Transfer learning for T-cell response prediction.

Area of Science:

• Immunology
• Computational Biology
• Bioinformatics

Background:

• Predicting T-cell responses to peptides is vital for developing personalized cancer vaccines.
• Limited and heterogeneous training data pose challenges, risking shortcut learning where models focus on data source rather than peptide characteristics.
• Existing transformer models for T-cell response prediction can exhibit inflated performance due to shortcut learning.

Purpose of the Study:

• To address the challenges of predicting T-cell responses using transformer models with heterogeneous, multi-domain data.
• To propose a domain-aware evaluation scheme to mitigate inflated performance metrics.
• To investigate transfer learning techniques for robust T-cell response prediction.

Main Methods:

• Utilized a transformer model architecture for T-cell response prediction.
• Developed and implemented a domain-aware evaluation scheme to assess model generalization.
• Applied and compared various transfer learning techniques, including per-source fine-tuning.

Main Results:

• Demonstrated that shortcut learning can indeed inflate predictive performance in T-cell response models.
• Showcased the effectiveness of the proposed domain-aware evaluation scheme.
• Found that per-source fine-tuning is a robust transfer learning strategy across diverse peptide sources.
• Achieved competitive performance compared to state-of-the-art methods for human peptide T-cell response prediction.

Conclusions:

• Transformer models for T-cell response prediction require careful evaluation to avoid shortcut learning.
• A domain-aware evaluation scheme is crucial for reliable assessment of model performance.
• Transfer learning, particularly per-source fine-tuning, significantly enhances model robustness and accuracy.
• The developed approach offers a promising strategy for advancing personalized cancer vaccine development through accurate T-cell response prediction.