Low-Quality Structural and Interaction Data Improves Binding Affinity Prediction via Random Forest.

Area of Science:

• Computational chemistry
• Structural biology
• Machine learning

Background:

• Protein-ligand binding affinity prediction is crucial in drug discovery.
• Scoring functions are essential tools for this prediction.
• The impact of low-quality training data on scoring function performance is poorly understood.

Purpose of the Study:

• To systematically investigate the effect of low-quality data on the predictive performance of docking scoring functions.
• To compare the performance of machine-learning scoring functions versus classical scoring functions when trained on varying data quality.

Main Methods:

• Machine learning and classical scoring functions were trained and tested using datasets with varying proportions of low-quality structural and binding data.
• Performance was evaluated based on predictive accuracy for protein-ligand binding affinity.

Main Results:

• Contrary to common belief, low-quality data was found to be beneficial, not detrimental, for machine-learning scoring function performance.
• High-quality data still yielded greater improvements, but larger datasets of any quality outperformed smaller, high-quality sets.
• Classical scoring functions showed limited improvement beyond an initial data threshold, irrespective of data quality.

Conclusions:

• Exploiting larger data volumes is more critical for machine-learning scoring function performance than strictly adhering to high-quality data.
• Machine-learning approaches demonstrate a superior ability to leverage extensive datasets compared to classical methods.