Positive Impression of Low-Ranking Microrn as in Human Cancer Classification

Li Fei-Fei, Yongjun Piao, Meijing Li, Minghao Piao, Keun Ho Ryu2014 · DOI

Abstract

Recently, many studies based on microRNAs (miRNAs) showed a new aspect of cancer classification, and feature selection methods are used to reduce the high dimensionality of miRNA expression data. These methods just consider the problem of where feature to class is 1:1 or n:1. But one miRNA may have influence to more than one type of cancers. However, these miRNAs are considered to be low ranked in traditional feature selection methods and they are removed at most of time. Therefore, it is necessary to consider the problem of 1:n or m:n during feature selection. In our wok, we considered both high and low-ranking features to cover all problems (1:1, n:1, 1:n, m:n) in cancer classification. After numerous tests, information gain and chi-squared feature selection methods were chosen to select the high and low-ranking features to form the m-to-n feature subset, and LibSVM classifier was used to do the multi-class classification. Our results demonstrate that the m-to-n features make a positive impression of low-ranking microRNAs in cancer classification since they lead to achieve higher classification accuracy compared with the traditional feature selection methods.

Introduction

In the field of bioinformatics, feature selection is a critical step for handling high-dimensional genomic data, such as microRNA (miRNA) expression profiles. Standard feature selection approaches typically prioritize features with strong 1:1 correlations to specific outcomes. However, biological systems often exhibit complex 1:n or m:n relationships, where a single miRNA may influence multiple cancer types.

Because these "multi-tasking" miRNAs do not correlate perfectly with a single class, they are often assigned low ranks by traditional algorithms and subsequently discarded. This study investigates the hypothesis that these overlooked low-ranking features are essential for high-accuracy multi-class classification, proposing an "m-to-n" approach that combines both high- and low-ranking signals.

Materials and Methods

The study utilized a well-known benchmark dataset from Lu et al. (2005), consisting of expression levels for 217 mammalian miRNAs. The dataset includes 73 samples categorized into five distinct tumor types: colon, pancreas, uterus, B-cell acute lymphoblastic leukemia (ALL), and T-cell ALL.

Table 1: Distribution of tumor samples across the five cancer types studied.

To identify the most relevant features, the researchers evaluated four attribute evaluators: Pearson’s correlation, chi-squared distribution, information gain, and gain ratio. The core methodology involved creating an "m-to-n" feature subset by combining 10 high-ranking features (strong 1:1 signals) with 17 low-ranking features (representing 1:n or m:n signals). Performance was evaluated using the LibSVM classifier with 10-fold cross-validation, and results were compared against other algorithms such as A1DE, J48, and IBK.

Results and Analysis

The experimental results clearly demonstrate the value of including low-ranking features. While a subset of 10 high-ranking features achieved an accuracy of 89.04%, the combined m-to-n subset (27 features) reached a peak accuracy of 94.52%. This outperforms previous benchmarks on the same dataset, such as the Default ARTMAP classifier, which reached 88.89%.

Figure 1: Comparison of classification accuracy across different feature selection methods as the number of features increases.

Analysis of the different evaluators showed that Information Gain and Chi-Squared were the most effective at identifying useful low-ranking features. As shown in the comparative performance data, these methods provided more stable accuracy across varying feature counts compared to Pearson’s correlation and Gain Ratio. The study found that while low-ranking features perform poorly in isolation (approx. 52.05% accuracy), they act as a vital fine-tuning mechanism when paired with high-ranking features, resolving ambiguities between similar cancer types.

Conclusion

This research proves that "low-ranking" does not equate to "useless" in the context of cancer classification. By adopting an m-to-n feature selection strategy, researchers can capture the complex, overlapping biological signals inherent in miRNA data. Future work will focus on developing automated algorithms to determine the optimal balance of high- and low-ranking features, potentially improving diagnostic tools for genomic medicine.

Summary updated

Abstract

Recently, many studies based on microRNAs (miRNAs) showed a new aspect of cancer classification, and feature selection methods are used to reduce the high dimensionality of miRNA expression data. These methods just consider the problem of where feature to class is 1:1 or n:1. But one miRNA may have influence to more than one type of cancers. However, these miRNAs are considered to be low ranked in traditional feature selection methods and they are removed at most of time. Therefore, it is necessary to consider the problem of 1:n or m:n during feature selection. In our wok, we considered both high and low-ranking features to cover all problems (1:1, n:1, 1:n, m:n) in cancer classification. After numerous tests, information gain and chi-squared feature selection methods were chosen to select the high and low-ranking features to form the m-to-n feature subset, and LibSVM classifier was used to do the multi-class classification. Our results demonstrate that the m-to-n features make a positive impression of low-ranking microRNAs in cancer classification since they lead to achieve higher classification accuracy compared with the traditional feature selection methods.
BibTeX
@misc{feifei2014positive,
  author = {Li Fei-Fei and Yongjun Piao and Meijing Li and Minghao Piao and Keun Ho Ryu},
  title = {{Positive Impression of Low-Ranking Microrn as in Human Cancer Classification}},
  year = {2014},
  doi = {10.5121/csit.2014.4239},
}
Figure