Modeling Small-Molecule Reactivity Identifies Promiscuous Bioactive Compounds.

Area of Science:

• Computational chemistry
• Drug discovery
• Cheminformatics

Background:

• High-throughput screening (HTS) is crucial for identifying drug candidates.
• Promiscuous bioactive compounds, or "frequent hitters," are often false positives in HTS.
• Pan assay interference compounds (PAINS) filters are commonly used to identify these problematic compounds.

Purpose of the Study:

• To evaluate deep learning models for predicting promiscuous bioactivity.
• To assess the performance of deep learning models compared to PAINS filters.
• To explore the complementary nature of deep learning reactivity modeling and PAINS filters.

Main Methods:

• Developed a deep learning model to predict small-molecule reactivity.
• Assessed model performance on PubChem assays without HTS data training.
• Compared deep learning model performance against PAINS filters.
• Evaluated combined deep learning and PAINS filter approaches.

Main Results:

• Deep learning model achieved 18.5% sensitivity and 95.5% specificity for promiscuous compounds.
• PAINS filters achieved 20.3% sensitivity at the same specificity.
• Combined approach yielded 24% sensitivity at 95.5% specificity.
• Reactivity model identified specific atoms involved in assay interference.

Conclusions:

• Deep learning reactivity modeling is a viable alternative/complement to PAINS filters.
• Combining deep learning and PAINS filters enhances promiscuous compound identification.
• Deep learning models offer mechanistic hypotheses for assay interference.