Search

Objective Random effects are often neglected when defining the control strategy for a biopharmaceutical process. In this article, we present a case study that highlights the importance of considering the variance introduced by random effects in the calculation of proven acceptable ranges (PAR), which form the basis of the control strategy. Methods Linear mixed models were used to model relations between process parameters and critical quality attributes in a set of unit operations that comprises a typical biopharmaceutical manufacturing process. Fitting such models yields estimates of fixed and random effect sizes as well as random and residual variance components. To form PARs, tolerance intervals specific to mixed models were applied that incorporate the random effect contribution to variance. Results We compared standardized fixed and random effect sizes for each unit operation and CQA. The results show that the investigated random effect is not only significant but in some unit operations even larger than the average fixed effect. A comparison between ordinary least squares and mixed models tolerance intervals shows that neglecting the contribution of the random effect can result in PARs that are too optimistic. Conclusions Uncontrollable effects such as week-to-week variability play a major role in process variability and can be modelled as a random effect. Following a workflow such as the one suggested in this article, random effects can be incorporated into a statistically sound control strategy, leading to lowered out of specification results and reduced patient risk.

Showing analytical similarity is key to license biosimilar products with reduced or circumvented clinical effort. Statistical procedures to assess analytical similarity of quality attributes at drug product level have been highly debated by academia, industry, and regulatory agencies. In the past, a tiered approach was recommended by regulators, consisting of equivalence tests and quality range tests. However, this approach has recently been withdrawn by FDA. New guidelines of FDA and EMA favour the usage of quality range tests. Moreover, it has recently been shown that simple range tests, such as the 3SD test, are flawed, since they do not control the agency risk of falsely declaring a non-biosimilar product as being biosimilar (Type I error). This has also been highlighted by regulators recently. In this contribution, we developed a novel bootstrapping test for assessing analytical similarity that overcomes current flaws of equivalence and range tests. The developed test shows the desired properties, that is, (i) similarity conditions can be easily defined, (ii) differences of mean and the variance between the biosimilar and the innovator can be studied simultaneously, and (iii) the Type I error of the test can be controlled at a low level, e.g. 5%, evenly along the entire similarity condition. Moreover, the test shows up to 10% higher mean power values in the similarity region compared to existing range tests that aim to control the Type I error. Hence, this test is superior to existing quality range tests and is perceived compliant with current regulatory requirements.

Intermediate acceptance criteria are the foundation for developing control strategies in process validation stage 1 in the pharmaceutical industry. At drug substance or product level such intermediate acceptance criteria for quality are available and referred to as specification limits. However, it often remains a challenge to define acceptance criteria for intermediate process steps. Available guidelines underpin the importance of intermediate acceptance criteria, because they are an integral part for setting up a control strategy for the manufacturing process. The guidelines recommend to base the definition of acceptance criteria on the entirety of process knowledge. Nevertheless, the guidelines remain unclear on how to derive such limits. Within this contribution we aim to present a sound data science methodology for the definition of intermediate acceptance criteria by putting the guidelines recommendations into practice (ICH Q6B, 1999). By using an integrated process model approach, we leverage manufacturing data and experimental data from small scale to derive intermediate acceptance criteria. The novelty of this approach is that the acceptance criteria are based on pre-defined out-of-specification probabilities, while also considering manufacturing variability in process parameters. In a case study we compare this methodology to a conventional +/- 3 standard deviations (3SD) approach and demonstrate that the presented methodology is superior to conventional approaches and provides a solid line of reasoning for justifying them in audits and regulatory submission.

Integrated or holistic process models may serve as the engine of a digital asset in a multistep-process digital twin. Concatenated individual-unit operation models are effective at propagating errors over an entire process, but are nonetheless limited in certain aspects of recent applications that prevent their deployment as a plausible digital asset, particularly regarding bioprocess development requirements. Sequential critical quality attribute tests along the process chain that form output–input (i.e., pool-to-load) relationships, are impacted by nonaligned design spaces at different scales and by simulation distribution challenges. Limited development experiments also inhibit the exploration of the overall design space, particularly regarding the propagation of extreme noncontrolled parameter values. In this contribution, bioprocess requirements are used as the framework to improve integrated process models by introducing a simplified data model for multiunit operation processes, increasing statistical robustness, adding a new simulation flow for scale-dependent variables, and describing a novel algorithm for extrapolation in a data-driven environment. Lastly, architectural and procedural requirements for a deployed digital twin are described, and a real-time workflow is proposed, thus providing a final framework for a digital asset in bioprocessing along the full product life cycle.

A wide variety of computational models covering statistical, mechanistic, and machine learning (locked and adaptive) methods are explored for use in biopharmaceutical manufacturing. Limited discussion exists on how to establish the credibility of a computational model for application in biopharmaceutical manufacturing. In this work, we tried to use the American Society of Mechanical Engineers (ASME) Verification and Validation 40 (V&V 40) standard and FDA proposed AI/ML model life cycle management framework for Software as a Medical Device (SaMD) in biopharmaceutical manufacturing use cases, by applying to a set of curated hypothetical examples. We discussed the need for standardized frameworks to facilitate consistent decision making to enable efficient adoption of computational models in biopharmaceutical manufacturing and alignment of existing good practices with existing frameworks. In the study of our examples, we anticipate existing frameworks like V&V 40 can be adopted.

The calculation of metabolic turnover rates is essential for the scalable design, analysis and control of bioprocesses. A novel rate calculation algorithm is presented that is based on the dynamic adaptation of window sizes in order to deliver robust and precise rates with uniform signal-to-noise ratios. Additionally a model-based generic algorithm for deriving optimal rate calculation workflows was developed. The generic algorithms delivered more precise and accurate rates for online and offline signals, which was demonstrated for both in silico- and real batch and fed-batch fermentation process data. The presented algorithms will strongly support bioprocess development and control as enabling tools for multivariate data analysis, mechanistic modeling and dynamic experimentation.

Risk assessments (RAs) are frequently conducted to assess the potential effect of process parameters (PPs) on product quality attributes (e.g., a critical quality attribute (CQA)). To evaluate the PPs criticality the risk priority number (RPN) for each PP is often calculated. This number is generated by the multiplication of three factors: severity, occurrence, and detectability. This mathematical operation may result in some potential errors due to the multiplication of ordinal scaled values and the assumption that the factors contribute equally to the PPs criticality. To avoid these misinterpretations and to assess the out of specification (OOS) probability of the drug substance, we present a novel and straightforward mathematical algorithm. This algorithm quantitatively describes the PPs effect on each CQA assessed within the RA. The transcription of severity and occurrence to model effect sizes and parameters distribution are the key elements of the herein developed approach. This approach can be applied to any conventional RA within the biopharmaceutical industry. We demonstrate that severity and occurrence contribute differently to the PP criticality and compare these results with the RPN number. Detectability is used in a final step to precisely sort the contribution of each factor. To illustrate, we show the misinterpretation risk of the PP critically by using the conventional RPN approach.

A new and generic approach of dynamic reaction–diffusion modeling was applied to analyze the catalytic performance of an industrially used cephalosporin C amidase immobilized on a series of insoluble carriers (epoxy-activated Sepabeads) that differed in particle or pore diameter. Enzymatic conversion of cephalosporin C into 7-amino cephalosporanic acid and d -α-amino adipic acid (DAAA) results in net proton release, and this was shown previously to bring about significant internal acidification of the enzyme immobilizates during reaction. In this study, we demonstrate first time model parameter estimation to in situ measured real-time pH data, recorded both inside the enzyme carrier as well as externally in the liquid bulk of a stirred tank reactor. An essential and novel feature of the model was applicability to intraparticle pH data derived from space-averaged opto-chemical measurements. Using model-based estimation, we determined the maximal intrinsic reaction rate of the immobilized amidase completely unmasked from diffusional and carrier acidification effects (vmax, specific) as well as the effective proton diffusion coefficient ( D e f f , H + ) . Neither parameter is accessible directly from the experiment. The resulting vmax, specific estimates showed that carrier-bound amidase had lost between 14% and 55% of the free enzyme's activity due to immobilization, depending on the carrier type used. The D e f f , H + estimate (≈1 × 10−10 m2/s) was about one order of magnitude lower than expected for proton diffusion in such porous structures, suggesting a role for co-diffusion of the proton and the corresponding DAAA anion in decreasing the proton diffusivity. The model was validated against independent time-course data sets, and it was shown to give useful prediction of immobilized enzyme effectiveness factors for the applied range of carrier characteristics and enzyme loadings. Simulation was used to characterize the influence of geometrical features of the carrier on distribution of substrate and products in the carrier pore, and it is demonstrated how these intraparticle concentration gradients affect enzyme reactor performance under different process conditions.

In this paper, we propose a new strategy for retrospective identification of feed phases from online sensor-data enriched feed profiles of an Escherichia Coli (E. coli) fed-batch fermentation process. In contrast to conventional (static), data-driven multi-class machine learning (ML), we exploit process knowledge in order to constrain our classification system yielding more parsimonious models compared to static ML approaches. In particular, we enforce unidirectionality on a set of binary, multivariate classifiers trained to discriminate between adjacent feed phases by linking the classifiers through a one-way switch. The switch is activated when the actual classifier output changes. As a consequence, the next binary classifier in the classifier chain is used for the discrimination between the next feed phase pair etc. We allow activation of the switch only after a predefined number of consecutive predictions of a transition event in order to prevent premature activation of the switch and undertake a sensitivity analysis regarding the optimal choice of the (time) lag parameter. From a complexity/parsimony perspective the benefit of our approach is three-fold: i) The multi-class learning task is broken down into binary subproblems which usually have simpler decision surfaces and tend to be less susceptible to the class-imbalance problem. ii) We exploit the fact that the process follows a rigid feed cycle structure (i.e. batch-feed-batch-feed) which allows us to focus on the subproblems involving phase transitions as they occur during the process while discarding off-transition classifiers and iii) only one binary classifier is active at the time which keeps effective model complexity low. We further use a combination of logistic regression and Lasso (i.e. regularized logistic regression, RLR) as a wrapper to extract the most relevant features for individual subproblems from the whole set of high-dimensional sensor data. We train different soft computing classifiers, including decision trees (DT), k-nearest neighbors (k-NN), support vector machines (SVM) and an own developed fuzzy classifier and compare our method with conventional multi-class ML. Our results show a remarkable out-performance of the here proposed method over static ML approaches in terms of accuracy and robustness. We achieved close to error free feed phase classification while reducing the misclassification rates in 17 out of 20 investigated test cases in the range between 39% and 98.2% depending on feature set and classifier architecture. Models trained on features based on selection by RLR significantly outperformed those trained on features suggested by experts and their predictive performance was considerably less affected by the choice of the lag parameter.

In biopharmaceutical process development and manufacturing, the online measurement of biomass and derived specific turnover rates is a central task to physiologically monitor and control the process. However, hard-type sensors such as dielectric spectroscopy, broth fluorescence, or permittivity measurement harbor various disadvantages. Therefore, soft-sensors, which use measurements of the off-gas stream and substrate feed to reconcile turnover rates and provide an online estimate of the biomass formation, are smart alternatives. For the reconciliation procedure, mass and energy balances are used together with accuracy estimations of measured conversion rates, which were so far arbitrarily chosen and static over the entire process. In this contribution, we present a novel strategy within the soft-sensor framework (named adaptive soft-sensor) to propagate uncertainties from measurements to conversion rates and demonstrate the benefits: For industrially relevant conditions, hereby the error of the resulting estimated biomass formation rate and specific substrate consumption rate could be decreased by 43 and 64 %, respectively, compared to traditional soft-sensor approaches. Moreover, we present a generic workflow to determine the required raw signal accuracy to obtain predefined accuracies of soft-sensor estimations. Thereby, appropriate measurement devices and maintenance intervals can be selected. Furthermore, using this workflow, we demonstrate that the estimation accuracy of the soft-sensor can be additionally and substantially increased. Electronic supplementary material The online version of this article (doi:10.1007/s00216-016-9711-9) contains supplementary material, which is available to authorized users.

Background Enzymatic hydrolysis of cellulose involves the spatiotemporally correlated action of distinct polysaccharide chain cleaving activities confined to the surface of an insoluble substrate. Because cellulases differ in preference for attacking crystalline compared to amorphous cellulose, the spatial distribution of structural order across the cellulose surface imposes additional constraints on the dynamic interplay between the enzymes. Reconstruction of total system behavior from single-molecule activity parameters is a longstanding key goal in the field. Results We have developed a stochastic, cellular automata-based modeling approach to describe degradation of cellulosic material by a cellulase system at single-molecule resolution. Substrate morphology was modeled to represent the amorphous and crystalline phases as well as the different spatial orientations of the polysaccharide chains. The enzyme system model consisted of an internally chain-cleaving endoglucanase (EG) as well as two processively acting, reducing and non-reducing chain end-cleaving cellobiohydrolases (CBHs). Substrate preference (amorphous: EG, CBH II; crystalline: CBH I) and characteristic frequencies for chain cleavage, processive movement, and dissociation were assigned from biochemical data. Once adsorbed, enzymes were allowed to reach surface-exposed substrate sites through “random-walk” lateral diffusion or processive motion. Simulations revealed that slow dissociation of processive enzymes at obstacles obstructing further movement resulted in local jamming of the cellulases, with consequent delay in the degradation of the surface area affected. Exploiting validation against evidence from atomic force microscopy imaging as a unique opportunity opened up by the modeling approach, we show that spatiotemporal characteristics of cellulose surface degradation by the system of synergizing cellulases were reproduced quantitatively at the nanometer resolution of the experimental data. This in turn gave useful prediction of the soluble sugar release rate. Conclusions Salient dynamic features of cellulose surface degradation by different cellulases acting in synergy were reproduced in simulations in good agreement with evidence from high-resolution visualization experiments. Due to the single-molecule resolution of the modeling approach, the utility of the presented model lies not only in predicting system behavior but also in elucidating inherently complex (e.g., stochastic) phenomena involved in enzymatic cellulose degradation. Thus, it creates synergy with experiment to advance the mechanistic understanding for improved application. Electronic supplementary material The online version of this article (doi:10.1186/s13068-016-0463-8) contains supplementary material, which is available to authorized users.

not Available.