Employing Automated Machine Learning (AutoML) Methods to Facilitate the In Silico ADMET Properties Prediction.

Area of Science:

• Computational Chemistry
• Pharmacology
• Drug Discovery

Background:

• Early-stage drug discovery relies on predicting Absorption, Distribution, Metabolism, Excretion, and Toxicity (ADMET) properties to minimize compound attrition.
• Artificial intelligence (AI) offers high-throughput and cost-effective solutions for *in silico* ADMET modeling.

Purpose of the Study:

• To develop and validate machine learning models for predicting 11 ADMET properties using automated machine learning (AutoML).
• To assess the performance of these models against existing predictive tools using external datasets.

Main Methods:

• Utilized the Hyperopt-sklearn AutoML method to combine 40 classification algorithms with optimized hyperparameters.
• Developed predictive models for 11 distinct ADMET properties, ensuring an Area Under the ROC Curve (AUC) greater than 0.8 for all models.

Main Results:

• All developed models achieved an AUC > 0.8, indicating high predictive accuracy.
• The models demonstrated superior performance for most ADMET properties compared to published predictive models.
• Performance was comparable to existing models for other ADMET properties when validated on external datasets.

Conclusions:

• Automated machine learning (AutoML) is a viable and effective approach for ADMET prediction in early drug discovery.
• The developed models can aid researchers in designing compounds with improved ADMET profiles, potentially reducing late-stage drug failures.