Thompson Sampling─An Efficient Method for Searching Ultralarge Synthesis on Demand Databases

Area of Science:

• Computational chemistry
• Drug discovery
• Machine learning

Background:

• Virtual screening of ultralarge synthesis on-demand libraries is crucial for hit identification in drug discovery.
• The exponential growth of chemical libraries (tens of billions of molecules) makes exhaustive virtual screening cost-ineffective.
• Heuristic search methods are needed to efficiently identify promising molecules from vast chemical spaces.

Purpose of the Study:

• To introduce and demonstrate Thompson sampling (TS) as an active learning approach for streamlining virtual screening.
• To showcase TS's ability to perform probabilistic searches in reagent space, avoiding full library enumeration.
• To illustrate the broad applicability of TS across various virtual screening modalities.

Main Methods:

• Thompson sampling (TS), an active learning algorithm, was employed for probabilistic virtual screening.
• The TS approach was applied to a docking-based virtual screen of a large combinatorial library.
• The method was evaluated on its efficiency in identifying top-ranked molecules compared to exhaustive screening.

Main Results:

• TS successfully identified over 50% of the top 100 molecules from a 335 million-molecule dataset.
• This achievement was accomplished by evaluating only 1% of the total dataset, demonstrating significant computational savings.
• The study confirmed TS's effectiveness in diverse virtual screening applications, including similarity searches and machine learning models.

Conclusions:

• Thompson sampling offers a computationally efficient and effective strategy for hit identification in drug discovery.
• TS significantly reduces the resources required for virtual screening of ultralarge chemical libraries.
• This active learning approach represents a paradigm shift in managing and searching massive molecular databases for drug development.