Tranquillyzer: A Neural Network Framework for Long-read Annotation and Demultiplexing.

Area of Science:

• Genomics
• Bioinformatics
• Molecular Biology

Background:

• Long-read single-cell RNA sequencing offers full-length transcript information but faces challenges in interpreting complex reads.
• Existing analysis pipelines struggle with high error rates, varied library designs, and molecular artifacts, leading to inaccurate barcode and UMI detection.

Purpose of the Study:

• To introduce Tranquillyzer, a deep learning framework for accurate structural inference of long-read sequencing molecules.
• To enable robust identification of adapters, barcodes, UMIs, and transcript segments despite sequencing noise and structural variations.

Main Methods:

• Developed Tranquillyzer, a deep learning framework utilizing context-aware structural inference.
• Modeled full architectural context of sequencing reads at base resolution for annotation.
• Validated performance on simulated benchmarks and standard long-read protocols.

Main Results:

• Tranquillyzer achieved >99.7% structural filtering accuracy, >91% demultiplexing efficiency, and >99.9% demultiplexing accuracy.
• Demonstrated superior performance over existing methods in handling sequencing noise and structural variability.
• Enabled systematic detection of complex molecular artifacts, including multi-fragment chimeras.

Conclusions:

• Tranquillyzer provides a generalizable framework for reliable structural parsing of long-read sequencing data.
• Enhances the interpretation of single-cell and bulk long-read RNA sequencing data.
• Facilitates accurate full-length transcript profiling by overcoming limitations in read structure interpretation.