Benchmark data sets for structure-based computational target prediction.

Area of Science:

• Computational chemistry
• Drug discovery
• Bioinformatics

Background:

• Structure-based computational target prediction methods aim to identify potential targets for bioactive compounds.
• Current protein-ligand docking methods struggle with ranking true targets in large datasets.
• A lack of standard evaluation datasets hinders method comparison and improvement demonstration.

Purpose of the Study:

• To propose two novel datasets and evaluation strategies for meaningful assessment of target prediction methods.
• To facilitate proof-of-concept, selectivity studies, and large-scale statistical evaluation.
• To enable robust comparison and validation of new computational target prediction techniques.

Main Methods:

• Development of a small dataset (three target classes) for proof-of-concept and selectivity studies.
• Creation of a large dataset (7992 protein structures, 72 ligands) for statistical evaluation.
• Proposal of performance metrics such as AUC, BEDROC, and NSLR for early enrichment recognition.

Main Results:

• The iRAISE inverse screening method achieved excellent or good enrichment in 55% of cases.
• A median AUC of 0.67 was obtained, with RMSDs below 2.0 Å for 74% of predictions.
• iRAISE successfully predicted the first true target within the top 2% for 59 out of 72 cases.

Conclusions:

• The proposed datasets and strategies provide a robust framework for evaluating target prediction methods.
• The iRAISE method shows significant promise for accurate and efficient target identification in drug discovery.
• These resources will accelerate the development and validation of next-generation computational drug discovery tools.