Abstract
Proteins are often considered the main biological element in charge of the different functions and structures of a cell. However, proteomics, the global study of all expressed proteins, often performed by mass spectrometry, is limited by its stochastic sampling and can only quantify a limited amount of protein per sample. Transcriptomics, which allows an exhaustive analysis of all expressed transcripts, is often used as a surrogate. However, the transcript level does not present a high level of correlation with the corresponding protein level, notably due to the existence of several post-transcriptional regulatory mechanisms. In this publication, we hypothesize that the missing protein values in proteomics could be predicted using machine learning regression methods, trained with many features extracted from transcriptomics, including known translational regulatory elements such as microRNAs and circular RNAs. After considering different machine learning algorithms applied on two different splitting strategies, we report that random forest can predict proteins in new samples out of transcriptomics data with good accuracy. The proposed pre-processing and model building scripts can be accessed on GitHub: https://github.com/jochotecoa/ml_proteomics.