LDNFSGB: prediction of long non-coding rna and disease association using network feature similarity and gradient boosting

Yuan Zhang, Fei Ye, Dapeng Xiong, Xieping GaoBMC Bioinformatics 2020 · OpenReview

TL;DR

This paper introduces LDNFSGB, a novel computational method for predicting associations between long non-coding RNAs (lncRNAs) and human diseases. By integrating network-based feature similarity with gradient boosting, LDNFSGB significantly outperforms existing methods, offering a more comprehensive approach to understanding these complex relationships and aiding in disease diagnosis and treatment.

What Problem Does This Paper Address?

Long non-coding RNAs (lncRNAs) are increasingly recognized for their crucial roles in various biological processes and their association with numerous human diseases. However, experimentally identifying these associations is a time-consuming and costly endeavor. While computational methods have emerged to address this, many existing approaches either focus on limited aspects of lncRNA and disease information or struggle with the inherent complexity and sparsity of biological data. This paper aims to bridge this gap by developing a more robust and comprehensive method that captures both local and global information from diverse biological networks to accurately predict lncRNA-disease associations.

Methods

The core of LDNFSGB lies in its innovative feature engineering and predictive modeling.

  • Comprehensive Feature Vector Construction: The method constructs a rich feature vector by integrating multiple similarity measures:
  • Disease Semantic Similarity (DISS): This is built upon two strategies: DISSS1 (local semantic information) and DISSS2 (global semantic information), derived from Medical Subject Heading (MeSH) terms. It also incorporates the Gaussian interaction profile kernel similarity for diseases (DISGS) to account for missing data.
  • lncRNA Function Similarity (LNCS): Similar to DISS, LNCS combines local (LNCFS1) and global (LNCFS2) functional similarity measures with the Gaussian interaction profile kernel similarity for lncRNAs (LNCGS).
  • lncRNA-Disease Interaction (LNCDIS): This represents the known associations between lncRNAs and diseases.
  • Dimensionality Reduction with Autoencoder: To handle the high dimensionality of the constructed feature vector and extract optimal features, an autoencoder is employed. This unsupervised learning technique learns a compressed representation of the data, enhancing efficiency and performance.
  • Gradient Boosting for Prediction: The dimensionality-reduced features are then fed into a gradient boosting algorithm. This ensemble method, known for its robustness and accuracy, builds a strong predictive model by sequentially adding weak learners (regression trees) that correct the errors of their predecessors.

Datasets

The study utilizes three publicly available datasets for training and evaluation:

  • LncRNADisease (v2017): This dataset contains known lncRNA-disease associations and was used for initial method development and extensive validation.
  • Lnc2Cancer: A manually curated dataset focusing on experimentally supported associations between lncRNAs and cancers.
  • LncRNADisease2.0: An updated version of the LncRNADisease dataset, featuring a larger number of lncRNA-disease associations.

Key Insights and Ideas

  • Holistic Feature Representation is Crucial: The paper strongly emphasizes that simply using one type of similarity (e.g., semantic or functional) is insufficient. Integrating diverse network features, including both local and global information, is key to accurately capturing the complex relationships between lncRNAs and diseases.
  • Autoencoders for Efficient Feature Extraction: The use of autoencoders proves effective in reducing the dimensionality of complex biological data, leading to more discriminative features and improving the efficiency of downstream prediction models.
  • Gradient Boosting's Power in Biological Prediction: The study highlights gradient boosting as a highly effective classifier for lncRNA-disease association prediction, outperforming other common machine learning algorithms. Its ability to iteratively refine predictions makes it well-suited for noisy biological data.
  • Uncovering Novel Associations: The method not only validates known associations but also predicts novel, potentially undiscovered lncRNAs related to diseases, opening new avenues for experimental research and therapeutic development.

Key Results

  • Superior Predictive Performance: LDNFSGB achieved state-of-the-art performance across all three datasets. On the LncRNADisease dataset, it achieved an AUC of 0.9761 using hold-out validation, outperforming other methods by a significant margin (2.59%-10.49%). Similar high AUC values were observed on Lnc2Cancer (0.9447) and LncRNADisease2.0 (0.9933).
  • Robustness Across Validation Methods: The model demonstrated consistent high performance across three validation strategies: hold-out, leave-one-out cross-validation (LOOCV), and 10-fold cross-validation.
  • Effective Classifier Comparison: Gradient boosting consistently outperformed other classifiers (SVM, Naïve Bayes, Logistic Regression, Random Forest, etc.) in terms of Accuracy, Sensitivity, MCC, and AUC.
  • Successful Case Studies: The method successfully predicted top candidate lncRNAs for various diseases, including colorectal cancer, osteosarcoma, cervical cancer, glioma, heart failure, and Alzheimer's disease. Notably, six of the top ten predicted lncRNAs for heart failure and Alzheimer's disease were confirmed by existing literature.

Limitations and Future Work

  • Limited Feature Scope: The authors acknowledge that the current feature extraction primarily relies on functional information of lncRNAs. They suggest that incorporating other characteristics like sequence, structure, and location information could further enhance prediction accuracy.
  • Supervised Learning Approach: The current method uses a supervised approach, treating unlabelled samples as negative. The authors propose that unsupervised learning could be explored to incorporate more information and potentially improve performance.
  • Unveiling Molecular Mechanisms: While LDNFSGB identifies potential novel associations, the molecular mechanisms underlying these relationships often remain to be elucidated, requiring further experimental validation.

The study opens doors for future research by providing a powerful tool for identifying potential lncRNA-disease associations and offering new perspectives for disease diagnosis and targeted treatment.

Summary updated

Abstract

A large number of experimental studies show that the mutation and regulation of long non-coding RNAs (lncRNAs) are associated with various human diseases. Accurate prediction of lncRNA-disease associations can provide a new perspective for the diagnosis and treatment of diseases. The main function of many lncRNAs is still unclear and using traditional experiments to detect lncRNA-disease associations is time-consuming.
BibTeX
@article{zhang2020ldnfsgb,
  author = {Yuan Zhang and Fei Ye and Dapeng Xiong and Xieping Gao},
  title = {{LDNFSGB: prediction of long non-coding rna and disease association using network feature similarity and gradient boosting}},
  journal = {BMC Bioinformatics},
  year = {2020},
  url = {https://openreview.net/forum?id=3NDbUunMiT},
}
Figure