A comparative study of multiple instance learning methods for cancer detection using T-cell receptor sequences.

Area of Science:

• Computational biology
• Machine learning
• Bioinformatics

Background:

• Multiple instance learning (MIL) is a machine learning technique for weakly supervised learning from data organized in bags.
• MIL has applications in image retrieval, object detection, and computer-aided diagnosis, with prior biomedical uses in medical imaging and molecule activity prediction.
• T-cell receptor (TCR) sequencing from peripheral blood offers a potential avenue for large-scale cancer screening.

Purpose of the Study:

• To investigate the applicability of 16 different MIL methods to cancer detection using TCR sequences.
• To evaluate MIL performance under simulated data conditions mimicking realistic scenarios and varying factors.
• To provide a proof of concept for distinguishing cancer patients from healthy individuals using peripheral blood TCR sequencing.

Main Methods:

• Reviewed and applied 16 distinct multiple instance learning algorithms.
• Simulated data based on two feasible data-generating mechanisms, varying key factors for performance evaluation.
• Applied selected MIL methods to TCR sequencing data from The Cancer Genome Atlas across ten cancer types.

Main Results:

• Satisfactory performance (Area Under the ROC Curve > 80%) was achieved for five out of ten cancer types when using appropriate MIL methods.
• Performance varied significantly across different MIL methods and cancer types.
• Identified specific MIL methods that performed well and those that were less effective for TCR-based cancer detection.

Conclusions:

• Multiple instance learning is a viable approach for cancer detection via TCR sequencing, particularly for certain cancer types.
• Method selection is crucial for achieving high performance; guidance is provided for choosing appropriate MIL techniques.
• Future research should focus on developing novel MIL methodologies to improve performance for challenging cancer types and enhance outcome explainability.