Deep neural network for prediction and control of permeability decline in single pass tangential flow ultrafiltration in continuous processing of monoclonal antibodies

Abstract

Single-pass tangential flow filtration (SPTFF) is a crucial technology enabling the continuous manufacturing of monoclonal antibodies (mAbs). By significantly increasing the membrane area utilized in the process, SPTFF allows the mAb process stream to be concentrated up to the desired final target in a single pass across the membrane surface without the need for recirculation. However, a key challenge in SPTFF is compensating for flux decline across the membrane due to concentration polarization and surface fouling phenomena. In continuous downstream processing, flux decline immediately impacts the continuous process flowrates. It reduces the concentration factor achievable in a single pass, thereby reducing the final concentration attained at the outlet of the SPTFF module. In this work, we develop a deep neural network model to predict the NWP in real-time without the need to conduct actual NWP measurements. The developed model incorporates process parameters such as pressure and feed concentrations through inline sensors and a spectroscopy-coupled data model (NIR-PLS model). The model determines the optimal timing for membrane cleaning steps when the normalized water permeability (NWP) falls below 60%. Using SCADA and PLC, a distributed control system was developed to integrate the monitoring sensors and control elements, such as the NIRS sensor for concentration monitoring, the DNN model for NWP prediction, weighing balances, pressure sensors, pumps, and valves. The model was tested in real-time, and the NWP was predicted within <5% error in three independent test cases, successfully enabling control of the SPTFF step in line with the Quality by Design paradigm.