LMFE: A Novel Method for Predicting Plant LncRNA Based on Multi-Feature Fusion and Ensemble Learning

TL;DR

The paper introduces LMFE, a new computational method that accurately predicts plant lncRNAs using multi-feature fusion and ensemble learning, outperforming existing tools.

Contribution

The novel LMFE method combines multi-feature fusion with XGBoost and SMOTE to improve plant lncRNA prediction accuracy and robustness.

Findings

LMFE achieved 99.42% accuracy on benchmark datasets and outperformed existing methods like CPC2 and PLEKv2.

The method showed robust cross-species performance with accuracy ranging from 89.30% to 99.81%.

LMFE improved accuracy by 12.29% on unbalanced datasets compared to traditional methods.

Abstract

Background/Objectives: Long non-coding RNAs (lncRNAs) play a crucial regulatory role in plant trait expression and disease management, making their accurate prediction a key research focus for guiding biological experiments. While extensive studies have been conducted on animals and humans, plant lncRNA research remains relatively limited due to various challenges, such as data scarcity and genomic complexity. This study aims to bridge this gap by developing an effective computational method for predicting plant lncRNAs, specifically by classifying transcribed RNA sequences as lncRNAs or mRNAs using multi-feature analysis. Methods: We propose the lncRNA multi-feature-fusion ensemble learning (LMFE) approach, a novel method that integrates 100-dimensional features from RNA biological properties-based, sequence-based, and structure-based features, employing the XGBoost ensemble learning algorithm for prediction. To address unbalanced datasets, we implemented the synthetic minority oversampling technique (SMOTE). LMFE was validated across benchmark datasets, cross-species datasets, unbalanced datasets, and independent datasets. Results: LMFE achieved an accuracy of 99.42%, an F1score of 0.99, and an MCC of 0.98 on the benchmark dataset, with robust cross-species performance (accuracy ranging from 89.30% to 99.81%). On unbalanced datasets, LMFE attained an average accuracy of 99.41%, representing a 12.29% improvement over traditional methods without SMOTE (average ACC of 87.12%). Compared to state-of-the-art methods, such as CPC2 and PLEKv2, LMFE consistently outperformed them across multiple metrics on independent datasets (with an accuracy ranging from 97.33% to 99.21%), with redundant features having minimal impact on performance. Conclusions: LMFE provides a highly accurate and generalizable solution for plant lncRNA prediction, outperforming existing methods through multi-feature fusion and ensemble learning while demonstrating robustness to redundant features. Despite its effectiveness, variations in performance across species highlight the necessity for future improvements in managing diverse plant genomes. This method represents a valuable tool for advancing plant lncRNA research and guiding biological experiments.

Figures