A Self-supervised Learning Framework for Classifying Microarray Gene Expression Data

Yijuan Lu, Qi Tian, Feng Liu, Maribel Sanchez, Yufeng WangInternational Conference on Computational Science 2006 · DOI

Introduction

Monitoring thousands of genes simultaneously via microarray technology has revolutionized biological research. However, analyzing this data presents significant computational challenges, primarily due to high dimensionality (thousands of genes vs. few samples) and the scarcity of labeled "ground truth" data. Traditional supervised learning algorithms, such as Support Vector Machines (SVMs), often struggle when labeled data is limited. This paper introduces Kernel Discriminant-EM (KDEM), a self-supervised learning framework that leverages vast amounts of unlabeled data to improve classification accuracy in gene expression profiles.

Methodology

KDEM evolves the linear Discriminant-EM (DEM) algorithm by incorporating nonlinear kernel mappings. The framework operates through an iterative three-step process:

  1. Nonlinear Projection: Data is projected into a high-dimensional feature space using the "kernel trick," making complex biological patterns more likely to be linearly separable.
  2. Dimensionality Reduction: The data is then projected into a lower-dimensional discriminating subspace to filter out noise and isolate relevant features.
  3. Expectation-Maximization (EM) Loop:
    • E-step: Assigns probabilistic labels to unlabeled data based on the current classifier.
  4. D-step: Optimizes the nonlinear subspace for class separation.
  5. M-step: Updates classifier parameters using both original labeled data and newly "pseudo-labeled" unlabeled data.

By iterating these steps, the model uses the underlying structure of unlabeled data to regularize the learning process, preventing overfitting to small labeled datasets.

Experimental Setup

The study utilized a time-course expression profile of the malaria parasite, Plasmodium falciparum, during its 48-hour red blood cell cycle. The dataset included 7,091 oligonucleotides representing over 4,000 Open Reading Frames (ORFs). A subset of 472 annotated genes across 12 functional classes served as the labeled set, while 3,776 genes with unknown functions were treated as unlabeled data to assist the self-supervision process.

Key Results

Performance Across Functional Classes

KDEM demonstrated superior performance compared to standard SVMs with RBF kernels and linear DEM. As shown in the table below, KDEM outperformed SVM in 11 out of 12 functional classes. Notably, in classes with very few samples (e.g., Transcription, Nucleotide Synthesis, and Mitochondrial), SVM often failed completely (f-measure of 0.0), whereas KDEM maintained significant predictive power.

Comparison of f-measure for SVM-rbf, DEM, and KDEM-rbf across twelve functional classes.

Robustness to Small Training Sets

A critical advantage of the KDEM framework is its stability when labeled data is scarce. While the performance of SVM and linear DEM plummeted as the training set portion was reduced from 2/3 to 1/5, KDEM’s f-measure remained remarkably stable at approximately 90%. This suggests that the self-supervised loop successfully captures the underlying biological manifold rather than simply memorizing training samples.

Performance stability (f-measure) of KDEM vs. SVM and DEM as the portion of the training set decreases.

Discussion and Conclusion

The results highlight that nonlinearity is essential for capturing the complex boundaries of gene expression patterns. While linear DEM performed slightly better in one specific class (DNA replication), the kernel-based approach was generally more robust. The primary limitation remains the sensitivity to kernel parameter selection (e.g., the RBF kernel's width). Future work will focus on automated parameter optimization and applying this framework to other high-dimensional "omics" datasets to further validate its utility in functional genomics discovery.

Summary updated

BibTeX
@inproceedings{lu2006self,
  author = {Yijuan Lu and Qi Tian and Feng Liu and Maribel Sanchez and Yufeng Wang},
  title = {{A Self-supervised Learning Framework for Classifying Microarray Gene Expression Data}},
  booktitle = {International Conference on Computational Science},
  year = {2006},
  doi = {10.1007/11758525\_93},
}
Figure