Experimentally anchored, MHC-agnostic prediction of immunogenicity and cytokine release
Antibody Therapeutics · 2026-10-10
DOI: 10.1093/abt/tbag050 ↗Abstract
Abstract Background Accurate prediction of anti-drug antibody responses remains a challenge in biologics development, as current tools primarily rely on major histocompatibility complex binding affinity, which does not capture the full complexity of immune activation. Methods We developed EPITOP™, a protein language model trained on epitopes experimentally verified via cytokine release assays, enabling direct immunogenicity prediction. This was combined with IMMUNOTOP™, which aggregates EPITOP™ outputs into protein-level scores. Results EPITOP™ achieves an Area Under the Curve (AUC) of 0.73 on the independent test set, demonstrating performance comparable to existing methods while providing biologically grounded and interpretable outputs for immunogenicity risk assessment. It also enables identification of immunogenic hotspots. Conclusions Beyond hotspot identification, the model can serve as a surrogate for peptide-level interferon-γ-associated immunogenicity when experimental measurement is impractical and may be useful for modalities in which experimental cytokine testing is difficult, although systemic cytokine release depends on additional biological factors not modeled here.
EUVIMED