Cross-task interpretability through unified modeling reveals a universal shortcut bias in neoantigen prediction.

Accurate neoantigen prediction is central to the design of personalized cancer immunotherapy. The immune recognition of neoantigens is a multi-step process involving peptide-major histocompatibility complex (MHC) binding, antigen presentation, and T cell activation. Despite extensive computational efforts, the overall accuracy of neoantigen identification remains unsatisfactory, limiting translational potential. Here, we present ImmUni, a unified transformer-based framework that models the three immunological stages within a consistent architecture. Through this unified modeling, we discovered that deep-learning-based immunogenicity predictors have inadvertently learned shortcut correlations driven by intra-human leukocyte antigen (intra-HLA) label imbalance, revealing a previously unrecognized model-agnostic source of bias. We quantified this bias using an information-theoretic metric and proposed a mutual-information-guided debiasing strategy that mitigates shortcut learning and improves mutation-level generalization. ImmUni not only identifies the shortcut bias issue in current neoantigen prediction but also defines a general methodological framework for diagnosing and correcting bias across data-limited, multi-step tasks in computational biology scenarios.