Abstract
BACKGROUND: Molecular measurements from cancer patients such as gene expression and DNA methylation can be influenced by several external factors. This makes it harder to reproduce the exact values of measurements coming from different laboratories. Furthermore, some cancer types are very heterogeneous, meaning that there might be different underlying causes for the same type of cancer among different individuals. If a model does not take potential biases in the data into account, this can lead to problems when trying to predict the stage of a certain cancer type. This is especially true when these biases differ between the training and test set. RESULTS: We introduce a method that can estimate this bias on a per-feature level and incorporate calculated feature confidences into a weighted combination of classifiers with disjoint feature sets. In this way, the method provides a prediction that is adjusted for the potential biases on a per-patient basis, providing a personalized prediction for each test patient. The new method achieves state-of-the-art performance on many different cancer data sets with measured DNA methylation or gene expression. Moreover, we show how to visualize the learned classifiers to display interesting associations with the target label. Applied to a leukemia data set, our method finds several ribosomal proteins associated with the risk group, which might be interesting targets for follow-up studies. This discovery supports the hypothesis that the ribosomes are a new frontier in genadaptivelearninge regulation. CONCLUSION: We introduce a new method for robust prediction of phenotypes from molecular measurements such as DNA methylation or gene expression. Furthermore, the visualization capabilities enable exploratory analysis on the learnt dependencies and pave the way for a personalized prediction of phenotypes. The software is available under GPL2+ from https://github.com/adrinjalali/Network-Classifier/tree/v1.0 .