Predicting Human Pharmacokinetic Parameters from Molecular Structures: Applications of Machine Learning in Drug Development

This project is a hybrid machine learning pipeline combining Random Forest, XGBoost, and Graph Neural Networks. It can directly predict key human pharmacokinetic parameters such as clearance (CL), volume of distribution (Vd), half-life (t½), and terminal elimination rate constant (λz) from SMILES chemical structure strings, and provides 95% confidence interval calibration. The project is sourced from GitHub's FHB_Human_PK_From_Structure_RShiny, supporting FastAPI interfaces and R Shiny interactive applications.

In new drug development, pharmacokinetic (PK) research is a key step for candidate drugs to enter clinical trials. Traditional methods require a large number of in vitro and in vivo experiments, which are time-consuming, labor-intensive, and costly. Human PK parameters (CL, Vd, t½, λz) directly affect dosing regimens, risk assessment, and safety windows. In recent years, AI technology can predict PK properties before synthesis by learning chemical structure-activity relationships, shortening the development cycle and reducing the risk of failure.

The project adopts a multi-model integration strategy:

1. Random Forest: Processes high-dimensional molecular descriptors and models nonlinear relationships;
2. XGBoost: Captures complex mapping relationships and prevents overfitting;
3. Graph Neural Network (AttentiveFP architecture): Treats molecules as graph structures and focuses on key substructures via attention mechanisms. Each parameter is trained and optimized independently, and the best models with GMFE <1.5 and R²>0.7 are selected to assemble a hybrid predictor.

The project introduces the split-conformal prediction method to provide 95% confidence intervals. This non-parametric technique uses the error distribution of the calibration set to determine the interval, ensuring that the probability of the true value falling within the interval reaches at least the set level. This allows researchers to obtain point estimates while understanding uncertainty, aiding decision-making (e.g., wide intervals require additional experimental verification).

Data Sources: Integrates public data from Lombardo Database, ChEMBL, Enamine, etc.; Feature Extraction: RDKit calculates 2D/3D molecular descriptors (molecular weight, LogP, etc.), and PyTorch Geometric converts SMILES into graph features; Deployment: Provides FastAPI RESTful API for programmatic access, and an R Shiny interface for non-technical users to input SMILES and obtain visualized results.

This project combines cutting-edge deep learning with traditional machine learning, provides reliable uncertainty estimates through rigorous statistical methods, and lowers the threshold for applying AI technology. For R&D institutions, it can quickly screen candidate compounds in the early stage, prioritize molecules with ideal PK properties, improve success rates, and reduce costs. In the future, with data accumulation and the evolution of GNN architectures, the model's accuracy is expected to further improve and become a standard tool in drug development.