Accurate and efficient target prediction using a potency-sensitive influence-relevance voter.

Area of Science:

• Chemoinformatics
• Computational Biology
• Drug Discovery

Background:

• Ligand-based algorithms using chemical similarity are common for predicting molecule-protein interactions.
• These methods are computationally efficient, making them suitable for large datasets.
• Assessing algorithm performance is crucial as data availability increases.

Purpose of the Study:

• To evaluate the performance of various target-prediction algorithms.
• To identify improvements for current drug-target prediction practices.
• To enhance the accuracy of predicting biological targets for diverse molecules.

Main Methods:

• A large-scale assessment was conducted using a ChEMBL-derived database of 490,760 molecule-protein interactions and 3236 protein targets.
• Performance was evaluated using threefold cross-validation, simulated screening with inactive molecules, and external test sets.
• A modified influence-relevance voter (IRV) algorithm, termed potency-sensitive IRV (PS-IRV), was developed.

Main Results:

• The study assessed multiple target-prediction algorithms on a substantial dataset.
• Incorporating molecule potency data improved target prediction accuracy.
• The inclusion of random inactive molecules during training enhanced algorithm performance in simulated screens.
• The PS-IRV algorithm achieved the best results on large test sets.

Conclusions:

• Potency data can significantly improve molecule target prediction accuracy.
• Training with inactive molecules boosts the performance of several prediction algorithms.
• The developed PS-IRV offers superior performance in realistic target-prediction experiments.
• Models and software are publicly available for broader use.