Gene Selection for Multiclass Prediction by Weighted Fisher Criterion

Jianhua Xuan, Yue Joseph Wang, Yibin Dong, Yuanjian Feng, Bin Wang, Javed I. Khan, Marina Bakay, Zuyi Wang, Lauren Pachman, Sara Winokur, Yi-Wen Chen, Robert Clarke, Eric P. HoffmanEURASIP J. Bioinform. Syst. Biol 2007 · DOI

TL;DR

This paper introduces a novel two-step gene selection method, the Weighted Fisher Criterion (wFC) approach, to identify a small, highly discriminative subset of genes for accurate multiclass disease prediction from gene expression data. By effectively handling the "curse of dimensionality" and emphasizing neighboring class separability, this method achieves high prediction accuracies on real-world datasets, offering a more efficient and cost-effective alternative to analyzing thousands of genes.

What Problem Does This Paper Address?

Gene expression profiling holds immense promise for understanding disease heterogeneity and developing molecular diagnostics. However, a major hurdle is the "curse of dimensionality": the vast number of genes (tens of thousands) compared to a small number of samples (typically 10-100) in microarray studies. This imbalance can degrade the performance of classification algorithms. Furthermore, existing gene selection methods often struggle with multiclass problems, where simply identifying individually discriminatory genes might not yield the most effective subset for distinguishing between multiple disease subtypes. The paper aims to address these challenges by developing a robust gene selection strategy that can pinpoint a minimal yet powerful set of genes for accurate multiclass classification.

Methods

The core of this paper is a two-step gene selection strategy:

  1. Individually Discriminatory Gene (IDG) Selection: This initial step uses a one-dimensional Weighted Fisher Criterion (1D wFC). The wFC is designed to give more weight to pairs of classes that are closer together (neighboring classes) and less weight to well-separated classes. This is achieved by incorporating a weighting function that depends on the Mahalanobis distance between class pairs. This approach is theoretically motivated to better approximate the true Bayes prediction error, especially in multiclass scenarios where traditional Fisher Criterion can be suboptimal. Genes exceeding an empirically determined threshold of 1D wFC are selected as IDGs.
  2. Jointly Discriminatory Gene (JDG) Selection: In the second step, a sequential search method, specifically Sequential Backward Floating Selection (SBFS) or Sequential Forward Floating Selection (SFFS), is employed. This step uses a multidimensional wFC to measure the joint class separability of gene subsets. The goal is to refine the IDG set into a smaller, more powerful JDG set by considering the correlation and combined discriminatory power of genes.

The performance of the selected gene subsets is then evaluated using Artificial Neural Networks (ANNs), specifically Multilayer Perceptrons (MLPs), and Support Vector Machines (SVMs), employing techniques like 3-fold and 10-fold cross-validation.

Datasets

The proposed method was evaluated on two publicly available microarray datasets:

  • NCI's SRBCTs dataset: This dataset comprises gene expression profiles for small round blue cell tumors, classified into four subtypes: Burkitt lymphoma (BL), Ewing sarcoma (EWS), neuroblastoma (NB), and rhabdomyosarcoma (RMS). It includes 63 training samples and 25 test samples.
  • CNMC's Muscular Dystrophies (MDs) dataset: This dataset contains gene expression profiles from 39 sample arrays for four types of muscular dystrophy: Becker muscular dystrophy (BMD), Duchenne muscular dystrophy (DMD), Dysferlin deficiency (LGMD 2B), and Fascioscapulohumeral dystrophy (FSHD).

Key Insights and Ideas

  • Neighboring Class Emphasis: The wFC's ability to prioritize distinguishing between similar classes (e.g., closely related disease subtypes) is crucial for accurate multiclass prediction, where traditional methods might overemphasize differences between very distinct classes.
  • Two-Step Efficiency: The IDG/JDG approach effectively reduces the gene search space, moving from tens of thousands of genes to a handful of highly informative ones, making the process computationally feasible and leading to more interpretable models.
  • "Peaking Phenomenon" Support: The study provides experimental evidence for the "peaking phenomenon" in ANNs, where adding too many features can degrade performance, reinforcing the need for effective gene selection.
  • Diagnostic Microarrays: The findings suggest that a small, curated set of genes can be sufficient for accurate diagnosis, paving the way for cost-effective diagnostic microarrays.

Key Results

  • SRBCTs: The proposed IDG/JDG method identified a subset of 9 genes that achieved a misclassification error rate (MCER) of 3.10% (3-fold CV) and 2.24% (10-fold CV) for SRBCT classification. This is comparable to or better than methods using 96 or 43 genes. On a blind test set, these 9 genes achieved 100% accuracy in classifying SRBCTs and correctly rejected non-SRBCT samples.
  • Muscular Dystrophies (MDs): The method identified 11 JDGs that resulted in an MCER of 14.8% (MLPs, 3-fold CV) and 7.69% (SVMs, 3-fold CV). This outperformed IDGs (15.5% MCER with 69 genes) and Dudoit's method (10.26% MCER with 94 genes).

Limitations and Future Work

  • Empirical Thresholds: The choice of thresholds for IDG selection (e.g., number of genes) was described as empirical, based on preliminary studies of MCER and stability analysis. The paper acknowledges that determining optimal thresholds can be challenging.
  • "Peaking Phenomenon" Mitigation: While the paper discusses the "peaking phenomenon" and mentions SVMs and subspace methods as potential solutions, further investigation into these approaches with large JDG sets is suggested for future work.
  • Subspace Classifiers: The authors express interest in investigating subspace classifiers in conjunction with a large number of JDGs to further improve generalization performance in diagnosing muscular dystrophies.
  • Blind Test Sample Rejection: For the SRBCT blind test, while the 9 JDGs correctly classified samples, the "95th percentile" criterion for rejecting diagnoses was not always definitive, indicating room for refinement in confidence assessment.

Summary updated

BibTeX
@article{xuan2007gene,
  author = {Jianhua Xuan and Yue Joseph Wang and Yibin Dong and Yuanjian Feng and Bin Wang and Javed I. Khan and Marina Bakay and Zuyi Wang and Lauren Pachman and Sara Winokur and Yi-Wen Chen and Robert Clarke and Eric P. Hoffman},
  title = {{Gene Selection for Multiclass Prediction by Weighted Fisher Criterion}},
  journal = {EURASIP J. Bioinform. Syst. Biol.},
  year = {2007},
  doi = {10.1155/2007/64628},
}
Figure