Artificial intelligence-driven prediction and validation of blood–brain barrier permeability and absorption, distribution, metabolism, excretion profiles in natural product research laboratory compounds Authors: Jai-Sing Yang, Eddie TC. Huang, Ken YK. Liao, Da-Tian Bau, Shih-Chang Tsai, Chao-Jung Chen, Kuan-Wen Chen, Ting-Yuan Liu, Yu-Jen Chiu, Fuu-Jen Tsai Published: 2024 ABSTRACT IntroductionOur previous research demonstrated that a large language model (LLM) based on the transformer architecture, specifically the MegaMolBART encoder with an XGBoost classifier, effectively predicts the blood–brain barrier (BBB) permeability of compounds. However, the permeability coefficients of compounds that can traverse this barrier remain unclear. Additionally, the absorption, distribution, metabolism, and excretion (ADME) characteristics of substances obtained from the Natural Product Research Laboratory (NPRL) at China Medical University Hospital (CMUH) have not yet been determined.ObjectivesThe study aims to investigate the pharmacokinetic ADME properties and BBB permeability coefficients of NPRL compounds.Materials and methodsA combined model using a transformer-based MegaMolBART encoder and XGBoost classifier was employed to predict BBB permeability. Machine learning (ML) tools from Discovery Studio were used to assess the ADME characteristics of the NPRL compounds. The CCK-8 assay was conducted to evaluate the cytotoxic effects of NPRL compounds on bEnd.3 brain endothelial cells after exposure to 10 μg/mL of the compounds. We assessed the permeability coefficient by subjecting bEnd.3 cell monolayers to the test compounds and measuring the permeability of FITC-dextran.ResultsThere were 4956 compounds that could cross the blood–brain barrier (BBB+) and 2851 that could not (BBB−) in the B3DB dataset that was utilized for training. A total of 2461 BBB+ and 2184 BBB− compounds were used in the NPRL-CMUH dataset for testing. The permeability coefficient of temozolomide (TMZ) and 21 other BBB + compounds exceeded 10 × 10−7 cm/s. Computational analysis revealed that NPRL compounds exhibited a variety of ADME characteristics.ConclusionComputer-based predictions for the NPRL of CMUH compounds regarding their capacity to traverse the BBB are verified by the findings. Artificial intelligence (AI) prediction models have effectively identified the potential ADME characteristics of various compounds. 1. Introduction Human health and well-being face substantial challenges due to neurological conditions, including Alzheimer’s disease (AD), Parkinson’s disease (PD), stroke (both ischemic and hemorrhagic), epilepsy, headaches, and brain tumors [1,2]. Identifying compounds capable of crossing the blood–brain barrier (BBB) and effectively targeting the central nervous system (CNS) remains a significant challenge in developing treatments for neurological disorders. This highly selective boundary restricts the passage of most substances into the brain [3,4]. To be considered viable therapeutic agent candidates for CNS diseases, compounds must demonstrate therapeutic efficacy and possess the ability to cross the BBB to reach their intended targets [4,5]. Candidates that can cross the BBB usually have certain traits: they are very attracted to fats, have a low molecular mass (usually less than 400–600 Da), do not bind well to plasma proteins very well, and do not break apart easily at physiological pH condition [6,7]. These strict requirements for BBB permeability make it challenging for larger molecules such as peptides, recombinant proteins, monoclonal antibodies, and gene therapies to traverse the barrier independently [8,9]. Furthermore, even when compounds capable of crossing the BBB are identified, their progression is often hindered by their suboptimal pharmacokinetic (PK) properties. These PK properties, which include absorption, distribution, metabolism, excretion, and toxicity (ADMET) profiles, contribute to high failure rates during new drug development [10,11]. In recent years, advancements in computational prediction models and artificial intelligence (AI) have significantly improved the efficiency of identifying compounds with desirable BBB permeability and ADMET characteristics [12–15]. Various AI research groups have developed predictive models for BBB penetration utilizing these tools to expedite CNS drug discovery [1,16–20]. For instance, Zhang et al. employed MolconnZ, MOE, and Dragon models to develop a quantitative structure–activity relationship (QSAR) model for BBB permeability, utilizing data from 159 compounds capable of passing through the BBB [21]. Similarly, Wang et al. applied six machine learning (ML) algorithms to categorize 2358 compounds using the scikit-learn framework (version 0.19.1) [22]. Miao et al. implemented POLY-Support Vector Machine (POLY-SVM), Sigmoid-Support Vector Machine (Sigmoid-SVM), K-Nearest Neighbor (KNN), Radial Basis Function-Support Vector Machine (RBF-SVM), Decision Tree (DT), and other deep learning techniques to predict BBB permeability [23]. Liang et al. investigated classification models to elucidate the structural features and transformation rules associated with compounds that can traverse the BBB [24]. Kumar et al. introduced “DeePred-BBB,” a deep learning-based model capable of predicting BBB permeability from simplified molecular input line entry system (SMILES) representations (Table 1) [25]. Our previous research demonstrated the ability of the large language model (LLM) in the transformer-based MegaMolBART encoder with XGBoost classifier to predict BBB permeability across an extensive compound library, although permeability coefficients remain undetermined. Furthermore, the pharmacokinetic ADME characteristics of compounds from the Natural Product Research Laboratory (NPRL) at China Medical University Hospital (CMUH) have not been fully explored. This study’s objective was to use the FITC-dextran extravasation model to develop an in vitro model to assess BBB permeability and predict the pharmacokinetic (PK) behavior of NPRL-CMUH compounds. Additionally, this study integrates machine learning (ML) techniques (Fig. 1). 2. Methods 2.1. Cell culture The bEnd.3 cells (immortalized mouse brain endothelial cells) from ATCC (Catalog no. CRL-2299; Manassas, VA, USA) were cultured in Dulbecco’s Modified Eagle’s Medium (DMEM) (American Type Culture Collection, Manassas, VA, USA) supplemented with 10% fetal bovine serum (R&D Systems, Minneapolis, MN, USA) and 100 μg/mL penicillin/streptomycin (Sigma–Aldrich, St. Louis, MO, USA) at 5% CO2 and 95% air [26,27]. 2.2. In vitro permeability assay The bEnd.3 cells (5.0 × 104 cells/cm2) were cultured onto 24-well transwell inserts (pore size 0.4 μm; Corning Inc., Corning, NY, USA). After 72 h, the cells formed a confluent monolayer and were then treated with NPRL compounds (10 μg/mL) or dimethyl sulfoxide (DMSO) (1 μL) as control for 12 h. Following the treatment, the inserts and chambers were PBS washed, and the medium was replaced with phenol-red free medium. Subsequently, 10 kDa FITC-dextran (10 mg/mL; 10 μL; Sigma–Aldrich) was added to the upper inserts and incubated at 37 °C for 2 h. Fluorescence was measured in the upper and lower chambers using a fluorescence plate reader (SpectraMAX M3; Molecular Devices, Sunnyvale, CA, USA) at an excitation/emission wavelength of 490 nm/520 nm [27]. The permeability coefficient was calculated using the following formula: RFUlower: relative fluorescent units in the lower well. RFUlower: relative fluorescent units in the lower well. RFUupper: relative fluorescent units in the upper well. RFUupper: relative fluorescent units in the upper well. V: volume of the bottom well. V: volume of the bottom well. T: time that the FITC-dextran was allowed to diffuse. T: time that the FITC-dextran was allowed to diffuse. A: total surface area of the monolayer (cm2). A: total surface area of the monolayer (cm2). 2.3. Cytotoxicity assay The bEnd 3 cells (5 × 103 cells/well) were seeded in 96-well plates, and then incubated for 24 h in a humidified incubator at 37 °C with 5% CO2. NPRL compounds (10 μg/mL) were then added to the culture media in the plates. The plates were incubated for an additional 20 h. Following this, 10 μL of CCK-8 solution was added to each well. The plates were incubated for 1–4 h, after which the absorbance was measured at 450 nm using an ELISA reader (detection: 450 nm, reference: 650 nm) [28]. 2.4. In silico BBB permeability prediction We employed the transformer-based Mega-MolBART encoder (NVIDIA, Version 1.0; URL: https://catalog.ngc.nvidia.com/orgs/nvidia/teams/clara/models/megamolbart) with an XGBoost classifier (URL: github.com/dmlc/xgboost) to predict BBB permeability [1]. This model was trained on the B3DB dataset and tested on the NPRL compound library. The MegaMolBART encoder processes the molecular structure information, while the XGBoost classifier uses this encoded information to make BBB permeability predictions. This approach allows efficient and accurate assessment of a molecule’s ability to cross the BBB based on its structural features. 2.5. In silico pharmacokinetic ADME properties prediction In silico pharmacokinetic properties prediction studies utilized the ADME algorithm of the Discovery Studio 2023 machine learning model (ML) software (BIOVIA), accessible at (URL: https://www.3ds.com/products/biovia/discovery-studio/qsar-admet-predictive-toxicology) (Fig. 1). This software calculates various ADME properties, including aqueous solubility (predicting the solubility of each compound in water at 25 °C), BBB penetration, CYP2D6 binding (predicting cytochrome P450 2D6 enzyme inhibition), human intestinal absorption (HIA) after oral administration, plasma protein binding (PPB) (predicting the likelihood that a compound will be highly bound to carrier proteins in the blood), as well as AlogP98 and PSA_2D [29,30]. The training set for the aqueous solubility prediction model comprised 775 compounds with molecular weights ranging from 50 to 800. These compounds included alkanes, alkenes, alkynes, halogens, amines, alcohols, nitrogen-containing compounds, ketones, aldehydes, and organic acids. The plot of predicted versus experimental data yielded a linear regression of LogSw (25 °C, pH = 7.0) with R2 = 0.84 and a standard deviation (SD) of 0.87. The pre-test dataset included 34 compounds, resulting in regression statistics of R2 = 0.88 and SD = 0.79. Additionally, a validation test dataset of 1615 compounds from the PDR and Comprehensive Medicinal Chemistry database (CMC) produced an overall RMSE (SD) of 1.0 [31]. The BBB penetration model makes educated guesses about how much of a drug will enter the brain after oral administration by using data from over 800 compounds known to penetrate the BBB. This model employs a quantitative linear regression for predicting blood–brain penetration, accompanied by 95% and 99% confidence ellipses in the ADMET_PSA_2D and ADMET_AlogP98 planes [32]. The cytochrome P450_2D6 (CYP2D6) model predicts CYP2D6 enzyme inhibition based on 2D chemical structures. Using modified Bayesian learning [33], training set of 151 structurally different compounds with known CYP2D6 inhibition constants were used to make it. The AI/ML model was created using 182 compounds in the training dataset for human intestinal absorption. This model incorporates AlogP98 and 2D polar surface area (PSA_2D). Compounds with well-documented human intestinal absorption exhibit at least 90% absorption into the human bloodstream and typically fall within the 95% and 99% confidence ellipse regions [34]. The plasma protein binding model predicts whether a compound is likely to be highly bound (≥90% bound) to carrier proteins in the blood. The training dataset comprised 854 compounds divided into 329 binders and 525 non-binders. A modified Bayesian learning approach created a binary classification model [35,36]. 2.6. Statistical analysis The results are shown as the mean ± standard deviation (SD). Statistical analysis is performed using a one-way analysis of variance or Tukey’s test. Levels of significance are represented as ***P < 0.001, indicating statistical discrepancies between the control group and the NPRL-treated groups [37–39]. 3. Results 3.1. In silico BBB permeability prediction In our prior research, we utilized a LLM-based on the transformer-based MegaMolBART encoder with XGBoost classifier architecture to forecast the BBB permeability of compounds sourced from the NPRL. We validated these predictions by conducting in vitro experiments with human BBB spheroid cells and evaluating BBB integrity using liquid chromatography-tandem mass spectrometry (LC-MS/MS). Our findings highlighted the significant potential of this approach for advancing BBB permeability research in new drug discovery. The LLM of the transformer-based MegaMolBART encoder with XGBoost classifier exhibited strong predictive performance, as demonstrated by the composition of the training and testing datasets. The training dataset sourced from the B3DB database, comprised 4956 BBB-permeable (BBB+) and 2851 BBB-impermeable (BBB−) compounds (Fig. 2A). The NPRL testing dataset included 2461 BBB+ and 2184 BBB− compounds (Fig. 2B). 3.2. In vitro BBB permeability and cytotoxicity analysis The effect of NPRL compounds on endothelial cell permeability was evaluated using an in vitro blood–brain barrier (BBB) model consisting of confluent bEnd.3 monolayers grown in transwell chambers. As shown in Fig. 3, exposure to 10 μg/mL of various NPRL substances (NPRL309, NPRL358, NPRL588, NPRL818, NPRL833, NPRL835, NPRL836, NPRL842, NPRL1089, NPRL1185, NPRL1188, NPRL1192, NPRL1195, NPRL1241, NPRL1958, NPRL2026, NPRL2029, NPRL2051, NPRL2059, NPRL2148, and NPRL3767) led to a notable increase in monolayer permeability, as evidenced by the greater diffusion of 10 kDa FITC-dextran across the transwell membrane. Conversely, five BBB-impermeable compounds (NPRL2359, NPRL2576, NPRL2646, NPRL3098, and NPRL3183) reduced the monolayer permeability (Fig. 3). To verify that the enhanced BBB permeability was not caused by the toxic effects on bEnd.3 cells, we assessed cell viability using CCK-8 assays. The results, depicted in Fig. 4, demonstrated that exposing bEnd.3 cells (10 μg/mL) of NPRL compounds had a negligible cytotoxic impact. This indicated that the observed permeability changes were not a result of compound-induced cell damage. These results corroborate our in silico predictions of the BBB permeability characteristics of the NPRL compound library. 3.3. In silico prediction of pharmacokinetic properties (ADME) To estimate the drug-like potential of the NPRL compounds, an in silico ADME study was conducted using Discovery Studio 2023 software within a ML model. This study predicted various pharmacokinetic properties, including intestinal absorption, compound solubility, metabolism by CYP2D6, plasma protein binding (PPB) capability, and blood–brain barrier (BBB) permeability, alongside the compound’s AlogP98a and PSA properties. Computer analysis of the ADME parameters showed that the NPRL compounds had different pharmacokinetic profiles, as shown in Fig. 5 and Table 2. Additional raw data are presented in Supplementary Table 1 (https://www.biomedicinej.com/cgi/editor.cgi?article=1474&window=additional_files&context=biomedicine). These results highlight the differences in ADME properties among the compounds, providing essential information for the future refinement of new drug development processes (see Table 2). 4. Discussion In vitro and in vivo experiments and clinical trials typically yield the most accurate BBB permeability data. However, obtaining experimental BBB permeability data for numerous compounds, particularly across extensive libraries, is both time consuming and expensive [20]. To address these challenges, AI-based computational methods have been developed to assess BBB permeability, providing a crucial and strategic alternative [1,40]. These AI approaches offer significant benefits, including improved efficiency, reduced costs, and the ability to quickly process large volumes of data [41,42]. In contemporary research, most scientists have utilized the chemical characteristics of substances, such as molecular fingerprints and two-dimensional molecular descriptors, to develop predictive models. Various datasets and algorithms assess the likelihood of compounds penetrating the BBB [43–48]. This method speeds up the process of finding possible drugs for disorders of the CNS by allowing many compounds to be screened simultaneously and reducing the need for extensive experimental testing [43–48]. We used a LLM built on the transformer-based MegaMolBART encoder with XGBoost classifier architecture to try to guess how compounds from the NPRL would pass through the BBB (Fig. 1 up, and Fig. 2). We also performed in vitro experiments (Fig. 1 down and Fig. 3) to see how NPRL compounds affected bEnd.3 brain endothelial cells and found BBB permeability coefficients (Fig. 3). The findings revealed that cytotoxicity did not affect the observed BBB permeability of the NPRL compounds, as the bEnd.3 cells exhibited no notable toxic effects at the concentrations tested (Fig. 4). Our research demonstrated the effectiveness of integrating computational predictive models with laboratory-based validation to enhance BBB permeability research in drug development. It is essential to use both these methods to find promising drug candidates for further study, as this study shows that the LLM model can accurately predict BBB permeability without being affected by cytotoxic effects [1]. Researchers continue to face challenges in verifying the mechanisms of action (MOA) for substances that cross the BBB and in validating the accuracy of predictive models without extensive in vivo testing [23,49,50]. Several critical factors influence the ability of a compound to penetrate the BBB, including lipophilicity, molecular weight, shape, and charge [51–55]. Additionally, a substance’s permeability may be affected by its interactions with various cellular components, such as active or passive membrane transporters, metabolic enzymes, and efflux pumps [52,54,56,57]. Pathological conditions or therapeutic agents can also alter the effectiveness and distribution of compounds by modifying BBB permeability [58,59]. When utilizing in vitro models to assess BBB permeability, as in this study, it is essential to consider the physicochemical properties of the compounds, the presence of an unstirred water layer, and the characteristics of the filter [60,61]. These factors can significantly affect the permeability coefficients obtained in cell-based experiments. The ADME parameters of NPRL compounds were analyzed to determine their pharmacokinetic characteristics (Fig. 5, Table 2 and Supplementary Table 1 (https://www.biomedicinej.com/cgi/editor.cgi?article=1474&window=additional_files&context=biomedicine)). These findings provide valuable insights for the development of pharmaceutical formulations, including those for oral and intravenous delivery, while also advancing the study of drug-metabolomics and pharmacokinetics (PK) and pharmacodynamics (PD). Our in vitro experimental results support computer-generated predictions regarding the BBB permeability of NPRL compounds, indicating that computational models can serve as reliable initial tools in new drug discovery. However, further in vivo research is needed to validate these predictions fully and to enhance our understanding of the mechanisms of action for potential clinical applications. 5. Conclusion We conducted research that supports the numerous advantages of employing deep learning techniques to predict the permeability of the BBB. However, it is important to recognize that the current methods are still insufficient to completely elucidate the complex processes involved in drug-transport across the BBB. At present, we are unable to accurately differentiate between the therapeutic effects, adverse effects, and secondary outcomes resulting from a compound’s ability to penetrate the BBB, which has significant implications for pharmacological studies. Future research will focus on experimental investigations using in vivo animal models and cellular assays to address these knowledge deficits. The objective of these endeavors is to clarify the MOA and explore the potential therapeutic applications of the identified compounds. This type of research is crucial for improving the precision of predictions made by deep learning models and broadening our understanding of pharmacokinetics at the BBB. Supplementary Information