The use of Raman spectroscopy for characterizing critical process parameters (CPPs) and critical quality attributes (CQAs) in bioprocessing has grown considerably during the last 10 years. Raman provides rich, chemical data and it’s largely a non-invasive and non-destructive technology, at least in terms of monitoring a bioprocess.

Most process monitoring implementations with Raman use an empirical multivariate calibration approach (sometimes called a “black box” method). Observational data on bioprocesses and the reference data collected with it are combined with multivariate methods to build a predictive model.

There are several advantages to this traditional approach, including the wide availability of user-friendly software and short courses to help build the models. Also, one does not need an in-depth understanding of the chemistry of the system under study, or the detailed physics/electronics of the measurement platform itself.

However, bioprocesses present some specific challenges for empirical multivariate calibration. One of the most significant is extreme covariates: the environment within a bioreactor is constantly changing, often in a highly interrelated way, making it very difficult for empirical models to isolate the right amount of spectral information from the covariates. Cross-validation is routinely used to objectively evaluate the performance of empirical models, but it hinges on the assumption of exchangeability of samples and that is severely violated in bioprocess data.

Another challenge for common empirical calibration methods is their assumption of stationarity, which means that the covariance structure is constant. This is also never the case for bioprocesses or for Raman spectrometers, making it difficult to apply empirical methods in a very robust way.  Lastly, key analytical figures of merit such as selectivity, sensitivity and precision can only be generally inferred by experience/test which leads to a lot of trial and error. In addition, every spectrometer is a little bit different so what works well on one spectrometer may not work well on another unit.

All the factors above have contributed to the somewhat notorious difficulty in producing Raman-based measurements that generalize well across process conditions, media systems, cell lines, and spectrometers. This was a problem that the 908 Devices team was very interested in tackling.

We recently launched MAVERICK, a Raman-based in-line analyzer that monitors and controls multiple bioprocess parameters. In terms of spectrometer details, it’s a very typical 785 nm Raman spectrometer. But we’ve built a turnkey platform designed to generalize very well across mammalian CHO and HEK cells and many different media and cell-line systems.

We used a different approach from empirical calibration; we refer to it as a de novo model, which means “from the beginning.” It’s more of a first principles approach that blends hard chemistry and physics information with Bayesian components representing the measuring instrument, measurement error physics/electronics, and knowledge of cell culture systems. No part of the de novo model is empirically calibrated/trained—it has never learned from a bioprocess. There are dynamic components to the de novo model as well, which react to ambient light and autofluorescence changes throughout a process.

It’s all de novo from the ground up and no empirical training on bioprocess runs is used, making it easier to avoid the challenges noted above with an empirical calibration approach.

Learn more about MAVERICK, our Raman-based PAT solution that was named one of the top 15 innovations in 2023 by The Analytical Scientist.

Download the application note on 908 Devices and Culture Biosciences’ collaboration to enable the digital transformation of biomanufacturing processes for Process Development.