E inference problem is solved by clustering overlapping reads such that each cluster corresponds to one viral haplotype [17,18,19]. In highly diverse virus populations, such as RNA or singlestranded DNA viruses, mutations can be so frequent that they maybe phased even if they are not observed on the same read. This global haplotype reconstruction problem becomes feasible if SNVs can be connected by a series of partially overlapping reads. It can be regarded as a sequence assembly problem from short reads, with the goal of reconstructing a viral quasispecies, i.e., a set of related sequences, rather than a single genome. Computational methods for viral quasispecies assembly include combinatorial optimization techniques [17,20,21,22,23] and generative probabilistic models [24,25,26]. SNV calling and local and global haplotype reconstruction assess viral genetic diversity at different spatial scales, ranging from single 1326631 sites to the whole genome. Long-range haplotype reconstructions are more informative than short-range inference, because the linkage between mutations often has important phenotypic consequences. On the other hand, the statistical power to detect variation is highest for local haplotypes, and the computational complexity of haplotype assembly increases with the length of the genomic region. The optimal scale of diversity estimation also depends on the employed NGS platform and the read data it generates. Among other factors, NGS technologies differ in the number of reads they produce per run, the read length, the error pattern, and the cost per base [3]. However, it is unknown how sequencing platforms compare across the different viral diversity estimation tasks. Here, we address this question and compare the two most commonly used NGS platforms for viral diversity estimation, namely 454/Roche pyrosequencing [27] and Illumina Genome Analyzer [28]. Previously, both platforms have been shown toViral Quasispecies Reconstructionexhibit similar mismatch error rates, while 454/Roche had an increased indel error rate in homopolymeric regions [29]. Instead of error profiles, we focus here on coverage and read length, two critical parameters for viral diversity estimation. Whereas 454/ Roche produces longer reads, Illumina reaches higher coverage per run at lower costs, suggesting more power to detect lowfrequency local variation with Illumina, but more power to assemble global haplotypes with 454/Roche data. We investigate this trade-off by analyzing a mixture of patient-derived viral clones that has been sequenced on both platforms and by simulated reads. We show how coverage, read length, and error rate jointly affect the performance of local and global haplotype inference. Our results provide guidance for the optimal choice of a NGS platform in viral diversity studies.Multiple sequence alignmentThe tool `s2f.py’ get (-)-Indolactam V included in the software 
package ShoRAH [30] was used to build a multiple sequence alignment (MSA) of the 454/Roche reads mapping the region corresponding to amino acids from 10 to 93 on the 223488-57-1 biological activity protease gene. Insertions were discarded as set in the default options. Only reads covering at least 80 of the region were retained. Illumina reads were aligned against the HXB2 reference sequence with the read mapper Novoalign (version 2.07.18, default parameter options, http:// www.novocraft.com/). The paired ends were aligned independently. The output, in SAM format, was parsed with a custom script to estimate the diversity at ea.E inference problem is solved by clustering overlapping reads such that each cluster corresponds to one viral haplotype [17,18,19]. In highly diverse virus populations, such as RNA or singlestranded DNA viruses, mutations can be so frequent that they maybe phased even if they are not observed on the same read. This global haplotype reconstruction problem becomes feasible if SNVs can be connected by a series of partially overlapping reads. It can be regarded as a sequence assembly problem from short reads, with the goal of reconstructing a viral quasispecies, i.e., a set of related sequences, rather than a single genome. Computational methods for viral quasispecies assembly include combinatorial optimization techniques [17,20,21,22,23] and generative probabilistic models [24,25,26]. SNV calling and local and global haplotype reconstruction assess viral genetic diversity at different spatial scales, ranging from single 1326631 sites to the whole genome. Long-range haplotype reconstructions are more informative than short-range inference, because the linkage between mutations often 
has important phenotypic consequences. On the other hand, the statistical power to detect variation is highest for local haplotypes, and the computational complexity of haplotype assembly increases with the length of the genomic region. The optimal scale of diversity estimation also depends on the employed NGS platform and the read data it generates. Among other factors, NGS technologies differ in the number of reads they produce per run, the read length, the error pattern, and the cost per base [3]. However, it is unknown how sequencing platforms compare across the different viral diversity estimation tasks. Here, we address this question and compare the two most commonly used NGS platforms for viral diversity estimation, namely 454/Roche pyrosequencing [27] and Illumina Genome Analyzer [28]. Previously, both platforms have been shown toViral Quasispecies Reconstructionexhibit similar mismatch error rates, while 454/Roche had an increased indel error rate in homopolymeric regions [29]. Instead of error profiles, we focus here on coverage and read length, two critical parameters for viral diversity estimation. Whereas 454/ Roche produces longer reads, Illumina reaches higher coverage per run at lower costs, suggesting more power to detect lowfrequency local variation with Illumina, but more power to assemble global haplotypes with 454/Roche data. We investigate this trade-off by analyzing a mixture of patient-derived viral clones that has been sequenced on both platforms and by simulated reads. We show how coverage, read length, and error rate jointly affect the performance of local and global haplotype inference. Our results provide guidance for the optimal choice of a NGS platform in viral diversity studies.Multiple sequence alignmentThe tool `s2f.py’ included in the software package ShoRAH [30] was used to build a multiple sequence alignment (MSA) of the 454/Roche reads mapping the region corresponding to amino acids from 10 to 93 on the protease gene. Insertions were discarded as set in the default options. Only reads covering at least 80 of the region were retained. Illumina reads were aligned against the HXB2 reference sequence with the read mapper Novoalign (version 2.07.18, default parameter options, http:// www.novocraft.com/). The paired ends were aligned independently. The output, in SAM format, was parsed with a custom script to estimate the diversity at ea.