A Novel Method Incorporating Gene Ontology Information for Unsupervised Clustering and Feature Selection

Shireesh Srivastava, Linxia Zhang, Rong Jin, Christina ChanPLoS 2008 · DOI

Abstract

A framework is presented that incorporates prior ontology information, which helped to (a) perform unsupervised clustering of the phenotypes, and (b) identify the genes relevant to each cluster of phenotypes. We demonstrate the proposed framework by applying it to identify physiologically-relevant feature genes that conferred differential toxicity to saturated vs. unsaturated FFAs. The framework can be applied to other problems to efficiently integrate ontology information and expression data in order to identify feature genes.

Introduction

Analyzing high-dimensional microarray data typically involves either unsupervised clustering to group samples or supervised feature selection to identify genes distinguishing known groups. However, unsupervised methods often struggle with noise, and supervised methods require pre-existing labels. Furthermore, standard workflows often relegate biological knowledge, such as Gene Ontology (GO), to post-hoc enrichment analysis. This paper introduces Integrative Mixture GO (IMGO), a Bayesian framework that integrates GO information directly into the clustering process to improve the discovery of physiologically relevant genes and phenotypes.

The IMGO Framework

The IMGO model bridges the gap between data-driven clustering and biological knowledge through three primary components:

  1. Bayesian Regression Mixture Models: The model assumes experimental conditions belong to latent subgroups, each characterized by a specific set of influential genes (regression weights).
  2. GO-Based Similarity Matrix with TF-IDF Weighting: Rather than treating genes as independent variables, IMGO utilizes a similarity matrix based on GO terms. To account for the varying specificity of biological functions, a Term Frequency-Inverse Document Frequency (TF-IDF) scheme—borrowed from information retrieval—is applied. This ensures that specific, informative functional labels carry more weight than broad, generic terms.
  3. Expectation-Maximization (EM) Algorithm: The model iteratively assigns samples to clusters and updates gene weights based on cluster membership and GO similarity until a stable solution is reached.

Case Study: Differential Lipotoxicity

The authors applied IMGO to study lipotoxicity in human hepatoblastoma (HepG2) cells. Cells were treated with saturated free fatty acids (FFAs), such as palmitate, and unsaturated FFAs, such as oleate and linoleate, alongside varying levels of the inflammatory cytokine TNF-$\alpha$. The objective was to identify the genetic drivers behind the toxicity of saturated fats compared to the relative benignity of unsaturated fats.

Key Results

  • Phenotype Discrimination: IMGO successfully separated toxic (saturated FFA) conditions from non-toxic ones in an unsupervised manner, achieving high separation scores purely by aligning patterns with known gene functions.
  • Feature Gene Identification: The model identified lysosomal ATPases and adenylate cyclase 9 (AC9) as top features associated with palmitate-induced toxicity.
  • Experimental Validation: A key prediction of the model was the involvement of the AC9/cAMP pathway. Follow-up laboratory experiments confirmed that intracellular cyclic AMP (cAMP) levels were significantly and specifically reduced by palmitate treatment, but not by non-toxic fats, validating the model's feature selection.

Discussion and Conclusion

IMGO demonstrates that incorporating functional constraints—specifically the assumption that genes within the same GO category should have similar effects—significantly reduces the search space and mitigates the impact of noise in genomic data. While the model currently assumes genes in a category act in the same direction, future iterations may incorporate more complex similarity matrices to account for activators and inhibitors. Despite the need for parameter tuning (specifically the sparsity and GO-weight parameters), IMGO provides a robust method for integrating prior biological knowledge into unsupervised learning workflows.

Summary updated

Abstract

A framework is presented that incorporates prior ontology information, which helped to (a) perform unsupervised clustering of the phenotypes, and (b) identify the genes relevant to each cluster of phenotypes. We demonstrate the proposed framework by applying it to identify physiologically-relevant feature genes that conferred differential toxicity to saturated vs. unsaturated FFAs. The framework can be applied to other problems to efficiently integrate ontology information and expression data in order to identify feature genes.
BibTeX
@inproceedings{srivastava2008novel,
  author = {Shireesh Srivastava and Linxia Zhang and Rong Jin and Christina Chan},
  title = {{A Novel Method Incorporating Gene Ontology Information for Unsupervised Clustering and Feature Selection}},
  booktitle = {PLoS},
  year = {2008},
  doi = {10.1371/journal.pone.0003860},
}
Figure