(W1030-08-52) Advancing Nanocarrier Therapeutic Development with Structural AI Analysis of Cryo-EM Images

Purpose: The emergency use authorization of the Pfizer and Moderna vaccines for COVID-19 was a scientific and engineering marvel. In addition to their rapid approval and importance in preventing the spread of the virus, the use of lipid nanoparticles (LNPs) as a delivery vehicle for mRNA has brought to the forefront of pharmaceutical technology a novel mode to deliver therapeutics. As the usage of LNPs and nanocarrier platforms grows, the importance of quantifying critical quality attributes for product development and regulatory understanding is paramount. While cryo-EM has served as a primary technique to visualize particle features like size and encapsulation efficiency, current techniques to measure these attributes are often manual and human labor intensive. A more efficient technique is required to extract these quantitative attributes directly, and with reduced time from human users. Such a technique should provide rapid assessment across a larger population of particles, while also outputting quantitative measurements to correlate particle structures with their manufacturing parameters and performance.

Methods: AI image analysis using a supervised machine learning algorithm based on human user vision was implemented on cryo-EM images. An iterative process was applied to differentiate background and particles, as well as individual particle features such as the interior and the wall of each particle. Volume fractions were reported after the classification and particle size distributions were computed using an equivalent spherical diameter method on the total population of segmented particles.

Results: The machine learning algorithm used successfully differentiated the particles from their background. Within the particles themselves, AI analysis distinguished empty and filled particles within a bleb structure. Within each filled portion, the regions with brighter pixels suspected to correspond to active ingredient domains were also classified. While a powerful result, further validation is needed to evaluate the meaning of this classification and confirm that these bright signals are indeed coming from the active ingredient. The machine learning algorithm was applied to other images with equivalent features and successfully classified the particles and their features without use of a new training. The particle size distribution, volume of drug rich regions, and surface area of each particle were also computed.

Conclusion: The use of cryo-EM will continue to remain paramount in the development of lipid nanoparticles and other nanocarriers platforms. Ultimately, the imaging data is only as valuable as the quantitative results that can be extracted from them. In this work we have demonstrated that AI image analysis is capable of classifying particles separately from background, despite the low signal to noise ratio. Going further, unique features such as walls and filled vs empty particles were classified within and across particles. Once classified, these features can be quantified across dozens of images and for hundreds of particles. This image quantification can serve as a pathway to connect structural critical quality attributes with the formulation, process, and performance of the carrier.

References: Figure 1a image adapted from Brader et al. (2021)

M.L. Brader, S.J. Williams, J.M. Banks, W.H. Hui, Z.H. Zhou, L. Jin
Encapsulation state of messenger RNA inside lipid nanoparticles
Biophys. J. (2021).

Figure 1. (a) Cryo-TEM image of mRNA-LNP showing ‘bleb’ structures with variations in electron density. Image adapted from Brader et al. (2021). (b) AI classification of empty and filled portions of the LNPs.