Machine learning approaches for predicting compounds that interact with therapeutic and ADMET related proteins

Area of Science:

• Computational chemistry
• Pharmacology
• Bioinformatics

Background:

• Drug discovery relies on predicting pharmacodynamic and ADMET properties.
• Machine learning (ML) methods are increasingly used for predicting compound-protein interactions.
• ML complements Quantitative Structure-Activity Relationship (QSAR) and structure-based methods, especially for diverse chemical structures or absent 3D receptor data.

Purpose of the Study:

• To review ML strategies for predicting protein binders.
• To highlight current progress and challenges in applying ML for drug discovery.
• To evaluate algorithms for representing compound properties in ML models.

Main Methods:

• Exploration of neural networks and support vector machines for prediction tasks.
• Application of ML to predict inhibitors, antagonists, blockers, agonists, activators, and substrates.
• Evaluation of algorithms for compound structural and physicochemical property representation.

Main Results:

• Demonstrated potential of ML in predicting substrates (e.g., P-glycoprotein, CYP isoenzymes) and inhibitors (e.g., protein kinases, CYP isoenzymes).
• Showcased ML's utility in predicting agonists (e.g., serotonin, estrogen receptors).
• ML methods show promise as virtual screening tools.

Conclusions:

• ML offers powerful computational tools for drug discovery and evaluation.
• Effective representation of compound properties is crucial for successful ML model development.
• Continued advancements in ML are expected to further enhance virtual screening and prediction of drug properties.