Data Augmentation to Support Biopharmaceutical Process Development through Digital Models—A Proof of Concept

Abstract

In recent years, monoclonal antibodies (mAbs) are gaining a wide market share as the most impactful bioproducts. The development of mAbs requires extensive experimental campaigns which may last several years and cost billions of dollars. Following the paradigm of Industry 4.0 digitalization, data-driven methodologies are now used to accelerate the development of new biopharmaceutical products. For instance, predictive models can be built to forecast the productivity of the cell lines in the culture in such a way as to anticipate the identification of the cell lines to be progressed in the scale-up exercise. However, the number of experiments that can be performed decreases dramatically as the process scale increases, due to the resources required for each experimental run. This limits the availability of experimental data and, accordingly, the applicability of data-driven methodologies to support the process development. To address this issue in this work we propose the use of digital models to generate in silico data and augment the amount of data available from real (i.e., in vivo) experimental runs, accordingly. In particular, we propose two strategies for in silico data generation to estimate the endpoint product titer in mAbs manufacturing: one based on a first principles model and one on a hybrid semi-parametric model. As a proof of concept, the effect of in silico data generation was investigated on a simulated biopharmaceutical process for the production of mAbs. We obtained very promising results: the digital model effectively supports the identification of high-productive cell lines (i.e., high mAb titer) even when a very low number of real experimental batches (two or three) is available.

Hybrid modeling in bioprocess dynamics: Structural variabilities, implementation strategies, and practical challenges

“…In this indirect method, the loss function is not directly linked to the neural network outputs (Pinto et al, 2022). As the data‐driven model is trained indirectly in the optimization framework, a regularization term can also be included in the fitness function (E wr ) to improve training convergence (Botton et al, 2022; Yang et al, 2011) as shown in Equation (6). $$${E}_{wr}=\frac{1}{n\times m}\sum _{i=1}^{n}\sum _{j=1}^{m}\frac{{({\xi }_{ij}-{\hat{\xi }}_{ij})}^{2}}{{\sigma }_{{\xi }_{i}}^{2}}+\lambda \Vert g\Vert ,$$$where λ is the regularization parameter (weight), and g is the parameter for the data‐driven component.…”

“…This exercise can be repeated unless the projected best process conditions get stabilized (Figure 6). In silico generation of simulated batch experiments with a semi‐parametric hybrid has been shown to improve product titer, while deciphering the correlation between critical process parameters and critical quality attributes (Botton et al, 2022).…”

“…The rate of production/consumption of culture variables (i.e., biomass, substrate, product) can be modeled as a product of maximum specific rates, stoichiometric coefficients (reaction rates on the cell concentration), and specific production/consumption rates estimated by the ANN from reduced concentration vector as input (Botton et al, 2022). A single hidden layer containing 10 neurons, hyperbolic-tangent activation function, linear output transfer function, and minimization of the normalized sum of squared errors with regularization term in training convergence was adopted.…”

“…In this indirect method, the loss function is not directly linked to the neural network outputs (Pinto et al, 2022). As the data-driven model is trained indirectly in the optimization framework, a regularization term can also be included in the fitness function (E wr ) to improve training convergence (Botton et al, 2022;Yang et al, 2011) as shown in Equation ( 6).…”