Accurate quality prediction and reliable uncertainty quantification are critical to ensure safe, efficient, and optimal operation of industrial processes. However, industrial fermentation processes involve strongly coupled and dynamically evolving operating variables, making it difficult for a single deterministic predictor to capture both multi-target correlations and predictive uncertainty. Therefore, we propose an Advanced Uncertainty-Aware Deep Ensemble (A-UADE) framework to improve multi-target prediction accuracy and uncertainty reliability. First, an uncertainty-weighted loss function dynamically adjusts the contribution of each sample–task pair based on the model’s predictive uncertainty. Second, negative correlation learning with standardized error correlation fosters heterogeneous predictive behaviors among ensemble members. Finally, projective gradient surgery alleviates gradient interference during multi-task training. Experiments on two real-world fermentation processes and one dynamic gas-mixture dataset show that A-UADE provides competitive point prediction accuracy and reliable uncertainty quantification compared with the retained contemporary baselines.

The rapid digitalization of bioprocess engineering has stimulated interest in large language models (LLMs) as interactive decision-support tools. Although LLMs encode extensive biological and technical knowledge, their reliability in authentic bioprocess engineering scenarios, where operational constraints, multi-scale dynamics, and process trade-offs must be considered, remains largely unexamined. To address this gap, we introduce BioMeter, a domain-specific evaluation framework constructed from authentic bioprocess engineering problem-solving scenarios, including troubleshooting, optimization, and operational decision tasks. BioMeter enables systematic assessment of whether LLMs can generate operationally feasible and constraint-consistent responses. Comprehensive evaluation of state-of-the-art open-source and commercial LLMs reveals substantial limitations in process-level reasoning, particularly in knowledge integration and decision-making under dynamic constraints. Building on these findings, we propose BioManner, an enhancement framework that improves engineering reliability by integrating hierarchical domain knowledge through retrieval-augmented generation and employing structured reasoning for constraint-aware problem decomposition. Experimental results demonstrate that BioManner significantly improves response accuracy, consistency, and engineering relevance across multiple task types, achieving up to 87% accuracy improvement over baseline models while maintaining full relevance to practical engineering queries. The code and data are available at: https: //github.com/TrustMATIC-Lab/BioMM.

The accuracy of enzyme kinetic parameters, particularly enzyme turnover numbers (kcat), is critical for the predictive performance of enzyme-constrained genome-scale metabolic models. However, currently available kinetic datasets remain sparse and often fail to capture in vivo enzyme behavior, thereby limiting model accuracy. To address these limitations, we develop EnzymeTuning, a generative adversarial network-based framework for global kcat optimization. By further incorporating literature-derived protein degradation constants, we infer protein synthesis rates and systematically assess their impact on model performance. Here, we show that EnzymeTuning substantially improves prediction accuracy and expands proteome-level coverage across diverse organisms, including Saccharomyces cerevisiae, Kluyveromyces lactis, Kluyveromyces marxianus, Yarrowia lipolytica, and Escherichia coli. Furthermore, EnzymeTuning reveals context-dependent enzyme usage patterns and adaptive catalytic resource allocation under diverse carbon- and nitrogen-limited chemostat conditions, underscoring the substantial potential of this framework for integrative multi-omics analyses.

Abstract Understanding transporter kinetics is essential for deciphering metabolite exchanges in biosystems, particularly for cells subject to substrate gradients. Nevertheless, the prediction of transporter kinetic parameters, maximum rate per gram protein (V max ) and Michaelis-Menten constant (K m ), has not yet been tackled. Here, we developed the first compound-protein interaction machine learning model of transporter V max and K m , MMTKPred, which achieved R 2 =0.553, RMSE=1.155 mmol/hr/g Protein and R 2 =0.330, RMSE=0.935 mM for log10-scaled V max and K m prediction, respectively. Moreover, we demonstrated MMTKPred’s predictive power across biosystem scales, from capturing transporter kinetics modulated by point mutations and substrate changes at the molecular level, to enabling substrate-sensitive metabolic modelling of non-model yeasts at the cellular level, and rationalizing inter-species substrate competition in co-cultures. Collectively, MMTKPred effectively models metabolite transport spanning from molecular to multi- species scales, thereby offering a computational tool for rational microbial cell factory optimization. Graphical abstract Highlights MMTKPred, first transporter kinetics CPI model, reaches ∼1 log10 RMSE for V max and K m . MMTKPred captures the effects of point mutations and substrate changes on transporters. Predicted kinetics enables substrate sensitivity in metabolic flux modelling. Predicted kinetics explains inter-species substrate competition outcomes.

Soft sensing is critical for real-time state awareness and quality control in biomanufacturing processes. However, existing soft sensing methods often underperform because they overlook the intrinsic multistage characteristics of these processes. Moreover, interpretability based on global variable contributions cannot provide actionable guidance for process optimization. To address these issues, we proposed the Bayesian optimization-driven multimetric partitioning with generalized additive models (BOMP-GAM). First, the framework partitions process data into different stages by fusing spatiotemporal and causal similarity through Bayesian optimization. Second, a node-level generalized additive model is used to yield stage-specific predictions and shape function of measured variables. Third, the final predictions were obtained by dynamically weighting the stagewise predictions based on multimetric similarity. Experiments on both simulated and real biomanufacturing processes show that BOMP-GAM achieves state-of-the-art predictive accuracy while uncovering stage-specific shape functions of measured variables. These insights enable targeted operational adjustments and ultimately improve production efficiency.

Reliable scale-up of fermentation processes remains challenging due to changes in reactor size, operating conditions, and transport phenomena, which can introduce distribution shifts between laboratory and industrial environments. This study proposes a physics-informed cross-scale transfer learning framework that integrates conservation principles with deep learning for laboratory-to-industrial fermentation prediction. The approach is demonstrated using an industrial penicillin fermentation case study. A physics-informed generative adversarial network is employed to generate continuous low-fidelity data from laboratory-scale batches under physical constraints, while a deep neural network is pre-trained to capture transferable fermentation representations. To address cross-scale discrepancies, an element conservation-based transfer learning strategy is introduced by incorporating carbon, nitrogen, and PAA conservation constraints into model adaptation. The proposed framework improves synthetic data quality by reducing MMD and DTW by 55.6% and 29.9%, respectively, achieves high prediction accuracy for penicillin titer (R2 = 0.9956), and significantly improves residual PAA and NH4+ prediction compared with baseline models. These results demonstrate the potential of combining physical constraints and transfer learning for cross-scale fermentation prediction under limited industrial data availability.

Understanding how nutrient-specific limitations shape anaerobic metabolism in Saccharomyces cerevisiae is essential for defining the physiological limits of yeast growth. Integrating chemostat physiology, multi-omic profiling, and targeted metabolic engineering under strictly anaerobic conditions, we show that yeast maintains a conserved maximum glucose uptake (~14 mmol/gDW/h) under carbon (C), nitrogen (N), and phosphorus (P) limitation, while distinct regulatory bottlenecks constrain maximal growth rate: ATP insufficiency under C and P limitation, and aminoacyl-tRNA synthetase scarcity under N limitation. Under these stresses, S. cerevisiae reallocates proteomic resources toward anabolic functions, with nutrient-specific phosphorylation networks compensating for translational stress, most pronounced under N limitation. Building on these insights, a "push-pull" strategy enhancing energy supply (VMA3) and translational capacity (WRS1) increased the maximal anaerobic growth rate by 27.2%, 47.5% and 52.5% under C, N, and P limitation, respectively. These findings reveal energy-translation coupling as the central determinant of anaerobic growth limits and provide a framework for rational strain engineering.

Fault root cause diagnosis is crucial for ensuring safety and improving efficiency in industrial processes. Traditional methods, such as granger causality and partial cross mapping, cannot identify the causal polarity (positive or negative) between variables, resulting in incorrect causal relationships and inability to effectively locate the root cause variable of the fault. Therefore, this paper proposes a hierarchical direct causal polarity identification (HDCPI) framework that infers direct causality and provides causal polarity between variables for root cause diagnosis. First, a continual learning variational autoencoder based multimode monitoring approach is used to detect fault occurrences, and candidate root cause variables are identified through contribution plot. Then, sparse nonlinear dynamics identification improved by dual domain sampling is applied to mining dynamic. Finally, a model-based causal detection function is used to eliminate indirect causality and identify causal polarity, thereby constructing a causal graph for fault root cause diagnosis. Extensive evaluations on both simulated and real-world datasets demonstrate that the proposed HDCPI substantially improves diagnostic precision and interpretability, consistently outperforming existing methods.

Although neural network-based soft sensing in biomanufacturing processes shows substantial promise, many existing approaches either fail to encode physical prior knowledge or do not make effective use of unlabeled data. These limitations hinder both the generalization of the model and its adherence to physical consistency. To address these issues, this study proposes the Physics-Informed Variational Autoencoder Regression (PIVAER)-based soft sensing, which improves the VAE by incorporating: (1) monotonicity constraints on specified input-output relations to ensure that predictions respect known monotonic trends, and (2) embedded differential constraints derived from the kinetic equation. Soft penalties in the form of physics residual and monotonicity terms are embedded within a unified objective function, thereby enabling semi-supervised training that integrates both labeled and unlabeled data. Through this design, the PIVAER achieves improved physical consistency and generalization capability. Its effectiveness has been demonstrated by experiments conducted on both simulated and real penicillin fermentation datasets. Comparative experiments using state-of-the-art methods demonstrate superior predictive accuracy and enhanced generalization capability. (c) 2025 The Chemical Industry and Engineering Society of China, and Chemical Industry Press Co., Ltd. All rights are reserved, including those for text and data mining, AI training, and similar technologies.

The choice of pH neutralizer plays a critical role in determining the performance of industrial lactic acid fermentation. This study investigated the distinct phenotypic and metabolic responses of Bacillus coagulans during batch fermentation with mixed carbon sources (glucose and trehalose) when using sodium hydroxide versus calcium hydroxide for pH control. Compared to calcium hydroxide, sodium hydroxide resulted in a slightly lower lactic acid yield and induced significant cell lysis during the carbon source transition. Metal ion experiments confirmed that sodium ion stress is the primary factor responsible for extensive cell lysis. Omics analysis indicated that under sodium hydroxide conditions, cells require more energy to resist stress. To meet this demand, cells downregulated cellular growth and sodium ion uptake systems, while upregulating carbon source transport. However, during carbon source transition, cellular resource reallocation resulted in insufficient carbon source uptake, leading to stagnation of glycolysis and amino acid metabolism. The resulting energy deficit triggered upregulation of cellular hydrolases, ultimately causing population-wide cell lysis. In contrast, calcium hydroxide as a neutralizer does not impose sodium stress and contributes to maintaining cell wall stability under carbon source deficiency. This study contributes to the understanding of cellular regulatory mechanisms in response to different neutralizers, and provides insights for metabolic engineering and process optimization.IMPORTANCEpH neutralizer selection critically determines cell survival and process efficiency in industrial fermentation. Our study demonstrates that sodium hydroxide, although effective for pH control, imposes severe sodium ion stress that synergizes with the inherent energy vulnerability of cells during carbon source transition, triggering catastrophic population-wide lysis in Bacillus coagulans. Through integrated physiology, ion experiments, and multi-omics analyses, we reveal that the underlying cause is an energy crisis arising from the conflict between stress defense and metabolic reallocation. In contrast, calcium hydroxide not only avoids sodium stress but also confers structural and thermal stability to cells, thereby preserving cell density and improving carbon utilization. These findings shift the perspective on neutralizers from mere pH regulators to key determinants of cellular energy economy and survival. Our study thus offers a rational basis for neutralizer selection, strain engineering toward ion resistance, and the design of robust fermentation processes, with direct implications for the sustainable and cost-effective production of lactic acid and other microbial metabolites.

Accurate real-time prediction of key quality indicators remains a major challenge in industrial bioprocessing, where complex, time-varying kinetics and unmeasurable metabolic states hinder process optimization. Existing physics-informed neural networks (PINNs) offer a promising hybrid paradigm by integrating data and mechanistic knowledge, yet their reliance on fully calibrated models limits robustness under parameter drift and unmeasurable state variables. To address these limitations, this study proposes a calibration-free PINN framework that jointly estimates key quality indicators and critical, unmeasurable internal system states, enabling robust extrapolation under parameter drift and limited data. Validated by industrial penicillin fermentation (cross-strain/process), simulations, and chemical cases, the framework demonstrates robust extrapolation under noise and parameter variability, accurately predicting key quality indicators. Empirically, the accuracy of unmeasurable state variable estimation is bounded by that of the physical model parameters. This affirms a high-precision, strongly generalizable modeling paradigm for complex bioprocesses.

Respiratory oscillations are a hallmark of glucose-limited yeast chemostats, yet how growth rate shapes their emergence and collapse remains unclear. Here, we combined accelerostat cultivation with quantitative metabolomics and proteomics to characterize the transition from oscillatory to non-oscillatory metabolism in Saccharomyces cerevisiae under aerobic, glucose-limited conditions. Respiratory oscillations were maintained at low growth rates, attenuated at intermediate rates, and no longer observed at higher rates, coinciding with the onset of ethanol formation. Metabolomics analysis showed that oscillatory dynamics were most pronounced in tricarboxylic acid cycle intermediates and trehalose, whereas glycolysis and the pentose phosphate pathway exhibited weaker oscillations and instead adjusted pool sizes with growth rate. Quantitative proteomics further indicated that loss of oscillations was accompanied by non-uniform proteome reallocation, including increased representation of translation, glycolysis, energy metabolism, and amino acid biosynthesis, together with reduced relative allocation to buffering and proteostasis-related functions. Together, these results indicate a growth rate-associated physiological transition in glucose-limited yeast, in which the disappearance of oscillatory behavior during accelerostat cultivation is associated with a shift from respiratory to respiro-fermentative metabolism and coordinated reorganization of the proteome.

The Crabtree effect in Saccharomyces cerevisiae involves rapid metabolic switching between respiratory and fermentative states under dynamic conditions, yet its regulatory principles across different perturbation timescales remain poorly understood. Here, we imposed short-term glucose pulses and long-term dilution-rate perturbations on steady-state yeast cultures with extracellular physiological analysis, time-resolved intra- and extracellular metabolomics, proteomics, and constraint-based flux analysis to investigate dynamic Crabtree regulation. During short-term glucose pulses, ethanol accumulation scaled linearly with perturbation intensity, while approximately 33

Driven by consumer preferences for safety and environmental protection, the global cosmetics industry has an increasing demand for natural and sustainable ingredients. Saccharomyces cerevisiae has emerged as a powerful platform for the biosynthesis of cosmetic ingredients due to its strong metabolic capacity, genetic operability, and cost-effective production capabilities. This review focuses on the latest advances in S. cerevisiae for the production of high-value cosmetic compounds, including antioxidants, repair agents, moisturizers, and structure-maintaining ingredients. Key strategies, such as genetic and metabolic engineering, pathway modularity, and fermentation optimization, are discussed, demonstrating significant improvements in yield and efficiency. In addition, the integration of artificial intelligence and machine learning in strain design and process control is explored, providing promising solutions to overcome metabolic bottlenecks and scale up production. Despite challenges such as metabolic burden, S. cerevisiae shows great potential for sustainable and scalable biosynthesis of cosmetic ingredients, paving the way for the next generation of biobased cosmetics. This comprehensive review provides valuable insights and technical references for the development of the field of synthetic biology in the cosmetics industry.

The accuracy of enzyme kinetic parameters, particularly the turnover number ( k cat), is critical for the predictive power of enzyme-constrained genome-scale metabolic models (ecGEMs). However, current kinetic datasets remain sparse and often fail to capture in vivo enzyme behavior, compromising model predictive capacity. To address these challenges, we developed EnzymeTuning——a generative adversarial network (GAN)-based framework for the global kcat optimization. This approach significantly enhances both the accuracy and proteome-level coverage of ecGEM predictions. Moreover, by incorporating literature-derived protein degradation constants ( kdeg ), we inferred the protein synthesis rate and systematically evaluated their impact on model performance. The framework was validated across distinct yeast species, including Saccharomyces cerevisiae , Kluyveromyces lactis , Kluyveromyces marxianus , and Yarrowia lipolytica , demonstrating its generalizability. Further, we found that EnzymeTuning facilitates the identification of context-specific enzyme usage patterns and adaptive catalytic resource allocation under varying carbon-to-nitrogen (C/N) ratios, showcasing the substantial potential of our toolbox for integrative omics analysis. Overall, EnzymeTuning provides a robust and scalable solution for refining kinetic parameters in ecGEMs, thereby promoting the wide applications of these computational models in systems and synthetic biological studies. ### Competing Interest Statement The authors have declared no competing interest. Shanghai Municipal Science and Technology Major Project, 2022YFA0913000 the National Key R&D Program of China, 2020YFA0908300 Natural Science Foundation of Shanghai, 25ZR1402110 National Natural Science Foundation of China, 22208211, 22378263

This study introduces a novel method called mutual information (MI) and attention-based variable selection (MAVS) to address the challenges of irrelevant and redundant variables in industrial process soft sensing while providing interpretability in variable contribution analysis. First, irrelevant variables are eliminated based on low MI values with the quality variable. Second, attention scores are used to remove redundant variables, and the false discovery rate is used to determine the number of beneficial variables. Finally, this work provides an interpretable and accurate contribution of the selected variables by using kernelSHAP, a kernel-based Shapley analysis. Unlike traditional approaches, MAVS integrates MI with attention mechanisms to optimize variable selection dynamically and adaptively. MAVS obtains stronger robustness and higher accuracy than the existing state-of-the-art models through optimal variable selection. The former also obtains better superior generalization than the latter through adaptive adjustment of attention weights. The superiority of MAVS is demonstrated using two real-world datasets and one simulated dataset.

Ethanol, a high-demand clean energy source, is primarily produced via fed-batch fermentation in industrial settings. Although our previous study identified an optimal glucose concentration of 30 g/L for maximal ethanol yield, the mechanisms underlying glucose-dependent cellular adaptation remain unclear. Here, we performed an integrated multi-omics analysis, including transcriptomics, proteomics, metabolomics, and fluxomics, to compare yeast cells under glucose-controlled and uncontrolled conditions. Our results indicate that high glucose stress triggers the regulation of transporters with different affinities and the upregulation of heat shock proteins (HSPs), trehalose, and amino acids. In contrast, protein turnover was reduced under glucose-controlled conditions, suggesting more efficient resource allocation. This metabolic reallocation enhances carbon flux through glycolysis, potentially providing additional energy and NADH to support biomass growth and ethanol production. These findings advance our understanding of yeast regulatory mechanisms under glucose stress and provide insights for metabolic engineering and process optimization.

Adipose tissue, a pivotal player in whole-body energy homeostasis and insulin sensitivity, undergoes considerable remodelling throughout the ageing process, a facet that has garnered little attention until the past decade. This Review comprehensively summarizes the dynamic metabolic, cellular and functional changes that occur in white and thermogenic adipose tissue during distinct ageing stages, across different adipose tissue depots. We emphasize the influence of ageing on different cell types within adipose tissue, including adipocytes, adipocyte progenitors, immune cells and senescent cells, and their collective effect on adipose tissue function and systemic metabolism. We also decipher the correlation between adipose tissue ageing and prevalent age-related conditions such as metabolic dysfunction-associated fatty liver disease and cardiovascular diseases. Finally, the Review delves into the potential of current anti-ageing interventions to beneficially affect adipose tissue, encompassing caloric restriction, metformin, glucagon-like peptide 1 receptor agonists and senolytics. The discussion extends to the exploration of whether targeting adipose tissue through such interventions could emerge as a prominent therapeutic strategy for mitigating age-related diseases and enhancing the healthspan and lifespan of the ageing population. This Review outlines changes in endocrine and metabolic function that occur in various adipose tissue depots during ageing and explains how these alterations contribute to age-related decline in metabolic health. The Review also discusses how various treatments might slow this decline by improving adipose tissue function.

Industrial-scale microbial fermentation processes often face limitations in mixing and mass transfer, leading to the formation of environmental gradients within the bioreactor. These gradients expose microbes to heterogeneous conditions over time and space. In this study, we evaluated the effects of combined substrate and dissolved oxygen (DO) gradients on the metabolic response of Penicillium chrysogenum at an industrial scale. Three representative heterogeneous environments were simulated in scale-down systems: (1) feed inlet (high glucose, low oxygen (HGLO): C-S > 20 mM, DO < 0.012 mM), (2) aeration inlet (high oxygen, low glucose (HOLG): C-S < 0.8 mM, DO > 0.2 mM), and (3) global environment (periodic 360 s fluctuation cycle with 45 s of HGLO and 75 s of HOLG conditions). Results showed that prolonged exposure to feed inlet conditions led to a complete loss of penicillin production capacity, accompanied by significant excretion of intracellular metabolites, and this effect was largely irreversible. While, cells randomly walking under the top impeller zone did not lose production capacity but showed signs of premature degeneration due to increased energy demand. When exposed to the global environment, cells finely tuned their metabolism in a periodical manner, with nearly a 50% loss of penicillin productivity. In summary, substrate gradients alone did not cause irreversible effects, but large substrate gradients contributed to reduced productivity. Oxygen gradients, however, not only reduced production but also caused irreversible cellular damage. These findings provide valuable insights for developing scale-up criteria and strain engineering strategies aimed at improving large-scale culture performance.

Fed-batch fermentation has become the preferred strategy in many industrial biomanufacturing processes. However, a key challenge remains in optimizing the feeding strategy to achieve stable maximum yields. In this study, we present an online Raman spectroscopy-based monitoring and control system, using bioethanol production by Saccharomyces cerevisiae as a case study. To address the issue of limited labeled data, a pseudo-labeling approach based on semi-supervised learning was employed, expanding the available training data set by 100-fold compared to conventional labeling methods. In addition, we developed a spectral-temporal concatenation convolutional neural network (STC-CNN) that incorporates sequential spectral features. Comparative evaluations with multiple machine learning algorithms demonstrated the superior performance of STC-CNN, achieving a root mean square error (RMSE) of 3.63 g/L for glucose prediction. The system enabled rapid and automated glucose feeding to maintain various target concentrations. Notably, a glucose setpoint of 30 g/L yielded the highest ethanol concentration of 140.68 g/L-an increase of 3.85% over traditional Fed-batch fermentation-while reducing glycerol by 6.67%. These results highlight the significant potential of Raman spectroscopy combined with deep learning for automated bioprocess optimization and discovery of optimal operating strategies.