Modern bioprocess development employs statistically optimized design of experiments (DOE) and regression modeling to find optimal bioprocess set points. Using modeling software, such as JMP Pro, it is possible to leverage artificial neural networks (ANNs) to improve model accuracy beyond the capabilities of regression models. Herein, we bridge the gap between a DOE skill set and a machine learning skill set by demonstrating a novel use of DOE to systematically create and evaluate ANN architecture using JMP Pro software. Additionally, we run a mammalian cell culture process at historical, one factor at a time, standard least squares regression, and ANN-derived set points. This case study demonstrates the significant differences between one factor at a time bioprocess development, DOE bioprocess development and the relative power of linear regression versus an ANN-DOE hybrid modeling approach.

Background Culturing cells as cell spheres results in a tissue-like environment that drives unique cell phenotypes, making it useful for generating cell populations intended for therapeutic use. Unfortunately, common methods that utilize static suspension culture have limited scalability, making commercialization of such cell therapies challenging. Our team is developing an allogeneic cell therapy for the treatment of lumbar disc degeneration comprised of discogenic cells, which are progenitor cells expanded from human nucleus pulposus cells that are grown in a sphere configuration. Methods We evaluate sphere production in Erlenmeyer, horizontal axis wheel, stirred tank bioreactor, and rocking bag format. We then explore the use of ramped agitation profiles and computational fluid dynamics to overcome obstacles related to cell settling and the undesired impact of mechanical forces on cell characteristics. Finally, we grow discogenic cells in stirred tank reactors (STRs) and test outcomes in vitro (potency via aggrecan production and identity) and in vivo (rabbit model of disc degeneration). Results Computation fluid dynamics were used to model hydrodynamic conditions in STR systems and develop statistically significant correlations to cell attributes including potency (measured by aggrecan production), cell doublings, cell settling, and sphere size. Subsequent model-based optimization and testing resulted in growth of cells with comparable attributes to the original static process, as measured using both in vitro and in vivo models. Maximum shear rate (1/s) was maintained between scales to demonstrate feasibility in a 50 L STR (200-fold scale-up). Conclusions Transition of discogenic cell production from static culture to a stirred-tank bioreactor enables cell sphere production in a scalable format. This work shows significant progress towards establishing a large-scale bioprocess methodology for this novel cell therapy that can be used for other, similar cell therapies.