Ensuring the stability of enzymes is crucial for their industrial implementation, as it increases operational stability, prolong activity and thus increases total turnover number, and improve cost-efficiency of the process (Žnidaršič-Plazl, 2021b; Woodley, 2022). To systematically evaluate DES candidates for performance goals, we first narrowed the list to those documented in the literature as beneficial for dehydrogenases, focusing specifically on those known to stabilize FDH. Studies suggest that polyol-based DESs, containing either choline chloride or betaine as HBA, are the most effective stabilizing media for various dehydrogenases, including FDH (Bittner et al., 2022; Gajardo-Parra et al., 2023). First, seven polyol-based DESs with either choline chloride or betaine as HBA were tested at three water contents (up to 50%, w/w) for their ability to stabilize FDH. It was confirmed that glycerol-based DES containing both tested HBAs are optimal candidates for stabilizing the enzyme upon prolonged incubation at room temperature (data not shown).

Glycerol-based DESs were recently demonstrated to stabilize NAD coenzymes (Radović et al., 2022). Besides, Leron and Li (2013), Leron et al. (2013) and Biswas et al. (2023) reported that among several choline chloride-based DES containing urea, ethylene glycol, and glycerol as HBDs, the one with glycerol had the highest CO₂ solubility. The above studies also emphasize that water plays a crucial role in enzyme performance and CO₂ solubility. Finally, concerning the toxicological footprint, glycerol-based DES are considered non-toxic and biodegradable (Radošević et al., 2015).

Based on the above considerations, we selected two glycerol (Gly)-based DESs, with choline chloride (ChCl) or betaine (B) as HBA in a molar ratio of 1:2, and prepared the corresponding solutions in water (10%–90%, w/w). In parallel, we also prepared solutions in 50 mM potassium phosphate buffer (pH 7.5) to keep the pH close to the enzyme’s optimal value and to possibly prevent a pH drop when CO₂ is added to the reaction medium. As previously mentioned, DESs diluted with more than 50% water (w/w) can be considered aqueous solutions of DES components (Hammond et al., 2017). Nevertheless, these mixtures were included in the study, as a high water content within DES is often essential for enzymes to sustain their catalytic activity (Taklimi et al., 2023).

A total of 20 DES-based solvents were prepared and characterized for their physicochemical properties (pH, density, and viscosity) relevant to the reaction (Table 1). As expected, the densities and viscosities of the DES aqueous solutions were strongly influenced by the water/buffer content, peaking at mixtures with 10% water (up to 1.21 g cm ^- ³ and 353.70 mPa s for B:Gly_10%W, and 1.17 g cm ^- ³ and 82.63 mPa s for ChCl:Gly_10%B). In general, B:Gly-based mixtures were denser and more viscous than their ChCl:Gly-based counterparts. All mixtures tested had pH values ranging from 5.3 to 9.2, with betaine-based mixtures being more acidic than ChCl-based ones. Dissolving DES in buffer generally maintained the solutions at pH values between 7.5 and 8. All DES mixtures remained stable for 3 months under laboratory conditions and showed no signs of contamination or precipitation.

For statistical analysis and mathematical modelling (Section 3.4.), the identification of a molecular representation that converts the component structures into descriptive features for numerical evaluation is essential (Venkatraman et al., 2018). An advanced and accurate molecular representation is the σ-profile (sigma profile), an unnormalized histogram of the screened surface charge of a molecule (Klamt, 2005). σ-profiles are distinguished from other representations by the fact that they capture nuanced effects such as polarizability and electron density asymmetry (Abranches et al., 2022). The σ-profile can be divided into three key regions: (i) the HBD region with negative charge densities, (ii) the non-polar region with nearly neutral charge densities, and (iii) the HBA region with positive charge densities (Figure 3A). This division is based on the fact that each atom in an HBA or HBD molecule is identifiable by a distinct peak with a specific screening charge density (σ) value (Lemaoui et al., 2020).

Here, the σ-profile of each DES-based mixture was calculated using BOVIA COSMOtherm software: the σ-profile curves for each HBA and HBD were divided into 10 regions, and the area under each region was calculated considering the molar ratios of the components and the water content (Supplementary Table S1). For glycerol, the σ-profile reveals peaks at negative polar coordinates (left side) corresponding to the positively polar H atoms in the -OH group, while peaks at positive polar coordinates (right side) correspond to the O atoms in the -OH group (Figure 3A) (Cheng et al., 2018). These polar regions interact with opposite polar segments in a solution. The σ-profile’s extension into strongly polar regions (−0.022 e/Ų < σ < −0.01 e/Ų on the left and 0.01 e/Ų < σ < 0.013 e/Ų on the right) is asymmetric, indicating that glycerol has a larger positive polarity surface area. This electrostatic misfit suggests that glycerol tends to act as a HBD and thus shows greater affinity for HBA in a solution (Cui et al., 2021). In contrast, the σ-profile of water extends more symmetrically into strongly polar regions, indicating its balanced ability to act as both a HBD and HBA. Finally, for ChCl, as well as for betaine (Radović et al., 2024), the strongest peak appears in the non-polar region, corresponding to the cholinium cation, followed by a peak in the HBA region, associated with the Cl⁻ anion. Figure 3A illustrates the sigma surfaces of the DES components (choline chloride, glycerol and water) generated by TmoleX19. The colors represent a calculated charge gradient, ranging from charge-deficient to charge-dense regions: HBD regions are labelled as deep blue and HBA regions as deep red on the surface. Non-polar regions are marked in green (Quaid and Reza, 2023). Figure 3B shows that even small changes in the ChCl:Gly mixture, such as increasing the water content from 10% to 90%, result in solvents with different polarity distributions. This demonstrates that the software is capable of capturing nuanced phenomena, which is crucial for exploring the chemical landscape of these solvents and understanding their potential impact on enzyme and coenzyme behavior, as well as the solubility of reaction participants. We have recently demonstrated the same ability of the software for the system betaine:ethylene glycol with 3 water proportions (10, 30, 50% water, w/w) (Radović et al., 2024).

The activity of FDH in the reference buffer and in aqueous solutions of DESs at different concentrations were measured by monitoring the oxidation of sodium formate. For the stability test, enzyme solutions were incubated in selected solvent systems at 30°C for 14 days and residual enzyme activity (A _Res ) was measured at regular intervals (Table 1). Both in the reference buffer and in the mixtures with DESs, FDH inactivation followed first-order kinetics, allowing us to use this kinetic model to calculate the FDH half-life (t _{1/2, FDH} ) (Figure 4).

As anticipated, FDH showed little or no activity in the DES mixtures containing ≤50% (w/w) water or buffer, while the activity increased with the addition of water and peaked in the highly diluted mixtures (90% water, w/w), although the values were still lower than those observed in the buffer (residual activities, A _R , between 44% and 92%). Interestingly, it appears that dilutions with water generally resulted in better enzyme activity than their counterparts diluted with buffer, although buffered systems are closer to the enzyme’s optimal pH of 7.5 (determined experimentally, data not shown). For example, ChCl:Gly with 90% water content (ChCl:Gly_90%W) with a pH of only 6.3 had an A _R of about 92%, while its counterpart with the buffer (ChCl:Gly_90%B) with a pH of 7.5, yielded the A _R of only 44.5%.

DESs with a water/buffer content in the range of 30% - 80% (w/w) showed the best ability to stabilize the enzyme, with half-lives of up to 64 days, which is much higher than in the reference buffer, where t _{1/2, FDH} was 2.1 days (Figure 4). On the other hand, highly diluted DESs (90% water/buffer, w/w) demonstrated a stabilizing effect on the enzyme comparable to that of the reference buffer.

The stability of FDH in buffered DES solutions (80% buffer, w/w) was also significantly increased compared to the reference buffer, with t _{1/2, FDH} of 65.6 days and 29.9 days for B:Gly_80%B and ChCl:Gly_80%B, respectively. Again, DESs solutions with a water/buffer content of only 10% led to a faster destabilization of the enzyme. These results are consistent with previous findings that most dehydrogenases require more than 10% water content in DES to maintain their structural integrity. DESs absorb water in their hydrogen bonding network, reducing the availability of free water molecules required for enzyme hydration. This reduction in water activity can lead to dehydration and irreversible denaturation of the enzyme (Mourelle-Insua et al., 2019; Bittner et al., 2022; Radović et al., 2024). Overall, solvent systems with 80% buffer (w/w) exhibited an optimal balance between FDH activity and stability, which was particularly evident for B:Gly_80%B, where the A_R was 51.7% (Figure 4).

The saturated dissolved CO₂ concentrations (c _s ) in various DES solutions with water/buffer were evaluated after introduction of CO₂ at a flow rate of 100 mL min⁻¹ until saturation. The results are presented in Table 1. All DESs mixtures with ≤50% water or buffer were poor media for dissolving CO₂, with c _s values between 282 and 913 mg L⁻¹, which is lower than those observed in the reference buffer (1,029 mg L⁻¹). Highly diluted DESs led to similar CO₂ solubilities as in the buffer, with the highest improvements observed in ChCl:Gly_90%B and B:Gly_90%W, with values of 1,149 and 1,112 mg L⁻¹, respectively.

To investigate the long-term stability of the NADH coenzyme, the changes in the UV-Vis absorption spectra of NADH were observed during a 14-day incubation at 25°C in the solvent systems described above. During incubation, the absorbance loss at 340 nm followed the first-order kinetics used to calculate the NADH half-lives shown in Figure 4. The results clearly show that DES composition plays an important role in the coenzyme degradation rate. Virtually all DES aqueous solutions, except ChCl:Gly_90%B (t _{1/2, NADH} = 2.9 days) and B:Gly_90%W (t _{1/2, NADH} = 0.8 days), stabilized the coenzyme compared to the reference buffer (t _{1/2, NADH} = 4.6 days). In general, ChCl-based DESs were more suitable for coenzyme stabilization than betaine-based DESs. This was particularly pronounced for DESs with 10% buffer content (e.g., for ChCl:Gly_10%B, t _{1/2, NADH} was 44.2 days, while for B:Gly_10%B, t _{1/2, NADH} was 17.5 days). As evident from Figure 4, a higher DES content had a positive effect on the ability of solvent systems to stabilize the coenzyme for all DESs tested. This was most evidenced for B:Gly solutions in water, where the t _{1/2, NADH} for B:Gly_30%W and B:Gly_90%W was 65.7 and 0.8 days, respectively.

To further navigate the DES screening, the short-term NADH stability in DES aqueous solutions saturated with CO₂ was investigated. The presence of dissolved CO₂ not only affects the intrinsic properties of the DESs but also the behavior of the system after substrate addition. It has been previously reported that the acidification of the reaction medium by dissolving CO₂ leads to enhanced NADH degradation, which directly affects the conversion of CO₂ to formic acid (Zhang et al., 2018). Therefore, NADH solutions in the solvent systems described above were monitored over a period of 90 min, and the corresponding degradation constants (k _NADH* ) were calculated using a first-order kinetic model (Table 1). In general, buffer-diluted DESs were equally or more successful in stabilization of the coenzyme over the tested period (k _NADH* ≤ 0.003 min⁻¹) than the reference buffer (k _NADH* = 0.003 min⁻¹), while water-based solvent systems with DESs were poor media in this regard, especially at high water contents (k _NADH* up to 0.009 min⁻¹). This effect is directly related to the inability of water-diluted DESs to maintain pH close to neutral. For example, after the introduction of CO₂ into B:Gly_80%W, pH decreased to 3.75, resulting in the highest observed k _NADH* value of 0.008 min⁻¹. Moreover, buffered solutions of ChCl-based DESs maintained a higher pH than betaine-based DES solutions after CO₂ saturation, resulting in complete stabilization of NADH over the time tested, except for ChCl:Gly_50%B (k _NADH* = 0.003 min⁻¹).

The correlations between the physicochemical properties (pH, viscosity and density) of the DESs used, the DES descriptors (σ-profiles), FDH performance, long-term NADH stability and CO₂ solubility were analyzed using the Spearman correlation matrix, which was selected due to the non-normal data distribution. The analysis confirmed our above assumptions and findings from the experimental screening: the targeted properties of the reaction affected by the DES composition led to contradictory results (Table 2). First, both FDH activity and stability in tested solvent systems showed negative correlations with NADH stability. In addition, FDH activity values demonstrated negative correlations with all analyzed variables except CO₂ solubility, favouring aqueous/buffered media over DES solutions with lower amounts of aqueous phase. Besides, negative correlations were found also for FDH and NADH stability with CO₂ solubility (Table 2).

The Spearman correlation matrix shown in Table 2 highlights the complex interplay between the physicochemical properties of DES (pH, density, viscosity) and the DES descriptors with the performance characteristics of FDH and the stability of its cofactor NADH. As expected, significant negative correlations were observed between FDH activity and DES density/viscosity, likely due to the lower mobility and diffusion of substrates and enzymes. Conversely, NADH stability showed significant positive correlation with the density/viscosity of the solvent systems with DESs, which can be attributed to the slower overall dynamics of the solvents providing a less detrimental environment for dissolved NADH.

FDH activity and NADH stability are significantly impacted by pH value, while for FDH stability there is no correlation with this chemical property. It is well-established that enzyme activity (Bisswanger, 2017) and NADH stability (Zachos et al., 2019) are pH-dependent. However, the consistent reports on pH-independent ability of DESs to stabilize various enzymes remains puzzling. This intriguing observation hints at other mechanisms, such as direct interactions between DES components and proteins or with nearby water molecules, potentially altering the medium’s water activity (Damjanović et al., 2024).

Furthermore, the analysis revealed that nearly all DES descriptors significantly impacted the targeted properties. Specifically, S ¹ _mix and S ² _mix (HBD region, medium polarity), S ³ _mix - S ⁵ _mix (nonpolar region, positive charges), S ⁶ _mix – S ⁸ _mix (nonpolar region, negative charges) and S ⁹ _mix (HBA region) (Lemaoui et al., 2020), all demonstrated a significant influence on the properties being studied. FDH activity exhibited a negative correlation with all the descriptors, except for S ² _mix , which showed a positive correlation. Conversely, both FDH stability and NADH stability demonstrated positive correlations with all the descriptors, except for S ² _mix , which displayed a negative correlation. Interestingly, the analysis suggests an inverse relationship between enzyme activity and stability in DESs: solvents rich in HBA and non-polar domains stabilize the enzyme (and coenzyme), while HBD-rich solvents enhance enzyme activity but may lead to destabilization. These findings emphasize the delicate balance between enzyme activity and stability in DESs, driven by their specific compositional properties (as described by σ-descriptors). Similar interplay between the enzyme’s active and stable (but inactive) states, influenced by the water content in DESs, has been confirmed in our recent study on lysozyme behaviour in DESs based on various naturally occurring osmolytes (Damjanović et al., 2024). Insights into this relationship could guide the rational design of DESs for optimized biocatalysis applications: by performing similar statistical analyses on a larger set of DESs, it may be possible to predict an ideal σ-profile shape, and thereby identify or design the most suitable DES for a specific purpose.

According to discussed above, our next step was to see if it was possible to develop a simple QSAR model to summarize the relationship between the targeted properties (FDH activity, FDH stability, and NADH stability) and the DES descriptors (Figure 3; Supplementary Table S1) along with the physicochemical properties (Table 1) of the DESs using piecewise linear regression (PLR) (Figure 5). The latter is a powerful tool for modelling complex relationships in a simple and interpretable way, especially when the relationship between variables changes at a certain point. The input variables of the PLR models were selected based on the significant correlations in the Spearman correlation matrix. The relationship between the observed data and model predictions was also estimated using the coefficient of determination for prediction (R _pred ² ), the adjusted coefficient of determination for calibration (R _pred ² _adj ), the root mean square error of prediction (RMSEP), the ratio of prediction to deviation (RPD) and the ratio of the error range (RER).

As shown in Table 3 and Figure 5, the developed models describe the experimental data with high precision. The best agreement between the experimental data and the data predicted by the model was obtained for the FDH activity (Figure 5A) (R _cal ² = 0.914, R _cal ² _adj = 0.913, RMSEC = 7.203%, R _pred ² = 0.905, R _pred ² _adj = 903, RMSEP = 7.379%, RPD = 5.589, RER = 18.324). On the other hand, the largest scatter between the model and experimental data was found for the FDH half-life (Figure 5B) (R _cal ² = 0.829, R _cal ² _adj = 0.827, RMSEC = 4.775 days, R _pred ² = 0.738, R _pred ² _adj = 0.731, RMSEP = 4.944%, RPD 3.292, RER = 10.456).

According to Hussain et al., an R ² value of 0.75 is considered significant, an R ² value of 0.50 is considered moderate, and an R ² value of 0.26 is considered weak (Hussain et al., 2018). Furthermore, models with RPD < 1.4 are considered non-reliable, those with RPD in the range from 1.4 to 2 are considered fair, while models with RPD > 2 are described as excellent models (Chang et al., 2001). Models with RER > 4 are acceptable for data screening, models with RER > 10 can be used for quality control, while models with RER > 15 can be used for quantification (Sim et al., 2023). Therefore, the PLR models developed for the prediction of FDH activity and NADH half-life based on R _pred ² can be considered substantial, while the model developed for the prediction of FDH half-life can be considered moderate. Based on the RPD values, all three models developed can be considered reliable. And based on the RER values, the model developed for the prediction of FDH activity can be used for quantification (RER = 18.324), while the other two modes can be used for quality control. Therefore, it can be concluded that the feasibility of mathematical models for predicting targeted properties or applications of DES using easily measurable physicochemical properties and chemical descriptors, as demonstrated here, could be valuable for both industrial applications and research efforts focusing on these solvents. Additionally, utilizing these QSPR models may not only assist in predicting the properties of interest but also provide valuable insights into the relationship between the structure of the DES and its measurable properties. By analyzing how various structural features influence the targeted properties, these models can help unravel the underlying mechanisms driving behavior of biomolecules in DESs. This understanding can inform the design and optimization of DESs, leading to more effective and tailored applications in various fields.

We have demonstrated that the evaluation of different targets related to the tested reaction often leads to contradictory results regarding the optimal DES. For example, the enzyme dissolved in B:Gly_30%B showed remarkable stability with a half-life of 40.1 days, while its relative activity was less than 2% of that in the buffer. This DES also showed average performance in the long-term stability of the coenzyme, with a half-life of 12.1 days. In contrast, the enzyme dissolved in B:Gly_90%B maintained a high relative activity, which was 71.2% of that in the reference buffer, but was one of the worst candidates for enzyme stabilization with a half-life of only 6.0 days. In general, “concentrated” DESs (<50% water/buffer, w/w), with η > 18 mPa s showed high efficacy in stabilizing both the enzyme and the coenzyme, but also high viscosity, which poses significant challenges for scaling up processes with these solvents. Based on these observations, it is crucial to reconcile these results by finding a DES that optimally fulfils the desired properties and thus contributes to the overall efficiency of the process. A graphical representation shown in Figures 6A, B illustrates the trade-off between the performance of ChCl:Gly and B:Gly, respectively, with different buffer proportions with respect to the targeted property values (FDH activity/stability, CO₂ solubility, NADH stability in saturated CO₂ solutions, and compared to the reference buffer). The radar chart is bounded by the respective lower and upper limits of each target property, with ratings, ranging from 0 to 100, reflecting the performance of DES-based solvents relative to the best candidate for each target property. At this point, DESs diluted with water were omitted due to their poor ability to stabilise NADH after acidification of the medium due to CO₂ introduction (Table 1).

As can be seen from Figure 6A, ChCl:Gly_80%B has a balanced distribution across all target property values, which is crucial for the simultaneous optimization of the reaction where all design objectives are equally important. In particular, this DES, which in this case could be considered an additive rather than a solvent (Hammond et al., 2017), had the highest values for all target properties except for FDH activity, where the reference buffer resulted in the highest values. It should be emphasized that ChCl:Gly_80%B had an almost 15-fold higher t _{1/2, FDH} value compared to the buffer and stabilized the coenzyme more effectively when CO₂ was dissolved in the solvent system. B:Gly_80%B also showed similar balanced behavior to ChCl:Gly_80%B, but was ineffective in stabilizing NADH after dissolving CO₂.

Since ChCl:Gly dissolved with 80% (w/w) buffer (ChCl:Gly_80%B) was found as the most promising solvent system for the FDH-catalyzed conversion of CO₂ to formate, the reaction was performed in a medium pre-saturated with CO₂ and product formation was monitored over time. The amount of formate produced by the enzymatic CO₂ reduction in the selected solvent system and in the reference buffer is shown in Figure 7. The reaction performed in the DES-supplemented medium yielded 26.5 μmol mL⁻¹ of formate with a volumetric productivity of 6.6 μmol mL⁻¹·h⁻¹, while the reaction in the buffer yielded 22.7 μmol mL⁻¹ with a volumetric productivity of 5.6 μmol mL⁻¹·h⁻¹. These differences can be attributed to the slightly lower solubility of CO₂ in the buffer and the pronounced degradation of NADH in the buffer due to acidification by dissolved CO₂ (Wu et al., 1986).

NADH stability was monitored in a separate experiment without enzyme addition: after 4 h (corresponding to the reaction time), the NADH concentration in the buffer decreased to 78% of its initial value, while in the DES medium the NADH concentration remained above 98% (data not shown). The shape of the concentration vs. time curve reflects the interplay of several factors, including enzyme activity, CO₂ solubility and availability, and NADH degradation dynamics. The variations in formate concentration between the two solvents observed during the first 120 min may be attributed, on one hand, to the higher enzyme activity in the buffer compared to ChCl:Gly_80%B, and, on the other hand, to the enhanced CO₂ solubility and NADH stability provided by the DES. It is presumed that this dynamic interplay results in the reaction rate being sometimes higher in the buffer and at other times in the DES.

In addition to improving volumetric productivity, the true potential of using ChCl:Gly_80%B lies in its ability to stabilize FDH and thus extend the enzyme’s half-life by up to 15-fold compared to the buffer. Although this DES property is not fully utilized in a reaction lasting only 4 h, it is of great advantage in a continuous process (e.g., using an enzymatic membrane reactor). Under steady-state conditions, the volumetric productivity of the DES-assisted process could be significantly increased, as the productivity inversely correlates with the enzyme’s deactivation rate constant. Therefore, the productivity of the selected process under study was estimated for continuous operation mode comprising the enzyme inactivation rate. Considering the following assumptions: (i) that the process runs under steady-state conditions, (ii) that the enzyme is continuously deactivated over time according to first-order kinetics, and (iii) that the substrate concentration remains relatively constant, the reaction performed in buffer over 10-day period would yield 185.7 μmol mL⁻¹ formate, while the reaction carried out in ChCl:Gly_80%B over the same period would yield 639.0 μmol mL⁻¹ formate, which is approximately a 3.5-fold improvement. The calculation of overall productivity included estimation of the Michaelis–Menten kinetic parameters by fitting the NADH concentration profiles to the differential equation describing the change in substrate concentration over time. The results showed that both the maximum reaction rate (v _max) and the NADH saturation constant (K _s) were higher for the solvent system with DES. In the buffer, the constant values were v _max = 0.091 μmol mL⁻¹ min⁻¹ and K _s = 6.746 μmol mL⁻¹, while for the systems with DES, the constants were v _max = 0.133 μmol mL⁻¹ min⁻¹ and K _s = 12.133 μmol mL⁻¹.

Finally, downstream processing was not part of this study. Nevertheless, based on the available literature, liquid–liquid extraction with the green solvent 2-methyltetrahydrofuran (Slater et al., 2016; Laitinen et al., 2021) appears to be a promising option for the separation of dilute aqueous formate solutions in the context of developing sustainable formic acid production.

For the enzymatic reactions performed in this study, the NADH-dependent formate dehydrogenase (FDH) from Pseudomonas sp. 101 was used (see Supplementary Information) (Tishkov et al., 1993). Carbon dioxide (CO₂) with a purity of 99.5% was acquired from Messer Croatia Plin (Zaprešić, Croatia), while all other chemicals were purchased from Sigma-Aldrich (St. Louis, Missouri, United States). All materials had a purity of at least 99% and were used as received without further purification.

For the preparation of DESs based on betaine (B) and choline chloride (ChCl), the hydrogen bond acceptor (HBA) and the hydrogen bond donor (HBD) were mixed in a molar ratio of 1:2 with water or 50 mM potassium phosphate buffer (pH 7.5) in defined proportions (Table 1). Prior to use, ChCl was dried in a vacuum concentrator at 60°C for 24 h. The mixtures were stirred and heated to 60°C until a colourless and homogeneous liquid was formed. All prepared DESs were stored in sealed bottles until further use. The pH values of the prepared DESs were measured using a pH glass electrode (Mettler Toledo, Greifensee, Switzerland). The properties of the prepared DESs (pH, density and viscosity) were determined at 25°C. Density was determined using the pycnometric method and the viscosity using a rotary viscometer (Anton Paar ViscoQC 300, Ashland, Virginia, United States). All measurements were performed in triplicates. The σ-descriptors of the DESs were calculated according to Panić et al. (2022) (Supplementary Table S1).

To determine FDH activity, the FDH enzyme, NAD⁺ coenzyme, and formate substrate were added sequentially to various solvent systems to achieve final concentrations of 1.5 mg mL⁻¹, 0.1 mg mL⁻¹, and 10 mg mL⁻¹, respectively. The total working volume of the assay was 250 µL. The NADH formation rate was measured immediately in a 96-well plate for 5 min at 340 nm using a UV-Vis spectrophotometer (SpectraMax^® ABS Plus, Molecular Devices, San Jose, CA, United States).

To evaluate FDH stability, stock solutions of the enzyme (12.8 mg mL⁻¹) were prepared in the tested DES solutions and the reference buffer. These solutions were stored in sealed vials at 30°C in the dark. Aliquots were taken at regular intervals over a 14-day period and analyzed for FDH activity using the method described above.

The first-order deactivation rate constant (k _FDH , day^-1) was evaluated from the first-order kinetic model for the decrease of residual enzymatic activity over time (A _Res (t)) (Equation 1): k F D H = 1 t ln A R e s , 0 A R e s   t (1) where A _Res is residual enzyme activity (either at time zero, A _R,0, or at time t, A _R (t)). The kinetic parameters were estimated by fitting the experimental data to the nonlinear equation using the Levenberg–Marquardt algorithm implemented in WR Mathematica 10.0 (Wolfram Research, Champaign, United States).

The half-life of the enzyme (t _{1/2, FDH} , day) was calculated using the previously determined k _FDH (Equation 1), according to Equation 2: t 1 / 2 , F D H = ln ⁡ 2 k F D H (2)

Measurements of the stability of the coenzyme NADH in various solvent systems (0.03 mg mL⁻¹) were monitored for up to 14 days. The samples were stored in the dark at 25°C and the absorption spectra in the range from 230 to 400 nm were recorded regularly using a UV-Vis spectrophotometer (SpectraMax^® ABS Plus, Molecular Devices, San Jose, CA, United States). Each measurement was carried out in triplicate. The absorption spectrum of NADH shows a characteristic peak with an absorption maximum at 260 nm, which is due to the adenosine monophosphate moiety, and another peak at 340 nm, which is due to the neutral nicotinamide moiety. The decrease in absorbance at 340 nm followed first-order kinetics (Equation 3): k N A D H = 1 t   ln A 0 A t (3) where k _NADH is the first-order degradation rate constant (day^-1), A is the absorbance at 340 nm, either at time zero, A ₀, or at time t, A _t (Wu et al., 1986).

The NADH half-life (t _{1/2, NADH} , day) was calculated using the previously determined k _NADH (Equation 3), according to Equation 4: t 1 / 2 , N A D H = ln ⁡ 2 k N A D H (4)

Moreover, the stability of NADH (25 mg mL⁻¹), added in tested solvent systems after their saturation with CO₂, was assessed to evaluate the impact of dissolved CO₂ on the coenzyme. Monitoring was accomplished by measuring the absorbance decrease at 340 nm using a UV-Vis spectrophotometer for 90 min. The absorbance decrease at 340 nm followed the first-order kinetics, and the first-order rate degradation constants (k _NADH*, day^-1) were calculated in the same manner as described above. For the FDH-catalyzed reduction of CO₂ in the reference buffer and ChCl:Gly_80%B, NADH stability was monitored throughout the entire course of the reaction (6 h).

The CO₂ solubility measurement, based on the setup described by Obert and Dave (Obert and Dave, 1999) was conducted in a glass tube where CO₂ was bubbled through the tested solvent systems with DESs and the reference buffer (V = 5 mL) for 90 min (time sufficient to reach a saturation point) at a flow rate of 100 mL min⁻¹. The gas was introduced using a small nozzle with an approximate diameter of 1 mm. To prevent CO₂ loss, the glass tube was sealed with parafilm. The concentration of CO₂ was continuously monitored using a Mettler Toledo (Greifensee, Switzerland) CO₂ sensor InPro 5000i/120.

Prior to the reaction initiation, both 50 mM potassium phosphate buffer (pH 7.5) and ChCl:Gly_80%B were saturated with CO₂ by bubbling both solvents for 90 min, resulting in CO₂ dissolved concentrations of 1,029 mg mL⁻¹ and 1,057 mg mL⁻¹, respectively. The enzymatic reaction was initiated by adding NADH and FDH to the CO₂-saturated buffer or ChCl:Gly_80%B to reach final concentrations of 16.4 mg mL⁻¹ and 28.8 mg mL⁻¹, respectively. The reactions occurred at a room temperature using a magnetic stirrer in small tubes with a working volume of 250 µL. Aliquots were sampled at specific intervals to monitor NADH consumption, which was measured spectrophotometrically at 340 nm, indicating enzyme activity in reducing CO₂. To validate the results and assess the stability of NADH under experimental conditions, control reactions were conducted without the FDH enzyme. These control mixtures allowed us to monitor NADH stability in both the buffer and ChCl:Gly_80%B throughout the reaction period, as described in Section 5.4. To verify that the consumed NADH was utilized for formate synthesis, the concentration of formate was additionally determined using the method described by Lang and Lang (Singh et al., 2018). Briefly, samples (25 µL) containing formate were mixed with 50 µL of solution A, 2.5 µL of solution B, and 175 µL of 100% acetic anhydride. The mixture was incubated at 50°C for 2 h with occasional mixing. Formation of red color was subsequently measured spectrophotometrically at 515 nm using the SpectraMax^® ABS Plus (Molecular Devices, San Jose, CA, United States). Solution A was prepared by dissolving 0.5 g of citric acid and 10 g of acetamide in 100 mL of isopropanol; solution B was prepared by dissolving 30 g of sodium acetate in 100 mL of water. For standard calibration, sodium formate dissolved in 50 mM potassium phosphate buffer (pH 7.5) was used.

It was assumed that FDH activity, FDH and NADH stability can be described as a function of the DES physical properties and σ-profile of the mixture, expressed by a set of S ⁱ _mix descriptors: A R  or  t 1 / 2 , F D H  or  t 1 / 2 , N A D H = f ρ , η , p H , S mix 1 , S mix 2 , S mix 3 , S mix 4 , S mix 5 , S mix 6 , S mix 7 , S mix 8 , S mix 9 , S mix 10

Piecewise linear regression (PLR) models have been used to describe the relationship between input and output variables. Input variables were selected based on the Spearman correlation matrix (Equation 5). A R  or  t 1 / 2 , F D H  or  t 1 / 2 , N A D H = ( b 01 + b 11 · ρ + b 21 · η + b 31 · p H + b 41 · S mix 1 + b 51 · S mix 2 + b 61 · S mix 3 + b 71 · S mix 4 + b 81 · S mix 5 + b 91 · S mix 6 + b 101 · S mix 7 + b 111 · S mix 8 + b 121 · S mix 9 + b 131 · S mix 10 ) · A R  or  t 1 / 2 , F D H  or  t 1 / 2 , N A D H ≤ b n + ( b 02 + b 12 · ρ + b 22 · η + b 32 · p H + b 42 · S mix 1 + b 52 · S mix 2 + b 62 · S mix 3 + b 72 · S mix 4 + b 82 · S mix 5 + b 92 · S mix 6 + b 102 · S mix 7 + b 112 · S mix 8 + b 122 · S mix 9 + b 132 · S mix 10 ) · A R  or  t 1 / 2 , F D H  or  t 1 / 2 , N A D H > b n (5)

The PLR technique is based on estimating the parameters of two linear regression equations: one for dependent variable values (Y) less than or equal to the breakpoint (b _n) and the other for dependent variable values (Y) higher than the breakpoint. The PLR parameters were estimated using the Levenberg-Marquardt algorithm implemented in the software Statistica 14.0 (Tibco Software Inc., Palo Alto, United States). The data set (63 data points for each output variable) was randomly split 70:30 into a calibration and a prediction data set. The applicability of the developed calibration models was estimated using the coefficient of determination for calibration (R _cal ²), the adjusted coefficient of determination for calibration (R _cal ² _adj ), and cosmo (RMSEC). Predictive performance of the models was estimated using the coefficient of determination for prediction (R _pred ²), the adjusted coefficient of determination for calibration (R _pred ² _adj), the root mean square error of prediction (RMSEP), the ratio of prediction to deviation (RPD) and the ratio of the error range (RER) (Fearn, 2002).

The productivity of the FDH-catalysed CO₂ conversion to formate was estimated based on the enzyme reaction kinetics including the enzyme inactivation rate (Bisswanger, 2017). Michaelis–Menten kinetic parameters were estimated by fitting the NADH concentration profiles to the differential equation (Equation 6) using WR Mathematica 10.0 (Wolfram Research, Champaign, United States). The change in substrate concentration over time without enzyme inactivation reads: d c N A D H d t = − v max · c N A D H K S + c N A D H (6)

Considering enzyme inactivation, the following expression for the reaction rate v(t) (Equation 7) is obtained: v t = v max · c N A D H K S + c N A D H · e − k F D H · t (7)

The productivity of the FDH-catalyzed CO₂ conversion to formate can be calculated by integrating (Equation 7) from the beginning of the reaction (t = 0) to its end (t = t _f ) to obtain (Equation 8): p r o d u c t i v i t y = ∫ 0 t f v max · c N A D H K S + c N A D H · e − k F D H · t d t = v max · c N A D H K S + c N A D H · 1 − e − k F D H · t f k F D H (8)