and rescreening: high-throughput screening is performed on the top-ranked compounds to screen out the top three compounds; the compounds screened by high-throughput screening are rescreened, dose-response curves are drawn, and IC50 and anti-ER stress index are calculated.

[0082] 384-well high-throughput screening. The cells were divided into two groups: ①Dex-TM (treated with 200nM dexamethasone for 10 days); ②TM (normal control). 10 concentrations were set for each of the 20 drugs. Salubrinal was used as a positive control. High-throughput screening is expected to yield 1-3 drugs, which will enter the rescreening stage, draw dose-response curves, calculate IC50 and anti-ER stress index, and evaluate the similarity between the 1-3 rescreened drugs and the 20 high-throughput screening drugs.

[0083] Step S320, trabecular meshwork cell model verification: the rescreened compounds are verified using the trabecular meshwork cell model to detect TM cell activity and ER stress levels.

[0084] 4 cell models: ①Dex-TM (500nM dexamethasone + DMSO, treated for 10 days); ②TUN-TM (inhibits N-glycosylation and blocks GlcNAc phosphotransferase, 5μg / ml tunicamycin + DMSO (dimethyl sulfoxide) for 72 hours); ③BreA-TM (blocks protein transport from ER to Golgi apparatus, 5μg / ml brefeldin A + DMSO for 72 hours); ④Thap-TM (inhibits ER membrane (referring to the membrane of the endoplasmic reticulum) Ca2+-ATPase 5μg / ml Thapsigargin + DMSO for 72 hours). Dexamethasone is a hormone drug that simulates the aggregation of misfolded myocilin in glaucoma trabecular meshwork cells, and tunicamycin, brefeldin A and Thapsigargin are endoplasmic reticulum stress inducers.

[0085] Detection indicators: WB (Western Blot, protein blotting) and qPCR (quantitative PCR, real-time fluorescence quantitative polymerase chain reaction) methods were used to detect myocilin and ER stress markers (GRP78, sXBP1, CHOP, and GADD34); confocal microscopy was used to detect cell apoptosis under TUNEL staining, the distribution of myocilin, the formation of CLANs by the special structure of the actin skeleton, and the trabecular meshwork cell S. Aureus phagocytosis experiment; electron microscopy was used to detect the endoplasmic reticulum morphology and apoptotic bodies.

[0086] Preferably, the method for discovering anti-ER stress glaucoma intraocular pressure-lowering drugs based on machine learning provided in this embodiment, step S300 also includes:

[0087] Step S330, validation of chronic mouse ocular hypertension model: the rescreened compounds are validated using a chronic mouse ocular hypertension model to detect TM cell apoptosis, ER stress levels, and disease phenotypes, including IOP changes and retinal RGC survival.

[0088] Chronic mouse ocular hypertension model validation: Glucocorticoid-induced chronic ocular hypertension mouse model. Improved on the basis of the eLife2023 study, 0.2% dexamethasone sodium phosphate was administered to the ocular surface 3 times / day for 30 days. Icare Tonolab contact intraocular pressure was tested every 2 days.

[0089] Detection indicators: IOP (intra-Ocular pressure) changes, retinal RGC (Retinalganglion cell) survival. WB and qPCR methods were used to detect myocilin and ER stress markers (GRP78, sXBP1, CHOP, and GADD34); confocal microscopy was used to detect cell apoptosis and myocilin distribution under TUNEL staining; electron microscopy was used to detect endoplasmic reticulum morphology and apoptotic bodies.

[0090] The anti-ER stress glaucoma intraocular pressure lowering drug discovery method based on machine learning disclosed in this embodiment, compared with the prior art, constructs an anti-ER stress compound library: constructs an anti-ER stress compound library through a target-independent strategy; machine learning model training and drug screening: extracts multiple drug physicochemical characteristics and structural characteristics for machine learning model training, performs computer screening on multiple drugs that have been marketed and preclinical, and selects anti-ER stress compounds; glaucoma disease phenotype verification: the screened anti-ER stress compounds are verified in a preset glaucoma cell model and animal model. The anti-ER stress glaucoma intraocular pressure lowering drug discovery method based on machine learning provided in this embodiment will explore new glaucoma intraocular pressure lowering drugs, expand the mechanism of action of existing drugs, establish a target-independent drug discovery model, and fully explore glaucoma drug discovery strategies based on disease phenotypes.

[0091] Although preferred embodiments of the present invention have been described, additional changes and modifications may be made to these embodiments by those skilled in the art once the basic inventive concepts are known. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments and all changes and modifications that fall within the scope of the present invention. Obviously, those skilled in the art may make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.

Claims

1. A method for discovering anti-endoplasmic reticulum stress glaucoma intraocular pressure-lowering drugs based on machine learning, characterized in that: The following steps are involved: Construction of anti-ER stress compound library: Construction of anti-ER stress compound library through target-independent strategy; Machine learning model training and drug screening: extract multiple drug physical and chemical characteristics and structural characteristics for machine learning model training, perform computer screening on multiple marketed and preclinical drugs, and select anti-endoplasmic reticulum stress compounds; Glaucoma disease phenotype validation: The screened anti-ER stress compounds are validated in the preset glaucoma cell model and animal model; The steps of constructing an anti-endoplasmic reticulum stress compound library include: Compound collection: Determine endpoint indicators through existing basic experiments and clinical research data, and search for effective anti-ER stress compounds; Heterogeneity evaluation: clustering the anti-ER stress compounds by K-means algorithm; using Tanimoto distance to evaluate the similarity and heterogeneity of the clustered anti-ER stress compounds; Training set construction: construct ER stress (+) compound training set and ER stress (-) compound training set; Proportion of sample training set: The proportion of ER stress(+) compound samples is unbalanced. In the training set, ER stress(+) compounds account for 7.46%; RDKit feature extraction: Using Python and RDKit package, compound features are extracted through SMILES symbols, including physical and chemical features and structural features; Compound label marking: using ER stress (+) compounds and ER stress (-) compounds as labels, annotating the ER stress (+) compounds and the ER stress (-) compounds in the ERstress (+) compound training set and the ER stress (-) compound training set, respectively; The steps of machine learning model training and drug screening include: Compound feature optimization: feature selection for the complete data set; sorting the extracted features; Model training and optimization: The sorted features are trained with support vector machines and XGBoost to form a machine learning model, and the machine learning model is optimized using corresponding indicators.

2. The method for discovering anti-endoplasmic reticulum stress glaucoma intraocular pressure-lowering drugs based on machine learning as claimed in claim 1, characterized in that: The step of optimizing the compound characteristics comprises: Use random forest method to evaluate impurity metrics in classification tasks; Sort the extracted features and delete those with a score of 0.

3. The method for discovering anti-endoplasmic reticulum stress glaucoma intraocular pressure lowering drugs based on machine learning as claimed in claim 1, characterized in that: In the steps of model training and optimization, three indicators, namely precision, recall and F1 score, are used to evaluate the machine learning model. The precision, recall and F1 score are respectively: precision=TP / (TP+FP) recall=TP / (TP+FN) F1 score = TP / (TP+1 / 2(FP+FN)) Among them, precision represents accuracy, TP represents true positive rate, FP represents false positive rate, recall represents recall rate, FN represents false negative rate, and F1 score represents F1 value.

4. The method for discovering anti-endoplasmic reticulum stress glaucoma intraocular pressure lowering drugs based on machine learning as claimed in claim 3, characterized in that: The steps of machine learning model training and drug screening also include: Compound library construction: Establish L3800 Selleck compound library; Test set screening: The test set mainly composed of the L3800 Selleck compound library was screened using the preferred machine learning model, and the predicted probabilities were ranked using the z-score to screen out multiple compounds with high rankings.

5. The method for discovering anti-endoplasmic reticulum stress glaucoma intraocular pressure lowering drugs based on machine learning as claimed in claim 4, characterized in that: The steps of glaucoma disease phenotype verification include: High-throughput screening and rescreening: Perform high-throughput screening on the top-ranked compounds to screen out the top three compounds; rescreen the high-throughput screened compounds, draw dose-response curves, and calculate IC50 and anti-ERstress index; Trabecular meshwork cell model validation: The rescreened compounds were validated using the trabecular meshwork cell model to detect TM cell activity and ER stress levels.

6. The method for discovering anti-endoplasmic reticulum stress glaucoma intraocular pressure lowering drugs based on machine learning as claimed in claim 5, characterized in that: The step of verifying the glaucoma disease phenotype also includes: Validation in chronic mouse ocular hypertension model: The rescreened compounds were validated using a chronic mouse ocular hypertension model to detect TM cell apoptosis, ER stress levels, and disease phenotypes, including IOP changes and retinal RGC survival.

Images