Improving quantitative structure-activity relationship models using Artificial Neural Networks trained with dropout.

Area of Science:

• Computational chemistry
• Machine learning
• Drug discovery

Background:

• Artificial Neural Networks (ANNs) are powerful machine learning tools.
• Dropout is a known technique to improve ANN performance on standard datasets.
• Quantitative Structure-Activity Relationship (QSAR) datasets present unique challenges for ANNs, including data bias and a high descriptor-to-active ratio.

Purpose of the Study:

• To evaluate the effectiveness of dropout in improving ANN performance for QSAR modeling.
• To benchmark dropout's impact on QSAR datasets within Ligand-Based Computer-Aided Drug Discovery.

Main Methods:

• A benchmark study was conducted on nine large QSAR datasets.
• Artificial Neural Networks (ANNs) were trained with and without the dropout technique.
• Performance was assessed using metrics such as enrichment false positive rate and log-scaled area under the receiver-operating characteristic curve (logAUC).

Main Results:

• Dropout significantly improved both enrichment false positive rate and logAUC by 22-46% compared to conventional ANN implementations.
• Optimal dropout rates were identified as a function of the descriptor set's signal-to-noise ratio, showing independence from the specific dataset.
• ANNs utilizing dropout with 2D and 3D autocorrelation descriptors outperformed standard ANNs and optimized fingerprint similarity search methods.

Conclusions:

• Dropout is an effective technique for enhancing ANN performance in QSAR modeling for drug discovery.
• The findings suggest that dropout can help overcome challenges posed by QSAR datasets, leading to more accurate predictions.
• This study highlights the potential of dropout-enhanced ANNs as a valuable tool in Ligand-Based Computer-Aided Drug Discovery.