(204c) Deep Learning Method to Understand the Mechanisms of Fouling during Bioreactor Harvesting

Monoclonal antibodies and other protein therapeutics are often produced through the cultivation of mammalian cells such as Chinese hamster ovary cells (CHO). Currently cell densities above 100 million/mL with over 10 g/L product titers can be reached with perfusion bioreactors. While perfusion bioreactors have a significant advantage over traditional fed-batch reactors, the extended production time leads to fouling of the membrane filter during clarification and cell retention / recycle which reduces productivity in terms of flux, throughput, and product recovery.

It has been shown that operating the perfusion reactor using alternating tangential flow filtration (ATF) is able to reduce fouling (by periodically reversing the direction of the feed flow into the module), but membrane fouling remains a challenge. Mem­brane fouling is very complex and is affected by the properties of the feed and the membrane as well as the operating conditions. Previous studies have indicated that CHO cells, cell debris, host cell proteins (HCPs) and DNA as well as anti-foam can all contribute to membrane fouling, however, it is not clear what are the dominant factors controlling membrane fouling during ATF bioreactor harvesting. In addition, the interplay between the feed and operating conditions on filter fouling is not well understood.

Here proteomic methods were used to characterize and quantify the extracted HCP and IgG molecules fouled on the filters during bioreactor harvesting from both laboratory and industrial ATF perfusion runs. These proteomic methods include 2D SDS PAGE coupled with mass spectroscopy. Over 200 HCP molecules were identified. Based on the identity and amount of these HCP molecules fouled on the filter, statistical and deep learning approaches were used to determine key factors leading to the fouling of the membrane during industrial operations.