Imputation of Assay Bioactivity Data Using Deep Learning.

Area of Science:

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

Background:

• Sparse bioactivity data is common in public and commercial drug discovery databases.
• Traditional quantitative structure-activity relationship (QSAR) models struggle with this type of data.
• Developing methods to effectively impute missing bioactivity values is crucial for efficient drug discovery.

Purpose of the Study:

• To introduce a novel deep learning neural network method for imputing assay pIC50 values.
• To demonstrate the method's ability to learn from sparse bioactivity data.
• To compare the performance of the deep learning method against traditional QSAR models.

Main Methods:

• A novel deep learning neural network was developed.
• The network was trained on sparse bioactivity data, learning correlations between different assays.
• The method was applied to two public domain datasets for validation.

Main Results:

• The deep learning method outperformed traditional QSAR models and other leading approaches.
• Accuracy (R^2) exceeded 0.9 when focusing on the most confident predictions.
• Overall prediction accuracy was significantly higher than previously reported methods.

Conclusions:

• The developed deep learning method offers a powerful new approach for bioactivity imputation.
• This method is particularly effective for handling sparse bioactivity data.
• The approach has the potential to significantly enhance drug discovery efficiency by improving data utilization.