With the exponential growth of genomics data, the demand for reliable clustering methods is increasing every day. Despite the wide usage of many clustering algorithms, the accuracy of these algorithms has been evaluated mostly on simulated data sets and seldom on real biological data for which a "correct answer" is available. In order to address this issue, we use the manually curated high-quality Arabidopsis thaliana gene family database as a "gold standard" to conduct a comprehensive comparison of the accuracies of four widely used clustering methods including K-means, TribeMCL, single-linkage clustering and complete-linkage clustering. We compare the results from running different clustering methods on two matrices: the E-value matrix and the k-tuple distance matrix. The E-value matrix is computed based on BLAST E-values. The k-tuple distance matrix is computed based on the difference in tuple frequencies. The TribeMCL with the E-value matrix performed best, with the Inflation parameter (=1.15) tuned considerably lower than what has been suggested previously (=2). The single-linkage clustering method with the E-value matrix was second best. Single-linkage clustering, K-means clustering, complete-linkage clustering, and TribeMCL with a k-tuple distance matrix performed reasonably well. Complete-linkage clustering with the k-tuple distance matrix performed the worst.