Advancing the development of deep learning and machine learning models for oral drugs through diverse descriptor classes: a focus on pharmacokinetic parameters (Vdss and PPB).

Abstract

In the present study, we report a predictive deep learning (DL) and machine learning (ML) model for pharmacokinetics (PK) parameters such as volume of distribution (Vdss) and plasma protein Binding (PPB). Using DL & ML algorithms our study provides a deeper and novel insights into the role of molecular descriptors in determining the PK parameters such as Vdss and PPB. FDA approved drugs with oral route of administration and having reported PK parameters were taken as the dataset. This was used for establishment of the foundational datasets followed by computation of different molecular descriptor classes. Feature engineering by Boruta algorithm exhibited significant increase in accuracy of the models. Features identified by Boruta algorithm, were trained for different models separately for both Vdss and PPB. The highest predictive scores amongst the models were achieved in gradient boosting (GB) and Stacking Classifier with 80% and 78% for Vdss. In the case of PPB, random forest and GB algorithm predicted the highest scores of 73% and 71%, respectively, in comparison to all other algorithms. In summary we report here appropriate ML algorithms like Stacking Classifier-by utilizing an unreported feature engineering algorithm -to predict Vdss and PPB individually considering over 67 descriptors each with ≥ 80% accuracy and 73% accuracy, respectively. Additionally, we developed models based on the shared descriptors between Vdss and PPB. Quantum chemical descriptors like MLFERs (MLFER_BH, MLFER_BO & MLFER_E) and topological descriptors like piPC5, piPC6, piPC9 & TpiPC identified as the common drivers of the functional activity of Vdss and PPB together.