Drug permeability clinical prediction model, construction method thereof and non-specific drug permeability evaluation method

Abstract

The invention provides a drug permeability clinical prediction model, a construction method thereof and a non-specific drug permeability evaluation method, and belongs to the technical field of biological medicine. The invention provides a construction method of a drug permeability clinical prediction model. According to the construction method, intestinal microorganism characteristics are introduced into model construction for the first time. The prediction performance of the constructed LGBM model containing the intestinal microorganism characteristics in the aspect of NAC permeability prediction is superior to that of other traditional business models. The invention also provides a serine-derived sphingolipid de novo synthesis pathway, which is different from the traditional sphingolipid de novo synthesis pathway, and is named as a cysteine sphingolipid metabolism pathway, NAC-derived cysteine may participate in the biosynthesis of microbial sphingolipid through covalent binding with palmitoyl-CoA, thereby damaging intestinal permeation of NAC.

Description

Technical Field

[0001] The present invention belongs to the field of biomedicine technology, and in particular relates to a drug permeability clinical prediction model and a construction method thereof, and a non-specific drug permeability evaluation method. Background Art

[0002] Worldwide, the pharmaceutical industry has placed great demands on reducing the cost, time and number of clinical drug trials during the development / approval of new dosage forms of generic drugs. The BCS evaluation marks a major breakthrough in the classification of drugs based on their water solubility and intestinal permeability. However, the current intestinal permeability evaluation system ignores the metabolic effects of about 1013 commensals from more than 250 different bacteria as well as fungi, viruses and archaea. The overall picture of drug-bacteria interactions shows that bioaccumulation and metabolism by intestinal bacteria greatly alter drug bioavailability, thereby affecting drug pharmacokinetics, side effects and individual clinical responses. NAC, a derivative of the amino acid L-cysteine, has been approved by the FDA for decades as a powerful antioxidant for the treatment of respiratory diseases. In 2021, NAC was even added as a dietary supplement to prevent severe liver damage caused by acetaminophen / alcohol intoxication due to its high water solubility (163.9 mg / ml at pH 7.4). Nevertheless, due to the low bioavailability of NAC fraction, once NAC enters the blood, it binds to plasma proteins and forms disulfide bridges through sulfhydryl groups, leading to side effects such as facial flushing, nausea, and vomiting. Most importantly, BCS class I / III oral NAC is still undecided in the pharmaceutical industry. The reason for the inconsistent permeability should be the development of NAC delivery technology will be the main issue to improve its bioavailability.

[0003] Machine learning (ML) is an emerging medical field with a wealth of resources for integrating computer science and statistics with medical problems. ML proponents praise its ability to handle large, complex, and diverse data commonly found in medicine, and believe that ML is the future of biomedical research, personalized medicine, and computer-aided diagnosis, which can significantly advance global healthcare. Machine learning is a scientific discipline that focuses on how computers learn from data. It emerges at the intersection of statistics and computer science, with the former attempting to learn relationships from data and the latter emphasizing efficient computational algorithms. This marriage between mathematics and computer science is driven by the unique computational challenges of building statistical models from massive data sets that may include billions or trillions of data points. The types of learning used by computers can be conveniently subdivided into categories such as supervised learning and unsupervised learning. However, data models for clinical prediction of drug permeability are still lacking. Summary of the invention

[0004] The present invention provides a drug permeability clinical prediction model and a construction method thereof and a nonspecific drug permeability evaluation method. Based on the nonspecific drug permeability evaluation strategy of the intestinal microbiota, it is found that cysteine ​​palmitoylation of the intestinal microbiota has a great influence on the intestinal permeability of NAC, resulting in its BCS I / III class inconsistency, providing important treatment strategies and theoretical basis for clinical medicine.

[0005] The present invention provides a method for constructing a drug permeability clinical prediction model, comprising the following steps:

[0006] Obtain intestinal microbial characteristics data;

[0007] The intestinal microorganism characteristic data is input into the prediction model to obtain the predicted Fsys characteristic value; the prediction model is determined by training based on a feature training set using a machine learning algorithm; the feature training set is the trained intestinal microorganism characteristic data; the feature training set is obtained by feature processing the training data using a nearest neighbor clustering algorithm and a recursive feature elimination method;

[0008] Determine the drug permeability according to the predicted Fsys characteristic value; the drug permeability is used to provide treatment strategies and theoretical basis for clinical medicine;

[0009] Wherein, the method for determining the prediction model is:

[0010] Constructing a network model set; the network model set includes: multivariate linear regression model (MLR), neural network model (NN), k-nearest neighbor model (k-NN), random forest model (RF), decision tree model (DT), partial least squares model (PLS), support vector regression model (SVR), extreme gradient boosting model (XGB) and lightweight gradient boosting algorithm model (LGBM);

[0011] Input the feature training set into each network model in the network model set for training, and adopt a nested cross-validation method to select a model with the smallest absolute error from multiple network models as a preliminary optimization model;

[0012] The random search method and the grid search method are used to adjust the hyperparameters in the preliminary optimization model with the negative mean absolute error as the measurement benchmark to obtain the prediction model.

[0013] In a specific embodiment of the present invention, the intestinal microbial characteristic data include: abundance of Bacteroides thetaiotaomicron, Cysteine-Cer (d18:1 / 16:0), Cysteine-3-ketosphinganine, intestinal pH, Cysteine-Cer (d18:1 / 24:0), Cysteine-Sphingosine (d18:0), Cysteine-Cer (d18:1 / 24:1), intestinal blood flow, Cysteine-DHCer (d18:1 / 16:0), Rhabditis fragilis Bacteria abundance , Cysteine-DHCer (d18:0 / 24:0), Cysteine-DHCer (d18:0 / 24:1), and viscosity of intestinal contents.

[0014] In one embodiment of the present invention, the drug comprises N-acetylcysteine.

[0015] In a specific embodiment of the present invention, the intestinal microorganisms include at least one of the following: Bacteroides thetaiotaomicron and Bacteroides fragilis.

[0016] The present invention proposes that microbial serine palmitoyltransferase (SPT) catalyzes the covalent binding between NAC-derived cysteine ​​and palmitoyl-CoA to produce cysteine-3-ketosphinganine, thereby impairing the intestinal permeability of NAC.

[0017] The present invention also provides the application of the drug permeability clinical prediction model constructed by the above drug permeability clinical prediction model construction method in predicting drug permeability.

[0018] In a specific embodiment of the present invention, the drug permeability clinical prediction model includes a lightweight gradient boosting machine algorithm model.

[0019] The present invention also provides a method for screening oral drugs with individual variable permeability using a drug permeability clinical prediction model constructed by the above-mentioned drug permeability clinical prediction model construction method, comprising the following steps: constructing a control group using an animal model, wherein a first control group has no intestinal microbial colonization, and a second control group has regular intestinal microbial colonization;

[0020] The control group was subjected to microbiome and metabolome analysis before a single intestinal infusion of the drug, and the intestinal in situ P of the drug-perfused group was detected every 15 minutes after the infusion of the drug. effvalue, analyzing the interaction between the drug and the intestinal microbial strain to obtain potential bacteria-drug pairs; the interaction is defined by the false discovery rate, wherein the false discovery rate corrected P value critical value is 0.05, and the exhaustion threshold is 30%; the detection includes using the drug permeability clinical prediction model for detection;

[0021] Based on the in vivo communities obtained from the microbiome, bacteria-drug pairs with a Z factor > 0.8 were screened to identify oral drugs with individual variable permeability.

[0022] In one embodiment of the present invention, the animal model comprises Wistar rats.

[0023] In one embodiment of the present invention, six rats in each group were fasted for 12 hours and had free access to water. Before anesthesia, a mixture of ketamine and hydroxyzine (80 mg / kg and 20 mg / kg) was injected intramuscularly and maintained on a heated surface at 37±1°C. The intestine was exposed through a 3-4 cm midline incision, and the proximal part of the jejunum of about 10 cm was cannulated at both ends. Under 37°C, a blank perfusate was pumped into the intestine at a flow rate of 0.5 mL / min using a peristaltic pump for about 30 minutes, and then a perfusate containing 30 mg / L NAC was passed through the intestinal cavity at a constant flow rate of 0.2 mL / min under 37°C. In order to determine the steady state during the perfusion process, perfusate samples were collected from the distal jejunum every 15 minutes (15, 30, 45, 60, 75, 90, 105 and 120 minutes). These samples were collected in pre-weighted vials to monitor the effective permeability (P eff ) value remains constant, indicating that the system has reached a stable state. The sample was immediately frozen at -20°C for LC-MS / MS analysis. eff (cm / s) is calculated by the following formula:

[0024]

[0025] Where: Qin = inlet perfusion flow (cm 3 / s), A(2πrl) = absorbable intestinal area considering radius and intestinal segment length (cm2), Cout = outlet drug concentration (mg / mL), Qout = outlet perfusion flow (cm3 / s), Cin = inlet drug concentration (mg / mL).

[0026] In one embodiment of the present invention, the Z-factor value is a relative indicator for distinguishing between signal and background populations. It is a parameter without measurement, and its range can be from "1 to less than 0". When the Z-factor is equal to zero, the signal and the background begin to overlap. Generally speaking, an acceptable Z-factor should be greater than 0.4, and a variety of drugs including NAC show typical interactions with Bacteroides. Among the 32 in vivo communities tested, 12 drugs with Z factors>0.5 were identified in 27 pairs of individuals, and 3 drugs with Z factors>0.8 were identified in 36 pairs of individuals.

[0027] Beneficial effects: The present invention provides a method for constructing a clinical prediction model for drug permeability, and for the first time incorporates intestinal microbial characteristics into model construction. The LGBM model containing intestinal microbial characteristics constructed in the embodiment of the present invention has better predictive performance in predicting NAC permeability than other traditional commercial models. The present invention found that cysteine ​​palmitoylation of the intestinal microbiota has a great influence on the intestinal permeability of NAC, resulting in its BCS I / III class inconsistency. It has been experimentally verified that the composition of the intestinal microbiota may be related to the permeability of NAC, especially to symbiotic intestinal bacteria such as B. thetaiotaomicron and B. fragilis. The Pearson correlation coefficient showed that B. thetaiotaomicron / B. fragilis was significantly positively correlated with 3-ketosphingosine and sphingosine, but negatively correlated with palmitoyl-CoA and serine.

[0028] The present invention also proposes a de novo synthesis pathway of sphingolipids derived from serine that is different from the traditional one, and names it "cysteine ​​sphingolipid metabolic pathway". NAC-derived cysteine ​​may participate in the biosynthesis of microbial sphingolipids by covalently binding to palmitoyl-CoA, thereby impairing the intestinal permeability of NAC. Using the method described in the present invention, hierarchical clustering demonstrated that patient populations with different combinations of high B. thetaiotaomicron, homocysteine ​​Cer (d18:1 / 16:0), homocysteine ​​Cer (d18.1 / 24:0), high intestinal pH, and high B. fragilis tend to include population subgroups with the following characteristics: an increasing chance of predicting high Fsys, while patients with lower combinations of these characteristic values ​​tend to have a decreasing chance of predicting high Fsys. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] Figure 1 Schematic diagram of the workflow of the microbiome-mediated oral drug permeability screening platform;

[0030] Figure 2 Figure 2: Intestinal P of NAC (upper panel), lithium carbonate (middle panel), and Panicillin vk (lower panel) monitored every 15 minutes for 2 hours. effValue diagram;

[0031] Figure 3 Schematic diagram of the discovery and validation cohorts used to generate and test the multiparameter model;

[0032] Figure 4 Schematic diagram of potential factors associated with NAC absorption in patients;

[0033] Figure 5 The bee swarm diagram illustrates the effect of each feature on Fsys according to the SHAP value;

[0034] Figure 6 Schematic diagram of Fsys prediction by combining the relative contribution of each feature through the LGBM model;

[0035] Figure 7 To obtain the global map predicted by NACFsys, the force graphs of all patients in the cohort were rotated 90° counterclockwise;

[0036] Figure 8 Schematic diagram of sphingolipids in correlation with clinical indicators and intestinal function;

[0037] Fig. 9 Schematic diagram of the confusion matrix of the prediction results of the LGBM model and the ACAT, GITA and ADAM models;

[0038] Fig.10 Design a global picture for the study. DETAILED DESCRIPTION

[0039] The present invention provides a method for constructing a drug permeability clinical prediction model, comprising the following steps:

[0040] Obtain intestinal microbial characteristics data;

[0041] The intestinal microorganism characteristic data is input into a prediction model to obtain a predicted Fsys (absolute bioavailability) characteristic value; the prediction model is determined by training based on a feature training set using a machine learning algorithm; the feature training set is the trained intestinal microorganism characteristic data; the feature training set is obtained by feature processing the training data using a nearest neighbor clustering algorithm and a recursive feature elimination method;

[0042] Determine the drug permeability according to the predicted Fsys characteristic value; the drug permeability is used to provide treatment strategies and theoretical basis for clinical medicine;

[0043] Wherein, the method for determining the prediction model is:

[0044] Constructing a network model set; the network model set includes: multivariate linear regression model (MLR), neural network model (NN), k-nearest neighbor model (k-NN), random forest model (RF), decision tree model (DT), partial least squares model (PLS), support vector regression model (SVR), extreme gradient boosting model (XGB) and lightweight gradient boosting machine algorithm model (LGBM);

[0045] Input the feature training set into each network model in the network model set for training, and adopt a nested cross-validation method to select a model with the smallest absolute error from multiple network models as a preliminary optimization model;

[0046] The random search method and the grid search method are used to adjust the hyperparameters in the preliminary optimization model with the negative mean absolute error as the measurement benchmark to obtain the prediction model.

[0047] In a specific embodiment of the present invention, the intestinal microbial characteristic data include: abundance of Bacteroides thetaiotaomicron, Cysteine-Cer (d18:1 / 16:0), Cysteine-3-ketosphinganine, intestinal pH, Cysteine-Cer (d18:1 / 24:0), Cysteine-Sphingosine (d18:0), Cysteine-Cer (d18:1 / 24:1), intestinal blood flow, Cysteine-DHCer (d18:1 / 16:0), Rhabditis fragilis Bacteria abundance , Cysteine-DHCer (d18:0 / 24:0), Cysteine-DHCer (d18:0 / 24:1), and viscosity of intestinal contents.

[0048] To develop a machine learning (ML) model incorporating gut microbial signatures, common gut SLs signatures were initially incorporated into the risk-based model. That is, the training data included age, prescription, drug dosage, intestinal blood flow, intestinal content viscosity, intestinal pH, intestinal transit rate, intestinal ceramide metabolome (Cysteine-3-ketosphinganine, Cysteine-Sphingosine (d18:0), Cysteine-Cer (d18:1 / 16:0), Cysteine-Cer (d1 8:1 / 24:0), Cysteine-Cer (d18:1 / 24:1), Cysteine-DHCer (d18:1 / 16:0), Cysteine-DHCer (d18:0 / 24:0), Cysteine-DHCer (d18:0 / 24:1)) and 17 characteristic values ​​including intestinal microbial groups (Bacteroides thetaiotaomicron abundance, Bacteroides fragilis abundance).

[0049] After eliminating the four least important features (prescription, drug dosage, intestinal transit rate, age) through recursive feature extraction, 13 features (polymorphic rod Bacteria abundance 、Cysteine-Cer(d18:1 / 16:0), Cysteine-3-ketosphinganine, intestinal pH, Cysteine-Cer(d18:1 / 24:0), Cysteine-Sphingosine(d18:0), Cysteine-Cer(d18:1 / 24:1), intestinal blood flow, Cysteine-DHCer(d18:1 / 16:0), Rhabditis fragilis Junfeng Spend , Cysteine-DHCer (d18:0 / 24:0), Cysteine-DHCer (d18:0 / 24:1), intestinal content viscosity).

[0050] Then, nine different ML regression algorithms are introduced to improve stability and accuracy, including multiple linear regression (MLR), neural network (NN), k-nearest neighbor (k-NN), random forest (RF), decision tree (DT), partial least squares (PLS), support vector regression (SVR), extreme gradient boosting (XGB), and lightweight gradient boosting machine (LGBM).

[0051] Finally, the LGBM model with 13 features (excluding prescription, NAC dose, intestinal transit rate, and age) achieved the lowest MAE, thus confirming the applicability of the LGBM model.

[0052] The present invention uses an optimized LGBM model of 13 features and finds that B.thetaiotaomicron has the greatest impact on the individual dose fraction (Fsys) entering the systemic circulation. Cysteine-Cer (d18:1 / 16:0), cysteine-3-ketosphingine and intestinal pH contribute similarly to Fsys. Patient groups with higher levels of B.thetaiotaomicron, cysteine-Cer (d18:1 / 16:0), cysteine-Cer (d18:1 / 24:0), intestinal pH and B.fragilis often constitute a subgroup of the population with an increased chance of predicting high Fsys.

[0053] In a specific embodiment of the present invention, the drug includes N-acetylcysteine ​​(NAC); the dosage form of the drug can be a tablet. As an embodiment, in the process of preparing tablets, fillers include starch, microcrystalline cellulose, sucrose, dextrin, lactose, powdered sugar, glucose, etc.; lubricants include magnesium stearate, stearic acid, sodium chloride, sodium oleate, sodium lauryl sulfate, poloxamer, etc.; binders include water, ethanol, starch slurry, syrup, hydroxypropyl methylcellulose, sodium carboxymethyl cellulose, sodium alginate, polyvinyl pyrrolidone, etc.; disintegrants include starch effervescent mixtures, i.e., sodium bicarbonate and citric acid, tartaric acid, low-substituted hydroxypropyl cellulose, etc.; suspending agents include polysaccharides such as acacia gum, agar, alginic acid, cellulose ethers, and carboxymethyl chitosan, etc.; the solvent includes water, a balanced salt solution, etc.

[0054] The medicine of the present invention contains adjuvants, which include one or more of a carrier, an excipient, a diluent, a lubricant, a wetting agent, an emulsifier, a suspension stabilizer, a preservative, a sweetener, and a flavor. The adjuvants can be selected from one or more of lactose, glucose, sucrose, sorbitol, mannose, starch, gum arabic, calcium phosphate, alginate, gelatin, calcium silicate, fine crystalline cellulose, polyvinyl pyrrolidone (PVP), cellulose, water, syrup, methylcellulose, methyl hydroxybenzoate, propyl hydroxybenzoate, talc, magnesium stearate, or mineral oil.

[0055] In a specific embodiment of the present invention, the intestinal microorganisms include at least one of the following: Bacteroides thetaiotaomicron and Bacteroides fragilis.

[0056] The present invention also provides the application of the drug permeability clinical prediction model constructed by the above drug permeability clinical prediction model construction method in predicting drug permeability.

[0057] In one embodiment of the present invention, the drug permeability clinical prediction model includes the LGBM model. In the present embodiment, through the screening of intestinal flora-mediated permeability, it was found that cysteine ​​palmitoylation of intestinal flora greatly interfered with the absorption of NAC in the intestine; and the optimized 13-feature LGBM model provided clinically relevant enhancement in Fsys prediction performance.

[0058] The present invention also provides a method for screening oral drugs with individual variable permeability using a drug permeability clinical prediction model constructed by the above-mentioned drug permeability clinical prediction model construction method, comprising the following steps: constructing a control group using an animal model, wherein a first control group has no intestinal microbial colonization, and a second control group has regular intestinal microbial colonization;

[0059] The control group was subjected to microbiome and metabolome analysis before a single intestinal infusion of the drug, and the intestinal in situ P of the drug-perfused group was detected every 15 minutes after the infusion of the drug. eff value, analyzing the interaction between the drug and the intestinal microbial strain to obtain potential bacteria-drug pairs; the interaction is defined by the false discovery rate, wherein the false discovery rate corrected P value critical value is 0.05, and the exhaustion threshold is 30%; the detection includes using the drug permeability clinical prediction model for detection;

[0060] Based on the in vivo communities obtained from the microbiome, bacteria-drug pairs with a Z factor > 0.8 were screened to identify oral drugs with individual variable permeability.

[0061] In one embodiment of the present invention, the animal model comprises Wistar rats.

[0062] In one embodiment of the present invention, 32 oral drugs with individually variable permeabilities were screened and it was determined that the intestinal sphingomyelin (SL) production capacity of B. thetaiotaommicron and B. fragilis affected the intestinal absorption of NAC, which was attributed to SPT-catalyzed cysteine ​​palmitoylation. The present invention narrowed the scope of the impact of SPT deficiency on NAC permeability and confirmed that the P of NAC eff The values ​​showed a negative linear dependence on cystine-3-oxosphingosine through covalent binding with palmitoyl-CoA.

[0063] The present invention also recruited 79 prospective cohorts and 48 independent validation cohorts, initially incorporating the common intestinal SLs features into the risk-based model, establishing input features describing NAC Fsys, including available age, prescription, NAC dose, intestinal blood flow, intestinal content viscosity, intestinal pH, intestinal transit rate, intestinal ceramide metabolome and microbiota input feature profiles, and introduced 9 different ML regression algorithms to improve their stability and accuracy, including multivariate linear regression (MLR), neural network (NN), k-nearest neighbor (k-NN), random forest (RF), decision tree (DT), partial least squares (PLS), support vector regression (SVR), extreme gradient boosting (XGB) and lightweight gradient boosting (LGBM). The selected ML model panel was trained and evaluated using a nested cross-validation strategy. In the inner loop (Discovery Cohort) of each model, a randomized study and grid study method with k-fold (k=10) cross-validation was used for the hyperparameter optimization process. The range of all hyperparameters was considered, as well as their final selected values. Subsequently, the models with these fixed hyperparameter structures were evaluated using the mean absolute error (MAE) of the outer loop test dataset (Validation Cohort). Among them, the LGBM model had the best prediction effect, with a MAE value of 3.884 (±0.772), and the smallest deviation compared with the RF, DT, and XGB models with similar MAE values. In order to further verify the prediction accuracy of the four algorithms, the models constructed using the classifier algorithm were predicted again with Fsys=85% as the threshold. Receiver operating characteristic (ROC) curve analysis showed that the LGBM model had the best prediction performance, with the highest AUC value of 0.841. It was confirmed that the LGBM model containing intestinal microbial characteristics had better performance in predicting NAC permeability than other commercial models.

[0064] To further illustrate the present invention, a drug permeability clinical prediction model and its construction method and a non-specific drug permeability evaluation method provided by the present invention are described in detail below in conjunction with embodiments, but they should not be construed as limiting the scope of protection of the present invention.

[0065] In the embodiments of the present invention, for the contents without specific techniques or conditions, the techniques or conditions described in the literature in the field or the product instructions can be used. The reagents or instruments used without indicating the manufacturer are all conventional products that can be purchased through regular channels. The experimental methods in the embodiments are all conventional methods unless otherwise specified. Unless otherwise specified, the percentages and parts involved in the present invention are all weight percentages and weight parts. The overall scheme flow of the present invention is as follows Fig.10 shown.

[0066] Example 1 Animal Experiment

[0067] Materials and methods: All animal manipulations were performed in accordance with protocols approved by the Experimental Animal Center of Peking University First Hospital (ethics number 2023-59332). Wistar GF or SPF rats (6 weeks old, 250-300 g) were housed 2 per cage with a 12 h light-dark cycle. Six rats in each group were fasted for 12 h with free access to water and then anesthetized with an intramuscular injection of a ketamine-thiazide mixture (80 mg / kg and 20 mg / kg) and placed on a heated surface maintained at 37 ± 1 °C. The intestine was exposed through a 3-4 cm midline incision, and the proximal part of the jejunum was cannulated for approximately 10 cm at both ends. Blank perfusate at 37 °C was pumped into the intestine at a flow rate of 0.5 mL / min for approximately 30 min by a peristaltic pump (Minipuls-3, Gilson, France), and then perfusate containing 3 mg / kg NAC was injected at a constant flow rate of 0.2 mL / min at 37 °C. After steady state (i.e., flow rate and concentration at each point in the perfusion segment remained constant over time), perfusate was collected from the distal jejunum at 15-min intervals (15, 30, 45, 60, 75, 90, 105, and 120 min), placed in pre-weighed vials, and immediately frozen at −20 °C until LC-MS / MS analysis.

[0068] The net water flux (NWF, intestinal water absorption and outflow) was determined by gravimetric method, and the drug concentration in the perfusion solution of each substance was corrected to calculate the effective permeability (P eff ). After sampling, the animals were euthanized by intracardiac injection of saturated potassium chloride solution. After death, the perfused and non-perfused intestinal segments were removed for microbiome determination. Intestinal fluid from each intestinal segment (perfused and non-perfused) was collected for subsequent microbiological determination (see Metabolomic Analysis of Gut Microbiota and Taxonomic Profile of Gut Microbiota).

[0069] In order to establish a screening platform for 32 oral drugs with individual variable permeability that may be affected by the intestinal microbiota, this example used Wistar rats to perform microbiome and metabolome analysis before single intestinal perfusion (SPIP) in the absence of intestinal microbial colonization (bacteria-free, GF) and conventional colonization (specific pathogen-free, SPF) controls. The intestinal in situ P of these drugs was monitored every 15 minutes. effvalues. In the screening protocol, the interactions between drugs (29 out of 32) oral drugs and 18 representative microbial strains were systematically analyzed. 60 potential bacteria-drug pairs were tested in two independent screens. A false discovery rate (FDR)-corrected P value cutoff of 0.05 and an exhaustion threshold of 30% were used to define interactions. Among the 32 in-situ communities, 12 drug pairs with Z factors > 0.5 were identified in 27 pairs and 3 drug pairs with Z factors > 0.8 were identified in 36 pairs. The three oral drugs with the highest Z factors, NAC, lithium carbonate, and Panicillinvk, showed significant differences in intestinal P values ​​between SPF and GF rats. eff The values ​​are different.

[0070] The results are as follows Figures 1-2 As shown, for each drug, age-matched SPF and GF rats were used to perform a single intestinal perfusion (SPIP) model. There were 6 rats in each group. The absorption rates of the three drugs in SPF rats increased from 0.649×10 -4 cm / s slowly increased to 0.706×10 -4 In sharp contrast, the intestinal absorption of NAC in GF rats increased dramatically to 1.457×10 -4 cm / sP eff platform, which was 2 times higher than that of SPF rats.

[0071] Example 2

[0072] Materials and Methods

[0073] After fasting overnight, 127 participants (64 males and 63 females, aged 22-69 years) obtained intestinal fluid samples by simultaneous duodenal and jejunal aspiration through two double-lumen catheters. The catheter was introduced and placed orally with the proximal end in the duodenum (5-10 cm from the pylorus) and the distal end 90 cm below the duodenum. The position of the catheter was confirmed by fluoroscopy. The Ethics Committee of Peking University First Hospital approved the study protocol (Ethics Number: 2023-166). All participants in this example signed written informed consent.

[0074] The results are as follows Figure 3 As shown, the model was trained on a discovery cohort that included available age, prescription, NAC dose, intestinal blood flow, intestinal content viscosity, intestinal pH, intestinal transit rate, intestinal ceramide metabolome, and microbiota composition (n = 192 participants) and subsequently validated on an independent validation cohort (n = 48 participants).

[0075] Example 3 Data segmentation strategy for ML model training

[0076] The ML models were trained and evaluated using a nested cross-validation approach, which consisted of an inner (training and validation) and outer (testing) loop. The nested cross-validation approach was performed 10 times for each ML model to evaluate the overall performance of the algorithm. For each nested cross-validation implementation, 20% of the samples in the dataset were randomly reserved as the outer loop for model performance testing using the train-test-split method in the Scikit-learn library in Python41. The remaining 80% of the samples used for hyperparameter selection and model training were used as the inner loop. The random search cv and grid search cv methods in the Scikit-learn library in Python41 were combined for hyperparameter selection: first, random search cv (with 100 random hyperparameter configurations in each case) was used to narrow the range of hyperparameter selection, and then grid search cv was used to determine the best hyperparameters. During this process, the data was split again according to the k-fold cross-validation strategy (k=10). 90% of the samples in the inner loop were used to train various model architectures with different hyperparameters, and the other 10% were used for hyperparameter evaluation. The performance of the model structure (i.e., fixed hyperparameter structure) selected by the inner loop is tested by the outer loop using the holdout dataset (i.e., 20% of the samples in the initial dataset).

[0077] The results are as follows Figures 4 to 7 As shown, patients in the validation cohort were classified as having a Fsys ≥ 85% ( Figure 4 , left) and Fsys<85% ( Figure 4 , right); the honeycomb plot reveals the overall contribution of the input features in terms of SHAP values ​​( Figure 5 ); gut microbial characteristics were identified as the first three key features ( Figure 6 ); Hierarchical clustering showed that patient groups with different combinations of high B.thetaiotaomicron, homocysteine-Cer (d18:1 / 16:0), homocysteine-Cer (d18:1 / 24:0), high intestinal pH, and high B.fragilis tended to constitute subgroups with increasingly higher chances of predicting high Fsys. Whereas patient groups with combinations of these lower feature values ​​tended to predict increasingly lower chances of predicting high Fsys.

[0078] Example 4 Engineering Properties

[0079] Cluster analysis was performed using the hierarchical clustering method in the Scipy package in Python. The initial input features were arranged into a hierarchy of clusters using the nearest neighbor clustering algorithm. The performance of the optimal ML algorithm (LGBM) was evaluated by removing selected features from the Scikit-learn library using the recursive feature elimination (REF) method. The hyperparameter structure of the LGBM model, determined using an inner loop 10-fold cross validation, varied only with the number of input features. The results are shown in Figure 8As shown, the LGBM model was successfully constructed.

[0080] Example 5 Machine Learning Model Development and Evaluation

[0081] In this example, a total of 9 machine learning algorithms were trained and studied, including LGBM, MLR, NN, k-NN, RF, DT, PLS, SVR, and XGB models, all of which were built and calculated in Python. The LGBM and XGB models were built using the LightGBM package and the XGBoost package, and all other models were built using the Scikit-learn library. The learning curve designed using the Scikit-learn library was used to verify whether the sample size was sufficient to build a model with stable predictive performance. The hyperparameters in the ML model were tuned using the random search cv and grid search cv methods. In order to evaluate the predictive accuracy of the trained ML model, the predictions of each outer loop (k=10) separated from the nested cross validation were compiled into a data frame for analysis. By determining the absolute error (AE), R 2 The values ​​of AE, MAE, R 2 The calculation formula of RMSE is as follows:

[0082] Absolute error(AE)=|predicted yi-true yi;

[0083]

[0084]

[0085]

[0086] Among them, predicted yi is the predicted value; true yi is the true value; i is the sample number; n is the number of samples; mean true yi is the average true value.

[0087] The results are as follows Fig. 9As shown in the results, the average accuracy, sensitivity, specificity, positive predictive value (PPV) and negative predictive value (NPV) of the LGBM model were 67.1%, 78.6%, 64.6%, 32.4% and 93.3%, respectively, which were much better than the ACAT (68.4%, 21.4%, 78.5%, 17.5% and 82.3%), GITA (58.2%, 78.6%, 53.8%, 26.8% and 92.1%) and ADAM (58.2%, 71.4%, 55.4%, 25.6% and 90.0%) models. Compared with the ACAT model (NRI = 0.433, P = 0.042) and the ADAM model (NRI = 0.164, P < 0.001), the LGBM model significantly improved the performance of Fsys prediction. Taken together, these findings suggest that the LGBM model incorporating gut microbial signatures has superior predictive performance in predicting NAC permeability compared with other models.

[0088] Example 6 Potential Mechanisms of the Effect of Intestinal Microbiota Metabolism on NAC Permeability

[0089] Figure 4 The metabolome dataset of the heat map shown shows 21 intestinal metabolites with significant differences. Among them, a series of intestinal sphingolipid biosynthesis molecules such as palmitoyl-CoA and serine were derived from the intestinal fluid of GF rats, and the content was higher than that of SPF rats, while the content of 3-ketosphinganine, sphingosine, and 6-hydroxysphinegosine was lower than that of SPF rats. KEGG enrichment of these multiple differences mainly focused on sphingolipid metabolism and amino acid metabolism. Then, the present invention divided the microbial classification of intestinal fluid samples from NAC-perfused and NAC-unperfused segments of SPF rats. Through ultra-deep metagenomic sequencing, the LEfSe algorithm was applied to explore the most enriched intestinal microbial communities. The results showed that the most enriched bacteria in the NAC-perfused segments were mainly B. thetaiotaomicron and B. fragilis relative to the NAC-unperfused segments. Real-time fluorescence quantitative PCR confirmed that the significantly differential bacteria in SPF rats after NAC perfusion were B. thetaiotaomicron and B. fragilis. From a phylogenetic point of view, most bacterial species cannot produce sphingolipids. However, Bacteroides is one of the most dominant commensal genera in the microbiota, which can produce and provide ceramides containing odd (C17:0) and even (C18:0) hydrocarbons. This is the same as mammalian ceramide-induced persistent (moderate) deterioration of colonized mucosa to interfere with drug permeability.

[0090] Subsequently, the diversity of intestinal flora in each group of mice was compared for α-diversity evaluation, and it was found that the ACE index, Chao1 index, Shannon index, and Simpson index of the NAC-perfused group were comparable to those of the NAC-unperfused group. The β-diversity of the intestinal microbiota assessed by ANOSIM was also not different between the NAC-perfused and NAC-unperfused groups based on PCoA scores (PC1: 12.12% and PC2: 8.38%), NMDS stress scores (0.1883), or multi-sample dilution curves. The composition of the intestinal microbiota may be related to the permeability of NAC, especially to commensal intestinal bacteria such as B. thetaiotaomicron and B. fragilis. The Pearson correlation coefficient showed that B. thetaiotaomicron / B. fragilis was significantly positively correlated with 3-ketosphingosine and sphingosine, but negatively correlated with palmitoyl-CoA and serine (p < 0.01).

[0091] The induced matching docking of palmitoyl-CoA with serine and cysteine ​​in SPT was analyzed. Comparison with the serine / cysteine-bound SPT structure revealed similar recognition features as follows: the thiol moiety of cysteine ​​interacted with R123, Y401, L484, and A486 in almost the same way, corresponding to the serine-bound R123, Y401, E487, and A486 in SPT. Therefore, the present invention proposes a "cysteine ​​sphingolipid metabolic pathway" that is different from the traditional pathway for de novo synthesis of sphingolipids derived from serine. Through ultra-performance liquid chromatography coupled to time-of-flight mass spectrometry (UPLC-ESI-QTOF / MS) metabolite profiling, the present invention further quantified that palmitoyl-CoA was significantly increased and cysteine-3-ketosphingine was significantly decreased in GF rats compared with SPF rats. This suggests that oral NAC-derived cysteine ​​may participate in the biosynthesis of microbial sphingolipids through covalent binding to palmitoyl-CoA, thereby impairing the intestinal permeation of NAC.

[0092] Since the penetration of NAC in vivo is closely related to the sphingolipid biosynthesis of B. thetaiotaomicron and B. fragilis, mutant strains with inactivated SPT (BT△SPT and BF△SPT) were biosynthesized to disrupt their sphingolipid production. Differentiated Caco-2 cells were co-cultured with wild-type B. thetaiotaomicron (BTWT), B. fragilis (BFWT), BT△SPT or BF△SPT cells, and NAC and palmitic acid azide (PAA) were added to establish bioorthogonal chemical reactions; the immunofluorescence intensity of each group was continued to be quantified by cycloaddition of cyclopyrimidine coupling detection reagent with Alexa Flour 647 cyclopyrimidine (AF647-Cyc). The azide-labeled SLs of Caco2 cells co-cultured with BT△SPT / BF△SPT cells rose rapidly to a plateau at 3.227 h. In contrast, the level of azide-labeled SLs in the BTWT / BFWT co-culture group rose slowly within 4 h. Among them, NAC accelerated the transfer of PAA-derived SLs in the BTWT co-culture group, but not in the BT△SPT co-culture group. To further investigate the effect of bacterial SPT on NAC permeability, isotope-labeled 15 N-NAC was used to monitor newly synthesized SLs. At the end of the 2-h incubation, cystine-3-keto-sphingosine, cystine-sphingosine, cystine-sphingosine, and cystine-sphingosine-1-phosphate in the SL pool were significantly downregulated by SPT deficiency in co-culture with B. thetaiotaomicron and B. fragilis. High-resolution MS2 spectroscopy also showed that microbial SPT ablation reduced the concentration of azidocysteine-3-keto-sphingosine obtained from Caco-2 cells. Pearson correlation analysis revealed that the P-values ​​of azidocysteine-3-keto-sphingosine with NAC were significantly higher in the BTWT (p = 3.0 × e-3, r = -0.41) and BFWT (p = 3.7 × e-3, r = -0.40) co-culture groups. eff The values ​​are negatively linearly correlated.

[0093] In order to strictly regulate the synthesis of the microbiome, 6-week-old GF rats were colonized with BTWT / BFWT or BT△SPT / BF△SPT, and only PAA (palmitoyl acetate) was added before. BT△SPT Rats and GF BF△SPT Intestinal NAC in situ in rats eff The values ​​increased sharply, reaching 1.218×10 -4 cm / s and 1.407×10 -4 cm / sP eff Platform, and GF BTWT and GFBFWT The copper-catalyzed increase in the intensity of alkyne-labeled SLs labeled with Alexa Fluor-647 azide (AF647-Az) was then measured in the intestines of WT strain-colonized rats. BT△SPT and GF BF△SPT In rats, levels of alkyne-labeled SLs were almost undetectable within 2 h. To further identify these alkyne-containing sphingolipids, differential profiling was performed using tandem mass spectrometry. High-resolution MS2 spectra quantified GF BT△SPT and GF BF△SPT A dramatic decrease in the amount of alkyne-containing cysteine-3-ketosphingosine in the intestine of rats compared with wild-type controls. BTWT and GF BFWT In the rat cecum, the levels of two long-chain base SLs, cysteine-Sa (d18:0) and cysteine-So (d18:1), synthesized by B. thetaiotaomicron and B. fragilis, were significantly higher than those of GF BT△SPT and GF BF△SPT Rat. GF BTWT and GF BFWT Cysteine-ceramide (Cysteine-Cer) and cysteine-sphingomyelin (Cysteine-SM) with C16:0 and C24:0 acyl chains were elevated in rats. In contrast, cysteine-SLs with C18:0 and C24:1 acyl chains were similar between the two groups. Taken together, these findings suggest that the SL production capacity of intestinal microorganisms affects the level of NAC absorption in the intestine.

[0094] Although the above embodiment describes the present invention in detail, it is only a part of the embodiments of the present invention, not all of the embodiments. People can also obtain other embodiments based on this embodiment without creativity, and these embodiments all fall within the protection scope of the present invention.

Claims

1. A method for constructing a drug permeability clinical prediction model, characterized in that: The following steps are involved: Obtain intestinal microbial characteristics data; Inputting the intestinal microbial characteristic data into a prediction model to obtain a predicted Fsys characteristic value; The prediction model is determined by training based on a feature training set using a machine learning algorithm; the feature training set is the feature data of the trained intestinal microorganisms; the feature training set is obtained by processing the training data using a nearest neighbor clustering algorithm and a recursive feature elimination method; Determine the drug permeability according to the predicted Fsys characteristic value; the drug permeability is used to provide treatment strategies and theoretical basis for clinical medicine; Wherein, the method for determining the prediction model is: Constructing a network model set; the network model set includes: multivariate linear regression model, neural network model, k nearest neighbor model, random forest model, decision tree model, partial least squares model, support vector regression model, extreme gradient enhancement model and lightweight gradient boosting algorithm model; Input the feature training set into each network model in the network model set for training, and adopt a nested cross-validation method to select a model with the smallest absolute error from multiple network models as a preliminary optimization model; The random search method and the grid search method are used to adjust the hyperparameters in the preliminary optimization model with the negative mean absolute error as the measurement benchmark to obtain the prediction model.

2. The method for constructing a drug permeability clinical prediction model according to claim 1, characterized in that: The intestinal microbial characteristic data include: abundance of Bacteroides thetaiotaomicron, Cysteine-Cer (d18:1 / 16:0), Cysteine-3-ketosphinganine, intestinal pH, Cysteine-Cer (d18:1 / 24:0), Cysteine-Sphingosine (d18:0), Cysteine-Cer (d18:1 / 24:1), intestinal blood flow, Cysteine-DHCer (d18:1 / 16:0), Rhabditis fragilis Junfeng Spend , Cysteine-DHCer (d18:0 / 24:0), Cysteine-DHCer (d18:0 / 24:1), and viscosity of intestinal contents.

3. The method for constructing a drug permeability clinical prediction model according to claim 1 or 2, characterized in that: Such drugs include N-acetylcysteine.

4. The method for constructing a drug permeability clinical prediction model according to claim 1, characterized in that: The intestinal microorganisms include at least one of the following: Bacteroides thetaiotaomicron and Bacteroides fragilis.

5. Application of the drug permeability clinical prediction model constructed by the method for constructing a drug permeability clinical prediction model according to claims 1 to 4 in predicting drug permeability.

6. The use according to claim 5, characterized in that: The drug permeability clinical prediction model includes the LGBM model.

7. A method for screening oral drugs with individual variable permeability using a drug permeability clinical prediction model constructed by the method for constructing a drug permeability clinical prediction model according to claims 1 to 4, characterized in that: The following steps are involved: The animal model was used to construct control groups, wherein the first control group had no intestinal microbial colonization and the second control group had regular intestinal microbial colonization; The control group was subjected to microbiome and metabolome analysis before a single intestinal infusion of the drug, and the intestinal in situ P of the drug-perfused group was detected every 15 minutes after the infusion of the drug. eff value, analyzing the interaction between the drug and the intestinal microbial strain to obtain potential bacteria-drug pairs; the interaction is defined by the false discovery rate, wherein the false discovery rate corrected P value critical value is 0.05, and the exhaustion threshold is 30%; the detection includes using the drug permeability clinical prediction model for detection; Based on the in vivo communities obtained from the microbiome, bacteria-drug pairs with a Z factor > 0.8 were screened to identify oral drugs with individual variable permeability.

8. The method according to claim 7, characterized in that: The animal model includes Wistar rats.

Images