Adding Stochastic Negative Examples into Machine Learning Improves Molecular Bioactivity Prediction.

Area of Science:

• Computational chemistry
• Pharmacology
• Machine learning

Background:

• Predicting ligand-target binding is crucial for drug discovery.
• Existing pharmacological datasets are often sparse, imbalanced, and approximate, limiting deep learning model performance.
• Conventional data splitting methods may not accurately reflect real-world testing scenarios.

Purpose of the Study:

• To develop and evaluate a novel data augmentation technique for improving multitask deep neural networks in predicting ligand-target binding.
• To address the challenges posed by sparse, imbalanced, and approximate pharmacological data.
• To create realistic benchmarks for evaluating model performance in temporal and drug-screening contexts.

Main Methods:

• Construction of two hold-out benchmarks simulating temporal and drug-screening test scenarios.
• Development of a data augmentation procedure named Stochastic Negative Addition (SNA).
• SNA randomly assigns untested molecule-target pairs as transient negative examples during training.

Main Results:

• Stochastic Negative Addition (SNA) significantly improved drug-screening benchmark performance (R² from 0.1926 ± 0.0186 to 0.4269 ± 0.0272, a 122% increase).
• A modest decrease in performance was observed for the temporal benchmark (13%).
• Performance gains were consistent across classification and regression tasks and superior to y-randomized controls.

Conclusions:

• Leveraging data uncertainty through methods like SNA can substantially improve predictions of drug-target relationships.
• The developed benchmarks provide a more realistic evaluation of predictive models.
• The study highlights the importance of addressing data limitations in pharmacological datasets for machine learning applications.