RosettaLigand has been successfully used to predict binding poses in protein-small molecule complexes. However, the RosettaLigand docking protocol is comparatively slow in identifying an initial starting pose for the small molecule (ligand) making it unfeasible for use in virtual High Throughput Screening (vHTS). To overcome this limitation, we developed a new sampling approach for placing the ligand in the protein binding site during the initial 'low-resolution' docking step. It combines the translational and rotational adjustments to the ligand pose in a single transformation step. The new algorithm is both more accurate and more time-efficient. The docking success rate is improved by 10-15% in a benchmark set of 43 protein/ligand complexes, reducing the number of models that typically need to be generated from 1000 to 150. The average time to generate a model is reduced from 50 seconds to 10 seconds. As a result we observe an effective 30-fold speed increase, making RosettaLigand appropriate for docking medium sized ligand libraries. We demonstrate that this improved initial placement of the ligand is critical for successful prediction of an accurate binding position in the 'high-resolution' full atom refinement step.