Tuning gradient boosting for imbalanced bioassay modelling with custom loss functions.

Area of Science:

• Computational chemistry
• Machine learning in drug discovery
• Bioinformatics

Background:

• Bioassay datasets are crucial for drug discovery but often exhibit severe class imbalance.
• This imbalance hinders the performance of machine learning models in identifying active compounds.

Purpose of the Study:

• To investigate alternative loss functions for imbalanced classification within Gradient Boosting models.
• To benchmark these novel approaches against conventional methods on diverse bioassay data.

Main Methods:

• Exploration of computer vision-inspired loss functions for imbalanced classification.
• Application and benchmarking of these loss functions within a Gradient Boosting framework.
• Evaluation across six public and proprietary bioassay datasets, totaling 42 tasks and 2 million compounds.

Main Results:

• Statistically significant performance improvements were observed on five out of six datasets compared to standard cross-entropy loss.
• Gradient Boosting with optimized loss functions matched or surpassed existing classifiers and neural networks.
• Training convergence speed increased up to 8 times faster.

Conclusions:

• Tuning loss functions is an effective and efficient strategy for Gradient Boosting on imbalanced bioassay data.
• This approach achieves state-of-the-art performance without sacrificing interpretability or scalability.
• Offers a practical solution for improving drug discovery pipelines.