However, for datasets with errors, the NFI fell rapidly to below 10% of NFI for some mappers (PASS, BWA, and GSNAP), while others (TMAP, SSAHA2, SMALT, MOSAIK, and Novoalign) maintained a high NFI percentage for datasets with up to a 4% error rate and finished at between 50 and 20% NFI for an 8% error rate (Novoalign fell rapidly and finished below 10%).
All the mappers identified between 100% and 95% of the NFI for datasets with no errors.
All analyses involving Pfeiffer scores comprised pooled estimates of the 5 imputed datasets, with standard errors computed by taking the within- and between-imputation components into account [ 16].
We used simulated datasets with varying error rates to compare mapper robustness.
For datasets with varying error rates, indel sizes and read lengths existed; the results are shown in Figure 8.
For example, all datasets with an error rate >3.3%% are poorly corrected when using Coral default parameters.
Segemehl showed the best F-measures even for datasets with high error rates and for all read lengths considered in this study.
In another experiment, we produced four series of artificial datasets with different error models, using the ESTsim tool (Hazelhurst and Bergheim, 2003).
To indicate alignment accuracy evaluation for short-read datasets with varying error rate existing, the results are shown in Figure 5(a) for sensitivity and Figure 5(b) for precision comparison.
To assess the multiple aligners, real-life short-read datasets and long-read datasets sampled from various organisms and in silico datasets with varying error rates, indel sizes, and read lengths were considered as standard datasets for different applications and sequencing technologies.
GASSST, SHRiMP2, GenomeMapper, SeqMap, RazerS, mrFAST, mrsFAST, Bowtie, and BOAT could provide relatively high percentage of total multimapped reads (>20%) and high percentage of corrected multimapped reads as well when dealing with short-read datasets with varying error rates and indel sizes, especially SHRiMP2 and 8(b)).
