Predicting Neoantigen Immunogenicity Beyond Peptide MHC Binding Affinity

115 concepts76 papers1 proposal

To understand how the adaptive immune system detects threats, imagine a lock-and-key mechanism. For decades, computational immunology treated the problem of antigen discovery as a two-body matching game: if a mutated protein fragment (a peptide) fits securely into the slot of a cellular display molecule (the Major Histocompatibility Complex, or MHC), it was assumed the immune system would sound the alarm.

However, biology presents a frustrating paradox. As noted by Xiaoyun Yang et al. in DeepNetBim: deep learning model for predicting HLA-epitope interactions based on network analysis by harnessing binding and immunogenicity information and Yanqing Niu et al. in Accurate Prediction of Immunogenic T-Cell Epitopes from Epitope Sequences Using the Genetic Algorithm-Based Ensemble Learning, strong binding between a peptide and an MHC molecule is merely a prerequisite, not a guarantee, of an immune response. This decoupling of binding affinity from true immunogenicity—the ability to actually trigger a T-cell response—creates a significant computational-to-clinical gap that limits personalized cancer vaccines and immunotherapies (Jingcheng Wu et al. in TSNAdb v2.0: the updated version of tumor-specific neoantigen database).

Why does this gap exist? Bernhard Knapp et al. explored this at scale in Large Scale Characterization of the LC13 TCR and HLA-B8 Structural Landscape in Reaction to 172 Altered Peptide Ligands: A Molecular Dynamics Simulation Study, discovering through molecular dynamics that immunogenic and non-immunogenic peptides are often structurally and dynamically indistinguishable when bound to MHC alone. The missing link is the third actor in the triad: the T-cell receptor (TCR). To accurately predict immunogenicity, modern computational models must shift from a two-body binding problem to a three-body interaction landscape.

Resolving the Three-Body Interaction

Predicting how a TCR recognizes a peptide-MHC (pMHC) complex is an immense search problem due to the staggering diversity of the human immune repertoire (John-William Sidhom et al. in Author Correction: DeepTCR is a deep learning framework for revealing sequence concepts within T-cell repertoires). The challenge lies in mapping the specific amino acid sequences within the TCR's loops—particularly the Complementarity Determining Region 3 (CDR3)—to the exact epitopes they recognize (Shiva Dahal‐Koirala et al. in Comprehensive Analysis of CDR3 Sequences in Gluten-Specific T-Cell Receptors Reveals a Dominant R-Motif and Several New Minor Motifs).

Early computational attempts struggled because they evaluated only a fraction of this interface, often focusing exclusively on the TCR's $\beta$ chain sequence. However, the physical reality of T-cell recognition requires both the $\alpha$ and $\beta$ chains to wrap around the presented peptide.

   [ T-Cell Receptor (TCR) ]
     /                 \
[ Alpha Chain ]   [ Beta Chain ]
     \                 /
    [ Bound Peptide Epitope ]
               |
   [ MHC / HLA Display Molecule ]

Alessandro Montemurro et al. demonstrated in NetTCR-2.0 enables accurate prediction of TCR-peptide binding by using paired TCRα and β sequence data that incorporating paired $\alpha$ and $\beta$ chain sequence data drastically improves prediction accuracy. By capturing the full sequence context of the receptor, deep learning architectures can begin to learn the joint semantic vocabulary of TCR-pMHC engagement. Building on this sequence-level semantic analysis, frameworks like TPepRet (Wei Fan 0010 et al. in TPepRet: a deep learning model for characterizing T-cell receptors-antigen binding patterns) apply bidirectional sequence learning to extract the subtle molecular rules that govern whether a TCR will successfully bind a target peptide.

From Sequences to Surface Topography

While sequence-based deep learning models have advanced significantly, they hit a wall when encountering rare or highly polymorphic MHC variants. This is especially true for MHC class II molecules, which feature open binding clefts and highly variable peptide lengths that lack the rigid structural constraints of MHC class I (Jiarui Li et al. in TCR-EML: Explainable Model Layers for TCR-pMHC Prediction).

To overcome this, researchers are moving beyond text-like sequence representations toward geometric and structural deep learning. The core intuition is simple: a TCR does not read an amino acid sequence; it feels a physical, electrostatic surface topography.

Frameworks like NeoaPred (Dawei Jiang et al. in NeoaPred: a deep-learning framework for predicting immunogenic neoantigen based on surface and structural features of peptide–human leukocyte antigen complexes) explicitly model the 3D structural and surface differences between mutant neoantigens and their wild-type counterparts. By evaluating how a mutation physically alters the solvent-exposed surface of the pMHC complex, these models capture the precise spatial features that a TCR encounters.

This structural perspective provides a powerful inductive bias. As Dario F. Marzella et al. showed in Geometric deep learning improves generalizability of MHC-bound peptide predictions, structure-based geometric deep learning generalizes far better to rare, "unseen" MHC alleles than sequence-only models, achieving superior performance while requiring significantly less training data.

The Systemic View of Immunogenicity

Even a perfectly modeled structural interface cannot fully predict immunogenicity in isolation, because an antigen's capacity to trigger an immune response is fundamentally dependent on the host's immune ecosystem.

As argued by Qibin Leng et al. in Pre‐existing heterologous T‐cell immunity and neoantigen immunogenicity, an individual's history of microbial exposure shapes a unique "heterologous immunity." If a patient's pre-existing T-cell repertoire lacks the specific TCR architectures capable of recognizing a given neoantigen surface, the peptide will remain immunologically silent, regardless of how strongly it binds the MHC molecule.

To capture this systemic reality, the field is expanding toward multimodal frameworks. Models like TRIM (Chujun He et al. in Multimodal framework for the joint analysis of single-cell RNA and T cell receptor sequencing data predicts T cell response to cancer immunotherapy) integrate single-cell RNA sequencing with TCR sequencing to evaluate the actual functional state of the T cells alongside their sequence identity. By combining these systemic host insights with high-throughput functional validation platforms (William Chour et al. in Large libraries of single-chain trimer peptide-MHCs enable antigen-specific CD8+ T cell discovery and analysis), computational immunology is successfully moving past simple binding affinity to map the complete, multi-scale landscape of immune recognition.

Key Foundational Work

Shift-Invariant Adaptive Double Threading: Learning MHC II-Peptide Binding (2008) — contributed 2 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_neoantigen_immunogenicity_prediction_cd16ba84,
  title = {neoantigen immunogenicity prediction},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/neoantigen-immunogenicity-prediction/rdr_cd16ba8496984f458f9c81772ca4a7b6},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-22},
}

This exact version (version 1, frozen text):

@misc{lacuna_neoantigen_immunogenicity_prediction_cd16ba84_v1,
  title = {neoantigen immunogenicity prediction},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/neoantigen-immunogenicity-prediction/rdr_cd16ba8496984f458f9c81772ca4a7b6/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-22},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake