HASTEN

HASTEN accelerates structure-based virtual screening by integrating machine learning with docking scores to prioritize compounds for drug discovery.

Key Features:

• Machine Learning Integration: Initially supports Chemprop for model training and allows integration of user-supplied machine learning methods via custom shell scripts.
• Docking Program Compatibility: Supports Glide (Schrodinger) for docking, can integrate alternative docking programs, and operates under the assumption that lower docking scores indicate better binding affinity.
• Simulation Mode: Provides a simulation mode that benchmarks workflows using pre-existing docking scores without performing 3D docking simulations.
• Validation Datasets: Validated on 12 literature-derived datasets of three million molecules each docked with FRED and an in-house dataset of four million compounds docked with Glide.

Scientific Applications:

• Improving virtual screening recall: Achieved a mean recall of 0.78 for the top 1% scoring molecules on literature datasets after docking 10% of the dataset and a recall of 0.95 on in-house Glide data.
• Compound prioritization for drug discovery: Prioritizes promising compounds from large molecular libraries to streamline structure-based virtual screening campaigns.

Methodology:

Uses machine learning models (Chemprop by default, with support for custom methods via shell scripts) to predict binding affinity from docking scores and prioritize compounds, employs a simulation mode using pre-existing docking scores without 3D simulations, and was validated on datasets docked with FRED and Glide under the assumption that lower docking scores indicate better binding affinity.