PRELIMINARY ARTIFICIAL NEURAL NETWORK ANALYSIS OF SELDI MASS SPECTROMETRY DATA FOR THE CLASSIFICATION OF MELANOMA TISSUE

摘要

Over recent years studies have shown an increasing application of bioinformatics tools, and in particular, artificial intelligence techniques such as artificial neural networks (ANNs) for biological problems. Proteomic techniques such as SELDI-MS (Surface Enhanced Laser Desorption/Ionization Mass Spectrometry) may be used to distinguish patterns derived from diseased tissue used to identify biomarkers representative of a certain pathological state. This paper describes research building upon studies by Ball et al. (2002) in identifying potential biomarkers using ANNs for the analysis of mass spectrometry data from melanoma tissue. This approach utilised ANNs to model for 72 melanoma tissue samples (36 stage 1 and 36 stage 4) to identify ions as potential biomarkers and any possible interactions between these. Preliminary results have shown that approximately 20,000 inputs can be screened to just 20 molecular ions which are capable of accurately predicting tumour grade. Using the additive approach described by Ball et al. (2002) individual molecular ions were used to predict tumour grade and model performance evaluated using a receiver operating characteristic (ROC) curve. The accuracy of the model with 1 ion was 58.3% with a sensitivity of 63.9% and specificity of 52.8%. With a 2 ion model, ANN performance increased to an accuracy of 70.8%, sensitivity of 66.7% and specificity of 75%. With a 3 and 4 ion model, the accuracy increased yet further to 77.8 and 83.3%, sensitivity to 72.2 and 77.8% and specificity of 83.3 and 88.9% respectively. Preliminary findings indicate the ANN approaches adopted allow optimisation and determination of the minimum number of ions (derived from SELDI-MS data) which can successfully predict tumour grade. This work continues so that these key ions may be determined in order to identify if they have an important role in tumour progression from low to high grade.