Improved lung cancer classification by employing diverse molecular features of microRNAs

Shiyong Guo, Chunyi Mao, Jun Peng, Shaohui Xie, Jun Yang, Wenping Xie, Wanran Li, Huaide Yang, Hao Guo, Zexuan Zhu, Yun ZhengHeliyon 2024 · DOI

TL;DR

This study demonstrates that incorporating miRNA editing information, beyond just miRNA abundance, significantly improves the classification accuracy of lung adenocarcinoma (LUAD). By analyzing original miRNAs, edited miRNAs, and their editing levels, the researchers achieved perfect 100% classification accuracy using only three key miRNA features.

What Problem Does This Paper Address?

Lung cancer, particularly lung adenocarcinoma (LUAD), remains a leading cause of cancer death worldwide. While microRNAs (miRNAs) have shown promise as biomarkers for cancer classification due to their dysregulation in tumors, existing classification methods primarily rely on miRNA abundance. However, miRNAs undergo various modifications, including editing, which can alter their function and expression. This paper addresses the gap in understanding how these miRNA editing events, in addition to standard miRNA abundance, can be leveraged to improve the accuracy and robustness of LUAD classification. The authors hypothesize that integrating these diverse molecular features will lead to more effective diagnostic tools.

Methods

The study employed a multi-faceted approach:

  • Data Acquisition and Preprocessing: They utilized a combination of publicly available small RNA sequencing data and newly generated data, totaling 395 samples (LUAD and normal controls). Crucially, they implemented the ComBat-seq algorithm to correct for batch effects arising from different sequencing platforms and studies, a common challenge in multi-source omics data.
  • Feature Extraction: Three types of molecular features were extracted for each sample:
  • Abundances of Original miRNAs (AOM): Standard miRNA expression levels.
  • Abundances of Edited miRNAs (AEM): Expression levels of miRNAs that have undergone editing.
  • Editing Levels of miRNA Editing Sites (EL): Quantitative measures of how frequently specific miRNA editing events occur.
  • Machine Learning for Classification: Eight different classification algorithms were tested, including kNN, C4.5, RF, SVM, NB, XGB, and NN. A key algorithm used for both feature selection and classification was the Discrete Function Learning (DFL) algorithm, which leverages information theory to identify minimal feature sets with high predictive power.
  • Feature Selection: To identify the most informative features, three methods were employed: Correlation-based Feature Selection (CFS), Wrapper for Feature Subset Selection (WFS), and the DFL algorithm.

Datasets

The study utilized a total of 395 small RNA sequencing profiles for LUAD and control samples. These included:

  • 38 samples generated by the authors from the First People's Hospital of Yunnan Province.
  • 357 publicly available samples downloaded from the NCBI SRA database, originating from four different studies.

The data was split into an 80% training set (316 samples) and a 20% independent testing set (79 samples) for model validation.

Key Insights and Ideas

  • MiRNA Editing is More Than Just Noise: The study strongly suggests that miRNA editing is not merely a random modification but carries significant biological information relevant to disease states like LUAD.
  • Combined Features Outperform Individual Ones: Integrating AOM, AEM, and EL consistently led to better classification performance across multiple algorithms compared to using only AOM or only editing-related features.
  • The Power of Minimal Feature Sets: The DFL algorithm's ability to identify a very small set of highly predictive features highlights the potential for developing simplified, yet powerful, diagnostic tools.
  • Specific miRNAs as Key Biomarkers: The identification of specific miRNAs (e.g., hsa-miR-135b-5p, hsa-miR-210-3p, and hsa-miR-182_48u) as critical for perfect classification underscores their potential as diagnostic biomarkers.

Key Results

  • Near-Perfect Classification with Combined Features: When using all three feature types (AOM, AEM, EL), several classification algorithms achieved accuracies of 100% on the independent test set.
  • DFL Algorithm's Remarkable Performance: The DFL algorithm, after feature selection, identified just three features (hsa-miR-135b-5p, hsa-miR-210-3p, and hsa-miR-182_48u) that enabled seven classification algorithms to achieve 100% accuracy on the independent testing set of 79 samples. This is a significant improvement over using only AOM features, where accuracies were lower.
  • Batch Effect Mitigation: PCA analysis before and after ComBat-seq correction visually demonstrated the successful removal of batch effects, ensuring the reliability of the results across different data sources.
  • Functional Relevance of Key miRNAs: The paper discusses the known roles of the identified key miRNAs in LUAD development and progression, lending biological plausibility to the findings.

Limitations and Future Work

  • Sample Size and Diversity: While 395 samples is a reasonable number, the authors acknowledge that the samples were obtained from five different studies, necessitating robust batch effect correction. Future work could involve larger, more homogenous datasets.
  • Clinical Validation: The study provides strong evidence for the diagnostic potential of these miRNA features. However, further validation in larger, prospective clinical cohorts is essential before these findings can be translated into clinical practice.
  • Mechanism of Action: While the functional relevance of some identified miRNAs is discussed, a deeper investigation into the precise mechanisms by which miRNA editing influences LUAD pathogenesis and how these specific miRNAs contribute to classification could be beneficial.
  • Exploration of Other Editing Types: The study primarily focused on A-to-I editing and 3' modifications. Exploring other types of miRNA modifications could potentially reveal additional informative features.

Summary updated

Abstract

MiRNAs are edited or modified in multiple ways during their biogenesis pathways. It was reported that miRNA editing was deregulated in tumors, suggesting the potential value of miRNA editing in cancer classification. Here we extracted three types of miRNA features from 395 LUAD and control samples, including the abundances of original miRNAs, the abundances of edited miRNAs, and the editing levels of miRNA editing sites. Our results show that eight classification algorithms selected generally had better performances on combined features than on the abundances of miRNAs or editing features of miRNAs alone. One feature selection algorithm, i.e., the DFL algorithm, selected only three features, i.e., the frequencies of hsa-miR-135b-5p, hsa-miR-210-3p and hsa-mir-182_48u (an edited miRNA), from 316 training samples. Seven classification algorithms achieved 100% accuracies on these three features for 79 independent testing samples. These results indicate that the additional information of miRNA editing is useful in improving the classification of LUAD samples.
BibTeX
@inproceedings{guo2024improved,
  author = {Shiyong Guo and Chunyi Mao and Jun Peng and Shaohui Xie and Jun Yang and Wenping Xie and Wanran Li and Huaide Yang and Hao Guo and Zexuan Zhu and Yun Zheng},
  title = {{Improved lung cancer classification by employing diverse molecular features of microRNAs}},
  booktitle = {Heliyon},
  year = {2024},
  doi = {10.1016/j.heliyon.2024.e26081},
}
Figure