Computation and predictive modeling to increase efficiency and performance in cell line and bioprocess development

Abstract

A critical early step in the development of a new biopharmaceutical is the selection of the master cell bank. Per FDA requirements, the same master cell bank must be used for all toxicity and clinical trials, as well as all production of the drug should it be commercialized. Developing a master cell bank is a time and labor-intensive process where thousands of clones are screened through a series of experiments. The Berkeley Lights Beacon® platform can be used as a high-throughput screening tool in cell line development and has been shown to produce clonally-derived cell lines, suitable for the development of a master cell bank. In a typical use case, a Berkeley Lights chip is loaded with 1750 cells, data is collected related to cell growth and on-chip assays, and the top 50-100 are selected for further analysis. The methodology for selecting the top clones, however, is not standardized and individual users may select different top clones based on how they weigh the growth and assay data. As a relatively new tool, there is little literature outlining how to best use data collected on Berkeley Lights to select the "best" clones for further screening. In this project, we use Amgen's database of Berkeley Lights experiments to determine which parameters are most predictive of performance in future fed-batch experiments. Data from 9 chips (N=13,900 pens; N=305 fed-batch experiments) was analyzed using linear and non-linear machine learning models to identify feature importance and improve cell selection methodology. The models generated show an improved ability to rank top clones compared to the currently methodology, a finding that is expected to improve average clone quality in cell line development.