Iterative attack-and-defend framework for improving TCR-epitope binding prediction models.

Area of Science:

• Immunology
• Computational Biology
• Artificial Intelligence

Background:

• TCR-epitope binding prediction is vital for adoptive T cell therapy and vaccine design.
• Existing models suffer from false positives due to limited negative sample data.
• Current negative sample generation methods fail to address model-specific vulnerabilities.

Purpose of the Study:

• To develop a novel framework for systematically identifying and mitigating weaknesses in TCR-epitope prediction models.
• To enhance the robustness of TCR-epitope binding prediction models against false positives.
• To create a comprehensive adversarial negative dataset for model improvement.

Main Methods:

• Propose an iterative attack-and-defend framework using reinforcement learning from AI feedback (RLAIF).
• Attack phase: Generate biologically implausible sequences to deceive prediction models.
• Defense phase: Incorporate identified false positives into fine-tuning datasets to improve detection.

Main Results:

• Successfully applied the framework to five diverse TCR-epitope prediction models.
• Demonstrated significant improvement in models' ability to detect adversarial false positives.
• Created a combined dataset serving as a benchmarking tool for prediction models.

Conclusions:

• The attack-and-defend framework effectively enhances TCR-epitope prediction model robustness.
• The generated adversarial dataset improves model performance and reduces false positives.
• This approach offers a new strategy for improving biological prediction models with limited negative sampling.