High predictivity (r² = 0.88, s = 3.56 kJ/mol, q² = 0.84, s_press = 5.18 kJ/mol) is achieved when tested on the training dataset of 6 Cw*0401 peptide sequences. The Cw*0401 predictive model outperforms the predictive models done by Rognan et al. 16 on training datasets of 5 A*0204 (r² = 0.85, s_press = 2.40 kJ/mol) and 37 2K^k (r² = 0.78, s_press = 3.16 kJ/mol) peptide sequences and is comparable with our previous DRB1*0402 (r² = 0.90, s = 1.20 kJ/mol, q² = 0.82, s_press = 1.61 kJ/mol) and DQB1*0503 (r² = 0.95, s = 1.20 kJ/mol, q² = 0.75, s_press = 2.15 kJ/mol) prediction models on a training set of 8 peptides 17. The cross-validation coefficient q² and the standard error of prediction s_press are stable, with q² = 0.84 and s_press = 5.18 kJ/mol. This iterative regression procedure validates the internal consistency of the scoring function in the current model, rendering it suitable for predictions on the test dataset obtained from biochemical studies. The predictive performance of our model is further validated using the test dataset of 58 peptides. The external validation results indicate that our Cw*0401 predictive model is suitable for discriminating binding ligands from the background with high accuracy (A_ROC = 0.93) with sensitivity of 76% (SP = 0.95).

An in-depth analysis was performed to investigate the characteristics of Cw*0401 binding peptides. A panel of 2279 sequences was generated using an overlapping sliding window of size 9 across the entire p24^gag 5 and gp160^gag 18 glycoproteins and modeled into the binding groove of Cw*0401. From these sequences, a total of 877 binding sequences (predicted binders; SE = 76%, SP = 80%) were selected and a systematic analysis was performed to analyze the number of occurrence of individual amino acid residues and physico-chemical properties 19 at each position of Cw*0401 binding peptides.

Peptide position p2 is characterized by alanine (10%), glycine (13%), leucine (9%), serine (9%). 60% of the predicted residues at this position are hydrophobic in nature, while 93% are neutral. These properties correlate with the physico-chemical properties of existing binding motif (Tyr/Phe) at p2 11 as well as with the observed conservation in the test data (Table 1: Phe – 58% and Tyr – 26%). The p3 position shows a strong preference for glycine (11%) and threonine (8%). Existing anchor residues at this position are aspartic acid and histidine, which accounts for 9% of the total position-specific composition in the dataset. The p4 position shares similar characteristics as p3 (glycine: 9%; threonine: 8%), with additional preference for leucine (8%). Similar results were obtained at the p5 position (alanine: 8%; glycine: 9%; leucine: 8%). At p6, characteristic residues include glycine (9%) and leucine (8%). This position is favored by neutral (81%) acyclic (89%), medium/large (77%), and hydrophobic (51%) residues. The physico-chemical properties of these residues are in agreement with Val/Ile/Leu as reported in earlier studies 11 and is comparable with the conservation in the test dataset reported in Table 1 (Val – 29%, Ile – 12%, Leu and Pro – 10%). Finally, the p9 position was defined by six amino acids, including alanine (8%), glycine (10%), isoleucine (8%), leucine (8%), threonine (9%), and valine (8%). This position is primarily dominated by neutral (90%) and hydrophobic (58%) residues, and agrees with profiles of Leu/Phe as previously reported 11 and is consistent with the conservation in the test dataset reported in Table 1 (Leu and Phe – 39%,). Collectively, our data indicates that individual binding pockets may not be highly specific as previously reported 11 but can rather accommodate a wide-range of anchor residues with common physico-chemical properties. We attribute the discrepancies to two possibilities: i) the lack of extensive research on Cw*0401 peptides; and ii) natural peptides carrying these residues may be present in small amount and were thus not detected by experimental studies.

The potential T-cell epitope repertoire (data not shown) and immunological hot spots (Figure 1) for HIV-1 p24^gag and gp160^gag are well distributed throughout both glycoproteins. The known p24^gag epitope (p24 168–175) and gp160^gag epitope (gp160 375–383) were successfully predicted by our model at the threshold of -200 (SE = 76%, SP = 80%) 520. At this threshold, the number of predicted hot spots for p24^gag are fourteen (p24 20–75, 89–148, 143–262, 311–381, 447–506, 526–590, 664–727, 708–791, 873–930, 996–1053, 1054–1157, 1157–1316, 1308–1368, and 1378–1427), with an estimated FN = 3, and FP = 3 for SE = 0.76 and SP = 0.95 (Figure 1 and Table 2). For gp160^gag we predict nine hot spots (gp160 44–89, 102–171, 174–289, 341–504, 489–559, 581–639, 667–721, 708–766, and 793–852), with an estimated FN = 2, FP = 2 for SE = 0.76 and SP = 0.95 (Figure 1 and Table 3). The results presented here indicate that Cw*0401 can bind antigenic peptides with specificities comparable to HLA-A and -B molecules, and any variability in antigen expression may be directly related to the loss or reduced cell surface expression of the molecule by mechanisms as yet unknown 821.

The free energy scoring function was calibrated using 6 peptides with experimental IC₅₀ values and tested on a dataset 58 peptides (54 binders and 4 non-binders) obtained from biochemical studies (Table 1). The predictive performance of our model was assessed using sensitivity (SE), specificity (SP) and receiver operating characteristic (ROC) analysis 28. SE = TP/(TP+FN) and SP = TN/(TN+FP), indicate percentages of correctly predicted binders and non-binders, respectively. TP (true positives) represents correctly predicted experimental binders and TN (true negatives) for experimental non-binders incorrectly predicted as binders. FN (false negatives) denotes experimental binders predicted as non-binders and FP (false positives) stands for experimental non-binders predicted as binders. The accuracy of our predictions was assessed by the ROC analysis where the ROC curve is generated by plotting SE as a function of (1-SP) for a complete range of classification thresholds. The area under the ROC curve (A_ROC) provides a measure of overall prediction accuracy, A_ROC < 70% for poor, A_ROC > 80% for good and A_ROC > 90% for excellent predictions 15.