Discovering misannotated lncRNAs using deep learning training dynamics.

Area of Science:

• Genomics
• Computational Biology
• Molecular Biology

Background:

• Recent studies reveal long non-coding RNAs (lncRNAs) can encode functional micropeptides, challenging their non-coding annotation.
• Current detection methods like Ribo-Seq and mass spectrometry are resource-intensive and cell-type specific.

Purpose of the Study:

• To develop a computational method for identifying misannotated lncRNAs using only sequence data.
• To provide an accessible alternative to experimental techniques for lncRNA coding potential assessment.

Main Methods:

• Deep learning models were trained to distinguish between coding and non-coding transcripts.
• The models' training dynamics were utilized to pinpoint lncRNAs exhibiting coding potential.
• Sequence similarity searches (BLAST) and structural predictions (AlphaFold2) were employed for validation.

Main Results:

• The computational method successfully identified misannotated lncRNAs that significantly overlap with experimentally validated sets.
• Predicted micropeptides from candidate lncRNAs showed significant homology to known protein sequences.
• Analysis revealed high-confidence folded structures for ORFs within a subset of identified lncRNAs.

Conclusions:

• The proposed computational approach effectively identifies lncRNAs with coding potential from sequence data alone.
• This method can assist experimental research in characterizing the lncRNA-encoded proteome.
• The findings contribute to curating improved datasets for coding potential prediction models.