On the Analysis of Optical Mapping Data - University of Wisconsin ...

49
3.4 Discussion
3.4.1 Uses
Alignment is a fundamental problem in optical mapping. From a statistical perspective,
optical map alignment is more challenging than DNA sequence alignment because optical
map data are more noisy. Prior work in this area has mostly focused on developing score
functions that can be used in DP algorithms. Here we have proposed a general framework to
study the null distributions of optimal scores for an arbitrary score function. Its most obvious
use is to derive significance tests with direct control over error rates. We have demonstrated
the usefulness of such tests in improving assembly of large genomes.
Evaluating score functions: The methods described above are applicable to any score
function and provide a natural mechanism to evaluate them. We consider here the modelbased
likelihood ratio (LR) score proposed by Valouev et al. (2006) for aligning optical maps
to an in silico reference. Figure 3.7 plots the best spurious ungapped global alignment score
against two replications from P G using this score. The correlation is weaker, but a map
specific cutoff is still more appropriate than a constant cutoff. We apply the direct approach
as before with n = 4 replications to estimate µ(M). The results, shown in table 3.2, indicate
that at least for the particular sets of parameters used, the SOMA score is more sensitive at a
comparable specificity. This is somewhat surprising, since the LR score is based on a formal
likelihood ratio test whereas the SOMA score is largely heuristic. Informal experiments
suggest that this is at least in part due to the sizing model used by Valouev et al. (2006),
which does not consider scaling errors and consequently underestimates the marginal sizing
variance for large fragments.
More generally, this framework can be used for exploratory purposes, e.g. to compare the
performance of different scores, or to guide the choice of parameters for a given score. It is
helpful, particularly for overlap alignments (required in iterative assembly to extend flanks
of a contig), if the distribution of the optimal score under the null does not depend strongly
on the map, since otherwise significant alignments can be masked by spurious alignments.