Raman model development for the protein conformational state classification in different freeze-dried formulations.

The aim of this work is to build a multivariate calibration (MVC) model from Raman spectra for the prediction of the protein conformational state class (i.e. native-like or non-native) in different freeze-dried pharmaceutical formulations of a model protein lactate dehydrogenase (LDH). As this model would be intended to facilitate and better understand formulation and process development, it should allow acceptable classification performance despite variations in formulation type and batch. Therefore, it was attempted to (1) find which factors interfere the Raman spectra, (2) understand them, and (3) make the MVC model robust for them. A variance analysis within the Raman spectral data space identified significant spectral background variations among certain formulation types and batches in the studied samples. Raw material (i.e. LDH) batch variability and the presence of a Maillard reaction in formulations were the main reasons for this. We demonstrate the successful use of both exhaustive calibration and external parameter orthogonalization (EPO) pre-processing for making the Raman classification model more robust for the expected spectral interferences.