Combined data-driven and mechanism-based approaches for human-intestinal-absorption prediction in the early drug-discovery stage†

Abstract

It is important to precisely predict the intestinal absorption ratio (Fa) at an early stage in the discovery of orally available drugs because it directly influences drug efficacy. Gastrointestinal unified theoretical framework (GUTFW) and machine learning (ML) are commonly used to predict the percentage of Fa. In GUTFW, the Fa of a drug is estimated using an equation based on the mechanism of human intestinal absorption, dose, solubility, membrane permeability, and dissolution of the drug. The experimental values of these in vitro parameters are required to accurately predict Fa. However, most of these values are unavailable at early stages of development. ML uses a limited dataset of the observed Fa values of drugs in humans. In this study we combined GUTFW and ML to compensate for each defect. We collected published data on the chemical structures of 460 drugs, including Fa and dose amounts. The key parameters of the GUTFW (Do, dose number; Dn, dissolution number; Pn, permeation number), solubility, membrane permeability, and structural descriptors were calculated and used as explanatory variables for ML. ML algorithms, namely, the random forest (RF) and message-passing neural network (MPNN; Chemprop), were investigated. The GUTFW model was compared to the conventional ML method, which uses only structural descriptors, and combined ML method, which uses both structural descriptors and GUTFW parameters. In addition, using the Chemprop framework, we investigated important substructures of Fa. Our result suggested that combinational ML produced higher predictivity than the GUTFW model and conventional ML model in the test dataset (20% of the dataset) [R² value and RMSE in the combinational ML method: 0.611 and 19.7 (RF), 0.520 and 21.6 (Chemprop); in conventional ML: 0.339 and 25.4 (RF), 0.497 and 22.1 (Chemprop); in GUTFW: 0.353 and 31.9]. Additionally, most of the substructures indicated by the Chemprop framework were consistent with the common knowledge of medicinal chemistry. We developed an accurate prediction method for human Fa using a combination of data-driven ML and mechanism-based GUTFW, where the parameters could be calculated without experimental data, enabling the model to efficiently promote early drug discovery.