Loc4Lnc: Accurate prediction of long noncoding RNA subcellular localization via enhanced RNA sequence representation

Yujia Cheng, Xiaoyong Pan, Yang YangQuant. Biol 2025 · DOI

TL;DR

This paper introduces Loc4Lnc, a novel deep learning framework that accurately predicts the subcellular localization of long noncoding RNAs (lncRNAs). By integrating convolutional and transformer layers, Loc4Lnc effectively captures complex long-range interactions within RNA sequences, outperforming existing methods and providing a more robust tool for understanding lncRNA functions.

What Problem Does This Paper Address?

Long noncoding RNAs (lncRNAs) are crucial regulators of gene expression and cellular processes, but their diverse functions are intimately tied to their location within the cell (e.g., nucleus, cytoplasm, exosome). Accurately predicting these locations is essential for understanding their roles in health and disease and for developing therapeutic strategies. However, existing computational methods struggle to capture the intricate, long-range interactions within lncRNA sequences, which are often quite long. This limitation leads to suboptimal predictive accuracy, especially when distant sequence elements play a critical role in determining localization. Furthermore, inconsistencies in classification definitions and databases across different predictors hinder direct comparison and progress in the field.

Methods

Loc4Lnc employs a sophisticated deep learning architecture designed to overcome the limitations of previous methods. The core of the approach lies in its enhanced feature extraction strategy:

  • Hybrid Architecture: It combines convolutional layers (to capture local sequence motifs) with transformer blocks (to model long-range dependencies). This hybrid approach allows for a more comprehensive representation of the RNA sequence.
  • Enformer-Inspired Feature Extraction: The model is inspired by the Enformer architecture, known for its ability to handle long sequences and capture complex interactions. It includes a "stem" module for initial feature extraction, a "convolutional tower" with multiple convolutional and residual blocks, and a "transformer block" with multi-head attention mechanisms.
  • Custom Relative Positional Encodings: To better distinguish between nearby and distant regulatory elements, Loc4Lnc utilizes custom relative positional encodings, enhancing its ability to model spatial relationships within the sequence.
  • TextCNN for Classification: The extracted features are then fed into a TextCNN classifier, a popular choice for text classification tasks, to perform the final five-class prediction.

Datasets

The study utilizes the RNALocate v2.0 database as the primary source for constructing a standardized benchmark dataset. This newer version offers more comprehensive and up-to-date lncRNA subcellular localization records compared to the older RNALocate v1.0. The authors filtered the data to include only Homo sapiens lncRNAs and selected five distinct subcellular locations: cytoplasm, nucleus, cytosol, chromatin, and exosome. They also addressed potential redundancy using the CD-HIT tool and excluded multilabel entries. The final dataset comprises 3277 lncRNA samples. This dataset was split into training (80%) and testing (20%) sets.

Key Insights and Ideas

  • Long-Range Interactions are Key: The paper strongly emphasizes that accurately predicting lncRNA localization hinges on capturing long-range interactions within the RNA sequence, a capability that traditional methods often lack.
  • Hybrid Models Excel: Combining convolutional and transformer architectures proves highly effective, leveraging the strengths of both for local and global feature extraction.
  • Database Matters: The shift from RNALocate v1.0 to v2.0 significantly impacts model performance, highlighting the importance of using up-to-date and comprehensive datasets for training and evaluation.
  • Standardization is Crucial: The authors highlight the challenges posed by inconsistent classification definitions and varying numbers of predicted locations across different tools, underscoring the need for standardized benchmarks.

Key Results

Loc4Lnc demonstrates significant improvements in predicting lncRNA subcellular localization.

  • On an independent test set, Loc4Lnc achieved a prediction accuracy of 0.636, outperforming existing methodologies.
  • When compared against traditional feature extraction methods (k-mer, Word2vec, DNABERT) combined with classifiers, Loc4Lnc consistently surpassed them. For instance, DNABERT (5-mer) + SVM achieved an accuracy of 0.546, while Loc4Lnc reached 0.636.
  • In comparisons with state-of-the-art predictors (LncLocator, DeepLncLoc, GraphLncLoc, LncLocFormer), Loc4Lnc also showed superior performance across multiple evaluation metrics and tasks. For example, on Task 1 (predicting four locations), Loc4Lnc achieved an accuracy of 0.662, significantly outperforming LncLocator (0.257) and DeepLncLoc (0.253).
  • Loc4Lnc supports prediction across five distinct subcellular locations, which is more than some of the compared predictors.

Limitations and Future Work

The authors acknowledge a few limitations:

  • Computational Complexity: While the use of a pretrained model (Enformer) helps, the computational complexity still grows exponentially with sequence length, which can be a challenge for extremely long sequences.
  • Data Quality and Variety: The predictive performance is inherently limited by the quality and variety of subcellular localization labels in the training data. Underrepresented or newly discovered compartments may lead to reduced performance.

Future research directions suggested include:

  • Incorporating Additional Biological Contexts: Integrating information like RNA secondary structure and protein-binding data could provide a more comprehensive understanding of lncRNA functions.
  • Exploring Lightweight Architectures: Developing more lightweight architectures or hybrid models could reduce computational costs and improve accessibility.
  • Extending to Multilabel Classification: Supporting multilabel classification for lncRNAs localized to multiple compartments would enhance the model's utility.

Summary updated

BibTeX
@article{cheng2025loc,
  author = {Yujia Cheng and Xiaoyong Pan and Yang Yang},
  title = {{Loc4Lnc: Accurate prediction of long noncoding RNA subcellular localization via enhanced RNA sequence representation}},
  journal = {Quant. Biol},
  year = {2025},
  doi = {10.1002/qub2.100},
}
Figure