Artificial intelligence-driven prediction and validation of blood-brain barrier permeability and absorption, distribution, metabolism, excretion profiles in natural product research laboratory compounds.

Introduction: Our 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.

Objectives: The study aims to investigate the pharmacokinetic ADME properties and BBB permeability coefficients of NPRL compounds.

Materials And Methods: A 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.

Results: There 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 cm/s. Computational analysis revealed that NPRL compounds exhibited a variety of ADME characteristics.

Conclusion: Computer-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.