ChemFH

ChemFH identifies and predicts assay interferences and frequent hitters to reduce false-positive hits in high-throughput screening and support more accurate hit selection in drug discovery.

Key Features:

• Assay interference detection: Screens for colloidal aggregators, spectroscopic interference, firefly luciferase inhibitors, chemically reactive substances, and promiscuous compounds.
• Large-scale dataset: Models are built from a dataset of 823,391 chemical compounds.
• Model architecture and performance: Uses multi-task directed message-passing network (DMPNN) architectures with uncertainty estimation, achieving an average area under the curve (AUC) of 0.91.
• Substructure alerts and rules: Incorporates 1441 representative alert substructures and ten established screening rules for frequent hitters.
• Validation and application: Validated on an external set of 75 compounds and applied in virtual screenings across five libraries, including two natural products and FDA-approved drugs.

Scientific Applications:

• False-positive reduction in HTS: Identifies assay artifacts and frequent hitters to reduce false-positive rates in high-throughput screening.
• Hit triage for drug discovery: Prioritizes true positive hits by flagging compounds likely to cause assay interference.
• Library screening and annotation: Applied to virtual screening across multiple libraries and used to annotate natural product and FDA-approved drug datasets.

Methodology:

Prediction models were constructed from 823,391 chemical compounds using multi-task directed message-passing network (DMPNN) architectures with uncertainty estimation; 1441 representative alert substructures and ten established screening rules were used for interference prediction; validation was performed on an external set of 75 compounds and via virtual screenings across five libraries including two natural products and FDA-approved drugs.