Using the 94.3 kb (23.2 calibration + 71.1 kb mutation hotspots) of DNA sequenced for assay calibration, there were 2,770 SNP positions in only one of the four CAL samples recorded in the dbSNP131 database. We first masked these positions examining only invariant bases to estimate the position-dependent error rate. The error rate is defined, at each invariant base, as the fraction of non-reference bases. Any position with an alternative frequency greater than 0.05 was filtered out because this would most likely result from a non-annotated SNP or a PCR error occurring upstream of sequencing. After filtering out the positions covered by less than ten reads, we calculated the error rate as a function of the strand (forward or reverse), the position on the reads (1 to 122) as well as the type of substitution observed (4 × 3 = 12 possibilities when considering strands separately), which are usual predictors of error in sequencing by synthesis [6]. The retained positions were stratified into bins corresponding to the substitution type (reference to mismatch substitution), the read position, and the strand. An
