Comparison of Machine Learning Classifiers on Integrated Transcriptomic Data

Omics data are being generated for different conditions, and can be a valuable resource for building novel predictive models for medical diagnosis. Given the reduced number of samples in each dataset, the application of Machine Learning (ML) models requires data integration. At the same time, multiple ML models are available, and the best option for data integration is not known. These challenges have been addressed typically in restricted settings, i.e., for one single disease at a time. However, a thorough comparison of models on integrated data, for different conditions, is still missing. In this paper we confront 7 classifiers on integrated data for 6 diseases, over 14 datasets. We compared the models on single and integrated datasets, employing different pre-processing techniques. We also evaluated the effect of feature selection, analyzing the robustness and relevance of the features extracted. We observed that, even if integration slightly reduces predictive power, the models are still able to produce good classifications. When testing generalization abilities on new datasets, sometimes the performance decreases drastically, depending on the disease studied.