Phred-equivalent Base Quality Scores

6.6	Phred-equivalent Base Quality Scores

Quality scores for individual called bases are determined by a method developed in collaboration with the Broad Institute(Genome Research,18(5): 763-70, 2008), whereby the methodology described by Ewing and Green (Genome Research, 8: 186-194, 1998) for the creation of quality scores as part of the Phred basecalling algorithm is applied to generating quality scores for 454 Sequencing reads. The quality scores computed for each called base are written to the CWF and SFF files (and optionally to a file paralleling the basecall FASTA file). Briefly, the method compares the properties of each base’s flowgram signals against properties that have been found to correlate with accurate and/or error-prone signal information, using training sets of read data. A multivariate analysis of those properties determines the sets of property values that best describe “bins” of basecalls, then assigns the training set accuracy rates of the basecalls in each bin as a quality score using the following scale:

Q = -10 log10 (error rate)

Then, when the basecalling of a new read is performed, the trace properties for each read are used to determine the bin in which each base falls, and the quality score associated with that bin is assigned to the basecall. For 454 Sequencing reads, the property metrics based on the following characteristics are used for the training and the quality score generation:

•	Base position of the base in the read sequence

•	Flow signal intensity for the flow where the base is called

•	Flow signal intensity one cycle before and/or one cycle after the flow where the base is called

•	Local variation of the flowgram signals in a window surrounding the flow where the base is called (i.e., how far are those signals from the ideal 0.0, 1.0, 2.0, 3.0, … signals). This provides an estimate of the homopolymer accuracy.

•	Overall “separation” of the flowgram signals (i.e., how greatly separated are the 0-mer signals from 1-mer signals)

Training sets using multiple data sets from different sample Runs were used for training the quality scores, and lookup tables were generated from those data sets. Separate training was performed for GS 20 chemistry, GS FLX standard chemistry, and GS Junior and GS FLX Titanium chemistry data, to account for the slightly different error tendencies of the three chemistries. These lookup tables are used to generate the base quality scores.