Accelerating mechanistic model calibration in protein chromatography using artificial neural networks

In the manufacturing of therapeutic monoclonal antibodies (mAbs), mechanistic models can aid the evaluation and selection of suitable chromatography operating conditions during process development. However, model calibration remains a common bottleneck for model implementation in industrial settings. To accelerate the calibration process, the present study proposes a semi-automated, artificial neural network (ANN)-assisted calibration workflow for the efficient estimation of model parameters. The workflow is applied for the calibration of the multicomponent kinetic formulation of the steric mass-action (SMA) isotherm model for cation exchange chromatography (CEX). Three case studies using two mAb feedstocks of differing complexity regarding their structure and impurities are investigated. Different combinations of training data (low and/or high load density) and parameter groupings (one-step approach for estimation of all parameters simultaneously; two-step approach for estimation of (1) equilibrium and charge followed by (2) effective mass transfer coefficient, kinetic, and shielding) are applied for the target compounds and impurities. The ANN-assisted calibration workflow provided acceptable model parameter estimations and subsequent good agreement between experimental and simulated chromatograms with only minimal refinement by inverse fitting for the target compounds of both feedstocks. For the impurities, the one-step parameter estimation approach showed satisfactory prediction quality only for the simple feedstock. For the complex feedstock, the two-step approach using only high loading data improved parameter prediction for both the impurities and the target compound. The observed reduction in calibration effort suggests great potential for ANN applications to facilitate mechanistic model calibration, thus enhancing and streamlining downstream process development for complex antibodies.