Regression-Based Active Learning for Accessible Acceleration of Ultra-Large Library Docking

Area of Science:

• Computational Chemistry
• Drug Discovery
• Bioinformatics

Background:

• Structure-based drug discovery utilizes 3D biomolecular models for hit identification and optimization.
• Ultralarge virtual screening enables rapid discovery of high-affinity compounds but demands extensive computational resources.

Purpose of the Study:

• To demonstrate that active learning with simple linear regression models can accelerate virtual screening.
• To show that complex models like deep learning are unnecessary for predicting low-sampling-depth docking results.

Main Methods:

• Employed active learning with simple linear regression models for virtual screening.
• Investigated active learning meta-parameters, identifying constant batch size and ensembling as optimal.
• Validated the approach on an ultralarge virtual screening dataset.

Main Results:

• Achieved retrieval of up to 90% of the top-1% docking hits after screening only 10% of ligands.
• Successfully retrieved 70% of the top-0.05% of ligands after screening merely 2% of an ultralarge library.
• Demonstrated the efficacy of simple models over complex deep learning approaches for this task.

Conclusions:

• Active learning with basic models offers a computationally accessible method for accelerated virtual screening.
• This approach can serve as a blueprint for developing low-compute agents for large-scale docking.
• Enhanced accessibility for academic researchers in rapid hit compound discovery for diverse targets.