| The genomes of numerous organisms have been sequenced. This advancement creates new opportunities in biomedical research, specifically, for conducting large-scale studies of the behavior of tens of thousands of genes in different cell types or tissues. Microarrays are experimental tools that assay the relative abundance of transcripts from tens of thousands of genes at a time. Unfortunately, microarray data are very noisy. This thesis details procedures to collect and model the noise embedded in microarray datasets, and methods to filter it. The results lead us to an algorithm that yields highly accurate discovery of the relative abundance of genes in two biological samples.;We define a function whose zero set collects a sample of the noise embedded in a dataset. The noise sample is applied to build models and construct local filters that eliminate most of the noise remaining in the dataset. The model includes parameters that are optimized in reference to experimental data obtained from biological samples designed to yield expression ratios > 1 (true positives) or = 1 (true negatives). The algorithm offers significant improvements in both specificity and sensitivity; specificity is at least 1000-fold better and sensitivity is 2-fold higher than existing state-of-the-art methods. Highly specific discovery has numerous applications in biomedical research and medicine including: (1) molecular classification of tumors, (2) the discovery of signaling pathways, and (3) the discovery of molecular systems that create biological phenotypes. |
