Table 2.
Tests on verified cancer drivers from non-coding regions show that CScape-somatic predicts all variants correctly, while the original CScape correctly predicts all but one SDHD variant
| Mutation | CSS | CS | FSa | CADD |
|---|---|---|---|---|
| TERT | ||||
| 5:g1295228G>A | + (0.56) | + (0.52) | + (1.33) | + (0.34) |
| 5:g1295229G>A | + (0.51) | + (0.62) | + (1.69) | + (0.66) |
| 5:g1295250G>A | + (0.51) | + (0.58) | + (0.56) | + (0.31) |
| SDHD | ||||
| 11:g111957523C>T | + (0.52) | + (0.81) | + (1.00) | + (1.64) |
| 11:g111957541C>T | + (0.68) | + (0.67) | + (1.62) | + (0.82) |
| 11:g111957544C>T | + (0.87) | − (0.40) | + (1.00) | + (0.64) |
| PLEKHS1 | ||||
| 10:g115511590G>A | + (0.71) | + (0.65) | − (0.17) | − (-0.10) |
| 10:g115511593C>T | + (0.57) | + (0.71) | − (0.17) | − (-0.06) |
Note: FunSeq2 and CADD predict the TERT and SDHD examples correctly, but both misclassify the PLEKHS1 examples. For each method, we present the predicted label (+ = driver, – = passenger) with the associated score in parentheses. (Classifiers: CSS = CScape-somatic, CS = CScape, FS = FunSeq2.)
a
For FunSeq2, we use a threshold of 0.56 (Rogers et al., 2017a).