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

Long-read single-cell RNA sequencing enables full-length transcript profiling but remains limited by challenges in interpreting structurally complex sequencing reads. High error rates, heterogeneous library designs, and frequent molecular artifacts disrupt barcode and UMI detection, while existing pipelines rely on positional heuristics that fail when structural elements are shifted, truncated, rearranged, or concatenated. Here, we introduce Tranquillyzer, a deep learning framework for global, context-aware structural inference of long-read molecules. Tranquillyzer performs base-resolution annotation of sequencing reads by modeling their full architectural context, enabling accurate identification of adapters, barcodes, UMIs, and transcript segments even under substantial sequencing noise and structural variability. Across simulated benchmarks, Tranquillyzer achieved > 99.7% structural filtering accuracy, > 91% demultiplexing efficiency, and > 99.9% demultiplexing accuracy, substantially exceeding existing methods. Tranquillyzer supports standard long-read single-cell protocols and can be rapidly trained to interpret custom library architectures. Its scalable processing and read-level visualization enable systematic detection of complex molecular artifacts, including multi-fragment chimeras. By enabling reliable structural parsing of long-read libraries, Tranquillyzer provides a generalizable framework for structural interpretation of single-cell and bulk long-read sequencing data. Tranquillyzer is freely available at: https://github.com/huishenlab/Tranquillyzer.