[Paper Review] Algorithm for Finding Optimal Gene Sets in Microarray Prediction

J. M. Deutsch|ArXiv.org|Aug 8, 2001

Gene expression and cancer classification23 references3 citations

TL;DR

This paper introduces GESSES, a replication-based evolutionary algorithm that identifies minimal optimal gene sets for accurate microarray-based cancer classification. Applied to leukemia and childhood cancer data, it reduced genes from 96 to just 15 while achieving perfect test classification, demonstrating a robust method for minimizing biologically irrelevant genes without sacrificing predictive accuracy.

ABSTRACT

Motivation: Microarray data has been recently been shown to be efficacious in distinguishing closely related cell types that often appear in the diagnosis of cancer. It is useful to determine the minimum number of genes needed to do such a diagnosis both for clinical use and to determine the importance of specific genes for cancer. Here a replication algorithm is used for this purpose. It evolves an ensemble of predictors, all using different combinations of genes to generate a set of optimal predictors. Results: We apply this method to the leukemia data of the Whitehead/MIT group that attempts to differentially diagnose two kinds of leukemia, and also to data of Khan et. al. to distinguish four different kinds of childhood cancers. In the latter case we were able to reduce the number of genes needed from 96 down to 15, while at the same time being able to perfectly classify all of their test data. Availability: http://stravinsky.ucsc.edu/josh/gesses/ Contact: [e-mail]

Motivation & Objective

• To identify the minimal set of genes required for accurate classification of cancer types using microarray data.
• To overcome noise and redundancy in gene expression data by evolving diverse predictor ensembles.
• To develop a method that integrates gene selection directly into the prediction process, rather than treating it as a pre-processing step.
• To evaluate whether a small, biologically meaningful gene set can achieve high predictive accuracy comparable to larger gene panels.
• To explore whether evolutionary algorithms can outperform traditional gene selection methods in identifying optimal gene subsets for cancer diagnosis.

Proposed method

• The GESSES algorithm uses a replication algorithm—a type of evolutionary computation—originally developed for quantum and protein simulations.
• It evolves an ensemble of predictors, each using a different subset of genes selected from an initial pool of candidate genes.
• Each predictor's fitness is scored based on its ability to correctly classify training samples into distinct cancer categories.
• The algorithm iteratively evolves the ensemble by adding genes to predictors, using a scoring function that balances classification accuracy and gene set size.
• A nearest-neighbor classifier is used as the prediction engine due to its simplicity, low assumptions, and fast learning on training data.
• The method employs both stochastic and deterministic evolution strategies, with convergence monitored by tracking the average number of misclassifications over generations.

Experimental results

Research questions

• RQ1 Can an evolutionary algorithm effectively identify a minimal set of genes that maintains high classification accuracy in microarray-based cancer diagnosis?
• RQ2 What is the optimal number of genes required to achieve reliable classification in complex cancer datasets like SRBCT and leukemia?
• RQ3 How does the performance of GESSES compare to traditional gene selection methods such as ranking, PCA, or gene shaving?
• RQ4 To what extent can gene redundancy and noise be mitigated by evolving diverse predictor ensembles?
• RQ5 Is it possible to achieve perfect test classification with fewer than 20 genes in high-dimensional microarray data?

Key findings

• GESSES successfully reduced the number of genes required to classify four types of childhood cancer (SRBCT) from 96 to fewer than 15, achieving perfect classification of all 20 test samples.
• The algorithm identified a stable optimal gene set size of approximately 12±2 genes for the SRBCT dataset, indicating a clear minimal effective gene set.
• For the leukemia dataset, no single optimal gene count emerged; predictors varied widely in performance, with some making zero errors and others up to five, suggesting data limitations or noise.
• The ensemble approach revealed multiple high-performing predictors, but no statistically significant improvement was found by tuning parameters like initial gene pool size or scoring method.
• Convergence to low error rates occurred rapidly in deterministic mode, with only three genes needed to reach ~2 average misclassifications in the leukemia dataset.
• The method identified biologically distinct genes compared to the original 96-gene set in the SRBCT study, suggesting potential for novel biological insights.