Application of Deep Neural Network Models in Drug Discovery Programs.

Area of Science:

• Computational chemistry and cheminformatics
• Artificial intelligence in drug discovery
• Pharmacology and toxicology

Background:

• In silico methods are crucial for optimizing drug properties like pharmacokinetics, pharmacodynamics, and safety in modern drug discovery.
• Large, harmonized datasets enable the use of deep neural networks (DNNs) for predictive modeling.
• Different DNN architectures vary in their applicability and performance for lead optimization, particularly regarding temporal stability and interpretability.

Purpose of the Study:

• To compare the effectiveness of established DNN-based methods for predicting key ADME (Absorption, Distribution, Metabolism, Excretion) properties and biological activity.
• To evaluate the performance of Multilayer Perceptron (MLP), Graph Convolutional Network (GCN), and Mol2Vec models in an industrial drug discovery setting.
• To assess the stability and interpretability of DNN predictions over time for guiding medicinal chemistry efforts.

Main Methods:

• Implementation and comparison of three DNN architectures: MLP, GCN, and Mol2Vec.
• Prediction of key ADME properties including microsomal lability, CYP3A4 inhibition, and factor Xa inhibition.
• Validation using external datasets and a time series validation study to assess prediction stability.

Main Results:

• MLP and GCN models demonstrated statistically superior performance compared to Mol2Vec on external validation sets.
• GCN-based predictions exhibited the highest stability over extended periods in time series validation.
• DNNs were found to be valuable for guiding local Structure-Activity Relationship (SAR) studies in medicinal chemistry.

Conclusions:

• Established DNN architectures, particularly GCN and MLP, are effective tools for predicting critical compound properties in drug discovery.
• GCN models offer enhanced temporal stability, crucial for iterative lead optimization projects.
• DNNs provide significant value in directing medicinal chemistry efforts, contributing to a more realistic application of artificial intelligence in pharmaceutical research.