RobNorm: Model-Based Robust Normalization for High-Throughput Proteomics from Mass Spectrometry Platform

In the analysis of proteomics data from mass spectrometry (MS), normalization is an important preprocessing step to correct sample-level variation and make abundance measurements for each specific protein comparable across different samples. Under heterogenous samples such as in the Phase I study of the Enhancing Genotype-Tissue Expression (eGTEx) project (Jiang, et al., 2019), the samples coming from 32 different tissues, and without prior housekeeping protein information or spike-ins, how to robustly correct the bias but keep tissue internal variations becomes a challenging question. Majority of previous normalization methods cannot guarantee a robust and tissue adaptive correction. This motivates us to develop a data-driven robust normalization method in MS platform to adapt tissue sample heterogeneities. To robustly estimate the sample effect, we take use of the density power weight to down weigh the outliers and extend the one-dimensional robust fitting in (Windham, 1995) and (Fujisawa and Eguchi, 2008) to our structured data. We construct our robust criterion and build the algorithm to get our robust normalization (RobNorm). We focus our comparison to the PQN a widely used normalization method in MS. In the simulation studies and real data application, we conclude that our robust normalization method to estimate the sample effect performs better than PQN especially when the regulation magnitude and proportion are large and strong. We also discuss some limitations in our method.