TAPAS combines AlphaFold3 confidence metrics, structural geometry, and ESM-2 sequence embeddings via TabPFN to predict TCR-pMHC binding more robustly than any single zero-shot metric.
Adapted from @BiologyAIDailyTAPAS: Learned integration of AlphaFold3 confidence and geometric features for TCR-pMHC binding prediction 1. TAPAS addresses a practical issue in structure-based TCR-pMHC prediction: individual AlphaFold3 “zero-shot” confidence metrics (ipTM, pDockQ2, iPAE-derived confidence, pLDDT) are informative but none is consistently best across datasets, making metric selection unreliable for new benchmarks. 2. The method reframes TCR-pMHC binding prediction as tabular classification and uses TabPFN (a transformer foundation model for tabular data) to integrate complementary signals without task-specific gradient training or hyperparameter tuning, aiming for robust performance across evaluation settings. 3. TAPAS features combine three groups: (a) AlphaFold3 interface confidence (4 features: iPAE confidence and pDockQ2, each computed for TCR-side and pMHC-side interfaces), (b) structural geometry (11 features: peptide–CDR3 contact counts/fractions, “confident contacts” with pLDDT ≥ 70 and PAE ≤ 10 Å, centroid distances, and TCR orientation relative to pMHC), and (c) sequence embeddings (288 PCA-reduced ESM-2 features from peptide + six CDR loops). 4. Data design emphasizes generalization: VDJdb human paired αβ / HLA-I filtered to 2,149 pairs across 207 peptides, evaluated with 5-fold cross-validation under (i) random split and (ii) strict split where test peptides are unseen during training; negatives are generated by peptide–TCR re-pairing with peptide dissimilarity constraints (Levenshtein distance > 3) to reduce negative sampling bias. 5. On VDJdb random split, TAPAS reaches macro-AUC@0.1 = 0.8985, outperforming the best zero-shot metric (TCR-pMHC ipTM at 0.8393) and sequence baselines (TSpred 0.8311, NetTCR-2.2 0.8174); the paired improvement vs best zero-shot is +0.0592 with peptide-level bootstrap CI excluding 0 (p<0.001). 6. On VDJdb strict split (unseen peptides), TAPAS achieves macro-AUC@0.1 = 0.8425, essentially matching the best zero-shot metric (iPAE confidence 0.8371). Sequence-only models drop sharply (both ~0.717), while AF3-derived metrics remain comparatively robust—highlighting that structural confidence carries transferable signal across peptides. 7. External benchmark 1 (ePytope-TCR viral epitopes; after removing overlaps with VDJdb: 445 TCRs, 8 pMHCs, 3,560 pairs) shows TAPAS ensemble macro-AUC@0.1 = 0.5618, numerically similar to the strongest zero-shot metric (ipTM 0.5585), while sequence baselines are near-random (~0.50). 8. External benchmark 2 (IMMREP25 unseen viral peptides; 20 peptides, 10,000 pairs) shows TAPAS ensemble macro-AUC@0.1 = 0.5847, exceeding the best zero-shot metric (pDockQ2 0.5740) with a small but statistically supported gain (Δ=+0.0107; p=0.047). TAPAS also remains top after applying IMMREP25-style normalization + TCRdist cluster smoothing (macro-AUC@0.1 increases to 0.6020). 9. Ablations clarify when each signal matters: on random split, ESM-2-only is strongest single group (0.8634), but on strict split ESM-2-only collapses (0.7246) while confidence-only (0.7755) and geometry-only (0.7736) are more stable; combining all three yields best overall (0.8425 strict), suggesting integration is particularly helpful under peptide distribution shift. 10. Interpretation and cost: SHAP on a structure-only submodel indicates pDockQ2 interface scores dominate structural importance (pMHC-side pDockQ2 ranked #1 across settings), with iPAE next and geometry smaller/variable. Compute profiling shows AF3 inference dominates (~780 s/pair) vs feature extraction (~12 s/pair) and TabPFN ensemble inference (~151 s total), making structure generation the main bottleneck. 💻Code: https://t.co/gVwqt2Aq53 📜Paper: https://t.co/Ti2ZPgU6Ec #TCR #Immunology #ComputationalBiology #AlphaFold3 #ProteinStructure #MachineLearning #TabPFN #ESM2 #Bioinformatics #TCellReceptor
