Protein therapeutics are a highly successful class of drugs that are currently used to treat a number of serious and life-threatening conditions such as cancer, autoimmune disorders, and infectious diseases including COVID-19. These therapeutics have numerous critical quality attributes (CQA) that must be evaluated to ensure safety and efficacy, including that they must adopt and retain the correct structural fold without forming unintended aggregates. The entirety of structural elements from primary sequence to quaternary interactions is termed by the industry as the ‘higher order structure’ (HOS) of the therapeutic, and the development of analytical techniques for HOS characterization throughout the lifecycle of a protein therapeutic, from development to manufacture, has therefore emerged as a major priority in the pharmaceutical industry. To address this measurement gap, our group has developed, demonstrated, and optimized one-dimensional (1D) and two-dimensional (2D) nuclear magnetic resonance (NMR) spectroscopy methods as robust and fit-for-purpose approaches for spectral ‘fingerprinting’ the HOS of protein therapeutics and their formulations. To utilize these NMR fingerprints in HOS decision making, I will describe how nonlinear iterative partial lease squares (NIPALS) based principal component analysis (PCA) with quantitative similarity assessment using PC Euclidean distance and cluster confidence surface overlap can be employed for automated and quantitative classification of one-dimensional (1D) diffusion-edited 1H spectra and 2D 1H-13C methyl spectra as demonstrated using measurements on an IgG1k NIST reference mAb (NISTmAb), and mAbs from biopharma partners. Results from the new approach, which has been dubbed PROtein Fingerprint Obtained by NIPALS Decomposition, or PROFOUND, show that class separation is highly tolerant of low signal/noise (S/N). It therefore overcomes the disadvantages of earlier methods, such as PROFILE, which employed convolution difference and linear correlation statistics to assess spectral similarity that both reduce the operative signal-to-noise (S/N) while simultaneously rendering the method potentially sensitive to that S/N. While the approach is robust with respect to processing details and analysis region selection, systematic imperfections such as baseline distortion can have a pronounced effect, therefore measurement technique must be controlled carefully. Since this classification approach can be performed without the need to identify signals, results suggest that it is possible to use even more efficient measurement strategies that do not produce spectra that can be analyzed visually, but nevertheless allow useful decision-making that can be timely, objective and automated.
