How to Optimize Cell Culture Media to Speed Biopharma Development
Biopharmaceutical companies today are challenged to develop high producing cell lines as quickly as possible. Commercially available media may fall short of performance expectations required to meet targets. The alternative —fully customized media and feed development — requires significant funding, time and in-house expertise in media development.
This article is posted on our Sartorius Blog.
Using a reliable design of experiments data analysis process can help biopharma companies optimize the media mixture for cell-based antibody development.
Without an appropriate growth medium, no biopharmaceutical process can exist. The proper medium allows cells to grow and generate the product. Besides the fact that cell culture processes would simply not exist without cell culture media, the medium (and associated feeds for fed-batch processes) represent a large portion of the cost of a biopharmaceutical process. Cell culture media are complex entities that encompass large quantities of raw materials as well as a high level of technical expertise, driving up the cost per liter.
To address these challenges, Sartorius-Stedim offers an advanced Design of Experiments (DOE) approach to media optimization for CHO-based processes. The CHOptimizer® Media Builder service combines a modern, high-throughput approach using the Ambr® 15 workstation with on-site support from application specialists to deliver optimized media in a short timeline.
DOE analysis using MODDE® software (and its Design and Analysis Wizards) is an important component of this process, helping to determine what the optimal mixture for the media should be.
An Example Exercise: Antibody Media Development
We can learn more about the role DOE plays by taking a look at how CHOptimizer® was used to optimize a process for a CHO cell line producing an IgG1 antibody. In the first stage, media optimization was studied by setting up a mixture design consisting of four ingredients (i.e., four different base media). Twenty-four formulations were encoded by the mixture design to match the capacity of the Ambr®15 system. Cultures were run in a simple fed-batch mode. The batches ran between 10 and 14 days. Peak viable cell density, doubling time and final IgG titer were the three responses measured following fermentation completion.
The results were used in a design of experiments (DOE) analysis to find the optimal mixture. This process included:
- Generating a mixture design using MODDE® and its Design Wizard;
- Analyzing the mixture data using the Analysis Wizard of the MODDE® software;
- Understanding how changes in media composition correlate with the responses;
- Using trilinear contour plots for model interpretation;
- Using SweetSpot plots to propose media compositions optimized for different target responses.
Four mixture factors (“base media”) were mixed according to a mixture design to create media formulations that cover a wide range of mixture component compositions. Lower and upper bounds of the mixture factors (“base media”) were [0 – 1] in all cases. There were no additional relational constraints.
Three responses relating to titer, peak viable cell density and doubling time were defined. The first two responses should be maximized and the last minimized.
Reproducing the DOE Results
In order to better understand how design of experiments works as part of an optimization process for live-cell antibody production, we can recreate the tasks.
- Task 1: Create a MODDE® project and reproduce the experimental design that was used.
- Task 2: Use the Analysis Wizard to model the three responses.
- Task 3: Evaluate how changes in media composition affect the three responses using trilinear contour plots.
The first task is to create a MODDE® project and reproduce the experimental design that was used. In this investigation, the Ambr®15 system was set up such that one standard was run thus leaving room for 23 true mixture combinations. This means the DOE protocol used in reality deviates a little from the mixture designs available in MODDE®’s list of eligible experimental designs (see screenshot below). We will therefore copy and paste the real data from an Excel-file. But first we need to specify the design framework.
Use the Design Wizard to create a new MODDE® project:
Define the factors and responses according to the information given above. Select objective Optimization (RSM), select the Modified Simplex Centroid w/Face design and increase the number of center-point to 5. Click Finish.
Upon exiting the Design Wizard a preliminary worksheet with 23 rows (23 experiments) is created. Open the file Media Development.xlsx and copy/paste the real factor and response values into the worksheet. The real worksheet is seen below.
In the case of four base media in a mixture, the design region corresponds to a tetrahedron. The graph below is a visualization of the distribution of the experimental runs in the design region. Four base media are mixed in a systematic fashion to create media formulations that cover a wide range of component concentrations.
Use the Analysis Wizard to model the three responses.
For each response judge the replicate experiments quality, response normal distribution, model quality etc.
- Are there any deviators? Which factors (“base media”) have the highest / lowest influence?
- Are the investigated factors (“base media”) influencing the three responses in the same way?
- Which mixture composition is favorable for maximizing titer? Maximizing peak VCD? Minimizing doubling time?
Results Task 2
The results for one response at a time are presented here.
Note: The replicate plot shows there is small variability among the replicates, which is encouraging and shows we have good data to work with. The histogram plot indicates that the response does not need a transformation and can be analyzed using the untransformed metric. According to the coefficient plot media M4 has a small impact on Peak titer; in order to get more titer a formulation with high amount of M1 and low amounts of M2 & M3 is desirable. The summary of fit plot, the normal probability plot of residuals and the observed versus predicted plot all point to a very good model for the titer response.
Note: The replicate plot shows there is small variability among the replicates. The histogram plot shows a slight tendency for tailing to the left; however, this tailing is not strong enough to justify a response transformation. The regression coefficient plot shows Peak VCD to depend similarly on the mixture factors as does Peak titer. The influence of the higher-order model terms is more pronounced for Peak VCD, however.
In order to increase Peak VCD, the formulation media should contain high amount of M1 and low amounts of M2 & M3. As indicated by the Summary of fit plot, N-plot of residuals and the obs/pred plot we have obtained a rather good model, although not as strong as the model for Peak titer. The close-to-zero value for Model Validity is caused by the very small replicate error and is more of a “cosmetic” issue rather than a real model problem.
D3 DT (“doubling time”)
Note: For the doubling time, the replicate plot suggests small replicate error. Additionally, we can see that there are a few runs with rather large numerical values for the doubling time, corresponding to a slight skewness to the right. This skewness is visualized in the histogram plot and indeed the skewness test is just triggered and it is a borderline case whether to log-transform or not to transform the response.
You may experiment analyzing data with and without a transformation and compare the results. In order to keep things simple our account is based on not transforming the response. Compared with the results for the two foregoing responses, the modeling results for the doubling time is a lot weaker. This is seen from the size of the confidence intervals in the coefficients plot and the comparatively low R2 and Q2. The low Q2 is primarily caused by a deviation for experiment #15, as seen in the observed versus predicted plot. Both linear and higher-order terms have some influence on the response.
Use trilinear contour plots to visualize how changes in the base media proportions correlate with changes in titer, VCD and doubling time.
Note that with a trilinear contour plot the influence of three base media can be plotted. The setting of the fourth base medium must be locked at a fixed value.
(Hint: Examine the coefficients plots produced during the previous task. Try to identify the weakest mixture factor and set it as the constant factor.)
- Where would you continue if you wanted to…
- Maximize Peak titer?
- Maximize Peak VCD?
- Doubling time?
- A suitable compromise balancing the goals? (Hint: Create a SweetSpot plot).
Results Task 3
In this task we will evaluate how changes in media composition affect the three responses using trilinear contour plots. This means one of the four mixture ingredients must be held constant. One way of identifying which mixture ingredient to use as a constant is through looking at the regression coefficients. Other means could be to consider other properties such as cost, purity, stability, half-life, toxicity or some other important property relevant for the application at hand.
The plots of regression coefficients used in the previous task jointly indicate that Media 4 is the least influential one. In the first set of trilinear contour plots given below the proportion of Media 4 = 0 and in the second set of contour plots Media 4 = 0.25. You can of course exploit other settings of Media 4 and alternative constant mixture factors, but that option is not pursued here.
Let’s return to the questions that were asked. To maximize Peak titer we would move towards the top vertex in the displayed contour plot for Peak titer (the upper, left plot). However, that co-ordinate would not satisfy the goals for Peak VCD and D3 DT (doubling time).
For the latter two responses, we should stay in the interior top-to-mid-part of contour plots. The SweetSpot plot (lower, right) can be thought of as an overlay of the three trilinear contour plots that is colored according to how many response criteria are fulfilled. The SweetSpot plot shows there is a region where the demands for all three responses are met, i.e., we comply with the limits in each case but we are not necessarily at the ideal target setting. In the SweetSpot region we have an area where we achieve a compromise for all three responses.
In order to visualize the impact of increasing the proportion of Media 4 from 0 to 0,25 we can look at the next set of trilinear plots. Evidently, increasing the amount of Media 4 is beneficial for doubling time, but not the other two responses. For this setting of Media 4, no SweetSpot region is definable.
Design of experiments (DOE) is an efficient strategy for systematically and simultaneously changing proportions of mixture ingredients and exploring the consequences on critical quality attributes.
In the current example the ranking of the four base media is such that the overall influence of M1 > M2&M3 > M4.
Based on the presented methodology three media formulations were identified:
- One media was optimized for titer production
- One media was optimized for doubling time
- One media was optimized for all three responses weighted equally
The three chosen media were used in a second phase DOE looking into spent media analysis and process development.