Abstract
No method has been reported to predict, even approximately, the impact of mild-to-moderate changes in several immunological parameters on resistance to infection. The ability to make such predictions would be useful in risk assessment. In addition, equations that predict host resistance on the basis of changes in components of a complex biological system (the immune system) would fulfill one of the major goals of systems biology. In this study, multiple machine learning classification methods were used to predict the effects of a series of drugs and chemicals on host resistance to Listeria monocytogenes in mice on the basis of changes in several holistic immunological parameters. A data set produced under the sponsorship of the National Toxicology Program (NTP) was used in this study. The NTP data set was found to have a high percentage of missing data and to be noisy (probably due to the intrinsically stochastic nature of immune responses). Data preprocessing steps were used to mitigate these problems. In evaluating the machine learning classifiers, we first randomly partitioned the NTP data set into 10 subsets. Each time, we used nine subsets of the data to train the machine learning classifiers, and the remaining single subset to predict outcomes with regard to host resistance. This process was repeated until all 10 combinations of the 9-1 split of the subsets have been tested. The best of the classifiers predicted host resistance outcome correctly for 94.7% of cases, a result which indicates it is possible to identify mathematical expressions that will be useful for risk assessment and to establish a basis for systems immunology.