Ligand-based virtual screening using a genetic algorithm with data fusion

Substructural analysis provides a simple and effective way of ranking the 2D fingerprints representing the molecules in a database upon the basis of weights that denote a substructural fragment’s contribution to the overall activity or inactivity of a molecule. A substructural analysis method has been described recently that is based on the use of a genetic algorithm (GA), with the resulting sets of weights proving to be more effective for ligand-based virtual screening than existing approaches. However, the inherently non-deterministic nature of a GA means that different runs are likely to result in different sets of weights and hence in variations in the effectiveness of screening. This paper describes the use of data fusion to combine the rankings generated in multiple GA runs, and demonstrates that the resulting fused rankings are markedly superior to GA runs on average, and in some cases can even exceed the performance of the very best individual GA run.