In previous studies, highly accurate single and multi pair classifiers were generated in diverse cancer classification tasks. How ever, the extent to which other gene pairs are able to dis criminate between phenotypic states, beyond these top performing classifiers, has never been determined. We iteratively generated TSP and k TSP classifiers on the same GIST/LMS microarray data originally used to derive the OBSCN/PRUNE2 classifier. This dataset contains 68 clinical tissue specimens, and was assessed using a micro array with 43,931 oligonucleotide probes. Following each application of the search algorithm, we removed the top scoring pair of genes from the dataset, and then repeated the algorithm to determine the accuracy of each classifier derived from the reduced dataset without the original best scoring gene pairs.

As seen in Figure 1, the TSP and k TSP algorithms retain appreciable cross validation accu racy even after the removal of multiple top scoring gene pair classifiers, though there are reductions in perform ance. The presence of accurate non overlapping transcrip tional classifiers was also observed in other datasets examined using this process of iterative reduction of the probe set. The combinatoric k TSP classifier achieved higher predictive accuracy upon removal of top scoring pairs than the single pair TSP classifier. Two Transcript Classifier Accuracies in Diverse Diagnostic Tasks The top scoring pair algorithm generated classifiers that discriminate between diverse phenotypic states with vari ous degrees of apparent accuracy.

We also exam ined the performance of the combinatoric k TSP method on these datasets, and found that it outperformed the sin gle pair TSP method on some of the datasets using classi fiers involving three to seven gene pairs. Cross validation accuracies, an estimation of algorithm per formance on novel data, compared well with apparent top accuracy, with most LOOCV accuracies observed to be above 85%. A lower classifier performance in cross validation when the apparent accuracy is high does not necessarily imply that the functional accuracy of the algorithm for a particular phenotype separation is limited, but rather that the sample sizes obtained in these micro array studies may not be sufficient to determine the most accurate gene pairs for diagnosis.

Of course, in instances where the observed Brefeldin_A cross validation accuracy is low, there is low confidence in the ability of the selected TSP to accu rately classify future samples. Sensitivity and specificity were found to vary with the dataset. Two cases exhibit markedly low sensitivity cardi omyopathy and transformed colitis. This was likely due to the comparatively smaller number of positive than negative tissue samples present in these microarray datasets, which serves as an implicit weight for the algo rithm to selectively choose classifiers exhibiting correct negative diagnoses over correct positive diagnoses.