ncProFormer: A CNN-enhanced Transformer for ncRNA Coding-Potential Prediction.

Area of Science:

• Genomics
• Computational Biology
• Molecular Biology

Background:

• Noncoding RNAs (ncRNAs) can be translated into functional micropeptides, but identifying coding-capable ncRNAs is challenging.
• Weak translation signals, low conservation, and data heterogeneity complicate accurate prediction.

Purpose of the Study:

• To develop an advanced deep learning framework, ncProFormer, for predicting the coding potential of ncRNAs.
• To improve the accuracy and generalizability of identifying functional micropeptide-encoding ncRNAs.

Main Methods:

• ncProFormer integrates the GENA-LM nucleic-acid language model for contextual sequence embeddings.
• A convolutional neural network (CNN)-enhanced transformer encoder captures local and long-range sequence dependencies.
• The framework utilizes an all-token representation strategy for comprehensive analysis.

Main Results:

• ncProFormer significantly outperformed existing methods on human, external validation, and public benchmark datasets.
• The model demonstrated robust predictive performance across species (human, mouse, rat) without retraining.
• This highlights the transferability of learned biological representations and robustness to distributional shifts.

Conclusions:

• ncProFormer is an effective and generalizable deep learning framework for identifying coding-capable ncRNAs.
• The tool offers a promising computational approach for characterizing ncRNA functions in diverse transcriptomic contexts.
• This work advances the field of ncRNA research by providing a reliable method for discovering novel micropeptide-encoding ncRNAs.