Self-Focusing Virtual Screening with Active Design Space Pruning.

Area of Science:

• Computational chemistry
• cheminformatics
• Drug discovery

Background:

• High-throughput virtual screening (HTVS) is crucial for identifying small molecules.
• Exhaustive virtual screening of large libraries is often cost-prohibitive.
• Model-guided optimization (MGO) enhances sample efficiency but incurs surrogate model training and inference costs.

Purpose of the Study:

• To introduce design space pruning (DSP) as a method to mitigate inference costs in MGO.
• To evaluate the effectiveness of DSP in reducing computational overhead during optimization.

Main Methods:

• Proposed an extension to MGO incorporating DSP.
• DSP irreversibly removes underperforming candidates from the search space.
• Applied DSP to various molecular optimization tasks.

Main Results:

• Observed significant reductions in computational overhead costs.
• Achieved comparable performance to baseline MGO methods.
• DSP effectively limits overhead costs in MGO workflows.

Conclusions:

• DSP is an effective strategy for reducing MGO inference costs.
• This technique is particularly valuable when overhead costs are substantial relative to objective costs (e.g., docking).
• DSP offers a cost-efficient enhancement to MGO for virtual screening applications.