Insights into ionic liquid-enhanced membrane protein stability through machine learning and molecular simulations

Abstract

Protein stability plays a critical role in structural elucidation, enzyme activity, and the storage of protein drugs, where ionic liquids (ILs) have emerged as promising protein stabilizers due to their exceptional biocompatibility and superior solubility. However, the underlying mechanisms by which ILs modulate protein stability, particularly through the regulation of hydrogen bonding and interfacial structures, remain inadequately understood. Herein, a machine learning-based framework, integrating molecular docking, unsupervised learning, molecular dynamics simulations and correlation analysis, is applied to clarify the mechanism of ILs enhancing membrane protein stability. It is found that ILs form clusters that are adsorbed on the protein surface, with ILs entering the hydration layer of the protein and forming intermolecular hydrogen bonds with the protein surface, thereby improving stability, consistent with experiments. Furthermore, a predictive model for protein stability is established by supervised learning and verification of the mechanism through interpretability analysis. Our framework quantitatively reveals the influence of hydrogen bonds and interface structures on membrane protein stability. Overall, these quantitative results not only deepen our understanding of the interactions between ILs and protein but also shed light on the rational design of protein stabilizers.

---

1. Introduction

The structural analysis of membrane proteins remains a pivotal challenge in the life sciences.^1–3 Although advanced prediction tools like AlphaFold have achieved remarkable accuracy in determining the three-dimensional structure of proteins,^4,5 experimental validation is still needed to establish the structure–function relationships.⁶ However, a major obstacle is that membrane proteins tend to lose their native conformation during extraction and purification,⁷ which necessitates optimized methods to preserve their stability. Ionic liquids (ILs) have emerged as effective agents in the extraction and purification of membrane proteins,^8–11 owing to their excellent solubility, biocompatibility, and thermal stability.^12,13 ILs not only enhance extraction efficiency and increase protein yields but also act as stabilizers.^12,14 For instance, Brogan et al. demonstrated via circular dichroism spectroscopy that the half-denaturation temperature of myoglobin increased by 55 °C in pyrrole-based ILs, even surpassing water's boiling point.¹⁵ Similarly, Banerjee et al. found that insulin maintained its secondary structure in the presence of choline and ILs at 4 °C for up to 4 months.¹⁶ These results highlight the significant role of ILs in the field of protein stabilization.

Understanding the stabilization mechanism of proteins by ILs involves investigating two major factors: hydrogen bonding and protein–IL interface structure (Fig. 1). Some studies, such as those by Phillips et al., demonstrated using wide-angle X-ray scattering that imidazolium ILs primarily disrupt internal hydrogen bonds within the protein, thereby reducing the structural stability of proteins.¹⁷ In contrast, Ghanta et al. showed through molecular dynamics (MD) simulations that ILs predominantly break the hydrogen bonds between the protein and water, while forming new hydrogen bonds with the protein surface to enhance its stability.¹⁸ Moreover, while certain studies using UV-visible absorption spectroscopy and far-UV circular dichroism suggested that ILs mainly affect the hydration layer of proteins by replacing water molecules, thereby enhancing protein stability,^19–22 other investigations proposed that the aggregation of cations into ionic clusters encapsulates the protein surface, thereby reducing structural fluctuations.^8,9,23,24 The interplay between the internal and interfacial hydrogen bonds, along with the various protein–IL interfacial interactions, critically determines overall protein stability. As those interactions can enhance or disrupt structural integrity, developing a unified quantitative framework to explain IL-mediated stabilization remains a significant challenge.

Fig. 1 The dominant stability mechanism of protein in IL-containing aqueous solution is considered from two main perspectives: hydrogen bonding and interface structure. The general workflow includes five major steps: (i) dataset generation, (ii) classification, (iii) screening of IL candidates, (iv) mechanistic analysis, and (v) model verification.

In this study, the membrane was not included in the MD simulations, as our focus was on the storage process of the protein after extraction from the membrane. Nevertheless, our previous studies demonstrated that ionic liquids preferentially insert at the protein–lipid interface while preserving protein structural stability.^9,22 Building on these results, we develop a machine learning (ML)-based framework to quantitatively elucidate how ILs enhance membrane protein stability (Fig. 1). Our approach encompasses five major steps: (i) dataset generation: we compile a dataset of 317 different anions paired with imidazolium cations and perform molecular docking simulations with aquaporin-2 (AQP₂) and sodium channel protein (Nav) to calculate binding energies between anions and protein (E_B). (ii) Classification: an unsupervised learning model is used to classify the anions into nine clusters, integrating their structural features with the E_B data to construct a ML model. (iii) Screening of IL candidates: we sample and evaluate ILs using more rigorous MD simulations based on E_B and clusters. In contrast to every sample investigated, we screenedby focusing on anions with the highest E_B and representative candidates from each cluster. (iv) Mechanistic analysis: key descriptors related to hydrogen bonding and interface structure are extracted from the trajectories of MD simulations, and correlation analysis is performed to identify the key factors influencing protein stability. (v) Model verification: finally, we verify the proposed mechanism by constructing a supervised learning model utilizing an extra trees regressor, with interpretability provided by the Shapley additive explanation (SHAP) analysis.

2. Results and discussion

2.1. Dataset for ionic liquids and binding energies with protein

Initially, we compiled a dataset of 317 anions, including their SMILES codes and structural files (Fig. S1), obtained from the ionic liquid database.²⁵ We then selected aquaporin 2 (AQP₂) and sodium channel protein (Nav) as our model proteins because they play crucial roles in cellular water and ion transport (Fig. S2), making their stability essential for both physiological functions and biomedical research. Subsequently, we performed molecular docking studies of these anions with the target proteins AQP₂ and Nav to obtain binding energies (E_B) and identify the corresponding binding sites,²⁶ thereby establishing a comprehensive dataset (Fig. 2a and Fig. S3). This dataset includes the SMILES codes of the anions, their respective PDB structural files, E_B, and the corresponding binding sites. Anions that exhibit stronger binding energies to AQP₂ also display stronger binding energies to Nav. Accordingly, we selected AQP₂ for further analysis. Given that the E_B is closely related to the specific binding site,²⁷ an analysis of these sites (Fig. 2b) revealed four distinct regions: the outside cavity, channel center, inside cavity, and hydrophobic surface. Notably, the outside and inside cavities of AQP₂ are characterized by predominantly carrying negative and positive charges, respectively.²⁸ Among these, the inside cavity accounts for 73.7% of all binding sites (Fig. 2c), which can be attributed to the fact that under electrostatic interactions, anions tend to bind to the positively charged inside cavity of AQP₂.

Fig. 2 (a) Hierarchical clustering dendrogram. The leaf nodes show the position where all anions are partitioned into 9 clusters, marked as I–IX from left to right and distinguished by different colors, and representative structures of each cluster are displayed in the blank spaces. Mapping the dendrogram to the binding energy reveals the structure–property relationships of the anions. (b) Structure of AQP₂, where I, II, III, and IV represent the most stable binding sites for anions. (c) Probability distribution of binding sites. (d) Binding energy distribution of anions within each cluster; bar colors correspond to cluster categories in (a).

We next performed hierarchical clustering analysis on the 317 anions using the Ward algorithm, resulting in 9 distinct clusters containing 19, 19, 34, 21, 38, 13, 68, 18, and 87 anions, respectively (Fig. S1). Notably, anions grouped within the same cluster exhibit similar structural motifs, indicating a rational basis for this classification. When correlating these structural features with their corresponding E_B, we found that anions within the same cluster exhibit comparable E_B values, indicating that we have established a robust structure–activity model that validates the rationality of our classification.

To further refine the structure–activity model of the anions, we analyzed both the functional groups of the anions and the distribution of their E_B values (Fig. 2d). Anions in clusters I to III all contain carboxyl groups but differ in their overall composition.

Specifically, cluster I anions are primarily monocarboxylic compounds, exhibiting E_B values ranging from −3 to −5 kcal mol⁻¹. In contrast, cluster III anions not only contain carboxyl groups but also incorporate other functional groups such as ether, amine, and hydroxyl groups. Particularly, cluster VII anions generally exhibit higher E_B, ranging from −4 to −10 kcal mol⁻¹. This cluster mainly comprises fluorinated anions with sulfonic acid groups, which can form hydrogen bonds with the protein (Fig. S4), suggesting the significant role of hydrogen bonding in these interactions.⁷ Additionally, cluster IX anions, characterized by a wide structural diversity that includes sulfonate groups, imidazole rings, azides, halogens, etc., exhibit a broad distribution of E_B values spanning from −1 to −8 kcal mol⁻¹.

2.2. High-throughput screening through molecular dynamics simulations

To investigate the effects of ILs on protein stability through MD simulations, we selected imidazolium cations with varying alkyl chain lengths [C_nmim]⁺ (n = 4–14)^13,29 and chloride as the counter ion due to its high water solubility.⁸ As shown in Fig. S5, with the increase in alkyl chain length, the adsorption state of ILs on the protein surface evolves from disordered adsorption to membrane-like adsorption, while the clustering structures in solution gradually transition from free ions to large clusters. Protein stability was characterized by root-mean-square deviation (RMSD) and α-helix content; as shown in Fig. 3a and b, the two-dimensional plot of α-helices and RMSD shifts from the center to the bottom-left corner, before returning to the center as the side chain of the cation elongates. This trend suggests an initial enhancement in stability, followed by a decline, peaking at [C₆mim]⁺. Consequently, [C₆mim]⁺ was chosen for further investigation.

Fig. 3 (a) Time evolution of the RMSD of the protein in aqueous solutions containing [C_nmim]Cl. (b) Effects of cationic alkyl chain length on protein stability, reflected by α-helicity and RMSD. (c) Average RMSD of the protein in solutions with [C₆mim]⁺ based ILs containing different anions, selected based on top E_B and varying clusters of the same E_B. The inset plots represent the model of the MD simulation, with the protein positioned at the center. The [C₆mim]⁺ are orderly arranged around the protein, while the anions are randomly distributed and not shown. (d) Structures of representative anions corresponding to the conditions in (c).

Subsequently, representative ILs were selected for MD simulations based on E_B and classification. The criteria involved selecting the anion with the highest E_B across all structures, followed by gradient selection of anions with varying E_B within the same cluster, and choosing anions with identical E_B from different clusters. Firstly, we chose the top 7 anions with the highest E_B (Fig. 3, named T₁ to T₇), all exceeding −6.9 kcal mol⁻¹. These anions typically have large molecular masses and contain fluorine and oxygen groups, and were paired with [C₆mim]⁺. As shown in Fig. 3c, the ILs with the highest E_B all enhanced protein stability, with RMSD values of 1.53, 1.95, and 2.15 Å for T₄, T₃, and control groups, respectively. Simulation snapshots reveal that these ILs form large clusters and adsorb on the protein (Fig. S6). Additionally, due to the hydrophilic sulfonic acid groups and hydrophobic fluorinated alkyl chains present in T₂, T₄, and T₇ anions (Fig. 3d), they resemble phospholipid structures,³⁰ promoting the formation of an ordered membrane-like structure around the protein and further stabilizing it.

To assess the influence of anions from the same cluster with different E_B, 7 anions with different E_B from cluster IX (E₁ to E₇) were selected, as shown in Fig. S7. The E_B values for E₁, E₅, and E₇ were −2.7, −5.3, and −7.4 kcal mol⁻¹, respectively, with the corresponding protein RMSD values of 1.71, 1.59, and 1.58 Å (Fig. S8), indicating that the effect of anions from the same cluster on protein stability is not linearly correlated with the change of E_B. In systems containing E₁ to E₃, ILs limited adsorption on the protein surface (Fig. S9), resulting in minimal stabilization effects. However, ILs in systems with E₄ to E₇ form clusters around the protein, leading to marked improvements in stability. Therefore, anions with higher E_B within the same cluster significantly enhance protein stability.

Since E₅ with a binding energy of 5.3 kcal mol⁻¹ demonstrated the most significant enhancement in protein stability, we further selected anions with E_B values around 5.3 kcal mol⁻¹ from different clusters (C_I to C_IX), as shown in Fig. 3d. Among these, anions from clusters C_VIII, C_V, and C_IV exhibited the lowest RMSD values of 1.48, 1.48, and 1.52 Å, respectively. These ILs form clusters adsorbed on the protein surface (Fig. S10), contributing to enhanced stability. To further evaluate their performance, the five screened ILs with the lowest RMSD were compared to three conventional protein stabilizers: polysorbate 20 (Tw20),³¹ polyethylene glycol (PEG),³² and Nonidet P-40 (NP-40).¹² As shown in Fig. 3c and Fig. S11 and S12, the screened ILs outperformed commercial stabilizers, reducing RMSD by 12–18%. We further evaluated structural stability using root mean square fluctuation (RMSF) and solvent accessible surface area (SASA) of the hydrophobic core. As shown in Fig. S13a, the C_VIII system shows a markedly reduced RMSF relative to the control, with fluctuations ranging from 10 to 40 Å for the control and 0 to 10 Å for C_VIII. Consistently, the hydrophobic core is less exposed in C_VIII, with a SASA of 12 953 Ųversus 24 005 Ų for the control (Fig. S13b). These metrics corroborate that the candidate ILs enhance the structural stability of the membrane protein. Additionally, these ILs exhibited low synthetic accessibility score³³ (SAscore) (Fig. S14), with the C_v anion scoring 2.19, indicating favorable synthetic feasibility. Consequently, these ILs are potential protein stabilizers.

2.3. Machine learning revealing the mechanism of ionic liquids stabilizing membrane proteins

To elucidate the mechanism by which ILs enhance protein stability, two categories of descriptors were extracted from the simulation results: hydrogen bond and interface structure. As shown in Fig. 4a, hydrogen bond descriptors include the number of internal hydrogen bonds within the protein (HB_pro), the number of hydrogen bonds between the protein surface and ILs (HB_pro-IL), and the number of hydrogen bonds between the protein and water (HB_pro-wat). Interface structure descriptors include the number of water molecules in the first and second hydration layers (C_I-wat, C_II-wat), anions (C_I-an, C_II-an), cations (C_I-ca, C_II-ca) near the protein surface, as well as the overall radius of gyration of the ILs (R_IL). Through correlation analysis, the Pearson correlation coefficients (r) between these descriptors were calculated to quantify their influence on protein stability (Fig. 4b). The sign of r indicates the direction of correlation: a positive sign for a positive correlation and a negative sign for an inverse correlation.

Fig. 4 (a) Descriptors representing hydrogen bonds and interface structures in IL–protein systems. (b) Pearson's correlation coefficients among descriptors. (c) Pearson's correlation coefficients between RMSD and individual descriptors. (d) Canonical correlation analysis between RMSD and descriptor sets for hydrogen bonds (HB_com) and interface structures (S_com), with the first canonical correlation coefficients indicated. (e) Predictive performance of the machine learning model for RMSD. (f) Feature importance based on SHAP values, with mean SHAP values shown on the right.

As shown in Fig. 4b and c, HB_pro-IL exhibits a moderate negative correlation with RMSD (r = −0.39), indicating that an increased hydrogen bonds between the protein surface and ILs is associated with enhanced structural stability. This aligns with previous experiments.^34,35 However, HB_pro has a negligible correlation (r = −0.05), indicating that the ILs have limited influence on the internal hydrogen bonds of the protein and mainly affects surface interactions, consistent with reports that the structure of membrane proteins is stable and that ILs have difficulty disrupting their internal interactions.²² On the other hand, for the interface structure, the C_I-wat, C_II-wat, C_I-an, and C_II-an have moderate correlations with RMSD, with all |r| values exceeding 0.36, indicating the importance of the hydration layer in protein stability. Notably, C_II-an is strongly negatively correlated with C_II-wat (r = −0.94), suggesting that IL anions can stabilize proteins by replacing water in the hydration layer. Moreover, the correlation between C_II-ca and RMSD (|r| = 0.27) is weaker than that of C_II-an (|r| = 0.39), indicating that anions play a more significant role in stabilizing proteins, consistent with previous studies.³⁶ Additionally, R_IL shows a moderate positive correlation with RMSD (Fig. 4a) (r = 0.31), implying that more compact IL structures are associated with greater protein stability.^24,25

In addition, to comprehensively analyze the impact of hydrogen bonds and interface structures on protein stability, canonical correlation analysis was performed. The strength of correlations between the two descriptor sets was evaluated using the first canonical correlation coefficient (r_c) in Fig. 4d, which ranges from 0 (no linear relationship) to 1 (perfect linear relationship). The hydrogen bond set (HB_com) and the interface structure set (S_com) exhibit moderate correlations with RMSD, with r_c = 0.44 and 0.41, respectively, indicating that both contribute to protein stability, aligning with previous findings.⁷ Additionally, a strong correlation was observed between HB_com and S_com, with r_c = 0.88, further emphasizing the close relationship between hydrogen bonds and interface structures. Based on the above analysis, the results suggest that ILs enhance protein stability by forming ionic clusters adsorbed onto the protein surface, penetrating the hydration layer, and establishing intermolecular hydrogen bonds with the surface. Moreover, the role of anions is more pronounced than that of cations.

To validate the proposed mechanism, supervised learning was employed to predict the RMSD of protein using the extra trees regressor. Descriptors selected from MD simulations and correlation analysis were used as input features. The model was optimized to minimize the root-mean-squared error (RMSE) between predicted and actual RMSD values. A Pearson correlation coefficient of 0.913 was achieved on the test set (Fig. 4e and Fig. S15), indicating high predictive accuracy and supporting the relevance of the selected descriptors to protein stability. SHAP analysis indicated that the top four most influential features are R_IL, C_II-wat, HB_pro-IL and C_II-an (Fig. 4f). Particularly, decreased R_IL and C_II-wat, along with increased HB_pro-IL and C_II-an, were associated with enhanced predicted stability. These results provide strong support for the proposed mechanism.

This study systematically elucidates the mechanisms by which ILs influence protein stability. A notable strength lies in the broad selection of ILs with diverse chemical properties, enhancing the generalizability of the findings. Building on prior work, we integrated multiple stability-related factors and applied statistical analyses to quantitatively evaluate their contributions. The proposed mechanisms are consistent with experimental observations. However, as the analysis focused on membrane proteins with inherently strong internal interactions, future studies should extend to other protein classes, such as globular and fibrous proteins, to establish a more comprehensive understanding of IL-induced stabilization effects.

3. Conclusions

In conclusion, we developed a machine learning-integrated framework combining molecular docking, unsupervised learning, MD simulations, and correlation analysis to elucidate how ILs enhance membrane protein stability. This framework was successfully applied to elucidate the mechanisms by which ILs enhance membrane protein stability, focusing on hydrogen bonds and interface structures. Initially, we employed molecular docking and unsupervised learning to develop a structure–activity model that correlates IL structures with E_B. The screened ILs showed a 12–18% improvement in performance compared to conventional stabilizers. Next, correlation analysis reveals the mechanism that ILs form ionic clusters that are adsorbed on the protein surface, with ILs penetrating into the protein hydration layer. These ILs form intermolecular hydrogen bonds with the protein surface, thereby stabilizing the protein. Finally, the model predicts the RMSD of the protein using an extra trees regressor, and SHAP values were utilized to identify the key explanatory variables influencing RMSD, which verified the mechanism. This work reveals the mechanism by which ILs improve protein stability, which is of great significance for understanding the interaction between ILs and proteins and developing new membrane protein extraction and storage systems.

4. Models and methods

4.1. Molecular docking

The molecular docking between aquaporin-2 (AQP₂), sodium channel protein (Nav) and ILs was conducted using the AutoDock Vina program.³⁷ The crystal structure of AQP₂ (PDB: 4nef²⁸) and Nav (PDB: 5vb8) were utilized, and the protein and ILs’ input files were prepared by merging non-polar hydrogen atoms and adding charges. The target protein remained fixed. The docking results were visualized using PyMOL, while the protein–ligand interaction was calculated using the protein–ligand interaction profiler web service.³⁸

4.2. Unsupervised learning

The dendrogram function in the SciPy package is used to implement agglomerative hierarchical clustering (AHC).³⁹ In this method, each sample begins as its own individual cluster and progressively merges with others based on similarity as it moves up the hierarchy. The result of AHC is a bottom-up hierarchical tree diagram. In this diagram, the Euclidean distance between branches serves as the measure of similarity, while the Ward algorithm is employed to calculate the dissimilarity between nodes.⁴⁰

4.3. Molecular dynamics simulations

The IL–AQP₂ system was constructed as shown in Fig. 3c, [C_nmim]⁺ (n = 4–14) cations were orderly arranged around AQP₂,⁴¹ and anions were randomly placed using Packmol.⁴² Water molecules surrounded the protein and ILs. The simulation box dimensions were set to 13.6 × 13.6 × 13.6 nm³. Periodic boundary conditions were applied in all three directions. The MD simulations were performed using Amber22 software.⁴³ The time step was 2 fs. The protein was described using the protein.ff14SB force field, while the second-generation General AMBER Force Field (GAFF2) was employed to describe the ILs,^44,45 which has been shown to reliably reproduce their structural and dynamical behaviors.^8,46–48 The TIP3P force field was used for water molecules.⁴⁶ Non-bonded interactions included electrostatic and van der Waals terms. The former was computed using the particle-mesh Ewald algorithm,^47,48 while the latter was described using the Lennard-Jones potential with a cutoff distance of 1.0 nm. The SHAKE algorithm was employed to constrain high-frequency vibrations of hydrogen atoms.⁴⁹ Temperature and pressure were controlled using the Berendsen thermostat⁵⁰ and the Berendsen barostat,⁵¹ respectively, with a coupling constant of 1.0 ps. The system was first equilibrated for 10 ns under the NPT ensemble at 310 K and 1 bar, followed by 300 ns of equilibration under the NVT ensemble at 310 K. The last 50 ns trajectory was used to analyze the impact of ILs on protein stability.

4.4. Correlation analysis and the supervised learning model

Correlation analysis was performed using the Statistical Package for the Social Sciences (SPSS)⁵² to correlate the number of internal hydrogen bonds within the protein (HB_pro), the number of hydrogen bonds between the protein surface and ILs (HB_pro-IL), the number of hydrogen bonds between the protein and water (HB_pro-wat), and the number of water molecules in the first and second hydration layers of the protein surface (C_I-wat and C_II-wat), the number of anions in the first and second hydration layers of the protein surface (C_I-an and C_II-an), the number of cations in the first and second hydration layers of the protein surface (C_I-ca and C_II-ca), and the overall radius of gyration of the ILs (R_IL) for the root mean square deviation (RMSD) of protein. The above descriptors are extracted from the last 50 ns trajectory of each MD to build a dataset.
where d_i is the Euclidean distance between atom i in the two structures and N is the total number of atoms.

The model training was conducted using scikit-learn.⁵³ The dataset was randomly divided into two groups: 80% for model training and 20% for model testing. The extra trees regressor was used, the maximum depth of the trees is unrestricted, and each leaf node must contain at least one sample. The ensemble model consists of 100 trees, and its performance is measured using the root mean square error (RMSE) calculated by:

where n represents the number of samples, y_i is the actual target value, and z_i is the predicted value.

Author contributions

Yanlei Wang and Jing Ren conceived and led the project. Ju Liu conducted the simulations and drafted the manuscript. Ju Liu and Guiming Zhang conducted the research. Cheng Song, Yanlei Wang, and Hongyan He contributed to this work by analyzing the data and revising the manuscript. Yanlei Wang and Hongyan He supervised the project. All authors discussed the results and commented on the manuscript.

Conflicts of interest

The authors declare no competing financial interests.

Data availability

The data supporting this article have been included as part of the supplementary information (SI). Supplementary information: the relevant data on the enhancement of protein stability by ILs. See DOI: https://doi.org/10.1039/d5cp02130h.

Acknowledgements

We are thankful to the Beijing Natural Science Foundation (2252011), the National Natural Science Foundation of China (22522817 and 22508386), U35 (New Teacher Start up Fund) Project of Renmin University of China (25XNKJ13), the Beijing Nova Program (20230484478) and the China Postdoctoral Science Foundation (GZC20250782 and 2025M77150) for financial support. This research was supported by Public Computing Cloud and Big Data and Responsible Artificial Intelligence for National Governance, Renmin University of China. The authors sincerely appreciate Prof. Suojiang Zhang (IPE, CAS) for his careful academic guidance and great support.

References

1. Y. Lu, C.-X. Yue, L. Zhang, D. Yao, Y. Xia, Q. Zhang, X. Zhang, S. Li, Y. Shen and M. Cao, et al., Structural basis for inositol pyrophosphate gating of the phosphate channel XPR1, Science, 2024, 386(6723), eadp3252.
2. C. Shen, S. Chang, Q. Luo, K. C. Chan, Z. Zhang, B. Luo, T. Xie, G. Lu, X. Zhu and X. Wei, et al., Structural basis of BAM-mediated outer membrane β-barrel protein assembly, Nature, 2023, 617(7959), 185–193.
3. R. Sexton, M. Fazel, M. Schweiger, S. Pressé and O. Beckstein, Bayesian nonparametric analysis of residence times for protein–lipid interactions in molecular dynamics simulations, J. Chem. Theory Comput., 2025, 21(8), 4203–4220.
4. J. Abramson, J. Adler, J. Dunger, R. Evans, T. Green, A. Pritzel, O. Ronneberger, L. Willmore, A. J. Ballard and J. Bambrick, et al., Accurate structure prediction of biomolecular interactions with AlphaFold 3, Nature, 2024, 630(8016), 493–500.
5. D. Coskun, M. Lihan, J. P. G. L. M. Rodrigues, M. Vass, D. Robinson, R. A. Friesner and E. B. Miller, Using AlphaFold and experimental structures for the prediction of the structure and binding affinities of gpcr complexes via induced fit docking and free energy perturbation, J. Chem. Theory Comput., 2024, 20(1), 477–489.
6. H. Liu, P. Yan, Z. Zhang, H. Han, Q. Zhou, J. Zheng, J. Zhang, F. Xu and W. Shui, Structural mass spectrometry captures residue-resolved comprehensive conformational rearrangements of a G protein-coupled receptor, J. Am. Chem. Soc., 2024, 146(29), 20045–20058.
7. P. Bharmoria, A. A. Tietze, D. Mondal, T. S. Kang, A. Kumar and M. G. Freire, Do ionic liquids exhibit the required characteristics to dissolve, extract, stabilize, and purify proteins? Past-present-future assessment, Chem. Rev., 2024, 124(6), 3037–3084.
8. J. Liu, Y. Wang, C. Wang, J. Gao, W. Cui, B. Zhao, L. Zhang, H. He and S. Zhang, Thermodynamical origin of nonmonotonic inserting behavior of imidazole ionic liquids into the lipid bilayer, J. Phys. Chem. Lett., 2021, 12(40), 9926–9932.
9. J. Liu, J. Ren, S. Li, H. He and Y. Wang, Protein interface regulating the inserting process of imidazole ionic liquids into the cell membrane, J. Phys. Chem. B, 2024, 128(18), 4456–4463.
10. Q. Zhao, H. Chu, B. Zhao, Z. Liang, L. Zhang and Y. Zhang, Advances of ionic liquids-based methods for protein analysis, Trac-Trends Anal. Chem., 2018, 108, 239–246.
11. E. Suarez Garcia, C. F. Miranda, M. T. Cesario, R. H. Wijffels, C. van den Berg and M. H. M. Eppink, Ionic liquid-assisted selective extraction and partitioning of biomolecules from macroalgae, ACS Sustainable Chem. Eng., 2023, 11(5), 1752–1762.
12. H. Chu, Q. Zhao, J. Liu, K. Yang, Y. Wang, J. Liu, K. Zhang, B. Zhao, H. He and Y. Zheng, et al., Ionic liquid-based extraction system for in-depth analysis of membrane protein complexes, Anal. Chem., 2022, 94(2), 758–767.
13. Q. Zhao, F. Fang, Y. Liang, H. Yuan, K. Yang, Q. Wu, Z. Liang, L. Zhang and Y. Zhang, 1-dodecyl-3-methylimidazolium chloride-assisted sample preparation method for efficient integral membrane proteome analysis, Anal. Chem., 2014, 86(15), 7544–7550.
14. M. Nuruzzaman, B. M. Colella, C. P. Uzoewulu, A. E. Meo, E. J. Gross, S. Ishizawa, S. Sana, H. Zhang, M. E. Hoff, B. T. W. Medlock, E. C. Joyner, S. Sato, E. A. Ison, Z. Li and J. Ohata, Hexafluoroisopropanol as a bioconjugation medium of ultrafast, tryptophan-selective catalysis, J. Am. Chem. Soc., 2024, 146(10), 6773–6783.
15. A. P. Brogan and J. P. Hallett, Solubilizing and stabilizing proteins in anhydrous ionic liquids through formation of protein-polymer surfactant nanoconstructs, J. Am. Chem. Soc., 2016, 138(13), 4494–4501.
16. A. Banerjee, K. Ibsen, T. Brown, R. Chen, C. Agatemor and S. Mitragotri, Ionic liquids for oral insulin delivery, Proc. Natl. Acad. Sci. U. S. A., 2018, 115(28), 7296–7301.
17. D. M. Phillips, L. F. Drummy, D. G. Conrady, D. M. Fox, R. R. Naik, M. O. Stone, P. C. Trulove, H. C. De Long and R. A. Mantz, Dissolution and regeneration of bombyx mori silk fibroin using ionic liquids, J. Am. Chem. Soc., 2004, 126(44), 14350–14351.
18. K. P. Ghanta, T. Pal, S. Mondal and S. Bandyopadhyay, Microscopic understanding of the effect of ionic liquid on protein from molecular simulation studies, J. Phys. Chem. B, 2020, 124(19), 3909–3921.
19. A. Sanchez-Fernandez, M. Basic, J. Xiang, S. Prevost, A. J. Jackson and C. Dicko, Hydration in deep eutectic solvents induces non-monotonic changes in the conformation and stability of proteins, J. Am. Chem. Soc., 2022, 144(51), 23657–23667.
20. H. Cui, L. Zhang, C. B. Yildiz, L. Eltoukhy, L. Cheng, K.-E. Jaeger, U. Schwaneberg and M. D. Davari, Enzyme hydration: How to retain resistance in ionic liquids, ACS Sustainable Chem. Eng., 2022, 10(46), 15104–15114.
21. S. Sahoo, T. Pal, S. Mondal, K. P. Ghanta and S. Bandyopadhyay, Conformational properties of Aβ peptide oligomers in aqueous ionic liquid solution: Insights from molecular simulation studies, J. Phys. Chem. B, 2023, 127(51), 10960–10973.
22. J. Liu, Y. Wang, B. Gao, K. Zhang, H. Li, J. Ren, F. Huo, B. Zhao, L. Zhang and S. Zhang, et al., Ionic liquid gating induces anomalous permeation through membrane channel proteins, J. Am. Chem. Soc., 2024, 146(19), 13588–13597.
23. V. Piccoli and L. Martínez, Competitive effects of anions on protein solvation by aqueous ionic liquids, J. Phys. Chem. B, 2024, 128(32), 7792–7802.
24. Y. Wang, J. Liu, X. Yuan, M. Wang, Y. Nie, S. Zhang and H. He, Ionic liquid-based aggregation towards the mesoscale mechanism and their applications, Chem. Eng. J., 2023, 455, 140370.
25. Z. Q. Zhang, X. M. Lu, X. X. Wang and C. H. Lu, IPE ionic liquid database, Institute of process engineering, Chinese academy sciences, Beijing, 100190. https://www.ildatabase.com/.
26. Y. Li, X. Cui, Z. Xiong, B. Liu, B.-Y. Wang, R. Shu, N. Qiao and M.-H. Yung, Quantum molecular docking with a quantum-inspired algorithm, J. Chem. Theory Comput., 2024, 20(15), 6687–6694.
27. B. Ni, H. Wang, H. K. S. Khalaf, V. Blay and D. R. Houston, Autodock-ss: Autodock for multiconformational ligand-based virtual screening, J. Chem. Inf. Model., 2024, 64(9), 3779–3789.
28. A. Frick, U. K. Eriksson, F. de Mattia, F. Oberg, K. Hedfalk, R. Neutze, W. J. de Grip, P. M. Deen and S. Tornroth-Horsefield, X-ray structure of human aquaporin 2 and its implications for nephrogenic diabetes insipidus and trafficking, Proc. Natl. Acad. Sci. U. S. A., 2014, 111(17), 6305–6310.
29. K. S. Egorova, E. G. Gordeev and V. P. Ananikov, Biological activity of ionic liquids and their application in pharmaceutics and medicine, Chem. Rev., 2017, 117(10), 7132–7189.
30. S. J. Marrink, V. Corradi, P. C. T. Souza, H. I. Ingolfsson, D. P. Tieleman and M. S. P. Sansom, Computational modeling of realistic cell membranes, Chem. Rev., 2019, 119(9), 6184–6226.
31. N. A. Erkamp, M. Oeller, T. Sneideris, H. Ausserwoger, A. Levin, T. J. Welsh, R. Qi, D. Qian, N. Lorenzen and H. Zhu, et al., Multidimensional protein solubility optimization with an ultrahigh-throughput microfluidic platform, Anal. Chem., 2023, 95(12), 5362–5368.
32. S. Foser, A. Schacher, K. A. Weyer, D. Brugger, E. Dietel, S. Marti and T. Schreitmüller, Isolation, structural characterization, and antiviral activity of positional isomers of monopegylated interferon α-2a (PEGASYS), Protein Expr. Purif., 2003, 30(1), 78–87.
33. P. Ertl and A. Schuffenhauer, Estimation of synthetic accessibility score of drug-like molecules based on molecular complexity and fragment contributions, J. Cheminform., 2009, 1(1), 8.
34. L. Bui-Le, C. J. Clarke, A. Bröhl, A. P. S. Brogan, J. A. J. Arpino, K. M. Polizzi and J. P. Hallett, Revealing the complexity of ionic liquid–protein interactions through a multi-technique investigation, Comm. Chem., 2020, 3(1), 55.
35. S. K. Thayallath, S. M. Shet, M. Bisht, P. Bharadwaj, M. M. Pereira, G. Franklin, S. K. Nataraj and D. Mondal, Designing protein nano-construct in ionic liquid: A boost in efficacy of cytochrome C under stresses, Chem. Commun., 2023, 59(39), 5894–5897.
36. L. Bui-Le, C. J. Clarke, A. Bröhl, A. P. S. Brogan, J. A. J. Arpino, K. M. Polizzi and J. P. Hallett, Revealing the complexity of ionic liquid–protein interactions through a multi-technique investigation, Comm. Chem., 2020, 3, 1.
37. O. Trott and A. J. Olson, Autodock vina: Improving the speed and accuracy of docking with a new scoring function, efficient optimization, and multithreading, J. Comput. Chem., 2010, 31(2), 455–461.
38. S. Salentin, S. Schreiber, V. J. Haupt, M. F. Adasme and M. Schroeder, Plip: Fully automated protein-ligand interaction profiler, Nucleic Acids Res., 2015, 43(1), 443–447.
39. P. Virtanen, R. Gommers, T. E. Oliphant, M. Haberland, T. Reddy, D. Cournapeau, E. Burovski, P. Peterson, W. Weckesser and J. Bright, et al., Scipy 1.0: Fundamental algorithms for scientific computing in python, Nat. Methods, 2020, 17(3), 261–272.
40. F. Murtagh and P. Legendre, Ward's hierarchical agglomerative clustering method: Which algorithms implement ward's criterion?, J. Classif., 2014, 31(3), 274–295.
41. S. Lee, A. Mao, S. Bhattacharya, N. Robertson, R. Grisshammer, C. G. Tate and N. Vaidehi, How do short chain nonionic detergents destabilize G-protein-coupled receptors?, J. Am. Chem. Soc., 2016, 138(47), 15425–15433.
42. L. Martinez, R. Andrade, E. G. Birgin and J. M. Martinez, Packmol: A package for building initial configurations for molecular dynamics simulations, J. Comput. Chem., 2009, 30(13), 2157–2164.
43. D. A. Pearlman, D. A. Case, J. W. Caldwell, W. S. Ross, T. E. Cheatham, S. DeBolt, D. Ferguson, G. Seibel and P. Kollman, Amber, a package of computer programs for applying molecular mechanics, normal mode analysis, molecular dynamics and free energy calculations to simulate the structural and energetic properties of molecules, Comput. Phys. Commun., 1995, 91(1), 1–41.
44. J. Wang, R. M. Wolf, J. W. Caldwell, P. A. Kollman and D. A. Case, Development and testing of a general amber force field, J. Comput. Chem., 2004, 25(9), 1157–1174.
45. D. Vassetti, M. Pagliai and P. Procacci, Assessment of GAFF2 and OPLS-AA general force fields in combination with the water models TIP3P, SPCE, and OPC3 for the solvation free energy of druglike organic molecules, J. Chem. Theory Comput., 2019, 15(3), 1983–1995.
46. W. L. Jorgensen, J. Chandrasekhar, J. D. Madura, R. W. Impey and M. L. Klein, Comparison of simple potential functions for simulating liquid water, J. Chem. Phys., 1983, 79(2), 926–935.
47. T. Darden, D. York and L. Pedersen, Particle mesh ewald: An N.log(N) method for Ewald sums in large systems, J. Chem. Phys., 1993, 98(12), 10089–10092.
48. R. Salomon-Ferrer, A. W. Götz, D. Poole, S. Le Grand and R. C. Walker, Routine microsecond molecular dynamics simulations with AMBER on GPUs. 2. explicit solvent particle mesh Ewald, J. Chem. Theory Comput., 2013, 9(9), 3878–3888.
49. H. C. Andersen, Rattle: A “velocity” version of the shake algorithm for molecular dynamics calculations, J. Comput. Phys., 1983, 52(1), 24–34.
50. H. J. C. Berendsen, J. P. M. Postma, W. F. V. Gunsteren, A. DiNola and J. R. Haak, Molecular dynamics with coupling to an external bath, J. Chem. Phys., 1984, 81(8), 3684–3690.
51. Y. Zhang, S. E. Feller, B. R. Brooks and R. W. Pastor, Computer simulation of liquid/liquid interfaces. I. Theory and application to octane/water, J. Chem. Phys., 1995, 103(23), 10252–10266.
52. IBM SPSS statistics for windows, version 25.0*, IBM Corp., Armonk, NY, 2017.
53. F. Pedregosa, G. Varoquaux, A. Gramfort, V. Michel, B. Thirion, O. Grisel, M. Blondel, P. Prettenhofer, R. Weiss and V. Dubourg, et al., Scikit-learn: Machine learning in python, J. Mach. Learn. Res., 2011, 12, 2825–2830.

---