Probabilistic Random Forest improves bioactivity predictions close to the classification threshold by taking into...

Area of Science:

• Computational chemistry and cheminformatics
• Drug discovery and development
• Machine learning in bioinformatics

Background:

• Protein-ligand interaction measurements have inherent reproducibility limits due to experimental errors.
• These experimental uncertainties are often neglected in predictive modeling, impacting model performance.
• Existing models may overfit or be overconfident due to ignoring error margins.

Purpose of the Study:

• To present a novel Probabilistic Random Forest (PRF) classifier for predicting protein-ligand interactions.
• To incorporate experimental standard deviations (σ) into the modeling process.
• To improve the accuracy and reliability of in silico target prediction models.

Main Methods:

• Applied a Probabilistic Random Forest (PRF) algorithm to ~550 protein target prediction tasks from ChEMBL and PubChem.
• Evaluated predictions by considering various scenarios of experimental standard deviations in training and test sets.
• Used fivefold stratified shuffled splits for robust model validation.

Main Results:

• PRF demonstrated significant benefits when experimental deviations were incorporated, particularly for data near the binary threshold (0.4-0.6 probability).
• PRF achieved a median absolute error reduction of ~17% compared to standard Random Forest (RF) in these critical regions.
• Standard RF models showed higher confidence but produced errors smaller than experimental uncertainty, suggesting potential overtraining.

Conclusions:

• Probabilistic Random Forest (PRF) is a valuable tool for protein target prediction, especially with overlapping class boundaries and measurement uncertainty.
• PRF improves model reliability where data points are close to the classification threshold.
• Excluding putative inactives without assigned experimental values improved PRF model performance.