Using Decision Trees and Support Vector Machines to Classify Genes by Names

In this paper we report an application of machine learning methods to classify gene names into two categories: known and unknown ones. We acquired a data set of 1,624 genes by letting a human expert classify them manually. To capture the knowledge of classification, we also asked the expert to derive a set of rules. In parallel, we trained two machine learners to capture the same knowledge. Both decision trees (CART) and Support Vector Machines (SVMs) outperform the expert rules; the cross-validated error rates are below 1%, and the area under the curve of Receiver Operating Characteristic (ROC) curves reach higher than 0.99. In summary, CART and SVMs reduced the overall error rate of prediction by 40% and 88%, respectively, compared with classification using expert rule sets. In addition, the machine classifiers are able to find some errors made by the human expert himself. Finally, we used the expert system to classify 7,447 genes on the Affymetrix U74A microarray chip. Results show 70% of the genes on this chip are known ones. In conclusion, we successfully demonstrate that the machine-derived classifiers are more capable of handling the job efficiently than the expertderived classifier. It further supports the idea that in many application domains, experts can perform the task, but cannot tell how; whereas expert systems are able to capture the knowledge from the experts.