Vytautas Galvanauskas

In intensive care, the arterial oxygen saturation of a ventilated patient is followed through the bedside monitor’s pulse oximetry reading, which describes the present situation only; whether it will hold over the coming hours is left to the attention of the staff. Soft sensors exist for intermittently measured variables, but none can estimate the saturation hours ahead without a recent reading of it. We ask whether the continuously recorded ventilator and expired-gas signals can estimate saturation nine hours after the last available reading. Using 4017 hourly records of 31 ventilated children, we build a virtual sensor as three nested implementation scenarios of one code base: X0 (XGBoost 2.1.4 through our pipeline, no physics term active), XP (X0 plus a blend toward the last reading and the cumulative oxygen uptake rate over the gap per kilogram of body mass, derived from an oxygen balance of the arterial blood), and v4 (XP plus a patient-level anchor that recalibrates the nine-hour-old reading toward the patient’s own history). Out-of-fold predictions under patient-grouped five-fold cross-validation are scored, with paired patient bootstrap intervals, on the 3730 h of 30 patients whose inputs contain no saturation value younger than nine hours. From X0 to XP, the mean per-patient MAPE decreases from 3.313% to 3.282%, resolved for this summary only; from XP to v4, it decreases to 3.115%, the pooled MAPE decreases from 3.885% to 3.729% (both resolved), the pooled correlation rises from 0.754 to 0.793, hypoxaemia below 90% at the end of the gap is detected with an AUC of 0.939 against 0.900 for the last reading, and the ensemble shrinks from 55 to 19 inputs and from 9637 to 474 leaves. An exact decomposition of every prediction shows that the two-coefficient anchor earns 82% of the improvement over persistence; most of the nine-hour information is the patient’s own level, the gas-derived terms add a small but resolved contribution, and the sensor is a trend indicator rather than an alarm.

Arterial oxygen saturation measured by co-oximetry (SaO2m) is the clinical reference for oxygenation, but in the mechanically ventilated patients studied here the freshest usable value is nine hours old. We build a virtual sensor that bridges this gap from continuously recorded ventilator and gas-rig time series with gradient-boosted trees, and report it as three nested implementation scenarios of one code base: X0 (XGBoost 2.1.4 through our pipeline, no physics term active), XP (X0 plus a blend toward the anchor and the cumulative oxygen uptake rate over the gap per kilogram of body mass, derived from an oxygen balance of the arterial blood) and v4 (XP plus a patient-level anchor that recalibrates the nine-hour-old measurement toward the patient’s own history). Out-of-fold predictions under patient-grouped five-fold cross-validation are scored on the 3730 hours of 30 patients whose inputs contain no SaO2m value younger than nine hours. From X0 to XP, the mean per-patient MAPE decreases slightly, from 3.321% to 3.289%, an effect resolved by a paired patient bootstrap for this summary only. From XP to v4, it decreases further to 3.117%, the pooled MAPE decreases from 3.890% to 3.727% (both changes resolved by the bootstrap), and the pooled correlation increases from 0.754 to 0.793. Meanwhile, the model becomes substantially smaller, with the number of input features decreasing from 55 to 19 and the number of tree leaves decreasing from 9527 to 450. An exact decomposition of every prediction shows that in v4 the two-coefficient anchor earns 81% of the improvement over persistence and carries nearly all of the prediction variance. All formulas are matched line by line to the C++ and Python sources.

Hybrid modeling techniques are increasingly important for improving predictive accuracy and control in biomanufacturing, particularly in data-limited conditions. This study develops and experimentally validates a hybrid deep learning model predictive control (MPC) framework for recombinant P. pastoris fed-batch fermentations. Bayesian optimization and grid search techniques were employed to identify the best-performing hybrid model architecture: an LSTM layer with 2 hidden units followed by a fully connected layer with 8 nodes and ReLU activation. This design balanced accuracy (NRMSE 4.93%) and computational efficiency (AICc 998). This architecture was adapted to a new, smaller dataset of bacteriophage Qβ coat protein production using transfer learning, yielding strong predictive performance with low validation (3.53%) and test (5.61%) losses. Finally, the hybrid model was integrated into a novel MPC system and experimentally validated, demonstrating robust real-time substrate feed control in a way that allows it to maintain specific target growth rates. The system achieved predictive accuracies of 6.51% for biomass and 14.65% for product estimation, with an average tracking error of 10.64%. In summary, this work establishes a robust, adaptable, and efficient hybrid modeling framework for MPC in P. pastoris bioprocesses. By integrating automated architecture searching, transfer learning, and MPC, the approach offers a practical and generalizable solution for real-time control and supports scalable digital twin deployment in industrial biotechnology.

Hybrid modeling techniques are increasingly important for improving predictive accuracy and control in biomanufacturing, particularly under data-limited conditions. This study focuses on the development and experimental validation of a hybrid deep learning model predictive control (MPC) framework for fed-batch P. pastoris fermentations producing recombinant proteins. Bayesian optimization and grid search were employed to identify optimal network architectures, revealing that models combining LSTM layers with fully connected layers provided the best balance between prediction accuracy and computational efficiency. The top-performing architecture was adapted to a new dataset involving bacteriophage Qβ coat protein production using transfer learning, yielding strong predictive performance with low validation and test losses. Experimental implementation of the novel hybrid MPC system demonstrated robust real-time control of substrate feeding to maintain target specific growth rates. While moderate discrepancies were observed in biomass and product predictions—particularly during the methanol adaptation phase and late-stage cytotoxic conditions—the controller effectively regulated process dynamics. These findings suggest that hybrid neural networks, when integrated with MPC and refined through automated architecture selection, offer a practical and generalizable solution for real-time control in microbial bioprocesses. This work provides a validated framework for deploying hybrid digital twins in fermentation and highlights the need to account for physiological effects in future models.

A model predictive control (MPC) method was investigated as a route to optimize and control the growth of E. coli BL21 (DE3) and biosynthesis of two different recombinant proteins (nerve growth factor NGF and coat protein of bacteriophage Qβ (Qβ-CP)). To determine the target trajectory for the E. coli cultivation process and estimate the model parameters, the off-line run-to-run optimization method was used. The proven method allowed us to successfully control the growth of microbial biomass, with a deviation of 6–12% from the target trajectory. It was proven that it is possible to obtain a “Golden Batch” profile for the implementation of MPC using datasets from only four to eight fermentation runs. The method showed its robustness when the cultivation of E. coli was carried out with two different titrant supply control systems—volumetric and gravimetric. Furthermore, the MPC method exhibited high adaptability, reliability, and resistance to various types of disturbances. MPC proved to be a reliable and effective method for controlling the cultivation and recombinant protein biosynthesis of fast-growing microorganisms such as E. coli.

Accurate estimations of the concentrations of soluble compounds are crucial for optimizing bioprocesses involving Escherichia coli (E. coli). This study proposes a hybrid model structure that leverages off-gas analysis data and physiological parameters, including the average biomass age and specific growth rate, to estimate soluble compounds such as acetate and glutamate in fed-batch cultivations We used a hybrid recurrent neural network to establish the relationships between these parameters. To enhance the precision of the estimates, the model incorporates ensemble averaging and information gain. Ensemble averaging combines varying model inputs, leading to more robust representations of the underlying dynamics in E. coli bioprocesses. Our hybrid model estimates acetates with 1% and 8% system precision using data from the first site and the second site at GSK plc, respectively. Using the data from the second site, the precision of the approach for other solutes was as fallows: isoleucine −8%, lactate and glutamate −9%, and a 13% error for glutamine., These results, demonstrate its practical potential.

Docosahexaenoic acid (DHA) is one of the most important long-chain polyunsaturated fatty acids (LC-PUFAs), with numerous health benefits. Crypthecodinium cohnii, a marine heterotrophic dinoflagellate, is successfully used for the industrial production of DHA because it can accumulate DHA at high concentrations within the cells. Glycerol is an interesting renewable substrate for DHA production since it is a by-product of biodiesel production and other industries, and is globally generated in large quantities. The DHA production potential from glycerol, ethanol and glucose is compared by combining fermentation experiments with the pathway-scale kinetic modeling and constraint-based stoichiometric modeling of C. cohnii metabolism. Glycerol has the slowest biomass growth rate among the tested substrates. This is partially compensated by the highest PUFAs fraction, where DHA is dominant. Mathematical modeling reveals that glycerol has the best experimentally observed carbon transformation rate into biomass, reaching the closest values to the theoretical upper limit. In addition to our observations, the published experimental evidence indicates that crude glycerol is readily consumed by C. cohnii, making glycerol an attractive substrate for DHA production.

An adaptive control system for the set-point control and disturbance rejection of biotechnological-process parameters is presented. The gain scheduling of PID (PI) controller parameters is based on only controller input/output signals and does not require additional measurement of process variables for controller-parameter adaptation. Realization of the proposed system does not depend on the instrumentation-level of the bioreactor and is, therefore, attractive for practical application. A simple gain-scheduling algorithm is developed, using tendency models of the controlled process. Dissolved oxygen concentration was controlled using the developed control system. The biotechnological process was simulated in fed-batch operating mode, under extreme operating conditions (the oxygen uptake-rate’s rapidly and widely varying, feeding and aeration rate disturbances). In the simulation experiments, the gain-scheduled controller demonstrated robust behavior and outperformed the compared conventional PI controller with fixed parameters.

Microbial biomass concentration is a key bioprocess parameter, estimated using various labor, operator and process cross-sensitive techniques, analyzed in a broad context and therefore the subject of correct interpretation. In this paper, the authors present the results of P. pastoris cell density estimation based on off-line (optical density, wet/dry cell weight concentration), in-situ (turbidity, permittivity), and soft-sensor (off-gas O₂/CO₂, alkali consumption) techniques. Cultivations were performed in a 5 L oxygen-enriched stirred tank bioreactor. The experimental plan determined varying aeration rates/levels, glycerol or methanol substrates, residual methanol levels, and temperature. In total, results from 13 up to 150 g (dry cell weight)/L cultivation runs were analyzed. Linear and exponential correlation models were identified for the turbidity sensor signal and dry cell weight concentration (DCW). Evaluated linear correlation between permittivity and DCW in the glycerol consumption phase (+ strain) to significant (for Mut^S strain) linearity decline for methanol consumption phase. DCW and permittivity-based biomass estimates used for soft-sensor parameters identification. Dataset consisting from 4 Mut⁺ strain cultivation experiments used for estimation quality (expressed in NRMSE) comparison for turbidity-based (8%), permittivity-based (11%), O₂ uptake-based (10%), CO₂ production-based (13%), and alkali consumption-based (8%) biomass estimates. Additionally, the authors present a novel solution (algorithm) for uncommon in-situ turbidity and permittivity sensor signal shift (caused by the intensive stirrer rate change and antifoam agent addition) on-line identification and minimization. The sensor signal filtering method leads to about 5-fold and 2-fold minimized biomass estimate drifts for turbidity- and permittivity-based biomass estimates, respectively.

This article presents the development and application of a distinct adaptive control algorithm that is based on fuzzy logic and was used to control the specific growth rate (SGR) in a fed-batch biotechnological process. The developed control algorithm was compared with two adaptive control systems that were based on a model-free adaptive technique and gain scheduling technique. A typical mathematical model of recombinant Escherichia coli fed-batch cultivation process was selected to evaluate the performance of the fuzzy-based control algorithm. The investigated control techniques performed similarly when considering the whole process duration. The adaptive PI controller with fuzzy-based parameter adaptation demonstrated advantages over the previously mentioned algorithms—especially when compensating the deviations of the SGR. These deviations usually occur when the equipment malfunctions or process disturbances take place. The fuzzy-based control system was stable within the investigated ranges. It was determined that, regarding control quality, the investigated control algorithms are suited to control the SGR in a fed-batch biotechnological process. However, substrate feeding rate manipulation and limitation needs to be used. Taking into account the time needed to design and tune the controller, the developed controller is suitable for practical applications when expert knowledge is available. The proposed algorithm can be further adapted and developed to control the SGR in other cell cultivations while running the process under substrate limitation conditions.