Estimating Cluster Centers And Variances - Newport iServer MicroServer iTHX-M Operator's Manual

Appendix D | Bayesian Robust Linear Model with Mahalanobis (BRLMM) Distance Classifier Al-
σ σ
symmetric with = . The distance d is computed as defined
1,2 2,1
G
above.
The confidence we assign to this call is d /d , where d is the smallest
1 2 1
distance of the three and d is the second smallest distance. This
2
confidence is always between zero and 1. It is a rough measure of the
quality of the call, but it is not a p-value. We set a threshold for quality
of 0.5 for a call/no-call decision, based on the performance of several
test data sets. This can be adjusted by the user to tune the tradeoff
between call rate and accuracy – see the results section for a
comparison of performance at various thresholds. To set the confidence
score threshold, see BRLMM Mapping Algorithm Settings BRLMM
Mapping Algorithm Settings (page 371).
The next section describes how we find the prototypes and how they
vary from the data for each SNP.

Estimating Cluster Centers and Variances

The previous section describes how to call genotypes and assign
confidence values to those calls given an appropriate prototype. This
section describes how to derive these prototypes.
Prototypes are derived by using an ad hoc Bayesian procedure. First, a
generic prior describing genotype clusters and centers for the typical
SNP is derived. Second, the generic SNP prior is combined with
initial genotype estimates for each specific SNP to derive a posterior
estimate of cluster centers and variances. This posterior estimate, as
described in the previous section, is what is used to call genotypes.
Figure D.4 below provides examples of SNPs to which this procedure
has been applied.