Figure 8

Direct SNV candidate identification from DDiMAP filtered sequences compares favorably with VarScan2. Mapped BCL2 reads from the 12 FL specimens were individually analyzed by VarScan2 (1% threshold) and compared to Sanger sequencing of amplicons to identify founder clone genotypes for use as gold-standard SNV data (see Additional file 1 for details). Mapped SOLiD read data were pooled to generate a collection of variants at a wide range of low frequencies, with 24/240 at levels below 0.2% and 120/240 at frequencies below 2.1%. (A) Precision-recall curves were obtained using a threshold series in VarScan2 (0.1% to 1%) and DDiMAP (100 ppm to 800 ppm) to demonstrate their well-matched overall performance. VarScan2 requires a threshold of 0.5% to achieve 100% PPV with a corresponding sensitivity of 85.4% (F1-score = 0.92). Matching performance is obtained using DDiMAP with a threshold setting of 400 ppm. Peak F1-scores occur at lower threshold values of 0.2% for VarScan2 and 300 ppm for DDiMAP, reflecting different P-R trade-offs, with DDiMAP better for precision and VarScan2 better for recall and nearly matched to the 200 ppm position of DDiMAP. (B) DDiMAP logistic sensitivity models were obtained using two conservative thresholds that generate 100% PPV (400 ppm for upper plot, 800 ppm for lower plot) but different recall for the BCL2 variants. At 400 ppm, 50% of variants at 0.33% and 80% of variants at 0.8% are detected while at the more conservative setting of 800 ppm, 50% of variants at approximately 0.5% and 80% of variants at approximately 0.9% are detected. VarScan2 thresholds of 0.5% and 0.75% provide matching overall performance, shown as vertical green dashed lines. Note how DDiMAP has increased sensitivity of low frequency variants with a reduced sensitivity at higher frequencies compared to theVarScan2 single hard threshold.