DeepLION

DeepLION identifies cancer-associated T cell receptors from TCR-sequencing data using deep multi-instance learning to enable noninvasive cancer detection via immune-response analysis.

Key Features:

• Deep Learning Framework: Uses a deep learning architecture with alternative convolution filters and 1-max pooling to process amino acid fragments of varying lengths.
• Multi-Instance Learning: Implements a multi-instance learning framework that models correlations among TCRs within the same repertoire and assigns adjusted weights to each TCR sequence during prediction.

Scientific Applications:

• Cancer-associated TCR identification: Identifies cancer-associated T cell receptors from TCR-sequencing data for computational cancer detection.
• Cross-cancer validation: Validated on patient cohorts from nine different cancer types, demonstrating robust predictive performance across cancers.
• Performance benchmarks: Reported AUCs include 0.97 for thyroid cancer and 0.90 for lung cancer.

Methodology:

Computational methods include alternative convolution filters and 1-max pooling to handle variable-length amino acid fragments, together with a multi-instance learning approach that models inter-TCR correlations and assigns adjusted weights to sequences.