BERTrand

BERTrand predicts peptide:TCR binding interactions using BERT-based models augmented with random TCR pairing to assess T-cell receptor recognition of peptide antigens.

Key Features:

• Data sources: Trains on datasets derived from T-cell receptor sequencing experiments comprising known peptide:TCR binding interactions.
• Negative decoy augmentation: Augments training data with negative decoys generated from healthy donors' T-cell repertoires to improve binder/non-binder discrimination.
• Model architecture: Employs deep learning methods from natural language processing, specifically bidirectional encoder representations from transformers (BERT).
• Random TCR pairing: Incorporates random TCR pairing to enhance prediction accuracy and generalization across peptides.
• Cross-peptide generalization and performance: Demonstrates cross-peptide generalization with an AUROC of 0.69 and outperforms previously published methods on novel peptide sequences.

Scientific Applications:

• Vaccine Development: Prioritizes peptide candidates that are predicted to bind T-cell receptors and potentially elicit immune responses.
• Immunotherapy Research: Supports design and evaluation of personalized therapies by predicting interactions within patient-specific T-cell repertoires.
• Disease Pathogenesis Studies: Aids investigation of how specific peptide:TCR interactions contribute to disease mechanisms.

Methodology:

Trains BERT-based models on peptide:TCR interaction datasets from T-cell receptor sequencing, augments data with negative decoys from healthy donor repertoires and random TCR pairing, and evaluates cross-peptide generalization using AUROC (reported 0.69).