Skip to main content
. 2018 Jun 27;34(13):i429–i437. doi: 10.1093/bioinformatics/bty244

Table 5.

Coefficient of determination (R2) of COSSMO models and MaxEntScan

All events
Alternatively spliced
Model Acceptor n = 172 940 Donor n = 175 000 Acceptor n = 22 564 Donor n = 25 865
MaxEntScan 0.080 0.195 −0.105 0.038
Independent 0.389 (±0.003) 0.438 (±0.005) 0.187 (±0.003) 0.208 (±0.009)
Comm-net 0.421 (±0.007) 0.552 (±0.001) 0.229 (±0.003) 0.372 (±0.008)
LSTM 0.614 (±0.005) 0.628 (±0.003) 0.430 (±0.006) 0.453 (±0.005)
Resnet + LSTM 0.612 (±0.006) 0.618 (±0.004) 0.420 (±0.007) 0.448 (±0.004)

Note: The coefficient of determination is computed using the class prior 1/K (uniform distribution over alternative splice sites) as baseline. For COSSMO, R2 is computed as the average of five cross-validation folds with standard errors given in parentheses. The right columns show loss only for those events that contain two or more annotated splice sites.