Multiple instance learning method based on convolutional neural network and self-attention for early cancer...

Area of Science:

• Oncology
• Bioinformatics
• Artificial Intelligence

Background:

• Early cancer detection is crucial for improving patient outcomes.
• T-cell receptor sequencing (TCR-seq) shows promise for cancer diagnostics.
• Existing methods may lack the precision needed for early-stage identification.

Purpose of the Study:

• To develop an advanced multiple instance learning (MIL) method for enhanced cancer detection.
• To leverage deep learning, specifically convolutional neural networks (CNNs) and self-attention, for analyzing TCR-seq data.
• To improve the accuracy and efficiency of early cancer identification using genomic biomarkers.

Main Methods:

• Developed MICA, a multiple instance learning method integrating CNNs and self-attention.
• Preprocessed TCR-seq data using word vectors for feature extraction.
• Employed an enhanced self-attention mechanism to capture instance relationships within TCR-seq data.
• Utilized cross-validation for rigorous performance evaluation.

Main Results:

• MICA achieved an Area Under the Curve (AUC) of 0.911 for lung cancer detection.
• MICA achieved an AUC of 0.946 for thyroid cancer detection.
• Demonstrated significant performance improvements over existing methods, with AUC increases of 7.1% and 2.1% for lung and thyroid cancers, respectively.

Conclusions:

• MICA is a highly effective deep learning approach for early cancer detection via TCR-seq.
• The method demonstrates superior performance in identifying lung and thyroid cancers.
• MICA offers a promising tool for advancing precision oncology and cancer diagnostics.