Ablation analysis of FSA on protein structural databases.

Ablation analysis of FSA on the protein benchmarks of Table 1 : Comparisons of the accuracies (Acc), sensitivities (Sn) and positive predictive values (PPV) of FSA with different components enabled or disabled. From top to bottom, FSA was run in default mode, –fast mode, with learning disabled, with iterative refinement disabled, and with 1 set (rather than 2 sets) of indel states; these options were then repeated for maximum-sensitivity mode (–maxsn). As made evident by the results (PPV) on the BAliBASE 3+fp and SABmark 1.65 datasets, query-specific learning helps FSA to distinguish homologous and non-homologous sequences. The above figures understate the utility of iterative refinement: while it generally has little effect on these small protein alignments, it occasionally dramatically reduces the number of small gaps and thereby improves the alignment accuracy.