Probabilistic grammatical model for helix‐helix contact site classification

Table 6 Comparison of classifiers performance in 4‐fold CV in the WDG150NR dataset in terms of AUC ROC

Method/trained for	c1	c2	c3	c4	wavg	uavg
Our PCFG comb	0.56	0.67	0.64	0.57	0.60	0.61
Our PCFG best of 3	0.59	0.68	0.71	0.70	0.65	0.67
BLASTP concat	0.51	0.45	0.47	0.44	0.48	0.47
BLASTP 2‐parts	0.49	0.47	0.51	0.47	0.48	0.48
HMMER2 global concat	0.55	0.64	0.56	0.46	0.56	0.55
HMMER2 global 2‐parts	0.56	0.61	0.62	0.49	0.58	0.57
HMMER2 global divider	0.55	0.62	0.60	0.45	0.57	0.56
HMMER2 local concat	0.54	0.58	0.45	0.56	0.54	0.53
HMMER2 local 2‐parts	0.53	0.64	0.64	0.57	0.58	0.60
HMMER2 local divider	0.55	0.61	0.62	0.46	0.57	0.56
HMMER3 local concat	0.52	0.57	0.53	0.52	0.54	0.54
HMMER3 local 2‐parts	0.57	0.61	0.57	0.60	0.58	0.59
HMMER3 local 2‐parts best of 3	0.57	0.61	0.57	0.61	0.59	0.59
MSA closest neighbor	0.56	0.69	0.62	0.55	0.60	0.61

wavg and uavg are weighted and unweighted averages, respectively. In the case of our PCFGs, accessibility grammars are used for c1 and c4, and van der Waals grammars are used for c2 and c3. Column best results are shown in bold (PCFG’s and HMMER3’s best of 3 results are not considered).

ISSN: 1748-7188