Table 8.
Model | M. smegmatis | L. phytofermentans | B. amyloliquefaciens | R. capsulatus | Mean AUPRC |
---|---|---|---|---|---|
RF-HOT | 0.955 | 0.626 | 0.608 | 0.691 | 0.720 ± 0.161 |
RF-TETRA | 0.800 | 0.608 | 0.843 | 0.678 | 0.732 ± 0.108 |
GRU-0 | 0.646 | 0.486 | 0.486 | 0.588 | 0.552 ± 0.079 |
GRU-1 | 0.622 | 0.490 | 0.500 | 0.576 | 0.547 ± 0.063 |
LSTM-3 | 0.625 | 0.499 | 0.494 | 0.559 | 0.544 ± 0.061 |
LSTM-4 | 0.623 | 0.501 | 0.505 | 0.573 | 0.550 ± 0.059 |
MULTiPly | 0.649 | 0.474 | 0.653 | 0.591 | 0.592 ± 0.083 |
iPro70-FMWin | 0.652 | 0.582 | 0.774 | 0.594 | 0.65 ± 0.088 |
bTSSFinder | (0.512, 0.272) | (0.507, 0.944) | (0, 0) | (0.513, 0.250) | NA |
G4PromFinder | (0.506, 0.938) | (0.448, 0.216) | (0.382, 0.339) | (0.510, 0.960) | NA |
BProm | (0.781, 0.006) | (0.501, 0.560) | (0.701, 0.421) | (0.615, 0.011) | NA |
AUPRC is roughly the weighted average precision across all recall levels. A perfect classifier has an AUPRC of 1, while a random classifier has an AUPRC of 0.5 in a balanced data set. These results were obtained in balanced data sets (i.e., with a 1:1 ratio of positive to negative instances). The numbers in bold indicate the model with the highest AUPRC. For BPROM, bTSSFinder, and G4PromFinder, the numbers between brackets indicate precision and recall achieved as these tools did not provide a probability associated to each instance in the data set