Virtual screening of chemical libraries following experimental assays of drug candidates is a common procedure in structure-based drug discovery. However, virtual screening of chemical libraries with millions of compounds requires a lot of time for computing and data analysis. A priori classification of compounds in the libraries as low-and high-binding free energy sets decreases the number of compounds for virtual screening experiments. This classification also reduces the required computational time and resources. Data analysis is demanding since a compound can be described by more than one thousand attributes that make any data analysis very challenging. In this paper, we use the hyperbox classification method in combination with partial least squares regression to determine the most relevant molecular descriptors of the drug molecules for an efficient classification. The effectiveness of the approach is illustrated on a target protein, SIRT6. The results indicate that the proposed approach outperforms other approaches reported in the literature with 83.55% accuracy using six common molecular descriptors (SC-5, SP-6, SHBd, minHaaCH, maxwHBa, FMF). Additionally, the top 10 hit compounds are determined and reported as the candidate inhibitors of SIRT6 for which no inhibitors have so far been reported in the literature.