Virtual screening has become a successful alternative and complementary technique to experimental high-throughput screening technologies for drug design. Since the scoring function of docking software cannot predict binding affinity accurately, how to improve the hit rate remains a common issue in structure-based virtual screening. This paper proposed a target-specific virtual screening method based on ensemble learning named ENS-VS. In this method, protein–ligand interaction energy terms and structure vectors of the ligands were used as a combination descriptor. Support vector machine, decision tree and Fisher linear discriminant classifiers were integrated into ENS-VS for predicting the activity of the compounds. The results showed that the enrichment factor (EF) 1% of ENS-VS was 6 times higher than that of Autodock vina. Compared with the newest virtual screening method SIEVE-Score, the mean EF 1% and AUC of ENS-VS (mean EF 1% = 52.77, AUC = 0.982) were statistically significantly higher than those of SIEVE-Score (mean EF 1% = 42.64, AUC = 0.912) on DUD-E datasets; and the mean EF 1% and AUC of ENS-VS (mean EF 1% = 29.73, AUC = 0.793) were also higher than those of SIEVE-Score (mean EF 1% = 25.56, AUC = 0.765) on eight DEKOIS datasets. ENS-VS also showed significant improvements compared with other similar research. The source code is available at http://github.com/eddyblue/ENS-VS.