In addition to running FastHap on real HuRef data, we constructed several simulated read matrices based on HuRef data (Bansal and Bafna, 2008).
An example of the read matrix is shown in Table 1.
The length of a read in the read matrix is the number of heterozygous SNPs the read covers.
To evaluate how well the haplotype assembly can be done w.r.t. the sequencing protocols, we next construct a graph from the read matrix.
Although the mean of the insert length is 1000, the corresponding gap length in the read matrix is very small because only heterozygous SNPs are considered for assembly.
The number of changes we made is obviously 1. Therefore the 'haplotype assembly' problem is identical to finding a pair of haplotypes H such that the MEC score of the reads in the read matrix is minimized.
Since we generate only paired-end reads in our simulation, which consist of two segments, if only one segment of a read covers heterozygous SNPs, the resulting read in the read matrix will be considered as a single read, otherwise it is considered as a paired-end read.
The objective function we use is MEC, which is the minimum number of changes, or corrections, that need to be made in the read matrix such that the resulting matrix admits a perfect bi-partition, where each corrected read maps to either haplotype perfectly.
HASH (Bansal et al., 2008) and HapCut (Bansal and Bafna, 2008) algorithms are both based on the idea of building a graph from the read matrix, where each row corresponds to a read and each column corresponds to a position of the haplotype.
To be able to compare the present study and the former ones, we used the same method, estimating gene count richness using at 11 million unique reads matrix and a gene richness categorical variable computed by applying a threshold of 480 000 bacterial genes, to distinguish low from high gene count.
