Clair3-RNA: A deep learning-based small variant caller for long-read RNA sequencing data.

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Long-read RNA sequencing (lrRNA-seq) enables full-length isoform and gene expression analysis but faces challenges due to high error rates and transcript complexity.
• Accurate variant calling in lrRNA-seq data is crucial for understanding transcript diversity and RNA editing events.

Purpose of the Study:

• To introduce Clair3-RNA, the first deep learning-based variant caller specifically designed for lrRNA-seq data.
• To enhance variant-calling performance by addressing lrRNA-seq specific challenges like uneven coverage and RNA editing.

Main Methods:

• Development of Clair3-RNA, a deep learning model leveraging Clair series pipelines with lrRNA-seq specific optimizations.
• Implementation of techniques including uneven coverage normalization, refined training data, editing site discovery, and haplotype phasing.
• Evaluation on PacBio and Oxford Nanopore Technologies (ONT) platforms, including cDNA and direct RNA sequencing (dRNA).

Main Results:

• Clair3-RNA achieved high SNP F1-scores: ~91% on ONT dRNA004, ~92% on PacBio Iso-Seq/MAS-Seq (≥4 reads).
• Performance improved to ~95% (ONT) and ~96% (PacBio) with ≥10 reads, and ~97% (ONT) / ~98% (PacBio) with phased reads.
• Demonstrated superior performance over existing callers on GIAB samples and accurate discrimination of RNA editing sites.

Conclusions:

• Clair3-RNA offers a robust and accurate solution for variant calling in lrRNA-seq data across different platforms.
• The tool's ability to handle RNA editing events enhances its utility for comprehensive transcriptomic analysis.
• Clair3-RNA is an open-source resource, facilitating further research in RNA variant calling.