Balancing Data on Deep Learning-Based Proteochemometric Activity Classification.

Area of Science:

• Computational chemistry
• Cheminformatics
• Machine learning in drug discovery

Background:

• In silico analysis of biological activity data is crucial for pharmaceutical development.
• Proteochemometric models leverage machine learning for ligand-target activity prediction by sharing information across targets.
• Bioactivity datasets often exhibit imbalance, potentially compromising model performance.

Purpose of the Study:

• To investigate the impact of various data balancing strategies on deep learning proteochemometric models.
• To control for compound series bias using clustering techniques within these models.
• To evaluate balancing strategies across diverse protein families (kinases, GPCRs, nuclear receptors, proteases).

Main Methods:

• Exploration of four balancing strategies: no_resampling, resampling_after_clustering, resampling_before_clustering, and semi_resampling.
• Application of clustering to mitigate compound series bias.
• Evaluation of models on BindingDB datasets encompassing kinases, GPCRs, nuclear receptors, and proteases.
• Utilizing deep learning for target-compound activity classification.

Main Results:

• The predicted proportion of positive activity predictions correlated with the actual data balance in the test set.
• Data balancing strategies demonstrably influenced the performance estimates of proteochemometric models.
• The semi_resampling strategy, involving data augmentation and clustering, showed promise in mitigating imbalance effects.

Conclusions:

• Data balance significantly impacts the performance estimation of proteochemometric models.
• A semi-resampling approach, integrating data augmentation and clustering in the training set, is recommended for realistic scenarios.
• This strategy helps mitigate data imbalance effects, leading to more reliable predictions in pharmaceutical development.