Uncertainty-aware benchmarking reveals ambiguous transcripts in mRNA-lncRNA classification.

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Distinguishing long non-coding RNAs (lncRNAs) from protein-coding transcripts is a persistent challenge in genomics.
• lncRNAs often share mRNA-like processing, and current sequence-derived signals are insufficient for accurate classification.
• Recent GENCODE efforts highlight the need for improved characterization of features influencing transcript classification.

Purpose of the Study:

• To develop an uncertainty-aware benchmarking framework for evaluating transcript classifiers.
• To identify sequence and feature patterns associated with classification uncertainty and errors.
• To improve the accuracy and reliability of distinguishing lncRNAs from protein-coding transcripts.

Main Methods:

• Performed uncertainty-aware benchmarking on GENCODE v46-v47 data using eight transcript classifiers.
• Quantified inter-tool agreement and entropy-based uncertainty to stratify transcripts.
• Incorporated repeat-derived and non-B DNA motif features alongside standard sequence and ORF signals.

Main Results:

• Approximately 45% of transcripts exhibited inter-tool discordance, especially lncRNAs.
• Low-uncertainty predictions correlated with strong coding-like signals; high-uncertainty profiles showed mixed signatures.
• Repeat-derived features emerged as significant contributors to classification, alongside classical predictors.

Conclusions:

• A novel framework combining benchmarking, uncertainty stratification, and extended feature profiling identifies patterns in classifier disagreement.
• This approach offers practical guidance for interpreting predictions and developing more robust classifiers.
• The study sheds light on sequence properties that effectively distinguish lncRNA sequences.