Machine Learning Boosted Docking (HASTEN): An Open-source Tool To Accelerate Structure-based Virtual Screening...

Area of Science:

• Computational chemistry
• cheminformatics
• machine learning

Background:

• Structure-based virtual screening is crucial for drug discovery.
• Accelerating virtual screening can significantly reduce the time and cost of identifying potential drug candidates.
• Machine learning models offer a promising approach to enhance screening efficiency.

Purpose of the Study:

• To develop and validate a machine learning-boosted docking software named HASTEN.
• To accelerate structure-based virtual screening processes.
• To assess the performance of HASTEN using diverse datasets.

Main Methods:

• HASTEN software was developed integrating machine learning models with docking protocols.
• Validation was performed using 12 literature datasets (3 million molecules docked with FRED) and one in-house dataset (4 million compounds docked with Glide).
• Performance was evaluated based on recall values for top-scoring molecules after partial dataset docking.

Main Results:

• HASTEN demonstrated reasonable performance on literature data with a mean recall of 0.78 for the top 1% of molecules after docking 10% of the dataset.
• Excellent recall of 0.95 was achieved for the in-house dataset.
• The software is compatible with various docking and machine learning methodologies.

Conclusions:

• HASTEN effectively accelerates structure-based virtual screening using machine learning.
• The software shows high performance in identifying active compounds, particularly with in-house data.
• HASTEN is a valuable, freely available tool for computational drug discovery.