Methods based on paired-end read mapping identify large indel events by comparing the expected distance between read pairs to the actual mapped distance. Such methods include PEMer, Hydra, and BreakDancer (25,27). For example, in the case of a 50-bp insertion, if the distance between read pairs is normally distributed, with mean 200 bp and then multiple pairs aligning to the same area with a distance between read pairs of approximately 150 bp would result in an insertion call (Figure 3C). Paired-end read mapping methods are therefore able to detect medium-sized insertions and deletions from mapped data. However, in most cases, the exact inserted or deleted sequence will not be known. Another major drawback of paired-end read mapping methods is that they are insensitive to small insertion or deletion events, owing to the difficulty in separating small perturbations in read-pair distance from the normal background variability. Another new class of indel-detection software is those based on machine learning methods in which insertions and deletions identified by various methods are filtered against empirically derived training set data to reduce the false-positive rate
