Table 2. Performance on the VEST-indel test set for frameshift variations.

Method	MCCb	Sensitivity	Specificity	F-scorec	False positive rate	False discovery rate
Evaluation on the consensus subseta
ENTPRISE-X	0.626 (0.749)	0.943	0.916	0.620 (0.767)	8.4%	54%
VEST-indel	0.440 (0.585)	0.914	0.814	0.421 (0.615)	18.6%	73%
DDIG-in	0.321 (0.441)	0.943	0.663	0.297 (0.439)	33.7%	82%
Evaluation on the full test set
ENTPRISE-X	0.586	0.878	0.912	0.590	8.8%	55%
Baselined	0.323	0.988	0.621	0.294	37.9%	83%
Baselinee	0.224	0.598	0.775	0.271	22.5%	83%
ENTPRISE-X_1f	0.570	0.878	0.905	0.574	9.5%	57%
ENTPRISE-X_2f	0.555	0.854	0.905	0.562	9.5%	58%
ENTPRISE-X_10altf	0.587±0.006	0.887±0.006	0.910±0.003	0.590±0.006	9.0%±0.3%	55.8%±0.7%
ENTPRISE-X-nolocal	0.481	0.707	0.914	0.509	8.6%	60%
ENTPRISE-X-nonew	0.099	0.793	0.390	0.168	61.0%	90%
ENTPRISE-X-noratio	0.513	0.890	0.871	0.509	12.9%	64%
ENTPRISE-X-noessential	0.574	0.890	0.903	0.575	9.7%	58%
ENTPRISE-X-nopathogen	0.543	0.866	0.896	0.546	10.4%	60%
ENTPRISE-X-nodisease	0.368	0.683	0.859	0.396	14.1%	72%
ENTPRISE-X-nointeract	0.586	0.890	0.909	0.588	9.1%	56%

^a To be fair to all methods, only the consensus mutations of three methods are evaluated in comparison to the other methods.

^b Matthew’s Correlation Coefficient. The numbers in parentheses are the maximal possible values.

^c 2(precision×recall)/(precision+recall), where precision = (true positive)/(true positive + false positive), recall = (true positive)/(true positive + false negative). Numbers in parentheses are the maximal possible values.

^d When only the feature representing if the gene is disease-associated or not is used.

^e When only the feature representing if the gene is essential or not is used.

^f When using one of the 2 models trained on each half of the pathogenic data and training ENTPRISE-X for 10 different random partitions of the pathogenic part of the training set were used.