id
stringlengths 24
24
| idx
int64 0
402
| paragraph
stringlengths 106
17.2k
|
|---|---|---|
668937bdc9c6a5c07a733145 | 33 | Although our GA-predicted 6×6×6 supercell suggests that TiOF 2 does not exhibit NbO 2 F-type exclusive [O-F-F-O] ordering along ⟨100⟩ columns, this result might be a consequence of our choice of fitting procedure for the DFT-derived cluster expansion model used in the GA structure-prediction scheme. Because our DFT training set includes only 2 × 2 × 2 supercells, only anion-anion interactions that fit within this supercell size are included in the resulting cluster-expansion model. |
668937bdc9c6a5c07a733145 | 34 | To validate the predictions from our 6 × 6 × 6 GA structure prediction, we performed additional DFT calculations on a set of [O-F-F-O] ordered 3 × 3 × 3 TiOF 2 supercells and compared the resulting energies (per formula unit) to the energies of those of the 2 × 2 × 2 supercells in our CE training set with exclusive cis-TiO 2 F 4 cation coordination and to the DFT-optimised energies of our four 4 × 4 × 4 GA-predicted supercells. All three sets of structures have no collinear O-Ti-O units and, there- |
668937bdc9c6a5c07a733145 | 35 | To validate against the experimental PDF data, each GA-predicted structure was used as an initial structural model that was fitted to the experimental PDF data, with the atomic positions left unrefined to limit the number of fitting parameters. All four GA-predicted structures give a better fit for the experimental data than the average cubic P m 3m model (R w = 31.2 %), with the best fit obtained for GA structure 4 (R w = 16.4 %) (Fig. ). The improved fit to the experimental PDF data is particularly evident in the split peak at 1.82 Å and 1.99 Å, which we previously assigned to nearest-neighbour Ti-O and Ti-F distances, respectively, and in the region between 3.5 Å and 4.0 Å, which we assigned to Ti-X-Ti pairs, and which is consistent with the observation from our 2 × 2 × 2 supercell DFT dataset that ReO 3 -type TiOF 2 preferentially adopts anion configurations that allow the Ti-O and Ti-F bond-lengths to be shorter and longer, respectively, than the average Ti-X bond-length of 1.90 Å. Our assignment of these peaks in the experimental PDF spectrum is also validated by direct analysis of the GA-predicted structures (Supporting Information), which gives mean Ti-O and Ti-F distances of 1.81 Å and 1.99 Å, respectively, and a clear splitting in Ti-X-Ti distances. |
668937bdc9c6a5c07a733145 | 36 | To validate our GA-predicted structures against our experimental 19 F NMR MAS data, we performed DFT calculations on each of the four GA-predicted structures to calculate magnetic shielding tensors for each fluo- ride anion. Obtaining a simulated 19 F NMR spectrum from DFT calculations requires converting from calculated isotropic and anisotropic shielding values, σ iso and σ csa , to predicted isotropic and anisotropic chemical shift values, δ iso and δ csa , respectively. |
668937bdc9c6a5c07a733145 | 37 | For the isotropic values, the conventional approach to convert from σ iso to δ iso is to use a linear transformation function δ iso = aσ iso +b, where the parameters a and b are obtained by linear regression between computed σ iso and experimental δ iso data. Several different linear transformation functions for 19 F have been published for a wide range of cations bonded to fluorine (see Table ). One approach for the quantitative prediction of 19 F δ iso values for a disordered system containing one metal cation is to derive an appropriate linear transformation function from data for a reference ordered (oxy)fluoride, containing the same cation as the system of interest and with fluoride ions occupying multiple crystallographic sites. This approach has previously been used to simulate the 19 F NMR spectra of the disordered oxyfluorides MO 2 F and MOF 3 (M = Nb, Ta), with transformation functions derived by fitting to data for the corresponding MF 5 fluorides . |
668937bdc9c6a5c07a733145 | 38 | For obtaining a transformation function σ iso → δ iso for titanium (oxy)fluorides, a reasonable reference system is TiF 4 . TiF 4 is formed from corner-sharing TiF 6 octahedra (Fig. ) and contains twelve inequivalent fluorine sites that can be classified into two groups: bridging fluorine atoms, that are bonded to two titanium centres, and terminal fluorine atoms, that are bonded to only one titanium centre . High-resolution 19 F NMR data for TiF 4 have previously been reported by Murakami et al. , who derived a linear σ iso → δ iso transformation function by fitting this model function to DFT-calculated σ iso and experimental δ iso data [99]. While the DFT and FIG. . 19 F σiso and σcsa values for TiF4 from DFT calculations (VASP, PBE + DFT-D3), plotted against the corresponding experimental δiso and δcsa values, respectively . Panels (b) and (e) show all data, and corresponding linear-least-squares fits (dashed lines). Panels (a) and (c), and (d) and (f), show the same data, selecting values for terminal F and bridging F only, respectively; each panel shows the original linear model obtained from fitting to the full dataset (dashed lines) and a revised linear model obtained by fitting to the corresponding data subset only (solid lines). The resulting best-fit linear models are δiso = -0.830σiso + 44.1 and δcsa = -0.671σcsa for bridging F, and δiso = -1.116σiso + 67.2 and δcsa = -0.918σcsa for terminal F. |
668937bdc9c6a5c07a733145 | 39 | experimental σ iso and δ iso data of Murakami et al. follow an approximate linear relationship, close inspection of these data (reproduced in the Supporting Information (Fig. ) shows that a single linear relationship is not able to accurately describe the correlations between σ iso and δ iso simultaneously for both bridging and terminal fluorine atoms, with each subset of atoms showing systematic deviations from the linear model obtained from fitting to all the fluorine atoms. Murakami et al. were also unable to obtain a satisfactory quantitative relationship between their calculated and experimental for the magnetic which prevents the quantitative prediction of spinning sidebands. |
668937bdc9c6a5c07a733145 | 40 | To address limitations, and to derive mation functions for the simulation of 19 F NMR spectra of titanium (oxy)fluorides, we have revisited the analysis of Murakami et al. . Fig. shows calculated σ iso data for TiF 4 plotted against the corresponding experimental δ iso values reported by Murakami et al. . The data form two distinct clusters, corresponding to terminal and bridging F, at low and high σ iso values respectively. As reported by Murakami et al., the full dataset does show an approximately linear relation-ship between σ iso and δ iso values. Zooming in on the data for terminal and bridging fluorine atoms (Figs. and, respectively), however, highlights the deficiencies of fitting a single linear model to both groups of data. The observation that these 19 F NMR data are not well described by a single linear relationship is, perhaps, unsurprising, given the significant difference in local chemical environment and bonding for fluorine atoms directly bonded to two versus one Ti centres. To account for the categorical difference between bridging and non-bridging fluorine atoms, we fit separate linear models to the two clusters of data. This approach is much better able to quantitatively describe the correlation between σ iso and δ iso within each category of fluorine atoms (bridging versus non-bridging), and we obtain best-fit linear relationships of δ iso = -0.830σ iso + 44.1 for bridging F, and δ iso = -1.116σ iso + 67.2 for terminal F. Figs. ) show an equivalent treatment of the DFT-calculated σ csa and experimentally derived δ csa data. Fitting the linear relationship δ csa = aσ csa to the full σ csa versus δ csa dataset (Fig. ) gives a poor fit, with the data for terminal and bridging fluorine atoms showing systematic deviations from the average best-fit model. By fitting separate functions to the data for the terminal (Fig. ) fluorine atoms and for the bridging (Fig. ) fluorine atoms, respectively, however, we obtain two transformation functions that much more accurately describe the quantitative relationship between σ csa and δ csa in TiF 4 . |
668937bdc9c6a5c07a733145 | 41 | In our GA-predicted models, all F bridge between two Ti. To generate simulated 19 F NMR spectra for these GA-predicted TiOF 2 structural models, we use the σ iso → δ iso and σ csa → δ csa relationships derived for bridging F in TiF 4 . We note that the σ iso → δ iso and σ csa → δ csa relationships derived here for terminal F species in TiF 4 provide suitable transformation functions for non-bridging Ti-F-□ fluorines that may be present in other titanium and oxyfluorides. |
668937bdc9c6a5c07a733145 | 42 | For all four GA-predicted structures, we obtain simulated 19 F spectra that are in close agreement with the experimental TiOF 2 spectrum (see Fig. for structure 4 and Fig. for all four GA-predicted structures), with average chemical shift values ranging from 22.6 ppm to 25.4 ppm (Table ). Moreover, the chemical shift anisotropies are well-modelled, as evidenced by the correct reproduction of the spinning sidebands. The simulated and experimental spectra are not perfectly superimposed due to slightly different averages and larger spreads in the calculated chemical shift values compared to the experimental data. This discrepancy is not altogether unexpected given the large chemical shift range of 19 F (over 1000 ppm ), the inherent uncertainty in our σ iso → δ iso relationship, and the use of finite-size 4×4×4 structural models as approximations to the experimental structure. |
668937bdc9c6a5c07a733145 | 43 | Both the simulated PDF and 19 F NMR spectra show good agreement with the corresponding experimental data, with particularly good agreement in the case of the PDF data for GA-predicted structure 4. We therefore conclude that our DFT-derived cluster expansion model correctly describes the form of anion short-range order in ReO 3 -type TiOF 2 , and consider structure 4, obtained from our genetic algorithm structure prediction, as a representative structural model for the experimental samples considered in this work. |
668937bdc9c6a5c07a733145 | 44 | ReO 3 -type TiOF has previously been as a potential lithium-ion electrode material Our results indicate that ReO 3 -type TiOF 2 exhibits a specific short-range anion ordering consisting of preferential cis-TiO 2 F 4 titanium coordination, which gives rise to correlated disorder at longer length-scales. Our analysis of the relative energies of different anion configurations shows that there are a large number of low-energy anion configurations that are expected to be competitive under synthesis conditions (Fig. ), indicating that the precise anion configuration in ReO 3 -type TiOF 2 might depend experiment simulated FIG. . Experimental (solid line) and simulated (dashed line) 19 F MAS (64 kHz) NMR spectra of ReO3-type TiOF2 (GA-predicted structure 4). |
668937bdc9c6a5c07a733145 | 45 | Having predicted and validated structural models for "as synthesised" ReO 3 -type TiOF 2 , we now consider the effect of variation in local anion structure on lithium intercalation properties. To this end, we computed the dilute limit intercalation voltage for all possible interstitial sites in three exemplar structures with varying degrees of anion ordering: these structures comprise a 4 × 4 × 4 supercell of the lowest energy 2×2×2 structure, which is fully-ordered with all-cis-Ti[O 2 F 4 ] coordination, the partially disordered GA-predicted structure 4, and a maximally disordered 4 × 4 × 4 special quasirandom structure (SQS), which approximates the O/F correlations for an infinite lattice with a fully random (maximum-entropy) distribution of anions . |
668937bdc9c6a5c07a733145 | 46 | Fig. presents calculated lithium-intercalation energies and estimated mean values for each exemplar structure. The fully-ordered all-cis-Ti[O 2 F 4 ] structure has only three non-equivalent interstitial sites, and therefore has a relatively narrow distribution of lithiumintercalation energies, with a mean of -1.53 eV. GApredicted structure 4 is partially disordered, and all 64 cubic interstitial sites in the 4 × 4 × 4 supercell are therefore inequivalent by symmetry. This gives a broader spread in lithium intercalation energies of more than ∼ 1 eV, with an estimated mean of (-2.44 ± 0.05) eV. The SQS structure is, again, more disordered than the GA-predicted structure, and has an even broader spread in lithium intercalation energies (∼ 2 eV) and an estimated mean of (-3.06 ± 0.09) eV. |
668937bdc9c6a5c07a733145 | 47 | These results show that the lithium intercalation energy for ReO 3 -type TiOF 2 is sensitive to the precise arrangement of oxygen and fluorine atoms within the host structure, with the mean intercalation energy shifting by > 1.5 eV between the fully ordered all-cisTi[O 2 F 4 ] structure and the 4 × 4 × 4 special quasirandom structure considered here. In general, as the anion substructure becomes more configurationally disordered, the in-intercalation energy / eV FIG. 12. Effect of changing O/F substructure on lithium intercalation energies into cubic TiOF2. Data are shown as raincloud plots for three exemplar structures: (top) the fully-ordered all-cisTi[O2F4] structure; (middle) a 4×4×4 supercell genetic-algorithm-predicted structure (GA structure 4); (bottom) the 4×4×4-supercell special quasi-random structure (bottom). For each dataset, the points show individual calculated intercalation and the solid distribution shows a kernel density estimate of the distribution of intercalation energies. Error bars show the 95 % compatibility interval for the estimated mean of each dataset, obtained by bootstrap resampling of the original data . |
668937bdc9c6a5c07a733145 | 48 | tercalation energy shifts to more negative values, while the distribution of lithium intercalation energies becomes broader. This result suggests that the electrochemical properties of ReO 3 -type TiOF 2 , and, by analogy, other heteroanionic intercalation electrode materials, may be modulated through directed synthesis protocols that produce samples with different degrees of short-range anion order. |
668937bdc9c6a5c07a733145 | 49 | Heteroanionic materials offer a rich chemical space for developing new materials with targeted properties. To understand and control the properties of heteroanionic materials requires a detailed characterisation of their structures-in particular, the specific arrangement of the component anions. Resolving the anionic substructure of anion-disordered oxyfluorides is particularly challenging, because X-ray and neutron Bragg scattering experiments give only an average structural description. Resolving local structural details in anion-disordered oxyfluorides, therefore, requires using alternative complementary experimental or computational techniques. |
668937bdc9c6a5c07a733145 | 50 | Here, we have presented a study of the anionic substructure in the exemplar transition-metal oxyfluoride ReO 3 -type TiOF 2 , using a combination of X-ray PDF, 19 F NMR, DFT modelling, and genetic-algorithm structure-prediction. We find that ReO 3 -type TiOF 2 exhibits strong short-range ordering characterised by preferential cis-O 2 F 4 coordination around titanium. This cis-coordination of titanium allows titanium cations to move away from the centre of their coordination octahe-dra to give shorter Ti-O bonds (and longer Ti-F bonds), giving a net increase in total Ti-X bond strength, relative to more symmetric trans-O 2 F 4 titanium coordination. This preferential cis-TiO 2 F 4 coordination also gives rise to correlated anion disorder , where the configuration of oxygen and fluorine ions decorrelates with separation, resulting in long-range anion disorder that is consistent with the average P m 3m structure model previously proposed from X-ray powder diffraction data . |
668937bdc9c6a5c07a733145 | 51 | To obtain structural models that incorporate this correlated disorder, we have used genetic-algorithm structure prediction to generate partially disordered supercells. We then validated these structural models by generating simulated X-ray PDF and 19 F NMR data, which we compared to equivalent experimental data for our synthesised TiOF 2 sample. For the simulation of the 19 F NMR spectrum, we used new empirical linear transformation functions to convert from calculated shielding values, σ iso and σ csa , to predicted chemical shift values, δ iso and δ csa , which we derived by fitting calculated σ iso and σ csa values for bridging fluoride ions in TiF 4 to previously published experimental data . We expect the resulting linear transformation functions to be generally applicable for calculating 19 F NMR spectra of other titanium oxyfluorides. For both the X-ray PDF and 19 F NMR data, our simulated data agree well with the corresponding experimental data, indicating that our geneticalgorithm-predicted structures reproduce well the shortrange structure of our sample. |
668937bdc9c6a5c07a733145 | 52 | We then consider the effect of variations in anionic short-range order on the lithium intercalation properties of ReO 3 -type TiOF 2 . By performing additional DFT calculations, we show that the local anion substructure can have a significant effect on lithium intercalation voltages, with an example fully-ordered low-energy structure and the maximally disordered special-quasi-random 4 × 4 × 4 supercell structure showing a difference of mean lithium intercalation voltage of > 1.5 V as well as a large increase in the spread of intercalation voltage values. Because the precise short-range structure of ReO 3 -type TiOF 2 may be affected by different synthesis protocols, this result indicates that it may be possible to tune the electrochemical intercalation behaviour of TiOF 2 -and, by analogy, of other transition-metal heteroanionic materials-through careful design of synthesis routes. |
668937bdc9c6a5c07a733145 | 53 | The work presented here demonstrates how the detailed local structure of heteroanionic oxyfluorides can be resolved using a combination of experimental and computational methods. By combining X-ray PDF analysis and 19 F NMR spectroscopy with DFT modelling and genetic algorithm structure prediction we have identified a revised structural model for ReO 3 -type TiOF 2 that is consistent between our experimental and computational analyses. We have also identified how the details of local coordination geometry and bonding direct short-range order in this material, through anions adopting local configurations that maximise Ti-(O/F) bond strength. The general strategy presented here is expected to be generally applicable to other anion-disordered oxyfluorides, where similar short-range deviations from the average crystallographic structure obtained from conventional diffraction methods are also likely. |
668937bdc9c6a5c07a733145 | 54 | X-ray diffraction analysis for ReO 3 -type TiOF 2 ; Pair Distribution Function data for TiOF 4 between 8 Å to 40 Å; Additional details of the cluster expansion model fitting and model ECIs; Details of the Genetic Algorithm structure prediction scheme; Structural analysis of the GA-predicted structures; Correlation between calculated σ iso and experimental δ iso values for 19 F in titanium (oxy)-fluorides; Haeberlen convention used to define the shielding and chemical shift NMR parameters; Details about calculations using the NMR-CASTEP code; Previously reported relationships between calculated σ iso and experimental δ iso values for 19 F in inorganic fluorides; Derivation of an empirical linear relation between calculated σ iso and experimental δ iso values for 19 F in titanium (oxy)-fluorides; Effect of DFT calculation method on atomic positions for TiF 4 ; Simulated 19 F MAS NMR spectra for all four 4 × 4 × 4 GA-predicted structural models. |
624c577f804882e535045c53 | 0 | All MEPA compounds were synthesized by Knoevenagel condensation of appropriate benzaldehydes with 2-methoxyethyl cyanoacetate, catalyzed by base, piperidine (Scheme 1). The preparation procedure was essentially the same for all the MEPA compounds. In a typical synthesis, equimolar amounts of 2-methoxyethyl cyanoacetate and an appropriate benzaldehyde were mixed in equimolar ratio in a 20 mL vial. A few drops of piperidine were added with stirring. The product of the reaction was isolated by filtration and purified by crystallization from 2-propanol. The condensation reaction proceeded smoothly, yielding products, which were purified by conventional techniques. The compounds were characterized by IR, 1 H and 13 C NMR spectroscopies. No stereochemical analysis of the novel ring-substituted MEPA was performed since no stereoisomers (E or/and Z) of known configuration were available. |
624c577f804882e535045c53 | 1 | Copolymers of the ST and the MEPA compounds, P(ST-co-MEPA) were prepared in 25-mL glass screw cap vials at ST/MEPA = 3 (mol) the monomer feed using 0.12 mol/L of ABCN at an overall monomer concentration 2.44 mol/L in 10 mL of toluene. The copolymerization was conducted at 70ºC. After a predetermined time, the mixture was cooled to room temperature, and precipitated dropwise in methanol. The composition of the copolymers was determined based on the nitrogen content (cyano group in MEPA monomers). The novel synthesized MEPA compounds copolymerized readily with ST under free-radical conditions (Scheme 2) forming white flaky precipitates when their solutions were poured into methanol. The conversion of the copolymers was kept between 10 and 20% to minimize compositional drift (Table ). Nitrogen elemental analysis showed that between 6.9 and 31.7 mol% of MEPA is present in the copolymers prepared at ST/MEPA = 3 (mol), which is indicative of relatively high reactivity of the MEPA monomers towards ST radical which is typical of phenoxy ringsubstituted phenylcyanoacrylates. Since MEPA monomers do not homopolymerize, the most likely structure of the copolymers would be isolated MEPA monomer units alternating with short ST sequences (Scheme 2). |
676eb5befa469535b9fb001a | 0 | LLMs are opening new possibilities for leveraging natural language processing in chemistry and other scientific fields. These models can access and generate chemical information, potentially assisting researchers with tasks such as predicting molecular properties, extracting structured data from text, and even designing new molecules. However, using LLMs in chemical research comes with unique challenges. One prominent issue is "hallucination," where the model produces outputs that are confidently incorrect, often due to gaps or inconsistencies in its training data . Hallucinations present a substantial obstacle in chemistry, where even minor inaccuracies can lead to significant misinterpretations in predicting molecular properties or reactions . To fully integrate LLMs into chemical research workflows, these hallucinations must be addressed and it is critical to improve the models' ability to better handle chemical data. |
676eb5befa469535b9fb001a | 1 | Existing research efforts are exploring various ways to improve LLM performance on chemistryspecific tasks. Some groups have developed specialized models, like ChemLLM, which is trained on extensive chemical datasets to ensure it is proficient in a wide array of chemical tasks . This specialization helps ChemLLM perform well in chemical applications. Instruction tuning is another promising approach; models such as MolecularGPT pre-train models with Simplified Molecular Input Line Entry System (SMILES) strings connected to molecular properties to enhance few-shot learning on chemical properties, outperforming traditional models on certain tasks . Additionally, fine-tuned models have demonstrated success in converting unstructured chemical text into structured data for reaction databases, highlighting LLMs' potential to build organized and accessible chemical knowledge bases . Some studies have also assessed the performance of generalpurpose LLMs in chemistry-related programming tasks, such as generating code for chemical data analysis . Alternatively, custom models can be created from the same transformer architecture that powers LLMs but using molecular properties as the training data. |
676eb5befa469535b9fb001a | 2 | For example, Prompt-MolOpt uses prompt engineering to improve multi-property optimization and address data scarcity issues common to this field . This method excels in fewand zero-shot learning scenarios due to its ability to leverage single-property datasets to learn generalized causal relationships. Another area where LLMs are being used to automatically design more effective and efficient agentic systems is a novel research field called Automated Design of Agentic Systems (ADAS) ]. |
676eb5befa469535b9fb001a | 3 | These efforts underscore the progress being made with specialized chemical LLMs and instruction-tuned models, but they come with limitations. Developing or fine-tuning models on dedicated chemical datasets requires substantial computational and energy resources ] and domain-specific expertise . Furthermore, once models are fine-tuned for a specific chemical application, their generalizability may suffer, and their adaptability to other domains or newly emerging chemical knowledge can become constrained . Therefore, there is a need for time-of-prompt solutions that can enhance the accuracy of LLM predictions at inference time-without requiring extensive retraining or fine-tuning . Such techniques would allow LLMs to be applied to a wider range of chemical tasks, even in cases where the model's pre-existing knowledge may be incomplete or out-of-date. |
676eb5befa469535b9fb001a | 4 | Two emerging approaches that could address these limitations are Retrieval-Augmented Generation (RAG) and the Multiprompt Instruction PRoposal Optimizer (MIPRO). RAG combines a retrieval system with a generative model, enabling LLMs to dynamically fetch or calculate relevant, up-to-date information from external databases or knowledge sources . In the context of chemistry, RAG could draw on calculations or curated databases to supply the LLM with accurate molecular data or specific molecular properties in real time . This external grounding could significantly reduce the likelihood of hallucinations by ensuring that the LLM has access to precise chemical data instead of relying solely on its potentially limited training set. RAG is potentially valuable for tasks like predicting properties using group contribution methods, where relationships between molecular structure and molecular properties are complex and require detailed, accurate data that an LLM may not robustly encode . |
676eb5befa469535b9fb001a | 5 | MIPRO is a prompt optimization framework that creates and refines the LLM prompts for improved accuracy and consistency . MIPRO uses an LLM to generate additional instructions to add to the prompt and then selects few-shot examples that illustrate successful executions of the given task, optimizing the selection of both using a PyTorch powered ML framework [Opsahl-Ong, 2024]. MIPRO can bootstrap examples from training data and dynamically generate instruction candidates to provide structured, task-specific guidance . Through Bayesian optimization, MIPRO iteratively identifies the optimal combination of examples and instructions, evaluated against a user-generated quantitative metric. This prompt refinement reduces hallucinations by ensuring that the LLM has a clear and relevant framework for understanding underlying data, without the need for creating or fine-tuning a model for a specialized application . |
676eb5befa469535b9fb001a | 6 | TPSA is used as a molecular descriptor in drug research because it can efficiently predict a drug's ability to passively cross biological membranes, such as the intestinal lining or the blood-brain barrier . This efficiency is crucial in early drug discovery stages, where researchers need to evaluate a large number of potential drug candidates. Several studies have shown that TPSA correlates well with drug permeability . For instance, drugs that are readily absorbed from the gut or those that can penetrate the central nervous system typically have lower TPSA values . TPSA has also been used in a model that predicts drug exposure in pregnant women and their fetuses. This model relies on a "permeability-limited placenta model" that simulates drug transfer between the mother and fetus . |
676eb5befa469535b9fb001a | 7 | Together, RAG and MIPRO present a powerful solution for improving LLM performance. RAG addresses the issue of outdated or incomplete information by grounding the LLM's responses in current, high-quality data sources, ensuring that predictions are accurate and contextually relevant. MIPRO complements this by optimizing the prompt structure, allowing the LLM to interpret and utilize retrieved data more effectively through well-designed instructions and examples. Here, as an example of this approach, I describe a method for predicting TPSA that combines RAG and MIPRO using a commercially available LLM, ChatGPT-4o-mini. In tandem, these approaches enabled the LLM to make accurate, data-driven predictions at inference time, enhancing its reliability without fine tuning the weights of the base model. This approach reduced the root mean squared error (RMSE) from 59.41 for prediction using the GPT-4o-mini LLM directly to 7.44 RMSE when MIPRO and RAG were employed for predictions on a set of random molecules. The individual contribution of the various elements of this approach is described below. |
676eb5befa469535b9fb001a | 8 | Molecular data were acquired from PubChem by querying random compound identifiers and fetching properties through the PubChem PUG-REST API. A dataset was constructed by binning TPSA values in intervals of 5 units, with 20 molecules per bin to ensure even distribution across TPSA values and provide a robust dataset for prompt tuning. Since PubChem defines TPSA as "a simple method -only N and O are considered," [ accessed on Nov 12, 2024.] only molecules with C, N, O, and H were included and if the N and O functional groups could not be mapped to one of the specified functional groups ], they were excluded. Bins were populated by randomly sampling molecules from PubChem until each TPSA interval had 20 molecules. RDKit was used to parse SMARTS patterns, generating a list of functional groups. SMARTS patterns were loaded and iteratively applied to each SMILES string, with RDKit identifying the presence of targeted functional groups in each molecule. These functional group assignments were then linked to TPSA contributions using lookup data containing TPSA values associated with each group. |
676eb5befa469535b9fb001a | 9 | To focus on drug-like molecules, SMILES codes with more than 10 hydrogen bond acceptors or more than 5 hydrogen bond donors were removed. Additionally, molecules with a mass greater than 500 were filtered out, further aligning the dataset with criteria typically used for drug-like compound properties. Finally, molecules with a non-zero charge were excluded to maintain focus on neutral compounds. The training set contained 30 structured examples from this list for selecting bootstrap examples from, while the validation set contained a second set of 30 that were used to validate prompt performance. |
676eb5befa469535b9fb001a | 10 | DSPy examples serve as modular, query-answer pairs that allowed standardization of data inputs and generated a comprehensive dataset spanning a wide range of TPSA values. This dataset was balanced across TPSA intervals to prevent biases toward certain values and ensure that the LLM was exposed to a representative set of molecular features. A scaffold split was performed to ensure that the train or test sets would contain examples of any scaffolds that repeated across the data. The examples were then loaded into the LLM program as a structured dataset, where each Example provided the model with a consistent input-output relationship. |
676eb5befa469535b9fb001a | 11 | GPT-4-o-mini was used as the model for generating and testing prompts, ensuring that both prompt generation and task completion maintained consistent model behavior. GPT-4o-mini which has a reduced model size compared to GPT-4o was used here in part to minimize the risk that prior training data would contain direct answers to the questions being asked. The reduced parameter size means these direct connections are less likely. The most recent version of MIPRO, MIPROv2 from the DSPy package was used. 10 few-shot example sets were proposed during the optimization. By generating 10 sets, MIPRO can experiment with a range of examples, allowing it to assess which examples best aid the model in reducing TPSA prediction errors. An initial temperature of 1.2 was used. This increases prompt diversity at the start. This helps MIPRO to explore various prompt combinations early on, with a controlled decrease in diversity over time for convergence. 25 trials were run, allowing MIPRO to iteratively refine prompts based on validation performance. Each trial generates a new prompt set, and Bayesian Optimization identifies which sets perform best. Minibatch evaluation was performed in batches of 5 examples, enabling efficient prompt evaluation in each trial. This approach allows for broader prompt testing within the given trial limit of 30. The number of few-shot and labeled examples in each prompt was set to a maximum of 8, ensuring manageable prompt length and optimizing example diversity without overwhelming the model with too many examples at once. After every 5 minibatches, a full evaluation on the validation set was performed. This periodic full evaluation providef a more stable performance benchmark, allowing the Bayesian optimizer to adjust prompt selection based on more reliable performance data. |
676eb5befa469535b9fb001a | 12 | A custom metric was used to calculate the absolute error between the LLM predicted TPSA value and the true TPSA value, using this difference to guide prompt and example selection across bootstrap example selection and prompt optimization. During bootstrap example selection, the metric assesses the accuracy of candidate few-shot examples generated from the training dataset. A threshold-based approach was used, retaining only examples where the absolute error was below 20. This threshold ensures that the examples selected for bootstrapping are reliable representations of a good TPSA prediction, forming a solid foundation for the fewshot examples used in prompt optimization. In prompt optimization, the metric guides Bayesian Optimization by continuously measuring the accuracy of different prompt configurations. At each trial, the effectiveness of a prompt is evaluated by calculating the negative absolute error across a batch of examples. Additionally, every few minibatches the entire validation set was evaluated to confirm that the current prompt configuration performs well on a broader set of examples, enhancing stability and reducing noise in prompt selection. By calculating negative absolute error between predicted and actual TPSA values, this metric guides the optimizer towards more accurate prompt selections. |
676eb5befa469535b9fb001a | 13 | The TPSA predictor is derived from DSPy Module object and utilizes the TypedPredictor program to ensure responses with correct formatting. The predictor encapsulates the logic for preparing, formatting, and training the model on prompt-optimized TPSA prediction tasks. It utilizes a structured prompt that can integrate molecular descriptions, functional group data, and specific atom counts. These modules include 1) Describing Molecular Functional Groups: The method first calls describe_molecule with the SMILES code. This function returns an assignment of functional groups based on predefined SMARTS patterns. |
676eb5befa469535b9fb001a | 14 | The prompts are generated in segments that are removed selectively during the ablation study. The prompt segments include: 1) Functional Group Information: A function was created using rdkit to provide a list of functional groups present in the smiles code as matched to the list of TPSA contributors ], 2) Atom Counts: identifies the number of nitrogen and oxygen atoms in the molecule. 3) The total atom count is used to generate specific instructions on how each atom's presence should impact the TPSA value. 4) Data from the published group contribution table to the TPSA for each functional group present 5) Details that specify the response format, ensuring the LLM outputs a JSON object with a list of TPSA values. The predicted TPSA contributions are summed to avoid math hallucinations ] and to provide a single TPSA value. |
676eb5befa469535b9fb001a | 15 | The effectiveness of structured prompt optimization using RAG and MIPRO (tpsa model, Figure ) was compared to a basic prompt (direct model) for predicting the TPSA of a set of 140 molecules. The direct model, which uses a simple, non-augmented prompt without RAG or MIPRO optimizations, results in a mean RMSE of 59.41, with predictions showing little alignment to the actual TPSA values. The prompt used was "Predict the numerical value of the topological surface area, TPSA, for a molecule described by the SMILES code, {molecule}," where molecule was one SMILES code selected from a list. This basic prompt leads to poor model performance, as the LLM struggles to reliably relate molecular structure to TPSA without the additional context provided in the optimized prompt. SMILES codes are common but if the training data did not include the specific property connected to that specific form of the SMILES code, as appears to be the case, the LLM cannot infer what the values should be. (Figure ). The tpsa model, incorporating RAG and optimized with MIPRO's prompt structuring and fewshot example selection, achieves an RMSE of 7.44, with most predictions closely matching the calculated values obtained from PubChem. This suggests that incorporating functional group details and other contextual information and optimizing prompts through MIPRO significantly enhances prediction accuracy. For the tpsa model a multi-part prompt structure (Figure ) was used that incorporated RAG components as well as text designed to ensure the response of the LLM followed the request for typed format of a list of float values which were then summed to get the predicted TPSA value. This prompt was used in the MIPRO process which produced examples and a data description that were appended to the prompt at inference. The outliers that had a predicted TPSA > 1 different from the calculated TPSA (supporting info, Figure ) tended to have longer lists of functional groups (6.4 vs 4.6 mean) and more nitrogen and oxygen atoms suggesting that the more complicated molecules were harder to predict. and types of functional groups are removed. Without these critical details, predictions become widely dispersed from actual values. This configuration underscores the importance of functional group and functional group information for minimizing hallucinations and achieving reliable TPSA predictions. The outliers (supporting info, Figure ) contain many of the same molecules as the individual RAG removals as well as some new ones. Next, the text added by the MIPRO optimization was iteratively removed from the full model to assess the contribution of MIPRO to the improvement of the overall model. When the description of the dataset (termed signature in DSPy) was removed (Figure ) the mean RMSE value increased only slightly to 7.47. In contrast, when the bootstrapped examples were removed, the RMSE increased to 18.07 (Figure ). While some predictions remained close to the ideal value, many did not, including 15 values that were exactly doubled over the actual values, a hallucination not observed in the direct prediction. |
676eb5befa469535b9fb001a | 16 | This study provided a simple example of a molecular property prediction that allowed a detailed examination of strategies to reduce LLM hallucinations in scientific applications. The results demonstrate the effectiveness of both RAG and MIPRO individually and in combination to improve the accuracy and reliability of LLMs in predicting molecular properties, a critical aspect of drug research. By augmenting LLMs with both external data retrieval and optimized prompt structures, we observed a significant reduction in prediction errors. Specifically, the fully optimized model achieved an RMSE of 7.44, closely aligning with calculated TPSA values and outperforming models that used only a simple prompt or incomplete prompt components. |
676eb5befa469535b9fb001a | 17 | MIPRO iteratively identifies the optimal combination of examples and instructions, creating a prompt that enables the model to consider functional group contributions, functional group details, and additive rules when making predictions. The addition of MIPRO's optimized prompts allowed the model to better interpret molecular structure and contextual details, such as functional group contributions which are essential for accurate predictions. Our ablation studies showed that removing specific prompt components led to increased error rates, confirming the importance of each element in minimizing model hallucinations. For instance, omitting functional group descriptions or atom counts resulted in poorer alignment, with RMSE rising to 28.95 when both elements were removed. These findings underscore the necessity of detailed molecular context in LLM prompts, when property predictions depend on molecular features. |
676eb5befa469535b9fb001a | 18 | By integrating a retrieval step that generates relevant molecular properties and functional group information, this RAG mitigates the risk of hallucinations. This approach addresses the limitations of relying solely on static training data, which may be unavailable for all possible inputs or insufficiently detailed for specialized tasks. By integrating RAG and MIPRO, the LLM's applicability to the chemical task of TPSA prediction was improved, without retraining or fine-tuning the LLM. These results suggest that RAG and MIPRO can significantly improve the utility of general-purpose LLMs in chemical and other scientific research, providing a flexible, scalable solution that enhances prediction accuracy and contextual relevance. This combined approach offers a promising pathway for leveraging LLMs in chemistry and other fields where accurate, context-aware data interpretation is essential. By allowing the model to retrieve relevant information for each query, RAG helps ensure that its predictions are rooted in reliable data. |
676eb5befa469535b9fb001a | 19 | By combining RAG's data-driven retrieval with MIPRO's prompt optimization, LLMs can be transformed into more accurate and versatile tools for chemical research, capable of delivering reliable predictions even in complex or unfamiliar contexts . This approach holds promise not only for chemistry but also for other scientific domains that require precise, contextually informed data interpretation . Together, RAG and MIPRO can enhance the utility of general-purpose LLMs across a wide range of research applications, reducing the need for specialized models and allowing researchers to leverage LLM technology with greater flexibility and accuracy . |
676eb5befa469535b9fb001a | 20 | Training or fine-tuning models with up-to-date information is another powerful approach but comes with drawbacks. Fine-tuning requires significant advance work to prepare a model tailored to a specific need. In contrast, approaches that can be performed at inference time offer the advantage of being applicable to any model without retraining the weights, thereby preserving generalizability. This combination could be especially useful in drug discovery, where accurate molecular property predictions are crucial for assessing drug permeability and potential efficacy early in the development pipeline. |
676eb5befa469535b9fb001a | 21 | As LLMs and their training data grow in size, their capabilities can seem limitless, however, they cannot be trained on data that does not exist yet. The approach described here takes an LLM incapable of a specific molecular task and makes it substantially more capable through augmented generation and prompt optimization. This approach could allow LLMs to be used as research assistants even when handling data outside of their initial training while maintaining the utility of LLMs in handling language. |
60c74162702a9b3e6a18a234 | 0 | Empirical force fields are widely used in molecular simulation studies, mostly when chemical reactivity is not operative. Due to the availability of plentiful experimental and firstprinciples quantum mechanical data, water is a popular testing application for developing new force field models and new approaches to developing models. Among the most widely used physics-based water models today are the TIP3P and TIP4P models introduced in the 1980s, which employ well-established functional forms dating back to the 1930s 9 consisting of a rigid molecular geometry, fixed atomic partial charges, and Lennard-Jones interactions. |
60c74162702a9b3e6a18a234 | 1 | In recent years, several new water models have been published that are reparameterizations of the rigid TIP3P and TIP4P models; examples of this include TIP4P-Ew, 10 TIP4P/2005, 11 TIP4P/ , 12 TIP3P-FB, 13 TIP4P-FB, OPC, and OPC3. These new models more accurately reproduce a number of experimentally measured physical properties of water without increasing the computational cost of the simulation. Perhaps more importantly, some of the more recent models were developed using automated parameter optimization tools such as ForceBalance, making possible the systematic optimization of force fields for a wide range of molecular liquids given the availability of experimental data. |
60c74162702a9b3e6a18a234 | 2 | One important caveat in force field development is that the fundamental approximations in the functional form could make it impossible or inappropriate to reproduce certain physical properties. For example, it is well-known that classical models cannot reproduce the heat capacity due to the importance of nuclear quantum effects in high-frequency intramolecular and intermolecular degrees of freedom. Another relevant example is that all known simple three-point water models, i.e. those that use fixed partial charges, fail to reproduce the density anomaly at 4 • C even when they are fit to data for the temperature dependence in the density. In modeling the heat of vaporization, it is often necessary to apply post-hoc corrections to account for condensed phase polarization as well as nuclear quantum effects. |
60c74162702a9b3e6a18a234 | 3 | In the development of the SPC/E water model, the authors argued that because the atomic partial charges include the effects of mean-field polarization in the condensed phase, there exists an implicit energetic cost of polarization that should increase the potential energy of each water molecule in the liquid simulation. Thus, in order to fit the heat of vaporization, a polarization correction of +5.22 kJ mol was added to the simulated potential energy of each molecule in the liquid. The development of the TIP4P-Ew model included a polarization correction and a simple quantum correction derived from making harmonic approximations to the high-frequency vibrational modes of liquid water, as well as some more minor nonideality corrections. In summary, these corrections increase the complexity of the parameterization procedure, require additional experimental data for the compounds being parameterized, and introduce uncertainty because they only approximately model the effects they are supposed to correct. Moreover, classical force fields are not uniform in how or whether the corrections are applied; for example, the OPLS-AA force field for organic liquids was developed by fitting Monte Carlo simulated density and heat of vaporization to experiments without corrections. For these reasons, it is desirable to use physical properties that require fewer post-hoc corrections when fitting parameters to improve agreement with experiment. |
60c74162702a9b3e6a18a234 | 4 | The surface tension of the liquid/vapor interface originates from the energetic preference for molecules to be located in the bulk liquid compared to at the surface, thus it is a property that characterizes the cohesive forces in the liquid, similar to the heat of vaporization. Furthermore, because the surface tension calculation does not involve taking any energetic differences with molecules fully in the gas phase, we hypothesize that it can substitute for the heat of vaporization in the force field parameterization without requiring corrections for polarization or nuclear quantum effects. Indeed, the nuclear quantum effects are smaller for the surface tension compared to heat of vaporization, as the heat of vaporization increases by 2.1% from H 2 O (40.657 kJ mol -1 ) to D 2 O (41.521 kJ mol -1 ), while the surface tension only changes by 0.15 % between light and heavy water (from 71.98 mJ m -2 to 71.87 mJ m -2 ). This is further supported by established protocols for calculating surface tension in MD simulations, in which all post-hoc corrections are intended to account only for long-range dispersion interactions. |
60c74162702a9b3e6a18a234 | 5 | Because surface tension data is widely available and can be easily measured, there exists an opportunity to create more accurate models of a wide range of liquids by using surface tension as a training physical property instead of heat of vaporization. Nielsen first demonstrated the use of surface tension as a fitting property to parameterize a coarse-grained mixture of hydrocarbons. Salas and Alejandré 23 developed a procedure that scales the charge and Lennard-Jones parameters to reproduce the dielectric constant, surface tension, and density in stepwise fashion, and applied the approach to build all-atom and coarse grained models for four molecular liquids including methanol and ionic liquids. Martínez-Jiménez and Saint-Martin applied a similar procedure to refine a coarse-grained potential for methanol that included an off-center charge site. As for water, many popular models such as TIP3P, SPC/E, and TIP4P-Ew 10 utilize surface tension as a validation test in the sense that models fitted to some properties should accurately predict other known properties. |
60c74162702a9b3e6a18a234 | 6 | However, to the best of our knowledge, no water model has been developed by adjusting the parameters to reproduce the surface tension directly; 25 in particular, water was not one of the four liquids studied in Ref. 23. The development of such a water model is needed for testing the hypothesis that surface tension can effectively substitute for the heat of vaporization in the force field parameterization. Moreover, the utility of surface tension as reference data for force field devleopment creates a need for automated tools and procedures that can effectively use this data to generate models for molecular liquids in systematic fashion. |
60c74162702a9b3e6a18a234 | 7 | In this article, we describe how the fitting of surface tension is enabled by extending the ForceBalance optimization method to include surface tension as a fitting target. To demonstrate feasibility, we develop and characterize two new water models, namely TIP3P-ST and TIP4P-ST (here ST stands for "surface tension"), where the surface tension property replaces the heat of vaporization in the training data. The resulting TIP4P-ST model confirms our hypothesis by exhibiting high accuracy for thermodynamic properties across a range of temperatures, for both training and validation data that include the density, dielectric constant, isothermal compressibility, thermal expansion coefficient, and self-diffusion coefficient. The TIP3P-ST model offers moderate agreement with the full range of data, but reproduces the correct temperature of maximum density of water for the first time in models of this form. |
60c74162702a9b3e6a18a234 | 8 | In both cases, the optimization procedure is able to match the experimental surface tension within 10%, which is highly accurate in the context of existing water models. We conclude that when placed in the context of other fixed charge models with rigid geometries, that the four-point models will yield accurate predictions in studies involving liquid/vapor interfaces and extremes of temperature and pressure, whereas the functional form of three-point rigid models is too limited to simultaneously describe the temperature dependence of density and other structural and kinetic properties with equivalent accuracy across broad temperature ranges. The model parameterization approach of picking alternative properties such as surface tension that require minimal post-hoc corrections is also expected to be broadly useful in developing the next generation of force fields for other molecular liquids and small molecules where such corrections are not easily obtainable. |
60c74162702a9b3e6a18a234 | 9 | Five individual parameters were optimized: two weight parameters w O and w H that control the molecular geometry, the charge on hydrogen q H , and the Lennard-Jones parameters for oxygen σ LJ and LJ . In order to optimize the geometry of the rigid water model, all interactions are defined in terms of off-center interaction sites (virtual sites) whose positions r O , r H1 , r H2 are defined in terms of the rigid TIP3P molecular geometry r O , r H1 , r H2 and the weight parameters w O and w H as: |
60c74162702a9b3e6a18a234 | 10 | The same procedure was previously used to optimize the TIP3P-FB three-point model. The TIP4P-ST four-point model used the TIP4P-Ew functional form and four parameters were optimized: the virtual-site position that carries the negative charge, the hydrogen charge q H , and the Lennard-Jones parameters for oxygen σ LJ and LJ . Starting values of the parameters are given in Table . |
60c74162702a9b3e6a18a234 | 11 | where the total objective function L tot depends on the optimization variables or "mathematical parameters" k, and is equal to the sum of contributions from the parameterization targets L T (k) weighted by w T , plus a regularization term. A parameterization target consists of a collection of weighted least-squares residuals between the force field predictions and a training data set. In this study, all of the liquid thermodynamic properties including surface tension are included in a single target with a weight of 1.0. |
60c74162702a9b3e6a18a234 | 12 | In general L tot may contain many least-squares residuals, thus the objective function is organized in hierarchical fashion with each target containing ≥ 1 properties, and each property containing ≥ 1 data points. The objective function for a target is a weighted sum of contributions for one or more individual properties: |
60c74162702a9b3e6a18a234 | 13 | where λ is the index for the force field parameter being optimized, K (0) λ represents the original parameter value, and p λ is the prior width that represents the expected magnitude of variation of the parameter over the course of the optimization. Table shows the values of p λ for different parameter types. In the case of TIP3P-ST, all values of p λ were set equal to K |
60c74162702a9b3e6a18a234 | 14 | λ , which effectively makes k λ into a scaling of the original parameter. In cases where parameters need to satisfy functional relationships such as a constraint on the total charge of a residue or molecule, the parameters used directly in the energy expression may be specified as functions of K; the charge on oxygen was defined in this way as q O = -2q H . |
60c74162702a9b3e6a18a234 | 15 | Thus, increasing the prior width p λ allows the physical parameter K λ to have greater variations for the same contribution to the penalty function. Although the optimization result depends on the choice of p λ , in practice these values may be varied within a factor of 2 without incurring significant changes in the performance of the optimized model. method is used to treat the electrostatic interactions with a real-space cutoff of 9 Å, and the same cutoff was used for Lennard-Jones (LJ) interactions. The system was first equilibrated for 1 ns, followed by an 8 ns production run. Thermodynamic averages were obtained by averaging over trajectory frames spaced 0.1 ps apart for a total of 80,000 samples. |
60c74162702a9b3e6a18a234 | 16 | The surface tension γ was evaluated separately using a simulation setup consisting of a water film in the NVT ensemble with two liquid-vapor interfaces. We used a tetragonal simulation cell with dimensions 3 nm × 3 nm × 10 nm containing a 3 nm thick water layer normal to the z-dimension with 1024 water molecules in total. A real-space cutoff distance of 15 Å was chosen for nonbonded interactions because the surface tension calculations required accounting for Lennard-Jones interactions at large distances. The other simulation parameters matched the NPT simulations. To evaluate the surface tension for a trajectory frame, we adopted the test-area method with the formula |
60c74162702a9b3e6a18a234 | 17 | where E is the potential energy, β ≡ 1 k B T the inverse temperature, k B Boltzmann's constant, and T the temperature. ∆E + and ∆E -are calculated by making two perturbations to the surface area S ≡ L x L y , by ∆S = ±0.0005S as suggested in Ref. 27. In each perturbation, the x and y dimensions of the simulation box are scaled proportionally, while the z dimension is scaled in the opposite direction to keep the total volume constant. The scaling operation is also applied to the molecular centroids, and the molecules are rigidly translated without modifying the molecular geometry. The ensemble averages in the formula are evaluated as the arithmetic average over trajectory frames. |
60c74162702a9b3e6a18a234 | 18 | The procedure for evaluating surface tension was implemented into the ForceBalance automated parameter optimization software, which uses the OpenMM library to carry out the NVT and NPT MD simulations, thus allowing the entire optimization procedure to be carried out in a single reproducible calculation. Although thermodynamic fluctuation formulas were used to estimate the parametric derivatives of thermodynamic properties simulated in the NPT ensemble in previous parameterization of TIP3P-FB and TIP4P-FB models, we found that in the case of surface tension, the parametric derivatives estimated in this way contained such high levels of statistical noise that it was more efficient to calculate parametric derivatives numerically via a 3-point finite difference formula, which involved running two separate simulations for each parameter being optimized. Details of the error analysis are described in Section 3.3. |
60c74162702a9b3e6a18a234 | 19 | The gas-phase enthalpy needed for ∆H vap is simulated using a single water molecule in the NVT ensemble without periodic boundary conditions. The gas-phase simulation used a Langevin integrator with a time step of 0.5 ns and a collision frequency of 1 ps -1 , a 1.0 ns equilibration period, and a 10.0 ns producion run from which coordinates are saved every 0.1 ps (100,000 frames total). Self-polarization and nuclear quantum effect corrections for ∆H vap and nuclear quantum corrections for c P were applied following previously established procedures. These validation properties were evaluated automatically from the NPT simulations in the course of parameter optimization by setting w (T ) j = 0 in ForceBalance. |
60c74162702a9b3e6a18a234 | 20 | To evaluate the self-diffusion coefficient D 0 , we first carried out a 1 ns equilibration and 1 ns production simulation in the NPT ensemble, and saved 100 trajectory frames containing position and velocity information with 10 ps time resolution as initial conditions for energyconserving simulations. From each simulation snapshot, an energy-conserving simulation was propagated for 10 ps using the Verlet integrator and 1.0 fs time step to generate a trajectory of 100 frames with a 0.1 ps time interval. |
60c74162702a9b3e6a18a234 | 21 | The diffusion coefficient contains a known dependence on the size of the periodic box. To estimate the intrinsic diffusion coefficient at infinite box sizes, the diffusion coefficient calculation is repeated for six box sizes, 25, 30, 40, 50, 60, and 90 Å. The final self-diffusion coefficient for each temperature point is then computed as an extrapolation of the inverse box size towards infinity. The shear viscosity is also obtained from the slope of the linear fit. |
60c74162702a9b3e6a18a234 | 22 | To assist with model comparison, Figure displays the parameters of each model in terms of percentage differences from TIP3P and TIP4P-Ew for 3-point and 4-point models respectively. In both 3-point and 4-point models, the σ LJ parameter has the least variation among the three models, which may be expected given its important role in determining the excluded volume, and consequently the liquid density. TIP3P-FB and TIP3P-ST feature larger values of r OH , Θ HOH and q H compared to TIP3P, consistent with increasing the hydrogen bonding strength by decreasing the intermolecular O-H distance, increasing the Coulomb interaction strength, and bringing the bond angle closer to the ideal tetrahedral angle. The parameter with the largest variation is LJ where TIP3P-ST has a smaller value than the other three models. Among the other four parameters, the TIP3P-ST and TIP3P-FB parameter are closer, though we note the former has a slightly higher value of r OH . This indicates TIP3P-ST has stronger directional character in its intermolecular interactions, and could be further understood by examining the thermodynamic properties. On the other hand, the TIP4P-FB and TIP4P-ST are highly similar in the q H , σ LJ and LJ parameters, and both models place the virtual site closer to the O atom than TIP4P-Ew. The value of ω v in TIP4P-ST is smaller than TIP4P-FB, but the accuracy of these two models are highly similar. |
60c74162702a9b3e6a18a234 | 23 | The comparison of thermodynamic properties at room temperature and standard pressure for six models vs. experiment are listed in Table . The temperature dependence of fitted thermodynamic properties are plotted in Figure , while the validation properties are plotted in Figure . The three-point TIP3P-ST model accurately reproduces experimental thermodynamic properties with a level of accuracy that well exceeds the widely adopted TIP3P model. Notably, TIP3P-ST correctly reproduces the temperature of maximum density, which could not be accomplished by the other rigid three-point water models in our comparisons. The closer agreement with the experimental density curve in TIP3P-ST is a significant difference from TIP3P-FB, and possibly caused by stronger directional interactions resulting from reduced LJ and increased r OH . TIP3P-ST reproduces the experimental thermal expansion coefficient and surface tension more closely than TIP3P-FB, but also has a lower self-diffusion coefficient and higher viscosity compared to experiment. The four-point TIP4P-ST model agrees within 5% of the experimental value for most properties, and the fitted surface tension is surprisingly close to the TIP4P-FB model which did not include surface tension in the fitting targets. Generally speaking, the performance of TIP4P-ST is nearly identical to TIP4P-FB, except that TIP4P-ST achieves an even closer fit to the density amounting to < 0.1% deviations across the whole temperature range. Isobaric heat capacity; D 0 : Self-diffusion Coefficient; η: shear viscosity; TMD: Temperature of maximum density. The TMD for TIP3P model was from reference. The validation properties provide insights into the predictive power of models fitted to surface tension. The TIP3P-ST model yields a higher ∆H vap than experiment, and also has a relatively large self-polarization correction of 7.38 kJ mol -1 indicating a large dipole moment. |
60c74162702a9b3e6a18a234 | 24 | The TIP4P-ST model also predicts ∆H vap slightly higher than the experimental value with a self-polarization correction of 7.03 kJ mol -1 . Notably, the corrected ∆H vap is almost identical to TIP4P-FB, again indicating they are close in the model space. The correction for nuclear quantum effects is relatively small at -0.27 kJ mol -1 at 298 K. The self-diffusion coefficient is another property where TIP3P-ST is different from the other models included in our comparison. The lower self-diffusion coefficient indicates a slightly more structured liquid, with stronger hydrogen bonds needed to reproduce the surface tension. This behavior is also reflected in the radial distribution plot, where the TIP3P-ST curve shows a higher first peak and lower first trough. |
60c74162702a9b3e6a18a234 | 25 | There is a notable trend in the rigid three-point water models where TIP3P-ST has the highest surface tension, temperature of maximum density and heat of vaporization, the most highly structured O-O RDF, and the lowest self-diffusion coefficient. All of these properties correspond to stronger cohesion and a highly structured hydrogen-bonding network. TIP3P on the other hand has the lowest surface tension, temperature of maximum density and heat of vaporization, the least structured O-O RDF, and the highest self-diffusion coefficient, whereas TIP3P-FB is intermediate between TIP3P-ST and TIP3P for all of these properties. The physically motivated correspondence between all of these properties, coupled with the observation that none of the rigid three-point models can reproduce all of the experimental properties equally accurately across the whole temperature range, reveals a potential limitation of the functional form of rigid three-point rigid water models. Despite these limitations, the high accuracy of TIP3P-FB for all tested thermodynamic, structural and kinetic properties except for the temperature dependence of the density (Table ) indicates that it is suitable for simulating biomolecular systems near ambient conditions, especially in applications that benefit from the lower computational cost of three-point models. |
60c74162702a9b3e6a18a234 | 26 | The four-point models have one fewer parameter because the molecular geometry is not being optimized, but more accurate results are obtained for the validation properties; in particular, the diffusion coefficient of TIP4P-ST agrees closely with experiment, and the O-O radial distribution function of TIP4P-ST agrees with experiment at a similar level as the TIP4P-Ew, TIP3P-FB and TIP4P-FB models. The improved ability of four-point models to reproduce experimental properties has previously been attributed to the model's ability to predict the correct quadrupole moment of the water molecule. |
60c74162702a9b3e6a18a234 | 27 | Recent studies on electrostatic models have raised questions regarding whether the simulated dielectric constant requires post-hoc corrections. These studies posit that the effective electrostatic moments used to compute the MM interactions of ions and polar species should be reduced with respect to the physical charges used to compute electrostatic properties, due to the dielectric screening caused by the electronic polarization of the medium. This implies that the dielectric constant computed from the partial charges in the force field should be increased by a correction prior to comparing with experiment, or conversely, the experimental value should be reduced prior to making the comparison with the force field. |
60c74162702a9b3e6a18a234 | 28 | In Reference 31, the authors concluded that the missing polarizability in non-polarizable models scales the dielectric constant by a factor of 1.78. Under the assumption that the same correction factor would apply to our models, the reference dielectric constant should be reduced by a factor of 1/1.78 = 0.5618. Here we test the effective charge hypothesis by reducing the reference dielectric constants by a factor of 0.56 in the fitting of three-point water models. If the effective charge hypothesis is correct, then we expect the model fitted to a reduced dielectric constant should produce improved agreement with experiment for validation properties. Isobaric heat capacity; D 0 : Self-diffusion Coefficient; η: shear viscosity; TMD: Temperature of maximum density. The TMD for TIP3P model was from reference. The comparison of optimized parameters between TIP3P-ST and the model fitted to reduced dielectric constant, denoted, as TIP3P-ST-0.56 (0), is shown in Table . A main difference is that the atomic charges q H increase and the H-O-H angle widens to accommodate the reduced dielectric constants. The molecular dipole moment is 2.24 D and the selfpolarization correction is smaller at 2.951kJ mol -1 , compared to TIP3P-ST which has a dipole moment of 2.46 D and self-polarization correction of 7.498kJ mol -1 . Table shows the effect on preperty predictions by reducing the reference dielectric constant. The TIP3P-ST-0.56 (0) model is able to reach similar levels of agreement with experiment as the original TIP3P-ST. The heat of vaporization increases further with respect to both experiment and TIP3P-ST. These observations support our earlier assertion that the quality of fitting for dielectric constants mainly depends on the molecular structure parameters and does not have major impact on the ability to fit other thermodynamic properties. However, due to the mixed results in relative accuracy of the models fitted to the original and reduced dielectric constants, we cannot conclude from this study whether correction of the dielectric constant is necessary in general. |
60c74162702a9b3e6a18a234 | 29 | The accurate computation of gradients of simulated thermodynamic properties with respect to force field parameters is highly important for efficient model optimization. The thermodynamic property being differentiated contains statistical noise due to the finite length of the simulation, so we expect the parametric gradients to contain statistical noise as well. Moreover, different methods for computing the parametric gradient may exhibit varying levels of statistical noise for the same computational resources used in the calculation. Thus, we decided to compare the statistical noise in the surface tension gradients for two calculation methods: "semi-analytic" (i.e. the property gradient is computed from a thermodynamic fluctuation formula using finite-difference potential energy gradients), and "pure numerical" (i.e. by running separate simulations for each parameter). |
60c74162702a9b3e6a18a234 | 30 | Here, the new terms in the formula ∂E + ∂k and ∂E - ∂k may be recognized as the potential energy derivatives of the surface-perturbed trajectory frames. These potential derivatives are evaluated numerically using sufficiently small steps in k to avoid incurring machineprecision errors. All quantities in angle brackets representing ensemble averages are then evaluated as arithmetic averages over the trajectory frames. The computational cost of evaluating the full set of potential energy derivatives scales linearly with the number of parameters, and the added cost per parameter is significantly less than the simulation itself. |
60c74162702a9b3e6a18a234 | 31 | In practice, the calculation of a single gradient element is roughly equal to 20% of the original MD simulation. On the other hand, pure numerical gradients of the surface tension involve running separate simulations where the parameter is perturbed by a small step, and repeating this procedure for each parameter being optimized. We used a central difference approximation, which implies the computational cost of the gradient is 2N param times the cost of simulating the property itself. Compared to the semi-analytic gradients, the numerical gradients involve running separate simulations with nearly fully independent samples (save for the same initial condition). The noise in the gradients also increases with decreasing parameter step size because the statistical error in the property is roughly independent of parameter size, resulting in large numerical errors for steps that are too small. It is also important to avoid step sizes that are too large and no longer within the linear regime. |
60c74162702a9b3e6a18a234 | 32 | Figure compares the accuracy of the semi-analytic and numerical methods with a fixed simulation run length. The mean and standard error for each gradient is computed from five independent runs using the TIP3P parameters with a simulation length of 20ns. Finite difference step sizes of δk λ = 0.001, 0.01 and 0.1 in the mathematical parameters were tested for both methods. When numeric gradients were used, the statistical errors were largest for step sizes of 0.001 and smallest for 0.01. For σ LJ , increasing the step size to 0.1 resulted in a different mean and larger standard error, indicating this step size was outside the linear regime; we did not observe this for q H and LJ . The semi-analytic gradients are computationally less costly but also have higher uncertainty than the numerical gradients, thus we concluded that numerical gradients with a step size of 0.01 provide the most statistically precise surface tension gradients for a fixed simulation length. These conclusions are based on our choice of the prior widths for the parameters; for a different choice of prior width, the recommended step size may be obtained by ensuring the step size in physical parameters, i.e. δK λ = δk λ p λ matches the current results. Taking the total computational cost into account, we compared the numeric gradients from 5 ns simulation with the semi-analytic gradients from 20 ns simulation, both with the optimal step size 0.01, as shown in Figure . Two sets of parameters were used, namely the TIP3P parameters and the TIP3P-ST parameters. The results show that for the TIP3P parameters, the numeric and semi-analytic gradients agree relatively well with comparable standard errors. However, when evaluated at the final TIP3P-ST parameters, the semi-analytic gradients have larger errors than the numeric gradients for the q H and σ LJ parameters, while LJ exhibits the opposite behavior. The small errors for the semi-analytic gradients of LJ may be due to the intrinsically small value of the gradient (i.e. in the limit of infinite simulation time). An intrinsically small gradient would reduce the error bars of the analytic gradient but not the numerical gradient, as the latter contains statistical noise from independent estimations of the surface tension and contributes a constant term to the error. The scale-independent behavior of the numerical gradient error is confirmed by comparing the standard error across parameters; for TIP3P these errors are (15.2, 18.0, 13.3) for q H , σ LJ , LJ respectively, and for TIP3P-ST the errors are (59.0, 32.9, 41.4). The standard error for surface tension gradients are larger overall for TIP3P-ST compared to TIP3P, which may be due to the slower dynamics of the model causing slower convergence of the property. Based on our observation that the statistical errors were mostly smaller using numeric gradients, we decided to use numeric gradients for optimizing the water models in this study. |
60c74162702a9b3e6a18a234 | 33 | We evaluated the surface tension using the test-area method, for TIP3P at 298 K, 1.0 atm, using various van der Waals cutoff distances. Figure shows that as the cutoff distance increases, the simulated surface tension continues to increase even at the distance of 18 Å. To estimate the surface tension at infinite cutoff distance, we performed a linear extrapolation on the surface tension against the inverse of the cutoff value, as shown in Figure . The final value of 53.87 mJ m -2 obtained from the intercept of the linear extrapolation is about 4 mJ m -2 higher than the value computed with cutoff at 15 Å. This indicates that our previously shown surface tension evaluated with the cutoff distance of 15 Å may be underestimated by a small amount. Since our simulated value in Figure were below the experimental value, a larger cutoff brings the theoretical surface tension into a closer agreement. In Table , the surface tension extrapolated to infinite cutoff are reported as γ ∞ . The TIP3P-ST model achieved the best agreement of within 1 mJ m -2 to the experiment. However, since the extrapolation is computationally time-consuming, we decided to use the 15 Å cutoff distance through out our parameterization. As a result, the weight of the surface tension property |
60c74162702a9b3e6a18a234 | 34 | The overall results point to the validity of using surface tension as a replacement for heat of vaporization in force field development. Both water models correctly reproduce the temperature of maximum density, which in particular is notable for the three-point model, TIP3P-ST, because models of this functional form have had difficulty in accurately reproducing the density anomaly at ambient pressures. Whereas TIP4P-ST can accurately reproduce a broader range of kinetic and structural properties, consistent with more recent well-optimized 4-point rigid models, the TIP3P-ST generalizes more poorly to the validation set by producing somewhat over-structured radial distribution functions and lower diffusion coefficients. This indicates that rigid 3-point models need to make a compromise between accurate depictions of cohesion vs. structural and kinetic properties due to their limited functional form. We additionally found that the dielectric constant could be independently adjusted without impacting the quality of fit of other training parameters, leading to differences in the molecular geometry and mixed impacts on the validation properties. |
60c74162702a9b3e6a18a234 | 35 | Recent work by Milne and Jorge suggests that polarization corrections of the form utilized by Berendsen, this work, and many others is unnecessary -and perhaps undesirable -in order to reproduce experimental observables such as the enthalpy of vaporization and hydration free energy of water and other polar liquids. Interestingly, our results suggest that when these properties are not used in the parameterization of the water model, the resulting enthalpy of vaporization will still be significantly greater than the experimentally measured quantity. Specifically we note that the enthalpies of vaporization of TIP3P-ST and TIP4P-ST are both somewhat greater than experiment (by approximately 0.57 and 0.33 kcal/mol, respectively) even after correction. If the polarization correction were not included, then the simulated ∆H vap would be even more positive and further increase disagreement with experiment, as the polarization correction for moving from the condensed phase to the gas phase is always favorable. We are optimistic that the procedure described in this study can be applied broadly to develop future generations of force fields for organic liquids and the nonbonded energy terms in biomolecular and general small molecule force fields. |
6450f5f207c3f02937167e86 | 0 | A tumor contains various kinds of cancer cells and non-cancer cells including immune cells, stromal cells, and in close vicinity, microbiotas . There are strong suggestions that intracellular and intercellular shuttling of lactate may be required to meet the diverse metabolic, signaling, and epigenetic regulations of cancer and cancer associated cells . |
6450f5f207c3f02937167e86 | 1 | Lactate generated by cancer cells may be shuttled to the extracellular milieu so that lactate could influence metabolic adaptations by macrophages and microbiotas . Besides the clear role of lactate in the metabolic rewiring of cancer cells and cancer-associated cells such as pro-inflammatory macrophages, there is a clear gap in our understanding of lactate-mediated non-metabolic implications including lactylation form of epigenetic modifications of histones. |
6450f5f207c3f02937167e86 | 2 | The coherence between metabolic and epigenomic process that supports the various tumor hallmarks including proliferation, invasiveness, metastasis and drug resistance is being unraveled . Emerging observations support the role of lactate as a non-metabolic signaling molecule that promotes lactylation and post-translational modification of a lysine residue on histone and other target proteins such as PKM2 and beta-catenin . At the same time, lactylation is reported in certain types of cells including macrophages, and is linked to the transcriptional changes that help the M1 macrophages to be changed into M2 macrophages . |
6450f5f207c3f02937167e86 | 3 | In essence, M2 macrophages are shown to share similar attributes to that of proinflammatory and tumor-associated macrophages . A view on the role of extracellular lactate as a metabolic fuel by microbiotas is proposed. In this way, an indirect role of lactylation in cancer growth and proliferation is proposed by activating M1 macrophages in the tumor microenvironment which is in turn transformed into pro-tumor M2 macrophages . Among many distinctive features of M1 macrophages, acetyl-CoA synthetase is highly active and overexpressed which is known to catalyze the formation of acetyl-CoA from acetate . |
6450f5f207c3f02937167e86 | 4 | However, the role of acetyl-CoA synthetase in lactylation is not explored. The constraints in the detection of lactate and lactate-derived metabolic products could be due to the highly shuttled and dynamic nature of lactate in the tumor microenvironment. Therefore, a link that can connect the generation of lactyl-CoA from lactate by specific enzymes of cancer cells, tumor-associated macrophages, and microbiotas in the tumor niche is not evident. |
6450f5f207c3f02937167e86 | 5 | In spite of increasing understanding on the novel evidence on the lactylation in various cellular components of tumor microenvironment, there is a gap regarding the nature of the enzymes may use lactyl-CoA for lactylation and report of lactyl-CoA in the physiological setting. At the same time, the possibility of the involvement of histone acetyltransferase (HAT) p300 which is well-known as a regulator of chromatin remodeling is indirectly proposed, but not explored . |
6450f5f207c3f02937167e86 | 6 | In this study, we aimed to investigate lactate excretion in the extracellular conditioned medium of cancer cells treated with anticancer compositions. To explore the missing link in lactylation, we utilized molecular docking and molecular dynamics (MD) simulations to identify potential enzymes capable of generating lactyl-CoA from lactate. Additionally, we conducted molecular interaction studies to determine the binding affinities of lactate and its derived products with HAT p300. Our findings shed light on the significant gaps in our understanding of lactylation, particularly regarding the enzyme that can catalyze lactylation using lactyl-CoA. |
6450f5f207c3f02937167e86 | 7 | Cell culture reagents were purchased from Invitrogen India Pvt. Ltd. and Himedia Laboratories Pvt. Ltd. MCF-7 breast cancer cells were procured from the National Centre of Cell Science (NCCS) Pune, India. DMSO, agarose, acrylamide, and other chemicals were of molecular biology grade and obtained from Himedia Laboratories Pvt. Ltd. India and Merck India Pvt. Ltd. |
6450f5f207c3f02937167e86 | 8 | The extracellular conditioned medium of MCF-7 breast cancer cells treated by DMSO and cow urine DMSO fraction (CUDF) (50 µg/ml) final concentration was collected as per the previously published procedure . Then, 750 µl of extracellular conditioned medium was mixed with 4X loading buffer (0.5 M Tris, pH 6.8, and Glycerol). Next, the conditioned medium along with the loading buffer was loaded on the vertical tube gel electrophoresis (VTGE) purification system with a matrix of 15% polyacrylamide gel (acrylamide: bisacrylamide, 30:1) . The fractionated extracellular metabolites were collected in the running buffer (96 mM glycine, pH 8.3). The detailed procedure was adopted from previously published in-house VTGE-assisted purification of metabolites . Furthermore, LC-HRMS analyses of VTGE-purified extracellular metabolite elutes were performed by Agilent TOF/Q-TOF Mass Spectrometer station Dual AJS ESI ion source. During LC separation, RPC18 Hypersil GOLD C18 100 x 2.1 mm-3 µm and mobile phase of 100% Water (0.1% FA in water) and 100% Acetonitrile (90% ACN +10% H2O+ 0.1% FA) were used in the proportion of 95% and 5% (40). Mass spectrometry was performed in a positive mode and analyzed as per the procedure adopted from previously reported methodology . |
6450f5f207c3f02937167e86 | 9 | Potential oncometabolites including lactate (PubChem CID: 91435), acetate (PubChem CID: 175), lactyl-CoA (PubChem CID: 3081970), and acetyl-CoA (PubChem CID-444493) were retrieved from the database as ligands for molecular docking. The PubChem database was used to download the structure of ligands in SDF format. Then conversion of ligands into PDB format took place using the software OpenBable. Before performing molecular docking, both ligands were energy minimized to obtain stable conformation using Avogadro software with the steepest descent method and MMFF94s force field. Protein Data Bank (PDB) () was used to download the receptor protein. Here, HAT p300 (PDB ID-6GYR) and acetyl-CoA synthetase (PDB ID: 2P2F) were taken as target proteins. Hetatoms are removed from the protein before performing docking. This protein was subjected to the AutoDock Tool 4.2. to perform the steps of protein preparation, which include the removal of water molecules, bond correction, assigning AD4 type atoms, adding polar hydrogens, and adding Kollman charges . AutoDock Vina Software was used to perform molecular docking . After the successful docking, confirmation of the binding position of oncometabolite into the receptor protein and calculation of bond distance has been done by Discovery Studio Visualizer v3.0 (DSV3) and Accelrys software . |
6450f5f207c3f02937167e86 | 10 | The 20ns Molecular Dynamics (MD) simulation of metabolite lactate (PubChem CID: 91435), and acetate (PubChem CID: 175) with acetyl-CoA synthetase (PDB ID: 2P2F) was performed with the help of Desmond software to confirm the binding stability and strength of the complex (48). Desmond has inbuilt functions to add pressure, volume system, temperature, and many functionalities to accomplish protein-ligand binding. Ligand-protein complex was plunged into a water-filled orthorhombic box of 10 Å spacing. The conformational changes upon binding of ligands with acetyl-CoA synthetase were recorded by using the 1000 trajectories frames generated during the 20ns MD simulation and the Root Mean Square Deviation (RMSD) was calculated to reveal the binding stability of lactate and acetate. |
6450f5f207c3f02937167e86 | 11 | The relevance of intracellular, extracellular, and inter-tissue lactate as metabolic fuel and signaling molecules that link the epigenetic-metabolic axis in cancer is considered challenging . The reason behind the gaps in understanding the role of lactate and its derived metabolic products such as lactyl-CoA is the lack of detection of lactate and lactyl-CoA at the physiological concentration at intracellular, extracellular, and inter-tissue levels. The lack of clear evidence on the levels of lactate and lactyl-CoA could be linked with constraints such as specific and efficient metabolite profiling methodologies and the highly shuttling and diffusive nature of lactate and lactyl-CoA in the tumor microenvironment. |
6450f5f207c3f02937167e86 | 12 | In this regard, we have attempted to detect the levels of lactate and lactyl-CoA at the intracellular and extracellular levels of breast cancer cells treated by DMSO and anticancer drug compositions enriched with free fatty acids and tripeptides . The use of anticancer drug composition CUDF is taken as one of the candidate drug models to see the difference in the detection of lactate and lactyl-CoA compared to the DMSO control. Extracellular metabolite profiling of MCF-7 breast cancer cells suggested surprising observations that lactate was not detected in DMSO control (Figure ). At the same time, CUDF-treated MCF-7 breast cancer cells previously reported for inducing cell death indicated the presence of lactate (m/z 89.0239, RT-1.483, mass-90.0131) in the LC-HRMS derived total ion chromatogram of extracellular conditioned medium (Figure ). The detailed MS and MS/MS spectra displayed clear evidence of lactate-specific negative ESI fragment ion spectra such as 89.0235 and 96.9594. Interestingly, LC-HRMS profiling of VTGE-purified intracellular metabolites did not show a detectable level of lactate and lactyl-CoA. At the same time, extracellular profiling showed the presence of lactate in the case of breast cancer cells under drug-induced stress and cell death (Figure ). But, we did not detect traces of lactyl-CoA in the case of both DMSO and CUDF-treated breast cancer cells. |
6450f5f207c3f02937167e86 | 13 | In a quest to find the relevance of lactate as an extracellular signaling molecule, we found that acetyl-CoA synthetase is known to promote pro-tumor phenotype in macrophages. Also, acetyl-CoA synthetase is known for the metabolic activities in bacterial cells that could be linked with the microbiotas in the niche of the tumor microenvironment. By employing molecular interaction studies, data indicated that lactate (Figure ) occupies the same substrate site as in the case of acetate (Figure ). The binding affinity and interactions at the active sites of acetyl-CoA synthetase are projected to be almost similar from lactate (Figure ) over acetate (figure ) in terms of residues such as TRP413, TRP414, GLN415, ARG515, ASN512, and ARG526. |
6450f5f207c3f02937167e86 | 14 | Besides a missing link on the suitable enzyme that performs the lactylation process, a debate is pertinent on the existence of an enzyme that converts lactate to lactyl-CoA. By using molecular docking studies, we have screened various potential enzymes as potential players that may catalyze the formation of lactyl-CoA from the lactate (data not shown). During the screening, a key enzyme is predicated as acetyl-CoA synthetase that may potentially prefer the lactate as a substrate during the formation of lactyl-CoA that is similar to the acetyl-CoA formation from acetate (Figure ). Data obtained from DSV3 predicted the almost similar binding pockets with hydrogen bonds between acetate (GLN415, ARG515, ARG526) and lactate TRP413, TRP414, GLN415, ARG515, ASN521) for acetyl-CoA synthetase (Figure and Table ). |
6450f5f207c3f02937167e86 | 15 | To analyze the stability of the lactate-acetyl CoA complex, the MD simulations were carried out for a length of 20 ns. For comparison, a known substrate acetate is selected as a positive control for MD simulations of acetyl-CoA synthetase. We have analyzed the conformation of the protein-ligand complex obtained during the simulation period of 20ns. Root mean square deviation (RMSD) was calculated during the simulation trajectory of 20ns for the ligands such as lactate and acetate against acetyl-CoA synthetase. The RMSD evolution plot of acetyl-CoA synthetase on the Y-axis suggests that values (1.2 to 2.4 Å) for lactate (Figure ) is within the well-accepted range, 1-3 Å and almost similar to acetate (Figure ) RMSD value (0.9 to 2.7 Å) Furthermore, it is important to note that simulations are converged and appear to be stabilized at the end of simulations of length between 12 to 20 ns for lactate and a known substrate of acetyl-CoA synthetase. |
6450f5f207c3f02937167e86 | 16 | Next, root mean square fluctuation (RMSF) values of acetyl-CoA synthetase in complex with lactate suggested a well-acceptable range of fluctuations for lactate (0.4 to 1.8 Å) and acetate (0.5 to 1.6 Å) specifically in the region spanning from 280 to 550 amino acid residues (Figure and). These residues from 280 to 550 position of acetyl-CoA synthetase is known for substrate binding and catalytic activity. The MD simulations explain the interaction fraction of the acetyl-CoA enzyme residues with the lactate and acetate, which means how much % of the simulation time the specific interactions of these residues maintained ligand-protein complexes. |
6450f5f207c3f02937167e86 | 17 | Concerning the recent discovery of an epigenetic lactylation process, there is a substantial gap between the substrate and the enzyme which is known to play a vital role in lactylation for histone modification. However, suggestions are put forth on the involvement of lactate-derived metabolites such as lactyl-CoA during the lactylation process . |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.