Iterative hybrid model based optimization of rAAV production

Abstract

Changes in serotype or genetic payload of recombinant adeno associated virus (rAAVs) gene therapies require adapting the transfection conditions of the upstream HEK293 cultivations. This study adopts an iterative model-based experiment design approach, where increasing data availability is leveraged to evolve models of different complexity. Initial models based on data from shaker flask runs guided the design of the first round at Ambr250 scale. With Ambr250 data becoming available, hybrid models capturing process state evolutions and historical models incorporating these evolutions to predict rAAV titer, were developed. These models were then combined into a full model approach, which was utilized within a Bayesian Optimization framework for the design of a second round of Ambr250 scale runs. The iterative approach was tested across different projects applying transfer learning to enhance the predictive power and improve the subsequent optimization. The approach was benchmarked against a statistical Design of Experiment method. The results show that the model-based experiment design consistently (and across projects) produces higher rAAV titer values than the benchmark approach (Project C: 4.4% or 7.0% increases in titer values relative to the response surface modeling approach for ELISA and ddPCR, respectively; Project D: 32.4% or 10.9% increases in titer values relative to the standard DoE-screening pick for ELISA and ddPCR, respectively), effectively optimizing the transfection mixture composition. The combination of propagation and historical models, augmented by transfer learning and an ever-increasing amount of data, enhanced the process design workflow, contributing to improved rAAV production through efficient transfection strategies.

Keywords: Design of Experiments; Parallel Mini‐bioreactors; human embryonic kidney suspension cell; hybrid modeling; rAAV production.

FIGURE 1

Chronologic order of all four campaigns conducted at the Ambr scale. For each campaign, the available data at the point of creating the design is highlighted. The number of runs that were successfully performed are specified in the top row. The model types used in the design of a campaign are shown at the bottom in the red boxes.

FIGURE 2

Decision tree model on shaker flask runs of project C targeting ddPCR.

FIGURE 3

Elastic net model results for ELISA and ddPCR respectively for shaker flask campaign C1. LOOCV was used to evaluate model performance. Replicate pairs are treated as one group for the purpose of the cross‐validation. The rRMSE is depicted for every pair of replicates in the validation set.

FIGURE 4

Observed versus Predicted plots for the Elastic Nets. The experiments in the validation set are listed individually. The Depicted values for ELISA and ddPCR are Z‐score normalized.

FIGURE 5

Experimental results of ELISA (left) and ddPCR (right) for campaign C1. The analytics data was normalized according to Z‐Score normalization. The runs are sorted in ascending order. Each point is labeled by a marker denoting the method used for the design of the experimental condition. Manual: Using decision trees and a linear model; Hypercube: Corner point repetition from shaker flask scale; Center Point: Center point of hypercube; Screening DoE – Pick: Manual selection based on the screening DoE.

FIGURE 6

Decision tree model on shaker flask runs of project D.

FIGURE 7

Experimental results of ELISA (left) and ddPCR (right) for campaign D1. The analytics data was normalized according to Z‐Score normalization. The runs are sorted in ascending order. Each point is labeled by a marker denoting the rationale behind the selection of the experimental conditions. Manual: Using decision trees; Hypercube: Corner point repetition from shaker flask scale; Center Point: Center point of hypercube; Screening DoE – Pick: Manual selection based on the screening DoE for internal control of process performance.

FIGURE 8

Results of modeling for the C2 campaign design. Figure (a) shows the rRMSE of the propagation model for all X variables in the model, (b) the rRMSE of the full model for ELISA and ddPCR, and figures (c) and (d) depict the observed versus predicted values of the full model for ELISA and ddPCR, respectively. The latter demonstrates the model's capability to predict the viral titers from initial conditions and design parameters with sufficient accuracy.

FIGURE 9

Optimizer suggestions of the experimental conditions of the main process design parameters Z0, Z1 and Z2 for campaign C2. The model predicted ddPCR for these conditions given the reference run is displayed in the fourth column. Risk level and reference run were varied. Rows with bold font were selected conditions and labeled with numeric identifiers.

FIGURE 10

Z‐Score normalized experimental analytics results (left: ELISA, right: DdPCR) for campaign C2. The analytics results are sorted in ascending order. Each point is labeled by a marker denoting the rationale behind the selection of the experimental conditions. Manual: Adaptation based on optimizer suggestion; Optimizer: Design parameters directly proposed by the Bayesian optimizer; Center Point: Center point of hypercube; Surface Model: State of the art methodology.

FIGURE 11

rRMSE results of modeling the D1 campaign without (orange) and with (blue) project C data, respectively. The addition of project C data provides clear benefit in terms of predictive performance of both viral titers, as seen by the decrease in rRMSE.

FIGURE 12

Optimizer suggestions of the experimental conditions of the main process design parameters Z0, Z1 and Z2 for campaign D2. The model predicted ddPCR for these conditions given the reference run is displayed in the fourth column. Risk level and reference run were varied. Rows with bold font were selected conditions and labeled with numeric identifiers.

FIGURE 13

Z‐Score normalized experimental analytics results (left: (a), right: B) for campaign D2. The analytics results are sorted in ascending order. Each point is labeled by a marker denoting the rationale behind the selection of the experimental conditions. Manual: Adaptation based on optimizer suggestion; Optimizer: Design parameters directly proposed by the Bayesian optimizer; Center Point: Center point of hypercube; Screening DoE – Pick: Manual selection based on the screening DoE for internal control of process performance.

FIGURE A1

Noise to signal ratios of shaker flask experiments of project C for both ELISA and ddPCR. The noise to signal ratio for ELISA is lower than for ddPCR and has less variability, however both methods show good replicability.

FIGURE A2

Noise to signal ratios of shaker flask experiments of project D for both ELISA and ddPCR. As for project C, the noise to signal ratio for ELISA is lower than for ddPCR. Compared to project C, the noise to signal ratios for project D are much larger, suggesting more intrinsic variability in the process.

FIGURE A3

Noise to signal ratios of Ambr experiments of project C for both ELISA and ddPCR. As for project C, the noise to signal ratio for ELISA is lower than for ddPCR. Compared to project C, the noise to signal ratios for the Ambr scale for project C are comparable to shake flask.

FIGURE A4

Noise to signal ratios of Ambr experiments of project D for both ELISA and ddPCR. As for project C, the noise to signal ratio for ELISA is lower than for ddPCR. Compared to project C, the noise to signal ratios for the Ambr scale for project C are comparable to shake flask.