In recent studies involving NAPPA microarrays, extra-well fluorescence is used as a key measure for identifying disease biomarkers because there is evidence to support that it is better correlated with strong antibody responses than statistical analysis involving intraspot intensity. Because this feature is not well quantified by traditional image analysis software, identification and quantification of extra-well fluorescence is performed manually, which is both time-consuming and highly susceptible to variation between raters. A system that could automate this task efficiently and effectively would greatly improve the process of data acquisition in microarray studies, thereby accelerating the discovery of disease biomarkers. In this study, we experimented with different machine learning methods, as well as novel heuristics, for identifying spots exhibiting extra-well fluorescence (rings) in microarray images and assigning each ring a grade of 1-5 based on its intensity and morphology. The sensitivity of our final system for identifying rings was found to be 72% at 99% specificity and 98% at 92% specificity. Our system performs this task significantly faster than a human, while maintaining high performance, and therefore represents a valuable tool for microarray image analysis.
Keywords: bioinformatics; biomarker; image analysis; nucleic acid programmable protein array (NAPPA); protein array.