Comparison of Computational Algorithms for the Classification of Liver Cancer using SELDI Mass Spectrometry: A Case Study

INTRODUCTION: As an alternative to DNA microarrays, mass spectrometry based analysis of proteomic patterns has shown great potential in cancer diagnosis. The ultimate application of this technique in clinical settings relies on the advancement of the technology itself and the maturity of the computational tools used to analyze the data. A number of computational algorithms constructed on different principles are available for the classification of disease status based on proteomic patterns. Nevertheless, few studies have addressed the difference in the performance of these approaches. In this report, we describe a comparative case study on the classification accuracy of hepatocellular carcinoma based on the serum proteomic pattern generated from a Surface Enhanced Laser Desorption/Ionization (SELDI) mass spectrometer. METHODS: Nine supervised classification algorithms are implemented in R software and compared for the classification accuracy. RESULTS: We found that the support vector machine with radial function is preferable as a tool for classification of hepatocellular carcinoma using features in SELDI mass spectra. Among the rest of the methods, random forest and prediction analysis of microarrays have better performance. A permutation-based technique reveals that the support vector machine with a radial function seems intrinsically superior in learning from the training data since it has a lower prediction error than others when there is essentially no differential signal. On the other hand, the performance of the random forest and prediction analysis of microarrays rely on their capability of capturing the signals with substantial differentiation between groups. CONCLUSIONS: Our finding is similar to a previous study, where classification methods based on the Matrix Assisted Laser Desorption/Ionization (MALDI) mass spectrometry are compared for the prediction accuracy of ovarian cancer. The support vector machine, random forest and prediction analysis of microarrays provide better prediction accuracy for hepatocellular carcinoma using SELDI proteomic data than six other approaches.