Downstream processing

Process: strong cation exchange chromatography in “bind & elute” mode

Product: complex viral antigen (assembly of 5 proteins)

Results: The objective of this study was to compare mechanistic modeling with traditional Design of Experiments (DoE) methods in quantifying how Critical Process Parameters—specifically elution pH and salt concentration—influence Critical Quality Attributes such as purity and yield. These two approaches produced markedly different outcomes (Fig. 1). When compared to experimental data points, the mechanistic model demonstrated greater precision than the statistical model.

Fig. 1: Contour maps generated by mechanistic and statistical models in terms of yield and purity

The trends predicted by the mechanistic model were found to be more consistent with physical expectations than those generated by the statistical approach (Fig. 2). For example, for elution salt concentration versus yield, the mechanistic model accurately predicts that increasing salt concentration leads to higher yields up to a maximum of 100%, corresponding to complete product elution. In contrast, the statistical model predicts a bell-shaped curve where yield decreases at high salt concentrations—a result that is not physically plausible. This unrealistic trend is due to limitations in analytical precision and constraints imposed by fitting only linear or quadratic models.

Fig. 2: Trends of the yield as a function of NaCl concentration in elution buffer generated by the mechanistic and statistical models

Furthermore, this study confirms that statistical models should not be used for extrapolation beyond experimentally studied regions. Specifically, yield values exceeding 100% and falling below 0% were predicted in pH and salt concentration ranges that were not experimentally tested. A comparison summary of the two approaches is presented in the table below.

Key message: establishing trends of CPPs vs CQAs is more reliable and requires less experiments with mechanistic modeling than with statistical DoE approaches.

Further reading: Purification of a Viral Protein by Ion-Exchange Chromatography: Methodology for Mechanistic Modeling and Comparison with a Statistical Approach - Ypso Facto

Process: weak cation exchange chromatography

Product: serum albumin (S.A.) and serum transferrin (S.T.)

Results: The aim of this case study is to demonstrate how simulation can assist in identifying better separation conditions. Two proteins, serum albumin (S.A.) and serum transferrin (S.T.), were purified on a weak cationic resin using a pH step gradient. The column was first equilibrated with an acetate/phosphate buffer in a 2:1 ratio at pH 4.8 with 40 mM Na+. Subsequently, a step gradient was applied using the same buffer at pH 7.0. A significant overlap between the two peaks is observed (Fig. 1).

Fig. 1: Initial separation between serum albumin and serum transferrin

Investigations aiming at improving separation performances began by performing the pH step gradient with buffers alone. It was found that phosphate and acetate behave very differently (Fig. 2). The rise from pH 4.8 to pH 7.0 occurs rather steeply after around 8 bed volumes in the case of phosphate. In contrast, the pH rise is much smoother and takes much longer in the case of acetate, with the inlet pH not being reached even after 25 bed volumes. Remarkably, the simulations (lines) describe the experimental data (symbols) well.

Fig. 2: Outlet pH profiles obtained with a step gradient from pH 4.8 to 7.0 with acetate and phosphate buffer solutions

Next, the study focused on the proteins alone. The theoretical curve of the charge versus pH was calculated based on the sequence of amino acids (Fig. 3). It was found that the two proteins have very similar charges, which explains the difficult separation. However, there is a sweet spot at around pH 6, where the charges of the two proteins differ. This suggests that maintaining a pH around 6 in the column is expected to better exploit the differences between the two proteins and thus improve the separation. Considering the graph above, this can be achieved by increasing the acetate/phoshphate ratio.

Fig. 3: Calculated protein charge as a function of pH

Experiments confirmed that separation could be enhanced by simply increasing the acetate/phosphate ratio to 20:1, without changing the resin, which is often the first approach considered for improving separation (Fig. 4, left). While increasing the acetate/phosphate ratio significantly improves yield, it negatively impacts the Process Mass Intensity (PMI) due to the longer separation time and increased buffer usage (Fig. 4, middle and left). The model was then employed to find an optimum in terms of yield and PMI by evaluating different acetate/phosphate ratios and initial pH values. It was determined that maintaining the same initial pH and increasing the acetate/phosphate ratio to 2:1 resulted in the best PMI with maximum yield.

Fig. 4: Improved separation conditions

Key message: Our ion exchange model can predict the impact of a change in buffer type

Further reading: Application note: prediction of the impact of buffer type with mechanistic simulation

Process: hydrophobic interaction chromatography (HIC)

Product: several proteins

Results: The aim of this case study was to showcase the ability of a hydrophobic interaction chromatography (HIC) mechanistic model to predict the effects of pH and ionic strength on process performance. Initially, a model was developed to forecast how these two process parameters influence the retention time of two model proteins (Fig. 1). It is observed that retention first decreases (salting-in) and then increases (salting-out) with ionic strength. Additionally, retention is highest near the protein's isoelectric point (pI). The simulations (lines) align well with the experimental data (dots).

Fig. 1: Impact of ionic strength and solution pH on two the retention of two model proteins

In a second step, the model was applied to a protein produced at GSK (Fig. 2). Experiments were conducted using a One Factor At a Time approach for model calibration. Subsequently, simulations were employed to predict the impact of pH and ionic strength in a full factorial design. Two purity grades were considered: the first one with the removal of low molecular weight (LMW) species eluting before the main peak, the second one without removing the LMW species. Process conditions could be identified to improve product recovery, productivity and solvent consumption for both purity grades, as presented in the figure below.

Fig. 2: Simulations performed for a protein produced at GSK

Key message: mechanistic modeling can be used to study the impact of process parameters on key performance indicators

Further reading: Purification of protein by HIC: Mechanistic modeling for improved process understanding and optimization

Process: mixed-mode chromatography (combining ion-exchange and hydrophobic interactions)

Product: antibody fragment

Results: The aim of this work was to develop a mixed mode model for protein purification. The case study deals with the removal of a low molecular weight impurity (LMW, 99 kDa) from an antibody fragment (111 kDa). The protein charge as a function of pH was calculated for both the product and LMW impurity from the theoretical amino acid sequence (Fig. 1). It is seen in the figure below that the two species have very similar charge profiles, thus making the separation difficult.  

Fig. 1: Calculated protein charge as a function of pH

The mechanistic model developed in this study combines ion exchange and hydrophobic interaction models embedded in Ionic. It was calibrated using minimal experimental data, involving a series of pulse injections (e.g., very small injections) at different salt concentrations and two pH values (Fig. 2). In the case of pure ion exchange behaviour, the retention of the protein in pulse experiments is expected to vary linearly with salt concentration when plotted on a log-log scale. The figure below shows that the experimental data deviate from ion exchange behaviour (symbols), and the developed mixed-mode model accurately describes the experimental data (dashed lines). The model parameters were fine-tuned using one reference purification experiment with fraction analysis.

Fig. 2: Experimental and calculated retention times as a function of salt concentration

The model was used to predict the impact of two critical process parameters—loading volume and NaCl concentration in the wash—on four KPIs: purity, yield, productivity, and Process Mass Intensity (PMI), which is directly related to eluent consumption (Fig. 3). A low NaCl concentration (around 25 mM) improves purity but drastically reduces yield. While loading has little effect on purity and yield, increasing it enhances productivity and PMI due to reduced eluent consumption.

Fig. 3: Contour maps obtained with the mechanistic model

The operating conditions independently chosen by Sanofi during process development using the traditional Design of Experiments (DoE) approach are marked by stars on the contour plots. The model validates these choices, achieving sufficient purity, satisfactory yield, and acceptable productivity and PMI values. Notably, the mechanistic model's results were derived from just a series of pulse experiments (which don't require fraction analysis) and a single purification experiment (which does). In contrast, the classical DoE approach demands tens of purification experiments.

Key message: Mechanistic modeling can be used to speed up process development

Further reading: Speeding-up process development of mixed-mode chromatography with mechanistic modeling