OralAbsPredict: A data-driven framework to predict human intestinal absorption (HIA) and human oral bioavailability (HOB) from chemical structures.

Oral absorption is a key pharmacokinetic process for orally administered drugs, determining therapeutic efficacy and influencing distribution, metabolism, and excretion. One of the major reasons a drug fails in early development is poor oral absorption. Therefore, it is important to determine the oral absorption properties of drugs during early development, even before synthesis. Human intestinal absorption (HIA) and human oral bioavailability (HOB) are two important pharmacokinetic properties of orally administered drugs that provide a good measure of drug absorption through the oral route. The experimental determination of the HIA and HOB required substantial time, resources, and money. Therefore, an alternative approach is needed to quickly screen the HIA and HOB of orally administered compounds. In this study, we introduce a Python-based software tool, "OralAbsPredict," that predicts HIA and HOB from chemical structures provided as SMILES strings. This tool returns three predicted values, HIA and HOB at two different experimental cutoffs, i.e., 50% and 20%, based on the developed machine learning models. The tool is based on the developed machine learning models showing good performance not only on the training set (accuracytr > 0.9) but also on the test set (accuracyte > 0.7). These models are developed using Functional-Class Fingerprints (FCFP) and Mordred descriptors: a count-based FCFP2 fingerprint for the HIA model, a Mordred descriptor for HOB_50% model, and a count-based FCFP4 fingerprint for the HOB_20% model. One of the major drawbacks of existing models is that they do not adequately address class imbalance in the modeled data. However, we have addressed the class imbalance problem in this work using hyperparameters like class_weight. Here, we have identified and interpreted the important features using the SHAP method. Additionally, we have performed a substructure analysis to identify key substructures present in most of the active molecules. From this analysis, we have found that most of the orally active compounds contain the aromatic ring system with chlorine, fluorine, and amine groups. A comparative analysis with the benchmark expert systems was also conducted, demonstrating that our models achieved similar test performance.