Vertebrate gene finding from multiple-species alignments using a two-level strategy

Table 1 Prediction accuracies for vertical and horizontal components

	Acceptors	Donors	Starts	Stops
Train set size	204,021	221,421	7,571	25,071
Eval set size	52,605	57,179	1,805	6,162
%True sites	14.05	13.01	16.68	8.08
F scores (%)
Presence	52.72	48.77	39.70	34.64
Vertical	82.01	81.00	55.70	49.25
Horizontal	84.36	84.43	57.01	48.22
Both	84.86	84.60	58.22	49.60
ENCODE Cl	63.18	65.86	27.44	14.67
ENCODE GF	80.23	81.38	42.47	50.49
100-ROC (%)
Presence	12.41	12.66	20.62	23.98
Vertical	2.46	2.52	14.49	12.76
Horizontal	1.81	1.58	12.48	11.77
Both	1.74	1.54	10.41	10.90
ENCODE Cl	0.99	0.61	9.14	10.49

The table shows the F score (geometric mean of sensitivity and specificity, which are close to each other) for various classifier components. The test set for the presence, vertical, horizontal and 'both' conditions is 'challenging' data; we show results for a mixture of the classifiers trained on challenging and randomly selected data. The 'ENCODE Cl' and 'ENCODE GF' lines are for the 31 ENCODE test regions, using classifier scores and gene-finder scores, respectively. The table also shows the 100%-ROC (receiver operating characteristic) error value for each condition. This error value is the probability that if a true instance and a decoy are selected at random, the classifier will give the decoy a higher score than the true instance.

ISSN: 1474-760X