Boosting Virtual Screening Enrichments with Data Fusion: Coalescing Hits from Two-Dimensional Fingerprints, Shape, and Docking

Virtual
screening is an effective way to find hits in drug discovery, with
approaches ranging from fast information-based similarity methods
to more computationally intensive physics-based docking methods. However,
the best approach to use for a given project is not clear in advance
of the screen. In this work, we show that combining results from multiple
methods using a standard score (Z-score) can significantly
improve virtual screening enrichments over any of the single screening
methods. We show that an augmented Z-score, which
considers the best two out of three scores for a given compound, outperforms
previously published data fusion algorithms. We use three different
virtual screening methods (two-dimensional (2D) fingerprint similarity,
shape-based similarity, and docking) and study two different databases
(DUD and MDDR). The average enrichment in the top 1% was improved
by 9% for DUD and 25% for the MDDR, compared with the top individual
method. Improvements of 22% for DUD and 43% for MDDR are seen over
the average of the three individual methods. Statistics are presented
that show a high significance associated with the findings in this
work.