Mutual information and variants for protein domain-domain contact prediction

BMC Research Notes

Table 2 Precision for detecting contact versus non-contact residues at 20% recall, for sub-alignments of 70%

Contact prediction for sub-alignments
MI variant	70% AVG	70% STD DEV	100%
MIc	52.5*	2.1	54.8
MIp	46.0*	2.1	47.4
MIcRA	41.9*	1.8	41.4
MIpRA	38.2*	1.5	38.5
MIp3D	30.6	1.3	28.5
MIRA	30.2*	1.2	31.4
MIp3DRA	28.0*	1.2	30.9
MI	25.8*	0.8	27.6
MI3DRA	23.2*	1.0	25.5
Random	-	-	24.4
RandomRA	-	-	24.4
MI3D	20.0	0.6	21.8

70% of sequences in an MSA were randomly selected and the 10 MI variant scores based on the new sub-alignment were calculated. This subset selection and calculation procedure was repeated 100 times for those test cases that had ≥200 sequences to ensure ≥125 sequences in each sub-alignment[8]. Thus 24 test cases were used. For each of the 100 iterations a P-ROC curve similar to Figure2 was plotted for the 24 test cases (figures not shown), and the precision at 20% recall recorded. Columns one and two, respectively, contain the averages and standard deviations of these 100 precision values. Column three indicates the precision attained at 20% recall when all sequences in the 24 original MSAs were used. When considering this set of 24 cases, the probability of randomly selecting a contact residue from all surface residues is 24.4% generally and when using the reduced alphabet MSAs. The MI variants are listed in descending order of the average precision of the 100 70% sub- alignments. The presence of an ‘*’ at an MI variant indicates that the difference between the precision of this MI variant and the next lowest is significant at the 0.1% level, when using two-sample t-tests with a sample size of 24.

ISSN: 1756-0500