Co-cultivation of yeast to enhance carboxypeptidase enzyme production: biocalorimetric, respirometric, and deep learning investigation

Carboxypeptidase Y (CPDY) production by Kluyveromyces marxianus is typically limited by ethanol accumulation, which inhibits enzyme yield. In this study, this challenge was addressed by co-cultivating K. marxianus with Saccharomyces cerevisiae to scavenge ethanol and thus enhance CPDY production. CPDY is an industrially important exopeptidase used in food processing, pharmaceutical and biorefinery applications. Using Response Surface Methodology (RSM), key medium components (yeast nitrogen base, ammonium chloride, dextrose) in shake-flask fermentations and a 2 L biocalorimetric reactor were optimized. The co-culture showed a clear division of labor such that, K. marxianus produced CPDY and ethanol from dextrose, and S. cerevisiae rapidly consumed the ethanol, preventing accumulation. As a result, CPDY activity was sustained longer, achieving enzyme titers roughly double those of individual cultures. Maximum CPDY activities of 8 U/mL (shake flask) and 16 U/mL (biocalorimeter) were obtained, with corresponding biomass concentrations of 3 g/L and 4.5 g/L. Real-time metabolic heat flow (calorimetry) and off-gas analysis revealed distinct dextrose and ethanol utilization phases, confirming the co-culture’s synergistic metabolism. Additionally, a custom Long Short-Term Memory (LSTM) neural network (“CoYeast-LSTM”) was developed to model the bioprocess. The LSTM accurately predicted key metabolic transitions and CPDY yields with a mean squared error MSE of 0.0060 and mean absolute error MAE of 0.071. This integrated approach, combining co-culture fermentation with advanced monitoring yielded high CPDY activity and improved real-time process control. The findings illustrate a novel and sustainable strategy for enzyme production through microbial co-cultivation coupled with real-time bioprocess monitoring and predictive modelling.

Yeasts, particularly Kluyveromyces marxianus and Saccharomyces cerevisiae, are industrially relevant organisms valued for their robustness, ability to withstand temperatures up to 45 °C, utilization of diverse carbon sources, and GRAS (Generally Regarded As Safe) status (Fleet, 2006). These yeasts produce numerous industrially important enzymes, including CPDY, a serine carboxypeptidase that cleaves amino acids from the carboxyl terminus of peptides (Staniszewski and Kordowska-Wiater, 2021; Ye et al., 2021).

Carboxypeptidase Y has diverse applications in food fortification, flavour enhancement, and debittering of protein hydrolysates (Fonseca et al., 2008; Leonardi et al., 2024). Recent research demonstrates its potential in biomass conversion through hydrolysis of lignocellulose-bound proteins, converting agricultural residues into valuable amino acids and enhancing downstream fermentation yields for bioethanol and polyhydroxyalkanoates (PHAs) (Ramírez-Zavala et al., 2004; Sung et al., 2022). Immobilized CPDY has been applied in protein hydrolysis and polypeptide sequencing, with enzymatic technologies widely used in food processing for modifying food components and producing functional biochemicals (Ro et al., 2005).

A major limitation in CPDY production is ethanol inhibition. Although K. marxianus demonstrates superior CPDY production capability compared to S. cerevisiae under standard conditions, its low ethanol tolerance restricts broader industrial application (Müntz et al., 2001). In contrast, S. cerevisiae exhibits higher ethanol tolerance but produces lower CPDY activity. Mixed starter cultures of multiple yeast species represent a cost-effective strategy to enhance enzyme production while mitigating inhibitory byproducts (Liu et al., 2022). Co-cultivation leverages syntrophism and metabolic cross-feeding between organisms, enabling complex metabolic outputs such as enhanced enzyme production and improved fermentation efficiency (Liu et al., 2024).

Biocalorimetric and respirometric monitoring provide real-time assessment of metabolic activity during bioprocesses, with heat and gas exchange serving as indicators of substrate consumption, growth rate, and product formation (Das et al., 2024). Response Surface Methodology (RSM) offers a systematic approach for optimizing bioprocess conditions while minimizing resource consumption (Roldán-López et al., 2025). Additionally, deep learning models, particularly Long Short-Term Memory (LSTM) networks, have demonstrated effectiveness in bioprocess modeling, capturing temporal dynamics and enabling predictive process control (Agyei et al., 2024; Akparov et al., 2024).

To the best of our knowledge, this is the first biocalorimetric-respirometric investigation of K. marxianus and S. cerevisiae co-cultivation for enhanced CPDY production through ethanol scavenging. This study employs RSM to optimize media composition and growth conditions, combines biocalorimetric-respirometric monitoring with AI-driven analytics using a bidirectional LSTM model (CoYeast-LSTM), and aims to establish a systematic framework for controlling biosynthesis of enzymes in engineered microbial consortia.

A detailed description of the biocalorimetry is available in our earlier publications (Santharam et al., 2017; Adimoolam et al., 2020; Easwaran et al., 2022) and the biocalorimeter setup is shown in Supplementary Figure S1 K Marxianus (MTCC 4139) and S. cerevisiae (MTCC 170) was acquired from MTCC (The Microbial Type Culture Collection), Chandigarh, India (www.mtccindia.res.in).

Co-culture inoculation and consortium preparation

Pre-cultures of K. marxianus and S. cerevisiae were grown separately in Yeast-peptone-dextrose (YPD) broth (10 g/L yeast extract, 20 g/L peptone, 20 g/L dextrose) at 30 °C and 200 rpm until the mid-exponential phase (OD 600 = 1.0 ± 0.05, corresponding to approximately 0.5 × 10 8 CFU/mL for each strain). The co-culture was established at an inoculation ratio of 1:1 (v/v) of K. marxianus to S. cerevisiae, with each strain contributing an initial OD 600 of 0.5 (combined OD 600 = 1.0) to the production medium. This equal volumetric ratio was selected to permit K. marxianus to establish CPDY production during the initial dextrose-consumption phase before ethanol accumulation triggered S. cerevisiae metabolic activity. To verify that both strains co-existed and maintained the expected division of labour throughout the fermentation, samples withdrawn at 6 h intervals were plated on two selective media, YPD agar incubated at 45 °C (selective for the thermotolerant K. marxianus, on which S. cerevisiae forms no colonies) and lysine agar incubated at 30 °C (selective for S. cerevisiae). Colony-forming unit (CFU) counts from these plates confirmed the persistence of both species across all three metabolic phases and provided qualitative validation that S. cerevisiae remained the minority biomass contributor during the dextrose-rich initial phase and increased proportionally during the ethanol-respiration phase.

Metabolic heat estimation

Precise temperature measurements were made using gold-plated sensors. Heat loss was minimized by an exterior insulating chamber. The heat transfer rate was obtained from Equation 1. where, T r is the reactor temperature, T j is the jacket temperature, U is the heat transfer coefficient [W m -2 K −1 ], and A is the heat exchange area. The heat produced by biological processes was carefully separated from the heat lost due to aeration, ambient heat loss, heat-related to stirring, heat buildup, and dosing heat.

Carbon-dioxide evolution rate (CER)

A gas analyzer (Siemens ultra mat, India) was used to measure carbon-dioxide evolution. The mass balancing approach was used to determine CER.

Dry cell weight (DCW) and dextrose concentration

Samples are collected and centrifuged at 8,000 rpm for 10 min, dried overnight at 80 °C and weighed to determine Dry Cell Weight (DCW). Residual dextrose concentration in the supernatant collected after centrifugation was measured using dinitrosalicylic acid (DNS) technique (Jain et al., 2020).

Ethanol estimation

Supernatant collected after centrifugation was treated with tributyl phosphate in 1:1 ratio vortexed and centrifuged at 8000 RPM for 10 min. The aqueous phase containing ethanol was collected and used for quantification. Ethanol concentration was quantified by mixing samples with a dichromate reagent (10 g of potassium dichromate in 100 mL of 5 M sulfuric acid) and rested for 1 h. The color change was spectrophotometrically measured at 585 nm and compared with a standard chart (Ergüden, 2018).

Each assay run included two controls, a blank of cell-free supernatant from an uninoculated medium for baseline absorbance, and a sample spiked with 50 mM ethanol for recovery assessment. Prior to the dichromate reaction, samples were extracted with tributyl phosphate, which partitions ethanol into the aqueous phase and excludes non-volatile sugars from the measurement matrix. Residual dextrose in the assay samples, quantified independently by the DNS method, remained below 0.5 g/L at the time points where ethanol was measured (>6 h). The mean ethanol recovery from the spiked samples (n = 3) was 96.8% ± 2.1%.

Enzyme extraction and quantification

Yeast cells were collected from the growth media by centrifugation at 8,000 rpm for 10 min at 4 °C. 0.5 mL cell lysis buffer (‘CelLytic Y′) was used to lyse 0.1 g of yeast cells, which were incubated at room temperature for 20 min Cell debris was removed from the solution by centrifugation. The enzyme assay technique was performed using the supernatant solution containing crude enzyme. CPDY catalyze the release of p-nitroanilide from the N-benzoyl-L-tyrosine-p-nitroanilide (NBTpNA) substrate. This p-nitroanilide release is measured spectrophotometrically at 405 nm to determine CPDY activity (Ramírez-Zavala et al., 2004).

Yield calculations and calorespirometric ratio (CR)

Heat yield due to biomass (DCW) was the slope acquired by linear fitting DCW against cumulative heat values. Cumulative heat is proportional to biomass throughout the reaction. The heat yield values (Y Q X) varies for various substrate. The Calorespirometric ratio (CR) is the ratio between heat dissipation and respiration (Chakrawal et al., 2020).

Morphological studies by scanning electron microscope (SEM)

SEM characterization was used to understand better the morphological properties of microbial species. This type of microscope can also be used for cell lysis studies on yeast. Preparation of sample for SEM analysis was carried out as per the standard procedure (Minuti et al., 2023). SEM imaging was performed using “Phenom-World desktop SEM (Model 2009; Netherlands)”.

Data preparation and feature engineering

The time-series data for LSTM modeling was sourced directly from the biocalorimetric and respirometric data used in the study. The primary dataset includes, online parameters such as, metabolic heat (Q), dissolved oxygen (DO), turbidity (representing biomass concentration), and pH measured continuously throughout the fermentation process and offline parameters such as, dry cell weight (DCW), dextrose (substrate) concentration, CPDY enzyme activity and ethanol levels collected at regular intervals via sampling. The datasets used in this study were pre-processed to ensure temporal consistency, eliminate noise and extract meaningful features for modeling the co-cultivation process of K. marxianus and S. cerevisiae.

Offline parameters (DCW, dextrose, ethanol, and CPDY activity) were sampled at 2 h intervals during the active growth phase (0–14 h) and at 3 h intervals during the ethanol-respiration and stationary phases (14–31 h), yielding approximately 8 and 6 sampling time points per phase respectively (14 data points in total per variable across the 31 h fermentation). This sampling density was selected to ensure cubic spline interpolation operated well within the Nyquist criterion for the metabolic dynamics observed, avoiding artificial oscillations while preserving the biological relevance of phase transitions. CPDY activity and biomass concentration were interpolated using a time-based cubic spline method to maintain temporal continuity across the fermentation timeline. Outliers in online measurements, including metabolic heat (Q) and DO, were identified using a moving median filter with a window size of five and replaced with smoothed values to reduce sensor-induced noise.

The sensor drift on DO readings during the stationary phase was addressed by applying an exponential detrend to prevent loss of the biological relevance in the dataset. The most critical aspects of the dynamics of the fermentation process were captured through feature engineering. To account for the time delay involved in transitioning between metabolic phases, lagged variables of the heat generated (Q) and DO levels were used. Additionally, the rate of heat released (dQ/dt) and the rate of oxygen consumed (dDO/dt) were calculated to further highlight the transition from one phase to another. Biological ratios that describe metabolic states, and substrate use patterns, were derived. A calorespirometric ratio (CR) and a metabolic efficiency ratio (MER) were developed to reflect these characteristics. In order to make the variables comparable across each other, regardless of their unit, normalization was applied as per Equation 2. where Q is the instantaneous metabolic heat flow (W), Δ t is the sampling time interval (h), Δ DCW is the incremental change in dry cell weight (g/L) between consecutive time points, and Δ CPDY is the corresponding incremental change in CPDY enzyme activity (U/mL). MER thus expresses the metabolic heat energy dissipated per unit of combined biomass and enzyme gained over each interval, serving as a dimensionless proxy for anabolic efficiency.

Min-Max normalization (Scaling the values of the variable to be in the range 0–1) was used to place all features on the same numerical scale. Phase-specific normalization was applied using scale parameters derived from the training data for each phase. Critically, phase boundaries were identified exclusively from the continuously available online signals the bimodal metabolic heat flow (Q) profile and the dissolved oxygen (DO) dip-and-recovery signature using pre-defined threshold criteria, without any requirement for offline measurements or LSTM model inference. Phase identification is therefore fully prospective and causally consistent with real-time deployment. This processing is intended to create an optimal dataset for training the LSTM model while preserving the biological relevance of the data.

The 23 offline measurement time points constituted the independent samples underlying the dataset. To prevent information leakage between partitions, training, validation, and test splits were assigned by complete inter-sample interval, so that no interpolated segment spanned a split boundary. All error metrics reported in this study were computed on the Min-Max normalized [0,1] scale.

CoYeast-LSTM model architecture

CoYeast-LSTM’s architecture design is capable of modeling the co-fermentation of K. marxianus and S. cerevisiae for CPDY production as well as the three different phases that occur during the co-fermentation phase of the process. This model includes an input layer with multiple variables from each time step of the co-fermentation process. These include, normalized values of metabolic heat (Q), DO, turbidity, pH, dry cell weight (DCW), substrate concentration, ethanol concentration, and CPDY activity. A two-layered LSTM is utilized for this model with the first LSTM layer having 64 units with return sequence enabled in order to capture and maintain the temporal properties at each time-step. This is followed by the second LSTM layer having 32 units to capture higher level temporal patterns. In order to avoid overfitting, a dropout rate of 0.2 was used on the layers. Also, the first LSTM layer has a bidirectional wrapper that processes the sequences both forward and backwards in time during offline training to enable it to identify the relationship between dextrose consumption and ethanol production phases.

This bidirectional processing is employed exclusively during offline model training on complete historical fermentation trajectories. For on-line deployment and real-time inference, the model operates in a strictly forward-only (causal) mode, where predictions at time t are generated using only data from time points 0 through t, with no access to future measurements. Additionally, a recurrent dropout rate of 0.1 is applied to the LSTM cells to improve the model’s ability to generalize while maintaining long term dependency (Lee et al., 2025; Metcalfe, 2025).

To refine predictions and improve interpretability, an attention mechanism is added after the final LSTM layer to focus on the most relevant time steps during phase transitions. Two dense layers are incorporated, the first dense layer consists of 16 units with ReLU activation to introduce non-linearity, while the output dense layer is configured with appropriate units and linear activation for continuous prediction targets such as CPDY activity, DCW, and substrate concentration. Residual connections are implemented between certain layers to improve gradient flow during training, ensuring robust predictions over extended fermentation periods.

Statistical analysis

Statistical analysis of the central composite design (CCD) experiments was performed using Design Expert® software (Stat-Ease Inc., Minneapolis, USA). A second-order quadratic model was fitted to the experimental data and its adequacy was assessed by analysis of variance (ANOVA). Model significance was established on the basis of the F-value and the model probability value (p < 0.05), together with a signal-to-noise ratio (Adequate Precision >4). The coefficient of determination (R 2), adjusted R 2, and predicted R 2 were computed to evaluate goodness-of-fit and predictive accuracy. A difference of less than 0.2 between the adjusted and predicted R 2 was taken as evidence that the model was not over-fitted. The coefficient of variation (C.V. %) was used as a relative measure of experimental reproducibility. Canonical analysis was applied to identify the stationary point (maximum, minimum, or saddle) of each response surface. For the CoYeast-LSTM model, predictive performance on continuous variables (CPDY activity, dry cell weight, and substrate concentration) was quantified using Mean Squared Error (MSE) and Mean Absolute Error (MAE) on the holdout test set (20%). A modified 5-fold time-series cross-validation strategy was employed, preserving temporal continuity within each fold according to a 60: 20: 20 (training: validation: testing) split. Parity plots for biomass and enzyme yield were assessed by Deming regression to account for measurement error in both axes, deviations of the slope from unity and the intercept from zero were used to identify systematic prediction bias. The error distribution was characterised by the interquartile range (IQR) of model residuals, and the cumulative distribution function (CDF) of absolute errors was used to determine the proportion of predictions falling within specified tolerances of ±0.05 U/mL and ±0.10 U/mL. Model adequacy was established from the significance of the overall model (p < 0.0001), the high coefficients of determination, and the close agreement between the adjusted and predicted R 2 values (differences of 0.1719 for DCW and 0.1096 for CPDY, both well below 0.2), which together indicate that the second-order models adequately represent the response surfaces without over-fitting. All statistical tests were conducted at a significance level of α = 0.05. In Table 1, the standard deviation (SD %) values for the input factor levels denote the spread of factor levels across the CCD runs as a percentage of the factor range, while the SD values for the responses (DCW and CPDY activity) denote the standard error of prediction at the canonical stationary point, computed from the residual mean square and the prediction variance at that point. Statistical methods for RSM and bioprocess model evaluation followed established guidelines for design-of-experiments-based enzyme production studies (Assega et al., 2025).

Hyperparameter tuning, training, testing and evaluation metrics of CoYeast-LSTM

The model hyperparameters were tuned to optimize predictive performance. To preserve the temporal structure of the data, the dataset was split into 60% training, 20% validation, and 20% testing partitions. A modified 5-fold time-series cross-validation was employed, maintaining temporal continuity within each fold. Model performance was evaluated using the regression metrics Mean Squared Error (MSE) and Mean Absolute Error (MAE) for the continuous target variables, namely, CPDY activity, dry cell weight (DCW), and substrate concentration.

Response surface methodology

A second-order quadratic model equation was obtained using the central composite design (CCD) technique of response surface methodology (RSM), a statistical tool used to correlate the relationship between variables controlling a reaction. The biomass production and, in turn, the enzyme activity of CPDY of the mixed culture (K. marxianus and S. cerevisiae) was predicted using CCD. Three parameters having a significant role in the process were considered for optimization. Dextrose is the carbon source, ammonium chloride is the nitrogen source and yeast nitrogen base provides amines, vitamins, amino acids, some trace elements, and salts as a basic nutrient for most fermentations. The effects of yeast nitrogen base (A), ammonium chloride (B) and dextrose (C) upon dry cell weight (DCW) and enzyme activity (CPDY) were analyzed. The desirability of the optimum process is to achieve maximum CPDY yield while keeping the other variables at minimum level. Dextrose serves as the primary carbon source for both K. marxianus and S. cerevisiae in the mixed culture system, directly influencing biomass accumulation and enzyme production. The optimization of dextrose concentration represents a critical balance between providing sufficient carbon for growth while avoiding substrate inhibition or osmotic stress. The quadratic relationship between dextrose concentration and enzyme activity suggests an optimal point where substrate availability maximizes without triggering inhibitory mechanisms. The implementation of fed-batch strategies, analyzed by RSM optimization results, can enhance enzyme production yields at industrial scales (Dubencovs et al., 2021; Easwaran et al., 2022). RSM optimization must consider economic factors alongside technical performance to ensure commercial viability. The cost of nitrogen sources represents a significant portion of fermentation costs, making the optimization of YNB and ammonium chloride concentrations economically important (Schneider, 2018; Srinivasan et al., 2020).

Adequacy and model fit summary of DCW and CPDY activity

The interaction between the control variables A (Yeast nitrogen base), B (Ammonium chloride), C (Dextrose) and the two responses were obtained by validating the quadratic equations (Equations 3, 4) obtained in terms of coded factor.

The model equations were optimized using montgomery optimization technique and are used for predicting the optimum control variables for maximum DCW and CPDY activity (Shanmugam and Mahadevan, 2015; Santharam et al., 2017). The model F-values of 79.83 for DCW and 68.42 for CPDY activity (Supplementary Table S3) were each derived from the ratio of the model mean square to the residual mean square, confirming that both quadratic models are statistically significant (p < 0.0001). The high coefficients of determination (R 2 > 0.98), adjusted R 2 (>0.96), and low coefficients of variation (C.V. % below 7) indicate robust predictive capability suitable for bioreactor scale-up (Metcalfe, 2025). The adjusted and predicted R 2 values are in reasonable agreement (differences of 0.17 for DCW and 0.11 for CPDY), indicating that the models are not over-fitted and that the selected variables adequately explain the observed variability. The residual standard deviations of 0.2327 and 0.3289 for DCW and CPDY, respectively, further confirm the adequacy of the fitted models. It was observed that the predicted versus actual data points lying closer to the straight line reveals that the quadratic model is significant for both the responses (Supplementary Figure S2). Experimental range and levels of reaction variables for DCW and CPDY are shown in Supplementary Table S1. Fit Statistics for DCW and CPDY activity are shown in Supplementary Table S2.

Response surface analysis of CPDY and DCW

The response surface methodology was used to investigate the individual and interaction effects of the three factors on CPDY production and dry cell weight. Based on ANOVA, yeast nitrogen base (A) and dextrose (C) had significant effects on CPDY production, whereas ammonium chloride was found to be non-significant. Among these, dextrose (C) exerted the strongest influence on CPDY production. Yeast nitrogen base (A) had a comparatively smaller effect on CPDY production than dextrose. For dry cell weight, all three factors had significant effects, with dextrose (C) exerting the strongest influence.

To study the interaction between media components for yeast growth (DCW) and CPDY activity, a graphical representation of the regression equation was plotted. RSM analysis revealed that the CPDY activity decreased with an increase in dextrose and yeast nitrogen base concentrations. From the response-surface regression, a maximum CPDY activity of 6.15 U/mL was obtained at 7 g/L yeast nitrogen base and 15 g/L dextrose, with activity decreasing at higher concentrations. The interactive effect between factors (dextrose and ammonium chloride) and (yeast nitrogen base and ammonium chloride) on CPDY activity is shown in Figure 1. It is apparent that the two factors dextrose and yeast nitrogen base are incompatible with ammonium chloride which exhibited an inverse association to CPDY activity. The interactive effect was significant, but the ammonium chloride has less influence than yeast nitrogen base and dextrose. The ANOVA results and model fit statistics are summarised in Supplementary Tables S2, S3.

It was also observed (Supplementary Figure S3) that a maximum DCW of 4.55 g/L was predicted at the canonical stationary point (YNB 7 g/L, NH 4 Cl 3.8 g/L, dextrose 15 g/L), which was achieved in the 2 L biocalorimeter (4.5 g/L). A lower DCW of 3 g/L was obtained in shake-flask culture, consistent with the improved mass transfer, pH control, and aeration characteristic of the stirred bioreactor. The interactive effects of ammonium chloride and dextrose at constant yeast nitrogen base, and of yeast nitrogen base and dextrose at constant ammonium chloride, were similarly examined, with dextrose exerting a stronger influence on DCW than either ammonium chloride or yeast nitrogen base. The outputs from canonical analysis are furnished in Table 1.

Canonical analysis of the same response surface located the stationary point at a slightly higher predicted maximum CPDY activity of 7.15 U/mL (Table 1). The corresponding experimental value obtained under the optimized medium composition in shake-flask culture was 8 U/mL, in close agreement with this prediction. In comparison, our previous study reported markedly lower CPDY activity, biomass production, and substrate consumption in the individual strains (Easwaran et al., 2022).

Shake flask studies

Shake-flask experiments were performed in triplicate (n = 3 independent biological replicates), and the error bars in Figure 2 represent the mean ± standard deviation of these replicates. For the bioreactor cultivation, the SD = ±3% reported in Table 2 corresponds to the reproducibility of triplicate analytical measurements. From Figure 2, it can be observed that the mixed culture of K. marxianus and S. cerevisiae yeast system adapts well to the optimized media with a minimum lag phase. Dextrose was entirely consumed in systems within 12 h. Yeast growth was initially high when it consumed dextrose. DCW reached 3 g/L when dextrose was completely exhausted. Further, the increase in DCW was due to the consumption of ethanol. CPDY activity increased as dextrose was depleted and declined upon onset of ethanol consumption.

Maximum CPDY activity of 8 U/mL was observed at 12 h. Subsequently, the metabolic activities were investigated using biocalorimetric and respirometric devices in a bioreactor. Earlier studies have reported a CPDY activity of only 1.12 U/mL and noted that the folding, transport, secretion, and maturation of immature and mature CPDY forms into the culture broth remain largely unexplored in recombinant S. cerevisiae. While the research successfully demonstrated the enhancement of CPDY expression through the engineering of signal sequences and the coexpression of folding accessory proteins, it does not explore the potential effects of varying fermentation conditions or the optimization of culture media on CPDY production, which could further improve yields and activity levels (Shin et al., 2014).

Microscopic and SEM morphological structure of the co-culture withdrawn from the reactor at 12 h is shown in Figures 3, 4 respectively. The microscopic images show well-developed isolated and mixed colonies of K. marxianus and S. cerevisiae. The cells of K. marxianus are oval-shaped and smaller in size (Figures 3b, 4b), whereas S. cerevisiae cells are round and larger (Figures 3c, 4c). Figures 3a, 4a shows the mixed culture morphology at 12 h, for size comparison.

Biocalorimetric and respirometric studies

Biocalorimetric studies were performed in the 2 L BioRC1e device to characterize real-time heat release and its relationship with substrate consumption, biomass growth, and product formation. The mixed yeast culture performed consistently under the optimised medium conditions. Referring to Figure 5, the aerobic fermentation starts without a lag phase which can be seen from the heat release curve. The metabolic heat rate increases gradually when substrate dextrose is consumed and reaches a maximum of 2.2 W at 9 h, then decreases to a heat rate of 1.45 W at 14 h. The heat rate again starts to increase and shows a second peak starting from 14 h, reaches a maximum of 1.6 W at 17 h and death phase arrives at 31 h. The second increase is due to the consumption of ethanol.

The process was aerobic fermentation at an optimized oxygen flow rate of 1 LPM fed to the reactor. DO concentration within the reactor was measured and the DO curve is a reciprocal of heat rate curve. Initial DO was measured as 30 ppm. The concentration decreases as the reaction progresses. Yeast consumes a huge amount of oxygen in aerobic fermentation, due to which the DO concentration reduces gradually and reaches zero at 9 h and during ethanol consumption, DO curve shows a second peak. Turbidity is a direct measure of biomass concentration within the reactor. Turbidity values increase from 0 h to 9 h due to biomass formation and reach a maximum of 500 Nephelometric Turbidity Units (NTU). After 9 h when the yeast starts to consume ethanol as a substrate, turbidity values increase very slowly and reach 560 NTU at the end of the stationary phase (31 h). pH values decrease to 2.5 due to the formation of acetic acid from 7 h. This metabolic heat and DO curves help to study the different substrate consumption patterns of the mixed yeast culture.

The offline parameters such as CPDY enzyme activity, biomass growth (DCW), dextrose consumption and metabolite formation (ethanol) were depicted in Figure 6a. As dextrose was consumed by the mixed yeast system metabolites and enzymes were produced. Ethanol concentration increases with a decrease in dextrose. Ethanol concentration of 20.88 g/L was observed at sixth h, decreases gradually and reaches 3 g/L at 32 h. The RSM-optimized dextrose (15 g/L) was supplemented to the basal YPD medium, which itself contains 20 g/L dextrose, giving a total initial dextrose loading of approximately 35 g/L in the bioreactor together with additional fermentable carbon contributed by the yeast extract. This total carbon availability accounts for the peak ethanol concentration of 20.88 g/L observed during the fermentative phase. After 6 h, yeast system consumes remaining dextrose and ethanol present in the reactor. Exhaustion of dextrose makes ethanol as sole carbon source consumed by S. cerevisiae yeast in the mixed yeast culture, leading to an increase in biomass concentration and ethanol consumption. Ethanol concentration should be maintained under a minimum range because higher ethanol concentration decreases the activity of enzymes responsible for the hydrolysis by performing conformational changes of the peptidases involved in the hydrolysis (Bleoanca et al., 2013). Ethanol increases the permeability of the membranes by affecting the transport of amino acids and dextrose leading to the reduction of cell growth rate and viability. Kluyveromyces marxianus, unlike S. cerevisiae, does not survive high ethanol concentrations. Ethanol has a wide range of harmful effects on yeasts, including effects on cell membranes and proteins, as well as the formation of reactive oxygen species (ROS) (de Moura Ferreira et al., 2022).

The reaction process has three phases. The first phase was the consumption of dextrose by the yeast system (0–9 h), mainly by K. marxianus, the second phase (Phase II, 9–14 h) was the diauxic shift and transitional co-consumption of residual dextrose and ethanol, and the third phase (Phase III, 14–31 h) was the sole ethanol respiration phase, within which the second calorimetric peak reached its maximum at 17 h. As reported in previous studies, mixed microbial systems exhibit higher CPDY activity than either individual yeast, as the consortium consumes greater quantities of substrate and ethanol and attains higher biomass, which together yield more enzyme than the individual systems. Table 2 shows the CPDY yield comparison between the individual and co-culture systems.

The heat yield values were calculated for the three phases (Figure 6b). Y Q X values were found to be 12.63 kJ/g for initial phase, 65.27 kJ/g for second phase and 15.99 kJ/g for the final phase. This is because biomass growth at initial and final phases was higher when compared to the second phase (Figures 6a,b). The respirometric analysis supports biocalorimetric observations where CER values show a double peak at 8 h and 18 h (Figure 6c). CER is higher during dextrose consumption and comparatively lower during ethanol consumption because dextrose has a higher calorific value than ethanol. As discussed above the third phase is completely a respiration process, whereas the first and second phases were partially respirometric and partially fermentative.

The calorespirometric ratio shows the nature of microbial growth (Figure 6d). The CR value was determined as, 410 kJ Cmol -1 CO 2 at 12 h, where the dextrose was completely exhausted, which ensures the aerobic fermentation metabolism of the reaction. Then followed by respiration mechanism with ethanol consumption confirmed by CR values higher than 1,000 kJ C-mol -1. To increase CPDY yield, substrate limitation is required. Biothermodynamic analysis will reveal the different substrate consumption pattern which is helpful in predicting the CPDY and biomass. This prediction is useful in avoiding the chances of enzyme degradation on achieving maximum activity.

CoYeast-LSTM predictive performance

The functional block diagram, architecture details and hyperparameter details of CoYeast-LSTM are illustrated in Figure 7 and Table 3 respectively.

Model training dynamics

The CoYeast-LSTM converged rapidly and the training MSE fell below 0.01 after the first epoch and both training and validation curves stabilized by epoch 15 (Figure 8a). Final training/validation values were 0.0043/0.0060 (MSE) and 0.015/0.018 (MAE) respectively, giving a <20% generalization gap (Figure 8b). Comparable scores on a 20% hold-out test set (MSE 0.0062 and MAE 0.019) confirm robustness. The early plateau coincides with the transition from dextrose to ethanol metabolism identified in the metabolic-heat fingerprint (Figure 6a), showing that the bidirectional architecture captures phase-shift dynamics. The model is sufficiently precise for real-time feed or harvest decisions in biomass-conversion bioreactors. As the offline variables were measured at 23 independent sampling points and temporally aligned to the continuous calorimetric stream by cubic-spline interpolation, model performance was additionally evaluated on the independent measurement points alone; this yielded a normalized MSE of 0.021 and MAE of 0.112, higher than the values obtained across the full interpolated series, reflecting the limited size of the single-fermentation dataset.

Heat-signal segmentation

A univariate histogram of the 1 Hz calorimetric data (Figure 9) reveals a distinctly bimodal distribution with modes at Q / Q max ≈ 0.02 and 0.98. Kernels centered at these modes map precisely onto the two metabolic regimes defined in Figure 6a: (i) high-heat dextrose oxidation (0–9 h) and (ii) low-heat ethanol respiration (14–31 h). The valley between the peaks (0.35 < Q / Q max < 0.65) corresponds to the transition window (9–14 h) where residual dextrose and nascent ethanol coexist. From a bioprocess viewpoint this confirms that calorimetry alone can serve as a single-sensor phase discriminator, allowing feed-forward control strategies that switch aeration or harvest timing when Q / Q max drops below ∼0.3. For the CoYeast-LSTM, the bimodality justifies using heat (Q) and its derivative (dQ/dt) as primary attention features. Their sharp phase-demarcation enables the network to localise enzyme-yield peaks with ±0.1 (normalized) accuracy. Such calorimetric fingerprints are increasingly relevant to bioprocess modeling where soft-sensor coverage is limited, yet real-time phase tracking is critical for enzyme cost minimisation.

Error distribution and residual analysis

The residual fan-plot (Figure 10a) displays a tight band (IQR ≈0.04 on the normalized scale) with all points falling between −0.14 and −0.02, indicating a consistent over-prediction bias of ∼0.07 across the normalized operating range (0.1–0.35). The box-and-whisker summary (Figure 10b) confirms this shift (median value, −0.065 on the normalized scale) and shows no extreme outliers, even during the ethanol-transition window (9–12 h) where process noise is highest. The empirical CDF (Figure 10c) reveals that 80% of absolute errors are ≤0.05 and 95% are ≤0.10 on the normalized scale, outperforming previously reported soft-sensor models for enzyme fermentation, which typically retain 60%–70% of errors within the same band (Metcalfe, 2025). This accuracy is well within the ±0.1 (normalized) envelope required to trigger real-time harvest decisions in lignocellulosic biorefineries, thereby validating the model for on-line deployment.

Biomass and enzyme yield predictions

The parity plot (Figure 11a) demonstrates excellent agreement between measured and predicted dry-cell-weight (DCW) across the full fermentation trajectory (normalized range 0.70–1.00). A Deming regression yields a slope of 0.97 ± 0.02 and an intercept of 0.01 (normalized), indicating a slight under-prediction bias at the lowest biomass levels but no systematic drift at higher titres. The root-mean-square error (RMSE) is 0.024 (normalized), corresponding to <3% of the working range and well below the ±0.05 (normalized) threshold typically required for feed-rate control in industrial yeast fermenters. This level of accuracy allows the CoYeast-LSTM soft-sensor to replace at-line gravimetric sampling (∼30 min lag) with real-time biomass visibility, enabling tighter oxygen-transfer and substrate-pulse strategies that can shorten batch time by 4%–6%. The low biomass prediction error achieved here reflects the advantage of combining the bidirectional LSTM architecture with calorimetric inputs for biomass estimation.

The CPDY parity plot (Figure 11b) confirms that the CoYeast-LSTM captures the full normalized activity range of 0–0.33, a Deming regression produced a slope of 0.98 ± 0.03 and an intercept of 0.004 (normalized), indicating minor under-prediction at very low titres but no systematic drift near the harvest peak (0.25–0.33, normalized). 92% of points fall inside the ±0.05 (normalized) envelope shown, matching the error distribution. For a 2 L calorimetric reactor, this resolution translates to <2% uncertainty in batch end-point detection sufficient to trigger timely cooling or substrate-feed shutdown, thereby avoiding enzyme denaturation and reducing downstream purification losses. This resolution in enzyme-activity prediction demonstrates the value of calorimetry-informed sequence modelling for real-time end-point detection in bioprocess applications.

Statistical performance metrics of Co-Yeast LSTM

The bidirectional CoYeast-LSTM out-performed traditional time-series benchmarks, yielding a mean-squared-error (MSE) of 0.0060 dimensionless (normalized scale) and a mean-absolute-error (MAE) of 0.071 on the normalized scale. Analysis of the attention weights indicates that the network assigns highest salience to metabolic heat flow (Q) and the calorespirometric ratio (CR) during the ethanol-respiration period (Phase III, 14–31 h), mirroring the biocalorimetric fingerprints. Coupling real-time calorimetry with deep sequence learning can materially enhance soft-sensor accuracy for enzyme fermentations. The model’s peak performance during the 17 h calorimetric peak within the ethanol-respiration phase (Phase III: 14–31 h) aligns with established fermentation kinetics (Mathew et al., 2015). This timeframe corresponds to the diauxic shift where yeast cells transition from glucose fermentation to ethanol respiration (Luzia et al., 2024). During this critical period, respiratory capacity gradually increases, metabolic heat production patterns shift significantly and calorespirometric ratios become highly informative for process monitoring (Easwaran et al., 2022). The model’s ability to mirror biocalorimetric fingerprints during this phase demonstrates sophisticated pattern recognition capabilities for complex metabolic transitions. Biocalorimetry has been re-affirmed as a universally applicable process analytical tool, with metabolic heat serving as a quantitative proxy for substrate consumption rate, biomass growth, and enzyme activity across diverse aerobic and anaerobic bioprocesses (Saravana Raj et al., 2026). Concurrently, the deployment of LSTM-based soft sensors within machine learning operations (MLOps) frameworks for industrial-scale fed-batch fermentation has highlighted the practical necessity of real-time predictive models that are robust to concept drift and operational variability capabilities that the CoYeast-LSTM architecture is well-positioned to deliver (Metcalfe, 2025).

CoYeast-LSTM was evaluated against three baseline models trained on identical datasets such as ARIMA (autoregressive integrated moving average, order selected by AIC), a feedforward neural network (FNN) with two dense hidden layers (64 and 32 units) using the same input features, and a standard unidirectional single-layer LSTM without attention or residual connections. All models were evaluated on the same test partition and normalized data. As shown in Table 4, CoYeast-LSTM achieved the lowest test-set MSE (0.0060) and MAE (0.071), outperforming ARIMA (MSE = 0.031, MAE = 0.143), FNN (MSE = 0.019, MAE = 0.109), and the standard LSTM (MSE = 0.012, MAE = 0.089).

Design rationale and justification for the CoYeast-LSTM architecture

The biocalorimetric and respirometric data from yeast co-cultivation exhibits temporal characteristics that makes LSTM networks particularly suitable for modelling this bioprocess. In this study, the metabolic heat rate displays a biphasic pattern, with an initial peak at 9 h followed by a second peak at 17 h. These peaks correspond to the consumption of dextrose and ethanol, respectively which can be effectively captured by a LSTM network with a unique memory cell architecture. LSTMED (LSTM-Autoencoder) approaches for process monitoring demonstrate effectiveness in capturing uneven dynamic features and long-term time dependency (Deng et al., 2023). The co-cultivation process exhibits dynamics across different time scales - from rapid changes in DO levels to slower changes in biomass accumulation. Traditional time series models struggle with such multi-scale temporal patterns, whereas LSTM’s gating mechanisms can selectively remember or forget information, making them ideal for modeling processes with varying timescales (Espinoza et al., 2025). The complex interactions between metabolic heat, DO, substrate utilization, and enzyme production are highly non-linear. It was observed that the yeast first produces ethanol during fermentation. Later, the yeast consumes that ethanol. This change affects how well CPDY enzyme activity is maintained over time. LSTM networks can capture these non-linear relationships through their activation functions and recurrent structure (Lee et al., 2025). The co-cultivation process spans over 31 h with critical phase transitions. When modelling such extended temporal sequences, traditional RNNs suffer from vanishing or exploding gradient problems. LSTM’s specialized architecture was designed specifically to overcome these issues through its gating mechanisms, making it suitable for modelling the complete fermentation cycle (Rusch et al., 2022). The production of CPDY involves sequential biological events where earlier conditions significantly impact later outcomes. LSTM models have demonstrated success in learning optimal sequential decision-making patterns, which is relevant for understanding the metabolic shifts observed in the three phases of the reaction process (Helleckes et al., 2023).

This study introduces a transformative approach to enzyme production through co-cultivation of K. marxianus and S. cerevisiae, achieving a CPDY activity of 16 U/mL in the biocalorimeter, representing a 33% increase over K. marxianus monoculture (12 U/mL) and a 100% increase relative to S. cerevisiae monoculture (8 U/mL), demonstrating the synergistic advantage of the division-of-labour consortium design. Response surface methodology identified an optimized medium composition (7 g/L yeast nitrogen base, 3.8 g/L ammonium chloride, and 15 g/L supplemental dextrose) that supported these improved yields. Real-time calorimetric fingerprinting revealed biphasic metabolic transitions, with heat flux and calorespirometric ratios serving as precise phase markers for process control. The advanced bidirectional CoYeast-LSTM model delivered high predictive accuracy (MSE 0.0060 and MAE 0.071 on the normalized [0 to 1] scale, corresponding to 1.24 U/mL RMSE and 1.14 U/mL MAE for CPDY), leveraging attention mechanisms focused on metabolic heat flow and ratios during critical phases, enabling accurate CPDY prediction, significantly enhancing real-time process monitoring and harvest precision. Economically, this methodology reduces offline sampling and operational costs, with potential applicability to larger scales. The integration of mechanistic insights, biocalorimetric monitoring, and deep learning establishes a novel, scalable paradigm that addresses enzyme cost and process sustainability challenges, paving the way for next-generation biomanufacturing aligned with industry 4.0. Future directions include multi-omics integration, fed-batch strategies and multi-strain co-cultivation for further improvements.