DNA Microarrays: A Case Study

The great challenge in DNA microarrays problems consists in finding the genes that are relevant to a particular disease or cell state in the midst of thousands of other genes. And thinking of genes as variables, for tackling these problems successfully, in an ideal world, one would have required about an order of magnitude more of samples. However, in a typical DNA microarray study, the number of samples available (a few dozens or a few hundreds at best), when compared to the thousands of genes under study, is ridiculously small, which obviously poses some challenges.

First, it is common to try and narrow down the search space with sophisticated discretization algorithms to filter out the noise or irrelevant genes. And although good results have been reported (data size reductions of 50%-98%), the problem still remains: hundreds or thousands of variables to be mined using just a few dozen samples.

The great advantage of using GeneXproTools in DNA microarray studies is that you can use the raw data (obviously you can also use the filtered data and even use GeneXproTools to filter out the noise) and still obtain excellent results. Obviously this means that not all the models with a good accuracy on the training data will have good predictive accuracy, but by making several runs one can select the top 10-20 models and then cross-reference the genes (attributes) used in all of them. You can then select and copy these most important genes from the Data Panel and create a much smaller dataset for creating the final model.

Let's illustrate this with real-world DNA microarray data, using the well-studied ALL-AML Leukemia datasets (these same datasets are used in the DNA Microarray sample run of GeneXproTools 4.0, and we recommend you play with it as you'll be able to see everything including the generated code). In this problem, the training dataset consists of 38 bone marrow samples (27 ALL and 11 AML), over 7129 probes from 6817 human genes. And the testing dataset consists of 34 samples, with 20 ALL and 14 AML. For this analysis, "0" was used to represent "ALL" and "1" to represent "AML" and the 7129 genes were numbered d₀-d₇₁₂₈.

For instance, in one study, for the ALL-AML Leukemia problem, a total of 308 promising genes were identified in 25 good runs (that is, in this case runs with 100% training accuracy and testing accuracies between 91.18%-97.06%). Of these 308 promising genes, only 11 (genes 759, 1881, 2287, 2407, 4362, 4846, 5485, 6040, 6587, 6638, and 6854) appeared in two or more models; and of these, only five (genes 759, 1881, 2287, 4846, and 6854) appeared in more than three models, with the most prevalent being genes 1881, 2287, and 4846, with eight, six, and four appearances, respectively. So, it is a good guess that the genes mostly to be involved in ALL-AML leukemia are genes 1881, 2287, and 4846, which is an exceptionally good starting point for tackling leukemia.