Effectiveness of iterative in silico–in vitro cycles

Determine whether iterative in silico–in vitro cycles of AI model training and experimental testing improve effectiveness over single-pass training on medium-throughput in vitro screens for target-specific small-molecule lead discovery using learning algorithms such as random forest and directed message passing neural networks.

Background

The paper discusses strategies for overcoming small labeled datasets in drug discovery, including generating target-specific training data via medium-throughput in vitro screens and training models such as random forests or directed message passing neural networks on these datasets. While these approaches have yielded fast and effective discovery of potential drug leads, the authors explicitly note uncertainty regarding the benefit of iterative feedback loops that alternate computational predictions with new experimental data collection.