(474d) Machine Learning-Enabled Fingerprint Raman Microscopy Monitoring Pipeline for Biomanufacturing Processes

There has been a rising interest in renewable biocentric manufacturing pipelines for their many advantages such as high stereoselectivity, moderate reaction temperature and pressure, non-toxic processing chemicals and reduced carbon emissions.^[1] Due to the limits of our monitoring technologies and of our knowledge on cellular culture dynamics, most bioprocesses are empirically operated. A thorough monitoring system is vital in the development of robust system models, which will aid in further system control and optimization.

Raman spectroscopy has emerged as a promising analytical tool as it can provide label-free non-destructive chemical information. Previous works have demonstrated utilizing calibration curves for monitoring key external metabolites, but these methods are only viable over the design space in which they are trained.^{[2] [3]} A particularly promising spectroscopic technique is the Broadband Coherent Anti-Stokes Raman Scattering (BCARS), which can obtain hyperspectral images, with cellular level spatial resolution and high spectral resolution at much faster time scales than its spontaneous counterparts.^[4]

In this talk we will show how hyperspectral BCARS data in combination with computer vision algorithms, clustering algorithms, classification algorithms and spectral chemometrics analytics gives us a clear picture of the bioprocess. The data extracted from the above mentioned hyperspectral BCARS datasets should aid in the construction of sophisticated unstructured-segregated model. Pseudomonas putida was selected as a case study organism as it not only produces a broad range of value-added chemicals but is also shown to be relatively better at resisting effects of toxicity.^[5] Utilizing dimensionality reduction techniques, the three-dimensional (spatial-spectral) hyperspectral dataset is compressed to a two-dimensional (spatial) image. Computer vision algorithms are used to distinguish and identify pixels in which there are cells and medium and then further cluster together multiple cell-labelled pixels which belong to a single cell. With the aid of clustering and classification algorithms, we identify various metabolic states of the bacteria and the overall metabolic state of the culture.^[6] Using self-modeling mixture analysis techniques, we determine the component Raman profiles and their corresponding concentrations.^[7]

We will show that the presented pipeline will not only enable streamlined high-throughput data collection and monitoring. The given data can be used downstream for improved model understanding and robust model construction. These are necessary steps for effective control strategies, process design and optimization solutions to maximize throughput robustly.

References:

[1] J. M. Clomburg, A. M. Crumbley, and R. Gonzalez, “Industrial biomanufacturing: The future of chemical production,” Science, vol. 355, no. 6320, p. aag0804, Jan. 2017, doi: 10.1126/science.aag0804.

[2] “Comparison of multivariate data analysis techniques to improve glucose concentration prediction in mammalian cell cultivations by Raman spectroscopy | Elsevier Enhanced Reader.” Accessed: Mar. 15, 2023. [Online]. Available: https://reader.elsevier.com/reader/sd/pii/S0731708518305880?token=5C250E...

[3] K. A. Esmonde-White, M. Cuellar, C. Uerpmann, B. Lenain, and I. R. Lewis, “Raman spectroscopy as a process analytical technology for pharmaceutical manufacturing and bioprocessing,” Anal. Bioanal. Chem., vol. 409, no. 3, pp. 637–649, 2017, doi: 10.1007/s00216-016-9824-1.

[4] C. H. Camp Jr et al., “High-speed coherent Raman fingerprint imaging of biological tissues,” Nat. Photonics, vol. 8, no. 8, pp. 627–634, Aug. 2014, doi: 10.1038/nphoton.2014.145.

[5] M.-R. Park et al., “Response of Pseudomonas putida to Complex, Aromatic-Rich Fractions from Biomass,” ChemSusChem, vol. 13, no. 17, pp. 4455–4467, Sep. 2020, doi: 10.1002/cssc.202000268.

[6] C. García-Timermans et al., “Discriminating Bacterial Phenotypes at the Population and Single-Cell Level: A Comparison of Flow Cytometry and Raman Spectroscopy Fingerprinting,” Cytometry A, vol. 97, no. 7, pp. 713–726, 2020, doi: 10.1002/cyto.a.23952.

[7] Willem. Windig and Jean. Guilment, “Interactive self-modeling mixture analysis,” Anal. Chem., vol. 63, no. 14, pp. 1425–1432, Jul. 1991, doi: 10.1021/ac00014a016.