Parameterization of disorder predictors for large-scale applications requiring high specificity by using an extended benchmark dataset

Table 3 Performance benchmark with the SL dataset under parameters at comparable high specificity level (~0.950).

Method	threshold	sensitivity	specificity	MCC	PE
DISOPRED2	0.08	0.557	0.947	0.559	0.504
IUPred long	0.54	0.544	0.948	0.550	0.492
IUPred short	0.51	0.491	0.948	0.507	0.440
CAST	40	0.448	0.951	0.474	0.399
DisEMBL Rem465	1	0.348	0.969	0.418	0.317
SEG45	3.30;3.60	0.368	0.950	0.402	0.318
SEG25	2.94;3.24	0.335	0.946	0.364	0.281
SEG12	2.29;2.59	0.268	0.950	0.305	0.218
DisEMBL Hotloops	2.7	0.259	0.949	0.295	0.208
DisEMBL Coils	1.94	0.251	0.948	0.286	0.200

Predictors were run with parameters tuned to achieve a comparable specificity level close to 0.950, which corresponds to ~5% of false positive predictions. For DisEMBL Remark 465 (in italic), parameter tuning only allowed a specificity of 0.969 as closest value to our criterion. Ranking is based on the Matthews Correlation Coefficient (MCC) but remains essentially unchanged for other performance measures such as probability excess (PE).

ISSN: 1471-2164