Fig 1.
Flowchart of hybrid modeling approach.
Fig 2.
Effect of the rules and V3-loop:coreceptor interactions included in SVM models on prediction accuracy.
(A) The effect of the number of V3-loop:coreceptor interactions on accuracy. Accuracy is represented by the median AUC for 500 runs of 10-fold cross validation for both panels A and B. The accuracy at zero interactions is the accuracy based on only the rules. (B) Contribution of rules to accuracy when used in addition to the top 11 interactions Fig 3A. The naming scheme is as follows: Int Only–interactions only; Q–net charge; R – 11/24/25 rule; M–glycosylation motif; L–length. Dashed red line illustrates the accuracy when using all four rules and the top 11 interactions (QLMR, 0.977).
Table 1.
Summary of best coreceptor usage model.
Fig 3.
Diagrams of the top selected interactions for the cases of all rules + interactions and interactions only.
(A) Interaction map for the 11 interactions selected in combination with all rules. V3-loop is shown as an idealized loop with 35 amino acids where grey circles indicate positions for which no interactions were selected (inactive), while green circles indicate V3-loop positions with interactions selected (active). Red triangles represent residues of CCR5 and blue squares represent residues of CXCR4, with dashed lines representing interactions with V3-loop residues. Ordered lists of observed amino acids (based on occurrence with a minimum of 5%) in one-letter code for each active V3-loop residue are provided. Observed amino acids for CCR5 tropic sequences are in red and those observed for CXCR4 tropic sequences in blue. Bolded letters in the ordered list of observed amino acids indicate an amino acid that is observed in at least 50% of sequences at a given position. (B) Interaction map for the 18 interactions selected without rules. Color scheme and layout is the same as in (A). Faded triangles/squares indicate interactions that were also selected when including all rules. The crossed out interaction was selected when including rules, but not when using interactions only.
Table 2.
Comparison of methods on test set (1161 sequences).
Table 3.
Method accuracy comparison on unique patient test set (373 sequences).
Table 4.
Comparison of HIV-1 subtype accuracies for test set.
Table 5.
Comparison of HIV-1 subtype accuracies for unique patient test set.