Deep active learning with high structural discriminability for molecular mutagenicity prediction.

Area of Science:

• Computational chemistry
• Toxicology
• Drug discovery

Background:

• Mutagenicity assessment is vital in drug discovery to prevent cancer and germ cell damage.
• In silico mutagenicity prediction is hampered by limited labeled molecular data.
• Experimental testing is costly and time-consuming, necessitating cost-effective annotation strategies.

Purpose of the Study:

• To introduce muTOX-AL, a deep active learning framework for efficient mutagenicity prediction.
• To reduce the cost of molecular annotation in drug discovery.
• To enhance the performance of in silico mutagenicity prediction models with limited data.

Main Methods:

• Development of a deep active learning framework (muTOX-AL).
• Active exploration of chemical space to identify valuable molecules for annotation.
• Utilizing an oracle (e.g., human expert) for targeted molecular labeling.

Main Results:

• muTOX-AL achieved competitive performance with a small number of labeled samples.
• Reduced the number of required training molecules by approximately 57% compared to random sampling.
• Demonstrated superior ability to select molecules with high structural similarity but differing mutagenic properties.

Conclusions:

• muTOX-AL offers an efficient solution for mutagenicity assessment in drug discovery.
• The framework significantly lowers annotation costs while maintaining high predictive performance.
• muTOX-AL's structural discriminability aids in identifying critical molecular features for toxicity prediction.