HB-PLS: An algorithm for identifying biological process or pathway regulators by integrating Huber loss and Berhu penalty with Partial Least Square

Gene expression data features high dimensionality, multicollinearity, and the existence of outlier or non-Gaussian distribution noise, which make the identification of true regulatory genes controlling a biological process or pathway difficult. In this study, we embedded the Huber-Berhu (HB) regression into the partial least squares (PLS) framework and created a new method called HB-PLS for predicting biological process or pathway regulators through construction of regulatory networks. PLS is an alternative to ordinary least squares (OLS) for handling multicollinearity in high dimensional data. The Huber loss is more robust to outliers than square loss, and the Berhu penalty can obtain a better balance between the l_2 penalty and the l_1 penalty. HB-PLS therefore inherits the advantages of the Huber loss, the Berhu penalty, and PLS. To solve the Huber-Berhu regression, a fast proximal gradient descent method was developed; the HB regression runs much faster than CVX, a Matlab-based modeling system for convex optimization. Implementation of HB-PLS to real transcriptomic data from Arabidopsis and maize led to the identification of many pathway regulators that had previously been identified experimentally. In terms of its efficiency in identifying positive biological process or pathway regulators, HB-PLS is comparable to sparse partial least squares (SPLS), a very efficient method developed for variable selection and dimension reduction in handling multicollinearity in high dimensional genomic data. However, HB-PLS is able to identify some distinct regulators, and in one case identify more positive regulators at the top of output list, which can reduce the burden for experimental test of the identified candidate targets. Our study suggests that HB-PLS is instrumental for identifying biological process and pathway genes.