Unlocking DOE potential by selecting the most appropriate design for rAAV optimization

Abstract

Producing recombinant adeno-associated virus (rAAV) for gene therapy via triple transfection is an intricate process involving many cellular interactions. Each of the different elements encoded in the three required plasmids-pHelper, pRepCap, and pGOI-plays a distinct role, affecting different cellular pathways when producing rAAVs. The required expression balance emphasizes the critical need to fine-tune the concentration of all these different elements. The use of design of experiments (DOE) to find optimal ratios is a powerful method to streamline the process. However, the choice of the DOE method and design construction is crucial to avoid misleading results. In this work, we examined and compared four distinct DOE approaches: rotatable central composite design (RCCD), Box-Behnken design (BBD), face-centered central composite design (FCCD), and mixture design (MD). We compared the abilities of the different models to predict optimal ratios and interactions among the plasmids and the transfection reagent. Our findings revealed that blocking is essential to reduce the variability caused by uncontrolled random effects and that MD coupled with FCCD outperformed all other approaches, improving volumetric productivity 109-fold. These outcomes underscore the importance of selecting a model that can effectively account for the biological context, ultimately yielding superior results in optimizing rAAV production.

Keywords: AAV; Box-Behnken; DOE; biomanufacturing; design of experiments; gene therapy; mixture design; optimization; recombinant adeno-associated virus; response surface methodology.

Graphical abstract

Figure 1

Comparison of models Model comparison between RCCD, FCCD, Box-Behnken design (BBD), and mixture design (MD) with the selected limits, characteristics, and statistical parameters for both log(Vp) and viability responses for each model. ^aRuns for MD are calculated by adding MD (13 runs) and FCCD (12 runs). Two additional runs are added to every model, accounting for the two required controls for qPCR.

Figure 2

RCCD 72 hpt (A) Graphic representation of the experimental space of a four-factor RCCD (left) and limits used for the model (right). (B) Normal estimate (orthog t) against normal quantile at 72 hpt showing the absolute value of the effects to identify parameters that are deviating from normality. The red line has a slope of 1, whereas the blue line passes through the origin with a slope of Lenth’s estimate of σ. (C) Comparison between actual and predicted log(Vp) at 72 hpt with the line of fit in red and the mean value in blue. (D) Residual of log(Vp) vs. predicted log(Vp) at 72 hpt to check if the values are randomly scattered around 0 (blue). (E) Residual vs. row number visualization at 72 hpt with a line at 0 (blue). (F), (G), (H), and (I) are equivalent to (B), (C), (D), and (E) for viability. (J–O) Surface plots for log(Vp) and (J) pHelper and pRepCap, (K) pHelper and FectoVIR, (L) pRepCap and FectoVIR, (M) pHelper and pGOI, (N) pGOI and pRepCap, and (O) pGOI and FectoVIR. Contour lines with log(Vp) values are at the bottom for all plots.

Figure 3

FCCD 72 hpt (A) Graphic representation of the experimental space of a four-factor FCCD (left) and limits used for the model (right). (B) Normal estimate (orthog t) against normal quantile at 72 hpt showing the absolute value of the effects to identify parameters that are deviating from normality. The red line has a slope of 1, whereas the blue line passes through the origin with a slope of Lenth’s estimate of σ. (C) Comparison between actual and predicted log(Vp) at 72 hpt with the line of fit in red and the mean value in blue. (D) Residual of log(Vp) vs. predicted log(Vp) at 72 hpt to check if the values are randomly scattered around 0 (blue). (E) Residual vs. row number visualization at 72 hpt with a line at 0 (blue). (F), (G), (H), and (I) are equivalent to (B), (C), (D), and (E) for viability. (J–O) Surface plots for log(Vp) and (J) pHelper and pRepCap, (K) pHelper and FectoVIR, (L) pRepCap and FectoVIR, (M) pHelper and pGOI, (N) pGOI and pRepCap, and (O) pGOI and FectoVIR. Contour lines with log(Vp) values are at the bottom for all plots.

Figure 4

BBD 72 hpt (A) Graphic representation of the experimental space of a four-factor BBD (left) and limits used for the model (right). (B) Normal estimate (orthog t) against normal quantile at 72 hpt showing the absolute value of the effects to identify parameters that are deviating from normality. The red line has a slope of 1, whereas the blue line passes through the origin with a slope of Lenth’s estimate of σ. (C) Comparison between actual and predicted log(Vp) at 72 hpt with the line of fit in red and the mean value in blue. (D) Residual of log(Vp) vs. predicted log(Vp) at 72 hpt to check if the values are randomly scattered around 0 (blue). (E) Residual vs. row number visualization at 72 hpt with a line at 0 (blue). (F), (G), (H), and (I) are equivalent to (B), (C), (D), and (E) for viability. (J–O) Surface plots for log(Vp) and (J) pHelper and pRepCap, (K) pHelper and FectoVIR, (L) pRepCap and FectoVIR, (M) pHelper and pGOI, (N) pGOI and pRepCap, and (O) pGOI and FectoVIR. Contour lines with log(Vp) values are at the bottom for all plots.

Figure 5

MD 72 hpt (A) Graphic representation of the experimental space of a three-factor MD (left) and limits used for the model (right). (B) Normal estimate (orthog t) against normal quantile at 72 hpt showing the absolute value of the effects to identify parameters that are deviating from normality. The red line has a slope of 1, whereas the blue line passes through the origin with a slope of Lenth’s estimate of σ. (C) Comparison between actual and predicted log(Vp) at 72 hpt with the line of fit in red and the mean value in blue. (D) Residual of log(Vp) vs. predicted log(Vp) at 72 hpt to check if the values are randomly scattered around 0 (blue). (E) Residual vs. row number visualization at 72 hpt with a line at 0 (blue). (F), (G), (H), and (I) are equivalent to (B), (C), (D), and (E) for viability. (J) Experimental space showing how both log(Vp) in blue and viability in green change at different ratios. The threshold for log(Vp) is set at 10.4 and for viability at 53%. (K) Surface plot for log(Vp), total DNA, and FectoVIR. Contour lines with log(Vp) values are at the bottom.

Figure 6

Validations and process improvement (A and B) Predicted and actual experimental values for all validated conditions at 48 and 72 hpt for all models regarding (A) log(Vp) and (B) viability. Only viability for RCCD at 72 hpt presented a significant (p = 0.01) difference between predicted and obtained values using multiple comparison two-way ANOVA. The rest were non-significant, validating the models. (C and D) (C) Obtained volumetric (Vp) productivity and (D) obtained viability for all validated optimal points in each model compared to the non-optimized condition prior to any DOE optimization. Conditions framed in black are those showing the highest process improvement and the selected approach: MD + FCCD. Fold improvements are shown for the four studied DOE approaches at 72 hpt. Error bars represent standard error of the model for the predicted values and standard error, with N = 3 for most experimental values. MD 72 hpt had N = 7 and the non-optimized condition N = 9.