- Brian Patrick Bradley: Department of Biological Sciences, University of Maryland Baltimore County, Baltimore, MD 21250, USA. bbradley@umbc.edu
Single biomarkers are rarely accurate. Even suites of biomarkers can give conflicting results. Ideally potent combinations of variables are isolated which accurately identify specific analytes and their level of toxicity. The search for such combinations can be done by reducing the thousands of candidate variables to the small number necessary for treatment classification. When the key variables are recognized by machine learning (ML) the results are quite surprising, given the apparent failure of other searching methods to produce good diagnostics. Proteins seem especially useful for portable field tests of a variety of adverse conditions. This review shows how ML, in particular artificial neural networks, can find potent biomarkers embedded in any type of expression data, mainly proteins in this article. A computer does multiple iterations to produce sets of proteins which systematically identify (to near 100% accuracy) the treatment classes of interest. Whether these proteins are useful in actual diagnoses is tested by presenting the computer model with unknown classes. Finding the biomarkers is getting easier but there still must be confirmation, by multivariable statistics and with field studies.