Abstract

Microarray expression data is being generated by the gigabyte all over the world with undoubted exponential increases to come. Annotated genomic data is also rapidly pouring into public databases. Our goal is to develop automated ways of combining these two sources of information to produce insight into the operation of cells under various conditions. Our approach is to use machine-learning techniques to identify characteristics of genes that are up-regulated or down-regulated in a particular microarray experiment. We seek models that are (a) accurate, (b) easy to interpret, and (c) stable to small variations in the training data. This paper explores the effectiveness of two standard machine-learning algorithms for this task: ï (based on probability) and PF (based on building rules). Although we do not anticipate using our learned models to predict expression levels of genes, we cast the task in a predictive framework, and evaluate the quality of the models in terms of their predictive power on genes held out from the training. The paper reports on experiments using actual E. coli microarray data, discussing the strengths and weaknesses of the two algorithms and demonstrating the trade-offs between accuracy, comprehensibility, and stability.

Article Outline

1. Introduction

2. Algorithm descriptions

2.1. Naïve Bayes

2.2. PF

2.3. Improving stability of PF

3. Experimental methodology

4. Experimental results and discussion

5. Discussion

6. Related work

7. Future directions and conclusions

Acknowledgements

References