The Effect of Debiasing Protein-Ligand Binding Data on Generalization.

Area of Science:

• Computational chemistry
• Machine learning in drug discovery

Background:

• Machine-learning models for protein-ligand binding prediction risk overfitting due to structured chemical data.
• Overfitting impairs model generalization and accuracy for novel candidate ligands.

Purpose of the Study:

• To investigate the impact of data debiasing on the generalization ability of machine-learning models in predicting protein-ligand binding.
• To determine if debiased data splits improve model performance on novel chemical entities.

Main Methods:

• Utilized distance-based data splits to assess model generalization.
• Compared model performance on randomly split versus distant held-out test sets.
• Applied data debiasing algorithms to the training data.

Main Results:

• Models initially performed better on random splits than distant splits, confirming generalization challenges.
• Surprisingly, debiasing data typically reduced model accuracy on distant held-out test sets.
• Post-debiasing performance metrics did not accurately reflect the model's ability to generalize.

Conclusions:

• Data debiasing may reduce the information available to machine-learning models, impairing their generalization capabilities.
• Current debiasing strategies might not be optimal for improving predictive accuracy in protein-ligand binding tasks.
• Further research is needed to develop effective debiasing methods that enhance, rather than hinder, model generalization.