HLA-B*57:01 Peptidome Modeling Predicts Abacavir Risk

Key Takeaways

• A peptidome-wide HLA-B*57:01 analysis spanning 13,706 presented peptides was associated with 88% sensitivity and 100% specificity for distinguishing 15 abacavir analogues in leave-one-out cross-validation using a 366-peptide subset.
• With a conventional approach, fewer than 3% of modeled HLA-B*57:01-peptide complexes placed abacavir in the native cleft, whereas the alternative tripartite workflow reproduced crystallized HLA-B*57:01-abacavir-peptide complexes at RMSD≤2.2 Å.
• Across the modeled repertoire, 63.9% of peptides (8,753/13,706) were predicted to remain bound in the presence of abacavir, with lysine enriched at P1 and tryptophan at P9 and proline at P4 decreased; the paper presents the framework as proof-of-concept for altered-peptide-repertoire mechanisms rather than clinical validation.

In leave-one-out cross-validation across 15 abacavir analogues, a reduced 366-peptide panel yielded 88% sensitivity and 100% specificity for classifying compounds by in vitro T-cell activation profile. The work was a full computational modeling study of abacavir-associated HLA-B*57:01 hypersensitivity that linked compound-level prediction to peptidome-scale structural modeling across 13,706 HLA-B*57:01-presented peptides. The authors framed that design as a response to earlier computational efforts built around only a small number of crystallized tripartite structures and docking strategies that did not consistently recreate the native binding arrangement.

In Prediction of drug hypersensitivity by comprehensive modeling of HLA-peptidomes, the authors describe HLA-B*57:01 as having a positive predictive value of 47.9% and a negative predictive value of 100% for abacavir hypersensitivity syndrome, a context that they say has led the FDA and EMA to mandate HLA-B*57:01 screening before initiating or reinitiating abacavir treatment. They also framed the mechanism around abacavir binding in the HLA-B*57:01 F pocket through interactions with Asp114 and Ser116. In that setting, the paper cited a shift in the native peptide repertoire of about 20%–25%, linking altered presentation rather than direct covalent binding to the hypersensitivity signal.

For structural benchmarking, TFold modeled 73 previously unseen HLA-peptide crystal structures at RMSD<2 Å before the investigators extended the workflow to 13,706 HLA-B*57:01-presented peptides. When abacavir was introduced by conventional redocking, fewer than 3% of full-peptidome HLA-B*57:01-peptide complexes placed the drug in the binding cleft in a way consistent with crystallographic data. The selected Chai-plus-ADCP tripartite pipeline instead reproduced crystallized HLA-B*57:01-abacavir-peptide complexes at RMSD≤2.2 Å, establishing the structural basis for the later repertoire analysis.

Across the modeled peptidome, 63.9% of peptides, or 8,753 of 13,706, were predicted to remain bound in the presence of abacavir, with lysine enriched at P1 and tryptophan at P9 and proline at P4 decreased; among peptides containing lysine at P1 but not P4, at P4 but not P1, or at both P1 and P4, reported binding rates were 79.4%, 76.4%, and 96.2%. The authors then used a reduced 366-peptide subset to train an SVM that, in leave-one-out cross-validation across 15 abacavir analogues, achieved 88% sensitivity and 100% specificity while reducing compute demands from about 7,300 CPU core hours to about 200 core hours per candidate compound. They noted that the analysis did not capture de novo peptides presented after drug binding, that the relatively small compound set was evaluated by cross-validation rather than additional independent external datasets, and that the framework was developed for altered-peptide-repertoire mechanisms rather than hapten-mediated or p-i mechanisms.

Clinician Questions

How accurately did the HLA-B*57:01 tripartite modeling pipeline reproduce known abacavir-peptide structures?

In this computational study, the HLA-B*57:01-abacavir-peptide pipeline reproduced crystallized tripartite complexes at RMSD≤2.2 Å across a modeling program that covered 13,706 presented peptides. For broader structural context, TFold also modeled 73 previously unseen HLA-peptide crystal structures at RMSD<2 Å.

What peptide repertoire changes were reported after abacavir binding to HLA-B*57:01?

Among 13,706 modeled HLA-B*57:01-presented peptides, 63.9%, or 8,753 peptides, were predicted to remain bound in the presence of abacavir. The reported repertoire shift included enrichment of lysine at P1 and decreases of tryptophan at P9 and proline at P4; among peptides containing lysine at P1 but not P4, at P4 but not P1, or at both P1 and P4, the reported binding rates were 79.4%, 76.4%, and 96.2%.

How did the 366-peptide SVM perform for predicting immunogenic abacavir analogues?

In leave-one-out cross-validation across 15 abacavir analogues with positive or negative in vitro T-cell activation profiles, an SVM trained on binding energies from a 366-peptide subset achieved 88% sensitivity and 100% specificity. The reduced peptide panel also lowered computational requirements from about 7,300 CPU core hours to about 200 core hours per candidate compound.