Structural Basis of TCR Recognition of NRAS Cancer Neoantigens and AlphaFold Modeling Assessment

Original Authors: Wu D, et al.
Published in: Science Advances, 2024

Abstract

T cell receptor (TCR) recognition of cancer neoantigens is central to anti-tumor immune responses and adoptive cell therapy. In their 2024 Science Advances study, Wu et al. determined high-resolution crystal structures of human TCR N17.1.2 in complex with NRAS Q61K/Q61R neoantigen peptide-HLA-A1 complexes, revealing fine-specificity recognition mechanisms for mutant peptides. The team systematically evaluated AlphaFold2 (TCRmodel2 framework) and AlphaFold3 modeling capabilities for such immune recognition complexes, finding that increased sampling significantly improves prediction accuracy, though success rates remain limited for other NRAS Q61K-specific TCRs.

1. Background: Structural Requirements for Neoantigen-Targeted T Cell Therapy

Adoptive cell therapy (ACT) using tumor-specific T cells has achieved durable remissions in various metastatic cancers, including melanoma, cervical cancer, breast cancer, cholangiocarcinoma, and colorectal cancer. Therapeutic effects are primarily mediated by cytotoxic CD8+ T cells recognizing neoantigens generated by somatic mutations during malignant transformation.

Neoantigens derived from oncogenic driver mutations have special therapeutic value—these mutations are both tumor-specific and essential for cancer cell fitness and proliferation. The RAS protein family (KRAS, NRAS, HRAS) represents the most commonly mutated oncogenes in human cancers. NRAS Q61 mutations account for approximately 20% of melanomas, with Q61K and Q61R being the most common forms, associated with more aggressive clinical features and poorer prognosis.

2. Specific Recognition of Neoantigens by TCR N17.1.2

2.1 Binding Affinity and Specificity

Surface plasmon resonance (SPR) experiments showed that TCR N17.1.2 binding affinities for both mutant neoantigens fall within the typical range for TCR recognition:

2.2 Complex Crystal Structure Overview

The team determined structures of N17.1.2–NRAS Q61K–HLA-A1 (2.10 Å) and N17.1.2–NRAS Q61R–HLA-A1 (2.26 Å) complexes. The Cα RMSD between the two complexes is only 0.3 Å, indicating highly similar binding modes.

• Classical diagonal orientation: Vα above HLA-A1 α2 helix, Vβ above α1 helix
• TCR-pMHC crossing angle: 28° (identical in both complexes)
• Incident angle: 5° for Q61K complex, 7° for Q61R complex

2.3 TCR-MHC Interactions

N17.1.2-MHC interactions show significant Vα dominance. Vα contributes 71% of MHC contacts, with the germline-encoded CDR1α loop alone contributing 36%—more than any other CDR.

2.4 Recognition Mechanism for Neoantigen Mutation Sites

N17.1.2 buries 70% (273 Ų) of the peptide's solvent-accessible surface. Its specificity mechanism involves:

• Minimizing interactions with peptide N-terminus and central region: These regions are identical in mutant and wild-type peptides
• Focusing on C-terminal P7 Lys/Arg mutation site: The protruding side chain of this residue is the most solvent-exposed epitope feature
• Key salt bridge: CDR3β Glu103β forms a salt bridge with P7 Lys/Arg that wild-type P7 Gln cannot replicate
• Vβ-dominated peptide recognition: In the N17.1.2–NRAS Q61K complex, Vβ contributes 76% of peptide contacts

2.5 Conformational Changes During Binding

Comparing TCR bound state with unbound structure, the six CDR loops show backbone RMSD of 0.53 Å and all-atom RMSD of 0.62 Å, with no significant structural differences. Thus, N17.1.2 recognition follows conformational selection rather than induced fit.

3. AlphaFold Modeling Assessment for TCR-pMHC Complexes

3.1 Modeling Strategy and Evaluation Metrics

The team systematically compared three AlphaFold implementations:

• AlphaFold2.3 (default): 25 models per complex
• TCRmodel2: TCR-pMHC-specific framework based on AlphaFold2.3, default 5 models; extended sampling generates 1000 models
• AlphaFold3: 5 models via web server

3.2 N17.1.2 Complex Modeling Results

3.3 Model Confidence Threshold Establishment

Based on benchmark testing with 20 TCR-pMHC complexes, the team established confidence thresholds for model selection:

• Combined threshold: Model confidence > 0.875 AND I-pLDDT > 87.5

3.4 Modeling Test for Non-binding Complexes

Using TCRmodel2 (1000 models) for N17.1.2 with wild-type NRAS–HLA-A1 (no experimentally detected binding), the top-ranked model showed high similarity to mutant neoantigen complexes (I-RMSD 0.81 Å), with confidence scores (0.91 model confidence, 91.4 I-pLDDT) slightly lower than neoantigen complex models. This indicates AlphaFold cannot distinguish true binding pairs from non-binding pairs, even when they differ by only a single point mutation.

3.5 Prediction for Other NRAS Q61K-Specific TCRs

The team modeled four additional TCRs known to target NRAS Q61K–HLA-A1. These TCRs differ from N17.1.2 in germline genes and CDR3 sequences. Results showed:

• Using TCRmodel2 (1000 models), AlphaFold2.3 (200 models), and AlphaFold3 (5 models) all failed to produce sufficiently high confidence scores
• Did not reach preset thresholds for model confidence (>0.875) and I-pLDDT (>87.5)

4. Discussion and Limitations

4.1 Structural Principles of Neoantigen Recognition

This study reveals one of two main mechanisms for neoantigen recognition: TCR distinguishes mutant from wild-type peptides by directly forming key interactions with mutation residues. For NRAS Q61K/R, the P7 Lys/Arg mutation site side chain protrudes from the peptide-MHC surface, providing a clear recognition target for TCR. N17.1.2 achieves high specificity by minimizing contacts with conserved peptide regions and focusing on the mutation site.

4.2 AlphaFold Modeling Achievements and Boundaries

Achievements:

• TCRmodel2 with extended sampling (1000 models) achieves CAPRI "High" quality, I-RMSD ~0.9 Å
• AlphaFold3 achieves "Medium" quality with default settings, better than AlphaFold2.3 default sampling
• Model confidence scores and I-pLDDT serve as reliable accuracy indicators

Boundaries and Limitations:

• Sampling dependence: Prediction accuracy is generally low with default sampling (5-25 models); successful modeling requires extensive sampling (1000 models)
• Cannot distinguish binding from non-binding: Still produces high-confidence models for wild-type peptide complexes with no experimental binding
• Limited generalization: Low prediction success rates for other NRAS Q61K-specific TCRs with different sequences

4.3 Implications for Large-Scale T Cell Specificity Prediction

This study provides cautious assessment for applying AlphaFold to large-scale TCR-pMHC specificity prediction:

• Current sampling strategies (especially AlphaFold3 web server's 5-model limitation) may be insufficient for reliable prediction success rates
• Stricter confidence thresholds (model confidence > 0.875 AND I-pLDDT > 87.5) are needed to filter reliable models
• For unknown TCR-pMHC complexes, risks of both false negatives (accurate models with low confidence) and false positives (high confidence but inaccurate) coexist

5. Conclusion

Wu et al.'s study provides important insights into TCR recognition of NRAS cancer neoantigens from both structural biology and computational modeling perspectives. The crystal structure reveals N17.1.2's division of labor—Vα-dominated MHC recognition and Vβ-dominated peptide recognition—and precise targeting mechanism for the P7 mutation site. AlphaFold modeling assessment shows that deep learning models can accurately predict TCR-pMHC complex structures with sufficient sampling, but default sampling settings have limited reliability and currently cannot distinguish binding from non-binding complexes. These findings provide a valuable reference framework for optimizing TCR design and improving computational prediction methods.

References

Wu, D. et al. Structural characterization and AlphaFold modeling of human T cell receptor recognition of NRAS cancer neoantigens. Sci. Adv. 10, eadq6150 (2024).

This article is based on Wu et al.'s 2024 study published in Science Advances, aiming to provide objective technical interpretation for professional practitioners.