Application of recurrent neural network for online prediction of cell density of recombinant Pichia pastoris producing HBsAg

From Spatial-Temporal Multiscale Modeling to Application: Bridging the Valley of Death in Industrial Biotechnology

The Valley of Death confronts industrial biotechnology with a significant challenge to the commercialization of products. Fortunately, with the integration of computation, automation and artificial intelligence (AI) technology, the industrial biotechnology accelerates to cross the Valley of Death. The Fourth Industrial Revolution (Industry 4.0) has spurred advanced development of intelligent biomanufacturing, which has evolved the industrial structures in line with the worldwide trend. To achieve this, intelligent biomanufacturing can be structured into three main parts that comprise digitalization, modeling and intellectualization, with modeling forming a crucial link between the other two components. This paper provides an overview of mechanistic models, data-driven models and their applications in bioprocess development. We provide a detailed elaboration of the hybrid model and its applications in bioprocess engineering, including strain design, process control and optimization, as well as bioreactor scale-up. Finally, the challenges and opportunities of biomanufacturing towards Industry 4.0 are also discussed.

Integrating mechanistic and deep learning models for accurately predicting the enrichment of polyhydroxyalkanoates accumulating bacteria in mixed microbial cultures

The enrichment of polyhydroxyalkanoates (PHA) accumulating bacteria (PAB) in mixed microbial cultures (MMC) is extremely difficult to be predicted and optimized. Here we demonstrate that mechanistic and deep learning models can be integrated innovatively to accurately predict the dynamic enrichment of PAB. Well-calibrated activated sludge models (ASM) of the PAB enrichment process provide time-dependent data under different operating conditions. Recurrent neural network (RNN) models are trained and tested based on the time-dependent dataset generated by ASM. The accurate prediction performance is achieved (R² > 0.991) for three different PAB enrichment datasets by the optimized RNN model. The optimized RNN model can also predict the equilibrium concentration of PAB (R² = 0.944) and corresponding time, which represents the end of the PAB enrichment process. This study demonstrates the strength of integrating mechanistic and deep learning models to predict long-term variations of specific microbes, helping to optimize their selection process for PHA production.

Comparison of partial least square, artificial neural network, and support vector regressions for real-time monitoring of CHO cell culture processes using in situ near-infrared spectroscopy

The biopharmaceutical industry must guarantee the efficiency and biosafety of biological medicines, which are quite sensitive to cell culture process variability. Real-time monitoring procedures based on vibrational spectroscopy such as near-infrared (NIR) spectroscopy, are then emerging to support innovative strategies for retro-control of key parameters as substrates and by-product concentration. Whereas monitoring models are mainly constructed using partial least squares regression (PLSR), spectroscopic models based on artificial neural networks (ANNR) and support vector regression (SVR) are emerging with promising results. Unfortunately, analysis of their performance in cell culture monitoring has been limited. This study was then focused to assess their performance and suitability for the cell culture process challenges. PLSR had inferior values of the determination coefficient (R² ) for all the monitored parameters (i.e., 0.85, 0.93, and 0.98, respectively for the PLSR, SVR, and ANNR models for glucose). In general, PLSR had a limited performance while models based on ANNR and SVR have been shown superior due to better management of inter-batch heterogeneity and enhanced specificity. Overall, the use of SVR and ANNR for the generation of calibration models enhanced the potential of NIR spectroscopy as a monitoring tool.

Machine learning applied for metabolic flux-based control of micro-aerated fermentations in bioreactors

Various bio-based processes depend on controlled micro-aerobic conditions to achieve a satisfactory product yield. However, the limiting oxygen concentration varies according to the micro-organism employed, while for industrial applications, there is no cost-effective way of measuring it at low levels. This study proposes a machine learning procedure within a metabolic flux-based control strategy (SUPERSYS_MCU) to address this issue. The control strategy used simulations of a genome-scale metabolic model to generate a surrogate model in the form of an artificial neural network, to be used in a micro-aerobic fermentation strategy (MF-ANN). The meta-model provided setpoints to the controller, allowing adjustment of the inlet air flow to control the oxygen uptake rate. The strategy was evaluated in micro-aerobic batch cultures employing industrial Saccharomyces cerevisiae yeast, with defined medium and glucose as the carbon source, as a case study. The performance of the proposed control scheme was compared with a conventional fermentation and with three previously reported micro-aeration strategies, including respiratory quotient-based control and constant air flow rate. Due to maintenance of the oxidative balance at the anaerobiosis threshold, the MF-ANN provided volumetric ethanol productivity of 4.16 g·L^-1 ·h^-1 and a yield of 0.48 g_ethanol .g_substrate ^-1 , which were higher than the values achieved for the other conditions studied (maximum of 3.4 g·L^-1 ·h^-1 and 0.35-0.40 g_ethanol ·g_substrate ^-1 , respectively). Due to its modular character, the MF-ANN strategy could be adapted to other micro-aerated bioprocesses.

History and Evolution of Modeling in Biotechnology: Modeling & Simulation, Application and Hardware Performance

Biological systems are typically composed of highly interconnected subunits and possess an inherent complexity that make monitoring, control and optimization of a bioprocess a challenging task. Today a toolset of modeling techniques can provide guidance in understanding complexity and in meeting those challenges. Over the last four decades, computational performance increased exponentially. This increase in hardware capacity allowed ever more detailed and computationally intensive models approaching a "one-to-one" representation of the biological reality. Fueled by governmental guidelines like the PAT initiative of the FDA, novel soft sensors and techniques were developed in the past to ensure product quality and provide data in real time. The estimation of current process state and prediction of future process course eventually enabled dynamic process control. In this review, past, present and envisioned future of models in biotechnology are compared and discussed with regard to application in process monitoring, control and optimization. In addition, hardware requirements and availability to fit the needs of increasingly more complex models are summarized. The major techniques and diverse approaches of modeling in industrial biotechnology are compared, and current as well as future trends and perspectives are outlined.

Risk prediction platform for pancreatic fistula after pancreatoduodenectomy using artificial intelligence

BACKGROUND

Despite advancements in operative technique and improvements in postoperative managements, postoperative pancreatic fistula (POPF) is a life-threatening complication following pancreatoduodenectomy (PD). There are some reports to predict POPF preoperatively or intraoperatively, but the accuracy of those is questionable. Artificial intelligence (AI) technology is being actively used in the medical field, but few studies have reported applying it to outcomes after PD.

AIM

To develop a risk prediction platform for POPF using an AI model.

METHODS

Medical records were reviewed from 1769 patients at Samsung Medical Center who underwent PD from 2007 to 2016. A total of 38 variables were inserted into AI-driven algorithms. The algorithms tested to make the risk prediction platform were random forest (RF) and a neural network (NN) with or without recursive feature elimination (RFE). The median imputation method was used for missing values. The area under the curve (AUC) was calculated to examine the discriminative power of algorithm for POPF prediction.

RESULTS

The number of POPFs was 221 (12.5%) according to the International Study Group of Pancreatic Fistula definition 2016. After median imputation, AUCs using 38 variables were 0.68 ± 0.02 with RF and 0.71 ± 0.02 with NN. The maximal AUC using NN with RFE was 0.74. Sixteen risk factors for POPF were identified by AI algorithm: Pancreatic duct diameter, body mass index, preoperative serum albumin, lipase level, amount of intraoperative fluid infusion, age, platelet count, extrapancreatic location of tumor, combined venous resection, co-existing pancreatitis, neoadjuvant radiotherapy, American Society of Anesthesiologists’ score, sex, soft texture of the pancreas, underlying heart disease, and preoperative endoscopic biliary decompression. We developed a web-based POPF prediction platform, and this application is freely available at http://popfrisk.smchbp.org.

CONCLUSION

This study is the first to predict POPF with multiple risk factors using AI. This platform is reliable (AUC 0.74), so it could be used to select patients who need especially intense therapy and to preoperatively establish an effective treatment strategy.

Biomass soft sensor for a Pichia pastoris fed-batch process based on phase detection and hybrid modeling

A common control strategy for the production of recombinant proteins in Pichia pastoris using the alcohol oxidase 1 (AOX1) promotor is to separate the bioprocess into two main phases: biomass generation on glycerol and protein production via methanol induction. This study reports the establishment of a soft sensor for the prediction of biomass concentration that adapts automatically to these distinct phases. A hybrid approach combining mechanistic (carbon balance) and data-driven modeling (multiple linear regression) is used for this purpose. The model parameters are dynamically adapted according to the current process phase using a multilevel phase detection algorithm. This algorithm is based on the online data of CO₂ in the off-gas (absolute value and first derivative) and cumulative base feed. The evaluation of the model resulted in a mean relative prediction error of 5.52% and R² of .96 for the entire process. The resulting model was implemented as a soft sensor for the online monitoring of the P. pastoris bioprocess. The soft sensor can be used for quality control and as input to process control systems, for example, for methanol control.

Accurate and cost-effective prediction of HBsAg titer in industrial scale fermentation process of recombinant Pichia pastoris by using neural network based soft sensor

In the current work, the attempt was made to apply best-fitted artificial neural network (ANN) architecture and the respective training process for predicting final titer of hepatitis B surface antigen (HBsAg), produced intracellularly by recombinant Pichia pastoris Mut⁺ in the commercial scale. For this purpose, in large-scale fed-batch fermentation, using methanol for HBsAg induction and cell growth, three parameters of average specific growth rate, biomass yield, and dry biomass concentration-in the definite integral form with respect to fermentation time-were selected as input vectors; the final concentration of HBsAg was selected for the ANN output. Used dataset consists of 38 runs from previous batches; feed-forward ANN 3:5:1 with training algorithm of backpropagation based on a Bayesian regularization was trained and tested with a high degree of accuracy. Implementing the verified ANN for predicting the HBsAg titer of the five new fermentation runs, excluded from the dataset, in the full-scale production, the coefficient of regression and root-mean-square error were found to be 0.969299 and 2.716774, respectively. These results suggest that this verified soft sensor could be an excellent alternative for the current relatively expensive and time-intensive analytical techniques such as enzyme-linked immunosorbent assay in the biopharmaceutical industry.