Convolutional architectures for virtual screening.

Area of Science:

• Computational chemistry
• Machine learning in drug discovery
• Bioactivity prediction

Background:

• Virtual screening algorithms must adapt to different drug discovery stages.
• Early screening prioritizes identifying all bioactive compounds, accepting false positives.
• Later screening rounds focus on enhancing prediction accuracy.

Purpose of the Study:

• To develop a novel Convolutional Neural Network (CNN) architecture for improved virtual screening.
• To predict the bioactivity of candidate compounds against CDK1.
• To optimize performance for both early and late stages of screening.

Main Methods:

• Utilized a novel CNN architecture for bioactivity prediction.
• Employed molecular fingerprints as vector representations of compounds.
• Trained the model to balance enrichment factor and prediction accuracy.

Main Results:

• Achieved 98.55% accuracy in active-only selection for early screening.
• Reached 98.88% accuracy in high-precision discrimination for later screening.
• Demonstrated superior performance in different screening modes.

Conclusions:

• The proposed CNN architecture surpasses current state-of-the-art machine learning methods.
• The study provides novel insights into the utility of molecular fingerprints.
• The model effectively addresses the adaptive needs of virtual screening algorithms.