A meta-learning approach for B-cell conformational epitope prediction

Hu, Yuh-Jyh; Lin, Shun-Chien; Lin, Yu-Lung; Lin, Kuan-Hui; You, Shun-Ning

doi:10.1186/s12859-014-0378-y

BMC Bioinformatics

Table 7 Results of individual comparisons

From: A meta-learning approach for B-cell conformational epitope prediction

Classifier	TPR	FPR	Precision	Accuracy	F-score	MCC	AUC
SEPPA 2.0	0.155	0.045	0.161	0.913	0.158	0.112	0.697
3-level Stacking^a w/o SEPPA 2.0	0.418	0.030	0.384	0.935	0.386	0.351	0.820
Cascade^a w/o SEPPA 2.0	0.404	0.033	0.405	0.937	0.404	0.371	0.820
DiscoTope 2.0	0.917	0.625	0.090	0.409	0.164	0.148	0.748
3-level Stacking^a w/o DiscoTop 2.0	0.231	0.013	0.541	0.939	0.324	0.327	0.809
Cascade^a w/o DiscoTope 2.0	0.212	0.013	0.532	0.939	0.303	0.310	0.806
Bpredictor	0.045	0.028	0.067	0.933	0.054	0.021	0.683
3-level Stacking^b w/o Bpredictor	0.119	0.006	0.471	0.957	0.190	0.222	0.779
Cascade^b w/o Bpredictor	0.149	0.002	0.769	0.962	0.250	0.328	0.787
ElliPro	0.421	0.279	0.131	0.694	0.199	0.090	0.630
3-level Stacking^c w/o ElliPro	0.367	0.009	0.802	0.935	0.504	0.516	0.861
Cascade^c w/o ElliPro	0.346	0.010	0.770	0.932	0.478	0.488	0.857
CBTOPE^d	0.801	0.424	0.118	0.591	0.205	0.188	0.798
3-level Stacking^d	0.446	0.010	0.751	0.954	0.558	0.557	0.913
Cascade^d	0.446	0.010	0.762	0.954	0.562	0.562	0.908

^aMeta classifiers were trained and tested using the data sets that were used specifically to train and test SEPPA 2.0 (or DiscoTope 2.0), excluding the antigens with missing feature values. All the classifiers, including the base predictor (SEPPA 2.0 or DiscoTope 2.0), were tested on the same test data to conduct a consistent comparison.
^bMeta classifiers were trained on the data used specifically to train Bpredictor, excluding the antigens with missing feature values. Though Bpredictor provided the test data set of its own, the data lacked the epitope residues annotated in the IEDB. Alternatively, we used the independent test data set of 15 antigens (Table 15) to test all the classifiers, including Bpredictor, to conduct a consistent comparison.
^cElliPro only provided the test data set, but no training data. Meta classifiers were consequently trained on the training data set of 94 antigens (Table 16), and tested on the test data of ElliPro, excluding the antigens with missing feature values. All the classifiers, including ElliPro, were tested on the same test data to conduct a consistent comparison.
^dMeta classifiers were trained and tested using the non-redundant (<40% sequence identity) benchmark dataset previously used to evaluate CBTOPE, excluding the antigens with missing feature values. Following CBTOPE, we adopted 5-fold CV to compare the performances. All the classifiers, including CBTOPE, were tested on the same test data to conduct a consistent comparison. The parameter value (-0.3) we used for CBTOPE was the same as used previously to evaluate CBTOPE in [24].

Back to article page

ISSN: 1471-2105

Contact us

General enquiries: journalsubmissions@springernature.com