TITAN

TITAN predicts T-cell receptor (TCR) specificity for antigenic epitopes by modeling TCR–epitope interactions.

Key Features:

• Bimodal Neural Network Architecture: A bimodal neural network simultaneously encodes TCR sequences and epitopes to enable independent assessment of generalization to unseen TCRs and epitopes.
• Epitope Encoding with SMILES: Epitopes are encoded at the atomic level using SMILES (Simplified Molecular Input Line Entry System) to represent molecular structure.
• Transfer Learning and Data Augmentation: The model leverages transfer learning and data augmentation to expand the effective training data for improved predictive performance.
• Performance Metrics: The model achieved an ROC-AUC of 0.87 in a 10-fold cross-validation for unseen TCRs and outperformed the state-of-the-art model ImRex in reported benchmarks.
• Attention Mechanism Analysis: An attention mechanism produces attention heatmaps that provide insights into model decision-making and reveal effects of epitope data sparsity.
• Generalization Capabilities: The approach emphasizes attention on chemically relevant molecular structures and shows improved generalization compared to existing methods, with noted limitations on unseen epitopes.

Scientific Applications:

• Prediction of TCR-epitope interactions: Predicts which epitopes are recognized by specific TCR sequences for studies of adaptive immunity.
• Generalization studies: Enables analysis of model generalization to unseen TCRs and epitopes in large sequence spaces with sparse data.
• Translational immunology: Supports exploratory work relevant to vaccine target identification and personalized immunotherapy research by prioritizing candidate TCR-epitope pairs.

Methodology:

Computational methods include a bimodal neural network encoding TCRs and SMILES-encoded epitopes, use of transfer learning and data augmentation, attention mechanisms producing heatmaps, and evaluation by 10-fold cross-validation reporting ROC-AUC (0.87) with comparisons to ImRex.