ИСТИНА |
Войти в систему Регистрация |
|
ИСТИНА ИНХС РАН |
||
The pharmacokinetic properties and toxicity of potential drug compounds (ADMET properties: absorption, distribution, metabolism, excretion, toxicity) critically affect their efficacy, pharmacological profile, administration protocol and safety. Their optimization is an important aspect of drug discovery and development process, and the ability to predict these properties for new structures can substantially improve its speed and efficiency. We have developed a general methodology for the prediction of ADMET parameters based on the application of artificial neural networks and fragmental descriptors to extensive and verified experimental data sets. The fragmental descriptors for a structure are the occurrence counts of the paths, cycles and branches of varied size using a hierarchical atom type classification, providing a ‘holographic’ representation of a molecule. During the model construction, the GPU-based deep learning and double cross-validation are used to achieve optimal performance and model predictivity. During the prediction, a graphic map highlighting the parts of a molecule that make positive or negative contributions to the predicted property is generated as an additional guidance for the ADMET optimization. The models built by us are implemented in an integrated online service available on the Internet (http://qsar.chem.msu.ru/admet/). It supports convenient prediction of important properties (in particular, lipophilicity, blood-brain barrier permeability, human intestinal absorption, hERG-mediated cardiac toxicity, etc.) as well as qualitative and semi-quantitative estimation of their suitability for drug-like compounds. This integrated prediction system may be used in the research in various areas of medicinal chemistry and pharmacology.