Case study · With Thermo Fisher Scientific
Scale-up with historical data & in silico process optimization using DataHowLab
A study with Thermo Fisher Scientific tested whether hybrid models in DataHowLab can predict across scales and optimize a feed strategy in silico, starting from a legacy dataset with few large-scale runs.
-
+15%
Higher titer than the best previous 5 L run, with the model-recommended feed recipe
-
1
Validation run instead of a full campaign to confirm the in silico optimized recipe
-
Titer predicted at 50 L and 500 L from small-scale data alone, with lower error than MLR
The challenge
Transferring learnings across scales
Upstream development of CHO-based fed-batch processes is resource-intensive, with full campaigns spanning many months, up to 1.5 years. Small-scale systems such as the 15 mL ambr often do not cover the operating conditions of 50 L to 500 L bioreactors, and classical approaches like multiple linear regression fail to extrapolate reliably across scales, while fully mechanistic models are time-consuming to build and require deep a priori process knowledge.
The approach
Hybrid models trained on a legacy multi-scale dataset
The dataset held 28 bioreactor runs across seven scales, from 15 mL to 500 L, with ExpiCHO-S cells. Hybrid models were trained on it in DataHowLab to identify key process features and design optimized conditions in silico, and the result was verified in vitro on a 5 L glass bioreactor. The whole workflow, from data formatting to in vitro verification, took approximately 3 months.
- 28 bioreactor runs
- 7 scales up to 500 L
- ~3 months, end to end
Get the full results
Download the full case study, with project outline, analysis, and case results.