Multiple instance neural networks based on sparse attention for cancer detection using T-cell receptor sequences.

Area of Science:

• Immunology
• Computational Biology
• Machine Learning

Background:

• Early cancer detection is crucial for improving patient outcomes.
• T cell receptors (TCRs) are increasingly recognized for their role in tumor biology and immunity.
• Classical machine learning methods face challenges with the one-to-many nature of patient-to-TCR sequence data.

Purpose of the Study:

• To develop an enhanced computational model for cancer detection using TCR sequence data.
• To improve the performance and explainability of machine learning models in cancer detection.
• To address limitations of existing Multiple Instance Learning (MIL) approaches for TCR data analysis.

Main Methods:

• Proposed Multiple Instance Neural Networks based on Sparse Attention (MINN-SA).
• Utilized a sparse attention mechanism to filter uninformative instances within TCR data bags.
• Incorporated skip connections to enhance model interpretability and predictive power.

Main Results:

• MINN-SA achieved the highest Area Under the ROC Curve (AUC) scores on average across 10 cancer types.
• The model demonstrated improved performance compared to existing MIL approaches.
• Estimated attention weights identified specific TCRs associated with tumor antigens within T cell repertoires.

Conclusions:

• MINN-SA offers a promising approach for enhancing cancer detection accuracy and explainability using TCR data.
• The sparse attention mechanism effectively identifies relevant TCRs, contributing to better diagnostic performance.
• This method advances the application of MIL and neural networks in cancer immunology research.