Objectives: Accurate prediction of a new compound’s pharmacokinetic (PK) profile is pivotal for the success of drug discovery programs. An initial assessment of PK in pre-clinical species and humans is typically performed through allometric scaling and mathematical modeling. These methods use parameters estimated from in-vitro or in-vivo experiments, which although helpful for an initial estimation, require extensive animal experiments. Furthermore, mathematical models are limited by the mechanistic underpinning of the drugs’ absorption, distribution, metabolism, and elimination (ADME) which are largely unknown in the early stages of drug discovery. In order to overcome these issues and to reduce the need for animal experiments, we propose a novel machine learning framework that utilizes structure driven molecular properties of proposed compounds to predict concentration vs time profile of drug exposure in rats[1].
Methods: Data A dataset consisting of 391 compounds was used in this study. For each compound present in the study, the molecular structure was represented as a SMILES (Simplified molecular-input line-entry system) string. In terms of PK/ADME parameters, the in-vivo clearance and in-vivo Volume of Distribution generated using a non-compartmental analysis method, PKa and Lipophilicity were also available in this dataset, and finally, the concentration vs time data for 1 mg/kg intravenous (IV) administration in rats was also available in the dataset. Model Framework The molecular structure of the small molecule compound was converted into either a descriptor-based representation and/or a molecular fingerprint-based representation. This was followed by feature selection to identify the combination of descriptors, PK/ADME property or fingerprints that would help in predicting the PK profiles. For parameters selected from feature importance (Clearance (CL) and Volume of distribution (Vdss)), ML models were developed to predict them based on the molecular structure[2]. Finally, predicted CL and Vdss were used as an input to an additional machine learning model which predicts the PK profile of the compound[3].
Results: Our results demonstrate that the proposed algorithm can be used to adequately predict PK profiles in rats with an average mean absolute percentage error between predicted and observed PK profile calculated to be around 3 fold. The predictions based on proposed machine learning framework were comparable to those predictions that were generated using a PBPK framework.
Conclusion: In conclusion, the ML framework presented in this article can predict PK profiles with reasonable accuracy in rats. These efforts aim to enable PK profile predictions earlier in the drug discovery process thus helping scientists gain insights into exposure of the compounds in-vivo and helping them with prioritization and screening of compounds.
Citations: 1. Stewart, A., et al., The FDA modernisation act 2.0: Bringing non-animal technologies to the regulatory table. Drug Discovery Today, 2023. 28(4): p. 103496. 2. Panteleimon D. Mavroudis, D.T., Alexandra Abos, Nikhil Pillai, Application of Machine Learning in combination with Mechanistic Modeling to Predict Plasma Exposure of Small Molecules. Frontiers in Systems Biology, 2023. 3. Nikhil Pillai, A.A., Donato Teutonico, Panteleimon D. Mavroudis Machine learning framework to predict pharmacokinetic profile of small molecule drugs based on chemical structure. Clinical and Translational Science 2024.
