AI-powered TCR-peptide-MHC interaction prediction using AlphaFold3 and deep learning for therapeutic TCR discovery, neoantigen validation, and T cell immunogenicity assessment.
The TCR-pMHC Prediction Agent predicts T-cell receptor interactions with peptide-MHC complexes using AlphaFold3-based structural modeling and deep learning. Accurate TCR-pMHC prediction enables therapeutic TCR discovery, neoantigen vaccine validation, and identification of immunogenic epitopes for cancer and infectious disease applications.
Binding Prediction: Predict TCR-pMHC binding affinity/probability.
Structural Modeling: Generate TCR-pMHC complex structures with AlphaFold3.
Epitope Specificity: Determine which epitopes a TCR recognizes.
: Predict off-target self-peptide binding.
Immunogenicity Scoring: Rank peptide immunogenicity.
Therapeutic TCR Screening: Screen TCRs for desired specificity.
| Approach | Method | Strengths |
|---|---|---|
| AlphaFold3 | Structure prediction | High accuracy, interpretable |
| TCR-BERT | Sequence transformer | Fast, large-scale |
| ERGO-II | RNN-based | Established benchmark |
| pMTnet | Multi-task learning | Generalizable |
| NetTCR | CNN-based | HLA-specific |
| TITAN | Attention-based | State-of-art sequence |
Input: TCR sequence (alpha/beta CDR3), peptide, HLA allele.
Structure Prediction: Generate pMHC and TCR structures.
Docking: Model TCR-pMHC complex.
Scoring: Calculate binding probability/affinity.
Cross-Reactivity: Screen against self-peptide database.
Validation Features: Extract structural determinants.
Output: Binding predictions, structures, safety assessment.
User: "Predict whether this tumor-reactive TCR binds the identified neoantigen and check for cross-reactivity with self-peptides."
Agent Action:
python3 Skills/Immunology_Vaccines/TCR_pMHC_Prediction_Agent/tcr_pmhc_predict.py \
--tcr_alpha_cdr3 CAVSDRGSTLGRLYF \
--tcr_beta_cdr3 CASSLGQAYEQYF \
--tcr_v_genes TRAV12-1,TRBV7-9 \
--peptide KRAS_G12D_VVGADGVGK \
--hla HLA-A*11:01 \
--check_cross_reactivity true \
--self_peptide_db human_proteome_9mers.fasta \
--method alphafold3 \
--output tcr_pmhc_results/
| Input | Format | Required |
|---|---|---|
| TCR CDR3 alpha | Amino acid sequence | Yes |
| TCR CDR3 beta | Amino acid sequence | Yes |
| V gene usage | IMGT notation | Recommended |
| Peptide | 8-11mer amino acids | Yes |
| HLA allele | 4-digit resolution | Yes |
| Output | Description | Format |
|---|---|---|
| Binding Score | Probability of binding | .json |
| Complex Structure | TCR-pMHC model | .pdb |
| Contact Map | Residue interactions | .csv, .png |
| Cross-Reactivity | Self-peptide hits | .csv |
| Confidence Score | Prediction reliability | .json |
| Binding Determinants | Key residues | .csv |
| Component | Application | Output |
|---|---|---|
| pMHC Modeling | Peptide-MHC structure | Complex structure |
| TCR Modeling | Variable region structure | TCR structure |
| Complex Prediction | Full ternary complex | Docked model |
| pLDDT Scores | Confidence per residue | Quality metric |
| PAE | Positional error | Interface confidence |
| Score Range | Interpretation | Action |
|---|---|---|
| >0.9 | Strong predicted binder | High confidence |
| 0.7-0.9 | Moderate predicted binder | Likely positive |
| 0.5-0.7 | Weak/uncertain | Experimental validation needed |
| <0.5 | Predicted non-binder | Low priority |
Structural Prediction:
Sequence Models:
Cross-Reactivity:
| Method | Dataset | AUC | Notes |
|---|---|---|---|
| AlphaFold3 | VDJdb benchmark | 0.85 | Structural |
| TCR-BERT | IEDB | 0.82 | Fast screening |
| ERGO-II | McPAS-TCR | 0.78 | Established |
| Ensemble | Combined | 0.88 | Best overall |
| Application | Use Case | TCR-pMHC Role |
|---|---|---|
| Neoantigen Vaccines | Validate immunogenicity | Predict T cell response |
| TCR-T Therapy | Select therapeutic TCRs | Screen candidates |
| Safety Assessment | Check cross-reactivity | Avoid autoimmunity |
| Epitope Discovery | Find immunogenic peptides | Prioritize targets |
| Database | Content | Purpose |
|---|---|---|
| Human Proteome | All self-peptides | Primary safety |
| Tissue-Specific | Expression-weighted | Toxicity prediction |
| Viral Mimicry | Viral homologs | Infection mimics |
| Cancer-Testis | CT antigens | On-target activity |
| Feature | Location | Significance |
|---|---|---|
| CDR3 beta apex | Peptide contact | Specificity |
| CDR3 alpha | MHC/peptide | Fine-tuning |
| CDR1/2 | MHC helices | HLA restriction |
| Germline-encoded | Framework | Base recognition |
| Limitation | Impact | Mitigation |
|---|---|---|
| Training Data Bias | Common HLA over-represented | Use diverse training |
| Novel TCRs | Out-of-distribution | Lower confidence |
| Post-translational | PTM peptides not modeled | Experimental validation |
| Dynamics | Static structures | MD simulation |
AI Group - Biomedical AI Platform