Table 1.
Neural networks trained to recognize promoters of the 17-base-spacing class
| Network | Index | Index coding/levels | Prototype | No. test promoters >cutoff | Sequence† |
|---|---|---|---|---|---|
| NW1 (14.5) | RI ∗ BvH | Fine/11, 4 | 41 promoters 1× permuted (5,412 seq) | > 0.93 (28/34) | pBR >0.93 1, 4,135 pBR(c) >0.93 125, 1,2264,278 |
| NW2 (26) | RI ∗ BvH | Coarse/6, 4 | 41 promoters (41 seq) | >0.985 (28/34) | pBR >0.985 1, 4,135 pBR(c) >0.985 21, 125, 1226, 4278 |
| NW3 (21) | RI ∗ BvH | Coarse/6, 4 | 41 promoters (41 seq) | >0.96 (29/34) | pBR >0.96 1, 4135 pBR(c) >0.96 21, 125, 1226, 4278 |
| NW4 (192) | RI− ∗ (RI+ − RI−) | Fine/10, 0 | 41 promoters (41 seq) | >0.80 (29/34) | pBR >0.80 1, 4135 pBR(c) >0.80 125, 532, 1056, 1226, 4278 |
| NW5 (41.4) | RI− ∗ (RI+ − RI−) | Coarse/6,4 | 41 promoters (41 seq) | >0.77 (30/34) | pBR >0.77 1, 170, 3654, 4135 pBR(c) >0.77 125, 1056, 1226, 4278 |
| Average result | 29/34 | 1, 4135 | |||
| 85% TP | 125, 1226, 4278, 1FP | ||||
Numbers listed after the network in parentheses are the training iterations in thousands. NW3 had 15 middle-layer neurons; all others had 8. Learning rates and momentum values of the following networks: NW1, 0.5 and 0.2; NW2, 0.9 and 0.4; NW3, 0.6 and 0.4; NW4, 0.15 and 0.1; NW5, 0.6 and 0.4. Input neurons varied from 348 to 352. NW3 was tested on the first and last 16.4 kb of the λ r strand (the B region was omitted to avoid its many uncharted promoter sites) with performance virtually identical to that above for pBR322, 100% true positives (TP) and 0.024% false positives (FP). Underlined coordinants indicate known promoters.
pBR is the clockwise sequence of pBR322, beginning at position −5 to capture the tet promoter; pBR322(c) is the counterclockwise sequence beginning at the EcoRI site. The cutoff values for the minimum network output designating a promoter are listed for each network.