Is SVM a black box? Visualization and interpretation of SVR models to analyze glycosylation in pharmaceutical proteins (mAbs, Fc) based on infrared spectroscopy FT-IR

Almost 60% of commercial pharmaceutical proteins are glycosylated. They are considered a critical quality attribute as it affects the stability, bioactivity and safety of proteins. Hence, the regulatory authorities require systematic characterization of the composition and structure of glycoproteins throughout all development processes. Currently existing methods are time-consuming, expensive and require significant sample preparation steps which can alter the robustness of the analyses. Thus, we have suggested the use of a fast, direct and simple method based on Fourier transform infrared spectroscopy (FT-IR) and Support Vector Regression (SVR) [1]. SVR method is considered as a powerful alternative to Partial least squares (PLSR), thanks to its ability to flexibly model nonlinear relationships. Usually, SVR are applied as a black box method, unlike the PLSR where we can interpret the models by evaluating their loading. However, in the present study, we have investigated the methodology of visualizing and interpreting SVR models in order to quantify glycosylation [2].