Using Deep Neural Networks and LASSO Regression to Predict miRNA Expression Changes Based on mRNA Data

Franz Leonard Böge¹Helena U Zacharias²Stefanie Christine Becker¹Klaus Jung^1*

• ¹University of Veterinary Medicine Hannover, Hanover, Germany
• ²Hannover Medical School, Hanover, Lower Saxony, Germany

Since the rise of molecular high-throughput technologies, many diseases are now studied on multiple omics layers in parallel. Understanding the interplay between microRNAs (miRNA) and their target mRNAs is important to understand the molecular level of diseases. While for many diseases, lots of public data from mRNA experiments is available, few paired data sets with both, miRNA and mRNA expression profiles, are available. This study aimed to assess the possibility of predicting miRNA expression data based on mRNA expression data, serving as a proof of principle that such cross-omics predictions are feasible. Furthermore, current research relies on target databases where information about miRNA-target-relationships is provided based on experimental and computational studies.To make use of publicly available mRNA profiles, we investigate the ability of artificial deep neural networks and linear Least Absolute Shrinkage and Selection Operator (LASSO) regression to predict unknown miRNA expression profiles. We evaluate the approach using seven paired miRNA/mRNA expression data sets, four from studies on West Nile Virus infection in mice tissues and three from Human Immunodeficiency Virus infection in human tissues. We assessed the performance of each model first by within data evaluations and second by cross study evaluations. Furthermore, we investigated if data augmentation or separate models for data from diseased and non-diseased samples can improve prediction performance.In general most settings achieved strong correlations on the level of individual samples. In some data sets and settings, also correlations of log fold changes and p-value from DEA between true and predicted miRNA profiles can be observed. Correlation between log fold changes could also be seen in a cross-study evaluation for the Human Immunodeficiency Virus data sets. Data augmentation consistently improved performance in neural networks, while its impact on Lasso models was not significant. Overall, cross-omics prediction of expression profiles appears possible, even with some correlations on the level of the differential expression analysis.