Turbo similarity searching: Effect of fingerprint and dataset on virtual-screening performance

Turbo similarity searching uses information about the nearest neighbours in a conventional chemical similarity search to increase the effectiveness of virtual screening, with a data fusion approach being used to combine the nearest-neighbour information. A previous paper suggested that the approach was highly effective in operation; this paper further tests the approach using a range of different databases and of structural representations. Searches were carried out on three different databases of chemical structures, using seven different types of fingerprint, as well as molecular holograms, physicochemical properties, topological indices and reduced graphs. The results show that turbo similarity searching can indeed enhance retrieval but that this is normally achieved only if the similarity search that acts as its starting point has already achieved at least some reasonable level of search effectiveness. In other cases, a modified version of TSS that uses the nearest-neighbour information for approximate machine learning can be used effectively. Whilst useful for qualitative (active/inactive) predictions of biological activity, turbo similarity searching does not appear to exhibit any predictive power when quantitative property data is available.